content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
\begin{figure}[!htbp]
\begin{center}
\begin{tabular}{cccc}
\hspace{-8mm} \psfig{figure=.//toy_points.eps,width=1.0in} &
\hspace{-8mm} \psfig{figure=.//toy_points_adaboost_100_pos.eps,width=1.0in} &
\hspace{-8mm} \psfig{figure=.//toy_ada_ada.eps,width=1.0in} &
\hspace{-8mm} \psfig{figure=.//toy_points_adaor_10_pos.eps,width=1.0in} \\
(a) & (b) & (c) & (d) \\
\end{tabular}
\end{center}
\vspace{-3mm}
\caption{The xor problem. (a) shows the positive and negative points.
(b) are the points classified as positives by the AdaBoost algorithm using
100 decision stump weak classifiers. (c) shows
points classified as positives by the Ada-Ada algorithm which will be
discussed later. (d) shows the positive points classified
by the Ada-Or classifier using 10 OrBoost weak classifiers.}
\label{fig:toy}
\end{figure}
Classification algorithms such as
decision tree~\cite{C4.5,CART}, neural networks~\cite{Bishop},
support vector machine (SVM) have been widely used in many areas.
The classification procedure in many of these algorithms can be understood as performing
reasoning using logic operators (and, or, not), with deterministic or probabilistic formulations.
The recent development of the AdaBoost algorithm~\cite{AdaBoost}
has particularly advanced the performance of many applications in the field.
We focus on the AdaBoost algorithm in this paper (also called
boosting together with its variations~\cite{Breiman,Friedman98}).
Boosting algorithms have many advantages over the traditional classification algorithms.
Its asymptotical behavior when combining a large number of weak classifiers
is less prone to the overfitting problem.
Once trained, a boosting algorithm performs weighted sum on the
selected weak classifiers. This linear summation weakly performs
the `and' and `or' operations. In the discrete case, as long as the overall score
is above the threshold, a pattern is considered as positive. This may include
a combinotory combinations of the conditions. Some weak classifiers
may require to be satisfied together (`and'), and some may not as long a subset answer yes (`or').
In the literature, decision stump has been widely used as weak classifier
due to its speed and small complexity.
However, decision stump does not have strong discrimination power. A comprehensive empirical study
for a wide variety of classifiers including SVM, Boosting (using decision-tree and decision-stump),
neural networks, and nearest neighboorhood, was reported in \cite{Caruana}.
Each decision stump corresponds to a thresholded feature ($\ge$ is changeable to $<$):
\begin{equation}
h(F_j(x),tr)= \left\{ \begin{array}{ll}
+1 & \textrm{if } F_j(x) \ge tr \\
-1 & \textrm{otherwise} \\
\end{array} \right.
\end{equation}
We call stump based AdaBoost algorithm {\em Ada-Stump} for the remainder of this paper.
Fig.~(\ref{fig:toy}.b) displays a failure example of the Ada-Stump.
We see that it can not deal with the `xor' patterns, even with 100 stumps.
One solution to this problem is to adopt more powerful weak classifiers, such
as decision tree, to the boosting algorithm. It was proposed by several authors~\cite{Friedman98,Reyzin} and
we call it Ada-Tree here for notional convenience (it is different from the AdaTree method~\cite{AdaTree}).
However, using decision tree greatly increases the time and computational complexity of the boosting algorithm.
Many vision applications were trained on very large datasets with each sample having
thousands or even millions features~\cite{Viola}.
This limits the use of decision tree or CART, and Ada-Stump remains
mostly used in vision~\cite{Viola}.
In this paper, we show that Ada-Stump intrinsically can not deal with the `xor' problem.
We propose layered logic models for classification, namely {\em Ada-Or}, {\em Ada-And}, and {\em Ada-AndOr}.
The algorithm has several interesting properties:
(1) it naturally incorporates the `and', `or', and `not' relations in the algorithm; (2) it has much
more discrimination power than Ada-Stump; (3) it has much smaller computational complexity than tree
based AdaBoost with only slightly degraded classification performance.
A recent effort to combine `and' and `or' in AdaBoost has been proposed in~\cite{Dundar}. However, the
`and' and `or' relations are not naturally embedded in the algorithm and it requires very complex optimization
procedure in training. How the algorithm can be used for general tasks in machine learning
and computer vision is at best unclear.
We apply the proposed models, Ada-Or, Ada-And, and Ada-AndOr, on several typical datasets from the
Irvine repository and two challenging vision applications, object segmentation and pedestrian detection.
Among the models, Ada-AndOr performs the best nearly in all cases.
We observe significant improvements on all the datasets over Ada-Stump.
For pedestrian detection, the performance of Ada-AndOr is very close to HOG~\cite{HOG}
using simple Haar features, though the main objective of this paper
is not to develop a pedestrian detector.
\section{AdaBoost algorithm \label{sect:adaboost}}
In this section, we briefly review the AdaBoost algorithm and explain
why Ada-Stump fails on the `xor' problem.
\subsection{Algorithms and theory \label{sect:theory}}
Let $\{(x_i,y_i,D_1(i)), i=1...N\}$ be a set of training samples and $D_1(i)$ is the distribution
for each sample $x_i$. AdaBoost algorithm~\cite{AdaBoost} proposed by Freund and Schapire
learns a strong classifier $H(x)=sign(\sum_{t=1}^T \alpha_t h_t(x))$, based on the training set,
by sequentially combining a number of weak classifiers.
We briefly give the general AdaBoost algorithm\cite{AdaBoost} below:
\begin{figure}[!htb]
\begin{center}
\begin{tabular}{|c|}
\hline
\begin{minipage}[c]{80mm}{
{\footnotesize
Given: $(x_1,y_1,D_1(1)),...,(x_N,y_N,D_1(N));\; y_i \in \{-1, 1 \}$
For $t=1,...,T:$
\begin{itemize}
\item[$\bullet$] Train weak classifier using distribution $D_t$.
\item[$\bullet$] Get weak hypothesis $h_t: \chi \to \{-1,+1 \}$.
\item[$\bullet$] Calculate the error of $h_t: \epsilon_t=\sum_{i=1}^N D_t(i) {\bf 1}_{(y_i \ne h_t(x_i))}$.
\item[$\bullet$] Compute $\alpha_t=-\log \epsilon_t/(1-\epsilon_t)$.
\item[$\bullet$] Update: $D_{t+1}(i) \leftarrow D_t(i) \cdot \exp (-\alpha_t y_i h_t(x_i))$ with $\sum_i D_{t+1}(i)=1$.
\end{itemize}
Output the the strong classifier: $H(x)=sign(\sum_{t=1}^T \alpha_t h_t(x))$.
}
}\end{minipage} \\
\hline
\end{tabular}
\end{center}
\caption{Discrete AdaBoost algorithm. ${\bf 1}$ is an indicator function. The variations to the
AdaBoost algorithms such as arc-gv~\cite{Breiman}, RealBoost and GentalBoost~\cite{Friedman98}
differ mostly from the way $h_t$ and $\alpha_t$ are computed.}
\label{fig:AdaBoost}
\end{figure}
The AdaBoost algorithm minimizes the total error $\sum_i e^{-\sum_{t=1}^T \alpha_t y_i h_t(x_i)}$
by sequentially selecting $h_t$ and computing $\alpha_t$ in a greedy manner. At each step, it
is to minimize
\[
E_t = \sum_i D_t(i) e^{- \alpha_t y_i h_t(x_i)}
\]
by coordinate descent:
(1) Select the best weak classifier from the candidate pool which minimizes $E_t$.
(2) Compute $\alpha_t$ by taking $\frac{d E_t} {d \alpha_t}=0$, which yields
\[
\alpha_t=-\log \epsilon_t/(1-\epsilon_t).
\]
A very important property of AdaBoost is that after a certain number
rounds, the test error still
goes down even the training error is not improving~\cite{BreimanArc}. This makes AdaBoost
less prone to the overfitting problem than many other classifiers.
Schapire et al.~\cite{Margin} explained
this behavior of AdaBoost from the margin theory perspective. For any data $(x,y)$,
\[
margin(x,y) = \frac{y \sum_t \alpha_t h_t(x)} {\sum_t \alpha_t}.
\]
$margin(x,y)$ essentially gives the confidence of the estimation $y$ to $x$.
For any given $\theta$, the overall test error is bounded by
\begin{equation}
error_{test} (x, y) \le p[margin(x,y) \le \theta] + \tilde{O}(\sqrt{ \frac{d}{m \theta^2}}),
\label{eq:margin}
\end{equation}
where $d$ is the VC dimension of the weak classifier and $m$ is the number of training
samples. Eqn.~(\ref{eq:margin}) shows three directions to reduce the test error:
(1) increase the margin (related to training error but not exactly the same);
(2) reduce the complexity of the weak classifier;
(3) increase the size of training data.
Moreover, it is shown~\cite{Friedman98} that AdaBoost and its variations are asymptotically approaching
the posterior distribution (there are still some debates about
this probabilistic formulation of the AdaBoost algorithm).
\begin{equation}
p(y|x) = \frac{e^{2 y \sum_t \alpha_t h_t(x)}}{1+ e^{2 y \sum_t \alpha_t h_t(x)}}.
\label{eq:logistic}
\end{equation}
The margin is directly tied to the discriminative probability.
\subsection{The xor problem}
It is well-known that the points shown in Fig.~(\ref{fig:toy}.a) as `xor' are not
linearly separable. The red and blue points are the positive
and negative samples respectively. Each weak classifier makes
a decision whether a point lies above or below a line passing the original.
Using this type of weak classifier, the AdaBoost algorithm is not able to
separate the red points from the blue ones. It is easy to verify.
For any positive sample $(x_1, x_2)$ with $H(x_1,x_2)=\sum_{t=1}^T \alpha_t h_t(x_1, x_2) > 0$,
then $(-x_1, -x_2)$ is a positive sample also. However
\[
h_t(-x_1, -x_2) = -h_t(x_1, x_2), \forall h_t
\]
and therefore $H(-x_1,-x_2)<0$.
\begin{figure}[!htbp]
\begin{center}
\begin{tabular}{c}
\psfig{figure=.//ada_illustration.eps,width=3.3in} \\
\end{tabular}
\begin{tabular}{cc}
(a) initial weights & (b) re-weighted \\
\end{tabular}
\end{center}
\caption{The re-weighting scheme in the AdaBoost to cause a deadlock.}
\label{fig:reweight}
\end{figure}
Let
\[
A = \{(x_1,x_2), x_1>0 \} \; and \; \bar{A} = \{(x_1,x_2), x_1<0 \}, and
\]
\[
B = \{(x_1,x_2), x_2>0 \} \; and \; \bar{B} = \{(x_1,x_2), x_2<0 \}.
\]
We denote $\land$, $\lor$, $\bar{}$ as the `and', `or', and `not' operations respectively.
Thus, the positive samples in Fig.~(\ref{fig:toy}.a) can be denoted by
\[
(A \land B) \lor (\bar{A} \land \bar{B}), \; or \; (A \lor \bar{B}) \land (\bar{A} \lor B).
\]
One of the key properties in AdaBoost is that it re-weights
the training samples after each round by giving higher weights
to those which were not correctly classified by the previous
weak classifiers. We take a close look at the re-weighting scheme to
the points in Fig.~(\ref{fig:toy}.a).
Initially, all the samples
receive equal weights, shown in Fig.~(\ref{fig:reweight}.a).
For any weak classifier (line passing the origin), the error is $\epsilon=0.5$ which
means that they are equally bad.
In a computer simulation the value is usually slightly smaller than $0.5$ since the
training points are discretized samples. Once a weak classifier is selected, e.g.,
the line $x_1>0$ ($A$), then positive samples $(A \land B)$ and negative samples
$(\bar{A} \land B)$ are correctly classified, and they will receive lower weights.
Fig.~(\ref{fig:reweight}.b) shows the weights for the samples after the first step of the
AdaBoost. Clearly, the weak classifier to minimize the error for this round
would be $x_1<0$ ($\bar{A}$), which is a contradictory decision to the previous weak classifier ($A$).
The re-weighted points after this round essentially lead the situation back to Fig.~(\ref{fig:reweight}.a).
The combination of the two weak classifiers is $A \land \bar{A}=\phi$ where $\phi$ denotes
an empty set. The algorithm then keeps repeating the same procedure, which is a deadlock.
Due to this reason, AdaBoost is sensitive to outliers since
it keeps giving high weights to those miss-classified samples.
\subsection{Possible solutions}
The previous section shows that Ada-Stump cannot solve the `xor' problem (on the line features passing
the origin). The AdaBoost algorithm makes an overall decision based on a weighted sum
$H(x)=sign(\sum_{t=1}^T \alpha_t h_t(x))$.
It weakly performs the `and', and `or' operations on the weak classifiers. The `not' is often embedded in the stump classifier
by switching $>$ and $<$. We assume that all types of weak classifiers have the aspect of
`not' and we focus on `and', and `or' operations for the rest of this paper.
There are several possible ways to improve the algorithm:
\begin{enumerate}
\item Designing hyper features to allow the patterns to be linearly separable. For example, in the `xor'
case, it could be $x_1 \times x_2$. However, (1) it
is often very hard to find the meaningful features which will nicely separate the positive
and negative samples; (2) complex features often lead to the over-fitting problem.
\item Introducing the explicit `and' and `or' relations into the AdaBoost.
\end{enumerate}
We can put `and's on top of `or's, or vice versa, or completely mix the two together. The probabilistic
boosting tree (PBT) algorithm~\cite{PBT} is one way of recursively combining `and's with `or's.
The disadvantages of PBT however are: (1) it requires longer training time than cascade and, (2) it produces
complex classifier and may lead to overfitting (like the decision tree).
Another solution is to build weak classifiers
with embedded `and' and `or' operations.
Using decision tree~\cite{C4.5} as weak classifiers has been described in several papers~\cite{Friedman98,Reyzin}.
However, each tree is a complex classifier and it requires much longer time
in training than the stump classifier. Also, it has more algorithm complexity than decision stump.
\section{Layered logic classifiers \label{sect:adaorboost}}
Eqn.~(\ref{eq:logistic}) shows that the AdaBoost algorithm is essentially approaching
a logistic probability by
\[
p(y|x) \propto e^{y \sum_t \alpha_t h_t(x)} \propto \prod_t e^{y \alpha_t h_t(x)}.
\]
The overall discriminative probability is a product of the probability of each
$h_t$. Depending upon its weight $\alpha_t$, each $h_t$ makes a direct impact on $p(y|x)$.
Using decision tree requires much longer time than stump
classifier. This is particularly a problem in vision as we often face millions of image samples with
each sample having thousands features.
Instead of using one layer AdaBoost, we can think of using two-layer AdaBoost
with the weak classifier being stronger than decision stump, but simpler than decision tree.
One idea might be to use Ada-Stump as weak classifier for the AdaBoost again, which we
call Ada-Ada-Stump, or Ada-Ada for short notation. However, Ada-Ada still somewhat performs a linear summation and
has difficulty on the `xor' as well. Fig.~(\ref{fig:toy}.c) shows the positives classified by
Ada-Ada with 50 weak classifiers of Ada-Stump, which itself has 5 stump weak classifiers.
It is a failure example.
It is worth to mention that one can indeed to make Ada-Ada work on these points by using
very tricky strategies of randomly selecting a subset of points in training.
However this greatly increases the training complexity and the procedures are not general.
Our solution is to propose AndBoost and OrBoost algorithms in which
the `and' and `or' operations are explicitly engaged.
We give detailed descriptions below.
\subsection{OrBoost \label{sect:orboostd}}
For a combined classifier, we can use the `or' operation directly by
\begin{equation}
H(x) = sign (h_1(x) \lor ... \lor h_T(x)),
\end{equation}
where
\begin{equation}
h_i(x) \lor h_j(x) = \left\{ \begin{array}{ll}
+1 & if \; h_i(x)=+1 \; or \; h_j(x)=+1\\
-1 & \textrm{otherwise} \\
\end{array} \right.
\end{equation}
\begin{figure}[!htb]
\begin{center}
\begin{tabular}{|c|}
\hline
\begin{minipage}[c]{80mm}{
{\footnotesize
Given: $(x_1,y_1,D(1)),...,(x_N,y_N,D(N));\; y_i \in \{-1, 1 \}$
For $t=1,...,T:$
\begin{itemize}
\item[$\bullet$] Train weak classifier using distribution $D$.
\item[$\bullet$] Get weak hypothesis $h_t: \chi \to \{-1,+1 \}$.
\item[$\bullet$] Train weak classifier using weights $D$ to minimize error $\sum_i D(i) {\bf 1} \big (y_i \ne (h_1(x_i)\; \lor ... \lor h_t(x_i)) \big)$.
\item[$\bullet$] The algorithm also stops if the error is not decreasing.
\end{itemize}
Output the overall classifier: $H(x) = sign (h_1(x) \lor ... \lor h_T(x))$.
}
}\end{minipage} \\
\hline
\end{tabular}
\end{center}
\caption{OrBoost algorithm.}
\label{fig:orboost_d}
\end{figure}
Fig~(\ref{fig:orboost_d}) gives the detailed procedure
of the OrBoost algorithm, which is straight forward to implement.
The overall classifier is a set of `or' operations on weak classifier, e.g.
decision stump, and it favors positive answer. If any weak classifier provides
a positive answer, then the final decision is
positive, regardless of what other weak classifier will say.
Unlike in the AdaBoost algorithm where mis-classified samples are given
higher weights in the next round, OrBoost gives up some samples quickly
and focus on those which can be classified correctly.
This helps to solve the deadlock situation in AdaBoost
shown in Fig.~(\ref{fig:reweight}).
\begin{figure}[!htbp]
\begin{center}
\begin{tabular}{c}
\psfig{figure=.//or_illustration.eps,width=3.3in} \\
\end{tabular}
\begin{tabular}{cc}
(a) initial weights & (b) re-weighted \\
\end{tabular}
\end{center}
\caption{The re-weighting scheme in the OrBoost algorithm which breaks the deadlock
in the AdaBoost algorithm.}
\label{fig:reweight_or}
\end{figure}
Fig.~(\ref{fig:reweight_or}) shows the feature selection and re-weighting steps by the OrBoost algorithm
for the xor problem. The first weak classifier is selected the same as before ($x_1 > 0 = A$).
However, positives $AB$ and negatives $\bar{A}B$ receive low weights in the AdaBoost
since they have been classified correctly. This creates a deadlock.
In OrBoost, the situation is different. Note that although the weights for all the
samples $D$ are fixed, the error evaluation function $\sum_i D(i) {\bf 1} \big (y_i \ne (h_1(x_i)\; \lor ... \lor h_t(x_i)) \big)$
affects how $D(i)$ plays a role. This is similar to the re-weighting scheme in the AdaBoost.
For example, positives $AB$ and negatives $A\bar{B}$ have been
classified as positives by the first weak classifier, $x_1 > 0 = A$. The errors on $AB$ and $A\bar{B}$
are therefore decided already regardless what the later weak classifiers will be.
Therefore, the second weak classifier would be $x_2 < 0 = \bar{B}$. The total error
by the two combined weak classifiers is $0.25$.
\subsection{AndBoost \label{sect:andboost}}
If we swap the labels of the positives and negatives in training,
the `or' operations in OrBoost can be directly turned into `and' operations since
\[
A \land B = \bar{A} \lor \bar{B}.
\]
However, for a given set of the training samples, `and' operations may
provide complementary decisions to the `or' operations.
Similarly, we can use the `and' operation directly by
\begin{equation}
H(x) = sign (h_1(x) \land ... \land h_T(x)),
\end{equation}
where
\begin{equation}
h_i(x) \land h_j(x) = \left\{ \begin{array}{ll}
+1 & if \; h_i(x)=+1 \; and \; h_j(x)=+1\\
-1 & \textrm{otherwise} \\
\end{array} \right.
\end{equation}
Therefore, we can design an AndBoost algorithm in Fig.~\ref{fig:andboost} which
is very similar to the OrBoost algorithm in Fig.~\ref{fig:orboost_d}.
\begin{figure}[!htb]
\begin{center}
\begin{tabular}{|c|}
\hline
\begin{minipage}[c]{80mm}{
{\footnotesize
Given: $(x_1,y_1,D(1)),...,(x_N,y_N,D(N));\; y_i \in \{-1, 1 \}$
For $t=1,...,T:$
\begin{itemize}
\item[$\bullet$] Train weak classifier using distribution $D$.
\item[$\bullet$] Get weak hypothesis $h_t: \chi \to \{-1,+1 \}$.
\item[$\bullet$] Train weak classifier using weights $D$ to minimize error $\sum_i D(i) {\bf 1} \big (y_i \ne (h_1(x_i)\; \land ... \land h_t(x_i)) \big)$.
\item[$\bullet$] The algorithm also stops if the error is not decreasing.
\end{itemize}
Output the overall classifier: $H(x) = sign (h_1(x) \land ... \land h_T(x))$.
}
}\end{minipage} \\
\hline
\end{tabular}
\end{center}
\caption{AndBoost algorithm.}
\label{fig:andboost}
\end{figure}
The performance of the AndBoost on the `xor' problem is the same as the OrBoost algorithm.
\subsection{AdaOrBoost}
\begin{figure*}[!htbp]
\begin{center}
\begin{tabular}{c}
\psfig{figure=.//andor_illustration.eps,width=5.4in} \\
\end{tabular}
\begin{tabular}{cccc}
(a) \hspace{26mm} & (b) \hspace{26mm} & (c) \hspace{26mm} & (d)\\
\end{tabular}
\end{center}
\vspace{-3mm}
\caption{Two-layer logic models. (a) shows a traditional Ada-Stump algorithm. (b) displays
the Ada-Or algorithm with OrBoost being the choice of weak classifier in the second layer.
(c) and (d) give the illustrations of Ada-And and Ada-AndOr respectively.}
\label{fig:andor}
\end{figure*}
After the introduction of the OrBoost and AndBoost algorithms, we
are ready to discuss the proposed layered models. We simply use a two-layer
AdaBoost algorithm with the weak classifiers in the second layer being the choice of OrBoost, AndBoost, or
both. We call the models, {\em Ada-Or}, {\em Ada-And}, and {\em Ada-AndOr} respectively.
There are two levels of weak classifiers now. For Ada-Or, the OrBoost is its weak classifier.
For OrBoost, any type of classifier can be its weak classifier.
To keep the complexity of OrBoost and AndBoost under check, we simply use the decision stump.
As we mentioned before the `not' operation is naturally embedded in
the decision stump.
Therefore, the Ada-AndOr has all the aspects of logic operations, `and', `or', and `not'.
Again, we call the weak classifiers in OrBoost and AndBoost {\em operations to} avoid confusion.
Fig.~(\ref{fig:toy}.b) shows the points which are classified by Ada-Stump using
100 stump weak classifiers. This failure example verifies our earlier claim for the `xor' pattern.
Fig.~(\ref{fig:toy}.d) shows the result by Ada-Or using 10 OrBoost weak classifiers, in which
there are 2 or operations. As we can see, the positive samples
have been classified correctly.
Fig.~(\ref{fig:toy}.c) gives the result by Ada-Ada.
The margin theory of Ada-Or, Ada-And, and Ada-AndOr still follows the same as pointed by Schapire et al.~\cite{Margin}
in eqn.~(\ref{eq:margin}). The complexity $d$ of weak classifier is decided by the OrBoost and AndBoost algorithms,
which are just a sequence of `or' operations or `and' operations. It is slightly more complex than decision stump,
but much simpler than decision tree or CART.
It is worth to mention that both OrBoost and AndBoost include a special case where only one operation presents.
This happens when the training error is not improving by adding the second operation. Therefore, stump
classifier is also included in OrBoost and AndBoost, if stump is the choice of operator.
\subsection{Experiments \label{sec:sampling}}
There are several major issues we are concerned with for the choice of different
classifiers for applications in machine learning and computer vision.
\begin{enumerate}
\item {\em Classification power}: This is often referred to as
training error or margin in eqn.~(\ref{eq:margin}).
A desirable classifier should produce low error and large margin on the training data.
\item {\em Low complexity}: This is often called VC dimension~\cite{SVM} and a classifier
with small VC dimension often has a good generalization power, small difference between the training error
and test error.
\item {\em Size of training data}: In the VC dimension and margin theory, the overall
test error is also greatly decided by the availability of training data. The more training data
we have and the classifier can handle, the smaller difference is between training error and
test error. In reality, we often do not have enough training data since collecting them
is not a easy task. Also, some none-parametric classifiers can only deal with limited amount
of training data since they work on the kernel space, which explodes on large size data.
\item {\em Efficient training time}: For many applications in computer vision and data mining,
the training data size can be immense and each data sample also has a large number of features.
This demands an efficient classifier in training also. Fast training is more required in online learning
algorithms~\cite{Oza} which has recently received many attentions in tracking.
\item {\em Efficient test time}: Judging the performance of a classifier is ultimately done in the
test stage. A classifier is expected to be able to quickly give an answer. For many modern
classifiers, this is not particularly a problem.
\end{enumerate}
The first three criterion collectively decide the test error of a classifier.
Another major factor affecting the performance a classifier is feature design.
If the intrinsic features can be found, different types of classifiers will
probably have a similar performance. However, the discussion of feature design is out of the scope of
this paper. Next, we focus on the performance of AdaOrBoost with comparison to the
other classifiers.
\subsection{Results on UCI repository datasets}
One of the reasons that the AdaBoost algorithm is widely used is
due to nice generalization power. Schapire et al. gave an explanation based on the
margin theory after Breiman~\cite{BreimanArc} observed an interesting behavior of AdaBoost:
the test error of AdaBoost further asymptotically goes down even the training error is not decreasing.
This was explained
in the margin theory as to increase the margin with more weak classifiers combined.
Brieman~\cite{Breiman} then designed an algorithm called `arc-gv' which
tries to directly maximize the minimum margin in
computing the $\alpha_t$ for AdaBoost. The experimental results were however contradictory to the
theory since arc-gv produces bigger test error than AdaBoost. Reyzin and Schapire~\cite{Reyzin} tried to
explain this finding and showed that the bigger test error by arc-gv was
indeed due to the use of complex weak classifier, CART. Next we compare
Ada-Or, Ada-And, and Ada-AndOr with arc-gv and AdaBoost using CART and decision stump.
We use the same datasets shown in Reyzin and Schapire~\cite{Reyzin}, which are
all from the UCI repository: breast cancer, ionosphere, ocr49 and splice.
The datasets have been slightly modified the same way as in~\cite{Reyzin}.
The two splice categories were merged into one in the splice dataset to create
two-class data. Only digits 4 and 9 from the NIST database were used in the ocr49 dataset.
The cancer, ion, ocr49 and splice then have 699, 351, 6000, 3175 data points respectively.
Each sample usually has $20-60$ features, depending upon what dataset it belongs to.
The data samples are randomly split into training and testing for 10 trials. Table~\ref{tb:uci}
shows the corresponding numbers.
\begin{table}[!htb]
\begin{center}
{\scriptsize
\begin{tabular}{|c|c|c|c|c|}
\hline
& cancer & ion & ocr 49 & splice \\
\hline
training & 630 & 315 & 1000 & 1000 \\
\hline
test & 69 & 36 & 5000 & 2175 \\
\hline
\end{tabular}
}
\end{center}
\caption{The sizes of training and test data of four datasets from UCI repository.
The training and test data samples are randomly selected.}
\label{tb:uci}
\end{table}
To illustrate the effectiveness of the layered models, we first
compare its results to those by Ada-Stump.
Though there are other alternatives such as RealBoost and GentalBoost~\cite{Friedman98},
decision stump remains being widely adopted in the AdaBoost implementation.
Fig.~(\ref{fig:different_op}.a) shows the training and test errors on the splice dataset
by Ada-Stump, Ada-Or, Ada-And, and Ada-AndOr using different number of weak classifiers.
In the implementation of OrBoost and AndBoost, we use 5 `or' operations.
Each curve is averaged over 10 trials by randomly selecting
1000 samples for training and 2175 samples for testing.
The Ada-AndOr gives the best performance among all.
We also observe that the differences between the training and test errors
for Ada-Stump and others are very similar.
The results for real-world vision applications also show similar behavior of Ada-Or, Ada-and, and Ada-AndOr.
This suggests that the OrBoost and AndBoost algorithms are having similar
generalization power as decision stump.
\begin{figure}[!htbp]
\begin{center}
\begin{tabular}{cc}
\hspace{-6mm} \psfig{figure=.//different_weak.eps,width=1.8in} &
\hspace{-6mm} \psfig{figure=.//different_op.eps,width=1.8in} \\
(a) & (b) \\
\end{tabular}
\end{center}
\caption{Training and test errors on the splice dataset by proposed models using different number
of weak classifiers and operations. (a) displays comparison with different number of weak classifiers.
Each curve is is averaged over 10 trials by randomly
splitting the dataset into training and test samples.
(b) shows comparison of using different operations. Each algorithm
uses 50 weak classifiers.}
\label{fig:different_op}
\end{figure}
To show how the use of different number of operations is affecting the performance,
we conduct another experiment on the splice dataset.
We plot out the training and test errors by using 50 weak classifiers with varying number of operations.
The overall performance of the models, both in training and testing, is not
improving too much with more than 3 operations shown in Fig.~(\ref{fig:different_op}.b). Similar observations apply to
other datasets as well. This suggests that the significant improvement
can be achieved without introducing too much overhead.
\begin{table*}[!htb]
\begin{center}
{\scriptsize
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& $\frac{arc-gv-CART}{arc-gv-stump}$ & $\frac{Ada-CART}{Ada-Stump}$ & $\frac{Ada-Stump (2500)}{Ada-Stump}$ & $\frac{Ada-Or}{Ada-Stump}$ & $\frac{Ada-And}{Ada-Stump}$ & $\frac{Ada-AndOr}{Ada-stump}$ \\
\hline
breast cancer & $73.3\%$ & $57.3\%$ & $80.7\%$ & $57.4\%$ & $48.2\%$ & $56.0\%$ \\
\hline
ionosphere & $74.7\%$ & $36.1\%$ & $136.5\%$ & $63.6\%$ & $87.8\%$ & $66.7\%$\\
\hline
ocr 49 & $37.3\%$ & $32.5\%$ & $91.4\%$ & $57.2\%$ & $54.8\%$ & $38.2\%$\\
\hline
splice & $47.8\%$ & $46.8\%$ & $113.6\%$ & $83.8\%$ & $68.4\%$ & $60.8\%$ \\
\hline
\end{tabular}
}
\end{center}
\caption{Test error ratios on the UCI datasets by arc-gv-CART, Ada-CART, Ada-Or, Ada-And, and
Ada-AndOr over Ada-Stump. 500 weak classifiers are used in all cases except for Ada-Stump (2500)
and Ada-Stump, which use 2500 and 100 stumps respectively.
Ada-Or, Ada-And, and Ada-AndOr all contain 5 operations in the OrBoost and AndBoost which have roughly 2500
stumps for each. Ada-AndOr significantly outperforms Ada-Stump, and it shows to be comparable to
arg-gv-CART, and is only a bit worse than Ada-CART.
}
\label{tb:comparison}
\end{table*}
It has been suggested~\cite{Friedman98,Breiman,Reyzin} that the best performance of boosting algorithm
is achieved by AdaBoost using decision tree~\cite{C4.5} or CART~\cite{CART}.
Some of the confusions about generalization (test) error based on the margin theory has
recently been clarified by Reyzin and Schapire~\cite{Reyzin}. In table~(\ref{tb:comparison}),
we compare the our algorithms with AdaBoost and arc-gv using decision tree.
For a fair comparison, we show the improvement of AdaOrBoost, arc-gv using CART, and AdaBoost
using CART over those using decision stump.
Table~(\ref{tb:comparison}) shows the error ratio. As we can see, the improvement of AdaOrBoost
is comparable to arc-gv using CART, but is worse than Ada-CART.
However, each CART, after tree pruning, has around 16 leaf nodes with the tree depth being around 7.
Therefore, the complexity of CART is much bigger than that of OrBoost and AndBoost.
This is particularly an issue for applications in vision as the training data is massive with each
data sample having thousands or even millions of features. The good performance of
Ada-CART is achieved using an average of $7$ levels of tree.
This greatly limits its usage in many vision applications and leaves
the decision stump classifiers still being currently widely used~\cite{Viola}.
\begin{figure}[!htbp]
\begin{center}
\begin{tabular}{c}
\psfig{figure=.//horse_comparisions.eps,width=3.0in} \\
(a) \\
\psfig{figure=.//pedestrain.eps,width=2.4in} \\
(b) \\
\end{tabular}
\end{center}
\vspace{-3mm}
\caption{Precision and recall curves for the horse segmentation and error
curve for pedestrian detection. (a) Shows the training and test errors by
decision stump based Ada-Stump, Ada-Or, Ada-And, and Ada-AndOr. (b) displays
the error curve by various classifiers. 3 operations are used for
all classifiers. Ada-AndOr achieves the
best result in both the cases among all Ada-. The horse segmentation
result (F=0.8) outperforms that reported in~\cite{Ren} (F=0.66) and
the pedestrian detection result is slightly worse than HOG~\cite{HOG}.
}
\vspace{-5mm}
\label{fig:detection}
\end{figure}
To illustrate the effectiveness of proposed algorithms, we further demonstrate them in
two challenging vision problems, object segmentation and pedestrian detection.
First, we demonstrate it on the Weizmann horse dataset~\cite{Borenstein}.
We use 328 images and use 126 for training and 214 for testing.
Each input image comes with a label map in which the
pixels on the horse body and background are labeled as $1$ and $0$ respectively.
Given a test image, our task is to classify all the pixels into horse or background.
In training, we take image patch of size $21 \times 21$ centered on every pixel
as training samples. The background and horse body
image patches are the negatives and positives respectively.
For each image patch, we compute around $10,000$ features such as the mean, variance,
and Haar responses of the original as well as Gabor filtered.
We implement a cascade approach~\cite{Viola} and implement several versions. One uses Ada-Stump and
others use Ada-Or and Ada-AndOr.
Each cascade node selects and fuses 100 weak classifiers.
All the algorithms use an identical set of features and bootstrapping procedure.
Fig.~(\ref{fig:detection}.a) shows the precision and recall curve of the algorithms
on the training and test images. We observe similar result as that for the UCI repository
datasets. Ada-AndOr improves the results over Ada-Stump by a considerable amount.
The differences between the training and test errors are nearly the same in this cascade setting as well.
The F-value of the results by Ada-AndOr is around 0.8 which is better than the number 0.66 reported
in~\cite{Ren} which uses low and middle level information.
Next, we show the Ada-AndOr algorithm for pedestrian detection on dataset reported in~\cite{HOG}.
We use 8 level of cascade with different choices of weak classifiers for AdaBoost.
Fig.~(\ref{fig:detection}b) shows the results by Ada-Stump, Ada-Ada, Ada-Or, Ada-And, and Ada-AndOr.
The conclusion is nearly the same as before. Ada-AndOr achieves the best result among all with
Ada-And being on the second place.
Though we are not specifically addressing the pedestrian detection problem here, the result is nevertheless
close to that by the well-known HOG pedestrian detector~\cite{HOG}.
However, we only use a set of generic Haar features without tuning the system specifically
for the pedestrian detection task.
\subsection{Conclusions}
Many of the classification problems in machine learning and computer vision
can be understood as performing logic operations combining `and', `or', and `not'.
In this paper, we have introduced layered logic classifiers.
We show that AdaBoost can not solve the `xor' problem using decision stump type of weak classifiers.
We propose an OrBoost and AndBoost algorithms to study the `or' and `and' operations
respectively.
We demonstrate that the combined algorithm of two layers, Ada-AndOr, greatly outperformed
Ada-Stump which is widely used in the literature.
The improvement is significant in most the cases. We demonstrate the effectiveness of Ada-AndOr
on traditional machine learning datasets, as well as challenging vision applications.
Though decision tree based AdaBoost algorithm is shown to produce smaller
test error, its complexity in training often limits its usage.
The OrBoost and AndBoost algorithm only increases the time complexity slightly than
decision stump, but they significantly reduce the test error.
The Ada-AndOr algorithm is useful for a wide variety of applications
in machine learning and computer vision.
{\bf Acknowledgment}
This work is supported by NSF IIS-1216528 (IIS-1360566) and NSF CAREER award IIS-0844566 (IIS-1360568).
\bibliographystyle{ieee}
|
\section{Moments under exponential $P_0$}\label{app_mom}
We provide an explicit expression for \eqref{mom_ext_gamma} when $P_0(\mathrm{d} y)=\lambda \exp(-\lambda y )\mathrm{d} y$ and the hyperparameters $c$ and $\beta$ are considered random.
\begin{multline}\label{mom3}
\mathds{E}[\tilde S^r(t)\,|\,\bm{X},\bm{Y},c,\beta]=\\
\exp\left\{-c\mathrm{e}^{-f_{0,r}(0)}\left[\mbox{Ei}(f_{0,r}(t))-\mbox{Ei}(f_{0,r}(X_1\wedge t))\right]\right\}\left(\frac{f_{0,r}(X_1\wedge t)}{f_{0,r}(t)}\right)^{-c\mathrm{e}^{-\lambda (X_1\wedge t)}}\\
\times\prod_{i=1}^n\exp\left\{-c\mathrm{e}^{-f_{i,r}(0)}\left[\mbox{Ei}(f_{i,r}(X_i\wedge t))-\mbox{Ei}(f_{i,r}(X_{i+1}\wedge t))\right]\right.\\
\left. -c\mathrm{e}^{-f_{i,0}(0)}\left[\mbox{Ei}(f_{i,0}(X_{i+1}\wedge t))-\mbox{Ei}(f_{i,0}(X_{i}\wedge t))\right]\right\}\\[5pt]
\times\left(\frac{i}{i+r}\frac{f_{i,0}(X_i\wedge t)}{f_{i,r}(X_i\wedge t)}\right)^{-c\mathrm{e}^{-\lambda (X_i\wedge t)}}\left(\frac{i+r}{i}\frac{f_{i,r}(X_{i+1}\wedge t)}{f_{i,0}(X_{i+1}\wedge t)}\right)^{-c\mathrm{e}^{-\lambda (X_{i+1}\wedge t)}}\\
\times\prod_{j=1}^k \left(1+r\,\frac{(t-Y_j^*)\mathds{1}_{[Y_j^*,\infty)}(t)}{\sum_{i=1}^n (X_i-Y_j^*)\mathds{1}_{[Y_j^*,\infty)}(X_i)+1/\beta}\right)^{-n_j},
\end{multline}
where $\mbox{Ei}(\cdot)$ is the exponential integral function defined for non-zero real values $z$ by
$$\mbox{Ei}(z)=-\int_{-z}^\infty\frac{\mathrm{e}^{-t}}{t}\mathrm{d} t$$
and the function $f_{i,r}$, for $i,r\geq 0$ such that $i+r>0$, is defined by
$$f_{i,r}(x)=\lambda\left(\frac{\xi_{i}+1/\beta+rt}{i+r}-x\right).$$
\section{Full conditional distributions}\label{app_fc}
In this section we provide expressions for the full conditional distributions needed in the algorithm described in Section~\ref{algo} for extended gamma processes with base measure $P_0(\mathrm{d} y)=\lambda \exp(-\lambda y) \mathrm{d} y$. These distributions are easily derived, up to a constant, from the joint distribution of the vector $(\bm{X},\bm{Y},c,\beta)$, that can be obtained from \eqref{eq:joint.rhr.1}.
Therefore we start by providing the full conditional distribution for the latent variable $Y_i$, with $i=1,\ldots,n$, where $\bm{Y}^{(-i)}$ denotes the vector of distinct values $(\tilde{Y}^*_1,\ldots,\tilde{Y}^*_{k^*})$ in $(Y_1,\ldots,Y_{i-1},Y_{i+1},\ldots,Y_n)$ and $(n_1^{(-i)},\ldots,n_{k^{*}}^{(-i)})$ represent the corresponding frequencies.
\begin{equation}\label{yfc}
\P[Y_i = \mathrm{d} y \, | \, \bm{X}, \bm{Y}^{(-i)},c,\beta]=p_0 G_0 (\mathrm{d} y)+ \sum_{j=1}^{k^*} p_j \delta_{\tilde{Y}_j^{*}}(\mathrm{d} y),
\end{equation}
where
\begin{align*}
p_0&\propto c\,\lambda\, \sum_{j=i}^n \frac{1}{j}\mathrm{e}^{-\lambda \frac{\xi_j+1/\beta}{j}}\left[\mbox{Ei}\left(f_{j,0}(X_{j+1})\right)-\mbox{Ei}\left(f_{j,0}(X_j)\right)\right],\\
p_j&\propto\mathds{1}_{\{Y_j^*\leq X_{i}\}}\frac{n_j^{(-i)}}{\sum_{l=1}^n (X_l-\tilde Y_j^*)\mathds{1}_{[0,X_l)}(\tilde Y_j^*)+1/\beta}
\end{align*}
and
\begin{equation*}
G_0(\mathrm{d} y)\propto \mathds{1}_{[0,X_i)}(y)\mathrm{e}^{-\lambda y}\frac{1}{\sum_{j=1}^n (X_j-y)\mathds{1}_{[0,X_j)}(y)+1/\beta}\mathrm{d} y.
\end{equation*}
Finally, the full conditional distributions for the parameters $c$ and $\beta$ are given respectively by
\begin{multline}\label{cfc}
\mathcal{L}(c\,|\,\bm{X},\bm{Y},\beta)\propto \mathcal{L}_{0}(c) c^k \beta^{-c}\prod_{i=1}^n \exp\left\{-c \mathrm{e}^{-f_{i,0}(0)}\left[\mbox{Ei}(f_{i,0}(X_i))-\mbox{Ei}(f_{i,0}(X_{i+1}))\right] \right\}\\
\times \frac{\left(\xi_i+1/\beta-iX_{i+1}\right)^{-c\mathrm{e}^{-\lambda X_{i+1}}}}{\left(\xi_i+1/\beta-iX_{i}\right)^{-c\mathrm{e}^{-\lambda X_i}}}
\end{multline}
and
\begin{multline}\label{betafc}
\mathcal{L}(\beta\,|\,\bm{X},\bm{Y},c)\propto \mathcal{L}_0(\beta)\beta^{-c}\prod_{i=1}^n \exp\left\{-c \mathrm{e}^{-f_{i,0}(0)}\left[\mbox{Ei}(f_{i,0}(X_i))-\mbox{Ei}(f_{i,0}(X_{i+1}))\right] \right\}\\
\times \frac{\left(\xi_i+1/\beta-iX_{i+1}\right)^{-c\mathrm{e}^{-\lambda X_{i+1}}}}{\left(\xi_i+1/\beta-iX_{i}\right)^{-c\mathrm{e}^{-\lambda X_i}}}\prod_{j=1}^k \left(\sum_{i=1}^n(X_i-Y_j^*)\mathds{1}_{[Y_j^*,\infty)}(X_i)+1/\beta\right)^{-n_j},
\end{multline}
where $\mathcal{L}_0(c)$ and $\mathcal{L}_0(\beta)$ are the prior distributions of $c$ and $\beta$ respectively.
\section{Censored observations}\label{sec:censored}
The methodology we have presented in Section~\ref{sec:real} needs to be adapted to the presence of right-censored observations in order to be applied to the dataset in Section~\ref{sec:leukemia}. Here we introduce some notation and illustrate how the posterior characterization of Proposition~\ref{Jam} changes when data are censored. To this end, let $C_i$ be the right-censoring time corresponding to $X_i$, and define $\Delta_i = \mathds{1}_{(0,C_i]}(X_i)$, so that $\Delta_i$ is either 0 or 1 according as to whether $X_i$ is censored or exact. The actual $i$th observation is $T_i=\min(X_i,C_i)$ and, therefore, data consist of pairs $\bm{D} = \{(T_i,\Delta_i)\}_{i=1\ldots n}$.
In this setting, the likelihood in~\eqref{eq:likelihood} can be rewritten as
\begin{equation*
\mathcal{L}(\mt;\bm{D})=e^{-\int_\mathbb{Y} K^*_{\bm{D}}(y)\mt(\mathrm{d} y)}\prod_{i:\,\Delta_i=1}\int_\mathbb{Y} k(T_i;y)\mt(\mathrm{d} y),
\end{equation*}
where $$K^*_{\bm{D}}(y)=\sum_{i=1}^n\int_0^{T_i}k(s;y)\mathrm{d} s.$$
By observing that the censored times are involved only through $K^*_{\bm{D}}$, we have that the results derived in Proposition~\ref{Jam} under the assumption of exact data easily carry over to the case with right-censored data. The only changes refer to $K_{\bm{X}}$, that is replaced by $K^*_{\bm{D}}$, and the jump components which occur only at the distinct values of the latent variables that correspond to exact observations. For instance in Proposition~\ref{Jam}, the L\'evy intensity of the part of the CRM without fixed points of discontinuity is modified by
\begin{equation*}
\nu^*(\mathrm{d} s, \mathrm{d} y)= e^{-s K_{\bm{D}}^*(y)}\rho_y(s)\mathrm{d} s\,c P_0(\mathrm{d} y),
\end{equation*}
while the distribution of the jump $J_j$ has density function $f(\,\cdot\,|\, n_j^*,K_{\bm{D}}^*(Y_j^*),Y_j^*)$ with $f$ defined in \eqref{eq:jump_dens} and $n_j^* = \#\big\{i\,:\,Y_i=Y_j^* \text{ and } \Delta_i = 1\big\}$. Adapting the results of Proposition~\ref{mom} and Corollary~\ref{cor:moments}, as well as the full conditional distributions in \ref{app_fc}, is then straightforward.
\subsection{Moment-based density approximation and sampling}\label{sec:approx1}
Recovering a probability distribution from the explicit knowledge of its moments
is a classical problem in probability and statistics that has received great attention in the literature.
See, e.g., \cite{provost2005moment}, references and motivating applications therein.
Our specific interest in the problem is motivated by the goal of determining an approximation of the density function of a distribution supported on $[0,1]$ that equals the posterior distribution of a random survival function evaluated at some instant $t$. This is a convenient case since, as the support is a bounded interval, all the moments exist and uniquely characterize the distribution, see \citet{rao1965linear}.
Moment-based methods for density functions' approximation can be essentially divided into two classes, namely methods that exploit orthogonal polynomial series \citep{provost2005moment} and maximum entropy methods \citep{csiszar1975divergence,mead1984maximum}. Both these procedures rely on systems of equations that relate the moments of the distribution with the coefficients involved in the approximation. For our purposes the use of orthogonal polynomial series turns out to be more convenient
since it ensures faster computations as it involves uniquely linear equations. This property is particularly important in our setting since we will need to implement the same approximation procedure for a large number of times in order to approximate the posterior distribution of a random survival function. Moreover, as discussed in \citet{epifani2009moment}, maximum entropy techniques can lead to numerical instability.
Specifically, we work with Jacobi polynomials, a broad class which includes, among others, Legendre and Chebyshev polynomials. They are well suited for the expansion of densities with compact support contrary to other polynomials like Laguerre and Hermite which can be preferred for densities with infinite of semi-infinite support \citep[see][]{provost2005moment}.
While the classical Jacobi polynomials are defined on $[-1,1]$, we consider a suitable transformation of such polynomials so that their support coincides with $[0,1]$ and therefore matches the support of the density we aim at approximating. That is, we consider a sequence of polynomials $(G_i)_{i\geq 0}$ such that, for every $i\in \mathds{N}$, $G_i$ is a polynomial of order $i$ defined by $G_i(s)=\sum_{r=0}^i G_{i,r}s^r$, with $s\in [0,1]$. The coefficients $G_{i,r}$ can be defined by a recurrence relation
\citep[see for example][]{szego1967orthogonal}. Such polynomials are orthogonal with respect to the $L^2$-product
\begin{equation*}\label{eq:scalar_product}
\langle F,G \rangle=\int_0^1 F(s) G(s) w_{a,b}(s)\mathrm{d} s,
\end{equation*}
where
\[w_{a,b}(s)=s^{a-1}(1-s)^{b-1}\]
is named \emph{weight function} of the basis. Moreover, without loss of generality, we assume that the $G_i$'s are normalized and, therefore, $\langle G_i,G_j \rangle=\delta_{ij}$ for every $i,j\in \mathds{N}$, where $\delta_{ij}$ is the Kronecker symbol. Any univariate density $f$ supported on $[0,1]$ can be uniquely decomposed on such a basis and therefore there is a unique sequence of real numbers $(\lambda_i)_{i \geq 0}$ such that
\begin{equation}\label{eq:decomposition}
f(s)=w_{a,b}(s)\sum_{i=0}^\infty \lambda_i G_i(s).
\end{equation}
Let us now consider a random variable $S$ whose density $f$ has support on $[0,1]$.
We denote its raw moments by $\mu_r=\mathds{E}\big[S^r\big]$, with $r\in\mathds{N}$.
From the evaluation of
$\int_0^1 f(s)\, G_i(s)\,\mathrm{d} s$ it follows that each $\lambda_i$ coincides with a linear combination of the first $i$ moments, specifically $\lambda_i=\sum_{r=0}^i G_{i,r}\mu_r$.
Then, the polynomial approximation method consists in truncating the sum in \eqref{eq:decomposition} at a given level $i=N$. This procedure leads to a methodology that makes use only of the first $N$ moments and provides the approximation
\begin{equation}\label{eq:approx}
f_N(s)=w_{a,b}(s)\sum_{i=0}^N \left(\sum_{r=0}^i G_{i,r}\mu_r\right) G_i(s).
\end{equation}
It is important to stress that the polynomial expansion approximation~\eqref{eq:approx} is not necessarily a density as it might fail to be positive or to integrate to 1. In order to overcome this problem, we consider the density $\pi_N$ proportional to the positive part of $f_N$, i.e. $\pi_N(s)\propto\max(f_N(s),0)$. We resort to importance sampling \citep[see, e.g.,][]{robert2004monte} for sampling from $\pi_N$.
This is a method for drawing independent weighted samples $(\varpi_\ell, S_\ell)$ from a distribution proportional to a given non-negative function, that exempts us from computing the normalizing constant. More precisely, the method requires to pick a proposal distribution $p$ whose support contains the support of $\pi_N$. A natural choice for $p$ is the Beta distribution
proportional to the weight function $w_{a,b}$.
The weights are then defined by $\varpi_\ell\propto\max(f_N(S_\ell),0)/p(S_\ell)$ such that they add up to 1.
An important issue related to any approximating method refers to the quantification of the approximating error. As for the polynomial approach we undertake, the error can be assessed for large $N$ by applying the asymptotic results in
\citet{alexits1961convergence}. In our case, the convergence $f_N(s)\rightarrow f(s)$ for $N\rightarrow \infty$, for all $s\in(0,1)$, implies $\pi_N(s)\rightarrow f(s)$ for $N\rightarrow \infty$. Thus, if $S_N$ denotes a random variable with distribution $\pi_N$, then the following convergence in distribution to the target random variable $S$ holds:
\begin{equation*
S_N \stackrel{\mathcal{D}}{\longrightarrow} S \text{ as } N \rightarrow \infty.
\end{equation*}
However, here we are interested in evaluating whether few moments allow for a good approximation of the posterior distribution of $\tilde S(t)$. This question will be addressed by means of an extensive numerical study in the next section. See \citet{epifani2003exponential} and \citet{epifani2009moment} for a similar treatment referring to functionals of neutral-to-the-right priors and Dirichlet processes respectively.
\subsection{Numerical study}\label{sec:approx2}
In this section we assess the quality of the approximation procedure described above by means of a simulation study. The rationale of our analysis consists in considering random survival functions for which moments of any order can be explicitly evaluated at any instant $t$, and then compare the true distribution with the approximation obtained by exploiting the knowledge of the first $N$ moments. This in turn will provide an insight on the impact of $N$ on the approximation error. To this end, we
consider three examples of random survival functions $\tilde S_j$, with $j=1,2,3$. For the illustrative purposes we pursue in this Section, it suffices to specify
the distribution of the random variable that coincides with $\tilde S_j$ evaluated in $t$, for every $t>0$. Specifically, we consider a Beta, a mixture of Beta, and a normal distribution truncated to $[0,1]$, that i
\begin{align*}\label{eq:simulated_examples}
\tilde S_1(t)&\sim\Be\left(\frac{S_0(t)}{a_1}, \frac{1-S_0(t)}{a_1}\right), &\nonumber\\[0.3pt]
\tilde S_2(t)&\sim\frac{1}{2}\Be\left(\frac{S_0(t)}{a_2}, \frac{1-S_0(t)}{a_2}\right)+\frac{1}{2}\Be\left(\frac{S_0(t)}{a_3}, \frac{1-S_0(t)}{a_3}\right), &\\[0.3pt]
\tilde S_3(t)&\sim\truncNorm\left(S_0(t),\frac{S_0(t)(1-S_0(t))}{a_4}\right), & \nonumber
\end{align*}
where
$S_0(t)=\mathrm{e}^{-t}$ and we have set $a_1=20$, $(a_2,a_3)=(10,30)$ and $a_4=2$.
We observe that, for every $t>0$, $\mathds{E}[\tilde S_1(t)]=\mathds{E}[\tilde S_2(t)]=S_0(t)$ but the same does not hold true for $\tilde S_3(t)$.\\
For each $j=1,2,3$, we computed the first 10 moments of $\tilde S_j(t)$ on a grid $\{t_1,\ldots,t_{50}\}$ of 50 equidistant values of $t$ in the range $[0,2.5]$. The choice of working with 10 moments will be motivated at the end of the section. Then we used the importance sampler described in Section~\ref{sec:approx1} to obtain samples of size 10~000 from the distribution of $\tilde S_j(t_i)$, for each $j=1,2,3$ and $i=1,\ldots,50$.
In Figure~\ref{fig:simulated_examples1}, for each $\tilde S_j$, we plot the true mean as well as the $95\%$ highest density intervals for the true distribution and for the approximated distribution obtained by exploiting $10$ moments.
Notice that the focus is not on approximating the mean since moments of any order are the starting point of our procedure. Interestingly, the approximated intervals show a very good fit to the true ones in all the three examples. As for the Beta case, the fit is exact since the Beta-shaped weight function allows the true density to be recovered with the first two moments. As for the mixture of Beta, exact and approximated intervals can hardly be distinguished. Finally, the fit is pretty good also for the intervals in the truncated normal example.
Similarly, in Figure~\ref{fig:simulated_examples2} we compare the true and the approximated densities of each $\tilde S_j(t)$ for fixed $t$ in $\{0.1,0.5,2.5\}$. Again, all the three examples show a very good pointwise fit.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=.32\linewidth]{beta_example.pdf}
\includegraphics[width=.32\linewidth]{beta_mixture_example.pdf}
\includegraphics[width=.32\linewidth]{normtrunc_example.pdf}
\caption{Mean of $\tilde S_j(t)$ (dashed black) and $95\%$ highest density intervals for the true distribution (solid black) and the approximated distribution (dashed red) for the Beta ($j=1$), mixture of Beta ($j=2$) and truncated normal ($j=3$) examples (left, middle and right, respectively).}
\label{fig:simulated_examples1}
\end{figure}
\end{center}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=.98\linewidth]{beta_example_pointwise.pdf}
\includegraphics[width=.98\linewidth]{beta_mixture_example_pointwise.pdf}
\includegraphics[width=.98\linewidth]{normtrunc_example_pointwise}
\caption{True density (solid black) and approximated one (dashed red) at time values $t=0.1$ (left column), $t=0.5$ (middle column) and $t=2.5$ (right column), for the Beta ($j=1$, top row), mixture of Beta ($j=2$, middle row) and truncated normal ($j=3$, bottom row) examples.}
\label{fig:simulated_examples2}
\end{figure}
\end{center}
$\,$\\[-69pt]
We conclude this section by assessing how the choice of $N$ affects the approximation error. To this end, for each instant $t$ on the grid, we numerically compare the true and approximated distributions of $\tilde S_j(t)$, by computing the integrated squared error ($L^2$ error) between the two. Thus, we consider as measure of the overall error of approximation the average of these values. The results are illustrated in Figure~\ref{fig:L2_error}. As expected, the approximation is exact in the Beta example. In the two other cases, we observe that the higher is the number of exploited moments, the lower is the average approximation error. Nonetheless, it is apparent that the incremental gain of using more moments is more substantial when $N$ is small whereas it is less impactful as $N$ increases: for example in the mixture of Beta case, the $L^2$ error is 2.11, 0.97, 0.38 and 0.33 with $N$ equal to 2, 4, 10 and 20 respectively. Moreover, when using a large number of moments, e.g. $N>20$, some numerical instability
can occur.
These observations suggest that working with $N=10$ moments in~\eqref{eq:approx} strikes a good balance between accuracy of approximation and numerical stability.
\begin{center}
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=.5\linewidth]{L2_error.pdf}
\caption{Average across $t$ of the $L^2$ error between the true and the approximated densities of $\tilde S_j(t)$, in the Beta example (blue triangles), the mixture of Beta (red dots) and the truncated normal example (black squares). The approximation is exact in the Beta example.}
\label{fig:L2_error}
\end{center}
\end{figure}
\end{center}
\subsection{Algorithm}\label{algo}
The two main steps we take for drawing samples from the posterior distribution of $\tilde S(t)$, for any $t\in\{t_1,\ldots,t_q\}$, are summarized in Algorithm~\ref{algo:sampler}. First we perform a Gibbs sampler for marginalizing the latent variables $\bm{Y}$ and the hyperparameters $(c,\beta)$ out of \eqref{mom3} and therefore, for every $i=1,\ldots,q$, we obtain an estimate for the posterior moments $\mathds{E}[\tilde S^{r}(t_i)\vert \bm{X}]$, with $r=1,\ldots,N$. We run the algorithm for $l_{\max}=100~000$ iterations, with a burn-in period of $l_{\min}=10~000$. Visual investigation of the traceplots of the parameters, in the illustrations of Sections~\ref{sec:simulated}~and~\ref{sec:leukemia}, did not reveal any convergence issue. The second part consists in sampling from the posterior distribution of $\tilde S(t_i)$, for every $i=1,\ldots,q$, by means of the importance sampler described in Section~\ref{sec:approx1}. Specifically we sample $\ell_{\max} = 10~000$ values for each $t_i$ on the grid.
\begin{center}
\begin{algorithm}[h!]
\caption{Posterior sampling\label{algo:sampler}}\medskip
\textbf{Part 1.} \textbf{Gibbs sampler}
\begin{algorithmic}[1]
\STATE set $l=0$ and admissible values for latent variables and hyperparameters,\\i.e. $\{Y_1=Y_1^{(0)},\ldots,Y_n=Y_n^{(0)}\}$, $c=c^{(0)}$ and $\beta=\beta^{(0)}$\\[7pt]
\STATE while $l< l_{\max}$, set $l=l+1$, and
\begin{itemize}
\item update $Y_{j}=Y_j^{(l)}$ by means of \eqref{yfc}, for every $j=1,\ldots,n$
\item update $c=c^{(l)}$ and $\beta=\beta^{(l)}$ by means of \eqref{cfc} and \eqref{betafc}
\item if $l > l_{\min}$, compute
\begin{equation}\label{momgibbs}
\mu^{(l)}_{r,t}=\mathds{E}[\tilde S^r(t)\,\vert\,\bm{X},\bm{Y}^{(l)},c^{(l)},\beta^{(l)}]
\end{equation}
by means of \eqref{mom3} for each $r=1,\ldots,N$ and for each $t$ in the grid
\end{itemize}
\STATE for each $r=1,\ldots,N$ and each $t$ define
$\hat\mu_{r,t}=\frac{1}{l_{\max}-l_{\min}} \sum_{l=l_{\min}+1}^{l_{\max}} \mu_{r,t}^{(l)}$\\[7pt]
\end{algorithmic}
\textbf{Part 2.} \textbf{Importance sampler}
\begin{algorithmic}[1]
\STATE for each $t$, use \eqref{eq:approx} and define the approximate posterior density of $\tilde S(t)$ by
$f_{N,t}(\,\cdot\,)=w_{a,b}(\,\cdot\,)\sum_{i=0}^N \left(\sum_{r=0}^i G_{i,r}\hat\mu_{r,t}\right) G_i(\,\cdot\,)$, where $\hat\mu_{0,t}\equiv 1$\\[7pt]
\STATE draw a weighted posterior sample $(\varpi_{\ell,t}, S_{\ell,t})_{\ell = 1,\ldots,\ell_{\max}}$ of $\tilde S(t)$, of size $\ell_{\max}$, from $\pi_{N,t}(\,\cdot\,)\propto \max\big(f_{N,t}(\,\cdot\,),0\big)$ by means of the important sampler described in Section~\ref{sec:approx1}\\[7pt]
\end{algorithmic}
\end{algorithm}
\end{center}
The drawn samples allow us to approximately evaluate the posterior distribution of $\tilde S(t_i)$, for every $i=1,\ldots,q$. This, in turn, can be exploited to carry out meaningful Bayesian inference (Algorithm~\ref{algo:inference}). As a remarkable example, we consider the median survival time that we denote by $m$. The identity for the cumulative distribution function of $m$
\begin{equation*
\P\left(m\leq t\vert\bm{X}\right) = \P\big(\tilde S(t) \leq 1/2 \vert\bm{X}\big)
\end{equation*}
allows us to evaluate the CDF of $m$ at each time point $t_i$ as $c_i=\P\big(\tilde S(t_i) \leq 1/2 \vert\bm{X}\big)$. Then, we can estimate the median survival time $m$ by means of the following approximation:
\begin{equation}\label{mst}
\hat m=\mathds{E}_{\bm{X}}[m]=\int_0^\infty \P[m>t\vert\bm{X}]\:\mathrm{d} t\approx \frac{M}{q-1}\sum_{i=1}^q(1-c_i)
\end{equation}
where the subscript $\bm{X}$ in $\mathds{E}_{\bm{X}}[m]$ indicates that the integral is with respect to the distribution of $\tilde S(\cdot)$ conditional to $\bm{X}$. Equivalently,
\begin{equation}\label{mst2}
\hat m\approx \sum_{i=1}^q t_i (c_{i+1}-c_i),
\end{equation}
with the proviso that $c_{q+1}\equiv 1$. Moreover, the sequence $(c_i)_{i=1}^q$ can be used to devise credible intervals for the median survival time, cf. Part 1 of Algorithm~\ref{algo:inference}. Note that both in \eqref{mst} and in \eqref{mst2} we approximate the integrals on the left-hand-side by means of simple Riemann sums and the quality of such an approximation clearly depends on the choice of $q$ and on $M$. Nonetheless, our investigations suggest that if $q$ is sufficiently large the estimates we obtain are pretty stable and that the choice of $M$ is not crucial since, for $t$ sufficiently large, $\P\big(\tilde S(t) \leq 1/2\vert \bm{X}\big)\approx 0$. Finally, the posterior samples generated by Algorithm~\ref{algo:sampler} can be used to obtain a $t$-by-$t$ estimation of other functionals that convey meaningful information such as the posterior mode and median (together with the posterior mean), cf. Part 2 of Algorithm~\ref{algo:inference}.
\begin{center}
\begin{algorithm}
\caption{Bayesian inference\label{algo:inference}}\medskip
\textbf{Part 1.} \textbf{Median survival time}
\begin{algorithmic}[1]
\STATE use the weighted sample $(\varpi_{\ell,t_i}, S_{\ell,t_i})_{\ell = 1,\ldots,\ell_{\max}}$ to estimate, for each $i=1,\ldots,q$, $c_i=\P(\tilde S(t_i)\leq 1/2 \vert \bm{X})$\\[7pt]
\STATE plug the $c_i$'s in \eqref{mst2} to obtain $\hat m$\\[7pt]
\STATE use the sequence $(c_i)_{i=1}^q$ as a proxy for the posterior distribution of $m$ so to devise credible intervals for $\hat m$.\\[7pt]
\end{algorithmic}
\textbf{Part 2.} \textbf{$t$-by-$t$ functionals}
\begin{algorithmic}[1]
\STATE use the weighted sample $(\varpi_{\ell,t_i}, S_{\ell,t_i})_{\ell = 1,\ldots,\ell_{\max}}$ to estimate, for each $i=1,\ldots,q$, $a_{i}=\inf_{x\in [0,1]}\{\P(\tilde S(t_i)\leq x \vert \bm{X})\geq 1/2\}$ and $b_i=\mbox{mode}\{\tilde S(t_i)\vert \bm{X}\}$\\[7pt]
\STATE use the sequences $(a_{i})_{i=1}^q$ and $(b_{i})_{i=1}^q$ to approximately evaluate, $t$-by-$t$, posterior median and mode respectively\\[7pt]
\STATE use
the weighted sample $(\varpi_{\ell,t_i}, S_{\ell,t_i})_{\ell = 1,\ldots,\ell_{\max}}$ to devise $t$-by-$t$ credible intervals\\[7pt]
\end{algorithmic}
\end{algorithm}
\end{center}
\indent The rest of this section is divided in two parts in which we apply Algorithms~\ref{algo:sampler} and \ref{algo:inference} to simulated and real survival data. In Section~\ref{sec:simulated} we focus on the estimation of the median survival time for simulated samples of varying size. In Section~\ref{sec:leukemia} we analyze a real two-sample dataset and we estimate posterior median and mode, together with credible intervals, of $\tilde S(t)$. In both illustrations our approximations are based on the first $N=10$ moments.
\subsection{Application to simulated survival data}\label{sec:simulated}
We consider four samples of size $n=25,50,100,200$, from a mixture $f$ of Weibull distributions, defined by
\begin{equation*}
f = \frac{1}{2}\mbox{Wbl}(2,2)+\frac{1}{2}\mbox{Wbl}(2,1/2).
\end{equation*}
After observing that the largest observation in the samples is 4.21, we set $M=5$ and $q=100$ for the analysis of each sample. By applying Algorithms~\ref{algo:sampler} and \ref{algo:inference} we approximately evaluate, $t$-by-$t$, the posterior distribution of $\tilde S(t)$ together with the posterior distribution of the median survival time $m$. In Figure~\ref{fig:simulated_data} we focus on the sample corresponding to $n=100$. On the left panel, true survival function and Kaplan--Meier estimate are plotted. By investigating the right panel we can appreciate that the
estimated HPD credible regions
for $\tilde S(t)$ contain the true survival function. Moreover, the posterior distribution of $m$ is nicely concentrated around the true value $m_0=0.724$.
\begin{center}
\begin{figure}[ht!]
\includegraphics[width=.47\linewidth]{sim_data_KM_100.pdf}
\includegraphics[width=.51\linewidth]{sim_data_100.pdf}
\caption{(Simulated dataset, $n=100$.) Left: true survival function (red line) and Kaplan--Meier estimate (balk line). Right: true survival function (red line) and estimated posterior mean (black solid line) with 95\% HPD credible intervals for $\tilde S(t)$ (black dashed lines); the blue plot appearing in the panel on the right is the posterior distribution of the median survival time $m$.}
\label{fig:simulated_data}
\end{figure}
\end{center}
We have investigated the performance of our methodology as the sample size $n$ grows. Table~\ref{table_mst} summarizes the values we obtained for $\hat m$ and the corresponding credible intervals. For all the sample sizes considered, credible intervals for $\hat m$ contain the true value. Moreover, as expected, as $n$ grows, they shrink around $m_0$: for example the length of the interval reduces from 0.526 to 0.227 when the size $n$ changes from 25 to 200. Finally, for all these samples, the estimated median survival time $\hat m$ is closer to $m_0$ than the empirical estimator $\hat m_e$.
\begin{table}[h!]
\caption{(Simulated datasets.) Comparison of the estimated median survival time ($\hat m$) obtained by means of our Bayesian nonparametric procedure (BNP) and the empirical median survival time $\hat m_e$, for different sample sizes. For BNP estimation we show $\hat m$, the absolute error $|\hat m-m_0|$ and the 95\%-credible interval (CI); last two columns show the empirical estimate $\hat m_e$ and the corresponding absolute error $|\hat m_e - m_0|$. The true median survival time $m_0$ is 0.724.}
\begin{center}
\begin{tabular}{ l l l l l l l l }
\hline
& &\multicolumn{3}{c}{BNP} & & \multicolumn{2}{c}{Empirical}\\
\hline
sample size & & $\hat m$ & error & CI& &
$\hat m_e$ & error\\
\hline
25 && 0.803 & 0.079 & (0.598, 1.124
& & 0.578 & 0.146 \\
50 && 0.734 & 0.010 & (0.577, 0.967)
& &0.605 & 0.119 \\
100& & 0.750 & 0.026 & (0.622, 0.912
& & 0.690 & 0.034 \\
200 && 0.746 & 0.022 & (0.669, 0.896)
& & 0.701 & 0.023 \\
\hline
\end{tabular}\label{table_mst}
\end{center}
\end{table}
\subsection{Application to real survival data}\label{sec:leukemia}
We now analyze, with the described methodology, a well known two-sample dataset involving leukemia remission times, in weeks, for two groups of patients, under active drug treatment and placebo respectively. The same dataset was studied, e.g., by \cite{Cox72}.
Observed remission times for patients under treatment ($\mathsf{T}$) are
$$\{6,6,6,6^*,7,9^*,10,10^*
,11,13,16,17^*
,19^*
,20^*
,22,23,25^*
,32^*
,32^*
,34^*
,35^*\},$$
where stars denote right-censored observations. Details on the censoring mechanism and on how to adapt our methodology to right-censored observations are provided in \ref{sec:censored}. On the other side, remission times of patients under placebo ($\mathsf{P}$) are all exact and coincide with
$$\{1,1,2,2,3,4,4,5,5,8,8,8,11,11,12,12,15,17,22,23\}.$$
For this illustration we set $M=2\max(\bm{X})$, that is $M=70$, and $q=50$. For both samples we estimate and compare posterior mean, median and mode as well as 95\% credible intervals.
In the left panel of Figure~\ref{fig:simulated_examples} we have plotted such estimates for sample $\mathsf{T}$. By inspecting the plot, it is apparent that, for large values of $t$, posterior mean, median and mode show significantly different behaviors, with posterior mean being more optimistic than posterior median and mode. It is worth stressing that such differences, while very meaningful for clinicians, could not be captured by marginal methods for which only the posterior mean would be available. A fair analysis must take into account the fact that, up to $t=23$, i.e. the value corresponding to the largest non-censored observation, the three curves are hardly distinguishable. The different patterns for larger $t$ might therefore depend on the prior specification of the model. Nonetheless, we believe this example is meaningful as it shows that a more complete posterior analysis is able to capture differences, if any, between posterior mean, median and mode.
When relying on marginal methods, the most natural choice for estimating the uncertainty of posterior estimates consists in considering the quantiles intervals corresponding to the output of the Gibbs sampler, that we refer to as \emph{marginal intervals}. This leads to consider, for any fixed $t$, the interval whose lower and upper extremes are the quantiles of order $0.025$ and $0.975$, respectively, of the sample of conditional moments
$\{\mu_{1,t}^{(l_{\min}+1)},\ldots,\mu_{1,t}^{(l_{\max})}\}$
defined in \eqref{momgibbs}. In the middle panel of Figure~\ref{fig:simulated_examples} we have compared the estimated 95\% HPD intervals for $\tilde S(t)$ and the marginal intervals corresponding to the output of the Gibbs sampler. In this example, the marginal method clearly underestimates the uncertainty associated to the posterior estimates.
This can be explained by observing that, since the underlying completely random measure has already been marginalized out, the intervals arising from the Gibbs sampler output, capture only the variability of the posterior mean that can be traced back to the latent variables $\bm{Y}$ and the parameters $(c,\beta)$. As a result, the uncertainty detected by the marginal method leads to credible intervals that can be significantly narrower than the actual posterior credible intervals that we approximate through the moment-based approach. This suggests that the use of intervals produced by marginal methods as proxies for posterior credible intervals should be, in general, avoided.
We conclude our analysis by observing that the availability of credible intervals for survival functions can be of great help in comparing treatments. In the right panel of Figure~\ref{fig:simulated_examples} posterior means as well as corresponding 95\% HPD intervals are plotted for both samples $\mathsf{T}$ and $\mathsf{P}$. By inspecting the plot, for example, the effectiveness of the treatment seems clearly significant as, essentially, there is no overlap between credible intervals of the two groups.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=.323\linewidth]{mean_median_mode_1DW_a.pdf}
\includegraphics[width=.323\linewidth]{compare_marginal_1DW_a.pdf}
\includegraphics[width=.323\linewidth]{compare_subsets_1DW.pdf}
\caption{Left: comparison of posterior mean (solid line), median (dashed line) and mode (point dashed line) in dataset $\mathsf{T}$, with
95\% HPD credible intervals
(dashed line). The Kaplan--Meier estimate is plotted in red. Middle: comparison of the
95\% HPD credible interval
(dashed black line) with the marginal interval (dashed red line).
Right: comparison of samples $\mathsf{T}$ (black) and $\mathsf{P}$ (red), with posterior means (solid) and
95\% HPD credible intervals
(dashed).}
\label{fig:simulated_examples}
\end{center}
\end{figure}
\section{Introduction}
\input{introduction}
\section{Hazard mixture models}\label{sec:haz}
\input{hazard_mixture_models}
\section{Approximated inference via moments}\label{sec:approx}
\input{approximation}
\section{Bayesian inference}\label{sec:real}
\input{real_data}
\section*{Acknowledgment}
J. Arbel and A. Lijoi are supported by the European Research Council (ERC) through StG ``N-BNP'' 306406.
|
\section{General formalism}
One approach to modeling time machines is through post-selected ensembles.
The statistics of these ensembles can be equivalent to a deformation of quantum mechanics that includes projection and renormalization.
Treating time loops by selecting out inconsistent histories and renormalizing the wave-function can also be formulated this way.
The Bell state projection method consists of appending a maximally entangled pair of qubits,
then projecting against the same pair state later. One bit is the reference bit,
and the other the periodic channel or time machine state. The Bell state is given by
\begin{equation}
|\phi_{Bell}\rangle = \frac{1}{\sqrt{2}}\left( |00\rangle + |11\rangle \right)
\end{equation}
The out state that emerges from the loop is formed by the product of this with the environment.
\begin{equation}
|\psi_{B-out}\rangle = |\phi_{Bell}\rangle \otimes |\psi_{ex}\rangle\\
\end{equation}
The input state to the loop is obtained by normal unitary action of the circuit.
\begin{equation}
|\psi_{B-in}\rangle = U_t |\phi_{B-out}\rangle
\end{equation}
Finally the state is projected against the same Bell state added in the first step to produce a reduced state vector.
\begin{equation}
|\bar{\psi}_B\rangle = \langle \phi_{Bell} | \psi_{B-in}\rangle
\end{equation}
The norm of this projected state may be different from unity and so requires a renormalization factor to restore it.
\begin{eqnarray}
N^2 = \langle \bar{\psi}_B | \bar{\psi}_B \rangle\\
|\psi_{final}\rangle = N^{-1}|\bar{\psi}_B\rangle
\end{eqnarray}
As long as $N$ does not vanish, the evolution is unique and paradoxes are avoided. The value of $N$ however can have a physical meaning. It gives the relative frequency for the post-selection criteria to be met, and has measurable effects on the statistics of entanglement within the projected model.
This model has the advantage of mapping pure states to pure states,
but fails for evolutions that map the Bell state bits onto one of the three orthogonal states.
\begin{eqnarray}
|\perp_1\rangle = \frac{1}{\sqrt{2}}\left( |00\rangle - |11\rangle \right)\\
|\perp_2\rangle = \frac{1}{\sqrt{2}}\left( |01\rangle + |10\rangle \right)\\
|\perp_3\rangle = \frac{1}{\sqrt{2}}\left( |01\rangle - |10\rangle \right)
\end{eqnarray}
These are produced by the action of a phase flip gate, a not gate and the combination of the two respectively.
In each case $N$ vanishes.
Due to the condition that the external evolution $U_t$ leaves the reference bit unchanged,
we can express the projected state as a sum of two other projected states.
\begin{equation}
|\bar{\psi}_B\rangle = |\bar{\psi_{00}}\rangle + |\bar{\psi_{11}}\rangle = \langle 0 | U_t | 0,\psi_{ex} \rangle + \langle 1 | U_t |1,\psi_{ex}\rangle
\end{equation}
The two components are each formed by appending only single channel eigenstate to $\psi_{ex}$ and then projecting against it again later.
The classical limit is formed by instead taking the decoherent mixture of these two component states.
\begin{eqnarray}
\rho_{cl} = Z^{-1} \left( \bar{|\psi}_{00}\rangle \langle \bar{\psi}_{00} | + |\bar{\psi}_{11}\rangle\langle \bar{\psi}_{11} | \right)\\
Z = \langle \bar{\psi}_{00} | \bar{\psi}_{00} \rangle + \langle \bar{\psi}_{11} | \bar{\psi}_{11} \rangle
\end{eqnarray}
The classical partition function only vanishes if both terms vanish. The noisy periodic classical channel
can be described with a single bit error rate parameter $k$.
\begin{eqnarray}
|\bar{\psi}_{01}\rangle = \langle 0 | U_t | 1,\psi_{ex}\rangle \\
|\bar{\psi}_{10}\rangle = \langle 1 | U_t | 0,\psi_{ex}\rangle \\
Z = (1-k)\langle \bar{\psi}_{00} | \bar{\psi}_{00} \rangle + (1-k)\langle \bar{\psi}_{11} | \bar{\psi}_{11} \rangle + k \langle \bar{\psi}_{01} | \bar{\psi}_{01} \rangle + k \langle \bar{\psi}_{10} | \bar{\psi}_{10} \rangle \\
\rho_k = Z^{-1} \left( (1-k)| \bar{\psi}_{00} \rangle\langle \bar{\psi}_{00} | + (1-k)| \bar{\psi}_{11} \rangle\langle \bar{\psi}_{11} | + k | \bar{\psi}_{01} \rangle\langle \bar{\psi}_{01} | + k | \bar{\psi}_{10} \rangle\langle \bar{\psi}_{10} | \right)
\end{eqnarray}
The most general approach is to consider a joint distribution over two states
\begin{eqnarray}
|\bar{\psi}(\phi_a,\phi_b) \rangle = \langle \phi_a | U_t | \phi_b , \psi_{ex} \rangle\\
Z = \int \omega( \phi_a,\phi_b) \langle \bar{\psi}(\phi_a,\phi_b) |\bar{\psi}(\phi_a,\phi_b) \rangle d\phi_a d\phi_b\\
\rho_\omega = Z^{-1}\int \omega( \phi_a,\phi_b) |\bar{\psi}(\phi_a,\phi_b) \rangle \langle\bar{\psi}(\phi_a,\phi_b) | d\phi_a d\phi_b
\end{eqnarray}
Each combination of states $(\phi_a,\phi_b)$ represents a different history in the time machine picture, where a state $\phi_b$ emerges from the time machine channel and a state $\phi_a$ enters it after interaction with the environment via $U_t$. In the post-selection picture, we select histories to be a part of the ensemble based on a measurement of $\phi_a$ and $\phi_b$ at the appropriate times.
The above expression for the partition function $Z$ can be reduced to the sum over pairs of the eigenstates $e_i$.
\begin{equation}
Z = \sum_{ij} \omega_{ij} \langle e_i,\psi_{ex}| U_t^\dagger | e_j \rangle\langle e_j | U_t | e_i \psi_{ex} \rangle
\end{equation}
The standard Bell model can be seen as a delta function choice for $\omega$
\begin{equation}
\omega_{Bell} = \delta(\phi_a - \phi_{Bell})\delta(\phi_b -\phi_{Bell})
\end{equation}
And the noisy classical limit can be expressed as
\begin{eqnarray}
\omega_{Classical} = (1-k)\delta(\phi_a - 0)\delta(\phi_a - 0)+(1-k)\delta(\phi_a - 1)\delta(\phi_a - 1)+ \nonumber\\
k\delta(\phi_a - 0)\delta(\phi_a - 1)+k\delta(\phi_a - 1)\delta(\phi_a - 0)
\end{eqnarray}
This description has a high degree of redundancy.
The full distribution function can be replaced by its moments multiplied by delta functions at each combination of eigenvalues.
If all of the moments are positive then the partition function should be non-vanishing for any unitary evolution $U_t$. To add noise to the Bell model we can add three delta functions at the three states orthogonal to the Bell state. The final density matrix of the system will be a mixture of four orthogonal projections.
\begin{eqnarray}
|\bar{\psi}_B\rangle = \langle 00 + 11 | U_t | \phi_{Bell}, \psi_{ex} \rangle\\
|\bar{\psi}_-\rangle = \langle 00 - 11 | U_t | \phi_{Bell}, \psi_{ex} \rangle\\
|\bar{\psi}_N\rangle = \langle 01 + 10 | U_t | \phi_{Bell}, \psi_{ex} \rangle\\
|\bar{\psi}_{N-}\rangle \langle 01 - 10 | U_t | \phi_{Bell}, \psi_{ex} \rangle\\
\langle \bar{\psi}_B|\bar{\psi}_B\rangle + \langle \bar{\psi}_-|\bar{\psi}_-\rangle + \langle \bar{\psi}_N|\bar{\psi}_N\rangle + \langle \bar{\psi}_{N-}|\bar{\psi}_{N-}\rangle =1
\end{eqnarray}
To simulate the standard depolarizing channel we use a partition function of
\begin{equation}
Z_{\lambda} = (1-\lambda) \langle\bar{\psi}_B |\bar{\psi}_B\rangle + \lambda/4
\end{equation}
and a density matrix of
\begin{eqnarray}
\rho_\lambda = Z^{-1}(1-\frac{3\lambda}{4}) |\bar{\psi}_B \rangle\langle \bar{\psi}_B | + \frac{\lambda}{4Z} |\bar{\psi}_- \rangle\langle \bar{\psi}_- |\nonumber\\
+ \frac{\lambda}{4Z} |\bar{\psi}_N \rangle\langle \bar{\psi}_N |+ \frac{\lambda}{4Z} |\bar{\psi}_{N-} \rangle\langle \bar{\psi}_{N-} |
\end{eqnarray}
The classical noisy channel can be obtained by the choice
\begin{eqnarray}
Z_{cl} = (1-k)\langle \bar{\psi}_B|\bar{\psi}_B\rangle + (1-k)\langle \bar{\psi}_-|\bar{\psi}_-\rangle + k\langle \bar{\psi}_N|\bar{\psi}_N\rangle + k\langle \bar{\psi}_{N-}|\bar{\psi}_{N-}\rangle\\
\rho_{cl} = \frac{1-k}{Z} |\bar{\psi}_B \rangle\langle \bar{\psi}_B | +\frac{1-k}{Z} |\bar{\psi}_- \rangle\langle \bar{\psi}_- |\nonumber\\
+ \frac{k}{Z}|\bar{\psi}_N \rangle\langle \bar{\psi}_N |+ \frac{k}{Z} |\bar{\psi}_{N-} \rangle\langle \bar{\psi}_{N-} |
\end{eqnarray}
Where the equal weight given to $\bar{\psi}_-$ decoheres the state along the $0-1$ basis.
If we choose the weight function to be a constant then ensemble will be unskewed and the resulting statistics those of a completely random channel.
\begin{eqnarray}
Z_{flat} = \int \langle \phi_b \psi_{ex} | U_t^\dagger | \phi_a \rangle\langle \phi_a | U_t | \phi_b \psi_{ex} \rangle d\phi_a d\phi_b\nonumber\\
=\sum_{ij}^d |\langle e_j | U_t | e_i\psi_{ex}\rangle|^2\left((2\pi)^{d-1}\int_0^\pi\int_{S_{d-2}} \cos^2\theta \sin^{d-2}\theta d\theta \cdot dS_{d-2}\right)^2 \nonumber\\
= (2\pi)^{2d-2}S_{d-1}^2 \frac{1}{d}\\
\rho_{flat} = Z_{flat}^{-1}\int \langle \phi_a | U_t | \phi_b \psi_{ex} \rangle\langle \phi_b \psi_{ex} | U_t^\dagger | \phi_a \rangle d\phi_a d\phi_b \nonumber\\
= Z^{-1} \sum_{ij}^d \langle e_j|U_t|e_i\psi_{ex}\rangle\langle e_i \psi_{ex}|U_t^\dagger|e_j\rangle\left((2\pi)^{d-1}\int_0^\pi\int_{S_{d-2}} \cos^2\theta \sin^{d-2}\theta d\theta \cdot dS_{d-2}\right)^2 \nonumber\\
= \frac{1}{d}\sum_{ij}^d \langle e_j|U_t|e_i\psi_{ex}\rangle\langle e_i \psi_{ex}|U_t^\dagger|e_j\rangle
\end{eqnarray}
for the $n$-qubit channel. Other notable choices of weight function are
\begin{equation}
\omega_{quad}(\phi_a,\phi_b) = |\langle\phi_a|\phi_b\rangle|^2\\
\end{equation}
and
\begin{equation}
\omega_\delta(\phi_a,\phi_b) = \delta(\phi_a-\phi_b)
\end{equation}
This gives mixtures defined by
\begin{eqnarray}
Z_{quad}= \int |\langle \phi_a|\phi_b\rangle|^2 \langle \phi_b \psi_{ex} | U_t^\dagger | \phi_a \rangle\langle \phi_a | U_t | \phi_b \psi_{ex} \rangle d\phi_a d\phi_b \nonumber\\
=\sum_{ij}^d \omega_{ij} \langle e_i \psi_{ex} | U_t^\dagger | e_j \rangle\langle e_j | U_t | e_i \psi_{ex} \rangle\\
\omega_{ij}=\frac{2\delta_{ij} +1}{d(d+2)}(2\pi)^{2d-2}S_{d-1}^2\\
\rho_{quad}= Z^{-1}\sum_{ij}^d \omega_{ij} \langle e_j | U_t | e_i \psi_{ex} \rangle \langle e_i \psi_{ex} | U_t^\dagger | e_j \rangle
\end{eqnarray}
for the inner product correlated channel and
\begin{eqnarray}
Z_\delta = \int \delta(\phi_a-\phi_b) \langle \phi_b \psi_{ex} | U_t^\dagger | \phi_a \rangle\langle \phi_a | U_t | \phi_b \psi_{ex} \rangle d\phi_a d\phi_b \nonumber\\
=\sum_{ij}^d \omega_{ij} \langle e_i \psi_{ex} | U_t^\dagger | e_j \rangle\langle e_j | U_t | e_i \psi_{ex} \rangle\\
\omega_{ij} = \frac{1}{d}\delta_{ij}(2\pi)^{d-1}S_{d-1}\\
\rho_\delta = Z^{-1} \sum_i^d \omega_{ij}\langle e_i | U_t | e_i\psi_{ex}\rangle\langle e_i\psi_{ex}|U_t^\dagger | e_i \rangle
\end{eqnarray}
for the delta correlated channel. In the above expressions $S_d$ refers to the area of the unit $d-$sphere. The powers of $2\pi$ come from the mutual phases between components of the $d-$dimensional complex unit vectors. The spread of the correlation function then gives us the effective noise of the channel.
Later we will see that, $\omega_\delta$ gives us a channel with noise of $1/2$.
We will be using the noisy Bell and noisy classical channels for the majority of this paper.
\section{Simple loop}
This section covers the action of the bare channel. The circuit is a simple swap operation between the periodic channel and an external channel.
The treatment in the Bell state model begins with an arbitrary input qubit,
\begin{equation}
|\psi_1\rangle = \alpha|0\rangle + \beta|1\rangle
\end{equation}
we can add a maximally entangled pair in a product state with it.
\begin{equation}
|\phi_{Bell}\rangle = \frac{1}{\sqrt{2}}( |00\rangle + e^{i\theta}|11\rangle)
\end{equation}
with the left qubit as our reference bit, and the right qubit as the time machine 'out' state.
The combined state before acting with the rest of the circuit is simply
\begin{equation}
|\psi_{out}\rangle = |\phi_{Bell}\rangle \otimes |\psi_1\rangle = \frac{1}{\sqrt{2}}(\alpha|000\rangle + \alpha e^{i\theta}|110\rangle +\beta|001\rangle +\beta e^{i\theta}|111\rangle)
\end{equation}
Now to simulate sending channel $\psi_1$ into the time machine and observing what comes out, we swap the in/out channel with channel 1.
\begin{equation}
U_{swap} = \psi_1 \leftrightarrow \phi
\end{equation}
The state after the swap is now,
\begin{equation}
|\psi_{in}\rangle = U_{swap}|\psi\rangle = \frac{1}{\sqrt{2}}(\alpha|000\rangle + \alpha e^{i\theta}|101\rangle +\beta|010\rangle +\beta e^{i\theta}|111\rangle)
\end{equation}
Generally the reference bit will be unchanged by the circuit, unless we wish to model some internal dynamics of the time machine.
Now we project against the original product state $\phi_{Bell}$ removing the first two bits and leaving only the channel $\psi_1$.
\begin{equation}
|\bar{\psi}\rangle = \frac{1}{2}(\langle 00,x|\psi'\rangle + e^{-i\theta}\langle 11,x |\psi'\rangle) = \frac{1}{2}(\alpha|0\rangle + \beta|1\rangle) = \frac{1}{2}|\psi_1\rangle = N|\psi_1\rangle
\end{equation}
The $,x$ notation is a reminder that the third qubit is not contracted, but remains.
This is none other than an ordinary quantum teleportation in the case that the final measurement against $\langle \phi_{Bell} |$ returned 1.
The the square of the factor N represents the portion of ensembles that satisfy the selection condition represented by the state projection.\\
The phase of the Bell pair drops out of the final calculation but other representations are possible.
Consider a rotated pair as a reference state,
\begin{equation}
|\phi_{alt}\rangle = \frac{1}{2}\left( |00\rangle + |01\rangle + |10\rangle - |11\rangle \right)
\end{equation}
The calculation is more tedious, but gives the same result.
\begin{eqnarray}
|\psi_{out}\rangle = |\phi_{alt}\rangle \otimes |\psi_1\rangle = \frac{1}{2}( \alpha|000\rangle + \alpha|010\rangle + \alpha|100\rangle - \alpha|110\rangle +\beta|001\rangle + \beta|011\rangle + \beta|101\rangle - \beta|111\rangle )\\
|\psi_{in}\rangle = U_{swap} |\psi\rangle = \frac{1}{2}( \alpha|000\rangle + \alpha|001\rangle + \alpha|100\rangle - \alpha|101\rangle +\beta|010\rangle + \beta|011\rangle + \beta|110\rangle - \beta|111\rangle )\\
|\bar{\psi}\rangle = \frac{1}{2}\left( \langle 00,x| + \langle 01,x| + \langle 10,x| - \langle 11,x| \right) |\psi'\rangle = \frac{1}{2} (\alpha |0\rangle + \beta|1\rangle)\\
N=1/2
\end{eqnarray}
By using a non-maximally entangled pair we get a distorted signal. Consider the case
\begin{equation}
|\phi_{twist}\rangle = \frac{1}{\sqrt{2}}|00\rangle + \frac{1}{2}|01\rangle + \frac{1}{2}|11\rangle
\end{equation}
The product state after channel swapping is
\begin{equation}
|\psi_{in}\rangle = \frac{\alpha}{\sqrt{2}}|000\rangle + \frac{\alpha}{2}|001\rangle + \frac{\alpha}{2}|101\rangle + \frac{\beta}{\sqrt{2}}|010\rangle + \frac{\beta}{2}|011\rangle + \frac{\beta}{2}|111\rangle
\end{equation}
The projection of this state against $\langle \phi_{twist} |$ gives
\begin{equation}
|\bar{\psi}\rangle = \frac{1}{2}|\psi_1\rangle + \frac{\beta}{\sqrt{8}}|0\rangle + \frac{\alpha}{\sqrt{8}}|1\rangle
\end{equation}
So it is important in this case for the pair to be maximally entangled.
We can also consider a decoherent time machine model. In this case the out state is again an arbitrary product, but is a single qubit rather than a pair.
\begin{equation}
|\phi_{TM}\rangle = \gamma_0 |0\rangle + \gamma_1 |1\rangle = \cos{\theta}|0\rangle + e^{i\xi}\sin{\theta}|1\rangle
\end{equation}
Taking the product and then the swap operation, we have
\begin{eqnarray}
|\psi_{out}\rangle = |\phi_{TM}\rangle \otimes |\psi_1\rangle = \alpha\gamma_0|00\rangle + \alpha\gamma_1|10\rangle + \beta\gamma_0|01\rangle + \beta\gamma_1|11\rangle\\
|\psi_{in}\rangle = U_{swap} |\psi\rangle = \alpha\gamma_0|00\rangle + \alpha\gamma_1|01\rangle + \beta\gamma_0|10\rangle + \beta\gamma_1|11\rangle\\
|\bar{\psi}\rangle = (\gamma_0^{\dagger}\langle 0,x| + \gamma_1^{\dagger}\langle 1,x| ) |\psi'\rangle\\= (\alpha\gamma_0^{\dagger}+\beta\gamma_1^{\dagger})( \gamma_0|0\rangle + \gamma_1|1\rangle)\nonumber\\
\langle \phi_{TM}| \psi_1 \rangle \cdot |\phi_{TM}\rangle
\end{eqnarray}
Summing over an orthogonal set of states for $\phi_{TM}$ in this single qubit method is identical to the Bell state method mentioned earlier, if we fix the mutual phase of the terms. Here, the final state is the time machine out state, with a weight corresponding to how well channel $\psi_1$ matches the in state $\phi_{TM}$. Checking this with a coherent integration of $\bar{\psi}$ over $\phi_{TM}$ we have,
\begin{eqnarray}
|\bar{\psi}\rangle = \bar{\alpha}|0\rangle + \bar{\beta}|1\rangle\\
\bar{\alpha} = \alpha \cos^2{\theta} + \beta e^{-i\xi}\sin{\theta} \cos{ \theta} \\
\bar{\beta} = \alpha e^{i\xi}\sin\theta\cos\theta + \beta\sin^2\theta\\
Z = \int d\phi = 2\pi^2\\
\int \bar{\alpha} d\phi = \pi^2 \alpha\\
\int \bar{\beta} d\phi = \pi^2 \beta \\
Z^{-1}\int |\bar{\psi}\rangle d\phi = \frac{1}{2}( \alpha|0\rangle + \beta|1\rangle)= \frac{1}{2}|\psi_1\rangle
\end{eqnarray}
This should not be surprising since the operations here are all linear, and the contributions from rotated states will simply be linear combinations of the contributions from the basis states, providing the overall constant factor $Z$.\\ We can also form decoherent mixtures of the projected states $\bar{\psi}$.
\begin{eqnarray}
\rho = Z^{-1}\int |\bar{\psi}\rangle\langle \bar{\psi} | d\phi_{TM}\\
Z= \int | \langle \phi_{TM},x | \psi_{in}\rangle |^2 d\phi_{TM} =\int \langle \bar{\psi}|\bar{\psi}\rangle d\phi = \int N_\phi N_\phi^\dagger d\phi
\end{eqnarray}
This prescription comes from the $\omega_\delta$ correlation function introduced earlier. The system in this model is an ensemble average of decoherent histories. Each of those histories is characterized by what state emerges from the time machine, and weighted by the norm squared of the projection of the full state against the emitted state.
In the simple loop, this choice of weight function integrates to
\begin{equation}
Z = \int_0^\pi \int_0^{2\pi}[ \alpha^2\cos^2\theta + \beta^2\sin^2\theta + 2\alpha\beta\sin\theta\cos\theta\cos\xi ]d\theta d\xi = \pi^2
\end{equation}
The density matrix terms are
\begin{eqnarray}
\rho_{00} = Z^{-1} \int [\omega(\theta,\xi)\cos^2\theta]d\theta d\xi = \frac{1}{2}\alpha^2+\frac{1}{4}\\
\rho_{11} = Z^{-1}\int [\omega(\theta,\xi)\sin^2\theta]d\theta d\xi = \frac{1}{2}\beta^2+\frac{1}{4}\\
\rho_{10}=\rho_{01}^\dagger = \frac{1}{4}\alpha\beta
\end{eqnarray}
which can be expressed more simply as
\begin{equation}
\rho_{TM} = \frac{1}{2} |\psi_1\rangle\langle \psi_1 | + \frac{1}{4}I
\end{equation}
So this weight function gives a channel with a $1:1$ signal noise ratio.
The $\delta$-weight model is also investigated by \cite{mark}, under the formalism called T-CTCs.
The treatment of the unproven proof in the $\delta$-model also loses coherence, and can be thought of as a realization of the mixing guessed by Hawking in \cite{hawk}.
The two qubit simple loop circuit can be expressed in the Bell state model by taking the product of the in state with a maximally entangled pair of 2 qubit Hadamard states.
The projection state is given by
\begin{equation}
|\phi_{TM}\rangle = \frac{1}{2}\left( |0000\rangle + |0101\rangle + |1010\rangle + |1111\rangle\right)
\end{equation}
Defining the input state to be transported as
\begin{equation}
|\psi_1\rangle = \gamma_{00}|00\rangle + \gamma_{01}|01\rangle + \gamma_{10}|10\rangle + \gamma_{11}|11\rangle
\end{equation}
Then the out state is simply
\begin{equation}
|\psi_{out}\rangle = |\phi_{TM}\rangle \otimes |\phi_1\rangle
\end{equation}
The swap operation gives us the in state that will be projected against $\phi_{TM}$.
\begin{equation}
|\psi_{in}\rangle = U_{swap52}U_{swap64}|\psi_{out} \rangle
\end{equation}
Projecting gives the state
\begin{equation}
|\bar{\psi}\rangle = \langle \phi_{TM} | \psi_{in}\rangle = \frac{1}{4}|\psi_1\rangle
\end{equation}
The classical limit will decohere the incoming channel along the time machine's preferred basis. To see this we consider the weights.
First the different out states will be defined by the product of $\psi_1$ with each eigenstate of the time machine.
\begin{equation}
|\psi_{out}(i)\rangle = |e_i\rangle \otimes |\psi_1\rangle
\end{equation}
and the in states in now acting only on 4 qubits so that the swap operation is
\begin{equation}
|\psi_{in}(i)\rangle = U_{swap13}U_{swap24}|\psi_{out}(i)\rangle
\end{equation}
Now the weight of each eigenstate will be
\begin{eqnarray}
\omega_{00} = k/4 + (1-k)|\langle 00,xx |\psi_{in}(00)\rangle|^2 = k/4 +(1-k)\gamma_{00}\gamma_{00}^\dagger \\
\omega_{01} = k/4 + (1-k)|\langle 01,xx |\psi_{in}(01)\rangle|^2 = k/4 +(1-k)\gamma_{01}\gamma_{01}^\dagger \\
\omega_{10} = k/4 + (1-k)|\langle 10,xx |\psi_{in}(10)\rangle|^2 = k/4 +(1-k)\gamma_{10}\gamma_{10}^\dagger \\
\omega_{11} = k/4 + (1-k)|\langle 11,xx |\psi_{in}(11)\rangle|^2 = k/4 +(1-k)\gamma_{11}\gamma_{11}^\dagger
\end{eqnarray}
where we have renormalized the noise constant to reflect the larger number of eigenstates. This gives the classical partition function as
\begin{equation}
Z_{cl} = k + (1-k)(\gamma_{00}\gamma_{00}^\dagger+\gamma_{01}\gamma_{01}^\dagger+\gamma_{10}\gamma_{10}^\dagger+\gamma_{11}\gamma_{11}^\dagger) = 1
\end{equation}
And an output density matrix of
\begin{eqnarray}
\rho_{cl} = \omega_{00}|00\rangle\lr0 0| +\omega_{01}|01\rangle\langle 01| +\omega_{10}|10\rangle\langle 10| +\omega_{11}|11\rangle\langle 11| \nonumber\\
= \frac{k}{4}I + (1-k)diag (\gamma_{00}\gamma_{00}^\dagger ,\gamma_{01}\gamma_{01}^\dagger ,\gamma_{10}\gamma_{10}^\dagger ,\gamma_{11}\gamma_{11}^\dagger )
\end{eqnarray}
With a few basics out of the way, we now proceed to examine the most iconic scenario in acausal physics, the well known grandfather paradox.
\section{Grandfather circuit}
This and the faulty gun circuit are also discussed in \cite{seth2}, and included here for comparison and commentary.
One of the most basic and important circuits to analyze is the quintessential causal paradox, where whichever state emerges, an orthogonal state is sent in by the action of $U_t$. Beginning with the Bell state
\begin{equation}
|\phi_{Bell}\rangle = \frac{1}{\sqrt{2}}( |00\rangle + |11\rangle)
\end{equation}
the paradox circuit is defined by acting on $\phi_{out}$ with a simple not gate. To begin with we will consider any external system to remain in a product state.
\begin{equation}
|\psi_{in}\rangle = U_{not2}|\psi_{out}\rangle = U_{not2}|\phi_{Bell}\rangle \otimes |\psi_{ext}\rangle= \frac{1}{\sqrt{2}}( |01\rangle + |10\rangle) \otimes |\psi_{ext}\rangle
\end{equation}
Where $U_{not2}$ implements a not operation on the second bit of the Bell pair. In general we will assume the first bit of the Bell pair to be unchanged by any circuit.
The projection operation now gives zero, indicating that the selection criteria are never met.
\begin{equation}
\langle \phi_{TM} | \psi_{in}\rangle = 0 \otimes |\psi_{ext}\rangle
\end{equation}
\begin{figure}[h!]
\caption{Grandfather circuit with decohering measurements.}
\centering
\includegraphics[width=0.4\textwidth]{grandf2.png}
\end{figure}
To better understand the situation, we can look at a perturbation of the state, either with noise or with a small rotation away from an exact not gate.
First consider the perturbation of $U_t$ such that
\begin{equation}
\tilde{U}_t=(1-\epsilon)U_{not} + \epsilon I
\end{equation}
Now we have
\begin{equation}
|\tilde{\psi}\rangle = \tilde{U}_t |\psi\rangle = \left[\frac{1-\epsilon}{\sqrt{2}}\left(|01\rangle+|10\rangle \right) + \frac{\epsilon}{\sqrt{2}}\left( |00\rangle + |11\rangle \right)\right] \otimes |\psi_{ext}\rangle
\end{equation}
Now the projection gives
\begin{equation}
\langle \phi_{TM} | \tilde{\psi} \rangle = \epsilon |\psi_{ext}\rangle
\end{equation}
and so
\begin{equation}
N=\epsilon
\end{equation}
We can now renormalize $\bar{\psi}$, but there is still a problem.
A unique limit does not exist in this case. Different perturbations of $U_{not}$ will give different projections, as the direction of perturbation will determine the direction of $\bar{\psi}$. Consider a different perturbation
\begin{equation}
\tilde{U}_t=(1-\epsilon)U_{not} + \epsilon U_\xi
\end{equation}
this projection gives
\begin{equation}
\langle \phi_{TM} | \tilde{\psi} \rangle = \epsilon U_\xi|\psi_{ext}\rangle
\end{equation}
Paradoxes are singular points in the behavior of the projection. The closer one gets to a circuit that has a vanishing normalization factor, the more sensitive the projection result becomes to small perturbations.
In general we will see that the magnitude of the normalization factor will be a measure of how far from normal the behavior of the system will be. The same result can be found with a noise term. Here we perturb the projection operator.
\begin{equation}
\langle \phi_{TM}|_{eff} = \frac{1-k}{\sqrt{2}}\left(\langle 00| + \langle 11| \right) + e^{i\vartheta}\frac{k}{\sqrt{2}}\left( \langle 01| + \langle 10|\right)
\end{equation}
This represents a small amplitude for the qubit to flip during 'transit', such that
\begin{equation}
N=k
\end{equation}
If we take the classical limit for a decoherent time machine, the partition function gives
\begin{equation}
Z = \epsilon + \int \omega(\phi) d\phi = \epsilon + |\langle 0| U_{not} | 0\rangle |^2 + | \langle 1 | U_{not} | 1\rangle |^2 = \epsilon
\end{equation}
This gives still gives zero for the density matrix unless the weight function is also perturbed, in which case
\begin{equation}
\rho = Z^{-1} \left( \omega(|0\rangle)|0\rangle\langle 0| + \omega(|1\rangle) |1\rangle\langle 1| \right) = \frac{1}{2}I
\end{equation}
For a partially decoherent weight model such as
\begin{equation}
\omega = | \langle \phi,x | U_{not} | \phi\otimes\psi \rangle |^2
\end{equation}
where $\phi$ is given by
\begin{equation}
|\phi\rangle = \gamma_0 |0\rangle + \gamma_1 e^{i\xi} | 1 \rangle = \cos \theta |0\rangle + \sin \theta e^{i\xi}|1\rangle
\end{equation}
now
\begin{equation}
\omega_{not}(\phi) = 4\cos^2 \theta \sin^2 \theta \cos^2 \xi
\end{equation}
for a simple not gate or
\begin{equation}
\omega_{rot}(\phi) = 4\cos^2 \theta \sin^2 \theta \sin^2 \xi
\end{equation}
for a not and phase flip.
This gives us a partition function of
\begin{equation}
Z_{not}=Z_{rot} = \frac{\pi^2}{2}
\end{equation}
and density matrix of
\begin{equation}
\rho_{not}=\rho_{rot} = Z^{-1}\int \omega |\bar{\psi}\rangle\langle \bar{\psi} | = \frac{1}{2}(|0\rangle\langle 0| + |1\rangle\langle 1|)
\end{equation}
in more detail for $\rho_{not}$,
\begin{eqnarray}
\rho_{00} = \frac{2}{\pi^2}\int_0^\pi \int_0^{2\pi} 4[\cos^4\theta\sin^2\theta\cos^2\xi ]d\theta d\xi =\frac{1}{2}\\
\rho_{01} = \frac{2}{\pi^2}\int_0^\pi \int_0^{2\pi} 4e^{i\xi}[\cos^3\theta\sin^3\theta\cos^2\xi ]d\theta d\xi =0\\
\rho_{11} = \frac{2}{\pi^2}\int_0^\pi \int_0^{2\pi} 4[\cos^2\theta\sin^4\theta\cos^2\xi ]d\theta d\xi =\frac{1}{2}\\
\rho_{10} = \frac{2}{\pi^2}\int_0^\pi \int_0^{2\pi} 4e^{-i\xi}[\cos^3\theta\sin^3\theta\cos^2\xi ]d\theta d\xi =0
\end{eqnarray}
And for $\rho_{rot}$ ,
\begin{eqnarray}
\rho_{00} = \frac{2}{\pi^2}\int_0^\pi \int_0^{2\pi} 4[\cos^4\theta\sin^2\theta\sin^2\xi ]d\theta d\xi =\frac{1}{2}\\
\rho_{01} = \frac{2}{\pi^2}\int_0^\pi \int_0^{2\pi} 4e^{i\xi}[\cos^3\theta\sin^3\theta\sin^2\xi ]d\theta d\xi =0\\
\rho_{11} = \frac{2}{\pi^2}\int_0^\pi \int_0^{2\pi} 4[\cos^2\theta\sin^4\theta\sin^2\xi ]d\theta d\xi =\frac{1}{2}\\
\rho_{10} = \frac{2}{\pi^2}\int_0^\pi \int_0^{2\pi} 4e^{-i\xi}[\cos^3\theta\sin^3\theta\sin^2\xi ]d\theta d\xi =0
\end{eqnarray}
In each decoherent model, the paradox maximizes entropy for the circulating qubit. Another implementation of the paradox is the controlled phase flip circuit. It is essentially another rotation of the previous case, included here for completeness.
\begin{equation}
U_{cpf} = |x,0\rangle\langle x,0| - |x,1\rangle\langle x,1|
\end{equation}
The joint state in the standard Bell representation is
\begin{equation}
|\psi_{out}\rangle = |\phi_{Bell}\rangle\otimes |\psi_1\rangle = \frac{1}{\sqrt{2}}( |00\rangle + |11\rangle ) \otimes ( \alpha |0\rangle + \beta | 1\rangle )
\end{equation}
The phase flip gives the in state as
\begin{equation}
|\psi_{in}\rangle = U_{cpf}|\psi_{out}\rangle = \frac{\alpha}{\sqrt{2}} (|00\rangle + |11\rangle)\otimes |0\rangle + \frac{\beta}{\sqrt{2}} ( |00\rangle - |11\rangle ) \otimes |1\rangle
\end{equation}
The projection is then
\begin{equation}
\langle \phi_{Bell}^\dagger , x| \psi_{in}\rangle = \alpha|0\rangle
\end{equation}
Destructive interference in the looping qubit removes the part of the state along $|1\rangle$.
If $\alpha$ vanishes then the grandfather paradox occurs.
A third implementation of the grandfather circuit is available by combining the phase flip and not gates.
These three grandfather circuits represent mappings that take the initial Bell state to one of the three orthogonal entangled states.
In the classical case, only the not circuit realizes the paradox, since the classical limit makes the weight of the post-selected ensemble indifferent to the mutual phase between the classical eigenstates.
Testing the behavior of the controlled phase flip in the $\delta-$correlated model
\begin{eqnarray}
|\psi\rangle = |\phi\rangle \otimes |\psi_1\rangle \\
|\bar{\psi}\rangle = \langle \phi ,x| U_{cpf} |\psi\rangle \\
= \alpha|0\rangle + \beta(\cos^2\theta - \sin^2\theta)|1\rangle\\
\omega(\theta) = \langle \bar{\psi} | \bar{\psi}\rangle = 1 - 4\beta^2(\sin^2\theta -\sin^4\theta)
\end{eqnarray}
This gives us a partition function of
\begin{equation}
Z = 2\pi\int_0^\theta \omega(\theta) d\theta = 2\pi^2(1-\beta^2/2)
\end{equation}
The density matrix for the circulating qubit will be
\begin{eqnarray}
\rho_{TM} = Z^{-1}\int_0^{2\pi}\int_0^\pi \omega(\theta)|\phi \rangle\langle\phi|d\xi d\theta \\
= \int_0^{2\pi}\int_0^\pi\frac{1 - 4\beta^2(\sin^2\theta -\sin^4\theta)}{2\pi^2(1-\beta^2/2} \left( \begin{array}{cc} \cos^2\theta & e^{i\xi}\sin\theta\cos\theta \\\\ e^{-i\xi}\sin\theta\cos\theta & \sin^2\theta \end{array} \right)d\xi d\theta\\
=\frac{1}{2}I
\end{eqnarray}
And for the control qubit we have
\begin{eqnarray}
\rho_{cpf} = \frac{2\pi}{Z}\int_0^\pi |\bar{\psi}\rangle\langle\bar{\psi}|d\theta = \int_0^\pi\left( \begin{array}{cc} \alpha^2 & \alpha\beta(1-2\sin^2\theta) \\\\ \alpha\beta^\dagger(1 - 2\sin^2\theta) & \beta\beta^\dagger(1-2\sin^2\theta)^2 \end{array} \right)\frac{d\theta}{\pi(1-\beta^2/2)}\nonumber\\
=\left( \begin{array}{cc} 2\alpha^2/(\alpha^2+1) & 0 \\\\ 0 & \beta\beta^\dagger/(\alpha^2+1) \end{array} \right)
\end{eqnarray}
This particular decoherent time machine channel is half noise. Consequently the reduction in permitted states as well as the back-propagation effect onto the control qubit are both half strength, rather than complete elimination. The loss of off diagonal terms is due to the decoherent nature of the circulating bit.\\
\begin{figure}[h!]
\caption{Phase flip version of the grandfather circuit.}
\centering
\includegraphics[width=0.5\textwidth]{cpf.png}
\end{figure}
In contrast to the previous examples, the phase flip operation does not have as significant an effect on the classical time machine channel. For the classical channel the weights are
\begin{eqnarray}
\langle 0,x | U_{cpf} | 0,\psi_1\rangle = \alpha|0\rangle + \beta|1\rangle\\
\langle 1,x | U_{cpf} | 1,\psi_1\rangle = \alpha|0\rangle - \beta|1\rangle\\
\omega_0 = k/2 + (1-k)| \langle 0,x | U_{cpf} | 0,\psi_1 \rangle|^2= 1 - k/2\\
\omega_1 = k/2 + (1-k)| \langle 1,x | U_{cpf} | 1,\psi_1 \rangle|^2 = 1 - k/2\\
Z= 2-k
\end{eqnarray}
The partition function is independent of the control qubit, in contrast to both other cases. This is because the classical channel is only sensitive to the eigenvalues of the bit in its preferred decohering basis, and a the phase flip does not affect those values.
The partition function for the time machine channel is
\begin{equation}
\rho_{cl} = (2-k)^{-1}\left(\left (1-\frac{k}{2}\right)|0\rangle\langle 0| + \left(1-\frac{k}{2}\right)|1\rangle\langle 1|\right) = \frac{1}{2}I
\end{equation}
and the control channel is given by
\begin{equation}
\rho_{clfp} = \frac{1}{2}\left( \alpha^2 |0\rangle\langle 0| + \beta\beta^\dagger |1\rangle\langle 1| \right)
\end{equation}
Again the interaction with a decoherent channel destroys the off diagonal terms. The effect is identical to measurement projection, with no need for back-propagation. The controlling bit becomes entangled with a forever hidden channel, making it too 'classical'. Circuits with more noise or less coherence seem to have less extreme effects than coherent noiseless ones. The requirement of noise or perturbation hints at more to be learned however. This brings us to the next circuit.
\section{Faulty gun circuit}
Taking a que from the perturbation of the grandfather scenario, let us consider general rotations and controlled rotations of the circulating qubit. Beginning with the Bell out state,
\begin{equation}
|\phi_{TM}\rangle = \frac{1}{\sqrt{2}}( |00\rangle + |11\rangle)
\end{equation}
We act with a rotation
\begin{equation}
U_\zeta = \cos\zeta |0\rangle\langle 0| + \sin\zeta |1\rangle\langle 0| + \cos\zeta |1\rangle\langle 1| - \sin\zeta |0\rangle\langle 1|
\end{equation}
onto the out channel, giving
\begin{equation}
|\psi'\rangle = U_\zeta |\phi\rangle = \frac{\cos\zeta}{\sqrt{2}} ( |00\rangle + |11\rangle ) + \frac{\sin\zeta}{\sqrt{2}}( |01\rangle - |10\rangle )
\end{equation}
The normalization constant is then given again by the inner product with the original Bell state,
\begin{equation}
N = \langle \phi_{TM} | U_\zeta | \phi_{TM} \rangle = \cos\zeta
\end{equation}
The parameter $\zeta$ rotates between the identity and the version of the grandfather circuit given by $U_{rot}$ in the previous section. The grandfather paradox occurs in the exact model when $N$ vanishes.
\begin{figure}[h!]
\caption{Measuring the effect of a single rotation on the periodic channel $\phi$.}
\centering
\includegraphics[width=0.5\textwidth]{faulty2.png}
\end{figure}
In the noisy classical limit we have a partition function of
\begin{equation}
Z = k + 2(1-k)\cos^2\zeta
\end{equation}
This reflects the relative number of consistent histories available to the system. The $\delta-$correlated channel gives,
\begin{equation}
\omega = |\langle \phi | U_\zeta | \phi \rangle |^2
\end{equation}
and integrating over $|\phi\rangle$,
\begin{equation}
|\phi\rangle = \gamma_0|0\rangle + \gamma_1 e^{i\xi}|1\rangle = \cos\theta |0\rangle + e^{i\xi}\sin\theta|1\rangle
\end{equation}
we have
\begin{equation}
\omega = | \cos\zeta - 2i \cos\theta \sin\theta \sin\zeta\sin\xi |^2
\end{equation}
giving a partition function of
\begin{equation}
Z = \frac{\pi^2}{2}( 3\cos^2\zeta + 1 )
\end{equation}
and a density matrix of
\begin{eqnarray}
\rho_{00} = Z^{-1} \int_0^\pi \int_0^{2\pi} \cos^2\theta [ \cos^2\zeta + 4\sin^2\zeta \sin^2 \xi \sin^2 \theta \cos^2 \theta ]d\theta d\xi =1/2\\
\rho_{11} = Z^{-1} \int_0^\pi \int_0^{2\pi} \sin^2\theta [ \cos^2\zeta + 4\sin^2\zeta \sin^2 \xi \sin^2 \theta \cos^2 \theta ]d\theta d\xi =1/2\\
\rho_{10}=\rho_{01} = Z^{-1} \int_0^\pi \int_0^{2\pi} e^{i\xi}\sin\theta\cos\theta [ \cos^2\zeta + 4\sin^2\zeta \sin^2 \xi \sin^2 \theta \cos^2 \theta ]d\theta d\xi =0
\end{eqnarray}
The mixed state is mainly a function of the symmetry of the circuit, as well as its factorization. This effect will present itself in several other circuits as well. The role of the relative value of the partition function and normalization parameter as a sort of prior probability is more evident when we consider the controlled not version of the circuit.\\
Starting with an input control qubit defined as
\begin{equation}
|\psi_1\rangle = \alpha|0\rangle + \beta|1\rangle
\end{equation}
we can add the maximally entangled pair of the in state and referce qubit.
\begin{equation}
|\phi_{out}\rangle = \frac{1}{\sqrt{2}}( |00\rangle + |11\rangle)
\end{equation}
The joint state is acted on by a controlled not
\begin{equation}
U_{cn23} = |x,00\rangle\langle x,00| + |x,01\rangle\langle x,01| + |x,10\rangle\langle x,11| + |x,11\rangle\langle x,10|
\end{equation}
giving an in state of
\begin{equation}
|\phi_{in}\rangle = U_t|\phi_{out}\psi_1\rangle = \frac{\alpha}{\sqrt{2}}( |000\rangle + |110\rangle) + \frac{\beta}{\sqrt{2}}( |011\rangle + |101\rangle)
\end{equation}
projecting onto the pair, we get
\begin{equation}
\langle \phi_{out}^\dagger,x | U_t |\phi_{in}\rangle = \alpha |0\rangle = N|0\rangle
\end{equation}
The control qubit is forced into the state that avoids the paradox, regardless of the original amplitude or phase relative to the rest of the wave-function. Furthermore, the result holds even if we consider the control channel to be macroscopic. This is an indicator of how the nonlinear effects of state selection can propagate backwards into what we would normally consider the initial states of the system. Consider the action of $C_{not}$ from the outside channel onto a decoherent time machine channel.
\begin{figure}[h!]
\caption{Controlled not acting on $\phi$.}
\centering
\includegraphics[width=0.3\textwidth]{gran3.png}
\end{figure}
The $\delta$ model in the controlled not circuit gives,
\begin{eqnarray}
|\psi_1\rangle = \alpha|0\rangle + \beta|1\rangle\\
|\phi_{out}\rangle = \gamma_0 |0\rangle + \gamma_1 |1\rangle = \cos{\theta}|0\rangle + e^{i\xi}\sin{\theta}|1\rangle\\
|\psi\rangle = \alpha \cos\theta |00\rangle + \alpha e^{i\xi}\sin\theta |10\rangle + \beta\cos\theta |01\rangle + \beta e^{i\xi}\sin\theta |11\rangle
\end{eqnarray}
we get for the in state
\begin{equation}
|\phi_{in}\rangle = U_t|\psi\rangle = \alpha \cos\theta |00\rangle + \alpha e^{i\xi}\sin\theta |10\rangle + \beta\cos\theta |11\rangle + \beta e^{i\xi}\sin\theta |01\rangle
\end{equation}
which projects to the state,
\begin{equation}
|\bar{\psi}\rangle = \langle \phi_{out}^\dagger,x | U_t | \phi_{in} \rangle = \alpha |0\rangle + 2\beta \sin\theta \cos \theta \cos \xi |1\rangle
\end{equation}
the weight function is
\begin{equation}
N^2 = \omega(\theta,\xi) = | \langle \phi_{out}^\dagger,x | U_t | \phi_{in} \rangle|^2 = \alpha^2 + 4\beta^2 \cos^2 \xi \sin^2\theta \cos^2\theta
\end{equation}
and partition function of
\begin{equation}
Z = \int_0^\pi \int_0^{2\pi} N^2 d\theta d\xi = \frac{\pi^2}{2}( 3\alpha^2 + 1 )
\end{equation}
and a density matrix given by
\begin{eqnarray}
\rho_{00} = \frac{2}{\pi^2}( 3\alpha^2 + 1 )^{-1}\int_0^\pi \int_0^{2\pi} \omega(\theta,\xi)\cos^2\theta = \frac{1}{2}\\
\rho_{11} = \frac{2}{\pi^2}( 3\alpha^2 + 1 )^{-1}\int_0^\pi \int_0^{2\pi} \omega(\theta,\xi)\sin^2\theta = \frac{1}{2}\\
\rho_{01} = \rho_{10} = 0
\end{eqnarray}
We can also do an expectation value for the control bit, since the weight is also a function of $\alpha$ and $\beta$. The partition function $Z$ will serve as a weight that skews the statistics of a random mixture of control bits. In this case if we use $\zeta$ as the angle of the flip component of the control bit
\begin{eqnarray}
|\psi_1\rangle = \cos\zeta |0\rangle + e^{i\vartheta}\sin\zeta |1\rangle \\
\omega_{\zeta} = Z_{TM}(\alpha) = \frac{\pi^2}{2}(3\alpha^2 + 1) = \frac{\pi^2}{2}(3\cos^2\zeta + 1)
\end{eqnarray}
A linear combination of density matrices weighted by $Z(\alpha)$ will give us
\begin{eqnarray}
Z_{\zeta} = 2\pi\int_0^{\pi} \left[ \frac{\pi^2}{2}(3\cos^2\zeta + 1) \right] d\zeta = \frac{5}{2}\pi^4 \\
\langle 0|\rho_\zeta|0\rangle = \frac{2}{5}\pi^{-4} \int_0^{2\pi}d\vartheta \int_0^\pi d\zeta \left[\frac{\pi^2}{2}( 3\cos^4\zeta + \cos^2\zeta)\right] = \frac{13}{20}\\
\langle 1|\rho_\zeta|1\rangle = \frac{2}{5}\pi^{-4} \int_0^{2\pi}d\vartheta \int_0^\pi d\zeta \left[\frac{\pi^2}{2}( 3\sin^2\zeta \cos^2\zeta + \sin^2\zeta)\right] = \frac{7}{20}\\
\langle 0|\rho_\zeta|1\rangle = \frac{2}{5}\pi^{-4} \int_0^{2\pi}d\vartheta\int_0^\pi d\zeta \left[\frac{\pi^2}{2} e^{i\vartheta}( 3\sin\zeta\cos^3\zeta + \sin\zeta\cos\zeta)\right] = 0
\end{eqnarray}
rather than the normal expectation value of $I/2$. This is an example of soft back-propagation of the constraints of the selection affecting normally flat distributions of randomly generated initial conditions. The incoming states are biased against creating the paradox. It is a fairly generic feature that the effects of time machines are felt outside of the respective light cones.\\
We can evaluate the circuit in the classical limit as
\begin{eqnarray}
Z_{cl} = k|\langle 0,x | U_t | 1,\psi_1\rangle|^2 + k|\langle 1,x | U_t | 0,\psi_1\rangle|^2+ (1-k)|\langle 0,x | U_t | 0,\psi_1\rangle|^2 + (1-k)|\langle 1,x | U_t | 1,\psi_1\rangle|^2 \nonumber\\
= 2k(1-\alpha^2) + 2(1-k)\alpha^2
\end{eqnarray}
Due to symmetry the internal classical bit is mixed.
\begin{equation}
\rho_{cl} = \frac{1}{2}I
\end{equation}
The expectation value of the control qubit in the classical time machine case is
\begin{eqnarray}
Z_\zeta = \int_0^{2\pi} \int_0^\pi ( 2k\sin^2\zeta+ 2(1-k)\cos^2 \zeta ) d\vartheta d\zeta = 4\pi^2\\
\langle 0 | \rho_\zeta |0\rangle = \pi^{-1} \int_0^\pi d\zeta (2k\sin^2\zeta\cos^2\zeta + 2(1-k)\cos^4\zeta ) = (3-2k)/8\\
\langle 1 | \rho_\zeta |1\rangle = \pi^{-1} \int_0^\pi d\zeta (2k\sin^4\zeta + 2(1-k)\sin^2\zeta\cos^2\zeta ) = (1+2k)/8\\
\langle 0 | \rho_\zeta | 1 \rangle= 0
\end{eqnarray}
With $k$ ranging from 0 to 1. This is considering only the circulating bit to be classical. If we take the classical limit on both the control and the circulating bit we get,
\begin{eqnarray}
Z=2\\
\langle 0 | \rho_\zeta | 0\rangle = \frac{1}{2}(2-k)\\
\langle 1 | \rho_\zeta | 1 \rangle = \frac{k}{2}
\end{eqnarray}
The classical bit is forced up to the level of noise to avoid activating the $C_{not}$ gate. This is useful expression can be applied to determine approximate behavior of classical time machine circuits.
Finally the standard depolarizing channel should be analyzed under the action of each singular circuit. Recall the definition of the four orthogonal projections for the Bell state prescription.
\begin{eqnarray}
|\bar{\psi}_B\rangle = \langle 00 + 11 ,x| U_t | \phi_{Bell}, \psi_{ex} \rangle\\
|\bar{\psi}_-\rangle = \langle 00 - 11 ,x| U_t | \phi_{Bell}, \psi_{ex} \rangle\\
|\bar{\psi}_N\rangle = \langle 01 + 10 ,x| U_t | \phi_{Bell}, \psi_{ex} \rangle\\
|\bar{\psi}_{N-}\rangle \langle 01 - 10 ,x| U_t | \phi_{Bell}, \psi_{ex} \rangle
\end{eqnarray}
The partition function of the noisy Bell channel is,
\begin{equation}
Z_{\lambda} = (1-\lambda)\langle \bar{\psi}_B | \bar{\psi}_B\rangle + \lambda/4
\end{equation}
The relevent out states are all the same product of $\phi_{bell}$ and the control qubit $\psi_1$. The in states are given by the action of each circuit
\begin{eqnarray}
|\psi_{cpf-in}\rangle =\frac{\alpha}{\sqrt{2}} (|00\rangle + |11\rangle)\otimes |0\rangle + \frac{\beta}{\sqrt{2}} ( |00\rangle - |11\rangle ) \otimes |1\rangle \\
|\psi_{crot-in}\rangle =\frac{\alpha}{\sqrt{2}} (|00\rangle + |11\rangle)\otimes |0\rangle + \frac{\beta}{\sqrt{2}} ( |01\rangle - |10\rangle ) \otimes |1\rangle \\
|\psi_{cnot-in}\rangle =\frac{\alpha}{\sqrt{2}} (|00\rangle + |11\rangle)\otimes |0\rangle + \frac{\beta}{\sqrt{2}} ( |01\rangle + |10\rangle ) \otimes |1\rangle
\end{eqnarray}
In each of these cases we have
\begin{equation}
|\bar{\psi}_B\rangle = \alpha|0\rangle
\end{equation}
For each of the others, one other projection takes on a nonzero value. In the controlled phase flip circuit,
\begin{equation}
|\bar{\psi}_-\rangle =\beta|1\rangle
\end{equation}
In the controlled not circuit,
\begin{equation}
|\bar{\psi}_N\rangle = \beta|1\rangle
\end{equation}
In the controlled rotation circuit
\begin{equation}
|\bar{\psi}_{N-}\rangle = \beta|1\rangle
\end{equation}
This gives rise to three identical density matrices and partition functions.
\begin{eqnarray}
Z_{cpf}=Z_{cnot}=Z_{crot}= (1-\lambda)\alpha^2 + \lambda/4\\
\rho_{cpf} = \rho_{cnot} = \rho_{crot} = (1-\frac{3\lambda}{4})\frac{\alpha^2}{Z} |0\rangle\langle 0| + \frac{\lambda\beta\beta^\dagger}{4Z}|1\rangle\langle 1|
\end{eqnarray}
In the case of rotation through an arbitrary angle $\zeta$, or by an arbitrary phase $\xi$, the in states are given by
\begin{eqnarray}
|\psi_{\zeta-in}\rangle =\frac{1}{\sqrt{2}}\left(|00\rangle + |11\rangle\right)\otimes\left(\alpha|0\rangle + \beta\cos\zeta|1\rangle\right) + \sin\zeta\frac{\beta}{\sqrt{2}} (|01\rangle - |10\rangle ) \otimes |1\rangle \\
|\psi_{\xi-in}\rangle =\frac{\alpha}{\sqrt{2}} (|00\rangle + |11\rangle)\otimes |0\rangle + \frac{\beta}{\sqrt{2}} ( |00\rangle + e^{i\xi}|11\rangle ) \otimes |1\rangle
\end{eqnarray}
\begin{figure}[h!]
\caption{Controlled rotation of decoherent $\phi$.}
\centering
\includegraphics[width=0.3\textwidth]{gran4.png}
\end{figure}
Two projections are nonzero for the $\zeta$ circuit.
\begin{eqnarray}
|\bar{\psi}_{\zeta,B}\rangle = \alpha|0\rangle + \beta\cos\zeta|1\rangle\\
|\bar{\psi}_{\zeta,N-}\rangle = \beta\sin\zeta|1\rangle\\
N^2_B=\langle\bar{\psi}_{\zeta,B}|\bar{\psi}_{\zeta,B}\rangle = 1-\beta\beta^\dagger\sin^2\zeta\\
N^2_{N-}=\langle\bar{\psi}_{\zeta,N-}|\bar{\psi}_{\zeta,N-}\rangle = \beta\beta^\dagger\sin^2\zeta
\end{eqnarray}
The partition function is
\begin{equation}
Z_\zeta = (1-\lambda)N^2_B + \lambda/4 = 1 - \frac{3}{4}\lambda - (1-\lambda)\beta\beta^\dagger\sin^2\zeta
\end{equation}
and the projected density matrix is
\begin{eqnarray}
\rho_\zeta = \frac{\alpha^2}{Z}\left(1-\frac{3\lambda}{4}\right)|0\rangle\langle 0| + \frac{\beta\beta^\dagger}{Z}\left( \frac{\lambda}{4} + (1-\lambda)\cos^2\zeta\right)|1\rangle\langle 1|\nonumber\\
+\frac{\cos\zeta}{Z}\left(1 - \frac{3\lambda}{4}\right)\left( \alpha\beta |1\rangle\langle 0| + \alpha\beta^\dagger |0\rangle\langle 1| \right)
\end{eqnarray}
Two projections are also nonzero for the $\xi$ circuit
\begin{eqnarray}
|\bar{\psi}_{\xi,B}\rangle = \alpha|0\rangle + \frac{\beta}{2}(1+e^{i\xi})|1\rangle\\
|\bar{\psi}_{\xi,-}\rangle = \frac{\beta}{2}(1-e^{i\xi})|1\rangle\\
N^2_B= \langle\bar{\psi}_{\xi,B}|\bar{\psi}_{\xi,B}\rangle =1 + \frac{\beta\beta^\dagger}{2}(1 + \cos \xi)\\
N^2_{-}=\langle\bar{\psi}_{\xi,-}|\bar{\psi}_{\xi,-}\rangle = \frac{\beta\beta^\dagger}{2}(1-\cos\xi)
\end{eqnarray}
with partition function
\begin{equation}
Z_\xi = (1-\lambda)N^2_B + \lambda/4 =1 - \frac{3}{4}\lambda + (1-\lambda)\frac{\beta\beta^\dagger}{2}(1 + \cos \xi)
\end{equation}
and density matrix of
\begin{eqnarray}
\rho_\xi =\frac{\alpha^2}{Z}\left(1-\frac{3\lambda}{4}\right)|0\rangle\langle 0| +\frac{\beta\beta^\dagger}{4Z}\left( \frac{\lambda}{4} + 2(1-\lambda)(1+\cos\xi)\right)|1\rangle\langle 1|\nonumber\\
+\frac{4-3\lambda}{4Z} \left( \alpha\beta(1+e^{i\xi})|1\rangle\langle 0| + \alpha\beta^\dagger(1+e^{-i\xi})|0\rangle\langle 1| \right)
\end{eqnarray}
We will return later to the full implications of these expressions. For now let us examine how time machines interact with measurement.
\section{Unproven proofs}
The idea of the unproven proof is that a circulating bit contains represents an inaccessible component of the system at any time outside of the loop. In the standard theory of entanglement mixed states are produced by the action of a trace over hidden degrees of freedom. The hidden degrees of freedom are called the purification space of the system. In scenarios with time machines, the degrees of freedom of a system appear change. The internal states of the time machine represent a temporary increase in the dimension of the system. If they become entangled and then later disappear, a sort of information paradox is born. The purification space as such ceases to exist, so what becomes of the rest of the wave-function? Does it remain mixed or pure? In the coherent approach, the state can be made pure, leading to strong acausal effects. These effects are less pronounced with noise, but at the cost of leaving a mixed state. The final state projection model is exactly the application of this method of purification to the state left after a black hole has evaporated\cite{mald}. This case has also appeared in \cite{seth3}, included here for completeness and discussion with other treatments.
\begin{figure}[h!]
\caption{The unproven proof circuit uses $\phi$ as the control qubit.}
\centering
\includegraphics[width=0.5\textwidth]{unproov.png}
\end{figure}
The first circuit to consider here is a single external qubit acted on by a $C_{not}$ controlled by the time machine state. A simple version of a measurement. We initialize the incoming state,
\begin{equation}
|\psi_1\rangle = |0\rangle
\end{equation}
Then taking the product with the Bell pair state,
\begin{eqnarray}
|\phi_{Bell}\rangle = \frac{1}{\sqrt{2}}( |00\rangle + |11\rangle)\\
|\psi_{out}\rangle = |\phi_{Bell}\rangle \otimes |\psi_1\rangle = \frac{1}{\sqrt{2}}( |000\rangle + |110\rangle )
\end{eqnarray}
Now the measurement operation defined here as $C_{not}$ from qubit 2 to qubit 3 gives
\begin{equation}
|\psi_{in}\rangle = U_{cnot23} |\psi_{out}\rangle = \frac{1}{\sqrt{2}}(|000\rangle + |111\rangle)
\end{equation}
and the projection gives
\begin{equation}
|\bar{\psi}\rangle = \langle \phi_{Bell} | \psi_{in}\rangle = \frac{1}{2}(|0\rangle + |1\rangle)
\end{equation}
This looks like a fairly benign result. The probing qubit is simply rotated, the resulting superposition no different than the action of an ordinary beam splitter. For a more generic incoming state we have
\begin{eqnarray}
|\psi_1\rangle =\alpha |0\rangle + \beta|1\rangle \\
|\psi_{out}\rangle = \frac{\alpha}{\sqrt{2}} ( |000\rangle + |110\rangle ) + \frac{\beta}{\sqrt{2}}( | 001\rangle + |111\rangle ) \\
|\psi_{in}\rangle = U_{cnot23} |\psi_{out}\rangle = \frac{\alpha}{\sqrt{2}} ( |000\rangle + |111\rangle ) + \frac{\beta}{\sqrt{2}}( | 001\rangle + |110\rangle )\\
|\bar{\psi}\rangle = \langle \phi_{Bell} | \psi_{in}\rangle = \frac{\alpha +\beta}2(|0\rangle + |1\rangle)
\end{eqnarray}
This is quite a bit more shocking. The state final state, while pure, is extremely non-unitary. The relative amplitude and phase of the original state of the probe are absorbed into the normalization constant, giving an out state independent of the input. The circuit in this method 'forgets', making it an 'amnesia' type circuit. This many to one behavior is another indication that time machines or other post-selection type phenomena must be closely related to entropic phenomena if the second law of thermodynamics is to hold. Just like the grandfather paradox, an effective prior probability that back-propagates onto the initial state of $\psi_1$. The relative acceptance rate of the post-selection is
\begin{equation}
N^2 = \frac{1}2(1 + \alpha\beta + \alpha\beta^\dagger)
\end{equation}
This is basically the probability that a measurement of the out qubit along the rotated axis of $|0\rangle +|1\rangle$ will find the qubit in that state. The system appears to be equivalent to using post-selection directly on the channel $\psi_1$, fixing its value as that superposition. Indeed there is a certain similarity between time machines and classical measurement apparatus, due to the familiar 'collapse postulate'. We will return to this and related scenarios in the amnesia circuit section.\\
In the noisy Bell model, one other orthogonal projection is not identically zero.
\begin{eqnarray}
|\bar{\psi}_B\rangle = \langle 00 + 11 ,x| \psi_{in} \rangle = \frac12(\alpha +\beta)(|0\rangle + |1\rangle) \\
|\bar{\psi}_-\rangle = \langle 00 - 11 ,x| U_{cnot23} | \phi_{Bell}, \psi_{ex} \rangle=\frac12 (\alpha -\beta)(|0\rangle - |1\rangle)
\end{eqnarray}
Allowing the mixture to contain some portion of this second projection results in decoherence and is enough to avoid the measurement catastrophe.
The partition function is
\begin{equation}
Z_\lambda = (1-\lambda)\langle \bar{\psi}_B | \bar{\psi}_B\rangle + \lambda/4 = \frac12 - \frac\lambda4 + \frac{1-\lambda}2(\alpha\beta + \alpha\beta^\dagger)
\end{equation}
This gives a density matrix for the probe channel of
\begin{eqnarray}
\rho_{00}= \rho_{11} = \frac1{4Z}\left( (1-\lambda)|\alpha+\beta|^2 + \frac\lambda2 \right) = \frac12\\
\rho_{10}=\rho_{01}^\dagger = \frac1{4Z}\left( (1-\lambda)|\alpha+\beta|^2 + \frac\lambda2 \alpha(\beta+\beta^\dagger)\right) = \frac12 +\frac\lambda{4Z}(N_B^2-1)
\end{eqnarray}
The correction to the off diagonal term indicates some distortion towards a preferred superposition state, but some decoherence as well. For the classical limit of the unproven proof circuit, the states are
\begin{eqnarray}
|0,\psi_1\rangle = |0\rangle \otimes |\psi_1\rangle = \alpha|00\rangle + \beta|01\rangle\\
|1,\psi_1\rangle = |1\rangle \otimes |\psi_1\rangle = \alpha|10\rangle + \beta|11\rangle\\
U_m|0,\psi_1\rangle = |0,\psi_1\rangle\\
U_m|1,\psi_1\rangle = \alpha|11\rangle + \beta|10\rangle
\end{eqnarray}
The two weights for the time machine's classical internal states are
\begin{eqnarray}
\omega_0 = k| \langle 1,x|U_m|0,\psi_1\rangle |^2 + (1-k)| \langle 0,x|U_m|0,\psi_1\rangle |^2 = 1 - k\\
\omega_1 = k| \langle 0,x|U_m|1,\psi_1\rangle |^2 + (1-k)| \langle 1,x|U_m|1,\psi_1\rangle |^2 = 1 - k
\end{eqnarray}
and the partition function is just
\begin{equation}
Z = \omega_0 + \omega_1 = 2-2k
\end{equation}
giving a density matrix for the probe channel of
\begin{equation}
\rho_{final} = \frac{1-k}{Z}\left( |\psi_1\rangle\langle \psi_1^\dagger| + U_{not}|\psi_1\rangle\langle\psi_1^\dagger|U_{not}^\dagger \right) = \frac{1}{2}\left( \begin{array}{cc} 1 & \alpha\beta \\ \alpha\beta^\dagger & 1 \end{array} \right)
\end{equation}
The probe bit is partially randomized by the briefly accessible random classical bit from the loop, but may retain coherence based on how close it is to the gate invariant state $|0\rangle+|1\rangle$. In this case the classical limit is better behaved than the Bell state model, since the wave function amplitude never vanishes, and the partition function is independent of the incoming state. Another alternative implementation of measurement is the conditional rotation gate controlled by the circulating qubit.
\begin{equation}
U_{crot} = |00\rangle\langle 00| + | 01\rangle\langle 01| + |11\rangle\langle 10| - |10\rangle\langle 11|
\end{equation}
And the conditional phase flip which is symmetric between the two channels.
\begin{equation}
U_{cpf} = |00\rangle\langle 00| + |01\rangle\langle 01| + |10\rangle\langle 10| - |11\rangle\langle 11|
\end{equation}
Taking the product with the Bell state
\begin{equation}
|\psi_{out}\rangle = |\psi_{TM}\rangle\otimes |\psi_1\rangle = \frac{1}{\sqrt{2}}\left( \alpha|000\rangle + \alpha|110\rangle + \beta|001\rangle + \beta |111\rangle\right)
\end{equation}
The respective in states are
\begin{eqnarray}
|\psi_{crot-in}\rangle = U_{crot} |\psi_{out}\rangle = \frac{1}{\sqrt{2}}\left( \alpha |000\rangle +\alpha|111\rangle + \beta |001\rangle - \beta |110\rangle\right)\\
|\psi_{cpf-in}\rangle = U_{cpf} |\psi_{out}\rangle = \frac{1}{\sqrt{2}}\left( \alpha|000\rangle + \alpha|110\rangle + \beta|001\rangle - \beta |111\rangle \right)
\end{eqnarray}
An these project to the respective final probe states
\begin{eqnarray}
|\bar{\psi}_{crot}\rangle = \langle \phi_{TM} ,x|\psi_{crot-in}\rangle = \frac{\alpha -\beta}{2} |0\rangle + \frac{\alpha +\beta}{2}|1\rangle \\
|\bar{\psi}_{cpf}\rangle = \langle \phi_{TM},x|\psi_{cpf-in}\rangle = \alpha |0\rangle
\end{eqnarray}
For the controlled flip we still have an amnesia type circuit and a measurement catastrophe similar to the grandfather paradox when $\alpha$ vanishes. Without the ability to trace over $\phi$, destructive interference causes the wavefunction to vanish, and its renormalized direction to be indeterminate.
The normalization factors are
\begin{eqnarray}
N^2_{crot} = 1/2\\
N_{cpf} = \alpha
\end{eqnarray}
To study these circuits in the partially decoherent $\delta-$weight model , we begin with the standard 2-qubit product state
\begin{equation}
|\psi_{out}\rangle = \alpha\cos\theta |00\rangle + \beta\cos\theta|01\rangle + \alpha e^{i\xi}\sin\theta|10\rangle + \beta e^{i\xi}\sin\theta|11\rangle
\end{equation}
The action of $U_{cpf}$ is symmetric between the channels, and the analysis of it has already been covered a previous section. The action of $U_{crot}$ controlled by the circulating qubit, gives the state
\begin{equation}
|\psi_{in}\rangle = \alpha\cos\theta |00\rangle + \beta\cos\theta|01\rangle - \beta e^{i\xi}\sin\theta|10\rangle + \alpha e^{i\xi}\sin\theta|11\rangle
\end{equation}
Then projecting against the initial time machine qubit,
\begin{equation}
|\bar{\psi}\rangle = \langle \phi_{TM} ,x | \psi_{in}\rangle = (\alpha \cos^2\theta - \beta\sin^2\theta) |0\rangle + (\beta\cos^2\theta +\alpha\sin^2\theta)|1\rangle
\end{equation}
This gives us a weight of
\begin{equation}
\omega(\theta,\xi) =\langle \bar{\psi} | \bar{\psi}\rangle = \sin^4\theta + \cos^4\theta
\end{equation}
and partition function of
\begin{equation}
Z= 3\pi^2/2
\end{equation}
The density matrix for $\phi$ is
\begin{equation}
\rho_{TM} = Z^{-1}\int_0^\pi\int_0^{2\pi} \omega(\theta,\xi) \left( \begin{array}{cc} \cos^2\theta & e^{i\xi}\sin\theta\cos\theta \\ e^{-i\xi}\sin\theta\cos\theta & \sin^2\theta \end{array} \right) d\theta d\xi = \frac{1}{2}I
\end{equation}
and for the probe state,
\begin{eqnarray}
\rho_{cpf} = Z^{-1}\int_0^\pi\int_0^{2\pi} | \bar{\psi} \rangle\langle \bar{\psi}| d\theta d\xi \\
\rho_{00} = \frac{1}{2} - \frac{\alpha}{6}(\beta^\dagger + \beta)\\
\rho_{11} = \frac{1}{2} + \frac{\alpha}{6}(\beta^\dagger + \beta)\\
\rho_{01} = \frac{\alpha}{2}(\beta^\dagger -\beta) +\frac{1}{6}(\alpha^2 - \beta\beta^\dagger)
\end{eqnarray}
The classical limit weight function on the other hand is the same as before since the gate leaves the periodic channel unchanged.
\begin{eqnarray}
\omega_0 = k| \langle 1,x|U_{crot}|0,\psi_1\rangle |^2 + (1-k)| \langle 0,x|U_{crot}|0,\psi_1\rangle |^2 = 1 - k\\
\omega_1 = k| \langle 0,x|U_{crot}|1,\psi_1\rangle |^2 + (1-k)| \langle 1,x|U_{crot}|1,\psi_1\rangle |^2 = 1 - k\\
Z = 2-2k
\end{eqnarray}
The density matrix is
\begin{equation}
\rho_{TM} = Z^{-1} ( \omega_0 |0\rangle\langle 0| + \omega_1 |1\rangle\langle 1| ) = \frac{1}{2}I
\end{equation}
for the time machine and
\begin{equation}
\rho_p = Z^{-1} \left( \omega_0 |\psi_1\rangle\langle \psi_1 | + \omega_1 U_{crot}|\psi_1\rangle\langle \psi_1|U_{crot}^\dagger \right) = \frac{1}{2}\left( \begin{array}{cc} 1 & \alpha(\beta^\dagger-\beta) \\ \alpha(\beta - \beta^\dagger) & 1 \end{array} \right)
\end{equation}
for the probe channel.
\section{Twice watched pot}
In the modern understanding of decoherence, it can occur from interaction with a hidden subsystem of environment. In isolated systems, the decoherence of a particular qubit is simply the extent to which the bit is entangled with a large number of other observables. The classical states can be measured by testing any one of a large number of degrees of freedom.
In order to further explore the measurement catastrophe and grandfather paradoxes, multiple probe qubits may be used to decohere the system. Consider the unproven proof circuit using $C_{not}$ from the periodic qubit $\psi_{TM}$ onto a probe qubit. Let us add a second probe channel controlled either by $\phi$ or by the first probe's value after it measures $\phi$. \begin{figure}[h!]
\caption{Two measurements of $\phi$.}
\centering
\includegraphics[width=0.4\textwidth]{twpp.png}
\end{figure}
The standard out state in the Bell model gives
\begin{equation}
|\psi_{out}\rangle = |\phi_{TM}\rangle \otimes |\psi_1\rangle \otimes |\psi_2\rangle = \frac{1}{\sqrt{2}}( |0000\rangle + |1100\rangle)
\end{equation}
The action of either circuit will give the same in state
\begin{equation}
|\psi_{in}\rangle = \frac{1}{\sqrt{2}}( |0000\rangle + |1111\rangle)
\end{equation}
which projects down to an ordinary Bell state , at least for inputs of $|0\rangle$ for the probe channels.
\begin{equation}
|\bar{\psi}\rangle = \langle \phi_{TM},xx|\psi_{in}\rangle = \frac{1}{2}( |00\rangle + |11\rangle)
\end{equation}
Tracing over the second probe channel gives a full mixed state. Considering an arbitrary incoming product state for the probe qubits,
\begin{equation}
|\psi_{in}\rangle = |\phi_{TM}\rangle \otimes |\psi_1\rangle \otimes |\psi_2\rangle = \frac{1}{\sqrt{2}}(|00\rangle + |11\rangle) \otimes (\alpha_1|0\rangle + \beta_1|1\rangle) \otimes (\alpha_2|0\rangle + \beta_2|1\rangle)
\end{equation}
Acting with $C_{not}$ from $\phi$ onto both probes gives the in state
\begin{equation}
|\psi_{in}\rangle = \frac{1}{\sqrt{2}} \left[ |00\rangle \otimes (\alpha_1|0\rangle + \beta_1|1\rangle) \otimes (\alpha_2|0\rangle + \beta_2|1\rangle) + |11\rangle \otimes (\alpha_1|1\rangle + \beta_1|0\rangle) \otimes (\alpha_2|1\rangle + \beta_2|0\rangle) \right]
\end{equation}
which projects down to
\begin{equation}
|\bar{\psi}_{Bell}\rangle = \langle \phi_{TM},xx|\psi_{in}\rangle = \frac12(\alpha_1\alpha_2 + \beta_1\beta_2)(|00\rangle + |11\rangle) + \frac12 (\alpha_1\beta_2 + \alpha_2\beta_1)(|10\rangle + |01\rangle)
\end{equation}
This gives us a normalization factor of
\begin{equation}
N^2 = \langle \bar{\psi}_{Bell}|\bar{\psi}_{Bell}\rangle = 1 + \alpha_1\alpha_2(\beta_1 + \beta_1^\dagger)(\beta_2 + \beta_2^\dagger)
\end{equation}
This gives us two singular joint states, so the measurement catastrophe remains, however the final joint state is no longer independent of the initial probe values. The result should be quite surprising given the single probe unproven proof circuit result. We can attempt to apply the projection at different times, but the result is the same. The information loss creeps into all channels entangled with an erased channel.
\begin{figure}[h!]
\caption{Projection affects channels that do not directly interact with $\phi$.}
\centering
\includegraphics[width=0.3\textwidth]{twpp2.png}
\end{figure}
If we separate the two circuits, considering the first measurement of $\phi$ by $\psi_1$ as a projective circuit, and then the later measurement of $\psi_1$ by $\psi_2$ as a normal unitary quantum gate, we would expect a joint state before the second measurement but after the time loop of
\begin{equation}
|\bar{\psi}\rangle = \langle \phi_{TM},x | \psi_{in}\rangle = \frac{1}{\sqrt{2}} (|0\rangle + |1\rangle) \otimes |\psi_2\rangle
\end{equation}
The second channel then interacts after projection to give
\begin{eqnarray}
U_{cn12}|\bar{\psi}\rangle = \frac{1}{\sqrt{2}}( \alpha_2|00\rangle + \beta_2|01\rangle + \beta_2|10\rangle + \alpha_2|11\rangle)\nonumber\\
=\frac{\alpha_2}{\sqrt{2}}(|00\rangle +|11\rangle) + \frac{\beta_2}{\sqrt{2}}(|01\rangle +|10\rangle)
\end{eqnarray}
Carrying out the $U_{cn12}$ gate before projection gives the state as
\begin{eqnarray}
|\psi_{in}\rangle = \alpha_1\alpha_2|0000\rangle + \alpha_1\beta_2|0001\rangle + \beta_1\beta_2|0010\rangle + \beta_1\alpha_2|0011\rangle\nonumber\\
+\beta_1\alpha_2|1100\rangle+\beta_1\beta_2|1101\rangle + \alpha_1\beta_2|1101\rangle + \alpha_1\alpha_2|1111\rangle
\end{eqnarray}
Which then projects down to the previous result.
Returning to the simple double probe circuit,we wil consider the noisy case. The nonzero projections of the probe product states are
\begin{eqnarray}
|\bar{\psi}_B\rangle = \frac12(\alpha_1\alpha_2 + \beta_1\beta_2)(|00\rangle + |11\rangle) + \frac12 (\alpha_1\beta_2 + \alpha_2\beta_1)(|10\rangle + |01\rangle)\\
|\bar{\psi}_-\rangle = \frac12(\alpha_1\alpha_2 - \beta_1\beta_2)(|00\rangle - |11\rangle) + \frac12 (\alpha_1\beta_2 - \alpha_2\beta_1)(|01\rangle - |10\rangle)
\end{eqnarray}
The partition function is
\begin{equation}
Z_\lambda = (1-\lambda)N_B^2 + \lambda/4 = \frac12 - \frac14\lambda + \frac12(1-\lambda)\alpha_1\alpha_2(\beta_1+\beta_1^\dagger)(\beta_2+\beta_2^\dagger)
\end{equation}
and the density matrix is
\begin{eqnarray}
\rho_\lambda = Z_\lambda^{-1}\left(1-\frac{3\lambda}4\right)|\bar{\psi}_B\rangle\langle \bar{\psi}_B| + \frac\lambda{4Z_\lambda}|\bar{\psi}_-\rangle\langle\bar{\psi}_-|\nonumber\\
\end{eqnarray}
A few simple cases should help to see the physical picture here. For an initial probe value of $|00\rangle$ we have
\begin{eqnarray}
\alpha_1=\alpha_2=1\\
Z =\frac12 - \frac\lambda4\\
\langle 00 | \rho_{twp} | 00\rangle = \langle 11 | \rho_{twp} | 11\rangle = 1/2
\end{eqnarray}
with all other $\rho_{ij}$ vanishing. $Z$ has a local minimum at the two points where the projection against the Bell state vanishes. Instead of the measurement catastrophe, we have a factorable density matrix and pure out state
\begin{eqnarray}
\alpha_1=\alpha_2=\beta_1 = -\beta_2 = 1/\sqrt{2}\\
Z =\lambda/4\\
|\xi_{min}\rangle =|\bar{\psi}_-\rangle = \frac{1}{2} \left( |00\rangle - |01\rangle + |10\rangle - |11\rangle \right)\\
\rho_{min} = |\xi_{min}\rangle\langle \xi_{min}|
\end{eqnarray}
An orthogonal pure state gives the maximum value for $Z$
\begin{eqnarray}
\alpha_1=\alpha_2=\beta_1 = \beta_2 = 1/\sqrt{2}\\
Z = 1 - \frac34\lambda \\
|\xi_{max}\rangle =|\bar{\psi}_B\rangle= \frac{1}{2} \left( |00\rangle + |01\rangle + |10\rangle + |11\rangle \right)\\
\rho_{max} = |\xi_{max}\rangle\langle \xi_{max}|
\end{eqnarray}
To get the classical limit for the double measurement circuit we take the classical weights defined as the amplitude squared of the projection against each classical basis state.
\begin{eqnarray}
|\psi_{out}(i)\rangle = |e_i\rangle \otimes |\psi_1\rangle \otimes |\psi_2\rangle\\
|\bar{\psi}_{ij}\rangle = \langle e_i ,xx | U_c | \psi_{out}(j) \rangle\\
|\bar{\psi}_{00}\rangle = |\psi_1\rangle \otimes |\psi_2\rangle\\
|\bar{\psi}_{11}\rangle = U_{not}|\psi_1\rangle \otimes U_{not} | \psi_2\rangle \\
|\bar{\psi}_{01}\rangle = |\bar{\psi}_{10}\rangle = 0\\
\omega_0 = k\langle \bar{\psi}_{10} | \bar{\psi}_{10} \rangle + (1-k)\langle \bar{\psi}_{00} | \bar{\psi}_{00} \rangle = 1-k \\
\omega_1 = k\langle \bar{\psi}_{01} | \bar{\psi}_{01} \rangle + (1-k)\langle \bar{\psi}_{11} | \bar{\psi}_{11} \rangle = 1-k \\
Z = \omega_0 + \omega_1 = 2-2k
\end{eqnarray}
This leads to the density matrix of the outgoing states
\begin{equation}
\rho_{cl} = \frac{1}{2}|\bar{\psi}_0\rangle\langle\bar{\psi}_0| + \frac{1}{2}|\bar{\psi}_1\rangle\langle\bar{\psi}_1|
\end{equation}
If the incoming channels are initialized to $|00\rangle$ then this produces
\begin{equation}
\rho = \frac{1}{2}( |00\rangle\langle 00| + |11\rangle\langle 11|)
\end{equation}
Another interesting case is to consider the same circuit, but with the probe qubits initially in an entangled rather than product state.
\begin{equation}
|\psi_{out}\rangle = |\phi_{Bell}\rangle \otimes \left( \gamma_{00}|00\rangle + \gamma_{01}|01\rangle + \gamma_{10}|10\rangle + \gamma_{11}|11\rangle \right)
\end{equation}
which projects to
\begin{eqnarray}
\langle \phi_{Bell},xx| \psi_{in}\rangle = \langle \phi_{Bell},xx| \psi_{out}\rangle + \langle \phi_{Bell},xx| U_{not(1)} U_{not(2)} | \psi_{out}\rangle\\
|\bar{\psi}_B\rangle=\frac{1}{2}(\gamma_{00}+\gamma_{11})\left( |00\rangle +|11\rangle \right) +\frac{1}{2}(\gamma_{01}+\gamma_{10})\left( |01\rangle +|10\rangle\right)
\end{eqnarray}
which has a measurement catastrophe over the set of entangled initial probe states given by
\begin{equation}
|\psi_{null}\rangle =\cos\vartheta(|00\rangle - |11\rangle) + e^{i\xi}\sin\vartheta( |01\rangle - |10\rangle )
\end{equation}
The dimension of the null space for the incoming measurement qubits remains at half, or one qubit less than, the whole space, as the number of measurement bits increases.
When adding noise there is again only one other nonzero projection.
\begin{equation}
|\bar{\psi}_-\rangle = \frac12(\gamma_{00}-\gamma_{11})(|00\rangle - |11\rangle) + \frac12(\gamma_{01} - \gamma_{10})(|01\rangle - |10\rangle)
\end{equation}
The partition function for the generic probe state is
\begin{equation}
Z_\lambda = (1-\lambda)N_B^2 + \lambda/4 = \frac12 - \frac\lambda4 + (1-\lambda)\left(\gamma_{00}\gamma_{11}^\dagger + \gamma_{11}\gamma_{00}^\dagger + \gamma_{01}\gamma_{10}^\dagger + \gamma_{10}\gamma_{01}^\dagger\right)
\end{equation}
This is comparatively well behaved even for entangled in states, due to the effective noise contributed from the spread of the weight function. In the classical limit case the weights are unchanged by the addition of entanglement in the probe in states.\\
A second important point about decoherence is that it is not usually periodic. Entanglement between the circulating bits and the environment cannot generally be identical for in and out states. Consequently, if we have multiple CTC qubits, they cannot mutually decohere without the normalization parameter vanishing. Measurement circuits that entangle multiple time machine channels act similarly to grandfather circuits. For example a single $C_{not}$ connecting two periodic qubits immediately reduces the allowed states by half, as it fixes the value of the control qubit to avoid generating a paradox on the subject bit. The same qubit acting as a control for multiple periodic channels can effectively over-determine its state, creating a grandfather type paradox for all values of the incoming probe state. To analyze these circuits we will need two Bell states, $\phi_1$ and $\phi_2$ one for each periodic channel, and an input state $\psi_1$. First let us examine the interaction of two periodic channels via $C_{not}$.
\begin{figure}[h!]
\caption{Interaction between projected channels decreases the normalization factor.}
\centering
\includegraphics[width=0.4\textwidth]{2phireduc.png}
\end{figure}
Beginning with the out state in the Bell model
\begin{equation}
|\phi_{TM}\rangle = |\phi_1\rangle \otimes |\phi_2\rangle = \frac{1}{2}\left( |0000\rangle + |0011\rangle + |1100\rangle + |1111\rangle \right)
\end{equation}
We connect the two channels with a $C_{not}$ from bit 2 to bit 4. Bit 1 and bit 3 are the reference bits for 2 and 4 respectively.
\begin{equation}
|\phi_{in}\rangle = U_{cn24}|\phi_{TM}\rangle = \frac{1}{2}\left( |0000\rangle + |0011\rangle + |1101\rangle + |1110\rangle \right)
\end{equation}
If we add a probe as bit 5, which measures bit 2 via another $C_{not}$ gate $U_{cn25}$, then the projected final state of the probe is unchanged. This indicates that channel $\phi_1$ consisting of bits 1 and 2 is fixed as the state $|00\rangle$.
\begin{eqnarray}
|\psi_{out}\rangle = |\phi_{TM}\rangle \otimes |\psi_1\rangle\\
|\psi_{in}\rangle = U_{cn25}U_{cn24}|\psi_{out}\rangle\nonumber\\
= \frac{1}{2}\left( |0000\rangle + |0011\rangle\right) \otimes |\psi_1\rangle + \frac{1}{2}\left( |1101\rangle + |1110\rangle \right) \otimes U_{not}|\psi_1\rangle\\
|\bar{\psi}_B \rangle = \langle \phi_2 , x | \langle \phi_1 ,xxx | \psi_{in}'\rangle = \frac{1}{4}|\psi_1\rangle
\end{eqnarray}
Adding noise $\lambda$ requires us to use more projection states to cover all the degrees of freedom of the 4 qubit channel.
There are potentially 16 independent projections given by the products of pairs of the four single channel orthogonal vectors.
In modeling the noise we have the freedom to correlate the fluctuations between channels $\phi_1$ and $\phi_2$ or not.
Four of the projections are nonzero.
\begin{eqnarray}
|\bar{\psi}_{BB}\rangle = \langle \phi_{BB} | \psi_{in} \rangle = \frac12|\psi_1\rangle\\
|\bar{\psi}_{-B}\rangle = \langle \phi_{-B} | \psi_{in} \rangle = \frac12|\psi_1\rangle\\
|\bar{\psi}_{BN}\rangle = \langle \phi_{BN} | \psi_{in} \rangle = \frac12U_{not}|\psi_1\rangle\\
|\bar{\psi}_{-N}\rangle = \langle \phi_{-N} | \psi_{in} \rangle = - \frac12U_{not}|\psi_1\rangle
\end{eqnarray}
Weighting each projection by the number of independent errors, and assuming the error rate is $\lambda$ for each qubit, we get three terms.
\begin{eqnarray}
Z = (1-\frac34\lambda)^2\langle\bar{\psi}_{BB} |\bar{\psi}_{BB}\rangle + \frac\lambda4(1-\frac34\lambda)\left( \langle\bar{\psi}_{-B} |\bar{\psi}_{-B}\rangle+ \langle\bar{\psi}_{BN} |\bar{\psi}_{BN}\rangle \right) + \frac{\lambda^2}{16} \langle\bar{\psi}_{-N} |\bar{\psi}_{-N}\rangle\nonumber\\
=\frac14(1-\frac\lambda2)^2
\end{eqnarray}
This gives a density matrix for the outgoing probe of
\begin{eqnarray}
\rho = Z^{-1}\left((1-\frac34\lambda)^2|\bar{\psi}_{BB} \rangle\langle\bar{\psi}_{BB}| + \frac\lambda4(1-\frac34\lambda)\left( |\bar{\psi}_{-B} \rangle\langle\bar{\psi}_{-B}|+ |\bar{\psi}_{BN} \rangle\langle\bar{\psi}_{BN}| \right) + \frac{\lambda^2}{16}| \bar{\psi}_{-N} \rangle\langle\bar{\psi}_{-N}|\right)\nonumber\\
=\frac{4-3\lambda}{4-2\lambda}|\psi_1\rangle\langle\psi_1| + \frac\lambda{4-2\lambda}U_{not}|\psi_1\rangle\langle\psi_1|U_{not}^\dagger
\end{eqnarray}
Evaluating the classical limit the weights are simpler.
\begin{eqnarray}
\omega_{00} =\omega_{01} = k/4 - (1-k)| \langle 00,x | U_{cn12}U_{cn13} | 00,\psi_1\rangle |^2 = 1 - 3k/4\\
\omega_{10} = \omega_{11} = k/4 - (1-k)| \langle 10,x | U_{cn12}U_{cn13} | 10,\psi_1\rangle |^2 = k/4\\
Z = 2 - k
\end{eqnarray}
and measurement output of
\begin{equation}
\rho_{cl} = (1-\frac{k}{2Z})|\psi_1\rangle\langle \psi_1| + \frac{k}{2Z}U_{not}|\psi_1\rangle\langle \psi_1|U_{not}^\dagger
\end{equation}
\section{Mutual paradoxes}
To further explore the concept of over-determination of entanglement, we examine a variation of the grandfather paradox, pitting one periodic qubit against another.
Two periodic channels $\phi_1$ and $\phi_2$ are both acted on by a controlled not from a single external channel $\psi_1$. in between these interactions a not gate acts on the $\psi_1$ channel.
In the bell state model the out state is
\begin{equation}
|\psi_{out}\rangle = \frac{1}{2}\left( |00\rangle + |11\rangle \right) \otimes \left( |00\rangle + |11\rangle \right) \otimes \left( \alpha|0\rangle + \beta|1\rangle\right)
\end{equation}
The action of three cages determines the in state to be
\begin{eqnarray}
|\psi_{in} \rangle= U_{cn54}U_{not5}U_{cn52}|\psi_{out}\rangle = \alpha\left( |00011\rangle + |00101\rangle + |11011\rangle + |11101\rangle\right) +\nonumber\\
\beta\left( |01000\rangle + |01110\rangle + |10000\rangle + |10110\rangle\right)
\end{eqnarray}
This state is always orthogonal to the original product state for any value of the input state $\psi_1$.
If the not gate is replaced by a rotation of angle $\zeta$ then the paradox is averted and the normalization factor behaves the same
as it does in the faulty gun circuit.
\begin{eqnarray}
|\psi_{in}'(\zeta)\rangle = U_{cn54}U_{rot5}(\zeta)U_{cn52}|\psi_{out}\rangle = \cos\zeta U_{cn54}U_{cn52}|\psi_{out}\rangle - \sin\zeta |\psi_{in}\rangle\\
\langle \phi_1,x | \langle \phi_2,xxx | \psi_{in}'(\zeta) \rangle = \cos\zeta \langle \phi_1,x | \langle \phi_2,xxx | U_{cn54}U_{cn52}|\psi_{out}\rangle =\alpha\cos\zeta |0\rangle\\
N_B = \alpha\cos\zeta
\end{eqnarray}
The first controlled not gate removes the $|1\rangle$ part of $\psi_1$, and the second does the same for the superposition created by the rotation gate.
\begin{figure}[h!]
\caption{Mutual paradox due to two inconsistent post-selections.}
\centering
\includegraphics[width=0.5\textwidth]{mutual.png}
\end{figure}
The two loops do not have to be in causal contact with each other, the paradox is a property of the global structure of the circuit.
Adding noise again requires that we calculate the other nonzero projections of $|\psi_{in}'\rangle$. These are
\begin{eqnarray}
|\bar{\psi}_{NB}\rangle = -\beta\sin\zeta |0\rangle\\
|\bar{\psi}_{BN}\rangle = -\alpha\sin\zeta |1\rangle\\
|\bar{\psi}_{NN}\rangle =\beta\cos\zeta|1\rangle
\end{eqnarray}
Assuming independent noise of $\lambda$ in each periodic channel, we again have three terms in the partition function.
\begin{equation}
Z_\lambda(\alpha) = \left(1-\frac34\lambda\right)^2\alpha^2\cos^2\zeta + \frac\lambda4\left(1-\frac34\lambda\right)\sin^2\zeta + \frac{\lambda^2\beta\beta^\dagger}{16}\cos^2\zeta
\end{equation}
The interaction with the decoherent channels mixes the initially pure state $\psi_1$ into
\begin{eqnarray}
\rho_\psi = Z_\lambda^{-1}\left((1-\frac34\lambda)^2|\bar{\psi}_{BB} \rangle\langle\bar{\psi}_{BB}| + \frac\lambda4(1-\frac34\lambda)\left( |\bar{\psi}_{NB} \rangle\langle\bar{\psi}_{NB}|+ |\bar{\psi}_{BN} \rangle\langle\bar{\psi}_{BN}| \right) + \frac{\lambda^2}{16}| \bar{\psi}_{NN} \rangle\langle\bar{\psi}_{NN}|\right)\\
\rho_{00} =\frac1{Z_\lambda} \left(1-\frac34\lambda\right)^2\alpha^2\cos^2\zeta + \frac\lambda{4Z_\lambda}\left(1-\frac34\lambda\right)\beta\beta^\dagger\sin^2\zeta \\
\rho_{01}=0
\end{eqnarray}
The circuit also has back-propagation, due to the dependence of $Z_\lambda$ on $\alpha$. The skew of an otherwise maximally mixed control channel can be calculated using
\begin{equation}
\rho_{00} = \frac{ \int \rho(\alpha) Z(\alpha) d\psi_1 }{\int Z d\psi_1}= \frac1{Z_\rho}\int_0^\pi\int_0^{2\pi} Z_\lambda(\cos\vartheta,\xi) \rho_\lambda(\vartheta,\xi) d\vartheta d\xi
\end{equation}
where
\begin{eqnarray}
\alpha = \cos\vartheta\\
\beta = e^{i\xi}\sin\vartheta
\end{eqnarray}
Evaluating the denominator gives
\begin{equation}
Z_\rho = \int Z_\lambda(\cos\theta) d\theta = \pi^2\left(1-\frac34\lambda\right)^2\cos^2\zeta + \pi^2\frac\lambda2\left(1-\frac34\lambda\right)\sin^2\zeta + \pi^2\frac{\lambda^2}{16}\cos^2\zeta
\end{equation}
So an otherwise maximally mixed test channel $\psi_1$ will be distorted by post-selection into the mixed state
\begin{equation}
\bar{\rho}_{skew} = \frac{\pi^2}{Z_\rho} \left(1-\frac34\lambda\right)^2\cos^2\zeta + \frac{\lambda\pi^2}{4Z_\rho}\left(1-\frac34\lambda\right)\sin^2\zeta \\
\end{equation}
The classical limit similar
\begin{eqnarray}
\langle 00,x| U_t | 00,\psi_1\rangle =\langle 01,x| U_t | 01,\psi_1\rangle =\langle 10,x| U_t | 10,\psi_1\rangle =\langle 11,x| U_t | 11,\psi_1\rangle = \alpha\cos\zeta\\
\langle 01,x| U_t | 00,\psi_1\rangle =\langle 00,x| U_t | 01,\psi_1\rangle =\langle 10,x| U_t | 11,\psi_1\rangle =\langle 11,x| U_t | 10,\psi_1\rangle = -\beta\sin\zeta\\
\langle 00,x| U_t | 10,\psi_1\rangle =\langle 10,x| U_t | 00,\psi_1\rangle =\langle 11,x| U_t | 01,\psi_1\rangle =\langle 01,x| U_t | 11,\psi_1\rangle = \alpha\sin\zeta\\
\langle 00,x| U_t | 11,\psi_1\rangle =\langle 11,x| U_t | 00,\psi_1\rangle =\langle 10,x| U_t | 01,\psi_1\rangle =\langle 01,x| U_t | 10,\psi_1\rangle = \beta\cos\zeta\\
Z_{cl}(\alpha) = (1-k)^2\alpha^2\cos^2\zeta + k(1-k)\sin^2\zeta + k^2\beta\beta^\dagger\cos^2\zeta
\end{eqnarray}
carrying out the same calculation for skewing the input channel $\psi_1$ gives
\begin{equation}
Z_\rho = 2\pi\int_0^\pi Z_\alpha(\cos\theta) d\theta = \pi^2 \cos^2\zeta\left[ (1-k)^2 + k^2 \right] + 2\pi^2k(1-k)\sin^2\zeta
\end{equation}
The accessible states gives a skew onto the input qubit of
\begin{eqnarray}
\rho_{cl-\psi} = \frac{1}{Z_\rho}\int_0^\pi\int_0^{2\pi} Z_\alpha(\cos\theta) \left( \begin{array}{cc} \cos^2\theta & e^{i\xi}\sin\theta\cos\theta \\ e^{-i\xi}\sin\theta\cos\theta & \sin^2\theta \end{array} \right) d\theta d\xi\nonumber\\
=\frac{\pi^2}{Z_\rho}k(1-k)\sin^2\zeta\left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right) + \frac{\pi^2}{4Z_\alpha}\cos^2\zeta \left( \begin{array}{cc} 3(1-k)^2+k^2 & 0 \\ 0 & (1-k)^2 + 3k^2 \end{array} \right)
\end{eqnarray}
The greater the noise, the more unbiased the input $\psi_1$. Input bias will be further explored in the section on back-propagation.
\section{Third party paradox}
A second important example paradox spread out over multiple gates is the exclusive or grandfather circuit,
consisting of two independent input channels and a single periodic channel which is acted on by both inputs acting as controls to two
$C_{not}$ gates. Projection over the periodic channel creates a correlation between the two input channels. The scenario is
essentially the projection of a GHZ entanglement of three channels by the 'third party', the periodic channel, to create
a Bell pair without direct interaction between the input qubits. \\
The initial state in the Bell state description is
\begin{eqnarray}
|\psi_{out}\rangle = |\phi_{Bell} \rangle \otimes |\psi_1\rangle \otimes |\psi_2\rangle\nonumber\\
= \frac{1}{\sqrt{2}}\left( \alpha_1\alpha_2|0000\rangle + \alpha_1\alpha_2|1100\rangle +\beta_1\beta_2|0011\rangle + \beta_1\beta_2|1111\rangle \right)\nonumber\\
+\frac{1}{\sqrt{2}}\left(\alpha_1\beta_2|0001\rangle + \alpha_1\beta_2|1101\rangle +\beta_1\alpha_2|0010\rangle + \beta_1\alpha_2|1110\rangle\right)
\end{eqnarray}
The circuit consists of two controlled not gates giving,
\begin{eqnarray}
|\psi_{in}\rangle = U_{cn32}U_{cn42}|\psi_{out}\rangle \nonumber\\
= \frac{1}{\sqrt{2}}\left( \alpha_1\alpha_2|0000\rangle + \alpha_1\alpha_2|1100\rangle +\beta_1\beta_2|0011\rangle + \beta_1\beta_2|1111\rangle \right)\nonumber\\
+\frac{1}{\sqrt{2}}\left(\alpha_1\beta_2|0001\rangle + \alpha_1\beta_2|1101\rangle +\beta_1\alpha_2|0010\rangle + \beta_1\alpha_2|1110\rangle\right)
\end{eqnarray}
\begin{figure}[h!]
\caption{Third party paradox produces entanglement between $\psi_1$ and $\psi_2$.}
\centering
\includegraphics[width=0.5\textwidth]{3rdparty.png}
\end{figure}
The projection of this state against $|\phi_{Bell}\rangle$ gives
\begin{equation}
|\bar{\psi}_B\rangle = \langle\phi_{bell}, xx|\psi_{in}\rangle = \alpha_1\alpha_2|00\rangle + \beta_1\beta_2|11\rangle
\end{equation}
with a normalization factor
\begin{equation}
N^2 = \langle\bar{\psi}|\bar{\psi}\rangle = \alpha_1^2\alpha_2^2 + \beta_1\beta_1^\dagger\beta_2\beta_2^\dagger
\end{equation}
The product state is projected onto an entangled state.
This is the mechanism most cited in information paradox literature for allowing information to escape and averting the firewall paradox in the black hole final state model.\cite{unit}
The singular states all correspond to orthogonal control bit inputs.
\begin{equation}
N^2 = |\langle \psi_1 | \psi_2 \rangle|^2
\end{equation}
This is another indication of the global nature of paradoxes, and of the phenomenon of back-propagation.
The noisy case adds only one other nonzero projection.
\begin{equation}
|\bar{\psi}_N\rangle = \alpha_1\beta_2|01\rangle + \alpha_2\beta_1|10\rangle
\end{equation}
The partition function and density matrix are
\begin{eqnarray}
Z_\lambda = (1-\lambda)N^2 + \lambda/4\\
\rho = \frac1{Z_\lambda}(1-\frac34\lambda)|\bar{\psi}_B\rangle\langle\bar{\psi}_B| + \frac\lambda{4Z_\lambda}|\bar{\psi}_N\rangle\langle \bar{\psi}_N|
\end{eqnarray}
The effect is similar in the classical limit. The out states are given by
\begin{eqnarray}
|\psi_{out}(0)\rangle = |0 \rangle \otimes |\psi_1\rangle \otimes |\psi_2\rangle\nonumber\\
=\alpha_1\alpha_2|000\rangle + \beta_1\beta_2|011\rangle + \alpha_1\beta_2|001\rangle +\beta_1\alpha_2|010\rangle\\
|\psi_{out}(1)\rangle = |1 \rangle \otimes |\psi_1\rangle \otimes |\psi_2\rangle\nonumber\\
=\alpha_1\alpha_2|100\rangle + \beta_1\beta_2|111\rangle + \alpha_1\beta_2|101\rangle +\beta_1\alpha_2|110\rangle
\end{eqnarray}
and in states by
\begin{eqnarray}
|\psi_{in}(0)\rangle = U_{cn31}U_{cn21}|\psi_{out}(0)\rangle \nonumber\\
=\alpha_1\alpha_2|000\rangle + \beta_1\beta_2|011\rangle + \alpha_1\beta_2|101\rangle +\beta_1\alpha_2|110\rangle\\
|\psi_{in}(1)\rangle = U_{cn31}U_{cn21}|\psi_{out}(1)\rangle \nonumber\\
=\alpha_1\alpha_2|100\rangle + \beta_1\beta_2|111\rangle + \alpha_1\beta_2|001\rangle +\beta_1\alpha_2|010\rangle
\end{eqnarray}
Then project each against the corresponding eigenstate to give
\begin{eqnarray}
|\bar{\psi}(0,0)\rangle =|\bar{\psi}(1,1)\rangle = \langle 0, xx|\psi_{in}(0)\rangle = |\bar{\psi}_B\rangle \\
|\bar{\psi}(1,0)\rangle =|\bar{\psi}(0,1)\rangle = \langle 1, xx|\psi_{in}(1)\rangle = |\bar{\psi}_N\rangle
\end{eqnarray}
The resulting weights are then
\begin{eqnarray}
\omega_0 = (1-k)\langle \bar{\psi}(0,0) | \bar{\psi}(0,0) \rangle + k\langle \bar{\psi}(1,0) | \bar{\psi}(1,0)\rangle = (1-k)N^2 + k(1-N^2)\\
= \omega_1 = (1-k) \langle \bar{\psi}(1,1) | \bar{\psi}(1,1) \rangle + k\langle \bar{\psi}(0,1) | \bar{\psi}(0,1)\rangle\\
Z_{cl} = 2\omega_0
\end{eqnarray}
And the density matrix as
\begin{equation}
\rho_{cl} = \frac{1-k}{Z_{cl}}|\bar{\psi}_B\rangle\langle\bar{\psi}_B| + \frac{k}{Z_{cl}}|\bar{\psi}_N\rangle\langle \bar{\psi}_N|
\end{equation}
If we fix one of the two controlling channels into a definite eigenstate, then its twin takes on the same state in the projected ensemble.
The channels could be said to be 'super-entangled', since changes to the doublet state will still project onto another doublet.
\section{Stubborn spin effect}
There are several circuit representations of what has been called the unproven theorem paradox. The simplest model is a single qubit circulating along a CTC, which we then measure, effectively acting on it with a projection operator, obtaining classical information. The use of $P_\pm$ for measurement here means that we have effectively absorbed the tracing over external bits step into the normal evolution.
\begin{equation}
\alpha|+\rangle + \beta|-\rangle = |\phi\rangle \rightarrow P_\pm|\phi\rangle
\end{equation}
Just after a measurement, the system is in one of the two eigenstates. To compare the relative frequency of the two measurement outcomes, $|+\rangle$, and $|-\rangle$, we consider the weights. Both are eigenvectors of $P_\pm$, and the noise operator is symmetric between them, so the weights will be the same. A single measurement will yield a random bit.\\
To consider an analog of the twice watched pot, we may imagine two successive measurements on a single circulating qubit.
Consider the probability of finding the system in the state $|+\rangle$ at one time, and $|\phi\rangle$, at some earlier time.
First projecting the state $|\phi\rangle$ onto $|+\rangle$ gives a factor of $|\alpha^2|$. Just after a measurement result of $|+\rangle$, the qubit is 'recycled' into the mixed state
\begin{equation}
\rho_i = |\alpha|^2A(|+\rangle\langle+|) = |\alpha|^2(1-k)|+\rangle\langle +| + |\alpha|^2k|-\rangle\langle -|
\end{equation}
with a weight of
\begin{equation}
\omega(\phi) = \langle\phi|\rho_i|\phi\rangle = |\alpha|^4(1-k) + |\alpha^2\beta^2|k
\end{equation}
where we have described the noise $A$ as a linear operator on density matrices with an effective bit error rate $k$. Clearly it is proportional to the ordinary transition probability for a state $\phi$ to scatter onto $|+\rangle$ and then back onto $\phi$. Here it is easy to see that integrating out $\phi$ will give the same weight for both classical measurement results. The result of the classical measurement is thus a random bit. Furthermore, we can see that the relative conditional probability of measuring the initial state $|\phi\rangle$, given a later or earlier measurement of $|+\rangle$ is the weight given here.
\begin{figure}[h!]
\caption{Transition probabilities for successive measurements and rotations of $\phi$ depend on future gates }
\centering
\includegraphics[width=0.4\textwidth]{stspin1.png}
\end{figure}
In the limit of small noise, two successive measurements of a qubit at a mutual angle $\theta$ will give the same result with probability
\begin{equation}
\tilde{p} = \frac{p^2}{p^2 + (1-p)^2} = ((\tan{\theta})^4 +1)^{-1}.
\end{equation}
Notice the quartic term, as opposed to $\tan^2\theta$ that would appear without the CTC. This is a strong deviation from normal unitary evolution.
We can model this scenario using a series of five gates acting on a single periodic channel and three probe channels.
Two rotation gates in between three $C_{not}$ measurement gates. Each measurement gate is controlled by the periodic bit and acts on a separate probe bit.
We will take each probe to be pre-selected in the $|0\rangle$ state to isolate this effect from back-propagation. This gives us a loop out state of
\begin{equation}
|\psi_{out} = |\phi_{TM}\rangle \otimes |000\rangle
\end{equation}
Acting with the first measurement we get
\begin{equation}
|\psi_{out}'\rangle = U_{cn23}|\psi_{out}\rangle = \frac1{\sqrt{2}}(|00000\rangle + |11100\rangle)
\end{equation}
Next comes rotation by $\theta_1$.
\begin{equation}
|\psi_{out}''\rangle = U_{\theta_1}|\psi_{out}'\rangle = \frac{1}{\sqrt{2}}(\cos\theta_1|00000\rangle + \cos\theta_1|11100\rangle + \sin\theta_1|01000\rangle - \sin\theta_1|10100\rangle)
\end{equation}
The second measurement maps the state to
\begin{equation}
|\psi_{out}'''\rangle = U_{cn24}|\psi_{out}''\rangle = \frac{1}{\sqrt{2}}(\cos\theta_1|00000\rangle + \cos\theta_1|11110\rangle + \sin\theta_1|01010\rangle - \sin\theta_1|10100\rangle)
\end{equation}
The second rotation by angle $\theta_2$ gives
\begin{eqnarray}
|\psi_{out}''''\rangle = U_{\theta_2}|\psi_{out}'''\rangle\nonumber\\
= \frac{1}{\sqrt{2}}(\cos\theta_1\cos\theta_2|00000\rangle + \cos\theta_1\cos\theta_2|11110\rangle + \sin\theta_1\cos\theta_2|01010\rangle - \sin\theta_1\cos\theta_2|10100\rangle)\nonumber\\
+\cos\theta_1\sin\theta_2|01000\rangle - \cos\theta_1\sin\theta_2|10110\rangle - \sin\theta_1\sin\theta_2|00010\rangle - \sin\theta_1\sin\theta_2|11100\rangle) &
\end{eqnarray}
Finally the in state is found after the third measurement
\begin{eqnarray}
|\psi_{in}\rangle = U_{cn25}|\psi_{out}''''\rangle\nonumber\\
= \frac{1}{\sqrt{2}}(\cos\theta_1\cos\theta_2|00000\rangle + \cos\theta_1\cos\theta_2|11111\rangle + \sin\theta_1\cos\theta_2|01011\rangle - \sin\theta_1\cos\theta_2|10100\rangle)\nonumber\\
+\cos\theta_1\sin\theta_2|01001\rangle - \cos\theta_1\sin\theta_2|10110\rangle - \sin\theta_1\sin\theta_2|00010\rangle - \sin\theta_1\sin\theta_2|11101\rangle) &
\end{eqnarray}
The Bell projection of this state gives
\begin{equation}
|\bar{\psi}_B\rangle = \langle \phi_{Bell} | \psi_{in} \rangle = \frac12( \cos\theta_1\cos\theta_2|000\rangle + \cos\theta_1\cos\theta_2|111\rangle - \sin\theta_1\sin\theta_2|010\rangle - \sin\theta_1\sin\theta_2|101\rangle)
\end{equation}
and normalization factor
\begin{equation}
N^2 = \frac12(\cos^2\theta_1\cos^2\theta_2 +\sin^2\theta_1\sin^2\theta_2)
\end{equation}
Two things are of interest in this result. The first and third measurements are perfectly entangled. This is not normal behavior, but not completely unsurprising
since from the looping channel's perspective measurements 1 and 3 occur back to back with no rotation in between.
The stubborn spin effect manifests when we look at the probability of the intermediate transition.
Just as in the spin case, the renormalized projected state vector gives a probability of observing alternating bits as
\begin{equation}
p_{flip} = \frac{\sin^2\theta_1\sin^2\theta_2}{N^2} = \frac1{\cot^2\theta_1\cot^2\theta_2 +1}
\end{equation}
This holds even without the third probe channel.
In the noisy case the three orthogonal projections and norms are
\begin{eqnarray}
|\phi_-\rangle = \frac1{\sqrt{2}}( |00\rangle - |11\rangle)\\
|\phi_N\rangle = \frac1{\sqrt{2}}(|01\rangle + |10\rangle)\\
|\phi_{-N} = \frac1{\sqrt{2}}(|01\rangle - |10\rangle)\\
|\bar{\psi}_-\rangle = \langle \phi_- | \psi_{in} \rangle = \frac12( \cos\theta_1\cos\theta_2|000\rangle - \cos\theta_1\cos\theta_2|111\rangle - \sin\theta_1\sin\theta_2|010\rangle + \sin\theta_1\sin\theta_2|101\rangle) \\
|\bar{\psi}_N\rangle = \langle \phi_N | \psi_{in} \rangle = \frac12( \cos\theta_1\sin\theta_2|001\rangle + \sin\theta_1\cos\theta_2|011\rangle - \sin\theta_1\cos\theta_2|100\rangle - \cos\theta_1\sin\theta_2|110\rangle) \\
|\bar{\psi}_{-N}\rangle = \langle \phi_{-N} | \psi_{in} \rangle = \frac12( \cos\theta_1\sin\theta_2|001\rangle + \sin\theta_1\cos\theta_2|011\rangle + \sin\theta_1\cos\theta_2|100\rangle + \cos\theta_1\sin\theta_2|110\rangle) \\
N_-^2 = N^2\\
N_N^2 = N^2_{-N}= \frac12(\sin^2\theta_1\cos^2\theta_2 + \cos^2\theta_1\sin^2\theta_2)
\end{eqnarray}
The partition function and density matrix are given by
\begin{eqnarray}
Z_\lambda= (1-\lambda)N^2 +\lambda/4\\
\rho = \frac{4 -3\lambda}{4Z_\lambda}|\bar{\psi}_B\rangle\langle \bar{\psi}_B| + \frac\lambda{4Z_\lambda}\left( |\bar{\psi}_-\rangle\langle \bar{\psi}_-| +|\bar{\psi}_N\rangle\langle \bar{\psi}_N| + |\bar{\psi}_{-N}\rangle\langle \bar{\psi}_{-N}| \right)
\end{eqnarray}
This results in a flip probability for the second measurement of
\begin{equation}
P_{flip-\lambda} = \frac{\sin^2\theta_1}{2Z_\lambda}\left( (1-\lambda)\sin^2\theta_2 + \frac\lambda{2}\right)
\end{equation}
The probability still depends nontrivially on the angle of the gate yet to be entered by the $\phi$ channel, but this effect decreases with noise.
The classical limit gives a similar result.
We use the leftmost bit as the classical periodic channel bit, and the middle and right bits as the first and second probes respectively.
The classical model in state is a mixture of the two possible unitary evolutions.
\begin{eqnarray}
U_t|000\rangle = \cos\theta_1\cos\theta_2|000\rangle + \sin\theta_1\cos\theta_2|101\rangle + \cos\theta_1\sin\theta_2|100\rangle - \sin\theta_1\sin\theta_2|001\rangle\\
U_t|100\rangle = \cos\theta_1\cos\theta_2|111\rangle - \cos\theta_1\sin\theta_2|011\rangle - \sin\theta_1\cos\theta_2|010\rangle - \sin\theta_1\sin\theta_2|110\rangle
\end{eqnarray}
The classical weights are the norms of the projections of these states onto each classical state. The four decoherent projections are
\begin{eqnarray}
|\bar{\psi}_{00}\rangle = \langle 0,xx| U_t | 000\rangle = \cos\theta_1\cos\theta_2|00\rangle - \sin\theta_1\sin\theta_2|01\rangle\\
|\bar{\psi}_{11}\rangle = \langle 1,xx|U_t|100\rangle = \cos\theta_1\cos\theta_2|11\rangle -\sin\theta_1\sin\theta_2|10\rangle\\
|\bar{\psi}_{10}\rangle = \langle 1,xx|U_t|000\rangle = \sin\theta_1\cos\theta_2|01\rangle + \cos\theta_1\sin\theta_2|00\rangle \\
|\bar{\psi}_{01}\rangle =\langle 0,xx|U_t|100\rangle = - \cos\theta_1\sin\theta_2|11\rangle - \sin\theta_1\cos\theta_2|10\rangle
\end{eqnarray}
The corresponding weights are
\begin{eqnarray}
\omega_{00} =\langle \bar{\psi}_{00} |\bar{\psi}_{00}\rangle = \omega_{11} =\langle \bar{\psi}_{11} |\bar{\psi}_{11}\rangle =\cos^2\theta_1\cos^2\theta_2 +\sin^2\theta_1\sin^2\theta_2\\
\omega_{01} =\langle \bar{\psi}_{01} |\bar{\psi}_{01}\rangle = \omega_{10} =\langle \bar{\psi}_{10} |\bar{\psi}_{10}\rangle =\sin^2\theta_1\cos^2\theta_2 + \cos^2\theta_1\sin^2\theta_2
\end{eqnarray}
The partition function is defined by
\begin{equation}
Z = (1-k)(\omega_{00}+\omega_{11}) + k(\omega_{01}+\omega_{10})
\end{equation}
and density matrix by
\begin{equation}
\rho_{cl} = \frac{1-k}{Z_{2}}|\bar{\psi}_{00}\rangle\langle\bar{\psi}_{00}| + \frac{k}{Z_{cl}}|\bar{\psi}_{01}\rangle\langle \bar{\psi}_{01}| + \frac{1-k}{Z_{2}}|\bar{\psi}_{11}\rangle\langle\bar{\psi}_{11}| + \frac{k}{Z_{cl}}|\bar{\psi}_{10}\rangle\langle \bar{\psi}_{10}|
\end{equation}
The flip probability for the classical limit channel is therefore
\begin{eqnarray}
P_{flip-cl} = \frac{(1-k)\sin^2\theta_1\sin^2\theta_2 + k\sin^2\theta_1\cos^2\theta_2}{\cos^2\theta_1\cos^2\theta_2 +\sin^2\theta_1\sin^2\theta_2}\nonumber\\
= (1-k)(1+ \cot^2\theta_1\cot^2\theta_2)^{-1} + k(\cot^2\theta_1+\tan^2\theta_2)^{-1}
\end{eqnarray}
This decreased probability for detecting a change in the state of the post-selected system is fairly generic. A typical state will not be both an eigenstate of the measurement and of the circuit as a whole. As a result, multiple measurements often decrease the partition function of the system by reducing the normalization factor $N$. Consider the dragging of a spin by a large number of nearly parallel measurements, an example an anti Zeno effect. A grandfather circuit could be made if the dragging moves the state onto one that is orthogonal to the initially observed state of the channel. The spin will appear to be coupled to an unknown field that resists the dragging effect. It can also serve as the disentangling mechanism thought to be required in the final state model\cite{presk}\\
\section{Amnesia paradox}
This is essentially the time reverse of the unproven theorem paradox, in which two different input external states map onto the same external out state. The nonunitary map generated violates the second law by reducing the entropy of a maximally mixed qubit. Here we begin with an arbitrary qubit coming in on state $\psi_1$, which then acts as the control qubit in a controlled not operation onto the unknown state $\phi_{TM}$. Then the bits are swapped, sending our original $\psi_1$ into the loop to be projected onto $\phi_{TM}$. The external bit has now been XORed with the temporary copy of itself. As usual the channels are from left to right, the reference qubit, the periodic channel $\phi$, and then any external qubits $\psi_i$.
\begin{equation}
|\psi_{out}\rangle = |\phi_{Bell}\rangle \otimes ( \alpha |0\rangle + \beta |1\rangle )
\end{equation}
The unitary evolution is
\begin{eqnarray}
|\psi_{in}\rangle = U_{cn23}U_{swap23}|\psi_{out}\rangle \nonumber\\
= \frac1{\sqrt{2}}\left( \alpha|000\rangle + \alpha|101\rangle + \beta|011\rangle + \beta|110\rangle \right)
\end{eqnarray}
The Bell projection of this state gives us
\begin{equation}
|\bar{\psi}_B\rangle = \frac{\alpha+\beta}2|0\rangle = N_B|0\rangle
\end{equation}
\begin{figure}[h!]
\caption{An amnesia type circuit made from two controlled not gates.}
\centering
\includegraphics[width=0.4\textwidth]{amnesia.png}
\end{figure}
The scenario here is very similar to the measurement catastrophe. Destructive interference in the bit to be erased can make the normalization factor vanish.
The other three projections are
\begin{eqnarray}
|\bar{\psi}_-\rangle = \frac{\alpha-\beta}2|0\rangle =N_-|0\rangle\\
|\bar{\psi}_N\rangle = \frac{\alpha+\beta}2|1\rangle =N_N|1\rangle\\
|\bar{\psi}_{-N}\rangle = \frac{\beta-\alpha}2|1\rangle =N_{-N}|1\rangle
\end{eqnarray}
The mixture of these gives us the noisy channel approximation.
\begin{eqnarray}
Z_\lambda = (1-\lambda)N_B^2 + \lambda/4 =\frac14\left( \lambda + (1-\lambda)|\alpha+\beta|^2\right)\\
\rho_\lambda(\alpha) = \frac{1-\lambda}{4Z_\lambda}|\alpha+\beta|^2|0\rangle\langle 0| + \frac{\lambda}{8Z_\lambda}I
\end{eqnarray}
Just as with the measurement catastrophe, we can approach the singular point in the noisy approximation.
In that case the result is a maximally mixed state for singular input.
In the classical limit we have the simple result.
\begin{eqnarray}
|\bar{\psi}_{00}\rangle = \langle 0,x|U_{cn12}U_{swap12}|0,\psi_1\rangle = \alpha|0\rangle\\
|\bar{\psi}_{10}\rangle = \langle 1,x|U_{cn12}U_{swap12}|0,\psi_1\rangle = \beta|1\rangle\\
|\bar{\psi}_{11}\rangle = \langle 1,x|U_{cn12}U_{swap12}|1,\psi_1\rangle = \beta|0\rangle\\
|\bar{\psi}_{01}\rangle = \langle 0,x|U_{cn12}U_{swap12}|1,\psi_1\rangle = \alpha|1\rangle\\
\omega_{00}=\omega_{01} = \alpha^2\\
\omega_{11}=\omega_{10} = \beta\beta^\dagger \\
Z_{cl} = (1-k)(\omega_{00}+\omega_{11}) + k(\omega_{01}+\omega_{10}) = 1\\
\rho_{cl} = (1-k)|0\rangle\langle 0| + k|1\rangle\langle 1|
\end{eqnarray}
Where it appears that decoherence has eliminated the back-propagation effects, since $Z_{cl}$ is a constant.
This same process can be used to effectively erase entanglement with only local operations, another violation of unitarity.
Adding a second external channel $\psi_2$ that begins maximally entangled with $\psi_1$, we have
\begin{equation}
|\psi_{out}\rangle = |\phi_{Bell}\rangle \otimes ( \alpha|00\rangle + \beta|11\rangle )
\end{equation}
We can think of the bit to be erased as entangled with a distant channel $\psi_2$, so that none of the gates act on that channel.
The state after evolution is
\begin{eqnarray}
|\psi_{in} \rangle= U_{cn23}U_{swap23}|\psi_{out}\rangle \nonumber\\
= \frac1{\sqrt{2}}\left( \alpha|0000\rangle + \alpha|1010\rangle + \beta|0111\rangle + \beta|1101\rangle \right)
\end{eqnarray}
The projected states are
\begin{eqnarray}
|\bar{\psi}_B\rangle = \alpha|00\rangle + \beta|01\rangle =\frac12 |0\rangle \otimes (\alpha|0\rangle +\beta|1\rangle)\\
|\bar{\psi}_-\rangle = \alpha|00\rangle - \beta|01\rangle = \frac12|0\rangle \otimes (\alpha|0\rangle -\beta|1\rangle)\\
|\bar{\psi}_N\rangle =\alpha|10\rangle + \beta|11\rangle = \frac12 |1\rangle \otimes (\alpha|0\rangle +\beta|1\rangle)\\\
|\bar{\psi}_{-N}\rangle = -\alpha|10\rangle + \beta|11\rangle = \frac12|1\rangle \otimes (-\alpha|0\rangle +\beta|1\rangle)\\
\end{eqnarray}
All four of these states are product states. The breaking of entanglement creates a superposition of local states for the distant entangled channel $\psi_2$. \\
\begin{figure}[h!]
\caption{Amnesia circuits acting on one of an entangled pair breaks the entanglement, resulting in a product of two pure states.}
\centering
\includegraphics[width=0.4\textwidth]{break.png}
\end{figure}
This can potentially be detected locally by an observer with access only to the $\psi_2$ channel.
Since $\psi_2$ has no interactions in this circuit, the effects of the erasure of $\psi_1$ will back-propagate along $\psi_2$ until it interacts with something that dilutes or cancels the effect.
Just like the partition function for the classical version of the unentangled amnesia circuit, the normalization factor is constant.
\begin{equation}
N^2 = \frac14\alpha^2 + \frac14\beta\beta^\dagger = \frac14
\end{equation}
This indicates no back-propagation onto the $\psi_1$ channel.
The partition function is also constant
\begin{equation}
Z = \frac14(1-\lambda) + \frac\lambda{4} = \frac14\\
\end{equation}
The density matrix has six nonzero components given by
\begin{equation}
\rho = \left( \begin{array}{cccc} (1-\frac\lambda{2})\alpha^2 & (1-\lambda)\alpha\beta & 0 & 0 \\\\ (1-\lambda)\alpha\beta^\dagger & (1-\frac\lambda{2})\beta\beta^\dagger & 0 & 0 \\\\ 0 & 0 & \frac\lambda{2}\alpha^2 & 0 \\\\ 0 & 0 & 0 & \frac\lambda{2}\beta\beta^\dagger \end{array} \right)
\end{equation}
In addition to decoupling the two channels, the distant channel is slightly decohered in proportion to the noise of the erasing circuit.
This is an important effect since it introduces noise into any communication between the erasing circuit and the distant channel $\psi_2$.
Such channels may be created by acting with a controlled phase flip on the $\psi_1$ channel before erasing it.
This changes the sign of $\beta$ in the superposition of the $\psi_2$ channel.
Since the effect propagates back to the last interaction of the $\psi_2$ channel, this allows communication from the vicinity of the time loop to an arbitrary point in spacetime,
such as the causal past of the creation of the loop itself, or across an event horizon.
This could be regarded as a principle of non-censorship, although other effects may interfere with such signaling.\\
\begin{figure}[h!]
\caption{The phase information of the broken entanglement becomes localized, and detectable with local measurements on the second channel. Since this phase can be affected by a controlled phase flip on the first channel, it leads to a secondary loop.}
\centering
\includegraphics[width=0.4\textwidth]{secondary.png}
\end{figure}
The classical version of the circuit with entangled external channels is given by
\begin{eqnarray}
|\psi_{ent}\rangle = \alpha|00\rangle + \beta|11\rangle\\
|\bar{\psi}_{00}\rangle = \langle 0,x|U_{cn12}U_{swap12}|0,\psi_{ent}\rangle = \alpha|00\rangle\\
|\bar{\psi}_{10}\rangle = \langle 1,x|U_{cn12}U_{swap12}|0,\psi_{ent}\rangle = \beta|11\rangle\\
|\bar{\psi}_{11}\rangle = \langle 1,x|U_{cn12}U_{swap12}|1,\psi_{ent}\rangle = \beta|01\rangle\\
|\bar{\psi}_{01}\rangle = \langle 0,x|U_{cn12}U_{swap12}|1,\psi_{ent}\rangle = \alpha|10\rangle\\
\omega_{00}=\omega_{01} = \alpha^2\\
\omega_{11}=\omega_{10} = \beta\beta^\dagger \\
Z_{cl} = (1-k)(\omega_{00}+\omega_{11}) + k(\omega_{01}+\omega_{10}) = 1\\
\rho_{cl} = (1-k)\left( \alpha^2|00\rangle\langle 00| + \beta\beta^\dagger|01\rangle\langle 01| \right) + k\left( \alpha^2|10\rangle\langle 10| + \beta\beta^\dagger|11\rangle\langle 11|\right)
\end{eqnarray}
In the classical case the distant channel is completely decohered along the preferred basis states, but its eigenvalues remain the same, so no change will be detected locally.
Likewise the addition of a controlled phase flip acting on $\psi_1$ before the swap with $\psi$ fails to communicate information to $\psi_2$.\\
Returning to the case of measurement catastrophe in the unproven proof circuit, it was noted that this is an amnesia type circuit.
The same decoupling behavior for entangled incoming probe states is also present in the single $C_{not}$ gate circuit. Let
\begin{equation}
|\psi_{out}\rangle = |\phi_{Bell}\rangle \otimes ( \alpha|00\rangle + \beta|11\rangle )
\end{equation}
and
\begin{eqnarray}
|\psi_{in}\rangle = U_{cn23}|\psi_{out}\rangle\nonumber\\
= \frac1{\sqrt{2}}\left( \alpha |0000\rangle + \alpha|1110\rangle + \beta|0011\rangle + \beta|1101\rangle \right)
\end{eqnarray}
Two projections are nonzero.
\begin{eqnarray}
|\bar{\psi}_B\rangle = \frac{\alpha}{2}\left( |00\rangle + |10\rangle \right) + \frac{\beta}{2}\left( |11\rangle + |01\rangle\right) = \frac12\left( |0\rangle + |1\rangle\right) \otimes \left( \alpha|0\rangle + \beta|1\rangle \right)\\
|\bar{\psi}_-\rangle = \frac{\alpha}{2}\left( |00\rangle - |10\rangle \right) + \frac{\beta}{2}\left( |11\rangle - |01\rangle\right) = \frac12\left( |0\rangle - |1\rangle\right) \otimes \left( \alpha|0\rangle - \beta|1\rangle \right)\\
N_B^2=N_-^2=1/4
\end{eqnarray}
The standard mixture gives the noisy channel.
\begin{eqnarray}
Z_\lambda = (1-\lambda)N^2 + \lambda/4 = 1/4\\
\rho =(4 -3\lambda)|\bar{\psi}_B\rangle\langle \bar{\psi}_B| + \lambda |\bar{\psi}_-\rangle\langle \bar{\psi}_-|
\end{eqnarray}
Just as in the other case, the noise propagate to the decoupled state as decoherence.\\
The secondary loop circuit consists of a Bell pair, $\pi/4$ rotation on one channel and a conditional phase flip on the other. The second channel then goes into an amnesia circuit creating a product state. If the control of the conditional phase flip begins as an arbitrary product state we have
\begin{equation}
|\psi_{out}\rangle = |\phi_{TM}\rangle \otimes \frac1{\sqrt{2}}\left( |00\rangle + |11\rangle \right) \otimes \left( \alpha|0\rangle + \beta|1\rangle \right)
\end{equation}
The secondary channel is created by the action of four gates
\begin{equation}
|\psi_{in}\rangle = U_{cn32}U_{cn23}U_{cfp53}U_{rot4}|\psi_{out}\rangle
\end{equation}
The intermediate states are
\begin{eqnarray}
U_{rot4}|\psi_{out}\rangle = |\phi_{TM}\rangle \otimes \frac12\left( |00\rangle - |01\rangle +|10\rangle + |11\rangle\right) \otimes \left( \alpha|0\rangle + \beta|1\rangle\right)\\
U_{cfp53}U_{rot4}|\psi_{out}\rangle = \nonumber\\
|\phi_{TM}\rangle \otimes \frac12 \left( \alpha|000\rangle -\alpha|010\rangle + \alpha|100\rangle + \alpha|110\rangle +\beta|001\rangle -\beta|011\rangle -\beta|101\rangle - \beta|111\rangle \right)\\
\sqrt{8}|\psi_{in}\rangle = \alpha|00000\rangle -\alpha|00010\rangle + \alpha|01100\rangle + \alpha|01110\rangle \nonumber\\
+\beta|00001\rangle -\beta|00011\rangle -\beta|01101\rangle - \beta|01111\rangle \nonumber\\
\alpha|10100\rangle -\alpha|10110\rangle + \alpha|11000\rangle + \alpha|11010\rangle \nonumber\\
+\beta|10101\rangle -\beta|10111\rangle -\beta|11001\rangle - \beta|11011\rangle
\end{eqnarray}
and the projected states are
\begin{eqnarray}
\bar\psi_B = \frac14\left(\alpha|000\rangle -\alpha|010\rangle +\beta|001\rangle -\beta|011\rangle\right)\nonumber\\
+\frac14\left(\alpha|000\rangle + \alpha|010\rangle - \beta|001\rangle - \beta|011\rangle\right)\nonumber\\
= \frac12\left( \alpha|000\rangle -\beta|011\rangle \right)\\
\bar\psi_N = \frac12\left( \alpha|100\rangle -\beta|111\rangle \right)\\
\bar\psi_- = \frac14\left(\alpha|000\rangle -\alpha|010\rangle +\beta|001\rangle -\beta|011\rangle\right)\nonumber\\
-\frac14\left(\alpha|000\rangle + \alpha|010\rangle - \beta|001\rangle - \beta|011\rangle\right)\nonumber\\
= \frac12\left( \beta|001\rangle -\alpha|010\rangle \right)\\
\bar\psi_{-N}= \frac12\left( \beta|101\rangle -\alpha|110\rangle \right)
\end{eqnarray}
The movement of the relative amplitudes $\alpha$ and $\beta$ from the $\psi_3$ channel to the entangled state with the $\psi_2$ channel, without direct interaction indicates that classical measurements of $\psi_2$ obtain information about $\psi_3$, up to the limit of the rate of phase errors in the $\phi$ channel. The mixed state is
\begin{eqnarray}
Z = \frac\lambda{4} + \left(1-\lambda\right)N_B^2 = \frac14\\
\rho = Z^{-1}\left( 1-\frac34\lambda\right)|\bar\psi_B\rangle\langle\bar\psi_B| + \frac\lambda{4Z}|\bar\psi_-\rangle\langle\bar\psi_-| + \frac\lambda{4Z}|\bar\psi_N\rangle\langle\bar\psi_N| + \frac\lambda{4Z}|\bar\psi_{-N}\rangle\langle\bar\psi_{-N}|
\end{eqnarray}
The 'second order' grandfather paradox can be formed by carefully choosing the entangled states. If we take the measurement of $\psi_2$ as the control of the phase flip of the $\psi_1$ channel, we create the particular two channel state
\begin{eqnarray}
|\psi_{GF}\rangle = U_{cpf21}U_{rot2}|\psi_{Bell}\rangle\nonumber\\
= \frac12 \left( |00\rangle + |01\rangle - |10\rangle - |11\rangle \right) \nonumber\\
= \frac12 \left( |0\rangle - |1\rangle \right) \otimes \left( |0\rangle + |1\rangle \right)
\end{eqnarray}
The grandfather paradox of the secondary channel is really just the creation of a measurement catastrophe in the amnesia circuit. The creation of secondary channels and local interference effects of a final state projection black hole has been investigated also by \cite{yuri1,yuri2}. Because these secondary channels are not limited to the forward light cone, they can be used to communicate with observers outside the black hole.
\section{Back-propagation}
In an ordinary experiment, one of the first things to understand is the idea of controlled conditions.
The reproducibility of results, the assumption of the ability to do independent tests, and the concept of a prepared initial state of the system
all stem in some way from our basic intuitions of causality. A system that does not respect these can reasonably be called acausal.
It would be tempting to regulate the effects of projection to a special spacelike surface such as an event horizon, but this would also be misleading.
One reason is because of the way that entanglement acts with projection. States that are highly entangled with a projected channel to the point
that we might begin considering them as classical states, even then can be projected into very different mixtures.
Consider the controlled not grandfather circuit. When we prepare the control channel in an eigenstate, we are selecting the initial value
in much the same way as when we choose to accept in the ensemble only runs that satisfy the post-selection measurement.
If the two measurements are not consistent with each other, then the ensemble is empty.
The size of the ensemble is proportional to the normalization factor.
It can be thought of as a measure of mutual consistency between preselected and post-selected states.
One could imagine that the projection surface acts on the control channel and forces it into the state $|0\rangle$, but it must also force any
states entangled with that channel as well. The extent of this is such that we would not be able to point to the surface at which the
projection occured, since any record would also be projected.
The difference between such extensive reshaping of the wave-function and true backwards causal action may be impossible to detect even in principle.
Let us return to the controlled not circuit, but add a few additional probe states.
We can use these to monitor the value of the control qubit before and after it passes through the $C_{not}$ and acts on the peridic qubit $\phi$.
The product state is
\begin{eqnarray}
|\psi_{out}\rangle = |\phi_{Bell}\rangle \otimes \frac1{\sqrt{2}}\left( | 0\rangle + |1\rangle \right) \otimes |0\rangle \otimes |0\rangle \nonumber\\
= \frac12 \left( |00000\rangle + |11000\rangle + |00100\rangle + |11100\rangle\right)
\end{eqnarray}
Here we have put the control channel in the Hadamard state, and the last two channels are probes in the $|0\rangle$ state.
The action of the circuit will sandwich the 'gun' $C_{not}$ between two measurement gates.
\begin{eqnarray}
|\psi_{in}\rangle = U_{cn35}U_{cn32}U_{cn34}|\psi_{out}\rangle \nonumber\\
= \frac12\left( |00000\rangle + |11000\rangle + |01111\rangle + |10111\rangle \right)
\end{eqnarray}
Then projections give us
\begin{eqnarray}
|\bar{\psi}_B\rangle = \frac1{\sqrt{2}}|000\rangle \\
|\bar\psi_N\rangle = \frac1{\sqrt{2}}|111\rangle \\
|\bar\psi_-\rangle=|\bar\psi_{-N}\rangle = 0\\
N^2 = \frac12
\end{eqnarray}
This clearly shows that the value of the control bit, as far as entangled observers are concerned does not change in the proximity of the the circuit.
Further, that this holds either with or without noise, as well as in a classical limit for the periodic bit.
By specifying the initial state of the probe and control qubits, we manually blocking the effects of post-selection from propagating further back.
The the amplitudes of states in post-selection quantum mechanics back-propagate like the probabilities of a Bayesian model.
If we follow the control channel further back, we can observe behavior similar to that of the periodic channel itself.
A grandfather circuit with a controlled not acting on an input channel behaves much like a simple projection of that channel onto the $|0\rangle$ state would.
Consider a stubborn spin circuit combined with a controlled not paradox circuit.
The system will contain the periodic channel $\phi$, the control channel and a probe channel.
Before the emergence of the $\phi$ state from it's loop, we take a the control and probe channels to be in the state
\begin{equation}
|\psi_{34}\rangle = |00\rangle
\end{equation}
To simulate a measurement at an angle $\theta$, we apply a $C_{not}$ sandwiched between two rotations, one by $\theta$ and one by $-\theta$.
\begin{eqnarray}
|\psi_{34}'\rangle = U_{rot3}(-\theta)U_{cn34}U_{rot3}(\theta)|\psi_{34}\rangle \nonumber\\
= \cos^2\theta|00\rangle + \sin\theta\cos\theta|11\rangle - \sin\theta\cos\theta|10\rangle + \sin^2\theta|01\rangle
\end{eqnarray}
The controlled grandfather circuit then effectively projects the control qubit against $\lr0|$, giving
\begin{equation}
|\bar\psi_B\rangle = \cos^2\theta|00\rangle + \sin^2\theta|01\rangle
\end{equation}
If the original circuit contained noise, then the equivalent projection onto the control channel input will contain at least as much noise.
The skewing effect of post-selection need not be as extreme as we follow the system backwards.
Noise acting on the control channel before it acts on the periodic channel can have a screening effect, as can certain configurations of gates.
If we change this setup so that the grandfather circuit is replaced by a faulty gun, then the control channel will only inherit some of the selective bias.
We replace the final controlled not gate from the control bit into the periodic channel with a controlled rotation by an angle $\theta_g$.
The angle of the stubborn spin gates shall be $\theta_s$.
The product state is
\begin{equation}
|\psi_{out}\rangle = |\phi_{Bell}\rangle \otimes |00\rangle
\end{equation}
After acting with the first three gates this gives
\begin{eqnarray}
|\psi'\rangle = U_{rot3}(-\theta_s)U_{cn34}U_{rot3}(\theta_s)|\psi_{out}\rangle \nonumber\\
= |\phi_{Bell}\rangle \otimes \left ( \cos^2\theta_s|00\rangle + \sin\theta_s\cos\theta_s|11\rangle - \sin\theta_s\cos\theta_s|10\rangle + \sin^2\theta_s|01\rangle \right)
\end{eqnarray}
Then the action of the faulty gun circuit gives
\begin{eqnarray}
|\psi_{in}\rangle = U_{crot32}(\theta_g)|\psi'\rangle\\
=|\phi_{Bell}\rangle \otimes \left ( \cos^2\theta_s|00\rangle + \sin^2\theta_s|01\rangle \right) \nonumber\\
+ \left( \cos\theta_g|\phi_{Bell}\rangle + \frac{\sin\theta_g}{\sqrt{2}}( |01\rangle - |10\rangle ) \right) \otimes \left(\sin\theta_s\cos\theta_s|11\rangle - \sin\theta_s\cos\theta_s|10\rangle \right)
\end{eqnarray}
Now the projections are
\begin{eqnarray}
|\bar\psi_B\rangle = \cos^2\theta_s|00\rangle + \sin^2\theta_s|01\rangle + \cos\theta_g\sin\theta_s\cos\theta_s|11\rangle - \cos\theta_g\sin\theta_s\cos\theta_s|10\rangle\\
|\bar\psi_{-N}\rangle = \sin\theta_g\sin\theta_s\cos\theta_s|11\rangle - \sin\theta_g\sin\theta_s\cos\theta_s|10\rangle \\
|\bar\psi_N\rangle=|\bar\psi_-\rangle = 0
\end{eqnarray}
and the normalization factor is
\begin{equation}
N_B^2= 1-\sin^2\theta_g\sin^2\theta_s\cos^2\theta_s
\end{equation}
This means that without noise we have a transition probability for the control qubit of
\begin{equation}
P_{flip} = \frac1{N^2}|\langle 01 | \bar\psi_B\rangle |^2 + \frac1{N^2}|\langle 11 | \bar\psi_B\rangle |^2 = \sin^2\theta_s \left(\frac{1-\sin^2\theta_g\cos^2\theta_s}{1-2\sin^2\theta_g\sin^2\theta_s\cos^2\theta_s}\right)
\end{equation}
When $\theta_g=0$ the control qubit has no interaction with $\phi$, and normal behavior is restored. At $\theta_g=\pi/2$ the circuit behaves as though the projection were acting on the control bit as before.
The angle $\theta_g$ controls to some degree how much the projection back-propagates onto the control channel. This is incidentally also how much influence the control channel has over $\phi$.
If we remove the control by another step, so that it acts with $U_{crot}(\theta)$ on a second control channel which then acts with another similar gate on the periodic channel, we can see that some of the
stubborn spin force is dissipated into the intermediate channels. The full calculation is
\begin{eqnarray}
|\psi_{out}\rangle = |\phi_{Bell}\rangle \otimes |000\rangle\\
|\psi'\rangle = U_{rot4}(-\theta_s)U_{cn45}U_{rot4}(\theta_s)|\psi_{out}\rangle \nonumber\\
= |\phi_{Bell}\rangle \otimes \left ( \cos^2\theta_s|000\rangle + \sin\theta_s\cos\theta_s|011\rangle - \sin\theta_s\cos\theta_s|010\rangle + \sin^2\theta_s|001\rangle \right)\\
|\psi_{in}\rangle = U_{crot32}(\theta_{g2})U_{crot43}(\theta_{g1})|\psi'\rangle\\
=|\phi_{Bell}\rangle \otimes \left ( \cos^2\theta_s|000\rangle + \sin^2\theta_s|001\rangle \right) \nonumber\\
+ |\phi_{Bell}\rangle \otimes \cos\theta_{g1}\sin\theta_s\cos\theta_s\left(|011\rangle - |010\rangle \right)\nonumber\\
+ |\phi_{Bell}\rangle \otimes \cos\theta_{g2}\sin\theta_{g1}\sin\theta_s\cos\theta_s\left(|111\rangle - |110\rangle \right)\nonumber\\
+ \frac1{\sqrt{2}}\left( |01\rangle - |10\rangle \right) \otimes \sin\theta_{g1}\sin\theta_{g2}\sin\theta_s\cos\theta_s\left(|111\rangle - |110\rangle \right)
\end{eqnarray}
Th right hand bit probes whether a the first control is flipped by the first rotation. The middle bit is the first control, which acts on the left bit which then acts on $\phi$.
The projection without noise eliminates only the last four terms.
\begin{eqnarray}
|\bar\psi_B\rangle = \left ( \cos^2\theta_s|000\rangle + \sin^2\theta_s|001\rangle \right) \nonumber\\
+ \cos\theta_{g1}\sin\theta_s\cos\theta_s\left(|011\rangle - |010\rangle \right)\nonumber\\
+ \cos\theta_{g2}\sin\theta_{g1}\sin\theta_s\cos\theta_s\left(|111\rangle - |110\rangle \right)
\end{eqnarray}
The normalization factor and stubborn spin flip probability are
\begin{eqnarray}
N^2 = 1 - 2\sin^2\theta_s\cos^2\theta_s\sin^2\theta_{g1}\sin^2\theta_{g2} \\
P_{flip} = \sin^2\theta_s\left( \frac{1 - \cos^2\theta_s\sin^2\theta_{g1}\sin^2\theta_{g2}}{1 - 2\sin^2\theta_s\cos^2\theta_s\sin^2\theta_{g1}\sin^2\theta_{g2}}\right)
\end{eqnarray}
This result is equivalent to the substitution
\begin{equation}
\sin^2\theta_g \rightarrow \sin^2\theta_{g1}\sin^2\theta_{g2}
\end{equation}
as far as deviation from normal behavior for the probe qubit is concerned. The more interactions between a given probe channel and the more diluted the effects of post-selection typically are.
For the two step control circuit the classical limit is
\begin{eqnarray}
\omega_{00} = \omega_{11} = N^2 \\
\omega_{01} = \omega_{10} = 2 \sin^2\theta_{g1}\sin^2\theta_{g2}\sin^2\theta_s\cos^2\theta_s
\end{eqnarray}
The noisy classical and coherent partition functions are
\begin{eqnarray}
Z_{cl}=(1-k)(\omega_{00}+\omega_{11}) + k (\omega_{01} + \omega_{10} ) = 2(1-k)N^2 + 2k(1-N^2) \nonumber\\
= 2 - 2k -4(1-2k)\sin^2\theta_{g1}\sin^2\theta_{g2}\sin^2\theta_s\cos^2\theta_s\\
Z_\lambda = (1-\lambda)N^2 + \lambda/4\nonumber\\
= 1 - \frac34\lambda -2(1-\lambda)\sin^2\theta_{g1}\sin^2\theta_{g2}\sin^2\theta_s\cos^2\theta_s
\end{eqnarray}
To find the density matrix and flip probability in these noisy cases we need the complementary projection
\begin{eqnarray}
|\bar\psi_{-N}\rangle = \sin\theta_{g1}\sin\theta_{g2}\sin\theta_s\cos\theta_s\left(|111\rangle - |110\rangle \right)\\
= \langle 1 ,xxx| U_t | 0 ,\psi \rangle = \langle 0,xxx| U_t | 1,\psi \rangle
\end{eqnarray}
The density matrices are
\begin{eqnarray}
\rho_{cl} = \frac{1-k}{Z_{cl}}|\bar\psi_B\rangle\langle \bar\psi_B | + \frac{k}{Z_{cl}}|\bar\psi_{-N}\rangle\langle \bar\psi_{-N}|\\
\rho_\lambda =\frac{4-3\lambda}{4Z_\lambda}|\bar\psi_B\rangle\langle \bar\psi_B | + \frac{\lambda}{4Z_\lambda}|\bar\psi_{-N}\rangle\langle \bar\psi_{-N}|
\end{eqnarray}
These give flip probabilities of
\begin{equation}
P_{cl-flip} = \sin^2\theta_s\left(\frac{1-k-(1-2k)\sin^2\theta_{g1}\sin^2\theta_{g2}\cos^2\theta_s}{1-k-2(1-2k)\sin^2\theta_{g1}\sin^2\theta_{g2}\sin^2\theta_s\cos^2\theta_s}\right)
\end{equation}
and
\begin{equation}
P_{\lambda-flip} = \sin^2\theta_s\left(\frac{4-3\lambda -(4-2\lambda)\sin^2\theta_{g1}\sin^2\theta_{g2}\cos^2\theta_s}{4-3\lambda-8(1-\lambda)\sin^2\theta_{g1}\sin^2\theta_{g2}\sin^2\theta_s\cos^2\theta_s}\right)
\end{equation}
Another case to study is to add additional controlled not channels directly to the periodic channel.
The flip probability will have a stubborn spin form when the sum of the other control channels is even, and a complementary behavior when the sum of the other channels is odd.
When tracing over the other control channels the probability for an odd total acts like the noise parameter $k$ in classical limit.
For the N-controlled not circuit we have
\begin{equation}
|\psi_{out}\rangle = |\phi_{Bell}\rangle \otimes |\psi_0\rangle \otimes |\psi_1\rangle \otimes ... \otimes |\psi_m\rangle
\end{equation}
After each other channel acts with $C_{not}$ on $\phi$,
\begin{eqnarray}
|\psi_{in}\rangle = |\phi_{Bell}\rangle \otimes \left( \alpha_0|0\rangle \otimes |\psi_{even}\rangle + \beta_0|1\rangle \otimes |\psi_{odd}\rangle \right) \nonumber\\
+ \frac{1}{\sqrt{2}}\left( |01\rangle + |10\rangle \right) \otimes \left( \alpha_0|0\rangle \otimes |\psi_{odd}\rangle + \beta_0|1\rangle \otimes |\psi_{even}\rangle\right)
\end{eqnarray}
Where we have taken the even and odd subscripts to refer to the even and odd combinations of the other channels.
The projections are
\begin{eqnarray}
|\bar\psi_B\rangle = \alpha_0|0\rangle \otimes |\psi_{even}\rangle + \beta_0|1\rangle \otimes |\psi_{odd}\rangle \\
|\bar\psi_N\rangle = \alpha_0|0\rangle \otimes |\psi_{odd}\rangle + \beta_0|1\rangle \otimes |\psi_{even}\rangle \\
|\bar\psi_-\rangle = |\bar\psi_{-N}\rangle =0
\end{eqnarray}
Tracing over the remaining channels we have
\begin{eqnarray}
E^2 = \langle \psi_{even} | \psi_{even} \rangle \\
D^2 = \langle \psi_{odd} | \psi_{odd}\rangle \\
N_B^2 = E^2\alpha_0^2 + D^2\beta_0\beta_0^\dagger \\
N_N^2 = D^2\alpha_0^2 + E^2\beta_0\beta_0^\dagger
\end{eqnarray}
The closer $E^2$ gets to $1/2$, the closer $N_B^2$ becomes to constant. This indicates less back-propagation onto the $\psi_0$ channel for $E^2$ closer to $1/2$.
If we add another control channel, then the value of $E^2$ will in general become closer to $1/2$ for any new product state we choose.
\begin{eqnarray}
E_{m+1}^2 = E_m^2\alpha_{m+1}^2 + D_m^2(1-\alpha_{m+1}^2)\\
D_{m+1}^2 = D_m^2\alpha_{m+1}^2 + E_m^2(1-\alpha_{m+1}^2)
\end{eqnarray}
Comparing this to the noisy classical controlled not circuit we find
\begin{equation}
k_{eff} = D^2
\end{equation}
Furthermore we can see the strictly increasing effective noise by looking at the quantity $\epsilon$, which strictly decreases with extra channels.
\begin{eqnarray}
\epsilon_{m} = |E_m^2-\frac12| = (2\alpha_{m-1}^2 - 1)\epsilon_{m-1} \\
k_{eff} = \frac12 - \epsilon
\end{eqnarray}
Since we can model noise in the control channel as the action of gates by hidden additional channels,
we can see that noise in between our measurement of the stubborn spin on
the control channel and the $C_{not}$ onto the periodic channel
will partially shield the control channel at earlier times from stubborn spin like effects.
Just as noise in a normal circuit decreases the effect of the initial state on post noise measurements, future noise
shields present measurements from post-selection bias.
In general, if we expand a particular circuit by adding gates to the input channels,
we can observe irregular behavior further and further prior to the circuit's interaction with the periodic qubits.
In a sense, preparing a precise input state for a postselected circuit is no more trivial than guaranteeing a particular output value.
The degree to which these effects would otherwise back-propagate can be seen in the dependence of the partition function on the input channel states.
In the controlled grandfather circuit without noise, the partition function is equal to the normalization factor squared, and thus the square of the $|0\rangle$ component of the control qubit.
In order to preselect the the control channel, we take a mixed qubit and perform a measurement to project it onto an eigenstate which we wish to use as the control input.
We could say the post-selection failed because our prepared control state was $|1\rangle$, or the converse.
In all cases where post-selection succeeded, we were unable to prepare the control state.
A measurement of our supposedly maximally mixed input state does not give the appropriate statistics.
Instead it behaves as a weighted ensemble, where the population of each pure state input is given by the partition function of the circuit.
The resulting skewed density matrix is a measure of the input bias of the circuit. For completeness this is given as
\begin{equation}
\bar\rho = \frac{\int Z(\psi)|\psi\rangle\langle\psi| d\psi}{\int Z d\psi}
\end{equation}
This is of course the same form as for the density matrix of final outgoing values for the non-periodic channels.
In a sense, both the initial and final values of the entangled channels are 'outputs' of the circuit, which is simply in a thermal state constrained by the pre and post-selected boundary conditions.
It should be noted that the stubborn spin and all the other pseudo-forces introduced by post-selection are $\emph{entropic forces}$.
\section{Entropy renormalized}
In several cases we have seen that selection can reduce the entropy of a system. The treatment here is given in \cite{me,me2}, but is included again here for completeness.
In the grandfather paradox unique states for the control qubits, and in the amnesia paradox, unique pure out states.
In order to describe larger systems and better understand the behavior post-selection induces it s important to quantify this effect.
If we limit ourselves to two qubit gates, the answer is simple. With four orthogonal projection states in the Bell model and two in the classical limit,
we can fix the value of two and one control qubits respectively.
In the classical grandfather circuit a single $C_{not}$ will fix the value of its control qubit to $|0\rangle$.
In the coherent version of the circuit we can add a controlled phase flip which will accomplish the same for its control qubit.
The $ C_{pf}+C_{not}$ circuit is described by
\begin{eqnarray}
|\psi_{out}\rangle = |\phi_{Bell}\rangle \otimes ( \alpha_1|0\rangle + \beta_1|1\rangle ) \otimes ( \alpha_2|0\rangle + \beta_2|1\rangle ) \\
|\psi_{in}\rangle = U_{cpf21}U_{cnot31}|\psi_{out}\rangle =\nonumber\\
\left( U_{not}|\phi\rangle \otimes \alpha_1\beta_2|01\rangle \right) +\left( U_{pf}|\phi\rangle \otimes \beta_1\alpha_2|10\rangle \right) + \left( U_{rot}|\phi\rangle \otimes \beta_1\beta_2|11\rangle \right) + \left(|\phi\rangle \otimes \alpha_1\alpha_2|00\rangle \right)
\end{eqnarray}
Each $\phi$ part is one of the four orthogonal Bell states. Taking the projection will single out one of these four joint states for the control qubits.
In the noiseless case only the $|00\rangle$ term will survive.
Adding additional controlled not and controlled phase flips dilutes this fixing effect over multiple qubits.
This is shown for $C_{not}$ in the third party paradox, and for $C_{pf}$ in the null space of the Bell projection of the in state for the
twice watched pot circuits.
The situation changes with the introduction of the NAND, or Deutch gates.
In these cases any number of external qubits can be fixed up to the level of noise.
Consider a circuit consisting of the periodic channel $\phi$ and two outside channels $\psi_1$, and $\psi_2$.
The double controlled rotation gate will take the $\psi_1$ and $\psi_2$ as controls and act on the state $\phi$ conditionally with the rotation $\theta_1$.
Then $\phi$ is further rotated by $\theta_2$ unconditionally.
The states are
\begin{eqnarray}
|\psi_{out}\rangle = |\phi\rangle \otimes ( \alpha_1|0\rangle + \beta_1|1\rangle ) \otimes ( \alpha_2|0\rangle + \beta_2|1\rangle ) \\
|\psi_{in}\rangle = U_{r1}(\theta_2)U_{ccr231}(\theta_1)|\psi_{out}\rangle =\nonumber\\
\left( U_r(\theta_2)|\phi\rangle \right) \otimes \left(\alpha_1\alpha_2|00\rangle + \alpha_1\beta_2|01\rangle + \beta_1\alpha_2|10\rangle \right) + \left(U_r(\theta_1+\theta_2)|\phi\rangle\right) \otimes \beta_1\beta_2|11\rangle
\end{eqnarray}
For the Bell state model we have
\begin{eqnarray}
|\phi\rangle = \frac1{\sqrt{2}}\left( |00\rangle + |11\rangle \right)\\
U_r(\theta)|\phi\rangle = \frac1{\sqrt{2}}\left( \cos\theta|00\rangle + \cos\theta|11\rangle + \sin\theta|01\rangle - \sin\theta|10\rangle \right)
\end{eqnarray}
The Bell projection of each part of the wave-function is proportional to the cosine of the total rotation of the $\phi$ part.
The four projections are,
\begin{eqnarray}
|\bar{\psi}_B\rangle = \cos\theta_2 \left(\alpha_1\alpha_2|00\rangle + \alpha_1\beta_2|01\rangle + \beta_1\alpha_2|10\rangle\right) + \cos(\theta_1+\theta_2)\beta_1\beta_2|11\rangle\\
|\bar{\psi}_-\rangle = |\bar{\psi}_{N}\rangle = 0 \\
|\bar{\psi}_{-N}\rangle = \sin\theta_2\left(\alpha_1\alpha_2|00\rangle + \alpha_1\beta_2|01\rangle + \beta_1\alpha_2|10\rangle\right) + \sin(\theta_1+\theta_2)\beta_1\beta_2|11\rangle\\
\end{eqnarray}
By choosing the angles $\theta_1$ and $\theta_2$ we can select or deselect the $|11\rangle$ state from the ensemble.
Setting $\theta_1=\theta_2=\pi/2$, we have
\begin{eqnarray}
|\bar{\psi}_B\rangle = -\beta_1\beta_2|11\rangle\\
|\bar{\psi}_{-N}\rangle = \alpha_1\alpha_2|00\rangle + \alpha_1\beta_2|01\rangle + \beta_1\alpha_2|10\rangle
\end{eqnarray}
Then the noisy case is given by
\begin{eqnarray}
N^2 = \beta_1\beta_1^\dagger\beta_2\beta_2^\dagger\\
Z = (1-\lambda)N^2 + \lambda/4\\
\rho = (1-\frac34\lambda)\frac{N^2}Z|11\rangle\langle 11| + \frac\lambda{4Z}|\bar{\psi}_{-N}\rangle\langle \bar{\psi}_{-N}|
\end{eqnarray}
The classical weights behave similarly,
\begin{eqnarray}
\omega_{00} = \omega_{11} = \beta_1\beta_1^\dagger\beta_2\beta_2^\dagger\\
\omega_{10} = \omega_{01} = 1 - \omega_{00}
\end{eqnarray}
for the same choice of $\theta_1$ and $\theta_2$. The classical limit partition function is
\begin{equation}
Z_{cl} = 2(1-k)N^2 + 2k(1-N^2)
\end{equation}
and density the matrix
\begin{equation}
\rho_{cl} = \frac{2(1-k)N^2}Z |11\rangle\langle 11| + \frac{2k}Z|\bar{\psi}_{-N}\rangle\langle \bar{\psi}_{-N}|
\end{equation}
These results quickly extend to larger numbers of control qubits
In the case of a triple controlled rotation gate given by
\begin{eqnarray}
U_{cccr}(\theta) = I - (1-\cos\theta)|1111\rangle\langle 1111| - (1-\cos\theta
)|1110\rangle\langle 1110| \nonumber\\
+ \sin\theta|1111\rangle \langle 1110| - \sin\theta|1110\rangle \langle 1111|
\end{eqnarray}
we can select the $|111\rangle$ state, giving the projections,
\begin{eqnarray}
|\psi_{123}\rangle = ( \alpha_1|0\rangle + \beta_1|1\rangle ) \otimes ( \alpha_2|0\rangle + \beta_2|1\rangle ) \otimes ( \alpha_3|0\rangle + \beta_3|1\rangle )\\
|\bar{\psi}_B\rangle = \beta_1\beta_2\beta_3|111\rangle\\
|\bar{\psi}_-\rangle = |\bar{\psi}_{N}\rangle = 0 \\
|\bar{\psi}_{-N}\rangle = |\psi_{123}\rangle - |\bar{\psi}_B\rangle
\end{eqnarray}
and mixture of
\begin{eqnarray}
N^2 = \beta_1\beta_2\beta_3 \beta_1^\dagger\beta_2^\dagger\beta_3^\dagger\\
Z = (1-\lambda)N^2 + \lambda/4 \\
\rho = \left(1-\frac34\lambda\right)\frac{N^2}Z |111\rangle\langle 111| + \frac{\lambda}{4Z}|\bar{\psi}_{-N}\rangle\langle \bar{\psi}_{-N}|
\end{eqnarray}
For any number of control qubits, the mixed state gives us a skew factor for the probability. Defining the coefficients as
\begin{equation}
\rho = a|111\rangle\langle 111| + b|\bar{\psi}_{-N}\rangle\langle \bar{\psi}_{-N} |
\end{equation}
comparing the ratio for the pure product state $a/b$, to the skewed ratio of the mixture of projected states ${\bar{a}}/{\bar{b}}$, gives
\begin{equation}
\Omega_\lambda = \frac{b}{a} \cdot \bar{a}/\bar{b} = \frac4\lambda -3
\end{equation}
for the noisy Bell model, and
\begin{equation}
\Omega_k = \frac{b}{a} \cdot \bar{a}/\bar{b} = \frac1k -1
\end{equation}
for the classical limit.
The entropy of a mixed ensemble with $n+1$ possible states can be written as
\begin{equation}
-S_0 = a \ln a + \sum_i^n b_i \ln b_i
\end{equation}
The entropy of the skewed ensemble created by selecting the state associated with $a$ is
\begin{eqnarray}
-S_\Omega = \bar{a}\ln \bar{a} + \sum_i^n \bar{b}_i \ln \bar{b}_i \nonumber\\
= \frac{\Omega a}{Z'} \ln\left(\frac{\Omega a}{Z'}\right) + \sum_i^n\frac{b_i}{Z'} \ln\left(\frac{b_i}{Z'}\right)
\end{eqnarray}
where we have defined
\begin{equation}
Z' = \left(\Omega - 1 \right) a +1
\end{equation}
We are interested in the difference of the two entropies.
\begin{eqnarray}
\Delta S = \bar{S}-S_0 = \ln Z' - \frac{\Omega a}{Z'}\ln \Omega - \frac\Omega{Z'}\left(a\ln a\right) - \frac1{Z'}\left( \sum_i^n b_i \ln b_i \right) - S_0\nonumber\\
= \left(\frac1{Z'}-1\right)S_0 + \ln Z' - \frac{a}{Z'}\Omega\ln\Omega - \frac{(\Omega-1)a}{Z'}\ln a \nonumber\\
= \left(\frac1{Z'}-1\right)\left( S_0 + \ln a \right) + \ln Z' - \frac{a}{Z'}\Omega\ln\Omega
\end{eqnarray}
Maximizing $S_0$ we have,
\begin{equation}
S_0 = -\ln a
\end{equation}
which eliminates the first term. Then maximizing for $a$ we have
\begin{eqnarray}
a_{max} = (\Omega-1)^{-2}\left( 1 - \Omega + \Omega \ln \Omega\right)\\
Z_{max} = \frac{\Omega\ln\Omega}{\Omega-1}\\
\Delta S_{max} = \ln Z_{max} - Z_{max} + 1 \nonumber\\
= -\ln(\Omega-1) - \frac{\ln\Omega}{\Omega-1} + \ln\ln\Omega + 1
\end{eqnarray}
This entropy difference can be exploited to do work. An example of this is a post-selected entangled Szilard engine.
In the classic Szilard engine a particle is trapped in a box that is then divided into two equal chambers.
An observer then measures which chamber the particle is in, and moves a piston that collapses the opposite chamber.
When the partition dividing the chambers is removed, the pressure of the gas particle pushes the piston out, doing work.
The process is then repeated, thus extracting useful work from a single heat bath.
The hitch in this scenario is that it is not a complete cycle until the bit of information gained by measurement is erased, increasing entropy.
The entropy cost for operating the engine balances out the work done, preserving the second law of thermodynamics.
This erasure cost is given by Landauer's principle.
We can attempt to circumvent this by changing the relative size of the chambers, but on average the cost will at least cancel and often outweigh any work gained.
However, if the observation of the particle is tied to a post-selected variable, then out ensemble can essentially cherry pick out those cases in which we come out ahead.
In this thought experiment the net work done is
\begin{equation}
W = -P_{left}\ln x - P_{right}\ln ( 1-x ) - \ln 2
\end{equation}
The partition is located a distance $ 0 < x < 1$ from the left side of the box. The constant term is from the erasure of the single bit of observation done as an intermediate step in entanglement.
The post-selected probabilities are
\begin{equation}
P_{left} = 1 - P_{right} = \frac{\Omega x}{ (\Omega -1)x + 1 }
\end{equation}
The maximum of $W$ requires solving a transcendental equation, so is best done numerically. An approximation shows that the leading term is of order $\ln\Omega$. It may seem that this bound could be violated in the following way. Suppose the partition is placed in the center, and then one of the pistons is removed to a large distance. The entropy difference between entangled and unentangled ensembles for this state may grow arbitrarily large. However, work will only be done if the particle is in the expanding chamber. By selecting the particle into the smaller chamber after the expansion, it may seem that entropy is reduced by a large amount, but the selection process merely back-propagates to before the insertion of the partition, rather than force any sort of 'tunneling' type event. Because of this, no additional work is done. It would be more appropriate to say that a logarithmic work bound exists. Another way to understand this effect is to consider Crook's fluctuation theorem.
\begin{equation}
P_{A\rightarrow B}( W) = P_{A \leftarrow B}( -W ) \exp\left[ \beta (W - \Delta F)\right]
\end{equation}
The ratio of probabilities of transitions between states of different free energies $\Delta F$ is exponential in the work done.
By skewing these probabilities with post-selection we effectively introduce a multiplier $\Omega$ for the skewed probability ratio. This manifests in an effective shift in the free energy of the system of
\begin{equation}
W_{TM} = \beta^{-1}\ln \Omega = TdS
\end{equation}
We may refer to this quantity as the negentropy or entropy potential of a time machine. It represents the amount of excess entropy that must be produced by any process that creates an effective time loop.
\section{Measurement and error}
In ordinary quantum mechanics we learn that the wave-function itself is not an observable.
Probabilities of events are normally only measurable by testing the frequency of the events.
The reason we have to comb through large data sets to test models that predict very unlikely events or small changes in
the relative probability of more common ones is that we generally have to wait and see to construct a frequentist model.
In post-selected systems the amplitude of some branch of the wave-function is enhanced or suppressed by a particular amount, depending on how
the system is entangled with the periodic channel. By controlling the degree of entanglement we can construct a sort of probability microscope.
This allows us to observe rare processes by renormalizing parts of the wave function that we want to interact more strongly with an observable.
To do this, we employ a 'fuse'. That is a probabilistic algorithm with a known exponentially small success probability.
This fuse is then placed in parallel with the problem we wish to estimate. We can then compare the success rates of the two paths to estimate the success rate of the unknown path.
If the fuse is 'blown' we can place an exponentially small upper bound on the probability of the unknown process.
There is a caveat that unless back-propagation is compensated for in some way, iterative measurements may interfere with each other.\\
Another distinguishing feature of nonlinear modifications to quantum mechanics is the ability to more reliably distinguish non-orthogonal states.
Any process which allows this typically leads back to the creation of effective CTCs and all the paradoxes of this paper.
In ordinary quantum mechanics the optimal method for distinguishing between two pure states in
a mixture using projective measurement is to choose one of the eigenvectors to be one of the two states to be distinguished. Some improvement over this is gained by
using the technique of positive operator valued measurements. It is a well known result that the three operators
\begin{eqnarray}
\Pi_a = \frac{1 - |b\rangle\langle b|}{1 + |\langle a| b \rangle |}\\
\Pi_b = \frac{1 - |a\rangle\langle a|}{1 + |\langle a| b \rangle |}\\
\Pi_n = 1 - \Pi_0 - \Pi_1
\end{eqnarray}
Allow for the maximum success probability for discerning the two states.
By acting with a not and two $C_{not}$ gates we can alight the post-selection to project out the 'inconclusive' demension of the expanded POVM space.
Rather than an inconclusive result probability of
\begin{equation}
P_n = | \langle a | b \rangle |
\end{equation}
we have
\begin{equation}
\bar{P}_n = \frac{P_n}{P_n + \Omega -\Omega P_n}
\end{equation}
Because the probability of a false identification is zero, it is unchanged by the renormalization. The relative ratio of detections is unaffected, only the ambiguous result decreases.
The mixture of the discerned bit is not otherwise affected, remaining a combination of the three orthogonal outputs. The lack of further back-propagation is due to the symmetry of the measurement process between the $|a\rangle$ and $|b\rangle$ states.
In a more general case, states in the mixture are renormalized away from the mean. If a third state is added to the mixture it will be suppressed in the ensemble by how close it is to the inconclusive projection.
For states that form an angle $\theta$ with the renormalized mean state of $|a\rangle+|b\rangle$,
\begin{equation}
P_n(\theta) =\frac{4P_n\cos^2\theta}{2 + \langle a | b \rangle + \langle b | a \rangle }
\end{equation}
and the relative weight in the ensemble will be skewed by a factor
\begin{equation}
w(\theta) = 1-P_n(\theta) + \bar{P}_n(\theta)
\end{equation}
In the limit as we attempt to discern nearer and nearer states, we re-create a grandfather paradox centered on the mean state, and amplifying small fluctuations away from it.\\
One of the more curious things in measurement theory is the non-uniqueness of the composition of density matrices.
The density matrix normally contains all the information needed for the prediction of measurement of a system.
In standard decoherence theory it also represents a degree of ignorance of the system's entanglement with the environment.
The complementary information resides in the hidden degrees of freedom and the mutual phase of the state components that are traced over to give the density matrix.
For post-selected quantum mechanics density matrices maintain a similar ambiguity as in regular quantum theory.
Each weight and projection are the same if the unprojected density matrix is the same, and the post-selected matrix is formed from these quantities.
Unlike the increased state discrepancy, there is no outcome to select against, since the value of no channel can depend on the difference between two equivalent density matrices.\\
The treatment of weak measurements is thoroughly covered elsewhere, and being designed for post-selected systems, needs no modification\cite{ahr1}.
In the presence of noise, we add to the ensemble the weak result of projecting against each orthogonal direction for the Bell state or classical channel,
and combine the sub-ensembles with a relative weight of $\Omega$ multiplying the desired projection's noiseless result.
Suppose a weak measurement is given by an operator $A$ inserted into the circuit between two unitary evolutions. Normally we would have,
\begin{equation}
\langle A \rangle_{weak} = \langle \psi_{in} | U_2 A U_1 |\psi_{out}\rangle
\end{equation}
for exact post-selection. The expectation value with noise is simply the weighted average of multiple weak measurements.
\begin{equation}
\langle A \rangle_\Omega = \frac1{Z}\left( \langle \phi_\perp | U_2 A U_1 | \phi \rangle + \Omega\langle \phi | U_2 A U_1 | \phi \rangle \right)
\end{equation}
The suppression of ambiguous measurement outcomes can also be used to improve error correction.
An example of this is a two qubit scheme, in which we periodically measure the parity of two entangled qubits, and select against odd states.
Two Bell states, one being the periodic channel, and the other the communication channel where we have a small probability of $\epsilon$ of flipping a bit.
We take the noisy product state to be
\begin{equation}
|\psi_{out}\rangle = |\phi_{TM}\rangle \otimes \left( (1-\epsilon) \left( \alpha|00\rangle + \beta|11\rangle \right) + \sqrt{\epsilon(1-\epsilon)} \left( \alpha + \beta\right)\left( |01\rangle + |10\rangle\right) + \epsilon\left( \beta |00\rangle + \alpha |11\rangle\right) \right)
\end{equation}
Then we use two cnot gates one from each part of the second Bell channel.
\begin{equation}
|\psi_{in}\rangle = U_{cnot32}U_{cnot42}|\psi_{out}\rangle
\end{equation}
Then the projected states are
\begin{eqnarray}
|\bar\psi_B \rangle = \left(\alpha(1-\epsilon) + \beta\epsilon\right)|00\rangle + \left(\beta(1-\epsilon) + \alpha\epsilon \right)|00\rangle \\
|\bar\psi_N \rangle = (1-\epsilon)\epsilon(\alpha + \beta\left( |01\rangle + |10\rangle \right)
\end{eqnarray}
The partition function will be
\begin{equation}
Z = \Omega\langle \bar\psi_B | \bar\psi_B\rangle + \langle \bar\psi_N | \bar\psi_N \rangle
\end{equation}
\begin{figure}[h!]
\caption{A scambler $U_t$ and selection based error correction.}
\centering
\includegraphics[width=0.4\textwidth]{errorc.png}
\end{figure}
In the limit of $\Omega\epsilon>>1$ it appears that we can correct $n$ errors with only $n+1$ qubits, up to the limit of fidelity
\begin{equation}
f_n =\frac{(1-\epsilon)^{n+1}}{(1-\epsilon)^{n+1} + \epsilon^n}
\end{equation}
If $\Omega$ is smaller, the ideal correction scheme depends upon it's value relative to the number of qubits and the individual error rate $\epsilon$.
\section{Tourist trap}
In a famous criticism of the science fiction of time travel, a few authors such as Hawking have remarked about the distinct lack of tourists from the future. The idea is that eventually someone will want to visit a particular moment, so we should immediately see them. The logic is similar to a combination of the Fermi and Obler paradoxes. Resolutions follow several avenues. First the asymptotic behavior of this 'time traveler expectation value' may easily yield a finite, even small result if it decays quickly enough. The work required to create a channel is extensive in the number of qubits, and with a low error rate it may be extremely large. At constant temperature the normalization factor for a loop in a thermal bath typically falls of exponentially. \\
To set up a model circuit for this problem will require several channels. In some capacity we are looking at a slightly more complex version of the unproven proof scenario, but with another twist, we want to model a conditionally post-selected channel. One channel we will designate $|p\rangle$, a qubit corresponding to whether the periodic channels are 'on'. Next we will use two sets of $n$-qubit control channels initialized to the $|0\rangle^n$ state. One qubit from each set of control channels will act on a single qubit from the set of periodic channels, creating an $n$-qubit version of the third party paradox circuit, with all null inputs.To send a message, a set of $n$ parallel Toffoli gates will conditionally copy a message onto one of the control channel sets if the power qubit is on. To receive the message a set of $n$ Bell pairs is used. One member of each pair is conditionally copied to the other set of control channels with a similar set of Toffoli gates, again taking the power qubit as one of the two controls in each gate. The Bell partners will, assuming post-selection is successful and the power qubit is on throughout, effectively relay the message. If we wish, we can use the Bell partners as the sending channel, or act on it with gates in between. If we want to simulate the idea of a time traveler bringing us the plans for the time machine itself, we can condition the value of the power qubit on the message received, as the receiving channel is measured before either set of Toffoli gates acts. Lastly we can post-select to give the power qubit an independent skew factor, representing an arbitrary degree of waste entropy production, or other back-reaction effects. The initial product state is composed of the periodic pairs, the receiving pairs, the power qubit, and the sending qubits.
\begin{equation}
|\psi_{out}\rangle = |\phi_{Bell}\rangle^n \otimes |p\rangle \otimes | m \rangle \otimes | \psi_{Bell} \rangle^n
\end{equation}
We can relate the message to be sent to the received message by a unitary evolution of received message and the environment.
\begin{equation}
|m\rangle = U_t ( |\psi_{bell}\rangle^n \otimes |e\rangle )
\end{equation}
The second members of the receiving Bell pairs can be traced over to simulate the decoherence of the received classical message due to the interaction with the rest of the world.
The normalization factor will be a sum over the projections against of each of the $2^{2n}$ orthogonal basis states for the $n$-qubit periodic channel, weighted by the number of errors. For a number of errors $n-q$ we have
\begin{equation}
\bar\psi_q = \langle \phi^q \otimes \phi_{\perp}^{n-q} | U_t | \phi_{bell}^n\rangle
\end{equation}
and therefore partition function of,
\begin{equation}
Z = \sum_{\bar\psi} \epsilon^n\Omega^q \langle \bar\psi_q|\bar\psi_q\rangle
\end{equation}
where $\epsilon$ is the per bit error rate. This is assuming all states are projected. But if only some are projected it becomes more complicated.
Effectively we are no longer projecting against the rotated Bell states, but against the product of those states and the power qubit. For projections where the power qubit is off the skew factor between the different orthogonal Bell states is unity. For the single qubit conditional periodic channel we have eight possible projections, and so eight possible weights. If we use $\omega_{on/off}$ for relative weight of normal and noisily post-selected subensembles,
the effective partition function will be
\begin{equation}
Z_p = \omega_{on}\left(\left(1-\frac34\lambda\right)N^2 + \frac\lambda{4}\right) + \omega_{off}
\end{equation}
Including members in the ensemble that are not post-selected effectively increases the noise.
\begin{equation}
\Omega_p = \frac{\omega_{on}}{\omega_{off}+\lambda/4}\Omega_\lambda\cdot\frac\lambda{4}
\end{equation}
This trade off can be seen in the third part paradox circuit if we add an additional rotation gate to the periodic channel. Both noise and any conditionality on the projection add a representative unbiased subensemble to our post-selected ensemble, decreasing skew effects.\\
Suppose that the power channel is set to some function of the message and environment. This could correspond to the recognition of a message separate from random noise, or some other physical prerequisite for post-selection. Because the normalization factor is always less than one, unless we insert an a priori bias for the power channel to be on, it will be set to on strictly less frequently than it would if there were no post-selection. Since only histories where post-selection occurs can be deselected from an ensemble, an arbitrary ensemble will have a bias against producing such events, if such selection is possible. Due to the dissipative nature of decoherence, any 'time tourist' would be exponentially unlikely to satisfy the periodic condition. Since mutual phase information would be lost when they interacted with the environment, they act like a large $n$ channel with a random distribution of phase flip gates. The normalization factor of the state corresponding to their appearance decays exponentially with interaction, down to an equilibrium value exponentially small in $n$. Since a thermal bath has a probability of producing any message that is exponentially small in $n$, the vacuum expectation of 'time tourists' may not be significantly different than it would otherwise be for a thermal bath in normal quantum mechanics. Unless the creation of projections and/or time machines is associated with a large increase in entropy, it will be as rare as the corresponding fluctuation of a thermal ensemble that decreases entropy by an amount equal to the entropy potential of the time machine. In order to realize a low error rate for a channel, the normalization factor for a typical message state must be close to unity, otherwise errors will be amplified, and our 'traveler' scrambled.\\
Returning to our toy model circuit, suppose there is a single string that can be formed by the $n$-qubits that will tell us how to build our post-selection/time machine. Since the third party message is received when we measure one of each Bell pair, we need not have built the rest of the circuit yet! What is the probability of getting this message compared to all the other random strings we might find by measuring a bunch of Bell pairs? It is actually $\emph{less}$ than $2^{-n}$. This is because the possibility that we will attempt to send a different message, accidentally or not, will only $\emph{decrease}$ the weight of the histories where the message is received in the first place. One can offset this by adding a skew factor of that history, but this is equivalent to adding a high a priori probability that the particular message will appear. This effect can be thought of as a soft chronology protection principle, that the expectation value of causality violations should drop off at least exponentially in both time and system size, similar to thermodynamic behavior given by the fluctuation theorem.
Another approach to this paradox is to generalize the renormalization process. When we perform a noisy post-selection we are dictating the ratio of particular sub-ensembles to form a skewed ensemble. In this case the ratios of sub-ensembles are given by the projection against the time machine eigenstates and an additional weight, the effective error rate, which prevents our ensemble size from vanishing in general. When looking at projection over larger numbers of states, there is more freedom to assign differences in sub-ensemble ratios. In the Bell model, for example, we could give a different probability for phase errors relative to bit errors in some preferred basis. In multi-qubit channels, there is no a priori reason that the error rates would be the same for different channels.
Since conditional projection is equivalent to projection against multiple variables, we may also consider limiting back-propagation with a rule. Given two sets of sub-ensembles, one in which a state projection occurs and another where it does not, we may fix the relative probability of the two within the total ensemble, and allow post-selection to only redistribute the weight of histories within sub ensembles that have the same set of 'projection events'.
Let us proceed to a more definite example.
Suppose we take three Bell pairs and measure one particle of each pair to obtain a classical 3 bit message.
Suppose we have a method of applying some state projection to these partner particles, but only if the first two bits of the classical message are 0. We can then choose the to project the last qubit however we wish, forming a single qubit periodic channel.
The initial state of the three pairs is assumed to be
\begin{equation}
|\psi_0\rangle = \frac1{\sqrt{8}}\left( |00\rangle + |11\rangle\right)^3
\end{equation}
Two sub ensembles exist after the measurement, one with the first two bits in the zero state, the other its complement.
\begin{equation}
|\psi_1\rangle = \frac1{\sqrt{8}}|00\rangle \otimes ( |0\rangle + |1\rangle ) + \frac1{\sqrt{8}}( |01\rangle + |10\rangle + |11\rangle ) \otimes ( |0\rangle + |1\rangle )
\end{equation}
With the 'insulation rule' in effect, we will have the final projected states
\begin{eqnarray}
|\bar\psi_B\rangle = \frac12|00\rangle \otimes ( \alpha_B|0\rangle + \beta_B|1\rangle ) + \frac1{\sqrt{8}}( |01\rangle + |10\rangle + |11\rangle ) \otimes ( |0\rangle + |1\rangle )\\
|\bar\psi_N\rangle = \frac12|00\rangle \otimes ( \alpha_N|0\rangle + \beta_N|1\rangle ) + \frac1{\sqrt{8}}( |01\rangle + |10\rangle + |11\rangle ) \otimes ( |0\rangle + |1\rangle )\\
|\bar\psi_-\rangle = \frac12|00\rangle \otimes ( \alpha_-|0\rangle + \beta_-|1\rangle ) + \frac1{\sqrt{8}}( |01\rangle + |10\rangle + |11\rangle ) \otimes ( |0\rangle + |1\rangle )\\
|\bar\psi_{-N}\rangle = \frac12|00\rangle \otimes ( \alpha_{-N}|0\rangle + \beta_{-N}|1\rangle ) + \frac1{\sqrt{8}}( |01\rangle + |10\rangle + |11\rangle ) \otimes ( |0\rangle + |1\rangle )
\end{eqnarray}
While in general this 'third qubit may be entangled with other variables, the premise that the weight in the ensemble of the $|00\rangle$ conditional projection state, as well as the relative amplitudes in the other branches of the wave-function, remains independent of the projection we choose. The final state is the normal mixture and partition of these four projections, or two in teh case of a classical channel. While the tourists are not suppressed relative to the thermal bath, they do not become more frequent either. While time machine effects may still back-propagate, they at least do not 'side-propagate' in this model. to be observed in branches of the wave function where there are no time machines.
Suppose we select against the $|000\rangle$ state. Two methods of renormalization are available, depending on if we wish the distribution of sub ensembles to skew the value of the 'power' qubits. Without partitioning we have
\begin{equation}
|\bar\psi\rangle_{bp} = \frac1{\sqrt{7}}|001\rangle + \frac1{\sqrt{7}}\left(|01\rangle+|10\rangle+|11\rangle\right) \otimes \left( |0\rangle + |1\rangle \right)
\end{equation}
and with partitioning the the projected sub-ensemble we have
\begin{equation}
|\bar\psi\rangle_{sp} = \frac12|001\rangle + \frac1{\sqrt{8}}\left(|01\rangle+|10\rangle+|11\rangle\right) \otimes \left( |0\rangle + |1\rangle \right)
\end{equation}
In the former case, the probability of conditional post-selection depends on the normalization factor.
This would give an overall bias against 'tourists', $p_{00}=1/7$, compared with the 'normal' value of $1/4$ for the 'insulated' model, or for no post-selection.\\
\section{Computation}
The new measurement effects allow for some powerful analog computations, but the pack-propagation phenomenon allows for the collapse of the verify-calculate divide in the computational hierarchy directly.
It has been demonstrated independently by several authors that post-selected computation collapses $NP$ to $P$, but it is further the case that this holds for each power individually. For example, verification in quadratic time leads to solutions in quadratic time, lists may be sorted in linear time, and searches conducted instantly.
Another point is that it holds $\emph{for every oracle}$ we might choose to introduce.
A post-selection assisted computer can physically reverse a calculation process by aligning its outcome with the selection condition.
As a result it can invert any oracle function, including many very sparse functions that are normally used to divide classes.
A blind password attack can normally only be done in exponential time for completely random keys. Each attempt is modeled as an oracle query, returning yes if the password is correct, and no otherwise.
While not normally considered an interesting problem in complexity theory, it is one of great practical importance.
With post-selection the inverse can be uncovered in effectively linear time. The only reason that the time would not be constant time is due to the amplification of errors.
An important result by \cite{scott} shows that post-selection efficiently solvesl $PP$ problems in polynomial time almost by definition. $PP$ can be expressed as problems for which a solution can possibly by verified in polynomial time using the best case of a probabilistic algorithm with access to coin flips. There are other definitions of $PP$, but this one makes the connection with post-selection the most clear. The method is we take an ordinary probabilistic algorithm and select along the case that it succeeds. One could call such a method a 'squeezed' probabilistic algorithm, since the normal distribution of the coin flips that the algorithm uses is squeezed onto those combinations of flips that result in an accept. It would be tempting to treat the system as an oracle for $PP$, but multiple queries can present a problem.
The primary difficulty in extending the individual results is that treating the squeezed probabilistic algorithms as oracles does not take into account interaction between projections. In simulating an interactive proof we may use a post-selected probabilistic algorithm to generate the proof to be reviewed, and attempt to use another such algorithm to find errors in the proof,
but adversarial systems will suffer from back-propagation that will cause them to mutually amplify each other's errors, possibly making the whole program unreliable.
In order to help insulate against back-propagation, we can employ amnesia type circuits to 'initialize' qubits into a Hadamard state or a classically random state if the channel is classical. Some back-propagation will still occur, but it is due to any singular points that the amnesia circuit has, rather than the action of later gates. In the classical limit, the unproven proof circuit can be used for this purpose, as a 'true random' source. It would be more effective since it has no singular points. \\
One consequence of the ability to efficiently measure small probabilities is that post-selection can efficiently solve some $\#P$ problems. If the number of solutions to a given NP problem is polynomial, we can find all of them in polynomial time, by using a fuse circuit whose success chance is sufficiently small, and iterating through a probabilistic search that tries random strings that have not been tried before until the fuse blows. However these results depend on access to such a low probability variable, with acts as an oracle of sorts. If the fuse is composed of coin flips, it should require polynomial time at least as large as the number of solutions, typically one degree higher.
Another example of this is searching. Given a long unsorted list, we can construct a fuse with a probability of less than $1/n$ of activating, using a logarithmic number of operations. A random guess circuit can move us to a particular random element in a similar number of moves. This suggests that we may find a single element in a list, or verify its absence in log time. These result may seem impressive, but remember, exact post-selection allows unlimited entropy reduction.
Let us explore the search problem with noisy post-selection.
Such hyper searching can only be done reliably if the noise ratio is small enough, that is
\begin{equation}
\Omega > n
\end{equation}
Given a probabilistic algorithm, we can use post-selection to improve its success rate, but only up to a certain factor.
\begin{equation}
\bar{P}_s = \frac{\Omega P_s}{\Omega P_s + 1 - P_s }
\end{equation}
Beyond that, we must simply execute the algorithm multiple times. For any fixed finite $\Omega$ we can still only solve problems in BQP in polynomial time.
However, using multiple independent periodic channels we can reduce the probability further. For two noisy selection circuits we have a sort of composition relation.
\begin{eqnarray}
\Omega_{12} \approx \Omega_1 \cdot \Omega_2 \\
\Omega_n \approx \Omega_0^n
\end{eqnarray}
If a noisy post-selector can perform state projections at a rate linear in time, this suggests that the error rate of a probabilistic algorithm entangled with a series of projections is
\begin{equation}
\epsilon_\Omega \approx \frac12(e^{\alpha t}P_0/(1-P_0) + 1)^{-1}
\end{equation}
If we compare this to the Chernoff bound error rate where each test requires a time $\gamma$,
\begin{equation}
\epsilon_{chrn} \approx \exp\left[ -2\left(p-\frac12\right)^2 \gamma t\right]
\end{equation}
For our probabilistic search, the item is found in a single attempt with probabilities
\begin{eqnarray}
P_0 = 1/n\\
p = \frac12 + \frac1{2n}
\end{eqnarray}
for the two methods respectively. Then the two error rates are
\begin{eqnarray}
\epsilon_\Omega \approx \frac12(1-1/n) e^{-\alpha t}\\
\epsilon_\gamma \approx e^{-n^2\gamma t /2}
\end{eqnarray}
Repeated projection acts like a heat engine, reducing the entropy of a random list element from $\ln n$ to close to zero, but increasing the entropy of the environment.
If the time is polynomial in $n$, then the skew factor from a constant rate of projection will be exponential in that polynomial.
\begin{equation}
\Omega_{poly} \approx e^{p(n)}
\end{equation}
This also happens to be the minimum nonzero probability bound for PP, or a fuse circuit using a polynomial number of gates.
Classical probabilistic algorithms should have an entropy bound since classical coin flipping is not reversible.
State projection is not a reversible process, and should introduce similar bounds for quantum computers that utilize it.
We can characterize probabilistic algorithms not only by time and space limitations but also by entropy production.
In many ways post-selected computing acts like adiabatic computing given an exponentially long cooling time.
\section{Remarks}
The variety of strange behaviors in quantum theory deformed by projections is at least as large as for ordinary quantum mechanics. If physics truly makes progress through paradoxes, then this is fertile ground. With projection the theory space available to a QFT becomes much larger, but also subject to more experimental constraints. The modeling of collapse toward a final state as a postselected computations has drawn considerable interest\cite{presk}. The creation of secondary channels and other back-propagation effects indicate that the exotic behaviors are not necessarily confined to the black hole interior or high curvature regions in the case of the final state model of black holes. However, the dilution of the skew as one moves away from the projection due to increasing entanglement provides some shielding effect. The diminishing returns of noisy postselection for a heat engine as one attempts to select states of smaller and smaller amplitude would indicate a partial information loss behavior as a compromise that limits the severity of exotic effects. The tourist suppression effect in the partial loss model could provide a Boltzmann like renormalizing factor for microscopic black hole fluctuations. This would also lead to loss of coherence in the UV, affecting the short distance behavior of gravity, and perhaps allowing a renormalization and indicating a minimum temperature for quantum gravity.
|
\section{Introduction}
Throughout the article we work over an algebraically closed field $k$ that is endowed with the trivial norm. In \cite{Hassett_weightedmoduli} Hassett introduces a class of modular compactifications $\overline{\mathcal{M}}_{g,\mathcal{A}}$ of the moduli space $\mathcal{M}_{g,n}$ of smooth curves with $n$ marked points parametrized by an \emph{input datum} $(g,\mathcal{A})$ consisting of a non-negative integer $g$ together with a collection $\mathcal{A}=(a_1,\ldots,a_n)$ of \emph{weights} $a_i\in\mathbb{Q}\cap(0,1]$ such that
\begin{equation*}
2g-2+a_1+\dots+a_n>0 \ .
\end{equation*}
The moduli space $\overline{\mathcal{M}}_{g,\mathcal{A}}$ parametrizes curves $(C,p_1,\ldots,p_n)$ with $n$ marked non-singular points on $C$ that are \emph{stable of type $(g,\mathcal{A})$}, i.e. nodal curves $(C,p_1,\ldots,p_n)$ with $n$ marked non-singular points that fulfill the following two conditions:
\begin{enumerate}
\item\label{item_amplecanonical} The twisted canonical divisor $K_C+a_1p_1+\ldots+a_np_n$ is ample.
\item A subset $p_{i_1},\ldots, p_{i_k}$ of the marked points is allowed to coincide only if the inequality $a_{i_1}+\ldots+a_{i_k}\leq 1$ holds.
\end{enumerate}
In the case $(a_1,\ldots,a_n)=(1,\ldots,1)$ this condition is nothing but the traditional notion of an $n$-marked stable curve and so the compactification $\overline{\mathcal{M}}_{g,\mathcal{A}}$ is exactly the well-known Deligne-Knudsen-Mumford compactification $\overline{\mathcal{M}}_{g,n}$ of $\mathcal{M}_{g,n}$ introduced in \cite{DeligneMumford_moduliofcurves} and \cite{Knudsen_projectivityII}.
In \cite[Theorem 2.1]{Hassett_weightedmoduli} Hassett shows that the moduli spaces $\overline{\mathcal{M}}_{g,\mathcal{A}}$ are connected Deligne-Mumford stacks that are proper and smooth over $\Spec \mathbb{Z}$ and whose coarse moduli spaces $\overline{M}_{g,\mathcal{A}}$ are projective over $\Spec\mathbb{Z}$. Denote by $\mathcal{M}_{g,\mathcal{A}}$ the open locus of smooth curves in $\overline{\mathcal{M}}_{g,\mathcal{A}}$. The following Theorem \ref{thm_SNC} is well-known to the experts, but, to the best of the author's knowledge, it has not appeared in the literature, so far. A discussion of its proof can be found in Section \ref{section_alternativeSNC}.
\begin{theorem}\label{thm_SNC}
The complement of $\mathcal{M}_{g,\mathcal{A}}$ in $\overline{\mathcal{M}}_{g,\mathcal{A}}$ is a divisor with (stack-theoretically) normal crossings.
\end{theorem}
So the open immersion $\mathcal{M}_{g,\mathcal{A}}\hookrightarrow \overline{\mathcal{M}}_{g,\mathcal{A}}$ has the structure of a toroidal embedding (see \cite{KKMSD_toroidal}). By the work of \cite{Thuillier_toroidal} and \cite{AbramovichCaporasoPayne_tropicalmoduli}, associated to this datum there is a natural strong deformation retraction $\mathbf{p}$ from the non-Archimedean analytic space $\overline{M}_{g,\mathcal{A}}^{an}$ associated to the coarse moduli space $\overline{M}_{g,\mathcal{A}}$ onto a closed subset $\mathfrak{S}(\overline{\mathcal{M}}_{g,\mathcal{A}})$ of $\overline{M}_{g,\mathcal{A}}^{an}$, called the \emph{skeleton} of $\overline{\mathcal{M}}_{g,\mathcal{A}}$, and this skeleton naturally carries the structure of an extended generalized cone complex in the sense of \cite[Section 2]{AbramovichCaporasoPayne_tropicalmoduli}.
In this article we define a notion of \emph{stability of type $(g,\mathcal{A})$} for tropical curves $\Gamma$ by imitating condition \eqref{item_amplecanonical}. Moreover we construct a set-theoretic moduli space $M_{g,\mathcal{A}}^{trop}$ parametrizing isomorphism classes of tropical curves that are stable of type $(g,\mathcal{A})$. Its natural extension $\overline{M}_{g,\mathcal{A}}^{trop}$ admits an interpretation as a set-theoretic moduli space of extended tropical curves that are stable of type $(g,\mathcal{A})$. The tropical moduli space $\overline{M}_{g,\mathcal{A}}^{trop}$ naturally carries the structure of a generalized extended cone complex.
Moreover, following \cite{BakerPayneRabinoff_nonarchtrop}, \cite[Section 2.2.3]{Viviani_tropcompTorelli}, and \cite[Section 1.1]{AbramovichCaporasoPayne_tropicalmoduli}, there is a naive set-theoretic \emph{tropicalization map}
\begin{equation*}
\trop_{g,\mathcal{A}}\mathrel{\mathop:}\overline{M}_{g,\mathcal{A}}^{an}\longrightarrow \overline{M}_{g,\mathcal{A}}^{trop}
\end{equation*}
from the non-Archimedean analytic space $\overline{M}_{g,\mathcal{A}}^{an}$ onto the extended tropical moduli space $\overline{M}_{g,\mathcal{A}}^{trop}$ defined as follows:
A point $x$ in $\overline{M}_{g,\mathcal{A}}^{an}$ can be represented by a morphism $\Spec K\rightarrow \overline{\mathcal{M}}_{g,\mathcal{A}}$ for a non-Archimedean field extension $K$ of $k$. Since $\overline{\mathcal{M}}_{g,\mathcal{A}}$ is a proper algebraic stack over $k$, the valuative criterion for properness implies that after a finite base change $K'\vert K$ this morphism extends uniquely to a morphism $\Spec R'\rightarrow\overline{\mathcal{M}}_{g,\mathcal{A}}$, where $R'$ denotes the valuation ring of $K'$. This datum is equivalent to a curve $\mathcal{C}\rightarrow \Spec R'$ that is stable of type $(g,\mathcal{A})$. Denote by $G_x$ the weighted dual graph of the special fiber $\mathcal{C}_s$ of $\mathcal{C}$; it is stable of type $(g,\mathcal{A})$ by Proposition \ref{prop_Astability}. At a node $p_e$ of $\mathcal{C}_s$ corresponding to an edge $e$ of $G_x$ the curve is defined by $xy=f_e$ in formal coordinates, where $f_e\in R'$. Endowing an edge $e$ with the length $l(e)=\val(f_e)$, where $\val$ denotes the valuation on $R'$, defines a tropical curve $\Gamma_x$ and we set
\begin{equation*}
\trop_{g,\mathcal{A}}(x)=[\Gamma_x]\in\overline{M}_{g,\mathcal{A}}^{trop}.
\end{equation*}
The main result of this article can now be stated as follows:
\newpage
\begin{theorem}\label{thm_modulartrop}
There is a natural isomorphism $J_{g,\mathcal{A}}\mathrel{\mathop:}\overline{M}_{g,\mathcal{A}}^{trop}\xrightarrow{\sim}\mathfrak{S}(\overline{\mathcal{M}}_{g,\mathcal{A}})$ of extended generalized cone complexes such that the diagram
\begin{center}\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=2em,column sep=3em,minimum width=2em]
{
& \overline{M}_{g,\mathcal{A}}^{an} & \\
\mathfrak{S}(\overline{\mathcal{M}}_{g,\mathcal{A}}) & & \overline{M}_{g,\mathcal{A}}^{trop} \\
};
\path[-stealth]
(m-1-2) edge node [above left] {$\mathbf{p}$} (m-2-1)
edge node [above right] {$\trop_{g,\mathcal{A}}$} (m-2-3)
(m-2-3) edge node [below] {$J_{g,\mathcal{A}}$} node [above] {$\sim$} (m-2-1);
\end{tikzpicture}\end{center}
commutes.
\end{theorem}
In the case of $\mathcal{A}=(1,\ldots,1)$ Theorem \ref{thm_modulartrop} is the same as \cite[Theorem 1.2.1]{AbramovichCaporasoPayne_tropicalmoduli}. Its proof, an adaption of the one in \cite{AbramovichCaporasoPayne_tropicalmoduli}, can be found in Section \ref{section_proof}. It relies on a careful analysis of the stratfication of $\overline{\mathcal{M}}_{g,\mathcal{A}}$ by dual graphs, which is undertaken in Section \ref{section_dualgraphs&boundary}. In the case of $\overline{\mathcal{M}}_{g,n}$ this analysis can be found in \cite[Section XII.10]{ArbarelloCornalbaGriffiths_moduliofcurves}. Similar proofs come up in \cite{CavalieriMarkwigRanganathan_admissiblecovers} and in \cite{Ranganathan_ratcurvesnonArch} treating moduli spaces of admissible covers of the projective line and of rational logarithmic stable maps into a toric variety respectively.
From Theorem \ref{thm_modulartrop} we immediately obtain the following:
\begin{corollary}\label{cor_tropwelldefcont}
The naive set-theoretic tropicalization map $\trop_{g,\mathcal{A}}\mathrel{\mathop:}\overline{M}_{g,\mathcal{A}}^{an}\rightarrow \overline{M}_{g,\mathcal{A}}^{trop}$ is well-defined and continuous.
\end{corollary}
The strong deformation retraction $\mathbf{p}$ restricts to a strong deformation retraction
\begin{equation*}
M_{g,\mathcal{A}}^{an}\longrightarrow\mathfrak{S}(\mathcal{M}_{g,\mathcal{A}}) \ ,
\end{equation*}
where $\mathfrak{S}(\mathcal{M}_{g,\mathcal{A}})=M_{g,\mathcal{A}}^{an}\cap\frak{S}(\overline{\mathcal{M}}_{g,\mathcal{A}})$ is the \emph{non-Archimedean skeleton} of $\mathcal{M}_{g,\mathcal{A}}$. Theorem \ref{thm_modulartrop} therefore immediately implies the following:
\begin{corollary}\label{cor_modulartrop}
The isomorphism $J_{g,\mathcal{A}}$ induces an isomorphism $M_{g,\mathcal{A}}^{trop}\xrightarrow{\sim}\mathfrak{S}(\mathcal{M}_{g,\mathcal{A}})$ of generalized cone complexes such that the diagram
\begin{center}\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=2em,column sep=3em,minimum width=2em]
{
& M_{g,\mathcal{A}}^{an} & \\
\mathfrak{S}(\mathcal{M}_{g,\mathcal{A}}) & & M_{g,\mathcal{A}}^{trop} \\
};
\path[-stealth]
(m-1-2) edge node [above left] {$\mathbf{p}$} (m-2-1)
edge node [above right] {$\trop_{g,\mathcal{A}}$} (m-2-3)
(m-2-3) edge node [below] {$J_{g,\mathcal{A}}$} node [above] {$\sim$} (m-2-1);
\end{tikzpicture}\end{center}
commutes.
\end{corollary}
It is a natural question whether the tautological maps of the moduli spaces $\overline{\mathcal{M}}_{g,\mathcal{A}}$, defined in analogy with \cite[Section 3]{Knudsen_projectivityII}, have tropical analogues and commute with tropicalization maps $\trop_{g,\mathcal{A}}$. We give a positive answer to this Question in Section \ref{section_troptaut}. In particular we are going to focus on the forgetful map from \cite[Theorem 4.3]{Hassett_weightedmoduli} and the gluing and clutching maps as defined in \cite[Proposition 2.1.1]{BayerManin_weightedstablemaps} for the moduli spaces of weighted stable maps.
Moreover, for two weight data $(g,\mathcal{A})$ and $(g,\mathcal{B})$ with $a_i\geq b_i$ for all $1\leq i\leq n$ Hassett constructs in \cite[Section 4]{Hassett_weightedmoduli} a proper birational \emph{reduction morphism}
\begin{equation*}
\rho_{\mathcal{A},\mathcal{B}}\mathrel{\mathop:}\overline{\mathcal{M}}_{g,\mathcal{A}}\rightarrow\overline{\mathcal{M}}_{g,\mathcal{B}}
\end{equation*}
that contracts those boundary divisors that parametrize $\mathcal{A}$-stable curves that are not $\mathcal{B}$-stable. These reduction morphisms play a central role in the birational geometry of the Deligne-Knudsen-Mumford moduli spaces $\overline{\mathcal{M}}_{g,n}$, since they form contractions onto certain log-canonical models of $\overline{\mathcal{M}}_{g,n}$. For these developments we refer the reader to \cite{Fedorchuk_weightedmoduli} and \cite{Moon_logcanonicalmodels} as well as to the survey \cite{FedorchukSmyth_alternatemoduli}. In Section \ref{section_forgetful&reduction} we define tropical analogues of these morphisms that commute with the tropicalization map.
In Section \ref{section_wallsandchambers} we investigate how the tropical moduli spaces $\overline{M}_{g,\mathcal{A}}^{\trop}$ vary in the weights $\mathcal{A}$ based on how the moduli spaces $\overline{\mathcal{M}}_{g,\mathcal{A}}$ vary in the weights $\mathcal{A}$.
In Section \ref{section_LosevManin} we finish with the classical example of Losev-Manin spaces, as defined in \cite{LosevManin_newmoduli}.
During the work on this project the author learned about the article \cite{CavalieriHampeMarkwigRanganathan_weightedmoduli}, which contains the $g=0$ case of Theorem \ref{thm_modulartrop} in \cite[Theorem 3.15]{CavalieriHampeMarkwigRanganathan_weightedmoduli}. The main goal of \cite{CavalieriHampeMarkwigRanganathan_weightedmoduli}, however, is to treat the tropicalization of $\overline{M}_{0,\mathcal{A}}$ from the point of view of geometric tropicalization, as developed in \cite{HackingKeelTevelev_modulidelPezzo} and further studied in \cite{Cueto_geometrictropicalization}. For this the authors of \cite{CavalieriHampeMarkwigRanganathan_weightedmoduli} embed $\overline{M}_{0,\mathcal{A}}$ into a toric variety $X$ and study the tropicalization $\Trop_X(\overline{M}_{0,\mathcal{A}})$ of $\overline{M}_{0,\mathcal{A}}$ with respect to $X$. By \cite[Theorem 1.1 and 1.2]{Ulirsch_functroplogsch} as well as Theorem \ref{thm_modulartrop} there is a natural continuous and surjective map
\begin{equation*}
\overline{M}_{0,\mathcal{A}}^{trop}\longrightarrow \Trop_X(\overline{M}_{0,\mathcal{A}})
\end{equation*}
that is, in general, not injective, as can be seen in \cite[Figure 4]{CavalieriHampeMarkwigRanganathan_weightedmoduli}. The main result of \cite{CavalieriHampeMarkwigRanganathan_weightedmoduli} is a characterization of those weights $\mathcal{A}$ for which this map is a bjiection, i.e. for which the geometric tropicalization of $\overline{M}_{0,\mathcal{A}}$ faithfully represents the full tropical moduli space $\overline{M}_{0,\mathcal{A}}^{trop}$.
\subsection{Acknowledgements}
The author would like to express his gratitude to Dan Abra\-movich for his constant support and encouragement. Thanks are also due to Renzo Cavalieri, Noah Giansiracusa, Simon Hampe, Diane MacLagan, Steffen Marcus, and Dhruv Ranganathan for several discussions related to this project. Finally, we would also like to thank the anonymous referee for many helpful remarks and suggestions.
\section{Weighted stable tropical curves and their moduli}
In this section we define tropical versions of Hassett's moduli spaces $\overline{\mathcal{M}}_{g,\mathcal{A}}$ of weighted stable curves. Our treatment of tropical moduli spaces in this section is strongly inspired by \cite[Section 3]{Caporaso_tropicalmoduli} and \cite[Section 4]{AbramovichCaporasoPayne_tropicalmoduli}.
\subsection{$\mathcal{A}$-stability for weighted graphs}
Recall (e.g. \cite[Section 3.2]{AbramovichCaporasoPayne_tropicalmoduli}, \cite[Definition 2.1]{Caporaso_tropicalmoduli}, or \cite[Definition 2.3 and 2.4]{Manin_quantumbook}) that a \emph{weighted graph $G$ with $n$ marked legs} is a sextuple
\begin{equation*}
\big(V(G),F(G),r,i,m,h\big)
\end{equation*}
consisting of:
\begin{itemize}
\item a finite set of vertices $V(G)$,
\item a finite set of \emph{flags} $F(G)$ together with a \emph{root map} $r\mathrel{\mathop:}F(G)\rightarrow V(G)$ associating to a flag of $G$ the vertex it emanates from,
\item an involution $i\mathrel{\mathop:}F(G)\rightarrow F(G)$ inducing a decomposition of $F(G)$ into the set $L(G)$ of fixed points of $i$, called the \emph{legs} of $G$, and a finite union of pairs of points, called the \emph{edges} of $G$,
\item a marking of $L(G)$, i.e. a bijection $m\mathrel{\mathop:}\{1,\ldots,n\}\xrightarrow{\sim} L(G)$ given by $L(G)=\{l_1,\ldots,l_n\}$, and
\item a \emph{weight function} $h\mathrel{\mathop:}V(G)\rightarrow \mathbb{N}$ associating to every vertex a nonnegative integer $h(v)$, referred to as the \emph{genus} of $v$.
\end{itemize}
We write $E(G)$ for the set of edges of $G$. Whenever there is no risk of confusion we sometimes drop the reference to $G$ from our notation and, for example, denote the set of vertices of $G$ by $V$ instead of $V(G)$.
The \emph{genus} $g(G)$ of $G$ is defined to be
\begin{equation}\label{eq_genusgraph}
g(G)=b_1(G)+\sum_{v\in V}h(v) \ ,
\end{equation}
where $b_1(G)=\dim_\mathbb{Q} H^1(G,\mathbb{Q})=\#E(G)-\#V(G)+1$ is the first Betti number of $G$. An \emph{automorphism} $\gamma\in\Aut(G)$ consists of bijective maps $V(G)\xrightarrow{\sim} V(G)$ and $F(G)\xrightarrow{\sim} F(G)$ making the obvious diagrams commute. So, in particular, we have $g\big(\gamma(v)\big)=g(v)$ for all vertices $v\in V(G)$, the induced map $L(G)\xrightarrow{\sim}L(G)$ is the identity, and $\gamma$ preserves incidences between edge and vertices as well as legs and vertices.
For a vertex $v\in V$, we write $L(v)$ for the set of marked legs emanating from $v$ and $\vert v\vert_E$ for the number of flags emanating from $v$ that are contained in an edge, i.e. the number of edges starting at $v$ counting loops with multiplicity two. Moreover, given an \emph{input datum} $(g,\mathcal{A})$, consisting of a non-negative integer $g$ together with a collection $\mathcal{A}=(a_1,\ldots,a_n)$ of numbers $a_i\in\mathbb{Q}\cap(0,1]$ such that
\begin{equation*}
2g-2+a_1+\dots+a_n>0 \ ,
\end{equation*}
we set $\vert v\vert_\mathcal{A}=\sum_{l_i\in L(v)}a_i$ for a vertex $v\in V$.
\begin{definition}\label{def_weightedstablegraph}
Let $(g,\mathcal{A})$ be an input datum. A weighted graph $G$ with $n$ legs is said to be \emph{stable of type $(g,\mathcal{A})$}, if it has genus $g$ and for all vertices $v\in V(G)$ we have:
\begin{equation*}
2h(v)-2+\vert v\vert_E+\vert v\vert_\mathcal{A}> 0 \ .
\end{equation*}
If $G$ is stable of type $(g,\mathcal{A})$ with $\mathcal{A}=(1,\ldots,1)$ we simply call it \emph{stable}.
\end{definition}
A \emph{weighted graph contraction} $\pi\mathrel{\mathop:}G\rightarrow G'$ is a composition of edge contractions that preserves the weight function in the sense that
\begin{equation*}
h'(v')=g\big(\pi^{-1}(v')\big)
\end{equation*}
for all vertices $v'\in V'=V(G')$, where we consider $\pi^{-1}(v')$ as a subgraph of $G$. Observe that, if $G$ is stable of type $(g,\mathcal{A})$, the contracted graph $G'$ is stable of type $(g,\mathcal{A})$ as well.
\begin{definition}
Let $(g,\mathcal{A})$ be an input datum. The category of $\mathcal{G}_{g,\mathcal{A}}$ of weighted graphs that are stable of type $(g,\mathcal{A})$ is defined as follows:
\begin{itemize}
\item Its objects are isomorphism classes of weighted graphs $G$ that are stable of type $(g,\mathcal{A})$.
\item The morphisms in $\mathcal{G}_{g,\mathcal{A}}$ are generated by weighted graph contractions $\pi\mathrel{\mathop:}G\rightarrow G'$ and automorphisms $\Aut(G)$ of every graph $G$.
\end{itemize}
\end{definition}
Note hereby that the set of isomorphisms between two weighted graphs $G_1$ and $G_2$ is either empty or a natural $\Aut(G_i)$-torsor for both $i=1,2$. In particular, a choice of an isomorphism $G_1\simeq G_2$ induces natural identifications between their automorphism groups $\Aut(G_i)$ and their weighted graph contractions $\pi_i\mathrel{\mathop:}G_i\rightarrow G_i'$.
\begin{remarks}
\begin{enumerate}[(i)]
\item The datum of the weight function $h\mathrel{\mathop:}V\rightarrow\mathbb{N}$ should be thought of as having $h(v)$ infinitely small loops at each vertex $v$. This, in particular, explains why the genus of $G$ is defined as in \eqref{eq_genusgraph}.
\item Expanding on \cite{BakerNorine_RiemannRoch} we can define a $\mathbb{Q}$-divisor on $G$ as a formal sum
$\sum_{v\in V}a_v(v)$ with coefficients $a_v\in\mathbb{Q}$. In this language a weighted graph $G$ is stable of type $(g,\mathcal{A})$ if and only if it has genus $g$ and the coefficients of the \emph{twisted canonical divisor}
\begin{equation*}
K_{G,\mathcal{A}}=\sum_{v\in V}\big(2h(v)-2+\vert v\vert_E+\vert v\vert_\mathcal{A}\big)(v)
\end{equation*}
on $G$ are strictly positive.
\item A theory similar to the one developed in this section has already been developed in \cite[Section 5]{BayerManin_weightedstablemaps}. We, in particular, refer the reader to \cite[Definition 5.1.6]{BayerManin_weightedstablemaps} which is equivalent to Definition \ref{def_weightedstablegraph}. Nevertheless note that the notion of a contraction in \cite[Definition 5.1.5]{BayerManin_weightedstablemaps} is different from ours, since the authors of \cite{BayerManin_weightedstablemaps} also allow legs to be merged.
\end{enumerate}
\end{remarks}
\subsection{Tropical moduli spaces}
Following \cite[Definition 2.2]{Caporaso_tropicalmoduli} and \cite[Section 4.1]{AbramovichCaporasoPayne_tropicalmoduli} a \emph{tropical curve $\Gamma$ of genus $g$ with $n$ legs} consists of a weighted graph $G$ of genus $g$ with $n$ legs together with a \emph{length function} $l\mathrel{\mathop:}E(G)\rightarrow\mathbb{R}_{> 0}$. By identifying an edge $e$ with an interval of length $l(e)$ we can associate to $\Gamma$ a metric space $\vert\Gamma\vert$, which is called the \emph{geometric realization} of $\Gamma$.
If we allow the length function $l$ to attain values in $\overline{\mathbb{R}}_{>0}=\mathbb{R}_{>0}\sqcup\{\infty\}$, we say that $\Gamma$ is an \emph{extended tropical curve} of genus $g$ with $n$ legs. Its \emph{geometric realization} has the structure of an extended metric space by identifying an edge $e$ with $l(e)=\infty$ with the double infinite line \begin{equation*}
\big(\mathbb{R}_{\geq0}\sqcup\{\infty\}\big)\cup \big(\mathbb{R}_{\leq0}\sqcup\{-\infty\}\big) \ ,
\end{equation*}
where the two points $\infty$ and $-\infty$ are identified.
\begin{center}\begin{tikzpicture}
\fill (0,0) circle (0.08 cm);
\fill (-2,0) circle (0.08 cm);
\fill (2,0) circle (0.08 cm);
\draw (-2,0) -- (-1,0);
\draw (1,0) -- (2,0);
\draw [dashed] (-1,0) -- (0,0);
\draw [dashed] (1,0) -- (0,0);
\node at (0,0.5) {$\pm\infty$};
\node at (3.5,-0.25) {$l(e)=\infty$};
\end{tikzpicture}\end{center}
We denote the category of rational polyhedral cone complexes as defined in \cite[Section 3.2]{Ulirsch_functroplogsch} by $\mathbf{RPCC}$.
\begin{definition}
Let $(g,\mathcal{A})$ be an input datum. We define a natural contravariant functor
\begin{equation*}
\Sigma\mathrel{\mathop:}\mathcal{G}_{g,\mathcal{A}}\longrightarrow \mathbf{RPCC}
\end{equation*}
as follows:
\begin{itemize}
\item Associated to an isomorphism class of a weighted graph $G$ that is stable of type $(g,\mathcal{A})$ is the rational polyhedral cone $\sigma_G=\mathbb{R}_{\geq 0}^{E(G)}$.
\item A weighted edge contraction $\pi\mathrel{\mathop:}G\rightarrow G'$ induces the natural embedding $i_\pi\mathrel{\mathop:}\sigma_{G'}\hookrightarrow\sigma_G$ of a face of $\sigma_G$ .
\item An automorphism of $G$ induces an automorphism of $\sigma_G$.
\end{itemize}
\end{definition}
Similarly there is also a natural functor $\overline{\Sigma}$ from $\mathcal{G}_{g,\mathcal{A}}$ into the category of extended rational polyhedral cone complexes that is given by $G\mapsto \overline{\sigma}_G=\overline{\mathbb{R}}_{\geq 0}^{E(G)}$. We denote the image of $\mathcal{G}_{g,\mathcal{A}}$ in $\mathbf{RPCC}$ simply by $\Sigma_{g,\mathcal{A}}$ and the category of its extensions by $\overline{\Sigma}_{g,\mathcal{A}}$.
\begin{definition}
The \emph{moduli space $M_{g,\mathcal{A}}^{trop}$ of $\mathcal{A}$-stable tropical curves of genus $g$ with $n$ marked legs} is defined to be the colimit
\begin{equation*}
M_{g,\mathcal{A}}^{trop}=\lim_{\longrightarrow} \sigma_G \ ,
\end{equation*}
taken over $(\mathcal{G}_{g,\mathcal{A}})^{op}$. The \emph{moduli space $\overline{M}_{g,\mathcal{A}}^{trop}$ of extended $\mathcal{A}$-stable tropical curves of genus $g$ with $n$ marked legs} is defined to be the colimit
\begin{equation*}
\overline{M}_{g,\mathcal{A}}^{trop}=\lim_{\longrightarrow} \overline{\sigma}_G
\end{equation*}
taken over $(\mathcal{G}_{g,\mathcal{A}})^{op}$.
\end{definition}
The tropical moduli spaces $M_{g,\mathcal{A}}^{trop}$ and $\overline{M}_{g,\mathcal{A}}^{trop}$ naturally carry the structure of a generalized cone complex and an extended generalized cone complex in the sense of \cite[Section 2]{AbramovichCaporasoPayne_tropicalmoduli} respectively.
Fix a weighted graph $G$ of genus $g$ with $n$ legs. We write $\sigma_G^\circ$ for the open cone $\mathbb{R}_{>0}^{E(G)}$. The set of tropical curves $\Gamma$ with underlying weighted graph isomorphic to $G$ can be parametrized by the quotient
\begin{equation*}
M_G^{trop}=\sigma_G^\circ/\Aut(G) \ ,
\end{equation*}
the \emph{moduli space of tropical curve of combinatorial type $G$}. Similarly, if we replace $\sigma_G^\circ$ by the extended open cone $\overline{\sigma}_G^\circ=\overline{\mathbb{R}}_{>0}^{E(G)}$, we obtain the set-theoretic \emph{moduli space
\begin{equation*}
\overline{M}_G^{trop}=\overline{\sigma}_G^\circ/\Aut(G)
\end{equation*}
of extended tropical curves of combinatorial type $G$}.
From this point of view one can interpret the quotients $\sigma_G/\Aut(G)$ (or $\overline{\sigma}_G/\Aut(G)$) as moduli spaces of tropical curves (or extended tropical curves), where we allow the edges to have zero length. For a weighted edge contraction $\pi\mathrel{\mathop:}G\rightarrow G'$ the faces $i_\pi\mathrel{\mathop:}\sigma_{G'}\hookrightarrow\sigma_G$ and $\overline{i}_\pi\mathrel{\mathop:}\overline{\sigma}_{G'}\hookrightarrow\overline{\sigma}_G$ parametrize those tropical curves, or extended tropical curves respectively, that have edge length zero for the edges that are collapsed by $\pi$.
As an immediate consequence of the construction we have:
\begin{proposition}
There are decompositions
\begin{equation*}
M_{g,\mathcal{A}}^{trop}=\bigsqcup_{G\in\mathcal{G}_{g,\mathcal{A}}}M_G^{trop}=\bigsqcup_{G\in\mathcal{G}_{g,\mathcal{A}}}\sigma^\circ_G/\Aut(G)
\end{equation*}
as well as
\begin{equation*}
\overline{M}_{g,\mathcal{A}}^{trop}=\bigsqcup_{G\in\mathcal{G}_{g,\mathcal{A}}}\overline{M}_G^{trop}=\bigsqcup_{G\in\mathcal{G}_{g,\mathcal{A}}}\overline{\sigma}^\circ_G/\Aut(G) \ .
\end{equation*}
\end{proposition}
\section{Dual graphs and the boundary of $\overline{\mathcal{M}}_{g,\mathcal{A}}$}\label{section_dualgraphs&boundary}
We study the structure of the boundary $\overline{\mathcal{M}}_{g,\mathcal{A}}-\mathcal{M}_{g,\mathcal{A}}$ of $\overline{\mathcal{M}}_{g,\mathcal{A}}$, i.e. the complement of the locus of non-singular curves, using the machinery of dual graphs. The main result of this section is the proof of Theorem \ref{thm_SNC} which states that the boundary has (stack-theoretically) normal crossings.
\subsection{Proof of Theorem \ref{thm_SNC}}\label{section_alternativeSNC}
The author is aware of at least two different proofs of Theorem \ref{thm_SNC}.
Following \cite[Section 3.3]{Hassett_weightedmoduli} one can undertake a detailed analysis of the formal deformation spaces of weighted stable curves, using the deformation theory of maps, as developed in \cite{Ran_defofmaps}. This approach has been carried out in an earlier version of this article.
Here we present an alternative approach to the proof of Theorem \ref{thm_SNC} that is much shorter and more elementary. The author would like to thank Dan Abramovich for communicating this proof to him.
It essentially reduces Theorem \ref{thm_SNC} to the analogous theorem for the Deligne-Knudsen-Mumford moduli spaces $\overline{\mathcal{M}}_{g,n}$, a case that has already been discussed in \cite[Theorem 2.7]{Knudsen_projectivityII}.
Set $N=\dim\overline{\mathcal{M}}_{g,\mathcal{A}}=3g-3+n$. Let $\mathfrak{o}_k$ be either equal to $k$, if $\characteristic k=0$, or to the unique complete regular local ring with residue field $k$ and and maximal ideal generated by $p$, if $\characteristic k=p\neq 0$. Theorem \ref{thm_SNC} immediately follows from the following.
\begin{theorem}\label{thm_nodecoordinates}
Let $z$ be a point of $\overline{\mathcal{M}}_{g,\mathcal{A}}$ corresponding to an $\mathcal{A}$-stable curve $(C,p_1,\ldots,p_n)$ with nodes $x_1,\ldots, x_k$. Then there are formal coordinates $t_1,\ldots t_N$ around $z$ such that the complete local ring $\widehat{\mathcal{O}}_{\overline{\mathcal{M}}_{g,\mathcal{A}},z}$ is isomorphic to $\mathfrak{o}_k\llbracket t_1,\ldots,t_N\rrbracket$ and the locus where $x_i$ stays a node is given by $t_i=0$ for $1\leq i\leq k$.
\end{theorem}
\begin{proof}
Denote by $\mathcal{S}_{g,n}$ the algebraic stack of nodal curves of genus $g$ with $n$ (possibly singular) marked points, as introduced in \cite[Section 5]{Olsson_logtwistedcurves}. Note that $\mathcal{S}_{g,n}$ is locally of finite type over $k$. Consider the universal curve $\mathcal{C}_g$ of $\mathcal{S}_g$ and denote by $\mathcal{C}_g^{sm}$ the open substack of $\mathcal{C}_{g}=\mathcal{S}_{g,1}$ where the marked point is not singular. We can identifiy $\overline{\mathcal{M}}_{g,\mathcal{A}}$ with an open substack of the $n$-fold fibered product $\mathcal{C}_g^{sm}\times_{\mathcal{S}_g}\cdots\times_{\mathcal{S}_g}\mathcal{C}_g^{sm}$, since $\mathcal{A}$-stability is an open condition. Therefore the natural forgetful morphism $f\mathrel{\mathop:}\overline{\mathcal{M}}_{g,\mathcal{A}}\rightarrow \mathcal{S}_g$ is smooth.
Denote the image of $z$ in $\mathcal{S}_g$ by $z'$ and let $N'=3g-3=N-n$. By \cite[Lemma 5.1]{Olsson_logtwistedcurves} there are formal coordinates $t'_1,\ldots,t'_{N'}$ around $z'$ such that
\begin{equation*}
\widehat{\mathcal{O}}_{\mathcal{S}_g,z'}\simeq \mathfrak{o}_k\llbracket t'_1,\ldots, t'_{N'}\rrbracket
\end{equation*}
and the locus where $x_i$ stays a node is given by $t'_i=0$ for $1\leq i\leq k$. Since $f\mathrel{\mathop:}\overline{\mathcal{M}}_{g,\mathcal{A}}\rightarrow \mathcal{S}_g$ is smooth, there are formal coordinates $t_1,\ldots, t_N$ around $z$ such that
\begin{equation*}
\widehat{\mathcal{O}}_{\overline{\mathcal{M}}_{g,\mathcal{A}},z}\simeq \mathfrak{o}_k\llbracket t_1,\ldots, t_{N}\rrbracket
\end{equation*}
and $f^\ast t_i'=t_i$ for all $1\leq i\leq N'$. Therefore the locus where $x_i$ stays a node is given by $t_i=0$.
\end{proof}
\begin{remark}\label{remark_nonNC}
In general, the complement of the smaller open subset $\mathcal{M}_{g,n}\subseteq\mathcal{M}_{g,\mathcal{A}}$ in $\overline{\mathcal{M}}_{g,\mathcal{A}}$ does not have normal crossings. This is due to the fact that marked points $p_{i_1},\ldots,p_{i_k}$ are allowed to coincide whenever the weights fulfill $a_{i_1}+\ldots+a_{i_k}\leq1$ without changing the combinatorial type of the curve. We refer the reader to the example of Losev-Manin spaces $L_n=\overline{M}_{0,\mathcal{A}}$ with $\mathcal{A}=\big(1,\frac{1}{n},\ldots,\frac{1}{n},1\big)\in\mathbb{Q}^{n+2}\cap(0,1]^{n+2}$ discussed in Section \ref{section_LosevManin} below, where the complement of $M_{0,n+2}$ in $L_n$ clearly does not have normal crossings (see Example \ref{example_L_3nonSNC}).
\end{remark}
\subsection{Stratification by dual graphs}\label{section_stratification}
Let $(C,p_1,\ldots,p_n)$ be a complete and connected nodal curve of genus $g$ with $n$ marked points. We can associate to $C$ its \emph{dual graph} $G_C$, a weighted graph with $n$ marked legs that is defined as follows:
\begin{itemize}
\item The set $V=V(G)$ of vertices of $G$ is the set of irreducible components $C_i$ of $C$.
\item The set of edges $E=E(G)$ is the set of nodes of $C$, where an edge $e$ connects two vertices $v_i$ and $v_j$ if and only if the corresponding components $C_i$ and $C_j$ meet each other in the node corresponding to $e$.
\item The set of legs $L=L(G)$ is the set of marked points of $C$. A leg $l_i$ emanates from a vertex $v$ if and only if the marked point $p_i$ corresponding to $l_i$ lies in the component $C_v$ corresponding to $v$.
\item The weight function $h\mathrel{\mathop:}V\rightarrow\mathbb{N}$ is defined by associating to a vertex $v$ the genus $g(C_v)$ of the corresponding component $C_v$.
\end{itemize}
It is well-known (see e.g. \cite[Proposition 2.6]{Manin_quantumbook}) that $g(C)=g(G_C)$ and that there is a natural homomorphism
\begin{equation*}
\Aut(C,p_1,\ldots,p_n)\longrightarrow \Aut(G_C)
\end{equation*}
of automorphism groups.
\begin{proposition}\label{prop_Astability}
Let $(g,\mathcal{A})$ be an input datum. For a complete and connected nodal curve $(C,p_1,\ldots,p_n)$ of genus $g$ with $n$ marked points the following properties are equivalent:
\begin{enumerate}[(i)]
\item The twisted canonical divisor $K_C+a_1p_1+\ldots+a_np_n$ on $C$ is ample.
\item The dual graph $G_C$ of $C$ is stable of type $\mathcal{A}$.
\end{enumerate}
\end{proposition}
\begin{proof}
The twisted canonical divisor $K_C+a_1p_1+\ldots+a_np_n$ on $C$ is ample if and only if its pullback to the normalization of each irreducible component of $C$ is effective and this is equivalent to
\begin{equation*}
2h(v)-2+\vert v\vert_E+\vert v\vert_\mathcal{A}> 0
\end{equation*}
for all vertices $v$ of $G_C$. But this is the exact definition of $G_C$ being stable of type $(g,\mathcal{A})$.
\end{proof}
Fix an input datum $(g,\mathcal{A})$. For a weighted graph $G$ of genus $g$ with $n$ legs that is stable of type $\mathcal{A}$ we denote by $\mathcal{M}_{G}$ the locally closed substack of $\overline{\mathcal{M}}_{g,\mathcal{A}}$ consisting of those $\mathcal{A}$-stable curves $C$ whose dual graph $G_C$ is equal to $G$. The closure of $\mathcal{M}_G$ in $\overline{\mathcal{M}}_{g,\mathcal{A}}$ will be denoted by $\overline{\mathcal{M}}_G$. Note that, if $G$ is the unique graph $\{\ast\}_{g,n}$ of genus $g$ with $n$ legs and one vertex, the stack $\mathcal{M}_{\{\ast\}_{g,n}}$ exactly parametrizes those $\mathcal{A}$-stable curves that are smooth and therefore coincides with $\mathcal{M}_{g,\mathcal{A}}$.
It is not hard to see that the locally closed substacks $\mathcal{M}_{G}$ are the strata of a stratification of $\overline{\mathcal{M}}_{g,\mathcal{A}}$, i.e.
\begin{equation*}
\overline{\mathcal{M}}_{g,\mathcal{A}}=\bigsqcup_{G} \mathcal{M}_G \
\end{equation*}
where the disjoint union is taken over all isomorphism classes of weighted graphs $G$ of genus $g$ with $n$ legs that are stable of type $\mathcal{A}$. As an immediate consequence of Theorem \ref{thm_SNC} we have:
\begin{corollary}\label{cor_codim=numedges}
The codimension of the locally closed stratum $\mathcal{M}_G$ is equal to the number of edges of the $\mathcal{A}$-stable weighted graph $G$.
\end{corollary}
\begin{proof}
The number $k=\#E(G)$ of edges of $G_C$ corresponds to the number of nodes of an $\mathcal{A}$-stable curve $C$ in $\overline{\mathcal{M}}_{g,\mathcal{A}}$. By Theorem \ref{thm_nodecoordinates} the complete local ring $\widehat{\mathcal{O}}_{\mathcal{M}_{g,\mathcal{A}},x}$ at a point $x=\big[(C,p_1,\ldots,p_n)\big]$ of $\mathcal{M}_G$ is equal to $\mathfrak{o}_k\llbracket t_1,\ldots,t_N\rrbracket$, where the coordinates $t_1,\ldots,t_N$ can be chosen such that the locus of the closure of $\mathcal{M}_G$ is given by $t_1\cdots t_k=0$. This implies
\begin{equation*}
\codim\mathcal{M}_G=k=\#E(G)\ .
\end{equation*}
\end{proof}
An alternative proof of Corollary \ref{cor_codim=numedges} can be found at the end of Section \ref{section_algebraicclutching&gluing} below.
\begin{remark}
Stratifications of $\overline{\mathcal{M}}_{g,\mathcal{A}}$ parametrized by the combinatorial data associated to weighted stable curves have already appeared in \cite{BayerManin_weightedstablemaps}, \cite{AlexeevGuy_weightedstablemaps}, and \cite{MustataMustata_weightedstablemaps}, all of which deal with moduli spaces of weighted stable maps. The crucial difference between their constructions and our approach is that their discrete data also contains information on whether marked points agree. Taking only dual graphs, we conveniently "forget" this information in order to obtain the stratification induced by the normal crossing boundary $\overline{\mathcal{M}}_{g,\mathcal{A}}-\mathcal{M}_{g,\mathcal{A}}$.
\end{remark}
\subsection{Clutching and gluing}\label{section_algebraicclutching&gluing}
In this section we study analogues of the clutching and gluing morphisms originally defined in \cite[Section 3]{Knudsen_projectivityII} for $\overline{\mathcal{M}}_{g,n}$. This construction is a special case of \cite[Proposition 2.1.1]{BayerManin_weightedstablemaps}, where these maps are defined for moduli spaces of weighted stable maps.
For a vertex $v$ of $G$ we denote by $\mathcal{A}(v)$ the tuple
\begin{equation*}
(a_{i_1},\ldots a_{i_k},1,\ldots,1)
\end{equation*}
consisting of those $a_i$ that correspond to legs $l_i$ emanating from $v$ and a $1$ for every flag of an edge incident to $v$. Moreover we write $n_v$ for the number of entries of $\mathcal{A}(v)$, i.e. the number of legs and edges emanating from $v$.
\begin{proposition}\label{prop_clutching&gluing}
For every weighted graph $G$ that is stable of type $(g,\mathcal{A})$ there is a natural \emph{clutching and gluing morphism}
\begin{equation*}
\phi_G\mathrel{\mathop:}\prod_{v\in V(G)}\overline{\mathcal{M}}_{h(v),\mathcal{A}(v)}\longrightarrow\overline{\mathcal{M}}_G\subseteq\overline{\mathcal{M}}_{g,\mathcal{A}}
\end{equation*}
of the moduli stacks that associates to a tuple consisting of stable curves $(C^v,p_1^v,\ldots,p_{n_v}^v)$ of type $\big(h(v),\mathcal{A}(v)\big)$ the stable curve $(C,p_1,\ldots, p_n)$ obtained by identifying two marked points, whenever they correspond to two flags defining an edge of $G$.
\end{proposition}
\begin{proof}
Let $S$ be scheme and $(C^v)$ be a tuple of complete nodal curves over $S$ with sections $p_1^v,\ldots,p_n^v$ such that each $(C^v,p_1^v,\ldots,p_{n_v}^v)$ is stable of type $\big(h(v),\mathcal{A}(v)\big)$. Then we can define a curve $C$ over $S$ by gluing the $C^v$ over two sections corresponding to two flags that are connected by an edge of $G$. Note that these sections do not intersect any other section, since they all have weight one and the $(C^v,p_1^v,\ldots,p_{n_v}^v)$ are stable of type $\big(h(v),\mathcal{A}(v)\big)$. The resulting curve $(C,p_1,\ldots, p_n)$ over $S$ is stable of type $(g,\mathcal{A})$, since each $(C^v,p_1^v,\ldots,p_{n_v}^v)$ is stable of type $\big(h(v),\mathcal{A}(v)\big)$ and the graph $\Gamma$ is stable of type $(g,\mathcal{A})$. This association commutes with arbitrary base changes $S'\rightarrow S$ and therefore defines a morphism of stacks.
\end{proof}
\begin{corollary}\label{cor_clutching&gluing}\begin{enumerate}[(i)]
\item Suppose $(g_1,\mathcal{A}_1)$ and $(g_2,\mathcal{A}_2)$ are two weight data. Set $g=g_1+g_2$ and $\mathcal{A}=\mathcal{A}_1\cup\mathcal{A}_2$. There is a natural \emph{clutching morphism}
\begin{equation*}
\kappa=\kappa_{g_1,\mathcal{A}_1,g_2,\mathcal{A}_2}\mathrel{\mathop:} \overline{\mathcal{M}}_{g_1,\mathcal{A}_1\sqcup\{1\}}\times\overline{\mathcal{M}}_{g_2,\mathcal{A}_2\sqcup\{1\}}\longrightarrow \overline{\mathcal{M}}_{g,\mathcal{A}}
\end{equation*}
that associates to a tuple consisting of the two $\mathcal{A}_i\cup\{1\}$-stable curves $(C_i,p_1^i,\ldots,p_{n_i+1}^i)$ the $\mathcal{A}$-stable curve $(C,p_1^1,\ldots, p_{n_1}^1,p_{1}^2,\ldots,p_{n_2}^2)$ obtained by identifying the two points $p_{n_1+1}^1$ and $p_{n_2+1}^2$ in a node.
\item Fix an input datum $(g,\mathcal{A})$ with $g>0$. There is a natural \emph{gluing morphism}
\begin{equation*}
\gamma\mathrel{\mathop:}\overline{\mathcal{M}}_{g-1,\mathcal{A}\sqcup\{1,1\}}\longrightarrow\overline{\mathcal{M}}_{g,\mathcal{A}}
\end{equation*}
obtained by gluing together the last two marked points of an $\mathcal{A}\cup\{1,1\}$-stable curve $(C,p_1,\ldots,p_{n+2})$ of genus $g-1$.
\end{enumerate}\end{corollary}
\begin{proof}
Both Part (i) and Part (ii) are special cases of Proposition \ref{prop_clutching&gluing}. For Part (i) we take the graph $G$ to consist of two vertices $v_1$ and $v_2$ with weights $g_1$ and $g_2$ connected by an edge and having $n_1$ or $n_2$ legs incident to $v_1$ or $v_2$ respectively. For Part (ii) the graph $G$ consists only one vertex with weight $g$, from which $n$ legs are emanating, and a loop.
\end{proof}
Set
\begin{equation*}
\widetilde{\mathcal{M}}_G=\prod_{v\in V(G)}\mathcal{M}_{h(v),\mathcal{A}(v)}
\end{equation*}
and note that the clutching and gluing morphism $\phi_G$ restricts to a morphism
\begin{equation*}
\widetilde{\mathcal{M}}_G\rightarrow \mathcal{M}_G \ .
\end{equation*}
\begin{proposition}\label{prop_MGstructure}
For a weighted graph $G$ that is stable of type $(g,\mathcal{A})$ the clutching and gluing morphism $\phi_G\mathrel{\mathop:}\widetilde{\mathcal{M}}_G\rightarrow \mathcal{M}_G$ induces an isomorphism
\begin{equation*}
\Big[\widetilde{\mathcal{M}}_G\big/\Aut(G)\Big]\simeq\mathcal{M}_G \ .
\end{equation*}
\end{proposition}
Our proof of Proposition \ref{prop_MGstructure} is a generalization of the proof of \cite[Proposition 10.11]{ArbarelloCornalbaGriffiths_moduliofcurves}.
\begin{proof}
We are going to show that both $\Big[\widetilde{\mathcal{M}}_G\big/\Aut(G)\Big]$ and $\mathcal{M}_G$ have the same groupoid presentation.
By the construction in \cite[Example 8.15]{ArbarelloCornalbaGriffiths_moduliofcurves} we can find a $\Aut(G)$-invariant surjective \'etale morphism $s,t\mathrel{\mathop:}U\rightarrow \widetilde{\mathcal{M}}_G$. In this case a groupoid presentation of $\Big[\widetilde{\mathcal{M}}_G\big/\Aut(G)\Big]$ is given by $Y_1\rightrightarrows Y_0$, where
\begin{equation*}
Y_1=\Aut(G)\times Y_0\times_{\mathcal{M}_G}Y_0 \ .
\end{equation*}
The \'etale atlas $Y_0$ is \'etale locally isomorphic to a product $\prod_{v\in V(G)}U_v$, where $U_v$ are local slices of the 'exhausting family' of $\mathcal{M}_{h(v),\mathcal{A}(v)}$ around curves $C_v$ as introduced in \cite[Section 3.4]{Hassett_weightedmoduli}. The clutching and gluing map is induced by the morphism $\prod_{v\in V(G)} U_v\rightarrow U$ into a slice $U$ of the 'exhausting family' of $\mathcal{M}_{g,\mathcal{A}}$ around $C=\phi_G\big((C_v)_{v\in V(G)}\big)$ that is determined by isomorphically mapping $U_v$ to one of the branches of $U_G$, the locus in $U$ parametrizing curves with dual graph $G$. In particular the composition $Y_0\rightarrow\mathcal{M}_G$ is surjective and \'etale.
The morphism $Y_0\rightarrow\widetilde{\mathcal{M}}_G$ gives rise to a family $\eta\mathrel{\mathop:}\mathcal{C}\rightarrow Y_0$ of curves with dual graphs equal to $G$. In this case we have natural isomorphisms
\begin{equation*}
Y_1=\Aut(G)\times Y_0\times_{\widetilde{\mathcal{M}}_G} Y_0\simeq\Aut(G)\times \Isom_{Y_0\times Y_0}^G(s^\ast\eta,t^\ast \eta) \simeq Y_0\times_{\mathcal{M}_G} Y_0 \ ,
\end{equation*}
where $\Isom^G$ denotes isomorphisms preserving the dual graph $G$.
\end{proof}
\begin{proof}[Alternative proof of Corollary \ref{cor_codim=numedges}]
The Betti number $b_1(G)$ can be can calculated by $b_1(G)=\#E(G)-\#V(G)+1$ and by Proposition \ref{prop_MGstructure} we have $\dim\mathcal{M}_G=\dim\widetilde{\mathcal{M}}_G$. Using $g=b_1(G)+\sum_v h(v)$ as well as $\sum_vn_v=n+2\cdot\#E(G)$, we therefore obtain:
\begin{equation*}\begin{split}
\dim\mathcal{M}_G=\dim\widetilde{\mathcal{M}}_G&=\sum_{v\in V(G)}3h(v)-3+n_v \\
&=3\big(b_1(G)+\sum_{v\in V(G)}h(v)\big)-3\big(\#V(G)-b_1(G)\big) +\sum_{v\in V(G)}n_v\\
&=3g-3+n +2\cdot\#E(G) -3\cdot\#E(G)\\
&=\dim\overline{\mathcal{M}}_{g,\mathcal{A}} -\#E(G) \ .
\end{split}\end{equation*}
\end{proof}
\section{Deformation retraction onto the non-Archimedean skeleton}
The goal of this section is to prove our main result, Theorem \ref{thm_modulartrop}. We begin with a quick review of the construction of the deformation retraction onto the skeleton of a simple toroidal scheme from \cite[Section 3.1]{Thuillier_toroidal} in Section \ref{section_skeletonsimple} and more generally of a toroidal Deligne-Mumford stack from \cite[Section 6]{AbramovichCaporasoPayne_tropicalmoduli} in Section \ref{section_skeletonstack}. Section \ref{section_proof} contains the proof of Theorem \ref{thm_modulartrop}.
\subsection{Skeletons of simple toroidal schemes}\label{section_skeletonsimple}
Suppose that $X_0\hookrightarrow X$ is a simple toroidal embedding, that is an open embedding such that for every point $x\in X$ there is an open neighborhood $U$ of $x$ and an \'etale morphism $\gamma\mathrel{\mathop:}U\rightarrow Z$ into a toric variety $Z$ with big torus $T$ such that $\gamma^{-1}(T)=X_0\cap U$. In \cite[Section 3.2]{Thuillier_toroidal} Thuillier defines a strong deformation retraction $\mathbf{p}$ from the non-Archimedean analytic space $X^\beth$ as defined in \cite[Section 1]{Thuillier_toroidal} onto a closed subset $\mathfrak{S}(X)$ of $X^\beth$, the \emph{non-Archimedean skeleton} of $X$.
We denote by $S_+$ the sheaf monoids associating to an open subset $U$ of $X$ the monoid $S_+(U)$ of effective Cartier divisors with support fully contained in $X-X_0$. As shown in \cite[Section 3.1]{Thuillier_toroidal} the natural stratification of the toric varieties $Z$ by $T$-orbits lifts to give a well-defined stratification of $X$ by locally closed subsets, henceforth called the \emph{toroidal strata} of $X$. Note that the unique open subset of this stratification is $X_0$. Denote by $F_X$ the set of generic points of the toroidal strata together with the induced topology and endowed with the restriction of the $S_+$. By \cite[Proposition 9.2]{Kato_toricsing} the monoidal space $F_X$ has the structure of what is called a Kato fan in \cite{Ulirsch_functroplogsch} and comes with a natural characteristic morphism $\phi_X\mathrel{\mathop:}(X,S_+)\rightarrow F_X$ sending every point in a toroidal stratum to its generic point. We refer the reader to \cite{Ulirsch_functroplogsch} for details on this construction.
In particular, by \cite[Theorem 1.2]{Ulirsch_functroplogsch} Thuillier's strong deformation retraction can be described as follows:
\begin{itemize}
\item The skeleton $\mathfrak{S}(X)$ is naturally homeomorphic to the set $\overline{\Sigma}_{X}=F_X(\overline{\mathbb{R}}_{\geq 0})$ of $\overline{\mathbb{R}}_{\geq 0}=\mathbb{R}_{\geq 0}\sqcup\{\infty\}$-valued points.
\item A point $x$ in $X^\beth$ gives rise to a morphism $\underline{x}\mathrel{\mathop:}\Spec R\rightarrow (X,S_+)$ of monoidal spaces, where $R$ is some non-Archimedean extension of $k$, and the image of $x$ in $\overline{\Sigma}_X$ is given by the composition
\begin{equation*}\begin{CD}
\Spec\overline{\mathbb{R}}_{\geq 0}@>\val^\#>>\Spec R@>\underline{x}>>(X,S_+)@>\phi_X>>F_X ,
\end{CD}\end{equation*}
where $\val^\#$ is the morphism induced by the valuation of $R$.
\end{itemize}
\subsection{Skeletons of toroidal Deligne-Mumford stacks} \label{section_skeletonstack}
Suppose now that $\mathcal{X}_0\hookrightarrow\mathcal{X}$ is toroidal embedding of separated Deligne-Mumford stacks of finite type over $k$, i.e. an open embedding of Deligne-Mumford stacks admitting a surjective \'etale morphism $U\rightarrow\mathcal{X}$ such that the base change $U_0\hookrightarrow U$ is a simple toroidal embedding. The toroidal stratification of $U$ induces a natural toroidal stratification of $\mathcal{X}$ by locally closed substacks $\mathcal{E}$ that does not depend on the choice of $U$. We write $S_+$ for the \'etale sheaf of effective Cartier-divisors with support in $\mathcal{X}-\mathcal{X}_0$.
The Keel-Mori Theorem \cite{KeelMori_groupoidquotients} implies that the stack $\mathcal{X}$ has a coarse moduli space $X$, which has the structure of separated algebraic space. By \cite{ConradTemkin_algspaces} the analytification $X^{an}$ of $X$ exists in the category of analytic spaces and, following \cite[Definition 6.1.2]{AbramovichCaporasoPayne_tropicalmoduli}, we can define $X^\beth$ as the subspace of $X^{an}$ that is locally given by unit balls in $X^{an}$. Note that the valuative criterion for properness yields $X^\beth=X^{an}$, whenever $\mathcal{X}$ is proper over $k$.
In Section \cite[Section 6.1]{AbramovichCaporasoPayne_tropicalmoduli} the authors extend Thuillier's \cite{Thuillier_toroidal} construction and show that this datum defines a strong deformation retraction $\mathbf{p}$ of $X^\beth$ onto a closed subset $\mathfrak{S}(\mathcal{X})$ of $X^\beth$, which is again called the \emph{non-Archimedean skeleton} of $\mathcal{X}$.
Consider now the category $\mathcal{H}_\mathcal{X}$ defined as follows:
\begin{itemize}
\item Its objects are the generic points of the toroidal strata of $\mathcal{X}$.
\item The morphisms in $\mathcal{H}_\mathcal{X}$ are generated by the natural homomorphisms $(S_+)_\eta\rightarrow (S_+)_\xi$, whenever $\eta$ specializes to $\xi$, and the \emph{monodromy groups} $H_\xi$ at $\xi$. \end{itemize}
Recall that the sheaf $S_+$ is \'etale locally trivial on the toroidal strata of $\mathcal{X}$ by \cite[Proposition 6.2.1]{AbramovichCaporasoPayne_tropicalmoduli}. The \emph{monodromy group} $H_\xi$ consists of those automorphisms of $(S_+)_\eta$ that are induced by the operation of $\pi_1^{et}(\mathcal{E}_\xi,\xi)$ on $(S_+)_\xi$, where $\mathcal{E}_\xi$ is the unique toroidal stratum containing $\xi$.
There is a natural functor $\Sigma\mathrel{\mathop:}\mathcal{H}_\mathcal{X}\rightarrow \mathbf{RPCC}$ given by
\begin{itemize}
\item the association $\xi\rightarrow\sigma_\xi=\Hom\big((S_+)_\xi,\mathbb{R}_{\geq 0}\big)$,
\item the embedding of a face $\sigma_\eta\hookrightarrow\sigma_\xi$, whenever $\eta$ specializes to $\xi$, and
\item an automorphism of $\sigma_\xi$ for every automorphism of the monodromy group $H_\xi$.
\end{itemize}
This functor naturally extends to a functor $\overline{\Sigma}$ into the category of extended rational polyhedral complexes, given by $\xi\rightarrow \sigma_\xi=\Hom\big((S_+)_\xi,\overline{\mathbb{R}}_{\geq 0}\big)$.
We can now rephrase \cite[Proposition 6.2.6]{AbramovichCaporasoPayne_tropicalmoduli} as follows:
\begin{proposition}[ \cite{AbramovichCaporasoPayne_tropicalmoduli} Proposition 6.2.6]\label{prop_skeleton=colimit} The skeleton $\mathfrak{S}(\mathcal{X})$ is the colimit
\begin{equation*}
\mathfrak{S}(\mathcal{X})=\lim_{\longrightarrow} \overline{\sigma}_\xi \ ,
\end{equation*}
taken over the category $\mathcal{H}_\mathcal{X}$. Let $x\in X^\beth$ be represented by a morphism $\underline{x}\mathrel{\mathop:}\Spec R\rightarrow \mathcal{X}$ from a valuation ring extending $k$ and write $p_x$ for the image of the closed point in $\mathcal{X}$. Then $(S_+)_{p_x}=(S_+)_{\xi_x}$ for the generic point $\xi_x$ of the unique stratum containing $p_x$ and the image $\mathbf{p}(x)$ in $\overline{\sigma}_{\xi_x}=\Hom\big((S_+)_{\xi_x},\overline{\mathbb{R}}_{\geq 0}\big)$ is given by the composition
\begin{equation*}\begin{CD}
\Spec \overline{\mathbb{R}}_{\geq 0}@>\val^\#>>\Spec R @>>>\Spec \widehat{\mathcal{O}}_{\mathcal{X},p_x}@>>>\Spec (S_+)_{p_x} \ .
\end{CD}\end{equation*}
\end{proposition}
\begin{remark}
Suppose that $\mathcal{X}$ is a proper toroidal Deligne-Mumford stack of finite type over $k$. By \cite[Section 1.5]{Ulirsch_nonArchstacks} the skeleton $\mathfrak{S}(\mathcal{X})$ is actually a deformation retract of the underlying topological space $\vert \mathcal{X}^{an}\vert$ of the analytic stack $\mathcal{X}^{an}$ associated to $\mathcal{X}$.
\end{remark}
\subsection{The skeleton of $\overline{\mathcal{M}}_{g,\mathcal{A}}$}\label{section_proof}
By Theorem \ref{thm_SNC} the open embedding $\mathcal{M}_{g,\mathcal{A}}\hookrightarrow\overline{\mathcal{M}}_{g,\mathcal{A}}$ defines a toroidal structure on $\overline{\mathcal{M}}_{g,\mathcal{A}}$. Note the toroidal stratification is exactly the stratification of $\overline{\mathcal{M}}_{g,\mathcal{A}}$ by dual graphs introduced in Section \ref{section_stratification}. Denote by $\xi_G$ the generic point of the stratum $\mathcal{M}_G$. The following Lemma \ref{lemma_pi1=Aut} is a generalization of \cite[Proposition 7.2.1]{AbramovichCaporasoPayne_tropicalmoduli}.
\begin{lemma}\label{lemma_pi1=Aut}
The association $G\mapsto \xi_G$ defines a natural equivalence between the categories $(\mathcal{G}_{g,\mathcal{A}})^{op}$ and $\mathcal{H}_{\overline{\mathcal{M}}_{g,\mathcal{A}}}$.
\end{lemma}
\begin{proof}
Note first that weighted graph contraction $G'\rightarrow G$ are in an unique order-reversing one-to-one correspondence with the specialization relations $\xi_G\rightarrow\xi_{G'}$. Therefore it is enough to show that the image of $\pi_1^{et}(\overline{\mathcal{M}}_{g,\mathcal{A}},\xi_G)$ acting on the the $(S_+)_{\xi_G}$ is precisely $\Aut(G)$. Consider the Galois cover $\widetilde{\mathcal{M}}_G\rightarrow\mathcal{M}_G$. The operation of $\pi_1^{et}(\overline{\mathcal{M}}_{g,\mathcal{A}},\xi_G)$ on the sheaf pullback of $S_+$ to $\widetilde{M}_G$ is trivial and it therefore factors through its quotient $\Aut(G)$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm_modulartrop}]
Recall that $\overline{M}_{g,\mathcal{A}}^{trop}$ is defined as the colimit
\begin{equation*}
\overline{M}_{g,\mathcal{A}}^{trop}=\lim_{\longrightarrow} \overline{\sigma}_G
\end{equation*}
taken over the category $(\mathcal{G}_{g,\mathcal{A}})^{op}$ and that by Proposition \ref{prop_skeleton=colimit} the skeleton $\mathfrak{S}(\overline{\mathcal{M}}_{g,\mathcal{A}})$ is given as the colimit
\begin{equation*}
\mathfrak{S}(\overline{\mathcal{M}}_{g,\mathcal{A}})=\lim_{\longrightarrow} \overline{\sigma}_\xi
\end{equation*}
taken over the category $\mathcal{H}_{\overline{\mathcal{M}}_{g,\mathcal{A}}}$. Therefore Lemma \ref{lemma_pi1=Aut} immediately implies that there is an isomorphism
\begin{equation*}
J_{g,\mathcal{A}}\mathrel{\mathop:}\overline{M}_{g,\mathcal{A}}^{trop}\longrightarrow\mathfrak{S}(\overline{\mathcal{M}}_{g,\mathcal{A}}) \ .
\end{equation*}
of generalized extended cone complexes.
We finally show that the strong deformation retraction $\mathbf{p}\mathrel{\mathop:}\overline{M}_{g,\mathcal{A}}^{an}\rightarrow\mathfrak{S}(\overline{\mathcal{M}}_{g,\mathcal{A}})$ can be given a modular interpretation as stated in the introduction. By the valuative criterion of properness a point $x\in \overline{M}_{g,\mathcal{A}}^{an}$ can be represented by a morphism $\Spec R\rightarrow\overline{\mathcal{M}}_{g,\mathcal{A}}$, which, in turn, gives rise to a $(g,\mathcal{A})$-stable curve $\mathcal{C}_x\rightarrow \Spec R$ over $R$. Denote the dual graph of its special fiber $\mathcal{C}_s$ by $G_x$ and the image of the closed point in $\Spec R$ in $\overline{\mathcal{M}}_{g,\mathcal{A}}$ by $p_x$.
By Theorem \ref{thm_nodecoordinates} we can choose coordinates $t_1,\ldots, t_N$ in $\widehat{\mathcal{O}}_{\overline{\mathcal{M}}_{g,\mathcal{A}},p_x}$ such that the locus, where $\mathcal{C}_s$ remains singular is given by $t_1\ldots t_k=0$. In formal coordinates we can describe $\mathcal{C}$ around a node $q_i$ of $\mathcal{C}_s$ by $xy=f_i$, where the $f_i\in R$ are precisely the images of $t_i$ in $R$. Now both the deformation retraction $\mathbf{p}$ and $\trop_{g,\mathcal{A}}$ are given by associating to $x$ the element in $\mathfrak{S}(\overline{\mathcal{M}}_{g,\mathcal{A}})=\overline{M}_{g,\mathcal{A}}^{trop}$ represented by $\big(\val(f_1),\ldots,\val(f_k),0,\ldots, 0\big)$ in $\overline{\sigma}_{G_x}$. This shows $\mathbf{p}(x)=\trop_{g,\mathcal{A}}(x)$ and finishes the proof of Theorem \ref{thm_modulartrop}.
\end{proof}
\section{Tropical tautological maps}\label{section_troptaut}
The purpose of this section is to define tropical analogues of the tautological maps between the moduli spaces $\overline{\mathcal{M}}_{g,\mathcal{A}}$ generalizing the constructions in \cite[Section 8]{AbramovichCaporasoPayne_tropicalmoduli}. We require the tropical tautological maps to commute with the tropicalization map
\begin{equation*}
\trop_{g,\mathcal{A}}\mathrel{\mathop:}\overline{M}_{g,\mathcal{A}}^{an}\longrightarrow\overline{M}_{g,\mathcal{A}}^{trop}\end{equation*}
as a basic principle to justify that our definitions make sense.
\subsection{Forgetful and reduction morphisms}\label{section_forgetful&reduction}
Fix an input datum $(\mathcal{A},g)$ and let $\mathcal{B}=(b_1,\ldots,b_n)$ be another tuple of weights such that $b_i\leq a_i$ for all $1\leq i\leq n$. In \cite[Theorem 4.1]{Hassett_weightedmoduli} Hassett constructs a natural birational \emph{reduction morphism}
\begin{equation*}
\rho_{\mathcal{A},\mathcal{B}}\mathrel{\mathop:}\overline{\mathcal{M}}_{g,\mathcal{A}}\longrightarrow\overline{\mathcal{M}}_{g,\mathcal{B}}
\end{equation*}
that takes an element $(C,p_1,\ldots, p_n)$ of $\overline{\mathcal{M}}_{g,\mathcal{A}}$ and collapses all the components along which the divisor $K_C=b_1p_1+\ldots+b_np_n$ fails to be ample.
Moreover consider a subset $\mathcal{A}'=\{a_{i_1},\dots,a_{i_r}\}\subseteq\mathcal{A}$ such that $2g-2+a_{i_1}+\ldots +a_{i_r}>0$. By \cite[Theorem 4.3]{Hassett_weightedmoduli} there is a natural \emph{forgetful morphism}
\begin{equation*}
\phi_{\mathcal{A},\mathcal{A}'}\mathrel{\mathop:}\overline{\mathcal{M}}_{g,\mathcal{A}}\longrightarrow\overline{\mathcal{M}}_{g,\mathcal{A}'}
\end{equation*}
that can be described by associating to an $\mathcal{A}$-stable curve $(C,p_1,\ldots, p_n)$ in $\overline{\mathcal{M}}_{g,\mathcal{A}}$ the curve $\phi_{\mathcal{A},\mathcal{A}'}(C,p_1,\ldots, p_n)$ given by deleting the marked points $p_i$ with $i\notin\mathcal{A}'$ and successively collapsing the components of $C$ such that $K_C+a_{i_1}p_{i_1}+\ldots +a_{i_r}p_{i_r}$ is not ample.
\begin{proposition}\label{prop_tropicalreduction&forgetful}
There is a natural \emph{tropical reduction map}
\begin{equation*}
\rho_{\mathcal{A},\mathcal{B}}^{trop}\mathrel{\mathop:}\overline{M}_{g,\mathcal{A}}^{trop}\longrightarrow\overline{M}_{g,\mathcal{B}}^{trop}
\end{equation*}
and a natural \emph{tropical forgetful map}
\begin{equation*}
\phi_{\mathcal{A},\mathcal{A}'}^{trop}\mathrel{\mathop:}\overline{M}_{g,\mathcal{A}}^{trop}\longrightarrow\overline{M}_{g,\mathcal{A}'}^{trop}
\end{equation*}
making the diagrams
\begin{equation*}
\begin{CD}
\overline{M}_{g,\mathcal{A}}^{an}@>\trop_{g,\mathcal{A}}>>\overline{M}_{g,\mathcal{A}}^{trop}\\
@V\rho_{\mathcal{A},\mathcal{B}}^{an}VV@VV\rho_{\mathcal{A},\mathcal{B}}^{trop}V\\
\overline{M}_{g,\mathcal{B}}^{an}@>\trop_{g,\mathcal{B}}>>\overline{M}_{g,\mathcal{B}}^{trop}
\end{CD}
\qquad\qquad
\begin{CD}
\overline{M}_{g,\mathcal{A}}^{an}@>\trop_{g,\mathcal{A}}>>\overline{M}_{g,\mathcal{A}}^{trop}\\
@V\phi_{\mathcal{A},\mathcal{A}'}^{an}VV@VV\phi_{\mathcal{A},\mathcal{A}'}^{trop}V\\
\overline{M}_{g,\mathcal{A}'}^{an}@>\trop_{g,\mathcal{A}'}>>\overline{M}_{g,\mathcal{A}'}^{trop}
\end{CD}
\end{equation*}
commutative.
\end{proposition}
Proposition \ref{prop_tropicalreduction&forgetful} is an immediate consequence of Hassett's description of the forgetful and reduction morphisms for $\overline{\mathcal{M}}_{g,\mathcal{A}}$ in \cite[Section 4.1]{Hassett_weightedmoduli} as well as the reasoning in \cite[Section 8.2]{AbramovichCaporasoPayne_tropicalmoduli}. We provide a proof in our language for the convenience of the reader.
\begin{proof}[Proof of Propostion \ref{prop_tropicalreduction&forgetful}]
We shall prove both statements simultaneously using the notation $\psi_{\mathcal{A},\mathcal{B}}$ for both the reduction morphism and the forgetful morphism. To make this notation consistent we follow \cite[Section 4.1]{Hassett_weightedmoduli} and formally set $\mathcal{B}=\mathcal{A}'\cup\{0,\ldots,0\}$ as well as:
\begin{itemize}
\item $\overline{\mathcal{M}}_{g,\mathcal{B}}=\overline{\mathcal{M}}_{g,\mathcal{A}'}$
\item $\overline{M}_{g,\mathcal{B}}^{trop}=\overline{M}_{g,\mathcal{A}'}^{trop}$
\item $\mathcal{G}_{g,\mathcal{B}}=\mathcal{G}_{g,\mathcal{A}'}$
\item $\overline{\Sigma}_{g,\mathcal{B}}=\overline{\Sigma}_{g,\mathcal{A}'}$
\end{itemize}
Our approach is to define natural functors
\begin{equation*}
\psi_{\mathcal{A},\mathcal{B}}^{\mathcal{G}}\mathrel{\mathop:}\mathcal{G}_{g,\mathcal{A}}\rightarrow\mathcal{G}_{g,\mathcal{B}}
\end{equation*}
and
\begin{equation*}
\psi_{\mathcal{A},\mathcal{B}}^{\overline{\Sigma}}\mathrel{\mathop:}\overline{\Sigma}_{g,\mathcal{A}}\rightarrow\overline{\Sigma}_{g,\mathcal{B}}\end{equation*}
that will induce $\psi_{\mathcal{A},\mathcal{B}}^{trop}$ by the universal property of colimits.
Let $G$ be a $\mathcal{A}$-stable weighted graph. If $G$ is not $\mathcal{B}$-stable we find ourselves in one of the following two situations:
\begin{enumerate}[(i)]
\item There is a vertex $v\in V(G^\ast)$ such that $h(v)=0$, $\vert v\vert_{E}=1$, and
\begin{equation*}
2h(v)-2+\vert v\vert_E+\vert v\vert_{\mathcal{B}}=\vert v\vert_{\mathcal{B}} -1\leq 0 \ .
\end{equation*}
In this case we contract the unique edge $e$ incident to $v$ and attach all the legs of $G^\ast$ incident to $v$ to the vertex on the other end of $e$.
\item There is a vertex $v\in V(G^\ast)$ such that $h(v)=0$, $\vert v\vert_E=2$, and
\begin{equation*}
2h(v)-2+\vert v\vert_E+\vert v\vert_{\mathcal{B}}=\vert v\vert_{\mathcal{B}} = 0 \ .
\end{equation*}
In this case the graph $G$ does not have any legs of positive weight incident to $v$ and we replace the two adjacent edges $e_1$ and $e_2$ by one edge connecting the two vertices $v_1$ and $v_2$ on the other end of $v$.
\end{enumerate}
Applying the algorithm described in (i) and (ii) possibly multiple times and deleting legs of zero weight we obtain a functor
\begin{equation*}\begin{split}
\psi_{\mathcal{A},\mathcal{B}}^\mathcal{G}\mathrel{\mathop:}\mathcal{G}_{g,\mathcal{A}}&\longrightarrow\mathcal{G}_{g,\mathcal{B}} \\
G&\longmapsto G^\ast \ ,
\end{split}\end{equation*}
since automorphisms of $G$ induce automorphisms of $G^\ast$ and weighted edge contractions of $G$ naturally induced weighted edge contractions of $G^\ast$.
Moreover, the projection maps $\overline{\sigma}_G\rightarrow\overline{\sigma}_{G^\ast}$ induce a functor $\psi_{\mathcal{A},\mathcal{B}}^{\overline{\Sigma}}\mathrel{\mathop:}\overline{\Sigma}_{g,\mathcal{A}} \rightarrow\overline{\Sigma}_{g,\mathcal{A}}$ making the diagram
\begin{equation*}\begin{CD}
\mathcal{G}_{g,\mathcal{A}}@>\overline{\Sigma}_{g,\mathcal{A}}>>\overline{\Sigma}_{g,\mathcal{A}}\\
@V\psi_{\mathcal{A},\mathcal{A}'}^\mathcal{G} VV @VV\psi_{\mathcal{A},\mathcal{A}'}^{\overline{\Sigma}} V\\
\mathcal{G}_{g,\mathcal{A}'}@>\overline{\Sigma}_{g,\mathcal{A}'}>>\overline{\Sigma}_{g,\mathcal{A}'}
\end{CD}\end{equation*}
commutative.
The map $\psi_{\mathcal{A},\mathcal{B}}^{trop}$ is defined to be the map $\overline{M}_{g,\mathcal{A}}^{trop}\rightarrow\overline{M}_{g,\mathcal{B}}^{trop}$ induced by $\psi_{\mathcal{A},\mathcal{B}}^{\mathcal{G}}$ and $\psi_{\mathcal{A},\mathcal{B}}^{\overline{\Sigma}}$ using the universal property of colimits.
Now note that the morphism $\psi_{\mathcal{A},\mathcal{B}}$ induces a functor $\psi_{\mathcal{A},\mathcal{B}}^\mathcal{H}\mathrel{\mathop:}\mathcal{H}_{\overline{\mathcal{M}}_{g,\mathcal{A}}}\rightarrow\mathcal{H}_{\overline{\mathcal{M}}_{g,\mathcal{B}}}$ making the natural diagram
\begin{equation*}\begin{CD}
\mathcal{H}_{\overline{\mathcal{M}}_{g,\mathcal{A}}} @<\simeq<< (\mathcal{G}_{g,\mathcal{A}})^{op}\\
@V\psi_{\mathcal{A},\mathcal{A}'}^\mathcal{H} VV @VV\psi_{\mathcal{A},\mathcal{A}'}^\mathcal{G} V\\
\mathcal{H}_{\overline{\mathcal{M}}_{g,\mathcal{B}}} @<\simeq<< (\mathcal{G}_{g,\mathcal{B}})^{op}
\end{CD}\end{equation*}
commutative. Because of that and Theorem \ref{thm_modulartrop} the diagrams in the statement of Theorem \ref{prop_tropicalreduction&forgetful} commute.
\end{proof}
\begin{proposition}\label{prop_reduction=subcomplex}
The tropical reduction morphism $\phi_{\mathcal{A},\mathcal{B}}^{trop}$ has a section identifying the moduli space $\overline{M}_{g,\mathcal{B}}^{trop}$ with the subcomplex of $\overline{M}_{g,\mathcal{A}}^{trop}$ given by removing those extended relatively open cones $\overline{\sigma}^\circ_G$ such that $G$ is not $\mathcal{B}$-stable.
\end{proposition}
\begin{proof}
Suppose that $G$ is weighted graph that is stable of type $(g,\mathcal{A})$ but not of type $(g,\mathcal{B})$. Then all other weighted graphs $G'$ that can be contracted to $G$ are also not stable of type $(g,\mathcal{B})$. On the other hand all graphs $G$ that are stable of type $(g,\mathcal{B})$ are also stable of type $(g,\mathcal{A})$ and their reduction $\rho_{\mathcal{A},\mathcal{B}}(G)$ is equal to $G$ itself.
\end{proof}
\begin{remark}
Similar sections exists for the forgetful morphism in both the algebraic and the tropical world (see \cite[Proposition 8.2.4]{AbramovichCaporasoPayne_tropicalmoduli} for the case $\mathcal{A}=(1,\ldots,1)$). Unlike this case, the section of the reduction morphism constructed in Proposition \ref{prop_reduction=subcomplex} does not have an algebraic analogue. One can, however, define a continuous section of $\phi_{\mathcal{A},\mathcal{B}}$ on the level of underlying topological spaces.
\end{remark}
\subsection{Clutching and gluing}\label{section_tropicalclutching&gluing}
Let $(g,\mathcal{A})$ be a fixed input datum and $G$ be a weighted graph that is stable of type $(g,\mathcal{A})$. Recall from Section \ref{section_algebraicclutching&gluing} that for a vertex $v$ of $G$ we denote by $\mathcal{A}(v)$ the tuple
\begin{equation*}
(a_{i_1},\ldots a_{i_k},1,\ldots,1)
\end{equation*}
consisting of those $a_i$ that correspond to legs $l_i$ emanating from $v$ and a $1$ for every flag of an edge incident to $v$.
\begin{definition}\label{def_tropicalclutching&gluing}
In analogy with the algebraic situation in Section \ref{section_algebraicclutching&gluing} we define the \emph{tropical clutching and gluing map}
\begin{equation*}
\phi_G^{trop}\mathrel{\mathop:}\prod_{v\in V(G)}\overline{M}_{h(v),\mathcal{A}(v)}^{trop}\longrightarrow \overline{M}_{g,\mathcal{A}}^{trop}
\end{equation*}
given by connecting two legs of tropical curves $[\Gamma_v]\in \overline{M}_{h(v),\mathcal{A}(v)}^{trop}$ by a bridge at infinity whenever the corresponding flags in $G$ are connected by an edge.
\end{definition}
\begin{proposition}\label{prop_tropicalclutching&gluing}
The natural diagram
\begin{equation*}\begin{CD}
\prod_{v\in V(G)}\overline{M}_{h(v),\mathcal{A}(v)}^{an}@>\prod\trop_{h(v),\mathcal{A}(v)}>> \prod_{v\in V(G)}\overline{M}_{h(v),\mathcal{A}(v)}^{trop}\\
@V\phi_G^{an}VV @VV\phi_G^{trop}V\\
\overline{M}_{g,\mathcal{A}}^{an}@>\trop_{g,\mathcal{A}}>>\overline{M}_{g,\mathcal{A}}^{trop}
\end{CD}\end{equation*}
is commutative
\end{proposition}
\begin{proof}
Let $(\mathcal{C}^v,p_1^v,\ldots,p_{n_v}^v)$ be families of $\big(h(v),\mathcal{A}(v)\big)$ stable curves over a valuation ring $R$ extending $k$ and denote the tropical curves associated to this data by $\Gamma_v$. Observe that the clutching and gluing map $\phi_G$ applied to the $(\mathcal{C}^v,p_1^v,\ldots,p_{n_v}^v)$ exactly corresponds to connecting two legs of the $\Gamma_v$, whenever they correspond to an edge in $G$. Since $\phi_G\big((\mathcal{C}^v,p_1^v,\ldots,p_{n_v}^v)\big)$ has a node over all of $\Spec R$, whenever two marked points have been glued, the special fiber is given by $xy=0$ in formal coordinates and therefore the connecting edge in $\phi_G^{trop}\big((\Gamma_v)\big)$ has to be of infinite length.
\end{proof}
As special cases of Definition \ref{def_tropicalclutching&gluing} we obtain the following two maps:
\begin{itemize}
\item The \emph{tropical clutching map}
\begin{equation*}
\kappa^{trop}=\kappa_{g_1,\mathcal{A}_1,g_2,\mathcal{A}_2}^{trop}\mathrel{\mathop:} \overline{M}_{g_1,\mathcal{A}_1\cup\{1\}}^{trop}\times\overline{M}_{g_2,\mathcal{A}_2\cup\{1\}}^{trop}\longrightarrow \overline{M}_{g,\mathcal{A}}^{trop}
\end{equation*}
is given by sending a pair of extended tropical curves $\Gamma_1$ and $\Gamma_2$ in $\overline{M}_{g_1,\mathcal{A}_1\cup\{1\}}^{trop}$ and $\overline{M}_{g_2,\mathcal{A}_2\cup\{1\}}^{trop}$ respectively to the extended tropical curve $\Gamma$ that is obtained by connecting the two legs $l_{n_1+1}$ and $l_{n_2+1}$ at infinity.
\item The \emph{tropical gluing map}
\begin{equation*}
\gamma^{trop}\mathrel{\mathop:}\overline{M}^{trop}_{g-1,\mathcal{A}\cup\{1,1\}}\longrightarrow\overline{M}^{trop}_{g,\mathcal{A}}
\end{equation*}
is defined by sending an extended tropical curve $\Gamma$ in $\overline{M}_{g-1,\mathcal{A}\cup\{1,1\}}$ to the tropical curve $\widetilde{\Gamma}$ obtained by connecting the two legs $l_{n+1}$ and $l_{n+2}$ at infinity.
\end{itemize}
Using this notation Proposition \ref{prop_tropicalclutching&gluing} yields the following generalization of \cite[Theorem 1.2.2]{AbramovichCaporasoPayne_tropicalmoduli}:
\begin{corollary}\label{cor_tropicalclutching&gluing}
The natural diagrams
\begin{equation*}\begin{CD}
\overline{M}_{g_1,\mathcal{A}_1\cup\{1\}}^{an}\times\overline{M}_{g_2,\mathcal{A}_2\cup\{1\}}^{an}@>\trop_{g_1,\mathcal{A}_1\cup\{1\}}\times\trop_{g_2,\mathcal{A}_2\cup\{1\}}>>\overline{M}_{g_1,\mathcal{A}_1\cup\{1\}}^{trop}\times\overline{M}_{g_2,\mathcal{A}_2\cup\{1\}}^{trop}\\
@V\kappa^{an}VV @VV k^{trop}V\\
\overline{M}_{g,\mathcal{A}}^{an} @>\trop_{g,\mathcal{A}}>> \overline{M}_{g,\mathcal{A}}^{trop}
\end{CD}\end{equation*}
and
\begin{equation*}\begin{CD}
\overline{M}_{g-1,\mathcal{A}\cup\{1,1\}}^{an}@>\trop_{g-1,\mathcal{A}\cup\{1,1\}}>>\overline{M}_{g-1,\mathcal{A}\cup\{1,1\}}^{trop}\\
@V\gamma^{an}VV @VV\gamma^{trop}V\\
\overline{M}_{g,\mathcal{A}}^{an}@>\trop_{g,\mathcal{A}}>>\overline{M}_{g,\mathcal{A}}^{trop}
\end{CD}\end{equation*}
are commutative.
\end{corollary}
\section{Variations of weight data}\label{section_wallsandchambers}
\subsection{Chamber decompositions}
In this section we compare how, given a fixed genus $g$, the two functions $\mathcal{A}\mapsto\overline{\mathcal{M}}_{g,\mathcal{A}}$ and $\mathcal{A}\mapsto\overline{M}_{g,\mathcal{A}}^{trop}$ vary in $\mathcal{A}$. Fix a genus $g\geq 0$ and a number $n\geq 0$. We denote by $\mathscr{D}_{g,n}$ the set of possible weight data
\begin{equation*}
\mathscr{D}_{g,n}=\big\{(a_1,\ldots,a_n)\in(0,1]^n\cap\mathbb{Q}^n\big\vert a_1+\ldots + a_n>2-2g \big\} \ .
\end{equation*}
As in \cite[Section 5]{Hassett_weightedmoduli} a \emph{chamber decomposition} $\mathscr{W}$ of $\mathscr{D}_{g,n}$ consists of a finite set of hyperplanes $w_S\subseteq\mathscr{D}_{g,n}$. We refer to the $w_S$ as the \emph{walls} of the chamber decomposition $\mathscr{W}$ and to connected components of the complement of the $w_S$ as the \emph{open chambers} of $\mathscr{W}$.
In \cite[Section 5]{Hassett_weightedmoduli} Hassett studies chamber decompositions of $\mathscr{D}_{g,n}$ with the property that the functions $\mathcal{A}\mapsto\overline{\mathcal{M}}_{g,\mathcal{A}}$ and $\mathcal{A}\mapsto\mathcal{C}_{g,\mathcal{A}}$ are constant on every open chamber, where $\mathcal{C}_{g,\mathcal{A}}$ denotes the universal curve of $\overline{\mathcal{M}}_{g,\mathcal{A}}$.
\begin{proposition}\label{prop_algcham>=tropcham}
Suppose that $\mathscr{W}$ is a chamber decomposition of $\mathscr{D}_{g,n}$ such that $\mathcal{A}\mapsto\mathcal{C}_{g,\mathcal{A}}$ is constant on open chambers. Then $\mathcal{A}\mapsto\overline{M}_{g,\mathcal{A}}^{trop}$ is constant on the open chambers of $\mathscr{W}$ as well.
\end{proposition}
\begin{proof}
If the two universal curves $\mathcal{C}_{g,\mathcal{A}}$ and $\mathcal{C}_{g',\mathcal{A}'}$ are isomorphic, there is an isomorphism between the moduli stacks $\overline{\mathcal{M}}_{g,\mathcal{A}}$ and $\overline{\mathcal{M}}_{g',\mathcal{A}'}$ that preserves the stratifications by dual graphs. Using Theorem \ref{thm_modulartrop} we see that the tropical moduli spaces $\overline{M}_{g,\mathcal{A}}^{trop}$ and $\overline{M}_{g',\mathcal{A}'}^{trop}$ are isomorphic.
\end{proof}
Moreover Hassett considers the \emph{coarse chamber decomposition}, which is given by
\begin{equation*}
\mathscr{W}_c=\Big\{\sum_{j\in S}a_j=1\big\vert S\subseteq\{1,\ldots, n\} \textrm{ and } 2<\vert S\vert \leq n-2\delta_{g,0}\Big\} \ ,
\end{equation*}
as well as the \emph{fine chamber decomposition}, which is given by
\begin{equation*}
\mathscr{W}_f=\Big\{\sum_{j\in S}a_j=1\big\vert S\subseteq\{1,\ldots, n\} \textrm{ and } 2\leq\vert S\vert \leq n-2\delta_{g,0}\Big\} \ .
\end{equation*}
Hereby $\delta_{i,j}$ denotes the Kronecker delta
\begin{equation*}
\delta_{i,j}=\left\{
\begin{array}{l l}
1 & \quad \text{if } i=j\\
0 & \quad \text{else.}\\
\end{array} \right .
\end{equation*}
In \cite[Proposition 5.1]{Hassett_weightedmoduli} Hassett shows that $\mathscr{W}_c$ is the coarsest chamber decompostion of $\mathscr{D}_{g,n}$ such that $\mathcal{A}\mapsto\overline{\mathcal{M}}_{g,\mathcal{A}}$ is constant on every open chamber and $\mathscr{W}_f$ is the coarsest chamber decomposition of $\mathscr{D}_{g,n}$ such that the map $\mathcal{A}\mapsto\mathcal{C}_{g,\mathcal{A}}$ is constant. Therefore Proposition \ref{prop_algcham>=tropcham} immediately implies that the association $\mathcal{A}\mapsto \overline{M}_{g,\mathcal{A}}^{trop}$ is constant on the fine chambers of $\mathscr{D}_{g,n}$.
The following Proposition \ref{prop_algfinecham=tropcoarsecham} is a partial analogue of \cite[Proposition 5.1]{Hassett_weightedmoduli} in the tropical world.
\begin{proposition}\label{prop_algfinecham=tropcoarsecham}
The fine chamber decomposition $\mathscr{W}_f$ is the coarsest chamber decomposition of $\mathscr{D}_{g,n}$ such that $\mathcal{A}\mapsto\overline{M}_{g,\mathcal{A}}^{trop}$ is constant on every open chamber.
\end{proposition}
\begin{proof}
The map $\mathcal{A}\mapsto \overline{M}_{g,\mathcal{A}}^{trop}$ is constant on the fine chambers of $\mathscr{D}_{g,n}$. So it is enough to note that $\overline{M}_{g,\mathcal{A}}^{trop}$ changes whenever we cross a wall of the fine chamber decomposition $\mathscr{W}_f$.
Suppose first that $g\geq 1$. Let $S\subseteq\{1,\ldots, n\}$ with $2\leq\vert S\vert\leq n$. Consider the graph $G_S$ containing one edge between two vertices $v_0$ and $v_g$, one of weight $0$ and the other of weight $g$, and legs $l_i$ incident to $v_0$, whenever $i\in S$ and incident to $v_g$, whenever $i\notin S$.
\begin{center}\begin{tikzpicture}
\fill (0,0) circle (0.05 cm);
\fill (3,0) circle (0.05 cm);
\draw (0,0) -- (3,0);
\foreach \a in {120,150,180,210,240}
\draw (\a:0) -- (\a:2);
\foreach \a in {300,330,0,30,60}
\draw (3,0) -- ({3+2*cos(\a)},{2*sin(\a)});
\node at (0.2,0.5) {$0$};
\node at (0.2,-0.5) {$v_0$};
\node at (2.8,0.5) {$g$};
\node at (2.8,-0.5) {$v_g$};
\node at (-3,0) {$i\in S$};
\node at (6,0) {$i\notin S$};
\end{tikzpicture}\end{center}
If $\sum_{i\in S}a_i>1$, then $G_S$ is stable of type $(g,\mathcal{A})$, since $\vert v_0\vert=1+\sum_{i\in S}a_i>2$ and $h(v_g)\geq 1$. But if $\sum_{i\in S}a_i\leq 1$, the graph $G_S$ is not stable of type $(g,\mathcal{A})$.
Consider now the case $g=0$. Let $S\subseteq\{1,\ldots,n\}$ with $2\leq\vert S\vert\leq n-2$ and consider again the same graph $G_S$ as above with two vertices of weight $0$ connected by an edge and legs incident to $v_i$ depending on whether they are in $S$ or not.
\begin{center}\begin{tikzpicture}
\fill (0,0) circle (0.05 cm);
\fill (3,0) circle (0.05 cm);
\draw (0,0) -- (3,0);
\foreach \a in {120,150,180,210,240}
\draw (\a:0) -- (\a:2);
\foreach \a in {300,330,0,30,60}
\draw (3,0) -- ({3+2*cos(\a)},{2*sin(\a)});
\node at (0.2,0.5) {$0$};
\node at (2.8,0.5) {$0$};
\node at (-3,0) {$i\in S$};
\node at (6,0) {$i\notin S$};
\end{tikzpicture}\end{center}
Suppose that $\sum_{i\in S}a_i\leq 1$. Then $\sum_{i=1}^{n} a_i>2$ implies $\sum_{i\notin S}a_i>1$. So when crossing the wall $\sum_{i\in S}a_i=1$ without changing the $a_i$ with $i\notin S$ we obtain that $G_S$ is stable of type $(0,\mathcal{A})$, if $\sum_{i\in S}a_i>1$, and not, if $\sum_{i\in S}a_i\leq 1$.
\end{proof}
\begin{remarks}\begin{enumerate}[(i)]
\item As noted in \cite[Remark 2.3]{AlexeevGuy_weightedstablemaps} there is a typographical error in the definitions of coarse and fine chamber decompositions in \cite[Section 5]{Hassett_weightedmoduli}. We fix this typo following the notation of \cite[Section 0.4]{BayerManin_weightedstablemaps}.
\item In \cite[Definition 2.8]{AlexeevGuy_weightedstablemaps} Alexeev and Guy propose an alternative to chamber decompositions: They associate a simplicial complex $\Delta_\mathcal{A}$ to the weights $\mathcal{A}$ that has a $\vert S\vert$-dimensional simplex for every subset $S\subseteq\{1,\ldots, n\}$ with $\sum_{i\in S}a_i\leq 1$. As seen in \cite[Section 4]{AlexeevGuy_weightedstablemaps} crossing a single wall from $\mathcal{A}$ to $\mathcal{B}$ with $\mathcal{B}\geq \mathcal{A}$ in $\mathscr{W}_{g,n}$ corresponds to adding a simplex to $\Delta_\mathcal{A}$ in order to obtain $\Delta_\mathcal{B}$.
\end{enumerate}\end{remarks}
\subsection{Kapranov's construction of $\overline{M}_{0,n}$}
By Theorem \ref{prop_tropicalreduction&forgetful} we can realize $\overline{M}_{g,n}$ as a composition of tropical reduction maps starting at a $\overline{M}_{g,\mathcal{A}}$ with a low weights $\mathcal{A}$ and Proposition \ref{prop_reduction=subcomplex} ensures that while we are crossing a wall in $\mathscr{W}_f$ from lower weight data to bigger ones, we are only adding additional extended cones.
In \cite[Section 6.1]{Hassett_weightedmoduli} Hassett identifies Kapranov's classical blow-up construction of $\overline{M}_{0,n}$ (see \cite{Kapranov_ChowquotientsI} and \cite[Section 4.3]{Kapranov_Veronese&M0nbar}) with a sequence of reduction maps. The weights of this sequence are given by
\begin{equation*}
\mathcal{A}_{r,s}[n]=\Bigg(\underbrace{\frac{1}{n-r-1},\ldots,\frac{1}{n-r-1}}_{n-r-1 \textrm{ times}},\frac{s}{n-r-1},\underbrace{1,\ldots,1}_{r \textrm{ times}}\Bigg)
\end{equation*}
for $r=1,\ldots, n-3$ and $s=1,\ldots, n-r-2$. The sequence starts with $\overline{M}_{0,\mathcal{A}_{1,1}[n]}\simeq\mathbb{P}^{n-3}$, at the $r$-th step the sequence of reduction maps is given by
\begin{equation*}\begin{CD}
\overline{M}_{0,\mathcal{A}_{r,n-r-2}[n]}@>>> \ldots @>>>\overline{M}_{0,\mathcal{A}_{r,2}[n]}@>>>\overline{M}_{0,\mathcal{A}_{r,1}[n]} \ ,
\end{CD}\end{equation*}
and Kapranov has shown that at the last step $\overline{M}_{0,\mathcal{A}_{n-3,1}[n]}$ is isomorphic to $\overline{M}_{0,n}$.
\begin{example}
The final weights in Hassett's interpretation of Kapranov's construction are given by
\begin{equation*}
\mathcal{A}_{n-3,1}[n]=\Bigg(\frac{1}{2},\frac{1}{2},\frac{1}{2},\underbrace{1,\ldots,1}_{n-3 \textrm{ times}}\Bigg) \ .
\end{equation*}
Let $n\geq 5$. As seen in Figure \ref{figure_counterexKapranov}, there is a rational weighted graph that is stable of type $\big(0,(1,\ldots,1)\big)$ but not of type $\big(0,\mathcal{A}_{n-3,1}[n]\big)$. Therefore the tropical moduli space $\overline{M}_{0,n}^{trop}$ contains an extended cone corresponding to this graph, but $\overline{M}_{0,\mathcal{A}_{n-3,1}}^{trop}$ does not and thus these two spaces cannot be isomorphic.
\end{example}
\begin{figure}
\begin{center}\begin{tikzpicture}
\fill (0,1) circle (0.05 cm);
\fill (-1,0) circle (0.05 cm);
\fill (1,0) circle (0.05 cm);
\fill (2,1,0) circle (0.05 cm);
\fill (7,0) circle (0.05 cm);
\fill (6,1) circle (0.05 cm);
\draw (-1,0) -- (0,1);
\draw (1,0) -- (0,1);
\draw (0,1) -- (0,2);
\draw (1,0) -- (1,-1);
\draw (-1,0) -- (-1,-1);
\draw (1,0) -- (2,1);
\draw (-1,0) -- (-1.7,0.7);
\draw (2,1) -- (2,2);
\draw (2,1) -- (2.7,0.3);
\draw (5.3,0.3) -- (6,1);
\draw (6,1) -- (6,2);
\draw (6,1) -- (7,0);
\draw (7,0) -- (7,-1);
\draw (7,0) -- (7.7,0.7);
\node at (0,2.5) {$l_3$};
\node at (2,2.5) {$l_5$};
\node at (1,-1.5) {$l_4$};
\node at (-1,-1.5) {$l_1$};
\node at (-2.1,1.1) {$l_2$};
\node at (6,2.5) {$l_{n-2}$};
\node at (7,-1.5) {$l_{n-1}$};
\node at (8.1,1.1) {$l_n$};
\node at (4,0.5) {$\ldots$};
\end{tikzpicture}\end{center}
\caption{A graph with $0$ vertex weights that is stable of type $\big(0,(1,\ldots,1)\big)$ but not of type $\big(0,\mathcal{A}_{n-3,1}[n]\big)$, whenever $n\geq 5$.}\label{figure_counterexKapranov} \end{figure}
The explanation for this behavior is that, although we have an isomorphism
\begin{equation*}
\overline{M}_{0,\mathcal{A}_{n-3,1}[n]}\simeq\overline{M}_{0,n} \ ,
\end{equation*}
the universal curves $\mathcal{C}_{0,\mathcal{A}_{n-3,1}[n]}$ and $\mathcal{C}_{0,n}$ of these two moduli spaces are not isomorphic. In other words $M_{0,\mathcal{A}_{n-3,1}[n]}$, the locus parametrizing smooth curves in $\overline{M}_{0,\mathcal{A}_{n-3,1}[n]}$, is bigger than $M_{0,n}$. One can, of course, deal with this phenomenon by further increasing the weights and, by the following example, a minimal increase will be enough.
\begin{example}
Let us now consider the weights
\begin{equation*}
\mathcal{A}_\epsilon=\Bigg(\frac{1}{2}+\epsilon,\frac{1}{2}+\epsilon,\frac{1}{2}+\epsilon,\underbrace{1,\ldots,1}_{n-3 \textrm{ times}}\Bigg) \ .
\end{equation*}
for $0<\epsilon\leq\frac{1}{2}$. Every rational stable tropical curve is also stable of type $\mathcal{A}_\epsilon$ and the tropical reduction map therefore induces a natural isomorphism
\begin{equation*}
\overline{M}_{0,n}^{trop}\simeq\overline{M}_{0,\mathcal{A}_\epsilon}^{trop} \ .
\end{equation*}
This, together with the fact that for every rational $\mathcal{A}_\epsilon$-stable $n$-marked curve none of the marked points are allowed to coincide, shows that there is also a natural isomorphism
\begin{equation*}
\mathcal{C}_{0,n}\simeq\mathcal{C}_{0,\mathcal{A}_\epsilon}
\end{equation*}
between the universal curves.
\end{example}
\section{Losev-Manin spaces}\label{section_LosevManin}
Let $g=0$. We are now going to consider the special case that the weights $\mathcal{A}=\{a_0,\ldots,a_{n+1}\}$ for $n\geq 1$ fulfill the two conditions:
\begin{enumerate}[(i)]
\item\label{itemi} $a_0+a_i > 1$ and $a_{n+1}+a_i>1$ for each $i\in\{1,\ldots,n\}$, and
\item\label{itemii} $a_{i_1}+\ldots +a_{i_r}\leq 1$ for all $\{j_1,\ldots,j_r\}\subseteq\{1,\ldots,n\}$.
\end{enumerate}
We begin with the following easy observation:
\begin{lemma}
For $n\geq 2$ and $g=0$ there is a unique fine chamber in $\mathscr{D}_{0,n+2}$ determined by the conditions \eqref{itemi} and \eqref{itemii} above.
\end{lemma}
\begin{proof}
We have to show that for every $S\subseteq\{0,\ldots, n+1\}$ with $2\leq\vert S\vert \leq n$ the above conditions imply that either $\sum_{i\in S}a_i\leq 1$ or $\sum_{i\in S}a_i>1$.
If $0\notin S$ and $n+1\in S$ or if $0\in S$ and $n+1\notin S$ Condition \eqref{itemi} immediately yields $\sum_{i\in S}Sa_i>1$, since $a_i\geq 0$ for all $i$. In the case that both $0\notin S$ and $n+1\notin S$ we have $\sum_{i\in S}a_i\leq 1$ by condition \eqref{itemii}. Now consider the case $S\supseteq\{0,n+1\}$. By Condition \eqref{itemi} we obtain
\begin{equation*}
a_0+a_i+a_j+a_{n+1}>2
\end{equation*}
for some $1\leq i,j\leq n$. Since $n\geq 2$ we may assume $i< j$. It follows from Condition \eqref{itemii} that $a_i+a_j\leq 1$ and therefore we obtain $a_0+a_{n+1}>1$, which, in turn, implies
\begin{equation*}
\sum_{i\in S}a_i\geq a_0+a_{n+1}>1 \ ,
\end{equation*}
since $a_i\geq 0$ for all $i$.
A tuple of weights that fulfills Conditions \eqref{itemi} and \eqref{itemii} is e.g. given by
\begin{equation*}
\mathcal{A}=\bigg(1,\frac{1}{n},\ldots,\frac{1}{n},1\bigg)
\end{equation*}
and therefore this chamber is non-empty.
\end{proof}
In \cite[Section 6.4]{Hassett_weightedmoduli} Hassett has identified the fine moduli spaces $\overline{M}_{0,\mathcal{A}}$ with the moduli spaces $L_n$, studied by Losev and Manin in \cite{LosevManin_newmoduli}, parametrizing chains $C$ of projective lines connecting the two end points $p_0$ and $p_{n+1}$ with $n$ additional marked points $p_1,\ldots, p_n$ such that
\begin{itemize}
\item the $p_1,\ldots,p_n$ are allowed to mutually coincide, but not to coincide with $p_0$, $p_{n+1}$, or the nodes, and
\item the normalization of every component of $C$ contains at least three special points.
\end{itemize}
By \cite{LosevManin_newmoduli} the moduli space $L_n$ is isomorphic to the smooth projective toric variety defined by the $(n-2)$-dimensional permutohedron $P_{n-1}$ as defined in \cite[Definition 1.3]{Kapranov_permutohedron} and the big torus $T=\mathbb{G}_m^{n-1}$ exactly parametrizes the locus of smooth curves in $L_n$.
\begin{example}\label{example_L_3nonSNC}
Let us now consider the case $n=3$. We may choose coordinates $(p,q)$ on $T\simeq\mathbb{G}_m^2$ such that up to $\mathbb{P}^1$-automorphism $(p_0,p_1,p_{4})=(0,1,\infty)$ and $(p,q)=(p_2,p_3)$ are free variables. In these coordinates the toric prime divisors are given by $p=0$, $p=\infty$, $q=0$, $q=\infty$, $p=q=0$, as well as $p=q=\infty$, and the toric boundary has normal crossings. The complement of $\overline{M}_{0,5}$ in $L_3$, however, also contains the divisors $p=1$, $q=1$, and $p=q$, which all intersect at $(1,1)$. Therefore $L_3-\overline{M}_{0,5}$ does not have normal crossings, as indicated in Remark \ref{remark_nonNC} above.
\begin{center}\begin{tikzpicture}
\draw (-0.5,1.5) -- (1.5,-0.5);
\draw (0,0.5) -- (0,6.5);
\draw (-0.5,6) -- (5.5,6);
\draw (4.5,6.5) -- (6.5,4.5);
\draw (6,5.5) -- (6,-0.5);
\draw (0.5,0) -- (6.5,0);
\draw[dashed] (-0.3,2) -- (6.3,2);
\draw[dashed] (2,-0.3) -- (2,6.3);
\draw[dashed] (0.3,0.3) -- (5.7,5.7);
\node at (3.5,6.3) {$q=\infty$};
\node at (4,-0.3) {$q=0$};
\node[rotate=270] at (6.3,3.5) {$p=\infty$};
\node[rotate=90] at (-0.3,4) {$p=0$};
\node[rotate=-45] at (6,6) {$p=q=\infty$};
\node[rotate=-45] at (0,0) {$p=q=0$};
\node[rotate=90] at (1.7,4) {$p=1$};
\node at (4,1.7) {$q=1$};
\node[rotate=45] at (4.3,3.7) {$p=q$};
\end{tikzpicture}\end{center}
\end{example}
Recall from \cite[Definition 1.3]{Kapranov_permutohedron} that the \emph{$(n-1)$-dimensional permutohedron} $P_n$ is the lattice polytope in $\mathbb{R}^n$ given as the convex hull of the points $\big(s(1),\ldots,s(n)\big)$, where $s$ runs through all elements in the symmetric group $S_n$. As explained in \cite[Definition 2.5.1]{LosevManin_newmoduli} the $l$-dimensional cones of its dual fan $\Delta_n$ are labelled by $(l+1)$-partitions of $\{1,\ldots,n\}$ and therefore naturally carries the structure of a moduli space $L_n^{trop}$ of stable \emph{rational tropical chain curves} with $n+2$ legs connecting the legs $l_0$ and $l_{n+1}$. Its canonical compactification $\overline{L}_n^{trop}$ parametrizes stable \emph{rational extened tropical chain curves} with $n+2$ legs connecting the legs $l_0$ and $l_{n+1}$.
\begin{figure}
\centering\begin{tikzpicture}
\fill (0,1) circle (0.05 cm);
\fill (-1,0) circle (0.05 cm);
\fill (1,0) circle (0.05 cm);
\fill (3,0) circle (0.05 cm);
\fill (2,1) circle (0.05 cm);
\fill (4,1) circle (0.05 cm);
\fill (5,0) circle (0.05 cm);
\draw (-1,0) -- (0,1);
\draw (1,0) -- (0,1);
\draw (0,1) -- (0,2);
\draw (1,0) -- (1,-1);
\draw (-1,0) -- (-1,-1);
\draw (5,0) -- (6,1);
\draw (-1,0) -- (-2,1);
\draw (1,0) -- (2,1);
\draw (2,1) -- (3,0);
\draw (2,1) -- (2,2);
\draw (3,0) -- (3,-1);
\draw (3,0) -- (4,1);
\draw (4,1) -- (5,0);
\draw (4,1) -- (4,2);
\draw (5,0) -- (5,-1);
\node at (0,2.5) {$l_2$};
\node at (6.5,1.5) {$l_8$};
\node at (3,-1.5) {$l_5$};
\node at (-1,-1.5) {$l_1$};
\node at (-2.5,1.5) {$l_0$};
\node at (1,-1.5) {$l_3$};
\node at (2,2.5) {$l_4$};
\node at (4,2.5) {$l_6$};
\node at (5,-1.5) {$l_7$};
\end{tikzpicture}
\caption{A stable rational tropical chain curve with nine legs.}
\end{figure}
\begin{figure}
\centering
\begin{tikzpicture}
\foreach \a in {60,120,180,240,300,360}
\draw (0,0) -- (\a:2);
\foreach \r in {0.33,0.66,1,1.33,1.66}
\foreach \a in {60,120,180,240,300,360}
\draw (\a:\r) -- (\a+60:\r);
\node at (30:2.5) {$(1,2,3)$};
\node at (90:2.5) {$(2,1,3)$};
\node at (150:2.5) {$(2,3,1)$};
\node at (210:2.5) {$(3,2,1)$};
\node at (270:2.5) {$(3,1,2)$};
\node at (330:2.5) {$(1,3,2)$};
\end{tikzpicture}
\caption{The tropical Losev-Manin space $L_3^{trop}$. We indicate the corners of the permutohedron $P_{3}$ dual to $2$-dimensional cones.}
\end{figure}
Moreover, there is a natural set-theoretic tropicalization map
\begin{equation*}
\trop_n\mathrel{\mathop:}L_n^{an}\longrightarrow\overline{L}_n^{trop}
\end{equation*}
defined analogously to $\trop_{g,\mathcal{A}}$ in the introduction.
Now recall that Kajiwara \cite[Section 1]{Kajiwara_troptoric} and, independently, Payne \cite[Section 3]{Payne_anallimittrop} construct a tropicalization map
\begin{equation*}
\trop_{\Delta}\mathrel{\mathop:}X(\Delta)^{an}\longrightarrow N_\mathbb{R}(\Delta)
\end{equation*}
into a partial compactification $N_\mathbb{R}(\Delta)$ of $N_\mathbb{R}$ for all toric varieties $X=X(\Delta)$ defined by a rational polyhedral fan $\Delta$ in $N_\mathbb{R}$, where $N$ denotes the cocharacter lattice of the big torus $T$ of $X$. On an $T$-invariant open affine subset $U=\Spec k[P]$ the partial compactifcation $N_\mathbb{R}(P)$ of $N_\mathbb{R}$ is given by $\Hom(P,\overline{\mathbb{R}})$ and $\trop_P$ by
\begin{equation*} \begin{split}
\trop_P\mathrel{\mathop:}U^{an}&\longrightarrow N_\mathbb{R}(P)=\Hom(P,\overline{\mathbb{R}})\\
x&\longmapsto \big(p\mapsto -\log\vert p\vert_x\big)
\end{split}\end{equation*}
We also refer the reader to \cite[Section 3]{Rabinoff_newtonpolygon} for further details on this construction.
\begin{corollary}
There is a natural homeomorphism $J_n\mathrel{\mathop:}\overline{L}_n^{trop}\xrightarrow{\sim}N_\mathbb{R}(\Delta_n)$ such that the diagram
\begin{center}\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=2em,column sep=3em,minimum width=2em]
{
&L_n^{an} & \\
N_\mathbb{R}(\Delta_n) & & \overline{L}_n^{trop} \\
};
\path[-stealth]
(m-1-2) edge node [above left] {$\trop_{\Delta_n}$} (m-2-1)
edge node [above right] {$\trop_n$} (m-2-3)
(m-2-3) edge node [below] {$J_n$} node [above] {$\sim$} (m-2-1);
\end{tikzpicture}\end{center}
commutes.
\end{corollary}
\begin{proof}
In view of \cite[Theorem 1.2]{Ulirsch_functroplogsch} and \cite[Proposition 7.1]{Ulirsch_functroplogsch} we can naturally identify $\trop_{\Delta_n}$ with the deformation retraction $\mathbf{p}$, since $L_n$ is proper over $k$ and thus the fan $\Delta_n$ is complete. Then the statement immediately follows from Theorem \ref{thm_modulartrop}.
\end{proof}
\bibliographystyle{amsalpha}
|
\section{Introduction} \label{intro}
An important problem in the statistical study of chaotic dynamical systems is obtaining a quantitative estimate on the rate of decay of correlations of the system. Such an estimate describes how fast the system looses memory of its past and opens the door to further statistical description of the system. Ideally, one would like to prove an exponential rate of mixing for systems with ``enough'' hyperbolicity.
The first results on exponential decay of correlations were obtained for uniformly expanding maps and hyperbolic maps (see \cite{Liv95} and references therein). Slower rates of mixing were also obtained for non-uniformly hyperbolic maps \cite{You98,You99,Sar02}.
For flows most of the existing results on exponential decay of correlations pertain to smooth systems or those with a Markov structure (see \cite{Dol98,Liv04,BalVal05,AviGouYoc06} and references therein). For systems with singularities, Chernov \cite{Che07} obtained a stretched-exponential rate of decay for certain Billiard flows, Baladi and Liverani \cite{BalLiv12} for piecewise cone-hyperbolic contact flows, while Obayashi \cite{Oba09} obtained exponential decay of correlations for suspension semiflows over piecewise expanding $\sC^2$ maps of the interval using a tower construction and applying the main result of \cite{BalVal05}.
The goal of this article is to introduce a method by which rates of decay of correlations can be obtained for systems with a neutral direction that have discontinuities and are of low regularity (without assuming the existence of a Markov structure). Such systems appear in practice and are of physical relevance. Indeed, the flow of the Lorenz system of ordinary differential equations (see \cite{But14}) and Billiard flows are examples of such systems. Our motivation is to put forward a method to eventually prove exponential mixing rates for these systems; however, in this article we consider the simplest case -- that of a skew product with a neutral direction. Also, we will prove only a stretched-exponential bound; however, with more delicate estimates one should be able to obtain an exponential bound. We will illustrate the method by considering a 2D skew-product map with an expanding piecewise $\sC^{1+\alpha}$ map in the base and a neutral direction on which the map is a rigid rotation. The method proposed here is a combination of the point of view of standard pairs due to Dolgopyat \cite{Dol04} and further developed by Chernov \cite{CheMar06,CheDol08,CheDol09}, and the oscillatory cancelation mechanism due to Dolgopyat \cite{Dol98}.
The skew product is introduced in \Cref{setting}. In \Cref{corr} the main theorem is introduced and proven assuming a crucial estimate. The rest of the article is devoted to the proof of this estimate. \Cref{iter} introduces the terminology of standard pairs and standard families. In section \Cref{trans} we introduce the notion of transversality. \Cref{modi} shows how one can use the oscillatory cancelation mechanism of Dolgopyat to modify standard families. Finally in \Cref{dens_decay} the main estimate is proven.
\section{The setting} \label{setting}
Consider the skew-product $F:\bT^{2}\to \bT^{2}$ defined by
\begin{equation} \label{eq:skewprod}
F:(x,y) \mapsto \left( f(x), y+\tau(x) \right).
\end{equation}
Assume that
\begin{equation}\label{eq:C1+}
f:\bT^1 \to \bT^1 \text{ is piecewise } \sC^{1+\alpha}.
\end{equation}
That is, $\bT^1$ can be partitioned into countably many open intervals (modulo a countable set of endpoints of the intervals) such that each open interval is a maximal interval of monotonicity of $f$ and $f$ is $\sC^{1+\alpha}$ on the open interval, extendable to the closed interval. Note that, for every $n \geq 1$, $f^n$ is also piecewise $\sC^{1+\alpha}$. Having the graph of $f^n$ in mind, we denote by $\cH^n$ the set of inverse branches of $f^n$. We choose to index partition elements of $f^n$ by elements of $\cH^n$. So, we denote the partition of $f^n$ by $\{O_h\}_{h \in \cH^n}$. Note that the domain and range of $h$ are $f^n(O_h)$ and $O_h$, respectively.
Assume that $f$ is expanding. That is, there exist $\lambda$ such that
\begin{equation} \label{eq:expansion}
\ln 2 < \lambda \text{ and } e^{\lambda n} \leq \abs{(f^n)'} .
\end{equation}
Assume that $f$ satisfies the following distortion bound. There exists a constant $D \geq 0$ such that for every $n \in \bN$
\begin{equation}\label{eq:dist}
\frac{\abs{h'(x)}}{\abs{h'(y)}} \leq e^{D\abs{x-y}^\alpha} \text{, for } h \in \cH^n\text{, and } x,y \in f^n(O_h).
\end{equation}
Note that this condition is implied by a similar condition for the first iterate of $f$, possibly with slightly worse constants.
Also, assume that, for every $n \in \bN$,
\begin{equation} \label{eq:summability}
\sum_{h \in \cH^n} \sup_{f^n(O_h)} \abs{h'}< \infty.
\end{equation}
This condition is trivially satisfied when $f$ has finitely many branches. This condition is also implied by the similar condition for the first iterate. Indeed, if $\sum_{h \in \cH} \sup_{f(O_h)} \abs{h'} \leq M < \infty$, then $\sum_{h \in \cH^n} \sup_{f^n(O_h)} \abs{h'}\leq M^n <\infty$.
Assume also that $f$ is covering.
\footnote{If $f$ is an expanding piecewise $\sC^2$ map on a finite partition, then covering and mixing are equivalent, see \cite{HofKel82_1,HofKel82_2}.}
That is,
\begin{equation} \label{eq:covering}
\forall n \in \bN, \exists N(n) \text{ such that } \forall h \in \cH^n, f^{N(n)}\bar O_h = \bT^1,
\end{equation}
where $\bar O_h$ denotes the closure of the open interval $O_h$.
Assume that
\begin{equation}
\tau:\bT^1 \to \bR \text{ is piecewise } \sC^1.
\end{equation}
Assume that $\tau$ is not Lipschitz-cohomologous to a piecewise constant function. That is,
\begin{equation} \label{eq:ncoh_pwc}\nexists \phi \in Lip (\bT^1, \bR) \text{ such that } \tau(x) = \phi \circ f(x) - \phi(x) + \psi(x),\end{equation}where $\psi(x)$ is a piecewise constant function on the joint partition of $f$ and $\tau$.
For every $n \in \bN$, define $\tau_n := \sum_{j=0}^{n-1} \tau \circ f^j$. Assume that there exists $C_\tau$ such that for every $n \in \bN$
\begin{equation} \label{eq:roof_regularity}
\abs{(\tau_n \circ h)'} \leq C_\tau \text{, for } h \in \cH^n.
\end{equation}
This condition is easily satisfied if $|\tau'|$ is bounded. Note that in more general cases mentioned earlier, e.g. the case of the Lorenz flow, this condition is not satisfied. For further details see \cite{But14}.
\textbf{Banach space assumptions.}
Let $\norm{\cdot}_{\sC^\alpha}= \abs{\cdot}_\alpha + \norm{\cdot}_{\sC^0}$, where $\abs{\cdot}_\alpha$ is the usual H\"older semi-norm. Suppose there exists a Banach space $\B \subset \L^{1}$ such that the following hold.
\begin{enumerate}
\item For every $g \in \B$, $\norm{g}_{\L^1} \leq \norm{g}_{\B}$. For every $g$ in $\sC^\alpha$, $\norm{g}_{\B} \leq \norm{g}_{\sC^{\alpha}}$.
\item For every $b$, the weighted transfer operator $\sL_b:\B \to \B$ associated to $f$ (see \eqref{eq:1dim_Lb}) with weight $\xi(x) = e^{ib\tau(x)}$ is bounded, has a spectral radius $\leq 1$ and has essential spectral radius strictly $< 1$.
\item It is possible to approximate $g \in \B$ with $g_\ve \in \sC^\alpha$ such that $\norm{g-g_\ve}_{\L^1} \leq \norm{g}_\B \ve$ and $\norm{g_\ve}_{\sC^\alpha}< \norm{g}_\B \ve^{-(1+\alpha^{-1})}$.
\end{enumerate}
Under the assumptions \eqref{eq:C1+}--\eqref{eq:summability}, it is known that the Banach space of functions of generalized bounded variation (see \cite[Section 4.2]{But14}) satisfies the above assumptions.
Suppose that $f$ preserves an absolutely continuous measure $\mu$ with a density $\wp \in \B$ and $(f,\mu)$ is mixing. It can be easily shown that $F$ preserves the absolutely continuous measure $\nu = \mu \times m$, where $m$ is the Lebesgue measure. We will also denote the density of $\nu$ by $\wp$ since it is constant in the vertical direction. The objective of this note is to prove a stretched-exponential decay of correlations for the skew product $F$. Of course, such an estimate implies $(F, \nu)$ is mixing.
\section{Decay of correlations} \label{corr}
For two observables $\phi$ and $\psi$ the correlation coefficients are defined by
\[
\operatorname{cor}_{\phi,\psi}(n) = \int_{\bT^2} \phi \cdot \psi \circ F^n \ d\nu- \int_{\bT^2} \phi \ d\nu\int_{\bT^2} \psi \ d\nu.
\]
Let $\sL:\L^1(\bT^2) \to \L^1(\bT^2)$ be the transfer operator associated to the skew product $F$. That is,
\begin{equation}\label{eq:2dim_L}
\sL g (x,y) =\sum_{z\in F^{-1}(x,y)} g(z)|\det DF^{-1}(z)| = \sum_{w \in f^{-1}(x)}\frac{1}{\abs{f'(w)}}g(w,y-\tau(w)).
\end{equation}
\begin{thm} Suppose the skew product $F$ satisfies assumptions \eqref{eq:skewprod}--\eqref{eq:roof_regularity}. Then, there exist constants $\gamma_3 > 0$ and $C$ such that for every $\phi \in \sC^{\alpha}(\bT^2)$, and $\psi \in \L^\infty(\bT^2)$,
\begin{equation}
\abs{ \operatorname{cor}_{\phi,\psi}(n)} \leq Ce^{-\gamma_3\sqrt{n}}\norm{\phi}_{\sC^\alpha} \norm{\psi}_{\L^\infty} .
\end{equation}
\end{thm}
\begin{proof}
It suffices to consider $\phi$ with $\int\phi \ d\nu=0$. Hence, it suffices to estimate $\abs{\int_{\bT^2} \phi \cdot \psi \circ F^n \ d\nu} $. Also, if the result holds with $\nu$ replaced by $m$, the 2D Lebesgue measure, then it will hold for $\nu = \wp dm$, by a standard approximation argument (since $\wp \in \B$ and it can be approximated by a $\sC^\alpha$ function using our assumption on the Banach space from \Cref{setting}). Finally, if the result, i.e. stretched-exponential decay, holds for $\phi \in \sC^3(\bT^2 ,\bR)$, then we can show by approximation that it holds for H\"older observables. Note that such approximations will worsen the rate of decay but the rate will remain stretched-exponential.
Using \eqref{eq:2dim_L}, we may write
\[
\abs{\int \phi \cdot \psi \circ F^n \ dm} = \abs{\int \sum_{b \in \bZ}\sL_b^n (\hat\phi_{b}) \cdot \hat\psi_{-b} \ dm},
\]
where $\{\hat \phi_b\}_{b\in \bZ}$ are the Fourier coefficients of $\phi$ in the $y$-direction, and
\begin{equation} \label{eq:1dim_Lb}
\sL_b^n g = \sum_{h \in \cH^n} e^{ib\tau_n \circ h }\cdot g \circ h \cdot |h'| \cdot \Id_{O_h} \circ h .
\end{equation}
Noting that the $\B$-norm is stronger than then $\L^1$-norm,
\begin{equation} \label{eq:b_split}
\begin{split}
\abs{\int \sum_{b \in \bZ}\sL_b^n ( \hat\phi_{b}) \cdot \hat\psi_{-b} \ dm}
&\leq \sum_{b\in \bZ} \norm{\sL_b^n ( \hat\phi_b)}_{\L^1} \norm{\hat\psi_{-b}}_{\L^{\infty}} \\
&\leq \sum_{|b|<b_0} \norm{\sL_b^n}_{\B}\norm{ \hat\phi_b}_{\B} \norm{\hat\psi_{-b}}_{\L^{\infty}}
\\
&+ \sum_{|b| \geq b_0} \norm{\sL_b^n}_{\mixnorm}\norm{ \hat\phi_b}_{\sC^\alpha} \norm{\hat\psi_{-b}}_{\L^{\infty}}.
\end{split}
\end{equation}
We estimate the second sum above using the following.
\begin{prop} \label{main_estimate}
There exists $\gamma_2 > 0$ and a constant $C$, such that for every $|b| \geq b_0$, for every $n \in \bN$,
\begin{equation}
\norm{ \sL_b^n}_{\mixnorm } \leq Ce^{-\frac{\gamma_2}{\ln{|b|}} n}.
\end{equation}
\end{prop} This is the main estimate of the article and the rest of the article is devoted to its proof. Assume that this statement holds and let us finish the proof.
Using the regularity of $\phi, \psi$, there exist constants $C$ and $d >1$ (actually $d=2$ works) such that for every $b \in \bZ$,
\begin{equation}\label{eq:Fourier_decay}
\norm{ \hat\phi_b}_{\sC^\alpha} \leq C \norm{\phi}_{\sC^d}|b|^{-d}, \norm{\hat\psi_{-b}}_{\L^\infty} \leq \norm{\psi}_{\L^{\infty}}.
\end{equation}
Using the estimate of \Cref{main_estimate},
\begin{equation}
\sum_{|b| \geq b_0} \norm{\sL_b^n}_{\mixnorm}\norm{ \hat\phi_b}_{\sC^\alpha} \norm{\hat\psi_{-b}}_{\L^{\infty}} \leq\sum_{|b| \geq b_0} C \norm{\phi}_{\sC^d} \norm{\psi}_{\L^\infty} e^{-\frac{\gamma_2 n}{\ln{|b|}}} |b|^{-d}.\end{equation}
For estimating the sum over $|b|<b_0$ \eqref{eq:b_split}, we use the following result, which is proven in \Cref{prf_small_b}. Note that the constants $C$ and $r$ below depend on $b$ and that is why for large $|b|$ we need a different argument.
\begin{prop} \label{small_b}
For all $b \neq 0$, there exists $C$ and $r>0$, both depending on $b$, such that for every $n \in \bN$,
\begin{equation}
\norm{\sL^n_b}_{\B} \leq Ce^{-rn}.
\end{equation}
\end{prop}
Using the estimate of \Cref{small_b} for $|b|<b_0$, and the estimate of \Cref{main_estimate} for $|b|>b_0$, it follows that
\[
\abs{\operatorname{cor}_{\phi,\psi}(n)} \leq \sum_{|b| \geq b_0} C_{b_0}C \norm{\phi}_{\sC^d} \norm{\psi}_{\L^\infty} e^{-\frac{\gamma_2 n}{\ln{|b|}}} |b|^{-d}, \text{ for all } n \in \bN.
\]
Estimating the above sum yields a stretched-exponential decay. Indeed, taking $d=2$, one way to estimate $\sum_{|b| \geq b_0}e^{-\frac{\gamma_2 n}{\ln{|b|}}} |b|^{-2}$ is to split the sum into two parts $|b| \leq L$ and $|b| \geq L+1$ to get
\[
\sum_{|b| \geq b_0} e^{-\frac{\gamma_2 n}{\ln{|b|}}} |b|^{-2} \leq Le^{-\frac{\gamma_2 n}{\ln{|L|}}} +L^{-1}.
\]
Now choose $L$ so that the two parts of the sum are equal. The solution is $L=e^{\sqrt{\frac{\gamma_2n}{2}}}$, and gives
\[
Le^{-\frac{\gamma_2 n}{\ln{|L|}}} +L^{-1} = 2L^{-1} \leq 2e^{-\sqrt{\frac{\gamma_2n}{2}}}.
\]
Therefore,
\[
\abs{\operatorname{cor}_{\phi,\psi}(n)} \leq 2 C_{b_0}C \norm{\phi}_{\sC^2} \norm{\psi}_{\L^\infty} e^{-\sqrt{\frac{\gamma_2}{2}}\sqrt{n}}.
\]
\end{proof}
\section{Iteration of Standard Families} \label{iter}
In this section we introduce standard families and their dynamics. We are essentially modelling the evolution of densities under $\sL_b^n$ with the iteration of standard families.
For $\alpha \in (0,1)$, and a function $\rho:I \to \bC$ define
\begin{equation}
H(\rho) = \sup_{x,y \in I} \frac{\abs{\ln \abs{\rho(x)} - \ln\abs{\rho(y)}}}{\abs{x-y}^\alpha}.
\end{equation}
Let $\arg(\rho) \in [0,2\pi)$ be the argument of $\rho$ written in polar form. All integrals where the measure is not indicated are with respect to the Lebesgue measure. For any measurable set $A$, $|A|$ denotes the Lebesgue measure of $A$.
\begin{defin}[Standard pair] \label{std_pair}
A \emph{standard pair} with associated parameters $a, b, \ve_0$ is a pair $(I, \rho)$ consisting of an open interval $I$ and a function $\rho \in \L^1(I, \bC)$ such that
\begin{equation}\label{eq:width}
|I| < \ve_0 \leq 1;
\end{equation}
\begin{equation}\label{eq:normalized}
\int_I \abs{\rho} =1;
\end{equation}
\begin{equation}\label{eq:mod_regularity}
H(\rho) \leq a;
\end{equation}
\begin{equation}\label{eq:arg_regularity}
\abs{\arg(\rho)'} \leq a|b|.
\end{equation}
\end{defin}
\begin{defin}[Standard family]
A \emph{standard family} $\cG$ is a set of standard pairs $\{(I_j, \rho_j)\}_{j \in \cJ}$ and an associated measure $w_\cG$ on a countable set $\cJ$. We require that there exists a constant $B>0$ such that,
\begin{equation} \label{eq:bd_def}
\abs{\partial_\ve \cG} := \sum_{j \in \cJ} w_\cG(j) \int_{\partial_\ve I_j}\abs{\rho_j} \leq B\ve, \text{ for all } \ve<\ve_0,
\end{equation}
where $\partial_\ve I_j$ denotes the $\ve$-boundary of the interval $I_j$.
If $w_\cG$ is a probability measure, then $\cG$ is called a \emph{standard probability family}. Each standard family induces an absolutely continuous (complex) measure on $\bT^1$ with the density\footnote{By the sum \eqref{eq:family_density} we really mean the sum of the trivially extended standard pairs to densities defined on all of $\bT^1$, that is we set them equal to zero outside their domain.}:
\begin{equation} \label{eq:family_density}
\rho_\cG = \sum_{j \in \cJ} w_\cG(j) \rho_j.
\end{equation}
The total weight of a standard family is denoted $\abs{\cG}:= \sum_{j \in \cJ} w_\cG(j)$. The set of standard families with associated parameters $a,b,B, \ve_0$ is denoted $\cM_{a,b,B, \ve_0}$.
\end{defin}
Suppose $a >0$. For positive quantities $A$, $B$, we shall write $A \asymp_a B$ and say that $A$ and $B$ are $a$-comparable if $e^{-a}A \leq B \leq e^aA$.
Note the following simple facts.
\begin{enumerate}
\item If $A\asymp_{a} B$ and $a'>a$, then $A\asymp_{a'} B$.
\item If $A\asymp_{a} B$, then $B\asymp_{a} A$.
\item If $A\asymp_{a} B$ and $A\leq C \leq B$, then $A\asymp_{a} C \asymp_{a} B$. That is, $A$, $C$, and $B$ are pairwise $a$-comparable. It follows that all values between $A$ and $B$ are pairwise $a$-comparable.
\item If $A \asymp_a B$ and $B \asymp_{a'} C$, then $A \asymp_{a+a'} C$. Therefore, $A \asymp_{a+a'} B \asymp_{a+a'} C$.
\item If $A_1 \asymp_a B_1$ and $A_2 \asymp_{a'} B_2$, then $A_1+A_2 \asymp_{\max\{a,a'\}} B_1 + B_2$.
\item If $A_1 \asymp_a B_1$ and $A_2 \asymp_{a'} B_2$, then $A_1A_2 \asymp_{a+a'} B_1 B_2$.
\end{enumerate}
\begin{lem} \label{Fed}
If $(I,\rho)$ satisfies \eqref{eq:mod_regularity}, then for every $J, J' \subset I$ with $\abs{J}\abs{J'}\neq 0$,
\begin{equation} \label{eq:comp1}
\inf_I \abs{\rho} \asymp_{a} Avg_J \abs{\rho} \asymp_{a} Avg_ {J'} \abs{\rho}\asymp_{a} \sup_I\abs{\rho},
\end{equation}
where $Avg_J \abs{\rho} = |J|^{-1} \int_J |\rho| $ is the average of $\abs{\rho}$ on $J$.
\end{lem}
\begin{proof}
Note that \eqref{eq:mod_regularity} implies that for every $x, y \in I$, $e^{-a\abs{x-y}^\alpha}\abs{\rho}(y) \leq \abs{\rho}(x) \leq e^{a\abs{x-y}^\alpha}\abs{\rho}(y)$. This implies $\inf_I \abs{\rho} \asymp_a \sup_I\abs{\rho}$. For the rest, observe that for every $J \subset I$, $\inf_I\abs{\rho} \leq \inf_J \abs{\rho} \leq Avg_J \abs{\rho} \leq \sup_J \abs{\rho} \leq \sup_I\abs{\rho}$; therefore, lying between $a$-comparable quantities, the averages are also $a$-comparable.
\end{proof}
Note that since \eqref{eq:dist} implies $H(h') \leq D$, we may apply \Cref{Fed} to $(f^n(O_h), h')$. It follows that
\begin{equation} \label{eq:int_dist}
\sup_{f^n(O_h)} \abs{h'} \asymp_D Avg_{f^n(O_h)} \abs{h'} = \frac{\abs{O_h}}{\abs{f^n(O_h)}} \text{, for every } n \in \bN, h \in \cH^n.
\end{equation}
For the last equality we have used that $f^n$ is one-to-one on $O_h$.
Given a standard family $\cG \in \cM_{a,b,B,\ve_0}$, we define its $n$-th iterate as follows.
\begin{defin} [Iteration] \label{iteration}
Let $\cG$ be a standard family with index set $\cJ$ and weight $w_{\cG}$.
For $(j,h) \in \cJ \times \cH^n$ such that $\abs{f^n(I_j \cap O_h)} \geq \ve_0$, let $\cU_{(j,h)}$ be the index set of a finite partition $\{U_{\ell}\}_{\ell \in \cU_{(j,h)}}$ of the interval $f^n(I_j \cap O_h)$ into open intervals\footnote{Modulo a finite set of endpoints.} of size
\begin{equation} \label{eq:chop_size}
\ve_0/3 \leq \abs{U_\ell} < \ve_0.
\end{equation}
For $(j,h) \in \cJ \times \cH^n$ such that $0<\abs{f^n(I_j \cap O_h)} < \ve_0$ set $\cU_{(j,h)}=\emptyset$. Define
\begin{equation}
\cJ_{n}:=\{(j,h,\ell) | (j,h)\in \cJ \times \cH^n, \ell \in \cU_{(j,h)}, I_j\cap O_h \neq \emptyset \}.\footnote{When $\cU_{(j,h)} = \emptyset$, by $(j,h,\ell)$ we mean $(j,h)$.}
\end{equation}
For every $j_{n}:=(j, h, \ell) \in \cJ_{n}$, define
\begin{eqnarray*}
I_{j_{n}} &:=& f^n(I_{j} \cap O_h) \cap U_\ell, \\
\rho_{j_{n}} &:=& e^{ib\tau_n \circ h}\rho_{j} \circ h |h'| z_{j_{ n}}^{-1}, \text{ where }z_{j_{ n}} :=\int_{I_{j_{ n}}} \abs{\rho_{j}} \circ h \abs{h'} .
\end{eqnarray*}
Define
$
\cG_{ n} := \left\{\left(I_{j_{ n}}, \rho_{j_{ n}} \right)\right\}_{j_{ n}\in \cJ_{ n}}
$
and associate to it the measure given by
\begin{equation}\label{eq:weight_evol}
w_{\cG_{ n}}(j_{n}) = z_{j_{ n}} w_{\cG}(j).
\end{equation}
\end{defin}
\begin{rem} Comparing \eqref{eq:1dim_Lb} with the definition of $\cG_n$ and the measure associated to it \eqref{eq:family_density}, we have
\begin{equation} \label{eq:connection}
\sL^n_b \rho = \rho_{\cG_n}.
\end{equation}
This is the main connection between the evolution of densities under $\sL^n_b$ and the evolution of standard families.
\end{rem}
\subsection{Invariance}
The first thing to show is the invariance of $\cM_{a,b,B,\ve_0}$ under iterations of $\sL_b^n$ for large enough $a,B,n$ and small enough $\ve_0$.
\begin{rem}\label{fixed_param}
In this section, by $a,B, n$ large and $\ve_0$ small we mean values that satisfy the following inequalities simultaneously.
\begin{enumerate}
\item $e^{-\lambda \alpha n} + D/a <1$,
\item $e^{-\lambda n} + C_\tau/a <1$,
\item $e^{4a+2D}(2^n e^{-\lambda n}+ e^{-\sigma}) <1$,
\item $\sigma>0$ is such that there exists $\cH_\sigma^n \subset \cH^n$ such that $\cH_0^n:=\cH^n \setminus \cH_\sigma^n$ is finite, and $\sum_{h \in\cH_\sigma^n}\sup\abs{h'} < e^{-\sigma}$,
\item $\ve_0$ is such that for every interval $I$ with $\abs{I}< \ve_0$, $
\#\left\{ h \in \cH^n_0: I \cap O_h \neq \emptyset \right\} \leq 2^n$.
\end{enumerate} The constants $D$, $C_\tau$ were introduced in \Cref{setting}. One may first choose $a$ and $n$ large enough that the first two inequalities hold. Then also choose $\sigma$ (and $n$) large enough that $e^{4a+2D}(2^n e^{-\lambda n}+ e^{-\sigma})<1$. The value of $\ve_0$ is then determined by $n$ and $\sigma$.
\end{rem}
\begin{prop}\label{invariance}
Suppose $\cG \in \cM_{a,b,B,\ve_0}$ is a standard family. For every $n \in \bN$, for every $(I_{j_n},\rho_{j_n}) \in \cG_n$ we have
\begin{equation} \label{eq:normalized_1}
\int_{I_{j_{n}}}\abs{ \rho_{j_{n}} } =1,
\end{equation}
\begin{equation} \label{eq:mod_reg_1}
H(\rho_{j_{n}}) \leq a (e^{-\lambda \alpha n} + a^{-1}D),
\end{equation}
\begin{equation} \label{eq:arg_reg_1}
\abs{\arg(\rho_{j_{n}})'} \leq a |b| (e^{-\lambda n} + a^{-1}C_\tau).
\end{equation}
For every $n \in \bN$,
\begin{equation} \label{eq:total_weight_inv}
\abs{\cG_n} = \abs{\cG}.
\end{equation}
For every $a$ and $n$ large, for every $\sigma>0$, if $\ve_0>0$ is small enough, then
\begin{equation} \label{eq:growth_lemma}
\abs{\partial_\ve \cG_{n}} \leq C_a(2^n+e^{-\sigma} e^{\lambda n})\abs{\partial_{e^{-\lambda n} \ve}\cG} + C_a \ve_0^{-1}\ve \abs{\cG}, \text{ for all } \ve < \ve_0.
\end{equation}\end{prop}
\begin{proof}
Property \eqref{eq:normalized_1} follows from the definition.
To show \eqref{eq:mod_reg_1}, note that
\begin{equation}
H(\rho_{j_{n}})= H \left( h' \cdot (\rho_j \circ h)\right).
\end{equation}
Using the definition of $H(\cdot)$ and noting its properties under multiplication and composition, it follows that
\[H(\rho_{j_{n}}) \leq H(h') + e^{-\lambda\alpha n}H(\rho_{j}). \]
By \eqref{eq:dist} we have $H(h') \leq D$, and by assumption $H(\rho_{j})\leq a$, finishing the proof of \eqref{eq:mod_reg_1}.
To show \eqref{eq:arg_reg_1}, note that $\arg(\rho_{j_{n}}) = b \tau_n \circ h + \arg(\rho) \circ h$. Therefore,
\[
\abs{\arg(\rho_{j_{n}})'} \leq |b| \abs{(\tau_n \circ h)'} + \abs{\arg(\rho)'} |h'| \leq |b| C_\tau + a|b|e^{-\lambda n}.
\]
To show \eqref{eq:total_weight_inv}, write
\begin{equation}
\begin{split}
\sum_{j_{n} \in \cJ_{n}} w_{\cG_{n}} (j_n)
&= \sum_{j_{n} \in \cJ_{n}} w_{\cG}(j) \int_{I_{j_n}} \abs{\rho_{j}} \circ h \abs{h'} \\
&= \sum_{(j,h) \in \cJ \times \cH^n} \sum_{\ell \in \cU_{(j,h)}} w_{\cG}(j) \int_{f^n(I_j \cap O_h) \cap U_\ell} \abs{\rho_{j}} \circ h \abs{h'} \\
&=\sum_{(j,h) \in \cJ \times \cH^n} w_{\cG}(j) \int_{f^n(I_{j} \cap O_h)} \abs{\rho_{j}} \circ h \abs{h'} \\
&=\sum_{j \in \cJ} w_{\cG}(j) \sum_{h \in \cH^n}\int_{f^n(I_{j} \cap O_h)} \abs{\rho_{j}} \circ h \abs{h'} \\
&= \sum_{j \in \cJ} w_{\cG} (j) \int_{I_j} \abs{\rho_{j}} = \sum_{j \in \cJ} w_{\cG} (j).
\end{split}
\end{equation}
Note that in the second line we wrote $\cJ \times \cH^n$ instead of $\{(j,h) \in \cJ \times \cH^n | I_j \cap O_h \neq \emptyset\}$. We can do this because if $I_j \cap O_h = \emptyset$, then the corresponding terms are zero.
The third equality follows by summing over all $\ell$ since the intervals $f^n(I_{j} \cap O_h) \cap U_\ell$ form a partition of the interval $f^n(I_{j} \cap O_h)$. The last line is a consequence of change of variables and $\int\abs{\rho_j} $ being equal to $1$.
To prove \eqref{eq:growth_lemma}, suppose $\sigma>0$. Then, \eqref{eq:summability} implies that there exists $\cH^n_\sigma \subset \cH^n$ such that
\begin{equation}
\cH^n_0 :=\cH^n \setminus \cH^n_\sigma \text{ is finite},
\end{equation}
and
\begin{equation} \label{eq:tale_sigma}
\sum_{h \in \cH^n_\sigma}\sup_{f^n(O_h)}\abs{h'} < e^{-\sigma}.
\end{equation}
Since $\cH^n_0$ is finite, choose $\ve_0 = \ve_0(\cH^n_0)$ such that\footnote{There is some freedom here to choose $\ve_0$. The value of $\ve_0$ depends on the partition $\{O_h\}_{h \in \cH^n}$ and the value of $\sigma$. Note that since we only use \eqref{eq:growth_lemma} with a fixed $n$, the value of $\ve_0$ causes no problems even if it is very small. The optimal value depends on the underlying system.} for every interval $I$ with $\abs{I}< \ve_0$,
\begin{equation}
\#\left\{ h \in \cH^n_0: I \cap O_h \neq \emptyset \right\} \leq 2^n.
\end{equation}
Suppose $\ve < \ve_0$. We have, by definition,
\[
\abs{\partial_\ve \cG_{n}} := \sum_{j_{n} \in \cJ_{n}} w_{\cG_{n}}(j_n)\int_{\partial_\ve I_{j_{n}}} \abs{\rho_{j_{n}}}.
\]
We split the sum into two parts according to whether $\cU_{(j,h)} \neq \emptyset$ or $\cU_{(j,h)} = \emptyset$. The two parts are respectively,
\[
\sum_{\{j_{n} \in \cJ_{n}| \cU_{(j,h) \neq \emptyset}\}} w_{\cG_{n}}(j_n)\int_{\partial_\ve I_{j_{n}}} \abs{\rho_{j_{n}}} = \sum_{j \in \cJ}\sum_{\{h \in \cH^n| f^n(I_j \cap O_h) \geq \ve_0\}} \sum_{\ell \in \cU_{(j,h)}} w_{\cG_{n}}(j_n)\int_{\partial_\ve I_{j_{n}}} \abs{\rho_{j_{n}}},
\]
and
\[
\sum_{\{j_{n} \in \cJ_{n}| \cU_{(j,h) = \emptyset}\}} w_{\cG_{n}}(j_n)\int_{\partial_\ve I_{j_{n}}} \abs{\rho_{j_{n}}} = \sum_{j \in \cJ}\sum_{\{h \in \cH^n| f^n(I_j \cap O_h) < \ve_0\}} w_{\cG_{n}}(j_n)\int_{\partial_\ve I_{j_{n}}} \abs{\rho_{j_{n}}}.
\]
\textbf{The case $\cU_{(j,h)} \neq \emptyset$:} First note that \eqref{eq:mod_reg_1} implies $H(\rho_{j_n}) \leq a$ for sufficiently large $n$. Therefore, \eqref{eq:comp1} implies
$\abs{\partial_\ve I_{j_n}}^{-1} \int_{\partial_\ve I_{j_n}} \abs{\rho_{j_n}} \asymp_a \abs{ I_{j_n}}^{-1} \int_{I_{j_n}}\abs{\rho_{j_n}} = \abs{ I_{j_n}}^{-1}$.
Hence,
\begin{equation}\label{eq:bd_int}
\int_{\partial_\ve I_{j_n}} \abs{\rho_{j_n}}\asymp_a \frac{\abs{\partial_\ve I_{j_n}}}{\abs{ I_{j_n}}} = \frac{\abs{\partial_\ve I_{j_n}}}{\abs{f^n(I_j\cap O_h)\cap U_\ell}}.
\end{equation}
Observe that by definition, $w_{\cG_{n}}(j_n) = w_{\cG} (j) z_{j_n}= w_{\cG} (j) \int_{f^n(I_j \cap O_h) \cap U_\ell} \abs{\rho_j} \circ h \abs{h'} $. Hence by change of variables $w_{\cG_{n}}(j_n) =w_{\cG} (j) \int_{I_j \cap O_h \cap f^{-n}(U_\ell)} \abs{\rho_j}$. Since $H(\rho_j)\leq a$ and $\int_{I_j} \abs{\rho_j}=1$, it follows by \eqref{eq:comp1} that
\begin{equation}\label{eq:bd_w}
w_{\cG_{n}}(j_n) \asymp_a w_{\cG}(j) \frac{\abs{I_j \cap O_h \cap f^{-n}(U_\ell)}}{\abs{I_j}}.
\end{equation}
Putting \eqref{eq:bd_int} and \eqref{eq:bd_w} together, and then using the distortion estimate \eqref{eq:int_dist},
\begin{equation} \label{eq:pre_sum_l}
\begin{split}
w_{\cG_{n}}(j_n) \int_{\partial_\ve I_{j_n}} \abs{\rho_{j_n}} &\asymp_{2a} w_{\cG}(j) \frac{\abs{\partial_\ve I_{j_n}}\abs{I_j \cap O_h \cap f^{-n}(U_\ell)}}{\abs{I_j}\abs{f^n(I_j\cap O_h)\cap U_\ell}} \\
&\leq_{2a+D} w_{\cG}(j) \abs{\partial_\ve I_{j_n}} \abs{I_j}^{-1} \sup_{f^n(I_j \cap O_h)} \abs{h'}.\\
&\leq_{2a+D} w_{\cG}(j) 2\ve \abs{I_j}^{-1} \sup_{f^n(I_j \cap O_h)} \abs{h'}.
\end{split}
\end{equation}
Note that by the notation $\leq_a B$ we mean $\leq e^aB$. To finish the estimate, we need to sum over $\ell$, then sum over $h$ such that $f^n(I_j \cap O_h) \geq \ve_0$ and then over $j \in \cJ$. Note that since each interval $f^n(I_j \cap O_h)$ is chopped into intervals of size $\geq \ve_0/3$, the number of elements in $\cU_{(j,h)}$ is bounded by $3\ve_0^{-1}\abs{f^n(I_j \cap O_h)}$. Therefore, summing \eqref{eq:pre_sum_l} over all $\ell \in \cU_{(j,h)}$, and then using the distortion estimate \eqref{eq:int_dist} yields,
\[
2\ve 3\ve_0^{-1} w_{\cG}(j) \abs{I_j}^{-1} \sup_{f^n(I_j \cap O_h)} \abs{h'} \abs{f^n(I_j \cap O_h)} \leq_{2a+2D} 6\ve \ve_0^{-1} w_{\cG}(j) \abs{I_j}^{-1}\abs{I_j \cap O_h}.
\]
Summing over $h$ such that $f^n(I_j\cap O_h) \geq \ve_0$ and noting that this is no greater than summing over all $h \in \cH^n$ yields
\[
\leq_{2a+2D} 6\ve \ve_0^{-1} w_{\cG}(j) \abs{I_j}^{-1}\abs{I_j}= 6\ve \ve_0^{-1}w_\cG(j).
\]
Finally, summing over $j \in \cJ$ yields,
\begin{equation}\label{eq:bd_chopped_pairs}
\sum_{\{j_n \in \cJ_n| \cU_{(j,h)} \neq \emptyset\}} w_{\cG_{n}}(j_n)\int_{\partial_\ve I_{j_{n}}} \abs{\rho_{j_{n}}} \leq_{2(a+D)}6\ve \ve_0^{-1}\abs{\cG}.
\end{equation}
\textbf{The case $\cU_{(j,h)} = \emptyset$:} We need to estimate:
\[
\sum_{\{j_{n} \in \cJ_{n}| \cU_{(j,h) = \emptyset}\}} w_{\cG_{n}}(j_n)\int_{\partial_\ve I_{j_{n}}} \abs{\rho_{j_{n}}} = \sum_{j \in \cJ}\sum_{\{h \in \cH^n| f^n(I_j \cap O_h) < \ve_0\}} w_{\cG_{n}}(j_n)\int_{\partial_\ve I_{j_{n}}} \abs{\rho_{j_{n}}}.
\]
We further split the second sum into two parts, one over $\cH^n_0$ and the other over $\cH^n_\sigma$.
For the sum over $\cH^n_0$, we have the bound
\[
\begin{split}
& \sum_{j \in \cJ} w_{\cG}(j) \sum_{\{h \in \cH_0^n| f^n(I_j \cap O_h) < \ve_0\}}\int_{\partial_\ve f^n (I_j \cap O_h)} \abs{\rho_{j}} \circ h \abs{h'} \\
&\leq
\sum_{j \in \cJ} w_{\cG}(j) \sum_{\{h \in \cH_0^n| f^n(I_j \cap O_h) < \ve_0\}} \left[ \int_{f^n\left(\partial_{e^{-\lambda n}\ve} I_{j} \cap O_h\right)} + \int_{f^n\left(\partial_{e^{-\lambda n}\ve} O_h\cap I_{j}\right)} \right] \abs{\rho_{j}} \circ h \abs{h'} \\
&\leq \abs{\partial_{e^{-\lambda n} \ve} \cG} + \sum_{j \in \cJ} w_{\cG}(j) \sum_{\{h \in \cH_0^n| f^n(I_j \cap O_h) < \ve_0\}} \int_{\partial_{e^{-\lambda n}\ve} O_h \cap I_{j}} \abs{\rho_{j}} . \\
\end{split}
\]
The first inequality holds because if a point is at a distance less than $\ve$ from the boundary of $f^n (I_{j} \cap O_h)$, then its preimage must be at a distance $e^{-\lambda n}\ve$ from the boundary of $I_{j}$ or from the boundary of $O_h$. The second inequality is a consequence of change of variables and $f^n$ being one-to-one on $I_j\cap O_h$.
Since $\abs{I_j} < \ve_0$, by the choice of $\ve_0$, it follows that $I_{j}$ intersects at most $2^n$ intervals $O_h$. Also, \eqref{eq:comp1} implies that $Avg_{\partial_{e^{-\lambda n}\ve} O_h \cap I_{j}}\abs{\rho_j} \asymp_a Avg_{\partial_{e^{-\lambda n}\ve} I_{j}} \abs{\rho_j}$. This in turn implies $\int_{\partial_{e^{-\lambda n}\ve} O_h \cap I_{j}}\abs{\rho_j} \leq_a \int_{\partial_{e^{-\lambda n}\ve} I_j }\abs{\rho_j}$ because $\abs{\partial_{e^{-\lambda n}\ve} O_h \cap I_{j}} \leq \abs{\partial_{e^{-\lambda n}\ve} I_j }$. Therefore,
\begin{equation} \label{eq:bd_short_pairs}
\begin{split}
& \sum_{j \in \cJ} w_{\cG}(j) \sum_{\{h \in \cH_0^n| f^n(I_j \cap O_h) < \ve_0\}}\int_{\partial_\ve f^n (I_j \cap O_h)} \abs{\rho_{j}} \circ h \abs{h'} \\ &\leq_a \abs{\partial_{e^{-\lambda n} \ve} \cG} + 2^n\sum_{j \in \cJ} w_{\cG_{k}}(j)\int_{\partial_{e^{-\lambda n}\ve} I_{j}} \abs{\rho_{j}} \\
&\leq_a 2^n 2\abs{\partial_{e^{-\lambda n} \ve} \cG}.
\end{split}
\end{equation}
For the sum over $\cH^n_\sigma$, similarly to \eqref{eq:pre_sum_l}, we have the bound
\[
\sum_{\{j_n | \cU_{(j,h)} = \emptyset\}} w_{\cG_{n}}(j_n)\int_{\partial_\ve I_{j_{n}}} \abs{\rho_{j_{n}}} \leq_{2a+D} \sum_{j \in \cJ} w_{\cG}(j) \sum_{\{h \in \cH_\sigma^n| f^n(I_j \cap O_h) < \ve_0\}}\abs{I_j}^{-1}\abs{\partial_\ve I_{j_n}}\sup_{f^n(I_j \cap O_h)} \abs{h'}.
\]
Multiplying and dividing the right hand side by $\abs{\partial_\ve I_j}$ and using $\abs{\partial_\ve I_{j_n}}\abs{\partial_\ve I_j}^{-1} \leq 1$, the right hand side is
\[ \leq_{2a+D} \sum_{j \in \cJ} w_{\cG}(j) \abs{\partial_\ve I_{j_n}} \abs{I_j}^{-1} \sum_{\{h \in \cH_\sigma^n| f^n(I_j \cap O_h) < \ve_0\}}\sup_{f^n(I_j \cap O_h)} \abs{h'}.
\]
Notice that $\abs{\partial_\ve I_{j_n}} \abs{I_j}^{-1} \asymp_a \int_{\partial_\ve I_j} \abs{\rho_j}$. Therefore, the above quantity is
\[
\leq_{3a+D} \sum_{j \in \cJ} w_{\cG}(j)\int_{\partial_\ve I_j} \abs{\rho_j} \sum_{\{h \in \cH_\sigma^n| f^n(I_j \cap O_h) < \ve_0\}}\sup_{f^n(I_j \cap O_h)} \abs{h'}.
\]
Using \eqref{eq:tale_sigma}, the estimate for the sum over $\cH^n_\sigma$ is
\begin{equation} \label{eq:bd_tail}
\leq_{3a+D} e^{-\sigma}\sum_{j \in \cJ} w_{\cG}(j) \int_{\partial_\ve I_j} \abs{\rho_j} = e^{-\sigma} \abs{\partial_\ve \cG} \leq_a e^{-\sigma}\abs{\partial_{e^{-\lambda n} \ve}\cG}.
\end{equation}
Finally, adding \eqref{eq:bd_chopped_pairs}, \eqref{eq:bd_short_pairs} and \eqref{eq:bd_tail} together, we arrive at
\begin{equation}
\abs{\partial_\ve \cG_{n}} \leq_{4a+2D} 2(2^n+e^{-\sigma} e^{\lambda n})\abs{\partial_{e^{-\lambda n} \ve}\cG} + 6\ve_0^{-1}\ve \abs{\cG}.
\end{equation}
\end{proof}
\begin{lem} \label{iterated_growth_lemma}
Suppose $\cG \in \cM_{a,b,B,\ve_0}$ with $a$ sufficiently large and $\ve_0$ sufficiently small. Then, there exist $C$, $\bar C$, and $0<\beta < \lambda$ such that,
\begin{equation} \label{eq:iterated_growth_lemma}
\abs{\partial_\ve \cG_{m}} \leq Ce^{\beta m} \abs{\partial_{e^{-\lambda m}\ve}\cG} + \bar C \ve, \text{ for all } m \in \bN \text{, and } \ve < \ve_0.
\end{equation}
\end{lem}
\begin{proof}
The result follows by choosing a fixed $n$ as in \Cref{fixed_param} and iterating \eqref{eq:growth_lemma} with this fixed $n$. Choose $n$, $\sigma$ large such that $2e^{4a+2D}(2^ne^{-\lambda n}+e^{-\sigma}) <1$ as in \Cref{fixed_param}. Let $\beta$ be such that $e^\beta=(2e^{4a+2D})^{1/n}(2^n+e^{-\sigma} e^{\lambda n})^{1/n}$. Then $e^{\beta n}e^{-\lambda n} <1$. That is, $\beta<\lambda$. For every $m \in \bN$, write $m=kn+r$, $0\leq r<n$. Applying \eqref{eq:growth_lemma}, we have
\begin{equation} \label{eq:once_gl}
\abs{\partial_\ve \cG_{n}} \leq 2e^{4a+2D}(2^r+e^{-\sigma}e^{\lambda r})\abs{\partial_{e^{-\lambda r} \ve}\cG_{kn}} + 6 \ve_0^{-1}\ve \abs{\cG_{kn}}.
\end{equation}
Let $C=2e^{4a+2D}(2^r+e^{-\sigma}e^{\lambda r})$. Applying \eqref{eq:growth_lemma} $k$ more times, we get
\begin{equation}
\begin{split}
\abs{\partial_\ve \cG_{m}} &\leq C e^{\beta kn}\abs{\partial_{e^{-\lambda r}e^{-\lambda kn} \ve}\cG} + 6\ve_0^{-1}\ve(e^{-\lambda r}/(1-e^{\beta-\lambda})+1) \abs{\cG} \\
&\leq C e^{\beta m}\abs{\partial_{e^{-\lambda m}\ve}\cG} + \bar C\ve \abs{\cG},
\end{split}
\end{equation}
where $\bar C = 6\ve_0^{-1}(e^{-\lambda r}/(1-e^{\beta-\lambda})+1)$.
\end{proof}
\begin{rem}
The invariance of $\cM_{a,b,B,\ve_0}$ under iterations by $\sL_b^m$ follows by taking $a$, $m$, $B$ large, and $\ve_0$ small as in \Cref{fixed_param}. Note that $m$ and $B$ must be large enough that $Ce^{(\beta-\lambda)m}+\bar C/B < 1$ to guarantee $\abs{\partial_\ve \cG_m} \leq B \ve$ for all $\ve <\ve_0$.
\end{rem}
\section{Transversality} \label{trans}
Due to the neutrality of the $y$-direction in our setting, it is possible that measure stays on $x$-direction invariant curves that simply rotate in the $y$-direction. In this scenario the skew-product would not be mixing. To avoid such a scenario we need an assumption that forces measure to spread in different directions. We shall refer to this property as transversality. In our setting we assume that $\tau$ is not Lipschitz-cohomologous to a piecewise constant function on the joint partition of $\tau$ and $f$. In this section, we show that this condition implies a uniform non-integrability condition, which in turn implies the transversality notion that we use later to obtain a stretched-exponential decay of correlations. The material of this section is influenced by \cite{Tsu08}.
Note that \begin{equation}
DF_{(x,y)} = \begin{pmatrix}
f '(x)& 0 \\
\tau'(x) & 1 \\
\end{pmatrix},
\end{equation}
which is independent of the second coordinate. Also, note that $DF$ preserves the cone $\sK_\eta=\{(u,v): \abs{v} \leq \eta\abs{u} \}$, where $\eta=\norm{\tau'/f'}_\infty/(1-\norm{1/f'}_\infty)$.
\begin{lem} \label{coho_tang_cones}
Suppose that for every $n \in \bN$, for every $x \in \bT^1$, and inverse branches $h_1, h_2 \in \cH^{n}$,\footnote{We really mean every pair of inverse branches that have $x$ in their domain.}
\[
DF^{n}_{h_1(x)} \sK_\eta \cap DF^{n}_{h_2(x)}\sK_\eta \neq \{0\}.
\]
Then, $\tau$ is Lipschitz-cohomologous to a piecewise constant function on the joint partition of $f$ and $\tau$.
\end{lem}
\begin{proof}
The hypothesis implies that for every $x \in \bT^1$, $\cap_{h \in \cH^{n}} DF^{n}_{h(x)}\sK_\eta \neq \{0\}$. That is, the intersection contains a common direction $\theta(x,n)$. Moreover, since the cones $DF^{n}_{h(x)}\sK_\eta$ contract uniformly under iteration, the intersection $\cap_{n \in \bN}\cap_{h \in \cH^{n}} DF^{n}_{h(x)}\sK_\eta $ contains a unique direction $(1,\theta(x))$. Since $\theta(x)$ is invariant under $DF$, we have \begin{equation} \label{eq:pre_coho}
f' \cdot \theta \circ f = \tau' + \theta.
\end{equation}
Define $\phi(x) = \int_{0}^x \theta(t)dt$. Note that since $\theta(x)$ is bounded, $\phi$ is Lipschitz. Let $p$ be the left endpoint of a partition element of the joint partition of $f$ and $\tau$ and $x$ be a point in the same partition element.
We may write $\tau(x)= \tau(p)+\int_p^x \tau'$, where $\tau(p)$ is interpreted as the value obtained by taking a one-sided limit. Substituting $\tau'=f' \cdot \theta \circ f - \theta$ from \eqref{eq:pre_coho} into this equation, and doing a change of variables yields,
\begin{equation}\label{eq:coho}
\tau(x) = \phi \circ f(x) - \phi(x) + (\tau(p) + \phi(p) - \phi\circ f (p)).
\end{equation}
It follows that $\tau$ is Lipschitz-cohomologous to a piecewise constant function on the joint partition of $f$ and $\tau$.
\end{proof}
\begin{lem} \label{trans_cones_UNI}
Suppose for $x \in \bT^1$, $n \in \bN$, and inverse branches $h_1, h_2 \in \cH^n$ holds
\[
DF^{n}_{h_1(x)} \sK_\eta \cap DF^{n}_{h_2(x)}\sK_\eta = \{0\}.
\]
Then, there exists $C_0:=C_0(n,x)$ such that
\[
\abs{(\tau_n \circ h_1)'(x)-(\tau_n \circ h_2)'(x)} > C_0.
\]
\end{lem}
\begin{proof}
The hypothesis implies that the two cones $DF^{n}_{h_1(x)} \sK_\eta$ and $DF^{n}_{h_2(x)}\sK_\eta$ are a distance $C_0:=C_0(n,x)$ apart. Suppose $v_1, v_2 \in \bR$ satisfy $\abs{v_1}, \abs{v_2} \leq \eta$. Then $(1,v_1), (1,v_2) \in \sK_\eta$. Observe that for $j \in \{1,2\}$, $DF^{n}_{h_j(x)}(1,v_j)=((f^n)'\circ h_j(x), \tau' \circ h_j(x) + v_j)$. Therefore $(1, (\tau_n \circ h_j)'(x)+v_j h_j') \in DF^{n}_{h_j(x)} \sK_\eta$. Therefore, the two vectors are also $C_0$ apart; that is,
\[
\abs{(\tau_n \circ h_1)'(x)+v_1 h_1'(x)- (\tau_n \circ h_2)'(x)-v_2 h_2'(x)} > C_0.
\]
Using the triangle inequality,
\[
\abs{(\tau_n \circ h_1)'(x)- (\tau_n \circ h_2)'(x))} > C_0 - \abs{v_1 h_1'(x)-v_2 h_2'(x)}.
\]
Taking $v_1=v_2=0$ implies the result.
\end{proof}
\begin{lem} \label{nbd_UNI}
Suppose $\tau$ is not Lipschitz-cohomologous to a piecewise constant function on the joint partition of $\tau$ and $f$. Then, there exists $x_0 \in \bT^1$, there exists $\ntrans \in \bN$, inverse branches $h_1, h_2 \in \cH^{\ntrans}$, a neighbourhood $V_{\ntrans}$ of $x_0$ contained in the open set $f^{\ntrans}(O_{h_1}) \cap f^\ntrans (O_{h_2})$, and a constant $C_1:=C_1(n_1,x_0)$ such that
\begin{equation} \label{eq:nbd_UNI}
\abs{\left(\tau_{\ntrans} \circ h_1 - \tau_{\ntrans} \circ h_2\right)'(x)} > C_1 \text{ for every } x \in V_{\ntrans}.
\end{equation}
\end{lem}
\begin{proof}
Suppose $\tau$ is not Lipschitz-cohomologous to a piecewise constant function on the joint partition of $\tau$ and $f$. \Cref{coho_tang_cones} implies that there exists $\ntrans \in \bN$, $x_0 \in \bT^1$,
and inverse branches $h_1, h_2 \in \cH^\ntrans$ such that
\begin{equation} \label{eq:trans_cones}
DF^{\ntrans}_{h_1(x_0)} \sK_\eta \cap DF^{\ntrans}_{h_2(x_0)}\sK_\eta = \{0\}.
\end{equation}
\Cref{trans_cones_UNI} implies that there exists $C_0 = C_0(n_1, x_0)$ such that
\begin{equation} \label{eq:distant_cones}
\abs{\left(\tau_{\ntrans} \circ h_1 - \tau_{\ntrans} \circ h_2\right)'(x_0)} \geq C_0.
\end{equation}
By continuity of $(f^{\ntrans})'$ and $\tau_{\ntrans}'$ at $h_1(x_0)$ and $h_2(x_0)$ the cones $DF^{\ntrans}_{h_1(x_0)} \sK_\eta$ and $DF^{\ntrans}_{h_2(x_0)}\sK_\eta$ vary continuously in a neighbourhood of $x_0$ and so does the distance between them (i.e. $C(n,\cdot)$ varies continuously in a neighbourhood of $x_0$). It follows that there exists a neighbourhood $V_{\ntrans}$ of $x_0$ and a constant, which we again denote by $C_1:=C_1(\ntrans,x_0)$ such that
\begin{equation} \label{eq:nbd_distant_cones}
\abs{\left(\tau_{\ntrans} \circ h_1 - \tau_{\ntrans} \circ h_2\right)'(x)} \geq C_1 \text{ for every } x \in V_{\ntrans}.
\end{equation}
\end{proof}
\begin{cor}\label{persistent_nbd_UNI}
Suppose $\tau$ is not Lipschitz-cohomologous to a piecewise constant function on the joint partition of $\tau$ and $f$. Then, there exists $x_0 \in \bT^1$, there exists $\ntrans \in \bN$, inverse branches $h_1, h_2 \in \cH^{\ntrans}$, a constant $C_1:=C_1(n_1, x_0)$, and for every $n \geq \ntrans$ and every $l_1,l_2 \in \cH^{n-\ntrans}$ there exists a neighbourhood $V_n$ of $x_0$ contained in $f^n(O_{l_1 \circ h_1}) \cap f^n(O_{l_2 \circ h_2})$ such that
\begin{equation}
\abs{\left(\tau_n \circ l_1 \circ h_1 - \tau_n \circ l_2 \circ h_2\right)'(x)} > C_1 \text{ for every } x \in V_{n}.
\end{equation}
\end{cor}
\begin{proof}
Let $x_0$, $\ntrans$, $V_{\ntrans}$ and $C_1$ be as in \Cref{nbd_UNI}. For every $n \geq \ntrans$ and $l_1,l_2 \in \cH^{n-\ntrans}$, by invariance of the cone, we have $DF^{n-\ntrans}_{l_j\circ h_j(x)}\sK_\eta \subset \sK_\eta$ for every $x \in f^n(O_{l_1 \circ h_1}) \cap f^n(O_{l_2 \circ h_2})$. If also $x \in V_{\ntrans}$ (the neighbourhood of $x_0$ from \Cref{nbd_UNI}), then $DF^{\ntrans}_{h_1(x)}DF^{n-\ntrans}_{l_1\circ h_1(x)}\sK_\eta \cap DF^{\ntrans}_{h_2(x)}DF^{n-\ntrans}_{l_2\circ h_2(x)}\sK_\eta = \{0\}$. That is, the cones $DF^{n}_{l_1\circ h_1(x)}\sK_\eta$ and $DF^{n}_{l_2\circ h_2(x)}\sK_\eta$ are at least distant $C_1$ apart. As in \Cref{trans_cones_UNI}, this transversality of the cones implies
\begin{equation}
\abs{\left(\tau_{\ntrans}\circ l_1 \circ h_1 - \tau_{\ntrans} \circ l_2 \circ h_2\right)'(x)} > C_1 \text{ for every } x \in V_n,
\end{equation}
where $V_n:=V_\ntrans \cap f^n(O_{l_1 \circ h_1}) \cap f^n(O_{l_2 \circ h_2})$.
\end{proof}
The following shows that any interval of positive length maps forward, while getting cut and expanded, in a way that at least two of its pieces overlap and simultaneously satisfy a condition similar to \eqref{eq:nbd_UNI}.
\begin{prop} \label{overlap_UNI}
Suppose $\tau$ is not Lipschitz cohomologous to a piecewise constant function on the joint partition of $\tau$ and $f$; and, in addition, $f$ is covering. There exists a constant $C_1$ such that for every interval $I$ with $0<\delta<\abs{I} \leq \ve_0 $, there exists $n_\delta$ such that for every $n \geq n_\delta$, there exist $h_1,h_2 \in \cH^n$, such that $O_{h_1},O_{h_2} \subset I$ and $f^n(O_{h_1}) \cap f^n(O_{h_2})$ contains an interval $I_*$ of size $0<\ovl \leq \abs{I_*}$ on which holds\footnote{The quantity $\ovl$ depends on $\delta$, $n_\delta$ and the choice of $I$. Later we will get rid of the dependence on $I$ by a compactness argument.}
\[
\abs{\left(\tau_n \circ h_1 - \tau_n \circ h_2 \right)'} > C_1.
\]
\end{prop}
\begin{proof}
\Cref{persistent_nbd_UNI} implies that there exists $x_0$, $\ntrans$, inverse branches $\tilde h_1, \tilde h_2 \in \cH^{\ntrans}$; there exists a constant $C_1$; and, for every $n \geq n_1$ and every $l_1,l_2 \in \cH^{n-n_1}$, there exists a neighbourhood $V_n$ of $x_0$ contained in $f^n(O_{l_1 \circ \tilde h_1}) \cap f^n(O_{l_2 \circ \tilde h_2})$, such that
\begin{equation}
\abs{\left(\tau_n \circ l_1 \circ \tilde h_1 - \tau_n \circ l_2 \circ \tilde h_2\right)'(x)} > C_1 \text{ for every } x \in V_{n}.
\end{equation}
Since $f$ is covering, there exists $N(\delta)$ (recall that $\delta$ is the lower bound on the length of $I$) such that for every $n \geq N(\delta)+n_1=:n_\delta$ and every $l_1, l_2 \in \cH^{n-n_1}$
\[
l_j \circ \tilde h_j (V_{n}) \subset O_{l_j \circ h_j} \subset I \text{, for }j \in \{1,2\}.
\]
Note that the first inclusion is a consequence of the property that $V_n$, $n\geq n_1$, is contained in $f^n(O_{l_1 \circ \tilde h_1}) \cap f^n(O_{l_2 \circ \tilde h_2})$. Set $h_1 := l_1 \circ \tilde h_1$, $h_2 := l_2 \circ \tilde h_2$ and $I_* = V_{n}$. Then, we have
\[
\abs{\left(\tau_n \circ h_1 - \tau_n \circ h_2 \right)'} > C_1 \text{ for every } x \in I_*.
\]
Denote the length of $I_*$, the overlap interval, by $\ovl$. Note that $\ovl$ depends on $\delta$, $n$ and the choice of the initial interval $I$.
\end{proof}
\subsection{Transversality of standard pairs}
In this subsection we will state the transversality condition of \Cref{overlap_UNI} in terms of standard pairs. We will also get rid of the dependence of $\Delta$ on the choice of the interval $I$ using a compactness argument.
\begin{cond}[Transversality of standard pairs] \label{finite_transversality}
Consider a standard family $\cG$. For every $\delta>0$, there exists $n_\delta \in \bN$, a finite number $k:=k_\delta$ of pairs of inverse branches
\[
\{(h_{1,1},h_{1,2}) \dots (h_{k,1},h_{k,2})\} \subset \cH^{n_\delta} \times \cH^{n_\delta}
\]
such that for any standard pair $(I, \rho) \in \cG$, with $\abs{I} >3\delta$, the image standard family $\cG_{n_\delta}$ contains two standard pairs, obtained from the above finite collection of inverse branches, which overlap and are transversal on an interval of length no smaller than $\ovl=\ovl(k_\delta, n_\delta)$.
More precisely, there exists $l \in \{1, \dots ,k_\delta\}$, $\ovl:=\ovl(k_\delta, n_\delta) >0$, standard pairs $(I_j, \rho_j)$, $j\in\{1,2\}$ such that there exists $U$, with $\ve_0/3 \leq \abs{U} \leq \ve_0$,\footnote{ $U$ is a choice of cutting and can be taken to be equal to $\bT^1$ if no cutting is necessary; that is, when $\abs{f^{n_\delta}(O_{h_{l,j}})}< \ve_0$.} such that
\[
I_j:=f^{n_\delta}(O_{h_{l,j}})\cap U, \rho_j := z_{ h_{l,j}}^{-1}e^{ib\tau_{n_\delta} \circ h_{l,j}}\rho \circ h_{l,j} |( h_{l,j})'|, \text{ and }
O_{h_{l,j}}\subset I.
\]
Furthermore, $I_1 \cap I_2$ contains an interval $I_*$ of size $\ovl$ on which holds
\[
\abs{\left(\tau_{n_\delta} \circ h_{l,1}- \tau_{n_\delta} \circ h_{l,2}\right)'} > C_1.
\]
Denote
\begin{equation} \label{least_trans_domain}
M(n_\delta) := \min_{\stackrel{j\in\{1,2\}}{ l\in\{1, \cdots k_\delta\}}}\{\abs{O_{h_{l,j}}\cap h_{l,j}(U_j)}\}.
\end{equation}
\end{cond}
\begin{proof}
Divide the interval into subintervals of length $\delta$. Denote the finite collection of intervals by $\{J_l\}_{l=1}^{k_\delta}$. For each interval apply \Cref{overlap_UNI}. It follows that there exists $n:=n_\delta$ and finitely many inverse branches
\[
\{(h_{1,1},h_{1,2}) \dots (h_{k,1},h_{k,2})\} \subset \cH^{n_\delta} \times \cH^{n_\delta}
\]
such that
$O_{h_{l,1}},O_{h_{l,2}} \subset J_l$ and $f^n(O_{h_{l,1}}) \cap f^n(O_{h_{l,2}})$ contains an interval of size $\Delta_l >0$\footnote{$\Delta_l$ depends on $\delta$ and $n_\delta$ in addition to $l$.} on which holds
\[
\abs{\left(\tau_n \circ h_{l,1} - \tau_n \circ h_{l,2} \right)'} > C_1.
\]
Let $\ovl = \min_{l\in \{1, \cdots, k_\delta \}} \Delta_l$. For any standard pair $(I,\rho)$ with $|I|> 3\delta$, $I$ contains at least one of the intervals $J_l$ of length $\delta$. As mentioned above, $J_l$ contains a pair of partition intervals $O_{h_{l,1}},O_{h_{l,2}}$ whose images overlap over an interval of length $\ve_0/3$ and are transversal. If these images are of length $< \ve_0$, by definition, they are the support of standard pairs:
\[
I_j:=f^n(O_{h_{l,j}}), \rho_j := z_{ h_{l,j}}^{-1}e^{ib\tau_n \circ h_{l,j}}\rho \circ h_{l,j} |( h_{l,j})'|.
\]
However, if one of the images is of size greater than $\ve_0$, it must be shortened. In this case we may choose a cutting interval $U$, with $\ve_0/3 \leq |U| \leq \ve_0$ that does not cut the overlap if $\ovl<\ve_0/3$. We also require that the cutting does not create other pieces of length $< \ve_0/3$. This can be done if $\ovl < \ve_0/3$ and if $\ovl \geq \ve_0/3$, we can consider a smaller overlap interval of size $< \ve_0/3$. With these considerations, we have obtained two standard pairs such that
\[
I_j:=f^n(O_{h_{l,j}})\cap U , \rho_j := z_{ h_{l,j}}^{-1}e^{ib\tau_n \circ h_{l,j}}\rho \circ h_{l,j} |( h_{l,j})'|,
\]
and such that $O_{h_{l,1}},O_{h_{l,2}} \subset I$ and $I_1 \cap I_2$ contains an interval of length $\ovl$ on which holds
\[
\abs{\left(\tau_n \circ h_{l,1} - \tau_n \circ h_{l,2} \right)'} > C_1.
\]
\end{proof}
\begin{rem}
The actual value of $3\delta$ for which the condition above is used, is determined in \Cref{family_decay}. Also, note that for $n>n_\delta$, \Cref{finite_transversality} still holds but depends on $n$. So as long as we keep $n$ fixed, we may use \Cref{finite_transversality}, repeatedly.
\end{rem}
\section{Weight reduction of standard families} \label{modi}
In this section our goal is to replace a standard family, after certain number of iterations, with an equivalent standard family of lower total weight.
\begin{defin}[Equivalence]
Two standard families $\cG$ and $\tilde \cG$ are said to be \emph{equivalent} if $\rho_\cG = \rho_{\tilde\cG}$, i.e. if
\begin{equation}
\sum_{j \in \cJ} w_{\cG}(j) \rho_j = \sum_{k \in \tilde \cJ} w_{\tilde \cG}(k) \tilde \rho_k.
\end{equation}
\end{defin}
\begin{rem}
In this section we need to slightly increase the value of the parameter $a$. More precisely, we need $\left(e^{-\lambda n} + C_\tau/a + C_\kappa/a \right)\alpha_1^{-1} <1$.
This does not cause any problems since we could have chosen $a$ larger to begin with in \Cref{fixed_param}. In regards to $n$, we need $n > n_\delta$ and we choose it large enough that the above inequality holds and also $ae^{-\lambda n} < C_1/4$. Finally, we assume that $\abs{b} \geq 4\pi/(C_1 \ovl)$, where $C_1$ and $\ovl$ are related to transversality and were defined in the \Cref{trans}.
\end{rem}
\begin{lem} \label{singlepair_decay}
Suppose $\cG = \{(I, \rho)\} \in \cM_{a,b,B,\ve_0}$ is a singleton standard family with $\delta \leq \abs{I}$ for which \Cref{finite_transversality} holds. Then there exists there exists constants $\gamma>0$, $C$ such that for large $b$, letting $n_b=C\ln |b|$, there exists a standard family $\cG_{n_b}^*$ equivalent to $\cG_{n_b}$ such that
\[
\sum_{j \in \cJ_{n_b}^*} w_{\cG_{n_b}^*}(j) \leq e^{-\gamma}w_\cG.
\]
\end{lem}
\begin{proof}
Let $(I_1,\rho_1),(I_2,\rho_2) \in \cG_{n}$ be the transversal standard pairs provided by \Cref{finite_transversality} applied to $\cG =\{(I, \rho)\}$. Let $w_1=w_{\cG_{ n}}(h_1)$ and $w_2=w_{\cG_{ n}}(h_2)$ be the weights of these standard pairs.\footnote{We are using $\cH^{ n}$ for the index set of $\cG_{ n}$ to keep the notation simpler. To be precise, write $w_{\cG_{ n}}(j_n)$ with $j_{ n} \in \cJ_n$ as defined above.} Let $I_*$ be the interval of length $\ovl$ on which \Cref{finite_transversality} holds. Let $\theta_1=\arg(\rho \circ h_1)$ and $\theta_2=\arg(\rho \circ h_2)$. Then, on the interval $I_*$, we may write $\abs{w_1\rho_1 + w_2\rho_2} = \abs{e^{i \Theta_b}w_1\abs{\rho_1} + w_2\abs{\rho_2}}$,
where
\begin{equation}
\Theta_b= b(\tau_n \circ h_1 - \tau_n \circ h_2) - (\theta_1 - \theta_2).
\end{equation}
Our goal is to take out $\rho_1$ and $\rho_2$ from the family $\cG_n$ and replace them with other standard pairs (formed by combining parts of $\rho_1$ and $\rho_2$) and obtain a standard family $\cG_n^*$ which is still equivalent to $\cG_n$, but has a total weight strictly less than that of $\cG_n$.
We will first show that the phase difference $\Theta_b$ grows at a certain rate.
\begin{claim}[Full phase oscillation]
For large $n$, there exists $C_1$ such that for $b \neq 0$, on the interval $I_*$, holds
\begin{equation} \label{eq:phase_bounds}
\frac{|b|C_1}{2} \leq \abs{\Theta_b'} \leq 2|b|(C_\tau +C_1) .
\end{equation}
\end{claim}
\begin{proof}
Note that by \eqref{eq:arg_regularity}, on $I_*$,
\begin{equation} \label{eq:internal_phase}
\begin{split}
\abs{\theta_1'}&=\abs{\arg {(\rho \circ h_1)}'} \leq a|b|h_1'| \leq a|b| e^{-\lambda n},\\
\abs{\theta_2'} &\leq a|b| e^{-\lambda n}.
\end{split}
\end{equation}
Choose $n$ large enough\footnote{In addition to previous constraints.} that
\begin{equation}
ae^{-\lambda n} < C_1/4.
\end{equation}
Then, $\abs{\theta_1 - \theta_2} < |b|C_1/2$. Also, \Cref{finite_transversality} implies $\abs{b(\tau_n \circ h_1 - \tau_n \circ h_2)'}> |b|C_1$ hence $\abs{\Theta_b'} > |b|C_1/2$. Finally, a simple estimate shows that $\abs{\Theta_b'} \leq 2|b|(C_\tau +C_1)$.
\end{proof}
Note that since $\Theta_b$ is $\sC^1$ and satisfies the bounds \eqref{eq:phase_bounds}, $\Theta_b'$ does not change sign in $I_*$. Divide the range of $\Theta_b$ into intervals of length between $2\pi$ and $3\pi$, then the bounds on $\Theta_b'$ imply that $I_*$ will be divided into corresponding intervals $I_m$ of length $K_1 |b|^{-1}\leq \abs{I_m} \leq K_2 |b|^{-1}$, where $K_1:=\pi/(C_\tau+C_1)$ and $K_2:=6\pi/C_1$. To clarify, these are intervals on which $\Theta_b$ makes one full oscillation, but less than one and a half full oscillations. Of course we must make sure $I_*$ is large enough to fit at least one such interval $I_m$. That is we need $K_1\abs{b}^{-1} \leq \abs{I_*}$. This can be accomplished by choosing $b$ large enough:\footnote{This is the only restriction on $b$.}
\begin{equation}\label{eq:b_0}
\abs{b} \geq K_1/\abs{I_*}=4\pi/(C_1\ovl):=b_0.
\end{equation}
We like to combine some part of $\rho_1$ and $\rho_2$ to take advantage of their cancellation. Since the modulus of these standard pairs are not smooth, if we combine them blindly, we might lose the $\sC^1$-smoothness required for the argument of a standard pair. For this reason we do the following splitting of the standard pairs into good parts, with a constant modulus, which we can combine; and bad parts, which we do not combine in this round.
For a function $\rho\in \L^1(I,\bC)$ with $\int_I\abs{\rho} \neq 0$, denote $N(\rho)=\rho/\int_I\abs{\rho}$. For $j\in \{1,2\}$, we split $(I_j,\rho_j)$ into two standard pairs $
\left(I_j,N(\tilde\rho_j) \right) $, $\left(I_j,N(\bar \rho_j) \right)$, such that:
\begin{equation}\label{eq:split}
\bar \rho_j = c e^{i \Theta_j},
\tilde\rho_j = (\abs{\rho_j}-c)e^{i\Theta_j}, \text{ where } c= \frac{e^{-a}}{2}, \Theta_j = b(\tau_n \circ h_j)+ \theta_j.
\end{equation}
Associate to them the weights $\bar w_j = w_j \int_{I_j}\abs{\bar \rho_j}$, $\tilde w_j = w_j \int_{I_j}\abs{\tilde \rho_j}$.
\begin{claim}[After splitting]
For $j \in \{1,2\}$,
\begin{eqnarray}
\rho_j &=& \bar \rho_j + \tilde \rho_j\\
w_j &=& \bar w_j + \tilde w_j \\
w_j \rho_j &=& \bar w_j N(\bar \rho_j)+ \tilde w_j N(\tilde \rho_j)\\
H(\bar \rho_j) &\leq& a \\
H(\tilde \rho_j) &\leq& 4a \label{eq:remainder_reg}\\
\abs{\Theta_j'} &\leq& a|b| \left(e^{-\lambda n} + \frac{C_\tau}{a}\right) \label{eq:new_phase}
\end{eqnarray}
\end{claim}
\begin{proof}
The first four statements are easy to prove. To prove \eqref{eq:remainder_reg}, note that $c =(1/2)e^{-a} \leq (1/2)\inf \abs{\rho_j}$.
Therefore,
\begin{equation}\label{eq:trick}
\frac{\abs{\abs{\rho_j(x)}-c}}{\abs{\abs{\rho_j(y)}-c}}\leq \frac{\abs{\abs{\rho_j(x)}-\abs{\rho_j(y)}}+ \abs{\rho_j(y)} - c}{\abs{\rho_j(y)}-c} = 1 + \frac{\abs{\abs{\rho_j(x)}-\abs{\rho_j(y)}}}{\abs{\rho_j(y)}-c}
\end{equation}
But, $\abs{\rho_j(y)} - (1/2) \abs{\rho_j(y)} = (1/2)\abs{\rho_j(y)} \geq \inf\abs{\rho_j} \geq c$. Hence, $
\abs{\rho_j(y)} - c \geq \frac{1}{2}\abs{\rho_j(y)}$,
and we have:
\begin{equation}
\begin{split}
\frac{\abs{\abs{\rho_j(x)}-c}}{\abs{\abs{\rho_j(y)}-c}} &\leq 1 + 2\frac{\abs{\abs{\rho_j(x)}-\abs{\rho_j(y)}}}{\abs{\rho_j(y)}} \leq 1+ 2\abs{\frac{\abs{\rho_j(x)}}{\abs{\rho_j(y)}}- 1}\\
&\leq 1 + 2 \abs{e^{a\abs{x-y}^\alpha} - 1} = 1+ 2 (2a|x-y|^\alpha)\\
&\leq e^{4a\abs{x-y}^\alpha}.
\end{split}
\end{equation}
The inequality \eqref{eq:new_phase} follows from \eqref{eq:internal_phase} and \eqref{eq:roof_regularity}.
\end{proof}
We now describe how to combine $\bar\rho_1$ and $\bar\rho_2$. Note that the modulus of these functions is constant and equal to $c$. We need the following result.
\begin{claim}[$J_m$. Preparing for a controlled cancellation of $\bar \rho_1$ and $\bar \rho_2$] \label{cases}
Suppose $w_2 \leq w_1$.\footnote{Otherwise, interchange indices and do the same proof.} For every $\alpha_1 \in (0,1/2]$ and $\alpha_2 \in [(\sqrt{7}-1)/2,1)$, for every $m$, there exists a subinterval $J_m \subset I_m$ with $K_3|b|^{-1}\leq \abs{J_m} \leq K_4\abs{b}^{-1}$ such that for every $\kappa_0 \geq w_2/(2w_1)$
\begin{equation} \label{eq:cancelation}
\alpha_1 c\left(\kappa_0 w_1+w_2 \right) \leq \abs{\kappa_0 w_1\bar \rho_1+w_2 \bar \rho_2} \leq c(\kappa_0 w_1 + \alpha_2 w_2).
\end{equation}
\end{claim}
\begin{proof}
Choose $K_3, K_4$ such that $1/4 \leq \cos(\Theta_b) \leq 1/2$ on $J_m$. This can be done because the phase difference $\Theta_b$ makes a full oscillation in $I_m$.
The left side of \eqref{eq:cancelation} is easy to prove and does not require a restriction on $\kappa_0$. Let us prove the right side.
Note that, on one hand, using $\cos(\Theta_b) \leq 1/2$,
\[
\begin{split}
\abs{\kappa_0 w_1\bar \rho_1+w_2 \bar \rho_2}^2 &= \kappa_0^2w_1^2\bar\rho_1^2+w_2^2\bar\rho_2^2+ 2\kappa_0 w_1w_2\abs{\bar\rho_1}\abs{\bar \rho_2} \cos(\Theta_b)\\
&=c^2\left(\kappa_0^2w_1^2+w_2^2+ 2\kappa_0 w_1w_2 \cos(\Theta_b)\right) \\
&\leq c^2\left(\kappa_0^2w_1^2+w_2^2+ \kappa_0 w_1w_2 \right).
\end{split}
\]
On the other hand,
\[
\left(c( \kappa_0 w_1 + \alpha_2 w_2)\right)^2 = c^2\left(\kappa_0^2w_1^2+ \alpha_2^2 w_2^2 + 2 \alpha_2\kappa_0 w_1 w_2\right).
\]
Hence it suffices to show
\[
0 \leq w_2 (\alpha_2^2 - 1) + \kappa_0 w_1(2\alpha_2 - 1).
\]
Solving for $\alpha_2$, it suffices to show
\begin{equation}\label{eq:alpha_2}
\alpha_2 \geq \sqrt{1+\frac{\kappa_0 w_1}{ w_2} + \left(\frac{\kappa_0 w_1}{w_2}\right)^2}- \frac{\kappa_0 w_1}{w_2}.
\end{equation}
If $w_2 / (2w_1) \leq \kappa_0 $, then $ (\kappa_0 w_1)/w_2 \geq 1/2$. From the graph of $x\mapsto \sqrt{1+x+x^2}-x$, for $x \in [0,\infty)$, it follows that the right hand side of \eqref{eq:alpha_2} is at most $(\sqrt{7}-1)/2$. Therefore \eqref{eq:alpha_2} is satisfied for any $\alpha_2 \geq (\sqrt{7}-1)/2$.
\end{proof}
Fix $\kappa_0=w_2/(2w_1)$. Choose a smooth function $\kappa \in \sC^1(I_*, [1-\kappa_0,1])$ such that $\kappa =1-\kappa_0$ on the middle third of $J_m$, denoted $J_m'$, and $\kappa = 1$ outside $J_m$. Note that, taking the length of a connected component of $J_m\setminus J_m'$ into account, $\kappa$ can be chosen such that $\abs{\kappa'}<C_\kappa \kappa_0 |b|$ on $I_m$. Define\footnote{Note that we are assuming $w_2 \leq w_1$.}
\begin{equation}\label{eq:new_pairs}
\bar\rho_{1*}:= \kappa \bar \rho_{1} \text{, and } \bar\rho_{2*} := \bar \rho_{2} + (1-\kappa)(w_1/w_2)\bar \rho_{1}.
\end{equation}
The domain of the definition above is the overlap interval $I_*$; however, for $j\in \{1,2\}$, we may extend the domain of $\bar\rho_{j*}$ to the interval $I_j$ so that $\bar\rho_{j*} = \bar\rho_j$. This should be clear from the definition of $\kappa$ and \eqref{eq:new_pairs}. We intend to replace $\bar \rho_1, \bar\rho_2$ with $\bar \rho_{1*}, \bar \rho_{2*}$. We will not touch $\tilde \rho_1, \tilde \rho_2$ except to normalize them.
Define a new family
\begin{equation}
\begin{split}
\cG_{n}^* &:=\left(\cG_{ n} \setminus \{(I_1,\rho_1), (I_2,\rho_2)\} \right) \cup \\ & \left\{\left(I_1,N(\bar\rho_{1*})\right), \left(I_2,N(\bar\rho_{2*})\right),\left(I_1,N(\tilde\rho_{1})\right), \left(I_2N(\tilde\rho_{2}\right)\right\}.
\end{split}
\end{equation}
with associated weight measure $w_{\cG_n^*}$ that is the same as $w_{\cG_n}$ except for the modified standard pairs. For the modified standard pairs, define the new weights by $\bar w_{1*} := w_1\int_{I_1}\abs{\bar \rho_{1*}}$, $\bar w_{2*} := w_2 \int_{I_2}\abs{\bar \rho_{2*}}$, $\tilde w_{1} := w_1\int_{I_1}\abs{\tilde \rho_{1}}$, $\tilde w_{2} := w_2 \int_{I_2}\abs{\tilde\rho_{2}}$.
Now we check that the new collection $\cG_n^*$ is a standard family equivalent to $\cG_n$.
\begin{claim}[After cancellation]
We have:
\begin{eqnarray}
w_1\rho_1 + w_2 \rho_2 &=& w_1 \bar \rho_{1*} + w_1 \tilde \rho_1 + w_2 \bar \rho_{2*} + w_2 \tilde \rho_2. \label{eq:equival}\\
H(\bar \rho_{j*}) &\leq& a |b|, \forall j \in \{1,2\} \label{eq:mod_reg_new}\\
\abs{\Theta_{j*}'} &\leq& a|b|, \forall j \in \{1,2\} \label{eq:arg_reg_new}\\
\abs{\partial_\ve \cG_n^*} &\leq& C_* \abs{\partial_\ve\cG_n} \label{eq:bd_of_mod}.
\end{eqnarray}
\end{claim}
\begin{proof}
The equality \eqref{eq:equival} follows from definition. Indeed, \eqref{eq:new_pairs} implies
\[
w_1\bar \rho_{1*} + w_2 \bar \rho_{2*}=w_1\bar \rho_1 + w_2 \bar \rho_2,
\]
which in turn implies \eqref{eq:equival}. Hence $\cG_n^*$ and $\cG_n$ are equivalent.
To prove \eqref{eq:mod_reg_new}, note that by construction $\bar\rho_{1*}, \bar\rho_{2*}$ are $\sC^1$. Hence it suffices to show $\abs{\bar\rho_{j*}'} \leq a|b| \abs{\bar\rho_{j*}}$.
For $\bar\rho_{1*}$ this condition is easier to check than for $\bar\rho_{2*}$. Let us check, for $\bar\rho_{2*}$ the stronger condition: $\abs{\bar\rho_{2*}'} \leq a|b| \abs{\bar\rho_{2*}}$ for $a$ and $n$ large enough.\footnote{The previous choices of $a$ and $n$ need to be updated.} Outside $J_m$, $\bar \rho_{2*}$ satisfies this condition because $\bar\rho_{2}$ does. On $J_m$, differentiating $\bar\rho_{2*}$ and using \eqref{eq:split} and \eqref{eq:new_phase} yield,
\[
\begin{split}
\abs{\bar\rho_{2*}'} &\leq c \left( \abs{\Theta_2'} + \abs{1- \kappa}\frac{w_1}{w_2} \abs{\Theta_1'}\right) + \frac{w_1}{w_2}\abs{\kappa'}\abs{\rho_{1}} \\
&\leq a|b|\left( e^{-\lambda n} + \frac{C_\tau}{a}\right) c\left(1 + \abs{1- \kappa}\frac{w_1}{w_2} \right) + c \frac{w_1}{w_2}\abs{\kappa'}.
\end{split}
\]
Now observe that since $K_3|b|^{-1}\leq \abs{J_m} \leq K_4|b|^{-1}$ and $1-w_2/(2w_1) \leq \kappa \leq 1$, we have $\abs{\kappa'}< C_\kappa w_2/(2w_1)|b|$. Also, $\abs{1-\kappa} \leq w_2/(2w_1)$. Therefore,
\[
\frac{w_1}{w_2}\abs{\kappa'} \leq C_\kappa \left(1+\abs{1-\kappa}\frac{w_1}{w_2}\right)|b|
\]
Therefore,
\[
\begin{split}
\abs{\bar\rho_{2*}'} &\leq a|b|\left( e^{-\lambda n} + \frac{C_\tau}{a} + \frac{C_\kappa}{a}\right)c\left(1 + \abs{1- \kappa}\frac{w_1}{w_2} \right) \\
&\leq a|b|\left( e^{-\lambda n} + \frac{C_\tau}{a} + \frac{C_\kappa}{a}\right) \frac{\abs{\bar\rho_{2*}}}{\alpha_1},
\end{split}
\]
where the last inequality follows from the left hand side of \eqref{eq:cancelation}. Recall that the left hand side of \eqref{eq:cancelation} requires no restriction on $\kappa_0$. It simply follows from the phase difference satisfying $\cos(\Theta_b)=\cos (\Theta_1-\Theta_2) \geq 1/4$ and $\alpha_1 < 1/2$.
Take $a,n$ large to conclude.
The inequality \eqref{eq:arg_reg_new} is a consequence of $\abs{\bar \rho_{j*}'} \leq a|b| \abs{\bar\rho_{j*}}$ since $\abs{\Theta_{j*}'} \leq \abs{\bar \rho_{j*}'}/\abs{\bar\rho_{j*}}$.
To prove \eqref{eq:bd_of_mod}, note that
\[
\abs{\partial_\ve \cG_n^*} \leq \abs{\partial_\ve\cG_n} + \int_{\partial_\ve I_1}w_1 \abs{\tilde \rho_1} + \int_{\partial_\ve I_2}w_2 \abs{\tilde \rho_2} + \int_{\partial_\ve I_1}w_1\abs{\bar\rho_{1*}} + \int_{\partial_\ve I_2}w_2\abs{\bar\rho_{2*}}.
\]
The first three terms are simply bounded by $3\abs{\partial_\ve \cG_n}$. The last two terms are bounded by $\abs{\partial_\ve \cG_n}+ \int_{\partial_\ve I_*} (w_1 \abs{\bar \rho_{1*}} + w_2 \abs{\bar \rho_{2*}})$. Note that using \eqref{eq:new_pairs}, $w_1 \abs{\bar \rho_{1*}} + w_2 \abs{\bar \rho_{2*}} \leq w_1\abs{\bar \rho_1}+ w_2\abs{\bar \rho_1}$. Putting all this together and noting that $\bar\rho_1=\bar\rho_2=c$, we have
\[
\begin{split}
\abs{\partial_\ve \cG_n^*} &\leq 3\abs{\partial_\ve\cG_n} + \abs{\partial_\ve\cG_n}+ \int_{\partial_\ve I_*}(w_2\abs{\bar\rho_{2}} +w_1 \abs{\bar \rho_1} ) \\
&\leq 4\abs{\partial_\ve\cG_n} + c \ve (w_1+w_2)\\
&\leq 5c \abs{\partial_\ve\cG_n}.
\end{split}
\]
\end{proof}
Let us check that the total weight of the new standard family is less than the old one.
\begin{claim}
$
\sum_{j \in \cJ_{n_b}'} w_{\cG_{n_b}'}(j) \leq e^{-\gamma}w_\cG.
$
\end{claim}
\begin{proof}
First, note that on $J_m'$, $\kappa = 1- \kappa_0$. Therefore on $J_m'$ (and using the right hand side of \eqref{eq:cancelation}) we have
\[
\begin{split}
w_1\abs{\bar \rho_{1*}} + w_2\abs{\bar \rho_{2*}} &\leq w_1 (1-\kappa_0) \abs{\bar \rho_{1}}+ w_2\abs{\bar \rho_{2} + \kappa_0\frac{w_1}{w_2}\bar \rho_{1}}\\
&\leq w_1 (1-\kappa_0) \abs{\bar \rho_{1}}+ \alpha_2 w_2\abs{\bar \rho_{2}} + w_1\kappa_0\abs{\bar \rho_{1}} \\
&= w_1 \abs{\bar \rho_1} + \alpha_2 w_2 \abs{\bar\rho_2}.
\end{split}
\]
Outside $J_m'$, by definition, $w_1\abs{\bar \rho_{1*}}+w_2\abs{\bar \rho_{2*}} \leq w_1\abs{\bar \rho_{1}}+w_2\abs{\bar \rho_{2}}$.
Then,
\[
\begin{split}
\int _{I_*} \left(w_1\abs{\bar \rho_{1*}} + w_2\abs{\bar \rho_{2*}} \right) &\leq \sum_{m } \int_{I_m \setminus J_m'} \left(w_1 \abs{\bar \rho_1} + w_2 \abs{\bar \rho_2}\right) + \int_{J_m'} \left( w_1\abs{\bar \rho_{1}} + \alpha_2 w_2\abs{\bar \rho_{2}} \right) \\
&= \int_{I_*}w_1 \abs{\bar \rho_1} + \sum_{m }\int_{I_m \setminus J_m'}w_2 \abs{\bar \rho_2} + \int_{J_m'} \alpha_2 w_2\abs{\bar \rho_{2}}\\
&= \int_{I_*}w_1 \abs{\bar \rho_1} + \sum_{m }\int_{I_m}w_2 \abs{\bar \rho_2} -(1-\alpha_2) \int_{J_m'} w_2\abs{\bar \rho_{2}} .
\end{split}
\]
Since $K_1|b|^{-1}\leq \abs{I_m} \leq K_2|b|^{-1}$ and $(1/3)K_3|b|^{-1}\leq \abs{J_m'} \leq (1/3)K_4|b|^{-1}$, and $\abs{\bar \rho_2}$ is a constant, \Cref{Fed} implies that $\int_{J_m'} w_2\abs{\bar \rho_{2}} \geq (\abs{J_m'}/\abs{I_m}) \int_{I_m}w_2 \abs{\bar \rho_2} \geq K_3/(3K_2) \int_{I_m}w_2 \abs{\bar \rho_2}$. Therefore,
\[
\int _{I_*} \left(w_1\abs{\bar \rho_{1*}} + w_2\abs{\bar \rho_{2*}} \right) \leq \int_{I_*}w_1 \abs{\bar \rho_1} + \sum_{m }\int_{I_m}w_2 \abs{\bar \rho_2} \left(1-(1-\alpha_2)K_3/(3K_2) \right)
\]
Set $\alpha_2':= 1-(1-\alpha_2)K_3/(3K_2)$. Note that $0< \alpha_2'<1$ and it does not depend on $m$. Pulling it out of the integral, we have
\[
\int _{I_*} \left(w_1\abs{\bar \rho_{1*}} + w_2\abs{\bar \rho_{2*}} \right) \leq \int_{I_* } w_1 \abs{\bar\rho_1} + \alpha_{2}'\int_{I_*}w_2 \abs{\bar \rho_2} .
\]
Add $\int_{I_1 \setminus I_*}w_1\abs{\bar \rho_{1*}} + \int_{I_2 \setminus I_*}w_2\abs{\bar \rho_{2*}} = \int_{I_1\setminus I_*}w_1\abs{\bar\rho_{1}} + \int_{I_2 \setminus I_*}w_2\abs{\bar \rho_2} $ to the above inequality and apply the same argument (using \Cref{Fed}) to conclude that there exists $0<\alpha_2''<1$ such that
\[
\bar w_{1*} + \bar w_{2*} = w_1\int _{I_1} \abs{\bar \rho_{1*}} + w_2\int_{I_2}\abs{\bar \rho_{2*}} \leq w_1\int_{I_1}\abs{\bar \rho_1} + \alpha_2'' w_2\int_{I_2}\abs{\bar \rho_2} = \bar w_1 + \alpha_2'' \bar w_2.
\]
Observe that since $\alpha_2''<1$ and the rest of the standard pairs in the family were not modified, the total weight of the new standard family is less than the original one. Estimating the total weight more precisely, for large $n$,
\[
\begin{split}
|\cG_n^*| &=
|\cG_n|-(w_1+w_2) + \bar w_{1*} + \tilde w_1 +\bar w_{2*} + \tilde w_2 \\
&\leq |\cG_n|-(w_1+w_2) + \bar w_{1} + \tilde w_1 + \alpha_2'' \bar w_{2} + \tilde w_2 \\
&\leq |\cG_n| - (1- \alpha_2'')\bar w_2 \\
&\leq w_{\cG} - (1- \alpha_2'')CM(n) w_\cG \\
&\leq e^{-\gamma} w_{\cG}.
\end{split}
\]
Recall that $w_{\cG}$ is the weight of the original standard family consisting of a single standard pair. In the next to last inequality we used the lower bound on $w_2$. Indeed, by definition of $\bar w_2$, \Cref{iteration}, change of variables, and finally the definition of $M(n)$ (see \Cref{finite_transversality}), we have
\[
\begin{split}
\bar w_2 = w_2 \int_{I_2} |\bar \rho_2| = c |I_2|w_2 &\geq c|I_2|w \int_{I_2} \abs{\rho} \circ h_2 \abs{h_2'} \\
&\geq c \ovl w \inf_I \abs{\rho} \abs{(I \cap O_{h_2}) \cap h_2(U_\ell)}\\
&\geq c \ovl w\inf_I \abs{\rho}M(n) := CM(n) w_\cG.
\end{split}
\]
\end{proof}
There is one last issue to resolve. The members of $\cG_{n}^*$ may not satisfy $H(\rho) \leq a $. In order to achieve this, we simply iterate the family for a time $C\ln|b|$. Indeed, suppose $(I, \rho) \in \cG_{n}^*$. Note that $H(\rho) \leq a |b|$. Following the proof of property \eqref{eq:mod_reg_1} in \Cref{invariance}, note that after $\tilde n$ more iterations, every image pair $\tilde \rho \in \cG_{n+\tilde n}^*$ satisfies:
\[
H(\tilde \rho) \leq a \left(|b| e^{-\lambda \tilde n} + \frac{D}{a}\right).
\]
Now, it is clear that, for $b \geq b_0$ if
$\tilde n > \lambda^{-1}\ln\left( |b|(1-Da^{-1})^{-1}\right)=:\tilde n_b$,
then $H(\tilde\rho) \leq a$. Finally, note that if $C$ is chosen large enough, then $n_b:=C\ln|b|$ dominates $n+\tilde n_b$. We have shown the existence of the standard family $\cG_{n_b}^*$ as claimed in \Cref{singlepair_decay}. \end{proof}
\begin{prop} \label{family_decay}
There exists $0< \gamma_1 < \gamma $ and for $|b| \geq b_0$ there exists $n_b=C \ln|b|$, such that for any standard probability family $\cG \in \cM_{a,b,B,\ve_0}$, there exists a standard family $\cG_{n_b}^* \in \cM_{a,b,B,\ve_0}$ equivalent to $\cG_{n_b}$ such that $\abs{\cG_{n_b}^*} \leq e^{-\gamma_1}$.
\end{prop}
\begin{proof}
Choose $\delta>0$ small enough that $B \delta < 1/4$. Then for large $m$, $\abs{\partial_\delta \cG_m} \leq B\delta <1/4$. It follows that the total weight of the standard pairs of $\cG_{m}$ that have width larger than $\delta$ is $L\geq 3/4$. Let $S$ denote the total weight of standard pairs that have width $\leq \delta$. Note that $\abs{\cG}=\abs{\cG_m}=S+L = 1$. Fix $\tilde n_b$ as in \Cref{singlepair_decay}. Using \Cref{singlepair_decay},
\[
\abs{\cG_{m+\tilde n_b}^*} \leq S + e^{-\gamma} L \leq (S+L)-(1-e^{-\gamma})L \leq 1-(1-e^{-\gamma})3/4 =: e^{-\gamma_1}.
\]
Take $n_b = m +\tilde n_b$.
\end{proof}
\section{Proofs of the main propositions} \label{dens_decay}
\begin{lem} \label{Lb_decay} There exists $C>0$, $\gamma_2>0$, such that if $\{(I,\rho)\} \in \cM_{a,b,B,\ve_0}$ with $a,b,B$ sufficiently large and $\ve_0$ sufficiently small, then for every $n \in \bN$,
\begin{equation}
\norm{ \sL_b^n \rho}_{\L^1 } \leq C e^{-\frac{\gamma_2}{\ln{|b|}} n}.
\end{equation}
\end{lem}
\begin{proof}
The result follows by repeatedly applying \Cref{family_decay} and renormalizing the total weight of the standard family at every step. Indeed, let $n_b=C\ln|b|$ as in \Cref{family_decay}. After $k+1$ repetitions we have
\[
\begin{split}
\norm{ \sL_b^{(k+1)n_b} \rho}_{\L^1 } &\leq \sum_{j_{k+1}\in \cJ_{k+1}} w_{\cG_{(k+1)n_b}^*}(j_{k+1}) \\
&\leq \sum_{j_k \in \cJ_k} e^{-\gamma_1} w_{\cG_{kn_b}^*}(j_k) \\
&\leq e^{-(k+1)\gamma_1}
\text{.}
\end{split}
\]
This means, for every $m \in \bN$ ($m=kn_b+r_b$, $0 \leq r_b<n_b$),
\[
\begin{split}
\norm{ \sL_b^{m} \rho}_{\L^1 } &\leq \norm{ \sL_b^{r_b} \sL_b^{kn_b}\rho}_{\L^1 } \leq \norm{ \sL_0^{r_b}\abs{ \sL_b^{kn_b}\rho}}_{\L^1 } = \norm{ \sL_b^{kn_b}\rho}_{\L^1 } \\
&\leq e^{k\gamma_1} \leq e^{-(\frac{m}{n_b}-1) \gamma_1}\leq e^{\gamma_1}e^{-\frac{\gamma_1}{n_b}m} \leq C_{\gamma_1}e^{-\frac{\gamma_1}{n_b}m}.
\end{split}
\]
Note that here by $\sL_b^0\rho$, we mean $\rho$. Set $\gamma_2=\gamma_1/C$, where $n_b = C\ln|b|$, to conclude.
\end{proof}
\begin{proof}[Proof of \Cref{main_estimate}] \label{prf_main_estimate}
We will show that there exists a constant $C$, such that for every $b \geq b_0$, for every $n \in \bN$,
\begin{equation}
\norm{ \sL_b^n}_{\mixnorm} \leq Ce^{-\frac{\gamma_2}{\ln{|b|}} n}.
\end{equation}
Write $g=g - c + c$, where $c = 1+ |g|_\alpha/a+\sup|g|$ ($\abs{g}_\alpha$ being the $\alpha$-H\"older constant of $g$). Note that both $|c|$ and $\abs{g-c}=1+|g|_\alpha/a+\sup{\abs{g}}-g$ are bounded below by $1$. Of course, $H(c)=0$. Calculating as in \eqref{eq:trick}, we get
\[
\frac{\abs{g(x)-c}}{\abs{g(y)-c}} \leq \frac{\abs{g(x)-g(y)}}{\abs{g(y)-c}}+1 \leq \frac{\abs{g}_\alpha \abs{x-y}^\alpha}{\abs{g(y)-c}} +1 \leq e^\frac{\abs{g}_\alpha \abs{x-y}^\alpha}{\abs{g(y)-c}}.
\] By the choice of $c$, it follows that $\abs{g}_\alpha/(\abs{g(y)-c}) \leq a$. Therefore, $H(g-c) \leq a$.
Let $g_c=g-c$. Partition the domains of $g_c$ and $c$ into intervals of length $\ve_0>0$ (small enough as in \Cref{fixed_param}) and renormalize the restricted functions. Then $g_c=\sum_{j=1}^{L} w_j \rho_j$ and $c=\sum_{k=1}^{L} \tilde w_k \tilde \rho_k$, where $L=\lceil\ve_0^{-1}\rceil$, $\{\rho_j\},\{\tilde\rho_k\}$ are standard families with parameters $a,b,B,\ve_0$ as defined before; and $\{w_j\}, \{\tilde w_k\}$ are their associated weights. Then $g = \sum w_j \rho_j + \sum \tilde w_k \tilde \rho_k$, where $\sum w_j+ \sum \tilde w_k = \norm{g_c}_{\L^1} + \norm{c}_{\L^1}$. Apply \Cref{Lb_decay} and obtain for $b \geq b_0$, for every $n \in \bN$,
\[
\begin{split}
\norm{ \sL_b^n g}_{\L^1} &\leq \sum_{j=1}^L w_j \norm{ \sL_b^n \rho_j}_{\L^1} + \sum_{j=1}^L \tilde w_k \norm{ \sL_b^n \tilde \rho_k}_{\L^1} \\
&\leq C_{\gamma_1}e^{-\frac{\gamma_2}{\ln{|b|}} n}\left(\sum_{j=1}^L w_j+ \sum_{j=1}^L \tilde w_k \right) \\
&\leq C_{\gamma_1}e^{-\frac{\gamma_2}{\ln{|b|}} n}\left(\norm{g_c}_{\L^1}+\norm{c}_{\L^1}\right) \\
& \leq CC_{\gamma_1} e^{-\frac{\gamma_2}{\ln{|b|}} n} \norm{g}_{\sC^\alpha}
\end{split}
\]
\end{proof}
\begin{proof}[Proof of \Cref{small_b}] \label{prf_small_b}
It suffices to show that $\sL_b$ has spectral radius $e^{-r}<1$ when $b \neq 0$. By the assumptions on $\B$, we know that $\sL_b$ has essential spectral radius $e^{-r'}<1$ and spectral radius at most $1$. This means that there are only finitely many eigenvalues outside the disk of radius $e^{-r'}$. Therefore, if we show that there are no eigenvalues on the unit circle, it follows that the spectral radius is strictly less than $1$. In turn, this implies the existence of $C$ such that $\norm{\sL^n_b}_{\B} < Ce^{-rn}$.
Let us show that there are no eigenvalues of $\sL_b$ on the unit circle for $b \neq 0$. Fix $b \geq b_0$ and suppose that there exists $g$ and $\lambda \in \bC$ satisfying $\abs{\lambda}=1$ such that
$\sL_b g = \lambda g$. Since $\abs{\sL_b g} \leq \sL_0 \abs g$, it follows that $\abs{g} \leq \sL_0 \abs{g}$. Since, $\int \abs{g} = \int \sL_0 \abs{g}$, this implies that
$\abs{g} = \sL_0 \abs{g}$. Since $\abs{\lambda}=1$, this means that $
\abs{ \sL_{b} g } = \sL_0 \abs{g}$.
Using the definition of $\sL_b$ observe that, for every $y$, the arguments of the complex numbers $g(x) e^{ib\tau(x)}$ must be equal for all $x$ such that $f(x)=y$. Choose some $k\in \bN$ such that $b k > b_0$, where $b_0$ is as in \eqref{eq:b_0}. The arguments of the complex numbers $g^k(x) e^{ibk\tau(x)}$ are equal for all $f(x)=y$. This means that $ \abs{ \sL_{kb} g^k } = \sL_0 \abs{g^k}$. It also means that $ \abs{ \sL_{kb}^n g^k } = \sL_0^n \abs{g^k}$ for any $n\in \bN$ (we could have considered the $n$-th power from the start of the argument).
We have that
\[
\int \abs{ \sL_{kb}^n g^k } dm= \int \sL_0^n \abs{g^k} dm= \int \abs{g^k} dm\quad \quad \text{for all $n\in \bN$}.
\]
Using the estimate for large $|b|$, if $g \in \B$, then the left hand side vanishes as $n\to \infty$ whereas the right hand side is fixed and non-zero.
\end{proof}
|
\section{Introduction}\label{intro_sec}
This article concerns the
transverse-field Ising model, introduced
in~\cite{lsm}
and a well-known generalization of
the familiar (classical) Ising model for ferromagnetism.
The model possesses
a phase transition and a critical
point, which may be identified using the
\emph{(residual) magnetization}.
The magnetization equals zero
below the critical point and is positive above it.
An important result for the classical
Ising model is that
the magnetization also vanishes
\emph{at} the critical point: in two
dimensions this goes back to the work of
Onsager~\cite{onsager}, in dimension $d\geq4$ it
was first proved by Aizenman and
Fern\'andez~\cite{af}, and recently the final
case $d=3$ was established by Aizenman, Duminil-Copin
and Sidoravicius~\cite{adcs}.
Building on the methods of~\cite{adcs}, the present
work shows that the magnetization
in the transverse-field model also vanishes
at the critical point. This implies that there
is a unique equilibrium state at the critical point.
We give precise statements
shortly, but first introduce the relevant
notation and definitions.
Let $n\geq1$ and write
\[
\L=\L_n=[-n,n]^d=\{-n,-n+2,\dotsc,n-1,n\}^d
\]
for a finite box in $\mathbb{Z}^d$. The transverse-field
Ising model is defined via its Hamiltonian, which in the
finite volume $\L$ takes the form
\[
H_\L=-\l\sum_{xy\in\L} \sigma_x^{(3)}\sigma_y^{(3)}-\d\sum_{x\in\L} \sigma_x^{(1)}
-\g \sum_{x\in\L} \sigma_x^{(3)}.
\]
Here the first sum is over all (unordered)
nearest neighbours in $\L$,
\[
\sigma^{(1)}=\begin{pmatrix}
1 & 0\\
0 & -1
\end{pmatrix},
\qquad
\sigma^{(3)}=\begin{pmatrix}
0 & 1\\
1 & 0
\end{pmatrix}
\]
are the spin-$\tfrac{1}{2}$ Pauli matrices,
and $\sigma^{(i)}_x=\sigma^{(i)}\otimes\mathrm{Id}_{\L\setminus\{x\}}$.
The parameters $\l$ and $\d,\g$
are nonnegative and represent spin-coupling
and field-strengths, respectively.
$H_\L$ is an operator on the Hilbert space
$\bigotimes_{x\in\L}\mathbb{C}^2$, and one defines
for each $\b\in(0,\infty)$ the state
$\langle\cdot\rangle_{\L,\b}$ by
\begin{equation}\label{Q-state}
\langle Q\rangle_{\L,\b}=\frac{\mathrm{tr}(Qe^{-\b H_\L})}{\mathrm{tr}(e^{-\b H_\L})}.
\end{equation}
The parameter $\b$ is referred to as \emph{inverse temperature}.
(Readers with a probabilistic background may prefer the `path
integral' definition of $\langle\cdot\rangle_{\L,\b}$
given in Section~\ref{spin_sec}.)
Of particular interest are the one- and two-point functions
\[
\langle \sigma_x^{(3)}\rangle_{\L,\b}\quad\mbox{ and }\quad
\langle \sigma_x^{(3)}\sigma_y^{(3)}\rangle_{\L,\b},
\]
and more general correlation functions
\begin{equation}\label{corr_eq}
\langle \sigma_A^{(3)}\rangle_{\L,\b},\quad\mbox{ where }
\sigma_A^{(3)}=\prod_{x\in A}\sigma_x^{(3)}.
\end{equation}
As written, these are analytical functions
of the model parameters $\l,\d,\g$. However,
one is interested in their their
limits as $n\to\infty$ or $\b,n\to\infty$.
(Existence of the limits is well-known,
see eg~\cite{akn} or~\cite{bjo_phd}.)
These need not be analytical, or even continuous:
for example if $\g=0$ then by symmetry
$\langle \sigma_0^{(3)}\rangle_{\L,\b}=0$, whereas the residual
magnetization
\begin{equation}\label{resid_eq}
\begin{split}
M^+_\b(\l,\d)&:=\lim_{\g\downarrow0}\lim_{n\to\infty}
\langle \sigma_0^{(3)}\rangle_{\L,\b}\\\mbox{or }
M^+_\infty(\l,\d)&:=\lim_{\g\downarrow0}\lim_{n\to\infty}\lim_{\b\to\infty}
\langle \sigma_0^{(3)}\rangle_{\L,\b}
\end{split}
\end{equation}
may be strictly positive. This leads to the definition of the
\emph{critical point}
\[
\l_\crit=\l_\crit(\d,\b):=\inf\{\l\geq 0: M^+_\b(\l,\d)>0\}.
\]
Note that $\l_\crit$ may also be defined in terms of
the uniqueness of the infinite-volume states (for all
boundary conditions), or the divergence of the
susceptibility, see~\cite{bjogr}.
We have that $0<\l_\crit<\infty$ if $d\geq2$, or if
$\b=\infty$ and $d\geq 1$.
The case $\b=\infty$ is referred to as the \emph{ground state}
and the case $\b<\infty$ as \emph{positive temperature}.
The following is the main result of this work:
\begin{theorem}\label{main_thm}
Let $\d>0$.
If $\b=\infty$ and $d\geq2$, or $\b<\infty$ and $d\geq3$, then
the residual magnetization satisfies
\[
M^+_\b(\l_\crit,\d)=0.
\]
\end{theorem}
If $\d=0$ one recovers the
classical Ising model; it is then standard to take $\l=1$
and vary the parameter $\b$, giving the critical point $\b_\crit$
which is positive and finite if $d\geq2$.
As remarked above, in this case the result is well-known.
For $\d>0$ the case when $\b=\infty$ and $d=1$
is also known and was
established in~\cite{pfeuty}
(and reproved in~\cite{bjogr} using graphical methods).
The other cases are new. Note that the
only nontrivial case left open by Theorem~\ref{main_thm} is when
$\b<\infty$ and $d=2$, which remains open for $\d>0$.
Like the previous works~\cite{adcs,af}
on the classical model, our proof of Theorem~\ref{main_thm} uses
\emph{graphical representations}. For the classical Ising model, and
related models such as the Potts model, the use of graphical
representations is a standard tool and has been a huge success
since the seminal work of
Fortuin and Kasteleyn~\cite{fk}.
In more recent times graphical representations have also been very
successful in the study of quantum models, not only the Ising
model~\cite{akn,bjo_irb,bjogr,campanino,driessler}
but also Heisenberg
models~\cite{aizenman_nacht,toth,ueltschi}
(see also~\cite{golds}).
The transverse-field
Ising model possesses at least three graphical representations,
which may be called firstly the space--time spin representation,
secondly the random-parity representation, and
thirdly the {{\sc fk}}-representation. The first
of these goes back to~\cite{driessler}, the
last to to~\cite{akn,campanino},
whereas the second was developed in~\cite{bjogr}
(see also~\cite{crawford_ioffe}
for the related random-current representation). These graphical
representations are obtained by applying the Lie--Trotter expansion to
the correlations functions~\eqref{corr_eq} using the eigenbasis for either
$\sigma^{(3)}$ or $\sigma^{(1)}$, see~\cite{ioffe_geom}.
Of primary importance for the
present work is the random-parity representation, which is described
in Section~\ref{rpr_sec}. It is a continuous version of the
random-current representation for the classical Ising model, developed
in~\cite{aiz82}. The key insight of the works~\cite{aiz82,adcs}
was that the phase transition in the (classical) Ising model relates to
a percolation transition for the random currents,
and the present work exploits a similar picture for the
quantum model.
In addition to graphical representations, the other key component of
the proof of Theorem~\ref{main_thm} is an \emph{infrared bound}
proved in~\cite{bjo_irb}, and stated below
in~\eqref{irb}.
This is a bound on the Fourier-transform of the
Schwinger function:
\begin{equation}\label{schwinger_eq}
c((x,s),(y,t))=
\frac{1}{\mathrm{tr}(e^{-\b H_\L})}
\mathrm{tr}(e^{-(\b-t+s) H_\L}\sigma_y^{(3)}e^{-(t-s) H_\L}\sigma_x^{(3)}).
\end{equation}
Infrared bounds go back to~\cite{fss}
and one of their great successes is the proof by Dyson, Lieb and
Simon~\cite{dls} of the existence of a phase transition
in the antiferromagnetic Heisenberg model.
See also~\cite{ueltschi} for a recent infrared
bound for the Heisenberg model in the same spirit as the bound
employed here.
The argument for proving Theorem~\ref{main_thm}
follows the general outline of the argument for the
classical model given in~\cite{adcs}.
The first step is to develop an infinite-volume
version of the random-parity
representation and study percolation under this
measure, see Section~\ref{infvol_sec}. The infrared bound is
used to show that when $\l=\l_\crit$ then
(for $\b<\infty$ and $d\geq3$ or $\b=\infty$ and $d\geq2$)
there is no unbounded percolation cluster,
see Proposition~\ref{crit-perc_prop}.
Combined with `local modifcations' of the random-parity
representation (Proposition~\ref{local_prop}) and the
switching lemma (Lemma~\ref{sw_lem}), this allows us
to deduce the result, as described in Section~\ref{proof_sec}.
Compared with the classical model~\cite{adcs},
the main difficulty in the present work arises from the
`continuous' nature of the graphical representations in the quantum
setting. For example, the configuration space of
the random-parity representation is non-compact so an argument
is needed to obtain tightness of the sequence of finite-volume
random-parity measures (see Proposition~\ref{weak_prop}).
Related difficulties arise in proving insertion tolerance
(see Proposition~\ref{local_prop}) and ergodicity
(see Lemma~\ref{mix_lem}).
In the rest of this article we
fix $\d>0$. We also set $\g=0$ and use the equality
of the residual and spontaneous
magnetization~\eqref{spont_resid}.
We use the abbreviation {{\sc tfim}} for transverse-field Ising
model, and we use the following probabilistic notation:
$\hbox{\rm 1\kern-.27em I}_A$ or $\hbox{\rm 1\kern-.27em I}\{A\}$ for the indicator function
taking value 1 if the event $A$ occurs, 0 otherwise, and
$\PP(X)$ for the expectation of the random variable $X$
under the probability measure $\PP$.
\section{Graphical representations}
\label{graphical_sec}
In this section we present two graphical representations of the
{{\sc tfim}}, namely the space--time spin
representation and the random-parity representation. The latter
represents the correlation functions~\eqref{corr_eq}
and Schwinger functions~\eqref{schwinger_eq}
in terms of expectations over random
`paths' and is of central importance to this work. A major technical
tool in this representation is the \emph{switching lemma}, which we
describe in Section~\ref{switch_sec}. In Section~\ref{local_sec}
we then prove some results on `local
modifications' in this representation, which are forms of insertion-
and deletion tolerance.
The space--time spin representation
has a less prominent role in the
main argument than the random-parity representation,
and is used mainly to
establish an ergodicity property of the infinite-volume random-parity
representation
(see Lemma~\ref{mix_lem} and Proposition~\ref{ergodic_prop}).
However, it provides a natural setting to
introduce a class of boundary conditions that are of central
importance to this work, and may
also provide a more intuitive description
of the {{\sc tfim}} to readers with a
probabilistic background than the definition given in
Section~\ref{intro_sec}.
Before proceeding we introduce some notation.
Recall that $\L_N$ denotes the box $[-N,N]^d\subseteq\mathbb{Z}^d$ and
let $\partial\L_N=\L_N\setminus\L_{N-1}$ denote the boundary
of $\L_N$.
For $r>0$ write $I_r$ for the interval $[-r/2,r/2]\subseteq\RR$ and
define $K(N,r)=\L_N\times I_r$. Also write
$I_\infty=\RR$, $\mathbb{K}_\b=\mathbb{Z}^d\times I_\b$
and $\mathbb{K}=\mathbb{K}_\infty=\mathbb{Z}^d\times I_\infty$.
We frequently think of $\mathbb{K}$ as a subset of $\RR^{d+1}$
and $K(N,r)$ as a subset of $\mathbb{K}$ in the natural way.
For elements $x\in\mathbb{Z}^d$ or $(x,t)\in\mathbb{K}$ we write $\|x\|$ and
$\|(x,t)\|=\|x\|+|t|$ for their $\ell^1$-norm. We
will also use the notation $\mathcal{E}=\{xy:x,y\in\mathbb{Z}^d,\|x-y\|=1\}$
for the set of unordered pairs of nearest neighbours in $\mathbb{Z}^d$,
$\mathcal{E}_N=\{xy:x,y\in\L_N,\|x-y\|=1\}$
for nearest neighbours in $\L_N$,
and $F(N,r)=\mathcal{E}_N\times I_r$ as well as
$\mathbb{F}=\mathcal{E}\times I_\infty$.
\subsection{Space--time spin representation}
\label{spin_sec}
Let $\S_\b$ be the set of functions
$\sigma(\cdot,\cdot):\mathbb{K}_\b\to\{-1,+1\}$ such that for all
$x\in\mathbb{Z}^d$, the restriction $\sigma(x,\cdot):I_\b\to\{-1,+1\}$
is right-continuous and changes value finitely often
in each bounded interval. Also let $\S=\S_\infty$
and let $\S(N,r)$ be
the set of restrictions of elements of $\S$ to $K(N,r)$.
The space--time spin representation is
based on probability measures on $\S(N,r)$.
To define these, let $E$ be a
probability measure governing
\begin{itemize}
\item[(a)] a collection $D=(D_x:x\in\mathbb{Z}^d)$ of independent
Poisson processes on $\RR$ of rate $\d$, and
\item[(b)] a collection $\xi=(\xi_x:x\in\mathbb{Z}^d)$
of independent random variables taking the values
0 or 1 with equal probability.
\end{itemize}
We let
\begin{equation*}
\sigma(x,t)=\left\{
\begin{array}{ll}
(-1)^{\xi_x+|D_x\cap (0,t]|} & \mbox{if } t\geq0,\\
(-1)^{\xi_x+|D_x\cap (t,0]|} & \mbox{if } t<0.
\end{array}\right.
\end{equation*}
Thus $\sigma(x,0)=(-1)^{\xi_x}$ and $\sigma(x,t)$ switches
value at the points $t\in D_x$. See Figure~\ref{graphical_fig}
for an example.
We sometimes write $\sigma_{x,t}$ for $\sigma(x,t)$.
In this way $E$
defines an `a-priori' measure on $\S$. We write $E_{N,r}$ for the
induced measure on $\S(N,r)$.
The general notation for space--time Ising
probability measures on $\S(N,r)$ will be of
the form
\begin{equation*}
\mu_{N,r}^{\mathfrak{s},\mathfrak{t}}(\cdot),\qquad\mbox{where }
\mathfrak{s},\mathfrak{t}\in\{\mathfrak{f},\mathfrak{w},\mathfrak{p}\}.
\end{equation*}
The superscripts $\mathfrak{s},\mathfrak{t}$
denote boundary conditions, `spatial' and
`temporal', respectively, and their
values $\mathfrak{f},\mathfrak{w},\mathfrak{p}$ stand for `free', `wired' and
`periodic'. The easiest to describe is
$\mu_{N,r}^{\mathfrak{f},\mathfrak{f}}$, which is given by its density
\begin{equation}\label{strn_eq}
\frac{d\mu_{N,r}^{\mathfrak{f},\mathfrak{f}}}{dE_{N,r}}(\sigma)=\frac{1}{Z^{\mathfrak{f},\mathfrak{f}}_{N,r}}
\exp\Big(\l\sum_{xy\in\mathcal{E}_N}\int_{I_r}\sigma(x,t)\sigma(y,t)dt\Big).
\end{equation}
Here
\begin{equation}\label{stpf_eq}
Z^{\mathfrak{f},\mathfrak{f}}_{N,r}=E_{N,r}\Big[\exp\Big(
\l\sum_{xy\in\mathcal{E}_N}\int_{I_r}\sigma(x,t)\sigma(y,t)dt\Big)\Big]
\end{equation}
is the appropriate normalization. To obtain the
boundary conditions
$\mathfrak{t}=\mathfrak{w}$ and $\mathfrak{t}=\mathfrak{p}$,
respectively, we include in the
density~\eqref{strn_eq} (and in the partition
function~\eqref{stpf_eq}) the following restrictions:
\begin{itemize}
\item for $\mathfrak{t}=\mathfrak{w}$, that $\sigma(x,-r/2)=\sigma(x,r/2)=+1$ for all $x\in\L_N$,
\item for $\mathfrak{t}=\mathfrak{p}$, that $\sigma(x,-r/2)=\sigma(x,r/2)$ for all $x\in\L_N$.
\end{itemize}
(Thus, intuitively, for $\mathfrak{t}=\mathfrak{w}$ the spins at
the endpoints of $I_r$ are `frozen' to be $+1$, whereas
for $\mathfrak{t}=\mathfrak{p}$ we think of $I_r$ as a circle
of length $r$.)
To obtain the
boundary condition $\mathfrak{s}=\mathfrak{w}$
we replace $\mathcal{E}_N$
by $\mathcal{E}_{N+1}$ in the sums
in~\eqref{strn_eq}--\eqref{stpf_eq} and set
$\sigma(x,t)=+1$ for any $x\in\partial\L_{N+1}$.
Finally, to obtain the
boundary condition $\mathfrak{s}=\mathfrak{p}$
we replace $\mathcal{E}_N$ by the set $\mathcal{E}_N^\mathfrak{p}$ obtained by adding
to $\mathcal{E}_N$ all pairs $xy$ such that $x$ and $y$ differ in exactly one
coordinate, this coordinate being $-N$ in one case and $N$ in the
other. (Intuitively this makes $\L_N$ `wrap around' in each
coordinate direction.)
\begin{figure}[tbp]
\includegraphics{critmag.3}
\hspace{1.3cm}
\includegraphics{critmag.4}
\caption
{\emph{Left:} A realization of the spin-representation $\sigma$.
The elements of $D$ are marked as $\times$, and the value of $\sigma$ is
indicated next to each line segment. (This realization is compatible
with $\mathfrak{t}=\mathfrak{f}$ or $\mathfrak{p}$ but not with $\mathfrak{t}=\mathfrak{w}$.)
\emph{Right:} A realization of the random-parity
representation $\psi^\mathfrak{p}_A$, where $A=\{(x,s),(y,t)\}$. `Odd'
intervals are drawn bold, bridges as horizontal lines. (This
realization is not compatible with $\mathfrak{t}=\mathfrak{f}$ or $\mathfrak{w}$.)}
\label{graphical_fig}
\end{figure}
The connection to the {{\sc tfim}} is that the correlation
functions~\eqref{corr_eq} satisfy
\[
\langle \sigma_A^{(3)}\rangle_{\L,\b}=
\mu_{n,\b}^{\mathfrak{f},\mathfrak{p}}\Big(\prod_{x\in A}\sigma(x,0)\Big),
\]
and the Schwinger function~\eqref{schwinger_eq}
\[
c((x,s),(y,t))=\mu_{n,\b}^{\mathfrak{f},\mathfrak{p}}\big(\sigma(x,s)\sigma(y,t)\big).
\]
(Recall that we write $\mu(f)$ for the
$\mu$-expectation of a function $f(\sigma)$.)
In light of this correspondence it is natural to use the notation
\begin{equation}\label{corr_eq2}
\langle \sigma_A\rangle_{N,r}^{\mathfrak{s},\mathfrak{t}}=
\mu_{N,r}^{\mathfrak{s},\mathfrak{t}}
\Big(\prod_{(x,t)\in A}\sigma(x,t)\Big)
\end{equation}
for general finite subsets $A\subseteq K(N,r)$. Henceforth we refer to the
quantities in~\eqref{corr_eq2} as \emph{correlation functions}.
The $\mathfrak{p}$-boundary condition
in `time' arises automatically from the cyclicity of the trace
in~\eqref{Q-state}. The other boundary conditions
$\mathfrak{t}=\mathfrak{f},\mathfrak{w}$ are
convenient when working with infinite-volume limits.
The correlation functions~\eqref{corr_eq2} have
a natural monotonicity in the boundary conditions.
In particular,
\begin{equation}\label{corr_comp}
\langle \sigma_A\rangle_{N,r}^{\mathfrak{f},\mathfrak{f}}\leq
\langle \sigma_A\rangle_{N,r}^{\mathfrak{f},\mathfrak{p}}\leq
\langle \sigma_A\rangle_{N,r}^{\mathfrak{p},\mathfrak{p}}\leq
\langle \sigma_A\rangle_{N,r}^{\mathfrak{w},\mathfrak{p}}\leq
\langle \sigma_A\rangle_{N,r}^{\mathfrak{w},\mathfrak{w}},\mbox{ etc.}
\end{equation}
(A detailed proof for the present model
may easily be devised using Theorem~2.2.12 and
Lemma~2.2.21 of~\cite{bjo_phd}.)
When $\b<\infty$ we fix $\mathfrak{t}=\mathfrak{p}$
and work with the measures
$\mu_{N,\b}^{\mathfrak{f},\mathfrak{p}}$ and $\mu_{N,\b}^{\mathfrak{w},\mathfrak{p}}$. When $\b=\infty$
we will primarily be working
with $\mu_{N,r}^{\mathfrak{f},\mathfrak{f}}$ and $\mu_{N,r}^{\mathfrak{w},\mathfrak{w}}$
where $r=2N$. The (weak) limits of these measures
exist as $N\to\infty$ and are related,
respectively, to the positive-temperature and ground-state limits
appearing in~\eqref{resid_eq}. In particular,
\begin{equation}\label{spont_resid}
M^+_\infty=\lim_{N\to\infty} \langle \sigma(0,0)\rangle_{N,r=2N}^{\mathfrak{w},\mathfrak{w}},\quad
M^+_\b=\lim_{N\to\infty} \langle \sigma(0,0)\rangle_{N,\b}^{\mathfrak{w},\mathfrak{p}}
\mbox{ for } \b<\infty,
\end{equation}
i.e.\ the residual and spontaneous magnetization
coincide, cf.~\cite{lml} and Section~2.5.2 of~\cite{bjo_phd}.
We write $\mu_{\l,\b}^{(\mathfrak{s},\mathfrak{t})}(\cdot)$
or $\el\cdot\rangle_{\l,\b}^{(\mathfrak{s},\mathfrak{t})}$ for any weak
limit obtained as above with boundary conditions
$\mathfrak{s},\mathfrak{t}$ (and $\b<\infty$ or $\b=\infty$).
When $M^+=0$ then the limit measure is unique:
\[
\mu_{\l,\b}^{(\mathfrak{f},\mathfrak{p})}=\mu_{\l,\b}^{(\mathfrak{w},\mathfrak{p})}\mbox{ for }\b<\infty,\qquad
\mu_{\l,\infty}^{(\mathfrak{f},\mathfrak{f})}=\mu_{\l,\infty}^{(\mathfrak{w},\mathfrak{w})}.
\]
(A detailed proof for the present model appears
in~\cite[Theorem~2.5.9]{bjo_phd}; it follows closely the
argument for the classical model~\cite{lml}.)
Thus our main result Theorem~\ref{main_thm} implies that the limit
measure is unique at the critical point.
For more information about the statements
in this subsection, and the spin-representation in
general, see~\cite{bjo_phd}.
\subsection{Random-parity representation}
\label{rpr_sec}
Like the space--time spin representation of the
previous subsection, the random-parity representation
expresses the correlation functions~\eqref{corr_eq2}
using Poisson processes in $K(N,r)$.
This time we write $E_{N,r}$ for a probability measure
governing a collection $B=(B_{xy}:xy\in\mathcal{E}_N)$
of independent Poisson processes on $I_r$ with
intensity $\l$, as well as an independent collection
$\tau=(\tau_x:x\in\L_N)\in\{0,1\}^{\L_N}$
of independent random variables taking values 0 or 1
with equal probability $1/2$. We sometimes refer to the points
of $B$ as \emph{bridges}.
When considering the correlation
function $\langle \sigma_A\rangle_{N,r}^{\mathfrak{s},\mathfrak{t}}$
we will use the term \emph{switching point}
for any point $(x,t)\in K(N,r)$
such that either (i) $(x,t)\in A$, or (ii)
there exists $y\in\L_N$ such that $t\in B_{xy}$
(ie, $(x,t)$ is the `endpoint' of some bridge).
The collection of switching points is denoted by
$S=(S_x:x\in\L_N)$, where $S_x$ is the set of $t\in I_r$
such that $(x,t)$ is a switching point.
The points of $A$ are referred to as \emph{sources}.
We say that $B$
is \emph{consistent} with $A$ if $|S_x|$ is
even for each $x$, and in this case we will also refer
to $S$ itself as consistent.
For each consistent $S$ we will define a labelling of
$K(N,r)$ using the labels `even' and `odd', and
for definiteness we use the convention that the
`odd' subset is closed. See Figure~\ref{graphical_fig} again for an
illustration of the description that follows.
The definition will depend on the boundary
conditions $\mathfrak{s}$
and $\mathfrak{t}$. We will not be using the
random-parity representation for
$\mathfrak{s}=\mathfrak{p}$ so we omit describing it.
For $\mathfrak{s}=\mathfrak{f}$ we will denote the
labelling $\psi^{\mathfrak{t}}_A$ and for
$\mathfrak{s}=\mathfrak{w}$ by
$\hat\psi^{\mathfrak{t}}_A$.
In what follows we assume for simplicity
that $A$ does not contain any point
of the form $(x,\pm r/2)$. We begin with the case
$\mathfrak{s}=\mathfrak{f}$:
\begin{list}{$\bullet$}{\leftmargin=1em}
\item If $\mathfrak{t}=\mathfrak{f}$
we label each point
$(x,\pm r/2)$ `even', and define the rest of
$\psi^\mathfrak{f}_A$
by requiring that labels switch between `even'
and `odd' at the points of $S$ and are constant
in between. This is possible
due to the assumption that $S$ is consistent.
\item If $\mathfrak{t}=\mathfrak{w}$ we instead
label all points of the form $(x,\pm r/2)$ `odd' and
apply the same rule for switching at points of $S$.
\item If $\mathfrak{t}=\mathfrak{p}$ we require the vector $\tau$.
We define $\psi_A^\mathfrak{p}$
by labelling each $(x,0)$ `even' if $\tau_x=0$
or `odd' if $\tau_x=1$, and letting the label switch
at the points of $S$ as before. Due to the consistency
of $S$ we can think of $\psi^\mathfrak{p}_A$ as a labelling
of the circle. (Clearly the choice $(x,0)$
is arbitrary, and one may equally well let
$\tau$ determine all the labels $(x,t)$ for any
fixed $t$.)
\end{list}
We now describe how to define the labelling in the
case when $\mathfrak{s}=\mathfrak{w}$. We now let
$\hat E_{N,r}$ denote a probability measure
which, in addition to processes $B$ and $\tau$
as above, also governs a process
$G=(G_x:x\in\partial\L_N)$ of independent Poisson
processes on $I_r$. The intensity of $G_x$ depends
on $x$, and equals $\l$ times the number of
$y\in\mathbb{Z}^d\setminus\L_N$ such that $xy\in\mathcal{E}$.
For simplicity we set $G_x=\varnothing$ for
$x\in\L_N\setminus\partial\L_N$.
We augment the switching points $S$ to
contain also each $(x,t)$ such that $t\in G_x$.
As before we say that $S$ is consistent if each
$|S_x|$ is even, and in that case we also say that the
pair $(B,G)$ is consistent with $A$.
Given a consistent $S$ we obtain the labelling, which we
now denote
$\hat\psi^\mathfrak{t}_A$, precisely as before.
Given a labelling $\psi^\mathfrak{t}_A$
(respectively, $\hat\psi^\mathfrak{t}_A$),
let $\epsilon$ denote
the total length of all intervals in $K(N,r)$
labelled `even', and let the
\emph{weight} $\partial\psi^\mathfrak{t}_A$
(respectively, $\partial\hat\psi^\mathfrak{t}_A$)
be defined as $e^{2\d\epsilon}$. In the case when
$S$ is not consistent we define the weight to be zero.
The random-parity
representation allows us to write
\begin{equation}\label{rpr_eq}
\langle \sigma_A\rangle^{\mathfrak{f},\mathfrak{t}}_{N,r}=
\frac{E_{N,r}(\partial\psi^\mathfrak{t}_A)}
{E_{N,r}(\partial\psi^\mathfrak{t}_\varnothing)}
\quad \mbox{and} \quad
\langle \sigma_A\rangle^{\mathfrak{w},\mathfrak{t}}_{N,r}=
\frac{\hat E_{N,r}(\partial\hat\psi^\mathfrak{t}_A)}
{\hat E_{N,r}(\partial\hat\psi^\mathfrak{t}_\varnothing)}.
\end{equation}
The first of these ($\mathfrak{s}=\mathfrak{f}$) was proved
in~\cite[Theorem~3.1]{bjogr}, and the second ($\mathfrak{s}=\mathfrak{w}$)
follows using similar arguments.
In fact, the representation~\eqref{rpr_eq}
holds for correlations in more general subsets
$K'$ of $\mathbb{K}$ than $K(N,r)$, in particular
for `regions with holes'. We briefly outline this
now (it will be used in Proposition~\ref{weak_prop}).
Write $K=K(N,r)$, let $J$ be a finite collection
of disjoint closed intervals in $K$, and let $K'=K\setminus J$.
Write $\partial J$ for the set of all endpoints
of intervals in $J$ (if $J$ contains some interval
consisting of a single point $(x,t)$ we
distinguish between $(x,t+)$ and $(x,t-)$).
For each $xy\in\mathcal{E}_N$ let $B_{xy}$ denote a Poisson
process on $I_r$ with variable intensity:
$\l$ if $(x,t),(y,t)\in K'$, otherwise 0. We
start by labelling
each point $(x,t)\in\partial J$ `even'. If $\mathfrak{t}=\mathfrak{f}$
(respectively, $\mathfrak{t}=\mathfrak{w}$) we label each point
$(x,\pm r/2)\in K'$ `even' (respectively, `odd').
If $\mathfrak{t}=\mathfrak{p}$ then for each $x$ such that
$\{x\}\times I_r\subseteq K'$ we let $\tau_x$ determine
the label $(x,0)$ as before. The full labelling is finally
obtained by letting
the labels switch at the points of $S$ as before.
(A labelling $\hat\psi_A^\mathfrak{t}$ of $K'$ is
obtained similarly, modifying also $G$ to have
zero intensity in $J$.) Writing $E_{K'}$ for the
corresponding probability measure, we note that
\begin{equation}\label{holes_eq}
E_{K'}(\partial\psi_\varnothing^\mathfrak{t})=
e^{-2\d |J|}
E_{N,r}(\partial\psi_\varnothing^\mathfrak{t}
\hbox{\rm 1\kern-.27em I}\{\psi_\varnothing^\mathfrak{t}\mbox{ `even' in }J\}),
\end{equation}
where $|J|$ denotes the total
length of the intervals comprising $J$,
see~\cite{bjogr}.
\subsubsection{Switching lemma}
\label{switch_sec}
Consider the case in~\eqref{rpr_eq}
when $A$ consists of the two points $(0,0)$
and $(x,t)$. The consistency constraint on $S$, and
the way the labels are defined, forces the existence
of an `odd path' between the two sources $(0,0)$
and $(x,t)$. A similar, but more complicated,
picture arises for general sets $A$. The main virtue
of the random-parity representation is that this
picture can be developed to the case of
\emph{pairs} of labellings in a way which also allows the
representation of \emph{differences} between correlation
functions. The tool for this is called the
switching lemma.
Let $\EE_{N,r}$ denote a probability measure governing
the following independent random variables:
\begin{itemize}
\item[(a)] two copies of the process
$B$, denoted $B$ and $\hat B$;
\item[(b)] two copies of $\tau$,
denoted $\tau$ and $\hat\tau$;
\item[(c)] one copy of the process $G$; and
\item[(d)] one copy of a process $\D=(\D_x:x\in\L_N)$,
where the $\D_x$ are independent Poisson processes
on $I_r$ with intensity $4\d$.
\end{itemize}
Thus we may write $\EE_{N,r}=E_{N,r}\times\hat E_{N,r}\times P_\D$,
where $P_\D$ denotes the distribution of $\D$.
We call $\D$ the process of `cuts'.
If we fix two finite sets $A_1$ and $A_2$ of
sources and two boundary conditions $\mathfrak{t}_1$
and $\mathfrak{t}_2$
then we obtain under $\EE_{N,r}$ a triple
$(\psi_{A_1}^{\mathfrak{t}_1},\hat\psi_{A_2}^{\mathfrak{t}_2},\D)$ whose
components are independent.
In what follows we will only be using the following
combinations of boundary conditions $\mathfrak{t}_1$, $\mathfrak{t}_2$:
in the case $\b<\infty$ we take $\mathfrak{t}_1=\mathfrak{t}_2=\mathfrak{p}$,
and in the case $\b=\infty$ we take $\mathfrak{t}_1=\mathfrak{f}$
and $\mathfrak{t}_2=\mathfrak{w}$, cf.\ the discussion below~\eqref{corr_comp}.
In the latter case we use the
notation $\hat G_x=G_x\cup \{-r/2,r/2\}$.
We next define a notion of
\emph{open paths}. The precise definition of
a path depends on the boundary
conditions $\mathfrak{t}_1,\mathfrak{t}_2$, and we start
with the case $\mathfrak{t}_1=\mathfrak{f},\mathfrak{t}_2=\mathfrak{w}$.
If $\k,\k'\in K(N,r)$
then an (open) path from $\k$ to $\k'$ is a sequence
of points $\k_0,\k_1,\dotsc,\k_{2m+1}\in K(N,r)$
satisfying the following:
\begin{enumerate}
\item $\k_0=\k$ and $\k_{2m+1}=\k'$;
\item\label{int_item}
for each $j\in\{0,\dotsc,m\}$, if $\k_{2j}=(x,s)$
then $\k_{2j+1}=(x,t)$ for some $t$ such that
the interval $[s\wedge t,s \vee t]$ contains no
point $s'\in\D_x$ which is labelled `even'
in both labellings $\psi_{A_1}^\mathfrak{f}$ and
$\hat\psi_{A_2}^\mathfrak{w}$;
\item\label{jump_item}
for each $j\in\{0,\dotsc,m-1\}$,
if $\k_{2j+1}=(x,s)$ and $\k_{2j+2}=(y,t)$
then either (i) $s=t\in B_{xy}\cup \hat B_{xy}$,
or (ii) $s\in\hat G_x$ and $t\in\hat G_y$.
\end{enumerate}
Intuitively this means that paths can
traverse bridges, `jump between' arbitrary
points of $\hat G$, and traverse subintervals of $K(N,r)$,
but are blocked by points of $\D$ which
fall where both labellings are `even'.
One way to think of the `jumping'
between points of $\hat G$ is that
the points in $\hat G$ are connections to and
from a `ghost-site' $\Gamma$.
To obtain the case $\mathfrak{t}_1=\mathfrak{t}_2=\mathfrak{p}$
we modify (ii) in item~\eqref{jump_item} above
by replacing $\hat G$ with $G$ in both places,
and we additionally allow in~\eqref{int_item}
that one may traverse either the inverval
$[s,t]$ or the interval $[t,s]$, where these
are to be regarded as intervals in the circle
(we keep the restriction on $\D$).
Intuitively these changes mean that
the `endpoints' $(x,\pm r/2)$ are no longer
connected to $\Gamma$ but are identified
with each other.
The event that there is an open path between $\k$
and $\k'$ is written $\{\k\leftrightarrow\k'\}$. The
possibility (ii) in~\eqref{jump_item} of
`jumping' via $\Gamma$ is special, and in some
cases we want to consider paths which do not
do this. We write
$\{\k\leftrightarrow\k'\mbox{ off }\Gamma\}$ for the event that
there is some path that does not
feature a pair $\k_{2j+1}=(x,s)$, $\k_{2j+2}=(y,t)$
with $x\neq y$ and $s,t\in\hat G$.
Note that when $\mathfrak{t}_1=\mathfrak{f},\mathfrak{t}_2=\mathfrak{w}$ this also excludes
jumping via the endpoints $(x,\pm r/2)$.
We also write $\{\k\leftrightarrow\Gamma\}$ for the event that
some open path connects $\k$ to a point in $\hat G$
(respectively, $G$).
Write 0 for the origin $(0,0)$ and
$\k$ for an arbitrary point in $K(N,r)$.
We will be using the following form of the
switching lemma:
\begin{lemma}\label{sw_lem}
$\EE_{N,r}(\partial\psi^{\mathfrak{t}_1}_{0\k}
\partial\hat\psi_\varnothing^{\mathfrak{t}_2})=
\EE_{N,r}(\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi_{0\k}^{\mathfrak{t}_2}
\hbox{\rm 1\kern-.27em I}\{0\leftrightarrow \k\mbox{ off }\Gamma\})$.
\end{lemma}
Here we have abbreviated $\{0,\k\}$
with $0\k$.
The proof of Lemma~\ref{sw_lem}
is a small modification
of~\cite[Theorem~4.2]{bjogr},
which we briefly outline now.
\begin{proof}[Proof sketch]
Note that
in the left-hand-side, the event
$\{0\leftrightarrow x\mbox{ off }\Gamma\}$ holds since there is an
odd path in $\psi_{0\k}^{\mathfrak{t}_1}$. Condition on the sets
$\overline B=B\cup\hat B$ and $G$, and note that the
conditional distribution of the pair $(B,\hat B)$
is given by assigning each element of $\overline B$
to $B$ or $\hat B$ with equal probability, independently.
Given $\overline B$ and $G$ there is a finite collection of
`possible' paths $\pi_1,\dotsc,\pi_n$ between $0$ and $\k$
off $\Gamma$ (the numbering is arbitrary but fixed,
and as noted above the collection is nonempty).
When we assign the
elements of $\overline B$ to $B$ and $\hat B$ and also sample
$\D$ and (in the case $\mathfrak{t}_1=\mathfrak{t}_2=\mathfrak{p}$) $\tau,\hat\tau$,
this will both determine the labellings
$\psi^{\mathfrak{t}_1}_{0\k}$ and $\hat\psi_\varnothing^{\mathfrak{t}_2}$, as
well as reveal which of the possible paths are indeed open.
Let $\pi_j$ be the first of these. Write
$\psi^{\mathfrak{t}_1}_{0\k}\triangle \pi_j$ and
$\hat\psi_\varnothing^{\mathfrak{t}_2}\triangle\pi_j$ for the labellings obtained
by switching `even' and `odd' along $\pi_j$.
It is easy to see that these new labellings
are consistent with sources $\varnothing$ and $0\k$,
respectively, and that they can be obtained
by switching certain elements between $B$ and $\hat B$ as well
as certain values of $\tau$ and $\hat\tau$. By symmetry,
the latter transformations are measure-preserving. Moreover,
it may be checked as in~\cite[p.~251]{bjogr} that the
change in \emph{weight} of the labellings exactly corresponds
to the change in \emph{probability} that $\pi_j$
is indeed the first open path. In particular,
in the new configuration there is still a path
between 0 and $\k$ off $\Gamma$.
\end{proof}
The following simple application of
Lemma~\ref{sw_lem} will be used in the proof of
Theorem~\ref{main_thm}. We have that
\begin{equation}\label{sw2}
\begin{split}
\EE_{N,r}(\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi_{0\k}^{\mathfrak{t}_2})
-\EE_{N,r}(\partial\psi_{0\k}^{\mathfrak{t}_1}\partial\hat\psi_\varnothing^{\mathfrak{t}_2})
&=\EE_{N,r}(\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi_{0\k}^{\mathfrak{t}_2}
[1-\hbox{\rm 1\kern-.27em I}\{0\leftrightarrow \k\mbox{ off }\Gamma\}])\\
&\leq
\EE_{N,r}(\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi_{0\k}^{\mathfrak{t}_2}
\hbox{\rm 1\kern-.27em I}\{0\leftrightarrow \Gamma\}).
\end{split}
\end{equation}
In the last step we used that $\hat\psi_{0\k}^{\mathfrak{t}_2}$
contains an odd path $\pi$ between 0 and $\k$, and if there
is no path between 0 and $\k$ off $\Gamma$ then $\pi$
must pass $\Gamma$.
Combined with~\eqref{rpr_eq} this has the following
consequence. Writing $0$ for $(0,0)$
and $\k$ for $(x,t)$ we have from~\eqref{sw2}
\begin{equation}\label{tp_rpr}
\begin{split}
0\leq
\langle \sigma_{0,0}\sigma_{x,t} \rangle^{\mathfrak{w},\mathfrak{t}_2}_{N,r}
-\langle \sigma_{0,0}\sigma_{x,t} \rangle^{\mathfrak{f},\mathfrak{t}_1}_{N,r}
&\leq \frac{
\EE_{N,r}(\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi_{0\k}^{\mathfrak{t}_2}
\hbox{\rm 1\kern-.27em I}\{0\leftrightarrow \Gamma\})}
{\EE_{N,r}(\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi_\varnothing^{\mathfrak{t}_2})}.
\end{split}
\end{equation}
\subsubsection{Local modifications}
\label{local_sec}
Of central importance to the proof of
Theorem~\ref{main_thm} is a probability measure
$\overline\PP_{N,r}$ defined as follows. Recall the
processes $B,\hat B, G, \D,\tau, \hat \tau$
as well as the labellings $\psi,\hat\psi$
of the previous subsection. If $A$ is an event
measurable with respect to these processes, we define
\begin{equation}\label{P_bar_def}
\overline\PP_{N,r}(A)=
\frac{\EE_{N,r}(\hbox{\rm 1\kern-.27em I}_A\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi_\varnothing^{\mathfrak{t}_2})}
{\EE_{N,r}(\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi_\varnothing^{\mathfrak{t}_2})}.
\end{equation}
In this section we prove two estimates relating
to $\overline\PP_{N,r}$. Before we state these, recall
our conventions: if $\b<\infty$ then we write $r=\b$
and let $\mathfrak{t}_1=\mathfrak{t}_2=\mathfrak{p}$, whereas if $\b=\infty$
then $r=2N\to\infty$ and $\mathfrak{t}_1=\mathfrak{f},\mathfrak{t}_2=\mathfrak{w}$. In the following
result we let $N_0<N$ and if $\b<\infty$ let $r_0=r=\b$,
or if $\b=\infty$ let $r_0=2N_0<2N=r$. If $J$
is a measurable subset of $K(N,r)$ we say that the
event $A$ is \emph{defined in} $J$ if it is
measurable with respect to the restrictions of the
processes $B,\hat B, G, \D,\psi,\hat\psi$ to $J$.
\begin{proposition}\label{local_prop}
\hspace{1cm}\\
(A) For each $(x,t)\in K(N,r)$ there is a constant
$C_{(x,t)}$ not depending on $N,r$ such that
\[
\langle \sigma_{0,0}\sigma_{x,t} \rangle^{\mathfrak{w},\mathfrak{t}_2}_{N,r}
-\langle \sigma_{0,0}\sigma_{x,t} \rangle^{\mathfrak{f},\mathfrak{t}_1}_{N,r}
\leq C_{(x,t)}\overline\PP_{N,r}((0,0)\leftrightarrow \Gamma).
\]\\
(B) There is a constant $c=c(N_0,r_0)$ such that the
following holds.
Let $A$ be an event defined in $K(N,r)\setminus K(N_0,r_0)$,
and let $\mathcal{C}$ be the event that there is an open
path inside $K(N_0,r_0)$ between every pair of
points in $K(N_0,r_0)$. Then
\[
\overline\PP_{N,r}(A)\leq c \,\overline\PP_{N,r}(A\cap\mathcal{C}).
\]
\end{proposition}
In proving Proposition~\ref{local_prop}
we will be using the following fact about
`local modifications' of point processes. Let
$X$ denote a point process
on the interval $[0,t]$. Let $\tilde X$ be another
point process on $[0,t]$ obtained from $X$
by a deterministic or random modification. For example,
$\tilde X$ may be obtained by adding a point somewhere in
$[0,t]$, or deleting one of the points of $X$.
Write $E,\tilde E,\EE$ for the law of $X$, the
law of $\tilde X$ and their joint law, respectively.
We will assume that $\tilde X$ is defined in such
a way that for some event $A$ we have
$\tilde X\in A$ with probability 1. Moreover, assume
that $f$ is some function and $c_1,c_2>0$ some constants
such that $\tfrac{f(X)}{f(\tilde X)}\leq c_1$ and
the Radon-Nikodym density $\tfrac{d\tilde E}{d E}\leq c_2$
almost surely. We then have that
\begin{equation}\label{modif_bound}
\begin{split}
E[f(X)]&= \EE[f(X)]=\EE[f(X)\hbox{\rm 1\kern-.27em I}\{\tilde X\in A\}]
\leq c_1 \EE[f(\tilde X)\hbox{\rm 1\kern-.27em I}\{\tilde X\in A\}]\\
&= c_1 \tilde E[f(\tilde X)\hbox{\rm 1\kern-.27em I}_A(\tilde X)]=
c_1 E\Big[\frac{d \tilde E}{d E}(X)f(X)\hbox{\rm 1\kern-.27em I}_A(X)\Big]\\
&\leq c_1c_2 E[f(X)\hbox{\rm 1\kern-.27em I}\{X\in A\}].
\end{split}
\end{equation}
We will be using~\eqref{modif_bound}
when $X$ is a Poisson process of intensity
$\a$, say, and $\tilde X$ is obtained in one of
the following three ways.
\begin{list}{$\bullet$}{\leftmargin=1em}
\item Firstly, if $\tilde X$ is
the trivial process obtained by deleting all
points of $X$ then
\begin{equation}\label{delete_RN}
\frac{d\tilde E}{d E}(X)=e^{\a t}\hbox{\rm 1\kern-.27em I}\{X=\varnothing\}
\leq e^{\a t}.
\end{equation}
\item Secondly, suppose $\tilde X$ is obtained from $X$
by adding two independent, uniformly distributed
points if $X=\varnothing$, but letting $\tilde X=X$
otherwise. Then
\begin{equation}\label{add_0or2_RN}
\begin{split}
\frac{d\tilde E}{d E}(X)&=
\hbox{\rm 1\kern-.27em I}\{X\neq\varnothing\}+\tfrac{2}{(\a t)^2}\hbox{\rm 1\kern-.27em I}\{|X|=2\}\\
&\leq 1+\tfrac{2}{(\a t)^2}.
\end{split}
\end{equation}
\item Thirdly, suppose $\tilde X$ is obtained from $X$
by adding uniformly a point if $|X|\in\{0,1\}$,
alternatively deleting a uniformly chosen point if $|X|\geq2$.
Then
\begin{equation}\label{add_del_RN}
\begin{split}
\frac{d\tilde E}{d E}(X)&=
\tfrac{1}{\a t}\hbox{\rm 1\kern-.27em I}\{|X|=1\}
+\tfrac{2}{\a t}\hbox{\rm 1\kern-.27em I}\{|X|=2\}+
\hbox{\rm 1\kern-.27em I}\{X\neq\varnothing\}\tfrac{\a t}{|X|+1}\\
&\leq \tfrac{2}{\a t}+\a t.
\end{split}
\end{equation}
\end{list}
One way to check~\eqref{delete_RN}--\eqref{add_del_RN}
is to approximate $X$ by a Bernoulli process
on $\{0,\tfrac{1}{n},\dotsc,\tfrac{\lfloor tn\rfloor}{n}\}$
with success probability $\a/n$ and look at the
limits of the corresponding likelihood ratios.
\begin{proof}[Proof of Proposition~\ref{local_prop}]
We begin with the easier part (B). The statement is
equivalent to
\begin{equation}\label{in1}
\EE_{N,r}(\hbox{\rm 1\kern-.27em I}_A\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi_\varnothing^{\mathfrak{t}_2})
\leq c\,
\EE_{N,r}(\hbox{\rm 1\kern-.27em I}_A\hbox{\rm 1\kern-.27em I}_\mathcal{C}\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi_\varnothing^{\mathfrak{t}_2}).
\end{equation}
We modify $\D$ by removing all points in $K(N_0,r_0)$
and we modify $B$ inside $K(N_0,r_0)$
as in~\eqref{add_0or2_RN}.
That is, whenever $B_{xy}\cap I_{r_0}=\varnothing$ we
add two bridges uniformly placed in $I_{r_0}$, otherwise
leave $B_{xy}$ unchanged. The resulting bridge-configuration
$\tilde B$
is then still consistent with source set $\varnothing$.
If $\b=\infty$ there is (due to our choice of boundary
conditions $\mathfrak{t}_1=\mathfrak{f}$, $\mathfrak{t}_2=\mathfrak{w}$) a unique labelling
$\tilde \psi_\varnothing^{\mathfrak{t}_1}$ consistent with $\tilde B$
which agrees with the original labelling $\psi_\varnothing^{\mathfrak{t}_1}$
in $K(N,r)\setminus K(N_0,r_0)$.
If $\b<\infty$ there is a unique
labelling $\tilde \psi_\varnothing^{\mathfrak{t}_1}$ which agrees with
$\psi_\varnothing^{\mathfrak{t}_1}$ in $K(N,r)\setminus K(N_0,r_0)$
and at each $(x,0)$ such that $x\in\L_{N_0}$.
Since we have removed all cuts and placed
bridges between all neighbouring pairs of intervals,
the event $\mathcal{C}$ holds after the modification.
Since all changes have been restricted to
$K(N_0,r_0)$ the change preserves the event $A$.
The total length $\tilde\epsilon$ of intervals labelled
`even' in $\tilde \psi_\varnothing^{\mathfrak{t}_1}$ satisfies
$\tilde\epsilon\geq\epsilon-r_0(2N_0+1)^d$.
Applying~\eqref{modif_bound} with
$f$ equal to the weight of the labelling,
as well as~\eqref{delete_RN} and~\eqref{add_0or2_RN}, we
obtain~\eqref{in1} with
\[
c=\exp(4\d r_0(2N_0+1)^d)
\exp(4\d r_0(2N_0+1)^d)
(1+2/(\l r_0)^2)^{2d(2N_0+1)^d}.
\]
(The first factor is due to the change in
the weight of the labelling,
the second to the change of measure of $\D$,
and the third to the change of measure of $B$.)
We now turn to part (A). By~\eqref{tp_rpr}
it suffices to show that
\begin{equation}\label{Cx_eq}
\EE_{N,r}(\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi_{0\k}^{\mathfrak{t}_2}
\hbox{\rm 1\kern-.27em I}\{0\leftrightarrow \Gamma\})\leq C_\k
\EE_{N,r}(\partial\psi_\varnothing^{\mathfrak{t}_1}\partial\hat\psi^{\mathfrak{t}_2}_\varnothing
\hbox{\rm 1\kern-.27em I}\{0\leftrightarrow \Gamma\}),
\end{equation}
where $\k=(x,t)$.
We begin with the (more delicate) case when
$\b=\infty$ and $r=2N$.
For each $y\in\L_N$ write
\begin{equation*}
I_{y,k}=\{y\}\times(k,k+1],\qquad -N\leq k\leq N-1.
\end{equation*}
Thus the $I_{y,k}$ form a partition of $K(N,r)$ into
intervals of length 1, and we have that
$(x,t)\in I_{x,\lceil t\rceil -1}$ and
$(0,0)\in I_{0,-1}$. We begin by defining a collection
$\Pi(x,t)$ of intervals of this form which `connect'
$(0,0)$ to $(x,t)$. There is some flexibility
in the choice of $\Pi(x,t)$, but for definiteness we
define it as follows. Firstly, let
$0=x_0,x_1,\dotsc,x_n=x$ be a fixed, shortest nearest-neighbour
path from $0$ to $x$ in $\L_N$; thus
$n=\|x\|$. Next, set $k_0=-1$ and
\begin{equation*}\begin{split}
k_1&=0, k_2=1,\dotsc,k_m=\lceil t \rceil-1\qquad
\mbox{ if } t>0,\\
k_1&=-2, k_2=-3,\dotsc,k_m=\lceil t \rceil-1\qquad
\mbox{ if } t<0.
\end{split}
\end{equation*}
We define
\begin{equation*}
\Pi(\kappa)=\Pi(x,t)=\{I_{0,k_0},I_{0,k_1},\dotsc,I_{0,k_m},
I_{x_1,k_m},I_{x_2,k_m},\dotsc, I_{x_n,k_m}\}.
\end{equation*}
Note that $(0,0)\in I_{0,k_0}$, that
$(x,t)\in I_{x_n,k_m}$, and that the number of intervals
in $\Pi(x,t)$ as well as their total length
are bounded by $|t|+2+\|x\|$.
We are going to modify $\D$, $\hat B$ and $\hat\psi$
along $\Pi(x,t)$
and apply the argument at~\eqref{modif_bound}.
We modify $\D$
by simply replacing $\D\cap I$ with $\varnothing$ for all
$I\in\Pi(x,t)$. By~\eqref{delete_RN} the
corresponding Radon--Nikodym density is at most
$\exp(4\d(|t|+2+\|x\|))$.
Next, define
\begin{equation*}
J_i=\{(x_ix_{i+1},s):(x_i,s)\in I_{x_i,k_m},\,
(x_{i+1},s)\in I_{x_{i+1},k_m}\}.
\end{equation*}
We modify $\hat B$ by applying the operation
in~\eqref{add_del_RN} in each
$J_i$ for $0\leq i\leq n-1$;
that is, if $J_i$ contains
0 or 1 bridge we add one uniformly, but if $J_i$ contains
2 or more bridges we delete one chosen uniformly.
Write $\tilde B$ for the modified process of bridges.
By~\eqref{add_del_RN}, the
corresponding Radon--Nikodym density is at most
$(2/\l+\l)^{|t|+2+\|x\|}$.
If $\hat B$ was
consistent with the sources $0,\k$
then $\tilde B$ is consistent with the source
set $\varnothing$. Due to the boundary
condition there is a unique labeling $\tilde\psi_\varnothing^{\mathfrak{w}}$
associated with $\tilde B$, and in fact
one obtains $\tilde\psi_\varnothing^\mathfrak{w}$ from
$\hat\psi_{0\k}^\mathfrak{w}$ by modifying the labels in intervals
belonging to $\Pi(x,t)$ only.
See Figure~\ref{local_fig}.
\begin{figure}[tbp]
\centering
\includegraphics{critmag.1}
\qquad\includegraphics{critmag.2}
\caption
{Part of the labellings $\hat\psi_{0\kappa}^\mathfrak{w}$ (left) and
$\tilde\psi_{\varnothing}^\mathfrak{w}$ (right) in the proof of
Proposition~\ref{local_prop}(A). The intervals comprising
$\Pi(\kappa)$ are indicated and highlighted in light
grey. Intervals labelled `odd' are drawn bold. Bridges
added (respectively, deleted) are drawn dashed
(respectively, marked with an $\times$).}
\label{local_fig}
\end{figure}
It follows that
\begin{equation*}
\frac{\partial\hat\psi_{0\k}^\mathfrak{w}}
{\partial\tilde\psi_\varnothing^\mathfrak{w}}\leq
\exp(2\d(|t|+2+\|x\|)).
\end{equation*}
Note that the modifications described above
do not destroy any connections
(but possibly creates some
new ones). In particular, if
$0\leftrightarrow\Gamma$ before
then also $0\leftrightarrow\Gamma$ after.
Applying the argument in~\eqref{modif_bound}
we therefore arrive at~\eqref{Cx_eq}, with
\[
C_{(x,t)}=\exp(6\d(|t|+2+\|x\|))
(2/\l+\l)^{|t|+2+\|x\|}.
\]
We now turn to the case $\b<\infty$,
and recall that we then have $r=\b$. We no longer
need to partition $I_\b$ into intervals of length 1,
but instead define $I_x=\{x\}\times I_\b$.
We now let
\begin{equation*}
\Pi(x,t)=\{I_{x_0},I_{x_1},I_{x_2},\dotsc, I_{x_n}\},
\end{equation*}
where as before
$0=x_0,x_1,\dotsc,x_n=x$ is a fixed, shortest, nearest-neighbour
path from $0$ to $x$ in $\L_N$. The number of
intervals in $\Pi(x,t)$ is now $n=\|x\|$, and their
combined length is $\b\|x\|$.
With this definition of $\Pi(x,t)$
we apply the same modifications to $\D$ and $\hat B$
as in the case $\b=\infty$. Now we
let $\tilde\psi_\varnothing^\mathfrak{p}$ be the unique labelling which
agrees with $\hat\psi_{0\k}^\mathfrak{p}$ outside the intervals
of $\Pi(x,t)$ and at all points $(x_i,0)$ for
$i=0,1,\dotsc,n$.
This time we get~\eqref{Cx_eq} with
\[
C_{(x,t)}=\exp(6\d\b\|x\|)(2/\l\b+\l\b)^{\|x\|}.
\]
\end{proof}
\section{Infinite-volume RPR}
\label{infvol_sec}
In this section we study the limit $\overline\PP$
of the measures $\overline\PP_{N,r}$ as $N\to\infty$
(and either $r=\b<\infty$ is fixed, or $r=2N\to\infty$).
In Section~\ref{inf_rpr_sec} we prove the existence
of $\overline\PP$ as well as basic properties such
as translation-invariance and ergodicity. Then in
Section~\ref{perc_sec} we show how the
argument of Burton and Keane~\cite{burton-keane}
can be adapted to show that, almost surely under
$\overline\PP$, there is either no or exactly one
infinite connected cluster.
\subsection{Existence and basic properties}
\label{inf_rpr_sec}
In proving existence of the limit $\overline\PP$
of the sequence $\overline\PP_{N,r}$ we will
need to pay attention to
the underlying point processes,
and we will loosely follow the notational
conventions of Daley and Vere-Jones~\cite{dvj}
for point processes. Recall that the
labelling $\psi_\varnothing^{\mathfrak{t}_1}$ is a function of the
pair $(B,\tau)$ where $B$ is a point process
on $\mathcal{E}_N\times I_r$ and $\tau\in\{0,1\}^{\L_N}$,
and similarly $\hat\psi_\varnothing^{\mathfrak{t}_2}$ is a function of
$(\hat B,\hat G, \hat\tau)$.
(If $\b=\infty$ then $\tau,\hat\tau$ are redundant
due to the boundary conditions.) Write
\[
\mathcal{X}=\mathcal{X}_\b= \Big[\bigcup_{i=0}^d(\mathbb{Z}^d+\tfrac{1}{2}e_i)\Big]
\times I_\b,\qquad
T=\{0,1\}^{\mathbb{Z}^d},
\]
where $e_0$ is the zero vector and, for $i\neq0$,
$e_i$ is the unit vector in the $i$:th coordinate.
Writing $\mathcal{N}=\mathcal{N}_\mathcal{X}$ for the set of boundedly
finite point processes
(counting measures) on $\mathcal{X}$ (denoted $\mathcal{N}_\mathcal{X}^{\#}$
in~\cite{dvj}), an obvious mapping
allows us to see both $B$ and $\hat B\cup\hat G$
as random elements of $\mathcal{N}$.
The measure $\overline\PP_{N,r}$ factorizes as
\begin{equation}\label{P_decomp}
\overline\PP_{N,r}=\PP_{N,r}^{\mathfrak{t}_1}\times
\hat\PP_{N,r}^{\mathfrak{t}_2}\times P_\D,
\end{equation}
where $P_\D$ is the law of $\D$ and
$\PP_{N,r}^{\mathfrak{t}_1}$, $\hat\PP_{N,r}^{\mathfrak{t}_2}$
are the measures on $\mathcal{N}\times T$ governing
$(B,\tau)$ and $(\hat B\cup \hat G,\hat\tau)$
respectively, given by
\[
\frac{d\PP_{N,r}^{\mathfrak{t}_1}}{dE_{N,r}}=
\frac{\partial\psi_\varnothing^{\mathfrak{t}_1}}
{E_{N,r}(\partial\psi_\varnothing^{\mathfrak{t}_1})},\qquad
\frac{d\hat\PP_{N,r}^{\mathfrak{t}_2}}{d\hat E_{N,r}}=
\frac{\partial\hat\psi_\varnothing^{\mathfrak{t}_2}}
{\hat E_{N,r}(\partial\hat\psi_\varnothing^{\mathfrak{t}_2})}.
\]
Write $\Omega=(\mathcal{N}\times T)^2\times\mathcal{N}$ and note
that $\mathcal{N}$, and hence also $\Omega$, is a
complete and separable metric
space~\cite[Proposition~9.1.IV]{dvj}.
We equip $\Omega$ with the Borel $\sigma$-algebra
$\mathcal{F}$, which
coincides with the $\sigma$-algebra generated by
finite-dimensional distributions.
Recall that $r$ is either
fixed (if $\b<\infty$) or $r=2N$ (if $\b=\infty$).
\begin{proposition}\label{weak_prop}
The measures $\overline\PP_{N,r}$ converge
weakly to a probability measure $\overline\PP$
on $\Omega$ as $N\to\infty$.
\end{proposition}
\begin{proof}
The measure $P_\D$ does not depend on $N$,
so by~\cite[Theorem~2.8]{billingsley-conv}
convergence of $\overline\PP_{N,r}$ follows once we
show convergence of $\PP_{N,r}^{\mathfrak{t}_1}$
and $\hat\PP_{N,r}^{\mathfrak{t}_2}$. We give full details for the
case of $\PP_{N,r}^{\mathfrak{t}_1}$, the case of
$\hat\PP_{N,r}^{\mathfrak{t}_2}$ is similar. We first show that
the sequence $\PP_{N,r}^{\mathfrak{t}_1}$ is tight, i.e.\ every
subsequence can be refined to a further subsequence
along which $\PP_{N,r}^{\mathfrak{t}_1}$ converges. We then show
that there is a $\pi$-system $\mathcal{A}_0$ such that
the limit of $\PP_{N,r}^{\mathfrak{t}_1}(A)$ exists for each
$A\in\mathcal{A}_0$. The result then follows in
the standard way from Prohorov's theorem.
Turning to the tightness of $\PP_{N,r}^{\mathfrak{t}_1}$,
note that since $T$ is compact it suffices
to show that the marginal of $\PP_{N,r}^{\mathfrak{t}_1}$
on $\mathcal{N}$ is tight. A criterion for this
is given in~\cite[Proposition~11.1.VI]{dvj}.
Fix $N_0,r_0$ and write $X=|B\cap F(N_0,r_0)|$
for the number of points of $B$ `inside'
$K(N_0,r_0)$. Tightness follows if we show
that for for each $\varepsilon>0$ there is $m$
such that
\[
\PP_{N,r}^{\mathfrak{t}_1}(X>m)<\varepsilon\mbox{ for all } N,r.
\]
By Markov's inequality it suffices to show that
there is a constant $C(N_0,r_0)$ not depending
on $N,r$ such that the expectation
\begin{equation}\label{tight_ineq}
\PP^{\mathfrak{t}_1}_{N,r}[X] \leq C(N_0,r_0)
\mbox{ for all } N,r.
\end{equation}
In proving~\eqref{tight_ineq} we will require
the following notation. Write
$K=K(N,r)$, $K_0=K(N_0,r_0)$, $K'=K\setminus K_0$,
$F=F(N,r)$, $F_0=F(N_0,r_0)$, $F'=F\setminus F_0$,
$B'=B\cap F'$, and $B^{(0)}=B\cap F_0$.
For briefer notation write $\psi$ for
the labelling $\psi_\varnothing^{\mathfrak{t}_1}$, and let
$\psi'$ denote the restriction of $\psi$
to $K'$. Recall that $\epsilon$ denotes
the total Lebesgue measure of $K$ labelled
`even' in $\psi$, and write $\epsilon'$
for the Lebesgue measure of the `even'
subset of $K'$. Letting $\mathcal{C}$ denote the event
that $B$ is consistent (with source set $\varnothing$)
we have that $\partial\psi=e^{2\d\epsilon}\hbox{\rm 1\kern-.27em I}_\mathcal{C}$.
Clearly
$\epsilon'\leq \epsilon\leq \epsilon'+|K_0|$,
where $|K_0|=(2N_0+1)^dr_0$ denotes the total
Lebesgue measure of $K_0$, and it follows that
\begin{equation}\label{E_ratio}
\PP^{\mathfrak{t}_1}_{N,r}[X]\leq
e^{2\d |K_0|}\frac{E_{N,r}(X e^{2\d \epsilon'}\hbox{\rm 1\kern-.27em I}_\mathcal{C})}
{E_{N,r}(e^{2\d \epsilon'}\hbox{\rm 1\kern-.27em I}_\mathcal{C})}.
\end{equation}
The difficulty lies in the fact that although
$X$ is a function of $B^{(0)}$ only,
both $\epsilon'$ and $\mathcal{C}$ depend both on both
$B^{(0)}$ and $B'$. To `separate' this dependence,
we introduce the notation
\[
Z_{xy}=|B_{xy}\cap I_r|,\quad Z^{(0)}_{xy}=|B_{xy}\cap I_{r_0}|,
\quad Z'_{xy}=|B_{xy}\setminus I_{r_0}|,\quad (xy\in\mathcal{E}_N).
\]
Thus $Z_{xy}=Z^{(0)}_{xy}+Z'_{xy}$,
and the $Z^{(0)}_{xy},Z'_{xy}$ are independent Poisson
random variables under $E_{N,r}$
(with parameters $\l r_0$ and $\l(r-r_0)$,
respectively). The number of switching points
on $\{x\}\times I_r$ can be written
\[
|S_x|=\sum_{\substack{y\in \L_N\\y\sim x}}Z_{xy},
\]
and letting $S^{(0)}_x$ denote the set of switching
points in $\{x\}\times I_{r_0}$ we can similarly write
\[
|S^{(0)}_x|=\sum_{\substack{y\in \L_{N_0}\\y\sim x}}Z^{(0)}_{xy}.
\]
The process $B'$ imposes parity constraints on
$B^{(0)}$ which can be described in terms of the
random vector $\pi=(\pi_x:x\in\L_{N_0})\in\{0,1\}^{\L_{N_0}}$
given by
\begin{equation*}
\pi_x\equiv \sum_{\substack{y\not\in\L_{N_0}\\y\sim x}} Z^{(0)}_{xy}
+\sum_{\substack{y\in \L_N\\y\sim x}} Z'_{xy}.
\end{equation*}
(Here and in what follows we write $\equiv$ for
congruence modulo 2.) Note that $\pi$ is a
function of $B'$ only, and that
$\mathcal{C}=\mathcal{C}'\cap\mathcal{C}^{(0)}\cap\tilde\mathcal{C}$, where
\[
\begin{split}
\mathcal{C}'&=\{|S_x|\equiv 0\;\forall x\in\L_N\setminus\L_{N_0}\},\\
\mathcal{C}^{(0)}&=\{|S_x^{(0)}|\equiv \pi_x\;\forall x\in\L_{N_0}\},\\
\tilde\mathcal{C}&=
\Big\{\exists\, z\in\{0,1\}^{\mathcal{E}_{N_0}}:\forall x\in\L_{N_0},
\sum_{\substack{y\sim x\\y\in\L_{N_0}}} z_{xy}\equiv \pi_x\Big\}.
\end{split}
\]
Strictly speaking the event $\tilde\mathcal{C}$ is redundant
as it is implied by $\mathcal{C}^{(0)}$, however
it is useful to keep since it, in contrast to $\mathcal{C}^{(0)}$,
depends on $B'$ only. For each realization of $\pi$
such that $\tilde\mathcal{C}$ holds we fix a deterministic
vector $z$ as in the definition of $\tilde\mathcal{C}$.
Note that the number of possible $\pi$ is at most
$2^{|\L_{N_0}|}$.
In the numerator of~\eqref{E_ratio} we have
\[\begin{split}
E_{N,r}(Xe^{2\d\epsilon'}\hbox{\rm 1\kern-.27em I}_{\mathcal{C}})&=
E_{N,r}\big(e^{2\d\epsilon'}\hbox{\rm 1\kern-.27em I}_{\mathcal{C}'\cap\tilde\mathcal{C}}
E_{N,r}(X\hbox{\rm 1\kern-.27em I}_{\mathcal{C}^{(0)}}\mid\psi')\big)\\
&\leq E_{N,r}(X)E_{N,r}(e^{2\d\epsilon'}\hbox{\rm 1\kern-.27em I}_{\mathcal{C}'\cap\tilde\mathcal{C}}),
\end{split}\]
where we bounded $\hbox{\rm 1\kern-.27em I}_{\mathcal{C}^{(0)}}$ by 1 and used the fact
that $X$ is independent of $\psi'$.
Note that $E_{N,r}(X)=\l r_0|\mathcal{E}_{N_0}|$.
In the denominator of~\eqref{E_ratio} we have
\[
E_{N,r}(e^{2\d\epsilon'}\hbox{\rm 1\kern-.27em I}_{\mathcal{C}})=
E_{N,r}\big(e^{2\d\epsilon'}\hbox{\rm 1\kern-.27em I}_{\mathcal{C}'\cap\tilde\mathcal{C}}
P_{N,r}(\mathcal{C}^{(0)}\mid \psi')\big),
\]
which we need to bound from below.
We claim that there is an $\varepsilon=\varepsilon(N_0,r_0)>0$
such that $P_{N,r}(\mathcal{C}^{(0)}\mid \psi')\geq \varepsilon$
for all realizations $\psi'$ such that $\tilde\mathcal{C}$
holds. Indeed, recall that we fixed a deterministic
vector $z$ for each $\pi$ such that $\tilde\mathcal{C}$ holds.
The event
$\tilde\mathcal{C}_z=\{Z^{(0)}_{xy}=z_{xy}\forall xy\in \mathcal{E}_{N_0}\}$
thus implies $\mathcal{C}^{(0)}$. Under
$P_{N,r}(\cdot\mid \psi')$ the $Z^{(0)}_{xy}$
are independent Poisson random variables,
so each $\tilde\mathcal{C}_z$ has positive probability.
The claim thus holds with $\varepsilon$
being the minimum of $P_{N,r}(\tilde\mathcal{C}_z\mid \psi')$
over the (at most $2^{|\L_{N_0}|}$)
choices of $z$.
Therefore~\eqref{tight_ineq} follows with
\[
C(N_0,r_0)=e^{2\d |K_0|}\l r_0|\mathcal{E}_{N_0}|/\varepsilon(N_0,r_0).
\]
Having proved tightness of the sequence $\PP_{N,r}^{\mathfrak{t}_1}$
we now turn to showing uniqueness of subsequential
limits. Let $\mathcal{A}_0$ denote the
collection of events of the form
\begin{equation*}
A=\{\psi\mbox{ is `even' in } J\},
\end{equation*}
where $J$ is any finite union of bounded
closed intervals
in $\mathbb{K}_\b$.
(We allow intervals of length 0, i.e.\ isolated
points.)
Then $\mathcal{A}_0$ is a $\pi$-system which
generates the $\sigma$-algebra $\mathcal{F}$
(note that the process $B$ can be recovered
from the labelling $\psi$).
We let $N$ be large enough that $J\subseteq K(N,r)$.
By~\eqref{holes_eq} and~\cite[Lemma~3.2]{bjogr}
we have
\begin{equation}\label{LJ_eq}
\PP^{\mathfrak{t}_1}_{N,r}(A)=c(J)\mu_{N,r}^{\mathfrak{f},\mathfrak{t}_1}
\big[\exp\big(-\l L_J(\sigma)\big)\big],
\end{equation}
for some constant $c(J)$ depending only on $J$,
and where
\begin{equation*}
L_J(\sigma)=\sum_{xy\in \mathcal{E}_N}
\int_{I_r}\sigma(x,t)\sigma(y,t)
\hbox{\rm 1\kern-.27em I}\{(xy,t)\in \tilde J\}dt
\end{equation*}
and $\tilde J$ is the
set of point $(yz,t)\in\mathbb{F}$ such that
$(y,t)\in J$ or $(z,t)\in J$ (or both).
By~\cite[Theorem~2.5.1]{bjo_phd}
the sequence of measures $\mu^{\mathfrak{f},\mathfrak{t}_1}_{N,r}$
converges weakly, hence the probability
in~\eqref{LJ_eq} converges.
\end{proof}
For later reference we note that the constant $c(J)$
in~\eqref{LJ_eq} can be written as
\begin{equation}\label{cJ_eq}
c(J)=2^{-n}e^{\d|J|+\l|\tilde J|}.
\end{equation}
Here $|J|$ and $|\tilde J|$ denote the total length
of the intervals comprising $J$ and $\tilde J$,
and $n$ is the difference between the number
of intervals comprising $K'=K\setminus J$ and
the number of intervals comprising $K$
(in counting the number of intervals we view
$I_\b$ as a circle when $\mathfrak{t}=\mathfrak{p}$).
For $(x,s)\in\mathbb{K}_\b$ define the translation
or shift $\tau_{(x,s)}:\mathbb{K}_\b\to\mathbb{K}_\b$ by
$\tau_{(x,s)}(y,t)=(y+x,t+s)$
where in the case $\b<\infty$ we view $t+s$
modulo $\b$. For a function $\zeta:\mathbb{K}_\b\to\RR$
(e.g.\ a labelling $\psi$
or a spin-configuration $\sigma$) we define
$\tau_{(x,s)}(\zeta)$
by $[\tau_{(x,s)}(\zeta)](y,t)=\zeta(y+x,t+s)$.
We write $\tau_x$ for $\tau_{(x,0)}$.
\begin{proposition}\label{ergodic_prop}
The measure $\overline\PP$ is invariant with respect
to the shifts $\tau_{(x,s)}$, and
ergodic with respect to the shifts $\tau_x$
for $x\neq0$.
\end{proposition}
\begin{proof}
We use the decomposition~\eqref{P_decomp}.
The measure $P_\D$ is translation-invariant
and ergodic, so it suffices to show that the weak
limits of $\PP_{N,r}^{\mathfrak{t}_1}$ and
$\hat\PP_{N,r}^{\mathfrak{t}_2}$ and are translation-invariant
and ergodic. Again, we give details for
$\PP_{N,r}^{\mathfrak{t}_1}$. Let $A\in\mathcal{A}_0$,
let $(x,s)\in\mathbb{K}_\b$, and let $N$ be large enough
that $\tau_{(x,s)}^{-1}J\subseteq K(N,r)$
(recall that $r=2N$ if $\b=\infty$ and $r=\b$
if $\b<\infty$). We have as in~\eqref{LJ_eq} that
\begin{equation*}
\PP^{\mathfrak{t}_1}_{N,r}(\tau_{(x,s)}A)=c(\tau_{(x,s)}^{-1}J)
\mu_{N,r}^{\mathfrak{f},\mathfrak{t}_1}
\big[\exp\big(-\l L_{\tau_{(x,s)}^{-1}J}(\sigma)\big)\big].
\end{equation*}
From~\eqref{cJ_eq} we see that
$c(\tau_{(x,s)}^{-1}J)=c(J)$.
By~\cite[Theorem~2.5.1]{bjo_phd},
\[\begin{split}
\lim_{N\to\infty} \mu_{N,r}^{\mathfrak{f},\mathfrak{t}_1}
\big[\exp\big(-\l L_{\tau_{(x,s)}^{-1}J}(\sigma)\big)\big]
&=\lim_{N\to\infty} \mu_{N,r}^{\mathfrak{f},\mathfrak{t}_1}
\big[\exp\big(-\l L_J(\sigma)\big)\big]\\
&=\mu^{(\mathfrak{f},\mathfrak{t}_1)}
\big[\exp\big(-\l L_J(\sigma)\big)\big],
\end{split}\]
and hence
$\lim_{N\to\infty}\PP^{\mathfrak{t}_1}_{N,r}(\tau_{(x,s)}A)=
\lim_{N\to\infty}\PP^{\mathfrak{t}_1}_{N,r}(A)$.
This proves translation-invariance on the
$\pi$-system $\mathcal{A}_0$, which by the $\pi$-systems
lemma implies full
translation-invariance.
Let $A_1,A_2\in\mathcal{A}_0$ be the events that
$\psi$ is `even' in $J_1$ and $J_2$, respectively.
For $\|x\|$ and $N$ large enough we have
from~\eqref{LJ_eq} and~\eqref{cJ_eq} that
\[
\frac{\PP_{N,r}^{\mathfrak{t}_1}(A_1\cap\tau_x A_2)}
{\PP_{N,r}^{\mathfrak{t}_1}(A_1)\PP_{N,r}^{\mathfrak{t}_1}(A_2)}
=\frac{\mu_{N,r}^{\mathfrak{f},\mathfrak{t}_1}
\big[\exp\big(-\l L_{J_1}(\sigma)\big)
\exp\big(-\l L_{\tau^{-1}_{x}J_2}(\sigma)\big)\big]}
{\mu_{N,r}^{\mathfrak{f},\mathfrak{t}_1}
\big[\exp\big(-\l L_{J_1}(\sigma)\big)\big]
\mu_{N,r}^{\mathfrak{f},\mathfrak{t}_1}
\big[\exp\big(-\l L_{\tau^{-1}_{x}J_2}(\sigma)\big)\big]}.
\]
Letting $N\to\infty$ and writing $\PP^{(\mathfrak{t}_1)}$ for
the weak limit of $\PP_{N,r}^{\mathfrak{t}_1}$ it follows that
\[
\frac{\PP^{(\mathfrak{t}_1)}(A_1\cap\tau_x A_2)}
{\PP^{(\mathfrak{t}_1)}(A_1)\PP^{(\mathfrak{t}_1)}(A_2)}
=\frac{\mu^{(\mathfrak{f},\mathfrak{t}_1)}
\big[\exp\big(-\l L_{J_1}(\sigma)\big)
\exp\big(-\l L_{J_2}(\tau^{-1}_{x}\sigma)\big)\big]}
{\mu^{(\mathfrak{f},\mathfrak{t}_1)}
\big[\exp\big(-\l L_{J_1}(\sigma)\big)\big]
\mu^{(\mathfrak{f},\mathfrak{t}_1)}
\big[\exp\big(-\l L_{J_2}(\sigma)\big)\big]}.
\]
It follows from Lemma~\ref{mix_lem}
in the Appendix that the
ratio on the right-hand-side converges to 1
as $\|x\|\to\infty$. Thus $\PP^{(\mathfrak{t}_1)}$
is mixing on $\mathcal{A}_0$, hence also mixing on $\mathcal{F}$
and hence ergodic.
\end{proof}
\subsection{Percolation}
\label{perc_sec}
The notions of paths and connectivity, defined
for $\overline\PP_{N,r}$, extend to $\overline\PP$.
Thus $\mathbb{K}_\b$ decomposes into a random
collection of connected components or clusters
(each of these is a union of intervals
bounded by certain elements of $\D$). Let $U$ denote the
number of these clusters which are \emph{unbounded}.
The random variable $U$ (which may be infinite)
is invariant under all translations $\tau_x$,
and hence by Proposition~\ref{ergodic_prop} it is
$\overline\PP$-a.s.\ constant. We will show:
\begin{proposition}\label{perc_prop}
Either $\overline\PP(U=0)=1$ or $\overline\PP(U=1)=1$.
\end{proposition}
\begin{proof}
As this type of argument is fairly standard in
percolation theory,
and most of the details are the same as for the classical
model~\cite{adcs}, we only sketch the proof and highlight
what adjustments are needed for the quantum case.
We focus on the case $\b=\infty$.
Let $k$ be such that $\overline\PP(U=k)=1$, we must show that
$k\leq1$. The possibility that $2\leq k<\infty$
may be ruled out using
Proposition~\ref{local_prop}. Roughly speaking,
a large enough box $K(N,r)$ will intersect all
$k$ unbounded components with positive probability.
Using part (B) of Proposition~\ref{local_prop}
one may deduce that, with positive probability,
all these components are in fact connected
inside $K(N,r)$, a contradiction.
Now assume that $k=\infty$. To get a contradiction
one considers what are called
\emph{coarse-trifurcations}. These are
points $(x,t)\in\mathbb{K}$ with
$x\in (2N_0+1)\mathbb{Z}^d$
and $t\in 2r_0\mathbb{Z}$, having the properties that
(i) all points in $K_0+(x,t)$
are connected in $K_0+(x,t)$, and
(ii) $\mathbb{K}\setminus (K_0+(x,t))$
contains at least 3 distinct unbounded connected
components. Here $N_0$ and $r_0$ are fixed,
and may (using Proposition~\ref{local_prop}(B)
again) be chosen such that each $(x,t)$
is a coarse-trifurcation with probability $p>0$.
Note that $p$ is the same for all $(x,t)$
by translation-invariance. As in~\cite{adcs}
one may construct a graph $F$ which reflects the
connectivity structure of the coarse-trifurcations
in some large box $K(N,r)$. Roughly speaking,
the edges of $F$ either connect distinct
coarse-trifurcations, or they connect a coarse-trifurcation
with an `element' on the boundary of $K(N,r)$.
The latter occurs if the
boundary can be reached without having to
go too close to another coarse-trifurcation,
otherwise the former occurs.
The main difference to~\cite{adcs} is the correct
notion of `element' on the boundary. In their
case, where the underlying graphical structure is
discrete, one may take the `elements' simply as vertices
on the boundary. In the present case we instead
take it to mean a maximal $\D$-free interval
intersecting the boundary of $K(N,r)$; that is an
interval of the form $\{y\}\times I$ with $I\subseteq I_r$
maximal such that $\D_y\cap I=\varnothing$, and such that either
$y\in\partial\L_N$, or one of the endpoints of $I$
is $\pm r/2$. With this convention the graph $F$
contains no cycles, as if it did contain a cycle
this would violate the definition of coarse-trifurcation.
(The important point is that the
collection of maximal $\D$-free
intervals refines the collection of connected components.)
Also, the coarse-trifurcations correspond to vertices
of $F$ of degree at least 3. This implies that
the number of coarse-trifurcations is at most the
number of leaves of $F$, which is in turn bounded above
by the number of maximal $\D$-free interval
intersecting the boundary of $K(N,r)$. The expectation
of the latter is easily seen to be at most
$2(2N+1)^d+4\d r(2N+1)^{d-1}$, whereas the
expected number of coarse-trifurcations is of the
order $p[r(2N+1)^d]/[r_0(2N_0+1)^d]$.
Letting $N,r\to\infty$ this contradicts $p>0$,
finishing the proof.
\end{proof}
\section{Proof of the main result}
\label{proof_sec}
\subsection{The infrared bound}
We now describe the infrared bound of~\cite{bjo_irb}
and use it to prove a result of key importance
for Theorem~\ref{main_thm}. For technical
reasons we will in this section redefine the box
$\L_n$ as $\{-n+1,\dotsc,n\}^d$ so that it has even
sidelength rather than odd.
Recall the Schwinger function~\eqref{schwinger_eq}
and its probabilistic representation~\eqref{corr_eq2}.
Write
\begin{equation*}
c_{N,r}(x,t)=\el\sigma(0,0)\sigma(x,t)\rangle_{N,r}^{\mathfrak{p},\mathfrak{p}}.
\end{equation*}
Note that we use periodic boundary conditions
in both directions.
Although $c_{N,r}(x,t)$ is defined for $(x,t)\in K(N,r)$,
we extend the definition
to all of $\mathbb{K}$ by periodicity. We write
\begin{equation*}
\L^\star_N=\tfrac{\pi}{N}\L_N,\quad
I_r^\star=\tfrac{2\pi}{r}\mathbb{Z},\quad
K^\star_{N,r}=\L^\star_N\times I_r^\star.
\end{equation*}
Elements of $K^\star_{N,r}$ will be denoted $\xi=(k,\ell)$
where $k\in \L_N^\star$ and $\ell\in I_r^\star$.
For large $N,r$ we may see $\L_N^\star$
as an approximation of $(-\pi,\pi]^d$
and $I_r^\star$ as an
approximation of $\RR$.
For $p=(p_1,\dotsc,p_d)\in(-\pi,\pi]^d$ let
$\hat L(p)=\sum_{j=1}^d(1-\cos(p_j))$
denote the Fourier transform of the graph Laplacian of $\mathbb{Z}^d$,
and define
\begin{equation*}
E_{\l,\d}(p,q)=\frac{2\l\hat L(p)+q^2/2\d}{48},
\qquad p\in(-\pi,\pi]^d, q\in\RR.
\end{equation*}
The Fourier transform of $c_{N,r}$ is
\begin{equation*}
\hat c_{N,r}(\xi)=\sum_{x\in \L_N}\int_{I_r} c_{N,r}(x,t)
e^{ik\cdot x} e^{i\ell t}\,dt,\quad
\xi=(k,\ell)\in K_{N,r}^\star,
\end{equation*}
where $k\cdot x$ denotes the usual scalar product.
Note that $\hat c_{N,r}(\xi)\geq0$.
The infrared bound of~\cite{bjo_irb} states that
\begin{equation}\label{irb}
\mbox{if } \xi\in K_{N,r}^\star\setminus\{0\}
\mbox{ then } \hat c_{N,r}(\xi)\leq \frac{1}{E_{\l,\d}(\xi)}.
\end{equation}
We will use this to show the following:
\begin{lemma}\label{avg_lem}
Suppose $\b<\infty$ and $d\geq 3$. Then
\begin{equation}\label{avg_pos}
\lim_{n\to\infty}\frac{1}{|\L_n|}\sum_{x\in \L_n}\int_{I_\b}
\el\sigma(0,0)\sigma(x,t)\rangle_{\l_\crit,\b}^{(\mathfrak{f},\mathfrak{p})}\,dt=0.
\end{equation}
Suppose $\b=\infty$ and $d\geq 2$. Then
\begin{equation}\label{avg_ground}
\lim_{n,r\to\infty}\frac{1}{|\L_n|r}\sum_{x\in \L_n}\int_{I_r}
\el\sigma(0,0)\sigma(x,t)\rangle_{\l_\crit,\infty}^{(\mathfrak{f},\mathfrak{f})}\,dt=0.
\end{equation}
\end{lemma}
\begin{proof}
We begin by showing that the stated conditions
on $d$ imply the following:
\begin{equation}\label{L1_pos}
\mbox{if $\b<\infty$ then }
\int_{(-\pi,\pi]^d}dp\sum_{\ell\in I_\b^\star} \frac{1}{E_{\l,\d}(p,\ell)}<\infty,
\end{equation}
and
\begin{equation}\label{L1_ground}
\mbox{if $\b=\infty$ then }
\int_{(-\pi,\pi]^d}dp\int_\RR dq \frac{1}{E_{\l,\d}(p,q)}<\infty.
\end{equation}
Firstly, in the case $\b<\infty$ we have that
\begin{equation*}\begin{split}
\int_{(-\pi,\pi]^d}dp\sum_{\ell\in I_\b^\star}
\frac{1}{E_{\l,\d}(p,\ell)}
&=\int_{(-\pi,\pi]^d}dp\sum_{\ell\in I_\b^\star}
\frac{48}{2\l\hat L(k)+\ell^2/2\d}\\
&=\frac{96\d}{(2\pi/\b)^2}\int_{(-\pi,\pi]^d}dp
\sum_{m\in\mathbb{Z}}\frac{1}{a(p)+m^2}\\
&\leq\frac{96\d}{(2\pi/\b)^2}\int_{(-\pi,\pi]^d}
\Big(\frac{1}{a(p)}+\frac{\pi}{a(p)^{1/2}}\Big)dp,
\end{split}
\end{equation*}
where $a(p)=\l\d\b^2\hat L(p)/\pi^2$. Note that
$1/\hat L(p)$ diverges for $p\to 0$, in the manner
of $1/\|p\|_2^2$, and that for $\a>0$
\begin{equation}\label{L-int-alpha}
\int_{(-\pi,\pi]^d}\frac{1}{\hat L(p)^\a}dp<\infty
\mbox{ if and only if } d>2\a.
\end{equation}
Thus~\eqref{L1_pos} holds for $d>2$, i.e.\ for $d\geq3$
as claimed. In the case $\b=\infty$ we have
\begin{equation*}\begin{split}
\int_{(-\pi,\pi]^d}dp\int_{-\infty}^\infty dq \frac{1}{E_{\l,\d}(p,q)}
&=96\d \int_{(-\pi,\pi]^d}dp\int_{-\infty}^\infty dq
\frac{1}{4\l\d\hat L(p)+q^2}\\
&=48\pi\sqrt{\d/\l}
\int_{(-\pi,\pi]^d}\frac{1}{\hat L(p)^{1/2}}dp.
\end{split}\end{equation*}
Thus by~\eqref{L-int-alpha} we have~\eqref{L1_ground}
for $d>1$, i.e.\ for $d\geq 2$ as claimed.
Now define the function
\begin{equation*}
G_{N,r}((x,s),(y,t))=\frac{1}{(2N)^dr}
\sum_{(k,\ell)\in K_N^\star\setminus\{0\}}
\frac{e^{-ik\cdot (x-y)}e^{-i\ell(s-t)}}{E_{\l,\d}(k,\ell)},
\end{equation*}
where $x,y\in\mathbb{Z}^d$ and $s,t\in\RR$. In the case when
$\b<\infty$ then by Riemann approximation
\begin{equation*}
\begin{split}
G_{N,\b}((x,s),(y,t))&\to\frac{1}{(2\pi)^d\b}
\int_{(-\pi,\pi]^d}dp\sum_{\ell\in I_\b^\star}
\frac{e^{-ip\cdot (x-y)}e^{-i\ell(s-t)}}{E_{\l,\d}(p,\ell)}\\
&=:G_{\b}((x,s),(y,t)),\mbox{ as }N\to\infty.
\end{split}
\end{equation*}
From~\eqref{L1_pos} we deduce that
\[
\frac{1}{(2n)^{d}}\sum_{x\in\L_n} \int_{I_\b}
dt \,G_\b((0,0),(x,t))\to 0\mbox{ as }n\to\infty,
\]
which in turn implies that
\begin{equation}\label{RL_pos}
\frac{1}{(2n)^{2d}}\sum_{x,y\in\L_n} \iint_{I_\b\times I_\b}
dsdt \,G_\b((x,s),(y,t))\to 0\mbox{ as }n\to\infty.
\end{equation}
Similarly, for $\b=\infty$ we have that
\begin{equation*}\begin{split}
G_{N,r}((x,s),(y,t))&\to\frac{1}{(2\pi)^{d+1}}
\int_{(-\pi,\pi]^d}dp\int_\RR dq
\frac{e^{-ip\cdot (x-y)}e^{-iq(s-t)}}{E_{\l,\d}(p,q)}\\
&=:G_{\infty}((x,s),(y,t)),\mbox{ as }N,r\to\infty,
\end{split}
\end{equation*}
and hence using~\eqref{L1_ground} that
\begin{equation}\label{RL_ground}
\frac{1}{(2n)^{2d}r^2}\sum_{x,y\in\L_n} \iint_{I_r\times I_r}
dsdt \,G_\infty((x,s),(y,t))\to 0\mbox{ as }n,r\to\infty.
\end{equation}
We now show how~\eqref{RL_pos} and~\eqref{RL_ground}
imply~\eqref{avg_pos} and~\eqref{avg_ground}, respectively.
Note that by Fourier inversion
\begin{equation*}
c_{N,r}(x,t)=\frac{1}{(2N)^dr}\sum_{k\in\L_N^\star}
\sum_{\ell\in I_r^\star} \hat c_{N,r}(k,\ell) e^{-ik\cdot x}e^{-i\ell t}.
\end{equation*}
Let $v:K(N,r)\to\mathbb{C}$ be an aribtrary bounded, measurable function.
It follows that
\begin{equation}\label{v_eq}
\begin{split}
\sum_{x,y\in\L_N}&\iint_{I_r\times I_r}dsdt \,v(x,s)\overline{v(y,t)}
c_{N,r}(x-y,s-t)\\
&=\frac{1}{(2N)^dr}\sum_{\xi\in K_{N,r}^\star}
\hat c_{N,r}(\xi) |z_v(\xi)|^2,
\end{split}
\end{equation}
where
\begin{equation*}
z_v(k,\ell)=\sum_{x\in\L_N}\int_{I_r}v(x,s) e^{-ik\cdot x}e^{-i\ell s}ds.
\end{equation*}
Using the infrared bound~\eqref{irb},
\begin{equation*}
\sum_{\xi\in K_{N,r}^\star}
\hat c_{N,r}(\xi) |z_v(\xi)|^2
\leq\sum_{\xi\in K_{N,r}^\star\setminus\{0\}}
\frac{1}{E_{\l,\d}(\xi)} |z_v(\xi)|^2+
\hat c_{N,r}(0) |z_v(0)|^2.
\end{equation*}
Note that
\begin{equation*}
\hat c_{N,r}(0) =\sum_{x\in\L_N}\int_{I_r} c_{N,r}(x,t)dt=:
\chi^{\mathfrak{p},\mathfrak{p}}_{N,r}
\end{equation*}
equals the (finite-volume) susceptibility.
Interchanging the order of summation again thus gives
\begin{multline}\label{v_rhs_eq}
\frac{1}{(2N)^dr}\sum_{\xi\in K_{N,r}^\star}
\hat c_{N,r}(\xi) |z_v(\xi)|^2\\\leq
\sum_{x,y\in\L_N}\iint_{I_r\times I_r} v(x,s) \overline{v(y,t)}
G_{N,r}((x,s),(y,t))+\frac{|z_v(0)|^2}{(2N)^dr} \chi^{\mathfrak{p},\mathfrak{p}}_{N,r}.
\end{multline}
Let $N_0<N$, and as usual let $r_0<r$
if $\b=\infty$, alternatively $r_0=r=\b$
if $\b<\infty$. Set
\[
v(x,s)=\hbox{\rm 1\kern-.27em I}\{x\in\L_{N_0},s\in I_{r_0}\}.
\]
In what follows we use the same notation
$\el\cdot\rangle_{N,r}^{\mathfrak{f}}$ for both
$\el\cdot\rangle_{N,r}^{\mathfrak{f},\mathfrak{f}}$ (in the case
$\b=\infty$) and $\el\cdot\rangle_{N,\b}^{\mathfrak{f},\mathfrak{p}}$ (in the case
$\b<\infty$). We also write
$\el\cdot\rangle_{\l,\b}^{(\mathfrak{f})}$ for both
infinite-volume limits $\el\cdot\rangle_{\l,\infty}^{(\mathfrak{f},\mathfrak{f})}$
and $\el\cdot\rangle_{\l,\b}^{(\mathfrak{f},\mathfrak{p})}$. By the monotonicity~\eqref{corr_comp}
of correlation functions we have that
\begin{equation*}
c_{N,r}(x-y,s-t)=\el\sigma(x,s)\sigma(y,t)\rangle_{N,r}^{\mathfrak{p},\mathfrak{p}}
\geq \el\sigma(x,s)\sigma(y,t)\rangle_{N,r}^{\mathfrak{f}}.
\end{equation*}
Thus for our choice of $v$ we have that
the left-hand-side of~\eqref{v_eq} is at
least
\begin{equation*}
\sum_{x,y\in\L_{N_0}}\iint_{I_{r_0}\times I_{r_0}}
\el\sigma(x,s)\sigma(y,t)\rangle_{N,r}^{\mathfrak{f}} dsdt.
\end{equation*}
By~\eqref{v_rhs_eq} it follows that
\begin{equation}\label{chi_bound}
\begin{split}
\sum_{x,y\in\L_{N_0}}&\iint_{I_{r_0}\times I_{r_0}}
\el\sigma(x,s)\sigma(y,t)\rangle_{N,r}^{\mathfrak{f}} dsdt\\
&\leq \sum_{x,y\in\L_{N_0}}\iint_{I_{r_0}\times I_{r_0}} dsdt
G_{N,r}((x,s),(y,t))+\frac{(2N_0)^dr_0}{(2N)^dr} \chi^{\mathfrak{p},\mathfrak{p}}_{N,r}.
\end{split}
\end{equation}
Now let $\l<\l_\crit$. This
implies that $\el\cdot\rangle_{\l,\b}^{(\mathfrak{f})}$ is the unique
infinite-volume limit of the
measures $\el\cdot\rangle_{N,r}^{\mathfrak{s},\mathfrak{t}}$, and
by finiteness of the susceptibility~\cite[Theorem~6.6]{bjogr}
and the dominated convergence theorem we have that
\begin{equation*}
\chi^{\mathfrak{p},\mathfrak{p}}_{N,r}\to
\sum_{x\in\mathbb{Z}^d}\int_{I_\b}\el\sigma(0,0)\sigma(x,t)\rangle_{\l,\b}^{(\mathfrak{f})}dt<\infty,
\end{equation*}
as $N\to\infty$ (for $\b=r<\infty$) or
$N,r=2N\to\infty$ (for $\b=\infty$).
Hence, letting $N\to\infty$ or $N,r\to\infty$ as appropriate, we obtain
from~\eqref{chi_bound} that
for all $\l<\l_\crit$ we have
\begin{equation}\label{G_bound}
\begin{split}
\sum_{x,y\in\L_{N_0}}&\iint_{I_{r_0}\times I_{r_0}}
\el\sigma(x,s)\sigma(y,t)\rangle_{\l,\b}^{(\mathfrak{f})} dsdt\\
&\leq \sum_{x,y\in\L_{N_0}}\iint_{I_{r_0}\times I_{r_0}}
G_\b((x,s),(y,t))\,dsdt.
\end{split}
\end{equation}
Letting $\l\uparrow\l_\crit$ and using the fact that
$\el\sigma(x,s)\sigma(y,t)\rangle_{\l,\b}^{(\mathfrak{f})}$ is left-continuous in $\l$
(since any two increasing limits can be interchanged),
we get that~\eqref{G_bound} holds also with $\l=\l_\crit$.
Letting
$N_0\to\infty$ or $N_0,r_0\to\infty$ as
appropriate we get from~\eqref{RL_pos} and~\eqref{RL_ground}
that
\[
\frac{1}{(2N_0)^{2d}r_0^2}
\sum_{x,y\in\L_{N_0}}\iint_{I_{r_0}\times I_{r_0}}
\el\sigma(x,s)\sigma(y,t)\rangle_{\l_\crit,\b}^{(\mathfrak{f})} \,dsdt\to0.
\]
Using translation-invariance and nonnegativity
of $\el\sigma(x,s)\sigma(y,t)\rangle_{\l_\crit,\b}^{(\mathfrak{f})}$, the
results~\eqref{avg_pos} and~\eqref{avg_ground} follow.
\end{proof}
\subsection{Proof of Theorem~\ref{main_thm}}
From this point the argument is almost identical
to that for the classical model~\cite{adcs},
however it is also short and elegant so we include
the remaining steps.
We begin by deducing from Lemma~\ref{avg_lem}
the following consequence for the number
$U$ of unbounded components under the measure $\overline\PP$.
\begin{proposition}\label{crit-perc_prop}
Under the conditions in Lemma~\ref{avg_lem}
and for $\l=\l_\crit$
we have that $\overline\PP(U=0)=1$.
\end{proposition}
\begin{proof}
Recall our convention on boundary conditions for the
measure $\overline\PP_{N,r}$: if $\b<\infty$ we write $\mathfrak{t}_1=\mathfrak{t}_2=\mathfrak{p}$,
if $\b=\infty$ we write $\mathfrak{t}_1=\mathfrak{f}$ and $\mathfrak{t}_2=\mathfrak{w}$. By the
definition of $\overline\PP_{N,r}$ and the Switching
Lemma~\ref{sw_lem} we have for any $N,r$
and $(x,s),(y,t)\in K(N,r)$ that
\[\begin{split}
\overline\PP_{N,r}((x,s)\leftrightarrow(y,t))&=
\el\sigma(x,s)\sigma(y,t)\rangle_{N,r}^{\mathfrak{f},\mathfrak{t}_1}
\el\sigma(x,s)\sigma(y,t)\rangle_{N,r}^{\mathfrak{w},\mathfrak{t}_2}\\
&\leq\el\sigma(x,s)\sigma(y,t)\rangle_{N,r}^{\mathfrak{f},\mathfrak{t}_1}.
\end{split}\]
By the convergence of the correlation function
and using Proposition~\ref{weak_prop}
(and a small, but standard, additional argument)
we get that
\begin{equation}\label{p-c}
\overline\PP((x,s)\leftrightarrow(y,t))\leq
\el\sigma(x,s)\sigma(y,t)\rangle_{\l_\crit,\b}^{(\mathfrak{f},\mathfrak{t}_1)}.
\end{equation}
Writing $\{(x,s)\leftrightarrow\infty\}$ for the event that
$(x,s)$ lies in an unbounded component we have
using Jensen's inequality and the fact that
$U\leq1$ (Proposition~\ref{perc_prop}) that
\[\begin{split}
\Big( r|\L_N| \overline\PP((0,0)\leftrightarrow\infty)\Big)^2
&\leq \overline\PP\Big( \Big(\sum_{x\in\L_N}
\int_{I_r} \hbox{\rm 1\kern-.27em I}\{(x,s)\leftrightarrow\infty\}ds\Big)^2\Big)\\
&= \sum_{x,y\in\L_N}
\iint_{I_r\times I_r} \overline\PP((x,s),(y,t)\leftrightarrow\infty)dsdt\\
&\leq \sum_{x,y\in\L_N}
\iint_{I_r\times I_r} \overline\PP((x,s)\leftrightarrow(y,t))dsdt\\
&= r|\L_N|\sum_{x\in\L_N}
\int_{I_r} \overline\PP((0,0)\leftrightarrow(x,s))ds.
\end{split}\]
Using~\eqref{p-c} we deduce that
\[
\overline\PP((0,0)\leftrightarrow\infty)^2\leq
\frac{1}{r|\L_N|}\sum_{x\in\L_N}
\int_{I_r} \el\sigma(0,0)\sigma(x,s)\rangle_{\l_\crit,\b}^{(\mathfrak{f},\mathfrak{t}_1)}
ds.
\]
Letting $N\to\infty$ or $N,r\to\infty$ as appropriate,
and using Lemma~\ref{avg_lem}, the result follows.
\end{proof}
Turning to the final steps in the
proof of Theorem~\ref{main_thm}, we recall from
Proposition~\ref{local_prop} that for each
$(x,t)\in\mathbb{K}_\b$ there is a constant $C_{(x,t)}$
such that for all $N,r$ we have
\[
\el\sigma(0,0)\sigma(x,t)\rangle_{N,r}^{\mathfrak{w},\mathfrak{t}_2}
-\el\sigma(0,0)\sigma(x,t)\rangle_{N,r}^{\mathfrak{f},\mathfrak{t}_1}\leq
C_{(x,t)} \overline\PP_{N,r}((0,0)\leftrightarrow\Gamma).
\]
Also note that for all $N_0\leq N$ and $r_0\leq r$
we have that
$\overline\PP_{N,r}((0,0)\leftrightarrow\Gamma)\leq\overline\PP_{N,r}((0,0)\leftrightarrow\partial K(N_0,r_0))$,
since any path to $\Gamma$ must leave $K(N_0,r_0)$.
Letting $N\to\infty$ (respectively, $N,r\to\infty$)
and then $N_0\to\infty$ (respectively, $N_0,r_0\to\infty$)
it follows from Proposition~\ref{crit-perc_prop} that
\[
\el\sigma(0,0)\sigma(x,t)\rangle_{\l_\crit,\b}^{(\mathfrak{w},\mathfrak{t}_2)}
-\el\sigma(0,0)\sigma(x,t)\rangle_{\l_\crit,\b}^{(\mathfrak{f},\mathfrak{t}_1)}\leq
C_{(x,t)} \overline\PP((0,0)\leftrightarrow\infty)=0.
\]
Thus $\el\sigma(0,0)\sigma(x,t)\rangle_{\l_\crit,\b}^{(\mathfrak{w},\mathfrak{t}_2)}
=\el\sigma(0,0)\sigma(x,t)\rangle_{\l_\crit,\b}^{(\mathfrak{f},\mathfrak{t}_1)}$.
By translation-invariance and
the Griffiths inequality
(proved in detail for the present model
in~\cite[Lemma~2.2.20]{bjo_phd}) it follows that
\[\begin{split}
\big(\el\sigma(0,0)\rangle_{\l_\crit,\b}^{(\mathfrak{w},\mathfrak{t}_2)}\big)^2&=
\el\sigma(0,0)\rangle_{\l_\crit,\b}^{(\mathfrak{w},\mathfrak{t}_2)}
\el\sigma(x,t)\rangle_{\l_\crit,\b}^{(\mathfrak{w},\mathfrak{t}_2)}\leq
\el\sigma(0,0)\sigma(x,t)\rangle_{\l_\crit,\b}^{(\mathfrak{w},\mathfrak{t}_2)}\\
&=\el\sigma(0,0)\sigma(x,t)\rangle_{\l_\crit,\b}^{(\mathfrak{f},\mathfrak{t}_1)}.
\end{split}\]
Thus using Lemma~\ref{avg_lem} again,
if $\b<\infty$ and $d\geq3$ or
$\b=\infty$ and $d\geq2$ then
\[
\big(\el\sigma(0,0)\rangle_{\l_\crit,\b}^{(\mathfrak{w},\mathfrak{t}_2)}\big)^2
\leq \frac{1}{r|\L_N|}\sum_{x\in\L_N}\int_{I_r}
\el\sigma(0,0)\sigma(x,t)\rangle_{\l_\crit,\b}^{(\mathfrak{f},\mathfrak{t}_2)}dt\to 0,
\]
hence by~\eqref{spont_resid} we have
$M^+_\b(\l_\crit)=\el\sigma(0,0)\rangle_{\l_\crit,\b}^{(\mathfrak{w},\mathfrak{t}_2)}=0$
as claimed.\qed
\section{Appendix: mixing in the space--time spin representation}
In this section we prove mixing results for
the infinite-volume space--time spin measures
$\mu^{(\mathfrak{s},\mathfrak{t})}_\b$ defined in Section~\ref{spin_sec}.
As usual we let $\mathfrak{t}=\mathfrak{p}$ if $\b<\infty$ and
$\mathfrak{t}\in\{\mathfrak{f},\mathfrak{w}\}$ if $\b=\infty$.
As a shorthand we write
\[
\mu^{(\mathfrak{w})}=\left\{\begin{array}{ll}
\mu^{(\mathfrak{w},\mathfrak{w})}_\infty & \mbox{if } \b=\infty,\\
\mu^{(\mathfrak{w},\mathfrak{p})}_\b & \mbox{if } \b<\infty,
\end{array}\right.\qquad
\mu^{(\mathfrak{f})}=\left\{\begin{array}{ll}
\mu^{(\mathfrak{f},\mathfrak{f})}_\infty & \mbox{if } \b=\infty,\\
\mu^{(\mathfrak{f},\mathfrak{p})}_\b & \mbox{if } \b<\infty,
\end{array}\right.
\]
and $\el\cdot\rangle^{(\mathfrak{w})}$, $\el\cdot\rangle^{(\mathfrak{f})}$ for
the corresponding expectation operators.
For simplicity of presentation we focus on
the case $\b=\infty$, similar results and
constructions hold for the case $\b<\infty$.
To state and prove our mixing results we need
to be precise about the topological set-up.
We define a metric $d$ on $\S$ as follows.
Firstly, for each $n\geq1$ define a `local' metric
\[
d_n(\sigma,\sigma')=\sum_{x\in\L_n}\int_{I_n} |\sigma(x,t)-\sigma'(x,t)|dt,
\qquad \sigma,\sigma'\in\S,
\]
and then extend this in a standard way by letting
\[
d(\sigma,\sigma') =\sum_{n\geq0}2^{-n}
\frac{d_n(\sigma,\sigma')}{1+d_n(\sigma,\sigma')}.
\]
Recall that a function $F:\S\to\RR$ is
\begin{list}{$\bullet$}{\leftmargin=1em}
\item \emph{uniformly continuous} if for each $\varepsilon>0$ there
is $\d>0$ such that if $d(\sigma,\sigma')<\d$ then
$|F(\sigma)-F(\sigma')|<\varepsilon$;
\item \emph{even} if $F(\sigma)=F(-\sigma)$ for all $\sigma\in\S$.
\end{list}
We will prove the following:
\begin{lemma}\label{mix_lem}
Let $0<\b\leq\infty$ and let $C_1,C_2:\S\to\RR$ be bounded,
uniformly continuous functions. Then
\[
\lim_{\|x\|\to\infty}\el C_1(\sigma) [C_2\circ\tau_x](\sigma)]\rangle^{(\mathfrak{w})}
=\el C_1(\sigma)\rangle^{(\mathfrak{w})}\el C_2(\sigma)\rangle^{(\mathfrak{w})}.
\]
If, in addition, $C_1,C_2$ are even then also
\[
\lim_{\|x\|\to\infty}\el C_1(\sigma) [C_2\circ\tau_x](\sigma)]\rangle^{(\mathfrak{f})}
=\el C_1(\sigma)\rangle^{(\mathfrak{f})}\el C_2(\sigma)\rangle^{(\mathfrak{f})}.
\]
\end{lemma}
The proof follows the strategy
in the appendix of~\cite{adcs}, and is based on
first proving the statement for functions of
the form $C(\sigma)=\sigma_A$ using the Griffiths inequality
and then extending to more general functions using the
Stone--Weierstrass theorem. However,
there are two diffculties associated with
this approach: firstly, the function $C(\sigma)=\sigma_A$
is not continuous; secondly, $\S$ is not compact.
(The locally compact version of the Stone--Weierstrass
theorem is not appropriate since the functions $C$
we want to consider do not `vanish at infinity'.)
Nonetheless, we have the following result.
Let $\cG$ denote the (real) algebra of functions
generated by the monomials of the form $\sigma_A$
for finite $A\subseteq\mathbb{K}$.
For tidier notation we drop the superscript $^{(\mathfrak{w})}$
or $^{(\mathfrak{f})}$ in the following result, which holds for
both cases.
\begin{proposition}\label{sw_prop}
Suppose $F:\S\to\RR$ is a bounded and
measurable function such that
for all $G\in\cG$
\[
\lim_{\|x\|\to\infty}\langle G(\sigma) [F\circ \tau_x](\sigma)\rangle
=\langle G(\sigma)\rangle \langle F(\sigma)\rangle.
\]
Then for all bounded,
uniformly continuous $C:\S\to\RR$ we also have
\[
\lim_{\|x\|\to\infty}\langle C(\sigma) [F\circ \tau_x](\sigma)\rangle
=\langle C(\sigma)\rangle \langle F(\sigma)\rangle.
\]
\end{proposition}
\begin{proof}
For each $\d>0$ let $\S_\d$ denote the set of functions
$\sigma\in\S$ which are constant on each interval of the form
$\{x\}\times[k\d,(k+1)\d)$ for $k\in\mathbb{Z}$. Then (by a diagonal
argument or otherwise) $\S_\d$ is compact. Define a mapping
$\S\to\S_\d$ by letting $\sigma\mapsto\sigma_\d$ where
$\sigma_\d(x,t)=\sigma(x,\d\lfloor t/\d\rfloor)$,
and for $F:\S\to\RR$ let $F_\d:\S\to\RR$ be given by
$F_\d(\sigma)=F(\sigma_\d)$. Note that if $G\in\cG$ then $G_\d\in\cG$.
Let $\mathcal{C}_\d$ denote the set of continuous functions
$\S_\d\to\RR$ and
$\cG_\d$ the set of restrictions of functions in $\cG$
to $\S_\d$. Then $\cG_\d$
is an algebra of functions in $\mathcal{C}_\d$,
and $\cG_\d$ separates
the points of $\S_\d$ (if $\sigma,\sigma'\in\S_\d$ differ at the
point $(x,k\d)$ then, by definition,
$\sigma(x,k\d)\neq\sigma'(x,k\d)$). Thus by the
Stone--Weierstrass theorem $\mathcal{C}_\d$ is the uniform closure
of $\cG_\d$, meaning that for each
bounded, uniformly continuous $C:\S\to\RR$
and each $\varepsilon>0$ there
is $G\in\cG$ such that
\[
\sup_{\sigma\in\S}|G_\d(\sigma)-C_\d(\sigma)|=
\sup_{\sigma\in\S_\d}|G_\d(\sigma)-C_\d(\sigma)|
<\varepsilon.
\]
Let $M$ be a uniform upper bound on both $|F|$ and $|C|$.
We have that
\begin{equation}\label{CMG2}
\begin{split}
| \langle C(\sigma) [F\circ\tau_x](\sigma)\rangle &
- \langle G_\d(\sigma) [F\circ\tau_x] (\sigma)\rangle |\\
&\leq M \langle |G_\d(\sigma) -C_\d(\sigma)|\rangle+M \langle |C(\sigma)
-C_\d(\sigma)|\rangle\\
&\leq M\varepsilon + M \langle |C(\sigma) -C_\d(\sigma)|\rangle.
\end{split}
\end{equation}
For $\eta>0$ sufficiently small,
\[\begin{split}
|C(\sigma) -C_\d(\sigma)|&=
|C(\sigma) -C(\sigma_\d)|\hbox{\rm 1\kern-.27em I}\{d(\sigma,\sigma_\d)<\eta\}\\
&\quad+|C(\sigma) -C_\d(\sigma)|\hbox{\rm 1\kern-.27em I}\{d(\sigma,\sigma_\d)\geq\eta\}\\
&\leq \varepsilon + 2M \hbox{\rm 1\kern-.27em I}\{d(\sigma,\sigma_\d)\geq\eta\}.
\end{split}\]
Thus
$\el |C(\sigma) -C_\d(\sigma)|\rangle \leq
\varepsilon+2M \mu(d(\sigma,\sigma_\d)\geq\eta)$,
and the last probability converges to 0
as $\d\downarrow 0$ (for example along a sequence
of the form $\d=2^{-m}$).
Hence~\eqref{CMG2} can be made arbitrarily small, uniformly in $x$.
The same bound applies to
$|\el C(\sigma)\rangle\el F(\sigma)\rangle - \el G_\d(\sigma)\rangle\el F(\sigma)\rangle |$.
Since $G_\d\in\cG$, the result follows.
\end{proof}
\begin{remark}\label{sw_rk}
Proposition~\ref{sw_prop} holds also if we assume in addition that
$F$, $G$ and $C$ are even functions.
To prove this in detail one may
pass to the quotient space $\S/\!\!\sim$, where the equivalence relation
$\sim$ consists of all pairs $\{\sigma,-\sigma\}$ for $\sigma\in\S$.
An even function on $\S$ may be identified
with
a function on $\S/\!\!\sim$ and this identifies continuous functions with
continuous functions. (This uses the fact that the mapping
$\sigma\mapsto-\sigma$ is an isometry and~\cite[Lemma~3.3.6]{burago}.)
The subspace $\S_\d/\!\!\sim$ is compact and the even functions
in $\cG$ separate the points of $\S_\d/\!\!\sim$, so we may apply the
Stone--Weierstrass theorem in the same way as in
Proposition~\ref{sw_prop}. The remaining estimates are the
same.
\end{remark}
\begin{proof}[Proof of Lemma~\ref{mix_lem}]
We allow ourselves to be rather brief and omit some details.
For the boundary condition $\mathfrak{w}$
it suffices to show that for all finite sets $A,B\subseteq\mathbb{K}$,
\begin{equation}\label{corr_mix_eq}
\el\sigma_A\sigma_{B+x}\rangle^{(\mathfrak{w})}\to \el\sigma_A\rangle^{(\mathfrak{w})}\el\sigma_B\rangle^{(\mathfrak{w})}
\mbox{ as }\|x\|\to\infty,
\end{equation}
by Proposition~\ref{sw_prop} and linearity.
For the boundary condition $\mathfrak{f}$
we need to show that~\eqref{corr_mix_eq}
holds (with $\mathfrak{w}$ replaced by $\mathfrak{f}$) when $A$ and $B$
are sets of even size, by Remark~\ref{sw_rk}.
Fix $N_0<N$ and $r_0<r$
large enough that $A,B\subseteq K(N_0,r_0)$,
and write $K$ for $K(N,r)$ and $K_0$ for $K(N_0,r_0)$.
We begin by showing that for each bounded,
measurable function $h:K_0\to[0,\infty)$ we have
\begin{multline}\label{exp_mix_eq}
\lim_{\|x\|\to\infty}\Big\langle
\sigma_A\exp\Big(\sum_{y\in\L_{N_0}}\int_{I_{r_0}}h(y,t)\sigma(y+x,t)dt\Big)\Big\rangle^{(\mathfrak{w})}
\\=\langle\sigma_A\rangle^{(\mathfrak{w})}
\Big\langle\exp\Big(\sum_{y\in\L_{N_0}}\int_{I_{r_0}}h(y,t)\sigma(y,t)dt\Big)\Big\rangle^{(\mathfrak{w})}.
\end{multline}
To go from~\eqref{exp_mix_eq} to~\eqref{corr_mix_eq}
one expands the exponentials as a power series.
Using the fact that~\eqref{exp_mix_eq} holds for arbitrary
$h$ and that correlation functions of the form
$\el\sigma(y_1,t_1)\dotsb\sigma(y_k,t_k)\rangle^{(\mathfrak{w})}$
are continuous in $t_1,\dotsc,t_k$ one may
deduce pointwise convergence of the
form~\eqref{corr_mix_eq} from the corresponding convergence
of repeated sums and integrals over $y_1,\dotsc,y_k$
and $t_1,\dotsc,t_k$.
We now show~\eqref{exp_mix_eq}.
Write $K(x)=\L_{\|x\|-N_0}\times I_{r_0+\|x\|}$ where
$x\in\mathbb{Z}^d$ is fixed with
$\|x\|$ large enough that $A\subseteq K(x)$.
Let $N,r$ be large enough that $K_0+x\subseteq K$.
Write
\[
h(\sigma)=\sum_{y\in\L_{N_0}}\int_{I_{r_0}}h(y,t)\sigma(y,t)dt
\]
and let $\el\cdot\rangle_{K;h\circ\tau_x}^\mathfrak{w}$ denote the
wired space--time Ising measure
defined as in~\eqref{strn_eq}--\eqref{stpf_eq}
but with the additional term
\begin{equation}\label{h-bar}
h(\tau_x(\sigma))=
\sum_{y\in\L_{N_0}}\int_{I_{r_0}} h(y,t)\sigma(y+x,t)dt
\end{equation}
in the exponential.
Using the shorthand
$\el\cdot\rangle_K^\mathfrak{w}$ for $\el\cdot\rangle_{N,r}^{\mathfrak{w},\mathfrak{w}}$ we have
\begin{equation}\label{KJ_eq}
\big\langle
\sigma_A\exp\big(h(\tau_x(\sigma))
\big)\big\rangle_{K}^\mathfrak{w}
=\langle\sigma_A\rangle_{K;h\circ\tau_x}^\mathfrak{w}
\big\langle\exp\big( h(\tau_x(\sigma))
\big)\big\rangle_{K}^\mathfrak{w}.
\end{equation}
The Griffiths inequality
(see~\cite[Lemma~2.2.20]{bjo_phd} for a proof for the present model)
implies that the correlation
$\el\sigma_A\rangle_{K;h\circ\tau_x}^\mathfrak{w}$
is increasing when viewed as a function of $h$
(under pointwise ordering of $h$). Comparison with the
case $h\equiv 0$ gives
\begin{equation*}
\langle\sigma_A\rangle_{K; h\circ\tau_x}^\mathfrak{w}\geq
\langle\sigma_A\rangle_{K}^\mathfrak{w}.
\end{equation*}
If we let $h(y,t)\to\infty$ for all $(y,t)\in K_0$
then $\langle\cdot\rangle_{K; h\circ\tau_x}^\mathfrak{w}$
converges to a state corresponding
to `wiring' the region $K_0+x$, and we deduce that
\begin{equation*}
\langle\sigma_A\rangle_{K; h\circ\tau_x}^\mathfrak{w}\leq
\langle\sigma_A\rangle_{K(x)}^\mathfrak{w}
\end{equation*}
(cf.\ \cite[Lemma~2.2.22]{bjo_phd}).
Letting $N,r\to\infty$ and applying translation-invariance
we obtain
\begin{equation}\label{Kx_eq}
\begin{split}
\langle\sigma_A\rangle^{(\mathfrak{w})}
\big\langle\exp\big(h(\sigma) \big)\big\rangle^{(\mathfrak{w})}
&\leq \big\langle
\sigma_A\exp\big(h(\tau_x(\sigma))
\big)\big\rangle^{(\mathfrak{w})}
\\&\leq \langle\sigma_A\rangle_{K(x)}^\mathfrak{w}
\big\langle\exp\big(
h(\sigma) \big)\big\rangle^{(\mathfrak{w})}
\end{split}
\end{equation}
Letting $\|x\|\to\infty$ we have $K(x)\uparrow\mathbb{K}$
and hence~\eqref{exp_mix_eq} follows.
For the case of boundary condition
$\mathfrak{f}$ let $J:F(N_0,r_0)\to[-\l,0]$ be measurable
and $q:K_0\to[0,\infty)$ be bounded and measurable.
Write
\[
J(\sigma)=\sum_{yz\in\mathcal{E}_{N_0}}\int_{I_{r_0}}
J(yz,t)\sigma(y,t)\sigma(z,t)dt,
\]
and (recalling the process $D$ of
discontinuities of $\sigma$)
\[
q(\sigma)=\sum_{(y,t)\in D\cap K_0} q(y,t).
\]
Note that $q(\sigma)$ is a function of $D$ only, and we
may therefore write $q(D)$ in place of $q(\sigma)$. With this notation
we have
$q(\tau_x(\sigma))= q(\tau_{-x}(D))$.
Let $\el\cdot\rangle_{K; (J,q)\circ\tau_x}^\mathfrak{f}$ denote the measure
defined as in~\eqref{strn_eq}--\eqref{stpf_eq}
but with the additional term
$J(\tau_x(\sigma))+q(\tau_x(\sigma))$
in the exponential. We have that
\begin{multline}\label{Jq_eq}
\big\langle
\sigma_A\exp\big(J(\tau_x(\sigma))+q(\tau_x(\sigma))
\big)\big\rangle_{K}^\mathfrak{f}\\
=\langle\sigma_A\rangle_{K;(J,q)\circ\tau_x}^\mathfrak{f}
\big\langle\exp\big( J(\tau_x(\sigma))+q(\tau_x(\sigma))
\big)\big\rangle_{K}^\mathfrak{f}.
\end{multline}
By standard properties of Poisson processes,
$\el\cdot\rangle_{K; (J,q)\circ\tau_x}^\mathfrak{f}$
may alternatively be obtained by first modifying the intensity
of $D$ under the a-priori measure $E_{N,r}$ from the constant
intensity $\d$ to the variable intensity
\[
\d(y,t)=\d\exp(q(y-x,t)\hbox{\rm 1\kern-.27em I}\{(y,t)\in K_0+x\}),
\]
and then having only the additional term $J(\tau_x(\sigma))$
in the exponential.
Thus the correlation $\langle\sigma_A\rangle_{K;(J,q)\circ\tau_x}^\mathfrak{f}$
is increasing in $J$ and decreasing in $q$,
and comparison with the cases $J\equiv-\l$,
$q\equiv\infty$, respectively $J\equiv0$, $q\equiv0$,
gives
\[
\langle\sigma_A\rangle_{K(x)}^\mathfrak{f}\leq
\langle\sigma_A\rangle_{K;(J,q)\circ\tau_x}^\mathfrak{f}\leq
\langle\sigma_A\rangle_{K}^\mathfrak{f}.
\]
Similarly to~\eqref{exp_mix_eq} we deduce that
\[
\lim_{\|x\|\to\infty}\big\langle
\sigma_A\exp\big(J(\tau_x(\sigma))+q(\tau_x(\sigma))\big)
\big\rangle^{(\mathfrak{f})}=\langle\sigma_A\rangle^{(\mathfrak{f})}
\big\langle
\exp\big(J(\sigma)+q(\sigma)\big)
\big\rangle^{(\mathfrak{f})}.
\]
Expanding the exponential
$\exp\big(J(\sigma)\big)$ as for~\eqref{corr_mix_eq}
we deduce that for all finite $B\subseteq\mathbb{F}$,
\begin{equation}\label{q-lim}
\el\sigma_A\sigma_{B+x}\exp\big(q(\tau_x(\sigma))\big)\rangle^{(\mathfrak{f})}
\to \el\sigma_A\rangle^{(\mathfrak{f})}\el\sigma_B\exp\big(q(\sigma)\big)\rangle^{(\mathfrak{f})}
\mbox{ as }\|x\|\to\infty.
\end{equation}
Let $x_1,\dotsc,x_n\in\mathbb{Z}^d$ and let
$s_1<t_1,\dotsc,s_n<t_n$ be real numbers
such that all points of the form $(x_j,s_j)$
or $(x_j,t_j)$ are distinct, and let
$B'$ be the set of these points. (Thus $B'\cap (\{x\}\times\RR)$
has even size for all $x\in\mathbb{Z}^d$.)
One may deduce from~\eqref{q-lim} that for any such set $B'\subseteq\mathbb{K}$ we
have that
\begin{equation}\label{q-lim-2}
\el\sigma_A\sigma_{B+x}\sigma_{B'+x}\rangle^{(\mathfrak{f})}
\to \el\sigma_A\rangle^{(\mathfrak{f})}\el\sigma_B\sigma_{B'}\rangle^{(\mathfrak{f})}
\mbox{ as }\|x\|\to\infty.
\end{equation}
This proves the claim of the lemma for the boundary condition $\mathfrak{f}$
since for any set $B\subseteq\mathbb{K}$ of even size one may write
$\sigma_B=\sigma_{B'}\sigma_{B''}$ for some set $B'$ as above, and some
finite $B''\subseteq\mathbb{F}$.
One way to see that~\eqref{q-lim} implies~\eqref{q-lim-2} is as
follows (we give only a sketch).
One may see~\eqref{q-lim} as a result about convergence of
the Laplace functionals of the point processes
$\tau_{x}^{-1}(D)\cap K_0$
with certain `skewed' distributions.
Using Theorem~11.1.VI and Proposition~11.1.VII
of~\cite{dvj} it follows that
\[
\el\sigma_A\sigma_{B+x}\hbox{\rm 1\kern-.27em I}\{\tau^{-1}_{x}(D)\in\mathcal{C}\}\rangle^{(\mathfrak{f})}
\to \el\sigma_A\rangle^{(\mathfrak{f})}\el\sigma_B\hbox{\rm 1\kern-.27em I}\{D\in\mathcal{C}\}\rangle^{(\mathfrak{f})}
\]
for each `stochastic continuity set' $\mathcal{C}$. These sets include the
events
\[
\mathcal{C}_j=\{|D\cap (\{x_j\}\times[s_j,t_j))|\mbox{ is even}\},
\]
and using the identity
$\sigma(x_j,s_j)\sigma(x_j,t_j)=2\hbox{\rm 1\kern-.27em I}_{\mathcal{C}_j}-1$
one may write $\sigma_{B'}$ as a linear combination of terms of the form
$\hbox{\rm 1\kern-.27em I}\{\bigcap_{j\in J}\mathcal{C}_j\}$ for $J\subseteq\{1,\dotsc,n\}$.
We deduce~\eqref{q-lim-2} by linearity.
\end{proof}
\subsection*{Acknowledgement}
The author thanks Geoffrey Grimmett for drawing his attention
to the article~\cite{adcs}, and the two anonymous referees for their
helpful comments, corrections and suggestions.
|
\section{Introduction}
Dirac semimetals in three spatial dimensions have gapless three-dimensional (3D) linear dispersions, i.e. 3D Dirac cones, in the bulk.
They can be regarded as a 3D analog of graphene.
After the theoretical predictions had been made \cite{Young2012,Wang2012,Wang2013}, the Dirac semimetals such as Na$_3$Bi and Cd$_3$As$_2$ were experimentally discovered recently \cite{Liu2014,Neupane2014,Borisenko2014,Cheng2014,He2014}.
These Dirac semimetals possess two Dirac nodes which are protected by crystalline symmetry.
One of the important meanings of the realization of Dirac semimetals is on the point that they can lead to various topological phases, since they lie next to various topological phases in the phase diagrams.
In 3D topological insulators, the bulk energy gap closing is required to make the system turn into normal band insulators \cite{Fu2007,Fu2007a,Zhang2009}.
At these transition points, Dirac semimetals can be realized.
Experimentally, such a continuous transition is observed in the solid-solution system TlBi(S$_{1-x}$Se$_x$)$_2$ \cite{Sato2011}.
Further, Weyl semimetals can be realized when time-reversal or inversion symmetry breaking occurs in Dirac semimetals \cite{Volovik2003,Volovik2007,Murakami2007,Burkov2011,Kurebayashi2014}.
Regardless of intensive searches, Weyl semimetal phases have not been experimentally observed so far.
Hence, it is expected that recent experimental realization of the Dirac semimetals also gives rise to the realization of Weyl semimetal phases.
Stability of topological phases against electron correlation is one of attractive themes.
It has been shown that strong short-range interactions break 2D topological insulator phases \cite{Rachel2010,Varney2010,Hohenadler2011,Yu2011,Yamaji2011,Zheng2011,Hohenadler2012,Yoshida2012,Tada2012,Hohenadler2013,Liu2013}, 3D topological insulator phases \cite{Pesin2010,Sekine2013}, and Weyl semimetal phases \cite{Wei2012,Sekine2013,Maciejko2013}.
On the other hand, recent studies have suggested that these topological phases can survive strong $1/r$ long-range Coulomb interactions \cite{Sekine2013a,Araki2013,Sekine2013b}.
What about in Dirac semimetals?
In Dirac semimetals, the effects of long-range interactions are expected to be important, since the screening effect is considered to be weak due to the vanishing density of states near the Fermi level.
Effects of long-range Coulomb interactions in graphene have been studied widely \cite{Kotov2012}.
Monolayer graphene on a substrate with sufficiently small dielectric constant has been predicted theoretically to be insulating (i.e. Dirac fermions become massive) due to strong $1/r$ Coulomb interactions \cite{Khveshchenko2001,Khveshchenko2004,Khveshchenko2007,Hands2008,Drut2009,Drut2009a,Araki2010,Armour2011,Giedt2011,Buividovich2012}.
However, as the number of layers is increased, it has been found that the semimetal phase survives strong $1/r$ Coulomb interactions \cite{Khveshchenko2001,Khveshchenko2004,Hands2008,Drut2009,Son2007}.
As a powerful method which enables us to treat strong $1/r$ Coulomb interactions properly, the U(1) lattice gauge theory has been applied to discuss the semimetal-insulator transition in graphene \cite{Drut2009,Drut2009a,Araki2010,Armour2011,Giedt2011,Buividovich2012}.
In this theory, the value of the chiral condensate is used as the order parameter for the transition.
It should be noted that the value obtained in an analytical calculation, the strong coupling expansion of the lattice gauge theory \cite{Araki2010}, and the value obtained in a numerical calculation \cite{Drut2009,Drut2009a} are in good agreement in the strong coupling region.
In this paper, we focus on the effects of strong $1/r$ long-range Coulomb interactions in multinode Dirac semimetals.
Due to the vanishing density of states near the Fermi level, the screening effect is considered to be weak in Dirac fermion systems.
Then it is expected that long-range interactions become important.
Based on the U(1) lattice gauge theory, we introduce two effective lattice models and take into account $1/r$ long-range Coulomb interactions between the bulk electrons.
Further we take into account the out-of-plane Fermi velocity anisotropy of the Dirac cones, since it is not small in experimentally observed Dirac semimetals \cite{Liu2014,Neupane2014}.
With the use of the strong coupling expansion of the lattice gauge theories and the mean-field approximation, we analyze the system from the strong coupling limit.
The value of the chiral condensate, which is equivalent to the dynamically generated mass of Dirac fermions, serves as the order parameter for the semimetal-insulator transition.
\section{Model}
Let us start from the effective continuum model for correlated $N$-node Dirac semimetals.
The model we consider is the (3+1)D four-component massless Dirac fermions of $N$ flavors interacting with the electromagnetic [U(1) gauge] field.
Compared to the usual quantum electrodynamics, our model is characterized by the Fermi velocity of Dirac fermions $v_{\rm F}$ which is much smaller than the speed of light $c$.
Due to this nature, the interactions via the vector potential (spatial components of the electromagnetic field) is suppressed by the factor $v_{\rm F}/c\sim 10^{-3}$.
Then the Euclidean action of the system can be written as
\begin{align}
\begin{split}
S=&\int d^4x \sum_{f=1}^N\bar{\psi}_f(x) \left[\gamma_0(\partial_0+iA_0)+\xi_j\gamma_j\partial_j\right]\psi_f(x)\\
&+\frac{\beta}{2}\int d^4x (\partial_i A_0)^2,\label{continuum-action}
\end{split}
\end{align}
where $\psi_f(x)$ is a four-component spinor with $f$ denoting the flavor of Dirac fermions, $\gamma_\mu$ ($\mu=0,1,2,3$) are the $4\times4$ gamma matrices which satisfy the Clifford algebra $\{\gamma_\mu,\gamma_\nu\}=2\delta_{\mu\nu}$, and $A_0$ is the scalar potential.
Here we have introduced parameters for the Fermi velocity anisotropy $\xi_j$ with $\xi_1=\xi_2=1$ and $\xi_3=v_{\rm F\perp}/v_{\rm F \parallel}$.
Note that we have rescaled variables as $v_{\rm F\parallel}x_0\rightarrow x_0$, $A_0/v_{\rm F\parallel}\rightarrow A_0$ in Eq. (\ref{continuum-action}).
A parameter $\beta$, which represents the effective strength of the $1/r$ Coulomb interactions, is given by
\begin{align}
\begin{split}
\beta=\frac{v_{\rm F\parallel}\epsilon}{e^2}=\frac{v_{\rm F\parallel}\epsilon}{4\pi c \alpha},
\end{split}
\end{align}
where $e$ is the electric charge, $\epsilon$ is the dielectric constant of the system, and $\alpha(\simeq 1/137)$ is the fine-structure constant.
The smallness of the Fermi velocity makes the Coulomb interactions effectively strong.
$\beta=0$ corresponds to the strong coupling limit.
In this study we consider the case of $\beta\ll 1$, i.e. the case of small dielectric constant.
In the following, we introduce two specific effective lattice models for $N$-node Dirac semimetals with $N=4$ and $N=16$.
We take advantage of the so-called ``fermion doubling problem'' which occurs when considering Dirac fermions on lattices.
It is known that the fermion doublers can emerge in the cases where lattice fermions possess chiral symmetry, which has been proved by the Nielsen-Ninomiya theorem \cite{NN-theorem}.
{\it The $N=16$ Dirac Semimetal.---}
First we consider a (3+1)D $N=16$ Dirac semimetal interacting via $1/r$ Coulomb interactions on a lattice.
As the noninteracting action, we adopt the doubled lattice fermions (in the chiral limit) which reproduce the four-component massless Dirac fermions of 16 flavors in the continuum limit \cite{Wilson1974}.
The Euclidean action of the system is given by $S^{(N=16)}=S^{(N=16)}_F+S_{G}$. The fermionic part $S^{(N=16)}_F$ is written as
\begin{align}
\begin{split}
S^{(N=16)}_F=&\frac{1}{2}\sum_{n}\left[\bar{\psi}_n\gamma_0 U_{n,0}\psi_{n+\hat{0}} - \bar{\psi}_{n+\hat{0}}\gamma_0 U^\dag_{n,0}\psi_n\right]\\
&+\frac{1}{2}\sum_{n,j}\xi_j\left[\bar{\psi}_n\gamma_j\psi_{n+\hat{j}} - \bar{\psi}_{n+\hat{j}}\gamma_j\psi_n\right],
\end{split}\label{S_Wilson}
\end{align}
where $\psi_n$ is a four-component spinor.
This action is understood as the naively discretized action of the four-component Dirac fermions of single flavor.
The U(1) gauge part $S_G$ is written as
\begin{align}
\begin{split}
S_G=\beta\sum_n\sum_{\mu>\nu}\left[1-\frac{1}{2}\left(U_{n,\mu}U_{n+\hat{\mu},\nu}U^\dag_{n+\hat{\nu},\mu}U^\dag_{n,\nu}+{\rm H.c.}\right)\right].
\end{split}\label{S_G}
\end{align}
Here $\hat{\mu}$ ($\mu=0,1,2,3$) denotes the unit vector along the $\mu$ direction, and
$n=(n_0,n_1,n_2,n_3)$ is a lattice site on a four-dimensional isotropic lattice.
The U(1) gauge link variables $U_{n,\mu}$ are given by $U_{n,0}=e^{iA_0(n)}\equiv e^{i\theta_n}$ ($-\pi\le\theta_n\le\pi$) and $U_{n,j}=1$.
{\it The $N=4$ Dirac Semimetal.---}
Next we consider a (3+1)D $N=4$ Dirac semimetal interacting via $1/r$ Coulomb interactions on a lattice.
As the noninteracting action, we adopt the staggered fermions (in the chiral limit) which reproduce the four-component massless Dirac fermions of $4$ flavors in the continuum limit \cite{Kogut1975,Susskind1977}.
The Euclidean action of the system is given by $S^{(N=4)}=S^{(N=4)}_F+S_{G}$. The fermionic part $S^{(N=4)}_F$ is written as
\begin{align}
\begin{split}
S^{(N=4)}_F&=\frac{1}{2}\sum_{n}\eta_{n,0}\left[\bar{\chi}_nU_{n,0}\chi_{n+\hat{0}} - \bar{\chi}_{n+\hat{0}} U^\dag_{n,0}\chi_n\right]\\
&\quad+\frac{1}{2}\sum_{n,j}\xi_j\eta_{n,j}\left[\bar{\chi}_n\chi_{n+\hat{j}} - \bar{\chi}_{n+\hat{j}}\chi_n\right],
\end{split}\label{S_staggered}
\end{align}
where $\chi_n$ is a single-component spinor, $\eta_{n,0}=1$, $\eta_{n,1}=(-1)^{n_0}$, $\eta_{n,2}=(-1)^{n_0+n_1}$, and $\eta_{n,3}=(-1)^{n_0+n_1+n_2}$.
The gauge part $S_G$ is the same as Eq. (\ref{S_G}).
The action (\ref{S_staggered}) can be understood as an action obtained by doing the spin diagonalization (the Kawamoto-Smit transformation) \cite{Kawamoto1981} to $\psi_n$ in the action (\ref{S_Wilson}) as
\begin{align}
\begin{split}
\psi_n=T_n\xi_n, \ \ \ \ \ \bar{\psi}_n=\bar{\xi}_n T^\dagger_n
\end{split}
\end{align}
with $T_n=(\gamma_0)^{n_0}(\gamma_1)^{n_1}(\gamma_2)^{n_2}(\gamma_3)^{n_3}$ and $\xi_n\equiv[\chi_n^1,\chi_n^2,\chi_n^3,\chi_n^4]^T$, and then by retaining one of the four components in $\xi_n$.
However, to be precise, the action of staggered fermions after recovering the spinor structure does not coincide with that of Wilson fermions.
This is known as the taste breaking of staggered fermions.
The (2+1)D staggered fermions have been used as an effective lattice model for graphene \cite{Drut2009,Drut2009a,Araki2010,Armour2011,Giedt2011,Buividovich2012}, since they reproduce the four-component massless Dirac fermions of 2 flavors in the continuum limit \cite{Burden1987}.
\section{Strong coupling expansion}
Let us derive the effective actions in the strong coupling limit ($\beta=0$). We can derive the effective action $S_{\rm eff}$ by integrating out the U(1) gauge link variable $U_{0,n}$ in the partition function $Z$ up to the arbitrary order in $\beta$ as follows:
\begin{align}
\begin{split}
Z^{(N=16)}=\int \mathcal{D}[\psi,\bar{\psi},U_0]e^{-S^{(N=16)}}=\int \mathcal{D}[\psi,\bar{\psi}]e^{-S^{(N=16)}_{\rm eff}}.\label{partition-function}
\end{split}
\end{align}
Here we have written down the case of the $N=16$ Dirac semimetal explicitly.
The same method can be applied to the case of $N=4$ by replacing $\psi$ to $\chi$.
In the strong coupling limit, $S_G$ vanishes and thus $U_{n,0}$ is contained only in $S^{(N=16)}_F$ and $S^{(N=4)}_F$.
Then the integral $\int \mathcal{D}U_0e^{-S^{(N=16)}_F}$ is performed as
\begin{align}
\begin{split}
&\prod_n\int_{-\pi}^\pi\frac{d\theta_n}{2\pi}\exp\left\{\frac{1}{2}\left[\bar{\psi}_n\gamma_0 U_{n,0}\psi_{n+\hat{0}} - \bar{\psi}_{n+\hat{0}}\gamma_0 U^\dag_{n,0}\psi_n\right]\right\}\\
&=\prod_n\left[1-\frac{1}{4}\bar{\psi}_n\gamma_0 \psi_{n+\hat{0}}\bar{\psi}_{n+\hat{0}}\gamma_0 \psi_n+\cdots\right]\\
&\approx e^{\frac{1}{4}\sum_n\mathrm{tr}\left[\gamma_0^T\bar{\psi}_n\psi_n\gamma_0^T\bar{\psi}_{n+\hat{0}}\psi_{n+\hat{0}}\right]},
\end{split}\label{U0-integration}
\end{align}
where we have used the fact that the Grassmann variables $\psi_\alpha$ and $\bar{\psi}_\alpha$ satisfy $\psi_\alpha^2=\bar{\psi}_\alpha^2=0$ with $\alpha$ denoting the component of the spinors.
In the second line, we have neglected the terms which consist of $8,12$ and $16$ different Grassmann variables.
As is mentioned in Sec. \ref{Results}, their contributions appear in higher orders of the order parameter, and do not affect the discussion on the semimetal-insulator transition in this model.
Further in the last line, we have rewritten the exponent as
\begin{align}
\begin{split}
\bar{\psi}_{n,\alpha}(\gamma_0)_{\alpha\beta} \psi_{n+\hat{0},\beta}\bar{\psi}_{n+\hat{0},\gamma}(\gamma_0)_{\gamma\delta} \psi_{n,\delta}\\
= -\mathrm{tr}\left[\gamma_0^T\bar{\psi}_n\psi_n\gamma_0^T\bar{\psi}_{n+\hat{0}}\psi_{n+\hat{0}}\right].
\end{split}
\end{align}
The subscripts $\alpha$ and $\beta$ denote the component of the spinors, and the superscript $T$ denotes the transpose of a matrix.
In the general cases of SU($N_c$) gauge field ($N_c\ge1$), we can perform the integration with respect to the gauge link variables $U$ in Eq. (\ref{partition-function}) by using the SU($N_c$) group integral formulas: $\int dU 1=1$, $\int dU U_{ab}=0$, $\int dU U_{ab}U_{cd}^\dagger=\delta_{ad}\delta_{bc}/N_c$, and so on.
Finally we obtain the effective action of the $N=16$ Dirac semimetal in the strong coupling limit given by
\begin{align}
\begin{split}
S^{(N=16)}_{{\rm eff}}&=\frac{1}{2}\sum_{n,j}\xi_j\left[\bar{\psi}_n\gamma_j\psi_{n+\hat{j}} - \bar{\psi}_{n+\hat{j}}\gamma_j\psi_n\right]\\
&\quad-\frac{1}{4}\sum_n\mathrm{tr}\left[\gamma_0^T\bar{\psi}_n\psi_n\gamma_0^T\bar{\psi}_{n+\hat{0}}\psi_{n+\hat{0}}\right].\label{Effective-action1}
\end{split}
\end{align}
From this equation, we see that the electron-electron interactions in the strong coupling limit is spatially on-site interaction but not in the (imaginary) time.
In the $N=4$ case, $\chi$ is a single-component Grassmann variable.
Therefore, due to the nature of Grassmann variables $\chi^2=\bar{\chi}^2=0$, the approximation done in the second line of Eq. (\ref{U0-integration}) is not needed.
Finally we obtain the effective action of the $N=4$ Dirac semimetal in the strong coupling limit as
\begin{align}
\begin{split}
S^{(N=4)}_{{\rm eff}}&=\frac{1}{2}\sum_{n,j}\xi_j\eta_{n,j}\left[\bar{\chi}_n\chi_{n+\hat{j}} - \bar{\chi}_{n+\hat{j}}\chi_n\right]\\
&\quad-\frac{1}{4}\sum_n\bar{\chi}_n\chi_n\bar{\chi}_{n+\hat{0}}\chi_{n+\hat{0}}.\label{Effective-action2}
\end{split}
\end{align}
Note that this action is exact in the strong coupling limit, although we call it ``effective action''.
\section{Free Energies in the strong coupling limit \label{FreeEnergy}}
Let us derive the free energies in the strong coupling limit at zero temperature.
To this end, we apply the extended Hubbard-Stratonovich transformation to the interaction terms.
First let us consider the case of the $N=16$ Dirac semimetal.
In this case, introducing the two complex matrix auxiliary fields $Q$ and $Q'$, $e^{\kappa\mathrm{tr}AB}$ with $\kappa>0$ and $A,B$ being matrices is deformed as follows \cite{Sekine2013a}:
\begin{align}
\begin{split}
&e^{\kappa\mathrm{tr}AB}={\rm (const.)}\times\\
&\int \mathcal{D}[Q,Q'] \exp\left\{-\kappa\left[Q_{\alpha\beta}Q'_{\alpha\beta}-A_{\alpha\beta}Q_{\beta\alpha}-B^T_{\alpha\beta}Q'_{\beta\alpha}\right]\right\},\label{EHS}
\end{split}
\end{align}
This integral is approximated by the saddle point values $Q_{\alpha\beta}=\langle B^T\rangle_{\beta\alpha}$ and $Q'_{\alpha\beta}=\langle A\rangle_{\beta\alpha}$.
In the case of the $N=4$ Dirac semimetal, we can apply Eq. (\ref{EHS}) with the subscripts removed, since there is no spinor structure in the action.
{\it Free Energy of the $N=16$ Dirac Semimetal.---}
We set $(\kappa, A, B)=(1/4, \gamma_0^T\bar{\psi}_n\psi_n, \gamma_0^T\bar{\psi}_{n+\hat{0}}\psi_{n+\hat{0}})$ to decouple the interaction term (the second term) of Eq. (\ref{Effective-action1}) to fermion bilinear form.
In this process, we need to assume the form of the $4\times 4$ matrix $\langle \bar{\psi}_n\psi_n\rangle$ by the mean-field approximation.
Recall that the purpose of this study is to discuss the semimetal-insulator transition induced by strong long-range Coulomb interactions.
Here let us consider the possible gapped phases in our model.
In the action (\ref{S_Wilson}) with $U_{n,0}=1$, only the identity matrix $\bm{1}$ and the matrix $\gamma_5$ can open energy gaps.
This is because, in the presence of these matrices, the single-particle Hamiltonian of the system is given by
\begin{align}
\begin{split}
\mathcal{H}(\bm{k})=\xi_j\alpha_j\sin k_j+m_4\alpha_4+m_5\alpha_5
\end{split}\label{MF-Hamiltonian}
\end{align}
with $\alpha_j=\gamma_0\gamma_j$, $\alpha_4=\gamma_0$ and $\alpha_5=i\gamma_0\gamma_5$, which leads to the gapped energy spectrum $E(\bm{k})=\pm\sqrt{\sum_j(\xi_j\sin k_j)^2+m_4^2+m_5^2}$.
Note that the action (\ref{S_Wilson}) possesses chiral symmetry, namely, the action is invariant under the chiral transformation $\psi_n\rightarrow e^{i\alpha\gamma_5}\psi_n$.
In such cases, as in the case of graphene \cite{Hands2008,Drut2009,Drut2009a,Araki2010,Armour2011,Giedt2011,Buividovich2012}, the identity matrix (i.e. the mass term) serves as the order parameter for the semimetal-insulator transition.
Therefore we can set $\langle\bar{\psi}_n\psi_n\rangle=-\sigma\bm{1}$.
If the value of $\sigma$ is nonzero in the strong coupling limit, then the value of $\sigma$ corresponds to $\sigma=\sqrt{m_4^2+m_5^2}$ in the energy spectrum $E(\bm{k})$.
Namely, we obtain the gapped spectrum.
We can regard the value of $\sigma$ as the dynamically generated mass of Dirac fermions.
In the lattice QCD, $\sigma$ is known as the ``chiral condensate''.
Then the terms $Q_{\alpha\beta}Q'_{\alpha\beta}$, $A_{\alpha\beta}Q_{\beta\alpha}$ and $B^T_{\alpha\beta}Q'_{\beta\alpha}$ in the integrand of Eq. (\ref{EHS}) are calculated explicitly as
\begin{align}
\begin{split}
Q_{\alpha\beta}Q'_{\alpha\beta}&=\langle B^T\rangle_{\beta\alpha}\langle A\rangle_{\beta\alpha}={\rm tr}\left[\langle B\rangle\langle A\rangle\right]=\sigma^2{\rm tr}\left[(\gamma_0^T)^2\right]\\
&=4\sigma^2,
\end{split}
\end{align}
\begin{align}
\begin{split}
A_{\alpha\beta}Q_{\beta\alpha}&=A_{\alpha\beta}\langle B^T\rangle_{\alpha\beta}={\rm tr}\left[A\langle B\rangle\right]
=-\sigma{\rm tr}\left[\gamma_0^T\bar{\psi}_n\psi_n\gamma_0^T\right]\\
&=-\sigma\bar{\psi}_n\psi_n,
\end{split}
\end{align}
\begin{align}
\begin{split}
B^T_{\alpha\beta}Q'_{\beta\alpha}&=B^T_{\alpha\beta}\langle A\rangle_{\alpha\beta}={\rm tr}\left[B\langle A\rangle\right]
=-\sigma{\rm tr}\left[\gamma_0^T\bar{\psi}_n\psi_n\gamma_0^T\right]\\
&=-\sigma\bar{\psi}_n\psi_n,
\end{split}
\end{align}
where we have used $\langle\bar{\psi}_{n+\hat{0}}\psi_{n+\hat{0}}\rangle=\langle\bar{\psi}_{n-\hat{0}}\psi_{n-\hat{0}}\rangle=\langle\bar{\psi}_n\psi_n\rangle=-\sigma\bm{1}$ and $(\gamma_0^T)^2=(\gamma_0^2)^T=\bm{1}$.
Combining these three equations, we obtain the interaction term decoupled to fermion bilinear:
\begin{align}
\begin{split}
\frac{1}{4}\sum_n\mathrm{tr}\left[\gamma_0^T\bar{\psi}_n\psi_n\gamma_0^T\bar{\psi}_{n+\hat{0}}\psi_{n+\hat{0}}\right]\sim -\frac{1}{4}\sum_{n}\left[4\sigma^2+2\sigma\bar{\psi}_n\psi_n\right].
\end{split}\label{int-term1}
\end{align}
We are now in a position to derive the free energy at zero temperature in the strong coupling limit.
Combining Eqs. (\ref{Effective-action1}) and (\ref{int-term1}), the effective action expressed by the auxiliary field $\sigma$ is given by
\begin{align}
\begin{split}
S^{(N=16)}_{\mathrm{eff}}(\sigma)=\frac{V}{T}\sigma^2+\sum_{k=-\pi}^\pi \bar{\psi}_k\mathcal{M}(\bm{k};\sigma)\psi_k \label{S_eff_SCL}
\end{split}
\end{align}
with $\mathcal{M}(\bm{k};\sigma)=\sum_j\xi_ji\gamma_j\sin k_j+\sigma/2$.
As mentioned above, the value $\sigma/2$ can be regarded as the dynamically generated mass.
Here $V$ and $T$ are the volume and temperature of the system, respectively, and we have done the Fourier transform from $n=(n_0,\bm{n})$ to $k=(k_0,\bm{k})$.
From this action, we derive the free energy at zero temperature per unit spacetime volume $\mathcal{F}^{(N=16)}(\sigma)$, according to the usual formula $\mathcal{F}^{(N=16)}=-\frac{T}{V}\ln Z^{(N=16)}$.
The partition function $Z^{(N=16)}$ is calculated by the Grassmann integral formula $Z^{(N=16)}=\int D[\psi,\bar{\psi}]e^{-\bar{\psi}\mathcal{M}\psi}=\mathrm{det}\mathcal{M}$.
The determinant of $\mathcal{M}$ is calculated by the formula $\mathrm{det}\mathcal{M}=\sqrt{\mathrm{det}(\mathcal{M}\mathcal{M}^\dagger)}$.
Finally, after a straightforward calculation, we arrive at the free energy in the strong coupling limit:
\begin{align}
\begin{split}
\mathcal{F}^{(N=16)}(\sigma)=\sigma^2-2\int_{-\pi}^{\pi}\frac{d^3k}{(2\pi)^3}\ln\left[\sum_j\xi_j^2\sin^2 k_j+\frac{1}{4}\sigma^2\right].
\end{split}\label{F1-SCL}
\end{align}
The ground state is determined by the stationary condition $d \mathcal{F}^{(N=16)}(\sigma)/d \sigma=0$.
{\it Free Energy of the $N=4$ Dirac Semimetal.---}
We set $(\kappa, A, B)=(1/4, \bar{\chi}_n\chi_n, \bar{\chi}_{n+\hat{0}}\chi_{n+\hat{0}})$ to decouple the interaction term (the second term) of Eq. (\ref{Effective-action2}) to fermion bilinear form.
Like in the $N=16$ case above, we need to assume the value of $\langle \bar{\chi}_n\chi_n\rangle$ by the mean-field approximation.
Note that the lattice action (\ref{S_staggered}) also possesses chiral symmetry, namely, the action is invariant under the chiral transformation defined by $\chi_n\rightarrow e^{i\alpha\epsilon(n)}\chi_n$, $\bar{\chi}_n\rightarrow e^{i\alpha\epsilon(n)}\bar{\chi}_n$ with $\epsilon (n)=(-1)^{n_0+n_1+n_2+n_3}$.
Hence we can set $\langle \bar{\chi}_n\chi_n\rangle=-\sigma$.
Then with this approximation, the interaction term is decoupled to fermion bilinear as
\begin{align}
\begin{split}
\frac{1}{4}\sum_n\bar{\chi}_n\chi_n\bar{\chi}_{n+\hat{0}}\chi_{n+\hat{0}}\sim -\frac{1}{4}\sum_{n}\left[\sigma^2+2\sigma\bar{\chi}_n\chi_n\right].
\end{split}\label{int-term2}
\end{align}
It is convenient to perform the Fourier transform only to the spatial directions, due to the factor $\eta_{n,j}$ in the noninteracting part of the effective action (\ref{Effective-action2}).
By introducing the eight-component spinor $\Psi_{n_0,\bm{k}}$ as
\begin{align}
\Psi_{n_0,\bm{k}}=
\begin{bmatrix}
\chi_{n_0}(k_1,k_2,k_3)\\
\chi_{n_0}(k_1-\pi,k_2,k_3)\\
\chi_{n_0}(k_1,k_2-\pi,k_3)\\
\chi_{n_0}(k_1,k_2,k_3-\pi)\\
\chi_{n_0}(k_1,k_2-\pi,k_3-\pi)\\
\chi_{n_0}(k_1-\pi,k_2,k_3-\pi)\\
\chi_{n_0}(k_1-\pi,k_2-\pi,k_3)\\
\chi_{n_0}(k_1-\pi,k_2-\pi,k_3-\pi)
\end{bmatrix},
\end{align}
and substituting Eq. (\ref{int-term2}) to Eq. (\ref{Effective-action2}), the effective action is rewritten as
\begin{align}
\begin{split}
S^{(N=4)}_{\mathrm{eff}}(\sigma)=\frac{1}{4}\frac{V}{T}\sigma^2+\sum_{n_0}\sum_{\bm{k}=0}^\pi \bar{\Psi}_{n_0,\bm{k}}^T\mathcal{V}(n_0,\bm{k};\sigma)\Psi_{n_0,\bm{k}}.
\end{split}
\end{align}
Note that the sum over the wave vector $k_j$ is from $0$ to $\pi$.
The procedure to derive the free energy at zero temperature per unit spacetime volume $\mathcal{F}^{(N=4)}(\sigma)$ is the same as the case of $N=16$.
The calculation of $\mathrm{det}\mathcal{V}$ is a little complicated but can be done analytically to be $\mathrm{det}\mathcal{V}=[\sum_j\xi_j^2\sin^2 k_j+\frac{1}{4}\sigma^2]^4$.
Finally we arrive at the free energy in the strong coupling limit:
\begin{align}
\begin{split}
\mathcal{F}^{(N=4)}(\sigma)=\frac{1}{4}\sigma^2-\frac{1}{2}\int_{-\pi}^{\pi}\frac{d^3k}{(2\pi)^3}\ln\left[\sum_j\xi_j^2\sin^2 k_j+\frac{1}{4}\sigma^2\right],
\end{split}\label{F2-SCL}
\end{align}
where we have used the fact that the integrand is an even function.
The ground state is determined by the stationary condition $d \mathcal{F}^{(N=4)}(\sigma)/d \sigma=0$.
\section{Numerical Results \label{Results}}
First we consider the result for the $N=4$ case.
The Fermi velocity anisotropy $v_{\rm F\perp}/v_{\rm F \parallel}(=\xi_3)$ dependence of the chiral condensate $\sigma$ in the strong coupling limit is shown in Fig. \ref{Fig1}.
It was found that the the value of $\sigma$ becomes zero when the ratio $v_{\rm F\perp}/v_{\rm F \parallel}$ is larger than about $0.24$, whereas the value of $\sigma$ is nonzero when $v_{\rm F\perp}/v_{\rm F \parallel}$ is smaller than about $0.24$.
As mentioned in Sec. \ref{FreeEnergy}, the value of $\sigma$, the chiral condensate, is regarded as the dynamically generated mass of Dirac fermions, and can be used as the order parameter for the semimetal-insulator transition.
Hence the result indicates that whether the system is insulating or semimetallic (i.e. gapped or gapless) in the strong coupling limit depends on the value of the Fermi velocity anisotropy.
Namely, the $N=4$ Dirac semimetals survive in the strong coupling limit when the anisotropy is weak, whereas they change to Mott insulators when the anisotropy is strong.
We see from Fig. \ref{Fig1} that the transition is of the second order.
The result, that the system becomes gapped in the strong coupling limit when the Fermi velocity anisotropy is strong (i.e. the ratio $v_{\rm F\perp}/v_{\rm F \parallel}$ is small), could be understood by the fact that in general quantum effects become stronger in lower dimensions.
In the case of monolayer graphene, theoretical studies have shown that the graphene suspended in vacuum (or the graphene on a substrate with sufficiently small dielectric constant) becomes gapped due to strong $1/r$ Coulomb interactions \cite{Khveshchenko2001,Khveshchenko2004,Khveshchenko2007,Hands2008,Drut2009,Drut2009a,Araki2010,Armour2011,Giedt2011,Buividovich2012}.
In the the $N=4$ case, the interaction term in the effective action [Eq. (\ref{Effective-action2})] describes spatially on-site interactions.
This means that our model with $v_{\rm F\perp}=0$ in the strong coupling limit is equivalent to a model for a stacked 2D system.
To be more precise, our model with $v_{\rm F\perp}=0$ in the strong coupling limit corresponds to an effective lattice model for monolayer graphene in the strong coupling limit, since the (2+1)D staggered fermions reproduce the four-component Dirac fermions of 2 flavors in the continuum limit \cite{Burden1987}.
Actually, the value of $\sigma$ when $v_{\rm F\perp}=0$ in our model, $\sigma\simeq 0.24$, is equal to the value obtained by a lattice strong coupling expansion study of monolayer graphene \cite{Araki2010}.
\begin{figure}[!t]
\centering
\includegraphics[width=1.0\columnwidth,clip]{Fig1.png}
\caption{(Color online) Fermi velocity anisotropy $v_{\rm F\perp}/v_{\rm F \parallel}$ dependence of the chiral condensate $\sigma$ for the $N=4$ case in the strong coupling limit ($\beta=0$). When $v_{\rm F\perp}/v_{\rm F \parallel}=0$, the system corresponds to monolayer graphene in the strong coupling limit.}\label{Fig1}
\end{figure}
Here we would like to mention the correctness of our value of $\sigma$ in the strong coupling limit.
As for monolayer graphene which is described by (2+1)D staggered fermions, the value of $\sigma$ obtained in a lattice strong coupling expansion study with the mean-field approximation \cite{Araki2010} is in qualitative agreement (within about 10$\%$ of difference) with the values obtained in Monte Carlo studies \cite{Drut2009,Drut2009a}.
Hence it is expected that our value of $\sigma$ for the $N=4$ case is quantitatively correct, because the mean-field approximation gives more proper results in higher dimensions in general.
Finally we consider the result for the $N=16$ case.
We see easily that $\mathcal{F}^{(N=16)}(\sigma)=4\mathcal{F}^{(N=4)}(\sigma)$.
Namely, within our calculation the values of $\sigma$ for both the $N=4$ and the $N=16$ cases, which serve as the order parameter for the semimetal-insulator transition, are equivalent in the strong coupling limit.
Here note that we have neglected the interaction terms which consist of $8,12$ and $16$ fermion fields when deriving the effective action of the $N=16$ Dirac semimetal [Eq. (\ref{Effective-action1})].
If the 8 fermion field term, $S_8\sim \bar{\psi}\psi\bar{\psi}\psi\bar{\psi}\psi\bar{\psi}\psi$, is taken into account, then we obtain $S_8\sim \sigma^4+\sigma^3\bar{\psi}\psi$ by a rough mean-field approximation.
The 12 and 16 fermion field terms can be approximated in the same way.
When $\sigma=0$ at the stationary point of the free energy $\mathcal{F}^{(N=16)}(\sigma)$, i.e. when the Fermi velocity anisotropy is weak, the effects of these higher order terms in the free energy can be neglected near $\sigma=0$.
In other words, the result that the $N=16$ Dirac semimetal survives in the strong coupling limit will not be changed even though such terms are taken into account.
However, when $\sigma\neq 0$ at the stationary point of $\mathcal{F}^{(N=16)}(\sigma)$, i.e. when the anisotropy is strong, such terms will modify the value of $\sigma$ at the stationary point.
Here note that the interaction term in the effective action [Eq. (\ref{Effective-action1})] describes spatially on-site interactions.
Namely, our model with $v_{\rm F\perp}=0$ in the strong coupling limit corresponds to a model of stacked (2+1)D four-component Dirac fermions of 8 flavors in the strong coupling limit.
It has been reported in the (2+1)D cases that the value of $\sigma$ becomes smaller as the number of Dirac fermion flavor $N^{\rm 2D}$ becomes larger and the semimetal phase with large $N^{\rm 2D}$ survives in the strong coupling limit \cite{Hands2008,Drut2009,Son2007}.
Therefore, when we take into account those higher order terms in the free energy, it is expected that the value of $\sigma$ is suppressed in the case of small $v_{\rm F\perp}/v_{\rm F \parallel}$.
To verify this prediction, further study is needed.
\section{Discussions}
Firstly, we note the relations between our models and the experimentally observed Dirac semimetals.
In the observed Dirac semimetals, there exist large out-of-plane Fermi velocity anisotropy such that $v_{\rm F\perp}/v_{\rm F \parallel}\approx 0.25$ in Na$_3$Bi \cite{Liu2014} and $v_{\rm F\perp}/v_{\rm F \parallel}\sim 0.1$ in Cd$_3$As$_2$ \cite{Neupane2014}.
Therefore we expect that our result gives some perception to realistic materials.
On the other hand, as for the number of the Dirac nodes $N$, $N$ is two in both Na$_3$Bi and Cd$_3$As$_2$.
From the results in the (2+1)D case, i.e. multilayer graphene \cite{Drut2009}, it is expected that the dynamically generated mass gap $\sigma$ in the strong coupling limit becomes larger with decreasing $N$.
However, it is difficult to show such a behavior explicitly in our study.
Hence the stability of the $N=2$ case is a remaining problem.
Secondly, let us discuss a possible global phase diagram of correlated $N$-node Dirac semimetals with $N=4$ and $N=16$.
We see from Fig. \ref{Fig1} that as the ratio $v_{\rm F\perp}/v_{\rm F \parallel}$ is increased from zero, the value of $\sigma$ becomes smaller and eventually reaches zero.
The chiral condensate $\sigma$ can be used as the order parameter for the semimetal-insulator transition.
Namely, the system is gapless in the strong coupling limit ($\beta=0$) when the ratio $v_{\rm F\perp}/v_{\rm F \parallel}$ is large, whereas the system is gapped when the ratio $v_{\rm F\perp}/v_{\rm F \parallel}$ is small.
We call the gapped phase with nonzero $\sigma$ the Mott insulator.
On the other hand, the system is obviously a Dirac semimetal in the noninteracting limit ($\beta=\infty$).
Therefore, there must exist the critical strength of the $1/r$ Coulomb interactions $\beta_c$, below which the system becomes semimetallic, i.e. the value of $\sigma$ becomes zero.
This critical value $\beta_c$ will become smaller as the value of $\sigma$ in the strong coupling limit becomes smaller.
A schematic global phase diagram for the $N=4$ case based on this analysis is shown in Fig. \ref{Fig2}.
In the case of $N=16$, as mentioned in the previous section, it is expected that the value of $\sigma$ in the strong coupling limit is suppressed when the ratio $v_{\rm F\perp}/v_{\rm F \parallel}$ is small.
Namely, it is expected that the region of the Mott insulator phase shrinks in the global phase diagram.
\begin{figure}[!t]
\centering
\includegraphics[width=1.0\columnwidth,clip]{Fig2.png}
\caption{(Color online) A possible global phase diagram of a correlated $N$-node Dirac semimetal with $N=4$.
The $\beta=0$ ($\beta=\infty$) line represents the strong coupling limit (noninteracting limit).
The Mott insulator phase is defined as the phase with nonzero value of $\sigma$.}\label{Fig2}
\end{figure}
In this study, we have focused only on the energy gap generation by strong $1/r$ Coulomb interactions, and thus the detailed information of the spinors in the low-energy effective model [Eq. (\ref{continuum-action})] is not required.
However, if we construct a low-energy effective model of some realistic material, then the spinors should be associated with the lattice structure and the spins of electrons, as in the case of graphene.
Hence it is expected that some order such as a magnetic or charge order is realized in the Mott insulator phase in Fig. \ref{Fig2}, although it is difficult in this study to identify what the order is.
This can be understood from the fact that the two possible orders in the Hamiltonian (\ref{MF-Hamiltonian}), the $\alpha_4$ order where no symmetry is broken and the $\alpha_5$ order where both time-reversal and inversion symmetries are broken, are energetically degenerate.
The result of this paper is not changed even though the $\alpha_5$ order is considered as the gapped order instead of the $\alpha_4$ order.
Thirdly, let us discuss the implications of our results to the stability of Weyl semimetals.
Weyl semimetals have gapless 3D linear dispersions in the bulk, and the effective Hamiltonians around the Weyl nodes are described by not the Dirac Hamiltonian but the Weyl Hamiltonian $\mathcal{H}_{\rm Weyl}(\bm{k})=\sum_{i=1}^3\bm{v}_i\cdot\bm{k}\sigma_i$ where $\sigma_i$ are the Pauli matrices.
At least either time-reversal or inversion symmetry breaking is required to realize a Weyl semimetal phase \cite{Volovik2003,Volovik2007,Murakami2007,Wan2011,Burkov2011a,Burkov2011,Kurebayashi2014,Halasz2012,Zyuzin2012a}.
Each Weyl node possesses chirality defined by $c={\rm sgn}[\bm{v}_1\cdot(\bm{v}_2\times\bm{v}_3)]=\pm1$. The number of the Weyl nodes with chirality $+1$ and that of chirality $-1$ must be equal in time-reversal symmetry broken Weyl semimetals.
For example, the Weyl semimetal phase predicted by a first-principles calculation in pyrochlore iridates possesses 24 Weyl nodes \cite{Wan2011}.
Here let us consider a simplified low-energy effective model for a Weyl semimetal with $2N$ nodes.
The Hamiltonian of such a system can be written as
\begin{align}
\begin{split}
H_{\rm Weyl}^{\rm eff}&=\sum_{\bm k}\sum_{f=1}^Nv_{{\rm F}\parallel}\left\{\psi^\dagger_{f+}(\bm{k})\left[\xi_ik_i\sigma_i\right]\psi_{f+}(\bm{k})\right.\\
&\quad\left.+\psi^\dagger_{f-}(\bm{k})\left[-\xi_ik_i\sigma_i\right]\psi_{f-}(\bm{k})\right\}\\
&=\sum_{\bm k}\sum_{f=1}^N\psi^\dagger_{f}(\bm{k})v_{{\rm F}\parallel}
\begin{bmatrix}
\xi_ik_i\sigma_i & 0\\
0 & -\xi_ik_i\sigma_i
\end{bmatrix}
\psi_{f}(\bm{k}),
\end{split}
\end{align}
where $\psi_f=[\psi_{f+},\psi_{f-}]^T$ with $\psi_{f\pm}$ being a two-component spinor, the subscript $\pm$ denotes the chirality of each Weyl node, and we have introduced the Fermi velocity anisotropy $\xi_i$.
By introducing the $4\times 4$ gamma matrices $\gamma_\mu$ in the chiral representation, we obtain the low-energy effective action (\ref{continuum-action}).
Namely, this indicates that in a rough approximation, the low-energy effective model of a $2N$-node Weyl semimetal is equivalent to that of a $N$-node Dirac semimetal.
Weyl semimetals have a topological property such that the energy gap opens only if the Weyl nodes with opposite chirality meet each other, since a single Weyl fermion cannot be massive by itself.
Due to this property, Weyl semimetals are expected to be more stable against strong $1/r$ Coulomb interactions than Dirac semimetals.
However, it is not easy to treat strong $1/r$ Coulomb interactions properly in multinode Weyl semimetals.
In this study, it was found that the $N$-node Dirac semimetals with $N=4$ and $N=16$ survive in the strong coupling limit.
Hence, it could be said that the $N_{\rm W}$-node Weyl semimetals with $N_{\rm W}=8$ and $N_{\rm W}=32$ also survive in the strong coupling limit when the Fermi velocity anisotropy is weak.
As for the cases of $N_{\rm W}<8$, a recent study has reported that a Weyl semimetal with $N_{\rm W}=2$ survives in this limit \cite{Sekine2013b}.
Finally, let us discuss the implications of our results to the stability of 3D topological insulators.
It is known that 3D topological insulators can be regarded as 3D Dirac fermion systems.
In the noninteracting cases, the bulk energy gap closes when the phase transition from the topological insulator phase to the normal band insulator phase occurs.
In other words, there exist Dirac point(s) in the bulk when the system is on the phase boundary between the topological insulator phase and the normal band insulator phase.
For example, the Fu-Kane-Mele model has three Dirac points \cite{Fu2007,Fu2007a}, and the effective model for Bi$_2$Se$_3$ has one Dirac point \cite{Zhang2009} on their phase boundaries.
What about in the interacting cases?
The phase transitions from the topological insulator phase to the other phases can occur without the gap closing, when accompanying the breaking of symmetry of the system such as time-reversal symmetry or inversion symmetry.
However, the gap closing is required when no symmetry is broken, as in the noninteracting cases.
From this viewpoint, our result, that the Dirac semimetals survives in the strong coupling limit, suggests that 3D topological insulator phases can be stable against strong $1/r$ Coulomb interactions.
Actually, a recent study has reported that a 3D topological insulator phase of Bi$_2$Se$_3$-type survives in the strong coupling limit when the spin-orbit interaction of the system is strong \cite{Sekine2013a}.
\section{Summary}
In summary, based on the U(1) lattice gauge theory, we have studied the stability of $N$-node Dirac semimetals in three spatial dimensions with $N=4$ and $N=16$ against strong $1/r$ long-range Coulomb interactions.
It was shown that the Dirac semimetals survive in the strong coupling limit when the Fermi velocity anisotropy is weak, whereas they change to Mott insulators when the anisotropy is strong.
This means that the three-dimensionality of the Dirac cones plays an important role in the stability.
The value of the dynamically generated mass gap at least for the $N=4$ case is expected to be quantitatively correct.
A possible global phase diagram of correlated Dirac semimetals was presented.
Dirac semimetals can lead to various topological phases by the change of parameters or symmetry breakings.
Our result, that Dirac semimetals are stable against strong $1/r$ long-range Coulomb interactions, implies the stability of other topological phases.
Namely, it is suggested that Weyl semimetals, which correspond to Dirac semimetals in a rough approximation, can survive in the strong coupling limit.
The existence of 3D topological insulator phases in the strong coupling limit is also suggested.
\begin{acknowledgments}
A.S. is supported by the JSPS Research Fellowship for Young Scientists.
This work was supported by Grant-in-Aid for Scientific Research (No. 25103703, No. 26107505 and No. 26400308) from the Ministry of Education, Culture, Sports, Science and Technology (MEXT), Japan.
\end{acknowledgments}
\nocite{*}
|
\section{Introduction}
It is well-established that the typical outcome of the collapse of a cloud is a small multiple system
(e.g. Duch\^ene
et al. 2006; Reipurth et al. 2014). In particular, at centimeter
wavelengths the studies of Reipurth et al. (2002a; 2004) have shown
that the cores of regions associated with outflow activity typically
reveal the presence of several radio continuum sources, each one
usually believed to be tracing the presence of a recently formed star.
However, these compact radio sources could also trace collisionally
ionized regions, that is, the radio equivalent of Herbig-Haro objects,
sources without an embedded star.
In the case of forming stars of low and intermediate mass, the
observed radio emission directly associated with the young star has
two possible origins. A first class of sources are those with resolved
structure at the scale of tenths of arc seconds that are usually
emitting free-free (thermal) radiation from ionized, collimated
outflows (e.g., Curiel et al. 1993). A second class of sources are young stars with
magnetospheric activity that produces detectable gyrosynchrotron
(non-thermal) emission (e.g., Dzib et al. 2013). The determination
of the characteristics of these sources can be of great help in
improving our understanding of the region studied, locating precisely
the position of young stars with either outflow or magnetospheric
activity. In particular, the detection of elongated free-free
sources, sometimes called thermal jets (Anglada 1996; Rodr\'\i guez
1997; Rodr\'\i guez et al. 1999; Rodr\'\i guez et al. 2000), locates the exciting source and the
origin and orientation of the outflow at sub-arcsec scales.
At the shortest wavelengths, e.g. 7 mm, dust emission also becomes
detectable, allowing additional high resolution imaging of
circumstellar material (Rodr\'\i guez et al. 2008a).
\subsection{The HH 92 Jet and its Driving Source}
Bally, Reipurth, \& Aspin (2002) reported the detection of a highly
collimated, low-excitation jet, HH~92, that emerges (with a position
angle of $311^\circ$) from an asymmetric infrared reflection nebula
spatially related with the source IRAS~05399-0121, an embedded
low-mass class~I protostar with a bolometric luminosity of 10
$L_\odot$. In their study, Bally et al. (2002) propose that HH~92,
together with the well-known objects HH 90/91 and HH~93 (Reipurth
1985; 1989) trace a single giant outflow lobe $\simeq$3 pc in length.
The driving source is located in a cloud clump labeled as LBS~30 (Lada
et al. 1991) with two cores, as seen in an 850~$\mu$m SCUBA image
(Figure~\ref{scubafig}) from the SCUBA Legacy Survey (Di Francesco et
al. 2008). Similar structure is seen in the 350~$\mu$m maps of
Miettinen \& Offner (2013).
However, despite several early efforts the exciting source of this
giant outflow was initially not firmly identified (Gredel, Reipurth,
\& Heathcote 1992; Davis, Mundt, \& Eisl\"oeffel 1994; Reipurth et al.
1993). With the advent of the Spitzer and Herschel missions, an
accurate position and photometry were achieved (Megeath et al. 2012;
Manoj et al. 2013). Most recently, the source has been studied in
detail at sub-mm wavelengths by Miettinen \& Offner (2013), who give
further references to the literature.
In this paper we present sensitive, high angular resolution Very Large
Array and Expanded Very Large Array observations at 8.46~GHz (3.6 cm) of the
core of the region that reveal the presence of a compact group of
radio sources. A distance of 415 pc to the L1630 cloud, where HH~92
is embedded, is adopted (Anthony-Twarog 1982).
\subsection{HH 34 and its Driving Source}
The Herbig-Haro object HH~34 was first noted by Guillermo Haro, see
Herbig (1974). It is part of a major complex of shocks, including a
highly collimated jet (Reipurth et al. 1986), and distant bow shocks
that make the complex span across almost 3~pc (Devine et al. 1997).
Proper motions and radial velocities indicate a dynamic age of the
entire complex of about 10,000~yr (Devine et al. 1997, Reipurth et al.
2002b). The driving source is not optically visible, although a
compact reflection nebula is visible at the base of the jet, and
spectroscopy of this nebula shows that the source is a rich
emission-line T~Tauri star (Reipurth et al. 1986). Imaging with NICMOS
on HST reveals the driving source in the H- and K-bands (Reipurth et
al. 2000). The source was also detected with low spatial resolution in
the radio continuum at 3.6~cm (Rodr\'\i guez \& Reipurth 1996). HH~34
is located in the northern part of the L1641 cloud, just south of the
Orion Nebula cluster, whose distance has been determined as 414$\pm$7
pc by Menten et al. (2007) and 418$\pm$6 pc by Kim et al. (2008), see
also the discussion in Muench et al. (2008). We here use the mean
distance of 416~pc.
In this paper we present a deep high angular resolution radio
continuum VLA map obtained at 7~mm.
\begin{figure}
\centering
\includegraphics[scale=0.34, angle=0]{scuba_HH92IRS.eps}
\caption{An 850~$\mu$m SCUBA map of the filamentary cloud clump LBS~30
in which the HH 92 driving source is located (white cross). The
elongated cloud clump has fragmented into two separate cores, of
which only the northwestern has begun to form stars (Miettinen \&
Offner 2013). The cross has the same dimensions (4.5$"$ $\times$
4.5$"$) as the panels in Figure~\ref{fighh92vla}. }
\label{scubafig}
\end{figure}
\begin{figure*}
\centering
\includegraphics[scale=0.60, angle=0]{HH92XALLNEW.eps}
\caption{Contour images of the 8.46 GHz (3.6 cm) continuum emission from the
core of the HH~92 outflow. (Top) Image for the 2003.03 epoch. The
contours are -4, -3, 3, 4, 5, 6, 8, 10, 12, 15, and 20 times
12.4 $\mu$Jy beam$^{-1}$, the rms noise of the image. The half power contour of
the synthesized beam ($0\rlap.{''}26 \times 0\rlap.{''}24$ with a
position angle of $-62^\circ$) is shown in the bottom left corner.
(Bottom) Image for the 2011.61 epoch. The contours are -4, -3, 3,
4, 5, 6, 8, 10, 12, 15, 20, 30, 40, 50, and 60 times 8.9 $\mu$Jy beam$^{-1}$, the
rms noise of the image. The half power contour of the synthesized
beam ($0\rlap.{''}27 \times 0\rlap.{''}23$ with a position angle of
$-45^\circ$) is shown in the bottom left corner. }
\label{fighh92vla}
\end{figure*}
\section{Observations}
\subsection{HH 92}
The Very Large Array (VLA) observations were made at 8.46 GHz (3.6 cm) in the B
(during epoch 2002 August 1) and A (during epoch 2003 June 21)
configurations. The VLA is part of the NRAO\footnote{The National
Radio Astronomy Observatory is operated by Associated Universities
Inc. under cooperative agreement with the National Science
Foundation.}. In both epochs the source J1331+305 was used as an
absolute amplitude calibrator (with an adopted flux density of 5.21
Jy) and the source J0541-056 was used as the phase calibrator (with
bootstrapped flux densities of 0.879$\pm$0.003 and 0.958$\pm$0.007 Jy
for the first and second epochs, respectively). The data were reduced
separately for each epoch using the standard VLA procedures and later
concatenated to produce images of high sensitivity and high angular
resolution. We refer to these concatenated data as having epoch
2003.03, the average epoch of the two observations.
With the purpose of studying the time evolution and kinematics of
these sources, we made Expanded Very Large Array (EVLA) observations
of the same region during 2011 August 11 in the A configuration. We
refer to these data as having epoch 2011.61, 8.58 years after the
first observations. The EVLA observations were made at 8.46 GHz (3.6 cm) using
two intermediate frequency (IF) bandwidths of 128 MHz each, separated
by 128 MHz, and containing both circular polarizations. Each IF was
split in 64 channels of 2 MHz each. For the continuum images we
averaged the central 54 channels, rejecting five channels at each end
of the bandwidth. The data reduction was made using the software
package AIPS of NRAO, following the recommendations for EVLA data
given in Appendix E of its CookBook (that can be found at
http://www.aips.nrao.edu/cook.html). The source J0137+3309 was used
as an absolute amplitude calibrator (with an adopted flux density of
3.15 Jy) and the source J0541-056 was used as the phase calibrator
(with a bootstrapped flux density of 0.699$\pm$0.003).
In Figure~\ref{fighh92vla} we show the images of the HH~92 region at
the two epochs studied. The top one is made from the 2003.03 data with
the \sl (u,v) \rm weighting parameter ROBUST (Briggs 1995) of the task IMAGR
set to -5. The bottom one is made from the 2011.61 data with the
\sl (u,v) \rm weighting parameter ROBUST of the task IMAGR set to 0. The
different weightings were needed to produce images of very similar
angular resolution, given that the 2003.03 data contains both A and B
configuration observations, while the 2011.61 data contains only A
configuration observations. These images clearly show three sources
that we label as sources VLA 1, 2, and 3. The three sources are
distributed in a region of $\sim 2{''}$ in extent (830 AU at a
distance of 415 pc). In Table 1 we list the parameters of these three
sources, taken from the images at both epochs.
\subsection{HH 34}
To obtain an image of the core of HH~34 at 43.3 GHz (7 mm) we used
two data sets from the VLA archive. The first data set was taken on 2004 October 05
in the A configuration under project AR552 and the second data set was taken on
2006 August 29 in the B configuration under project AK634. Both observations were
made with an effective bandwidth of 100 MHz. For the first epoch 0713+438 was used as flux calibrator
(with an adopted flux density of 0.29 Jy) and 0541-056 as phase calibrator, with a bootstrapped
flux density of 0.61$\pm$0.01 Jy. For the second epoch 1331+305 was used as flux calibrator
(with an adopted flux density of 1.45 Jy) and 0541-056 as phase calibrator, with a bootstrapped
flux density of 1.88$\pm$0.06 Jy. The data were calibrated following the standard VLA
procedures for high frequency observations and the two epochs were concatenated to obtain
an image with better sensitivity. In Figure 3 we show
the resulting image, made with a ROBUST weight of 5 (Briggs 1995) to emphasize sensitivity.
An extended source is clearly detected, with peak position $RA(2000) = 05^h~ 35^m~ 29\rlap.{^s}846 \pm 0\rlap.{^s}001$,
$DEC(2000) = -06^\circ~ 26'~ 58\rlap.{''}08 \pm 0\rlap.{''}02$, a total flux density of 3.5$\pm$0.4 mJy and
deconvolved dimensions
of $0\rlap.{''}39 \pm 0\rlap.{''}05 \times 0\rlap.{''}25 \pm 0\rlap.{''}04; PA = 55^\circ \pm 13^\circ$.
\begin{table*}[htbp]
\small
\setlength{\tabnotewidth}{1.8\columnwidth}
\tablecols{6}
\caption{Radio Sources at the Core of the HH~92 Outflow$^{\lowercase{a}}$}
\begin{center}
\begin{tabular}{lcccccc}\hline\hline
& &\multicolumn{2}{c}{Position$^b$} & Total Flux
& & \\
\cline{3-4}
VLA & Epoch & $\alpha$(J2000) & $\delta$(J2000) & Density (mJy) &
Deconvolved Angular Size$^c$ \\
\hline
1 & 2003.03 & 05 42 27.678 & -01 20 01.29 & 0.31$\pm$0.02
& $0\rlap.{''}23 \pm 0\rlap.{''}03 \times 0\rlap.{''}08 \pm 0\rlap.{''}03;~ +130^\circ\pm 9^\circ$ \\
1 & 2011.61 & 05 42 27.677 & -01 20 01.27 & 0.29$\pm$0.02
& $0\rlap.{''}20 \pm 0\rlap.{''}03 \times 0\rlap.{''}08 \pm 0\rlap.{''}03;~ +122^\circ\pm 10^\circ$ \\
& & & & & \\
2 & 2003.03 & 05 42 27.724 & -01 20 01.35 & 0.28$\pm$0.03
& $0\rlap.{''}36 \pm 0\rlap.{''}05 \times 0\rlap.{''}26 \pm 0\rlap.{''}06;~ +106^\circ\pm 34^\circ$ \\
2 & 2011.61 & 05 42 27.728 & -01 20 01.31 & 0.28$\pm$0.02
& $0\rlap.{''}37 \pm 0\rlap.{''}04 \times 0\rlap.{''}34 \pm 0\rlap.{''}04;~ +106^\circ\pm 12^\circ$ \\
& & & & & \\
3 & 2003.03 & 05 42 27.765 & -01 20 02.24 & 0.34$\pm$0.02
& $0\rlap.{''}17 \pm 0\rlap.{''}04 \times 0\rlap.{''}10 \pm 0\rlap.{''}04;~ +162^\circ\pm 28^\circ$ \\
3 & 2011.61 & 05 42 27.765 & -01 20 02.22 & 0.64$\pm$0.02
& $0\rlap.{''}06 \pm 0\rlap.{''}02 \times 0\rlap.{''}04 \pm 0\rlap.{''}03;~ +96^\circ\pm 45^\circ$ \\
\hline\hline
\tabnotetext{a}{Parameters taken from the images at each epoch.}
\tabnotetext{b}{Units of right
ascension are hours, minutes, and seconds
and units of declination are degrees, arcminutes, and arcseconds. Positional accuracy
is estimated to be $0\rlap.{''}01$.}
\tabnotetext{c}{Major axis $\times$ minor axis; position angle.}
\label{tab:1}
\end{tabular}
\end{center}
\end{table*}
\section{HH 92: ~~Results and Interpretation}
\subsection{Proper Motions}
The positions of the three sources at the two epochs (see Table 1) are
approximately coincident within the positional error (1$\sigma \simeq
0\rlap.{''}01$). Using the distance of 415 pc and the time separation
of 8.58 years between observations, we set a 3-$\sigma$ upper limit of
$\leq$7~km s$^{-1}$ for the proper motions of the sources. This rules
out an interpretation in terms of some of them being classic HH knots,
since such objects exhibit proper motions an order of magnitude or
more larger. Most probably, the radio sources are tracing young stars
or stationary HH knots.
\subsection{Flux Density Variations as a Function of Time}
The sources VLA 1 and VLA 2 do not exhibit flux density variations
within the observational noise of $\sim$0.02-0.03 mJy. This sets an
upper limit of $\sim$10\% to any possible flux density variations. In
contrast, source VLA 3 shows a strong increase of a factor of 2, from
0.34 to 0.64 mJy. Furthermore, the deconvolved size of this source
decreased for the second epoch, suggesting that most of the flux
density increase took place in a very compact region, perhaps due to
gyrosynchrotron activity in VLA 3.
\subsection{Morphology of the Sources}
The deconvolved angular dimensions of the sources VLA 2 and VLA 3 are
consistent, within noise, with a spherical geometry. In contrast, VLA 1 is clearly
elongated, with its major axis $\sim$3 times larger than its minor
axis. The average position angle of this source (from the two epochs
of observation) is $126^\circ \pm 7^\circ$, coincident within modulo
$180^\circ$ with the position angle of the optical jet, $311^\circ$
(Bally et al. 2002).
\subsection{Interpretation}
The resolution of the HH 92 driving source into
three radio continuum sources can be interpreted in
terms of the sources being either young stars or
representing thermal emission from HH shocks.
The radio emission could also have a non-thermal origin,
as has been observed in the knots of some jets (e.g. Carrasco-Gonz\'alez et al. 2010),
but we do not have the multifrequency data
required to test this possibility.
Recent submm observations using the SMA interferometer by Chiang et
al. (in prep.) have revealed that there is only one 850~$\mu$m source
in the field of the triple system, namely VLA~1. Neither VLA~2 nor
VLA~3 are detected. This strongly suggests that VLA~1 is the driving
source, and is consistent with the elongation along the flow axis seen
for VLA~1 (see above).
So what are VLA~2 and 3? Both lie approximately within the redshifted
outflow lobe from VLA~1, which would indicate that they are HH knots.
However, our two-epoch observations spanning 8~1/2 yr reveal no
measurable proper motions at all, and put very stringent limits on the
proper motions of less than 7~km~s$^{-1}$. Since HH flows typically
have space velocities of 100 km~s$^{-1}$ or more, such slow motions
would indicate that the motion is mostly along the line of sight.
However, this is inconsistent with the giant flow dimensions of about
3~pc, which points towards an outflow axis lying close to the plane of
the sky. It follows that if VLA~2 and 3 are shocks, then they are
stationary shocks, such as would occur if small ambient cloudlets are
interacting with a strong stellar wind or outflow (Schwartz 1978).
Possible stationary shocks with no detectable proper motions along the flow have been
found in regions of high ambient density (e.g., Mart\'\i\ et al. 1995;
Rodr\'\i guez et al. 2008b).
This interpretation gains weight because at least VLA~3 shows strong
flux-variability over the time span of the observations.
Alternatively, a possible scenario would be that the HH 92 driving source forms a small
non-hierarchical triple stellar system, and that an outflow from the active source VLA 1
runs over the two nearby stellar components of the triple system, creating shocks where
the outflow rams into the circumstellar material of these two stars. A similar case has
recently been identified in the young IRAS 16293–2422 triple system (Girart et al. 2014).
This scenario is not inconsistent with the lack of detection of VLA 2 and 3 at submm wavelengths,
because not all very young stars are detectable at sub-mm wavelengths, since the cool
circumstellar envelopes merely re-emit instantaneously the radiation from the central
protostars, and if the accretion is dormant, the re-emission from the envelopes will be weak.
The presently available observations do not allow us to distinguish
between the two scenarios in which VLA~2 and 3 are either small
cloudlets or protostellar companions.
\section{HH 34: ~~Results and Interpretation}
Our 7~mm map of the HH~34 driving source is seen in
Figure~\ref{hh34fig}. When compared to the beam, the source is clearly
extended in a NE-SW direction, but there is also an extension almost
perpendicular to the main elongation. In Figure~\ref{hh34hstfig} we
show the area of our 7~mm map shown in Figure~\ref{hh34fig} as a
square superposed on a high-resolution optical image ([SII] 6717/6731
transitions) of the HH~34 jet. It is immediately evident that the main
elongation of the source is perpendicular to the jet axis, suggesting
that what we see is dust emission from a circumstellar disk. The disk
has a total width of 0.4~arcsec, which corresponds to a radius of
$\sim$80~AU, matching well with expectations for Class~I disks. At
7~mm, both dust emission and free-free emission can be detected
(Rodr\'\i guez et al. 2008a), and so it
seems likely that the elongation we see perpendicular to the disk axis
results from the jet, which is monopolar at optical wavelengths, but
bipolar at mid-infrared wavelengths (Raga et al. 2011).
We estimate the total mass of the 7 mm source
by assuming optically thin, isothermal dust emission and a gas-to-dust ratio of 100
(Sodroski et al. 1997). The total mass opacity coefficient at 7 mm is poorly known and
values between $8 \times 10^{-3}$ cm$^2$ g$^{-1}$ (Rodr\'\i guez et al. 2007) and
$2 \times 10^{-3}$ cm$^2$ g$^{-1}$ (Isella et al. 2014) can be found in the literature.
Assuming an average value of $5 \times 10^{-3}$ cm$^2$ g$^{-1}$, the
total mass can be roughly estimated to be:
\footnotesize
$$ \Biggl[{{M} \over {M_\odot}}\Biggr] = 0.17
\Biggl[{{S_\nu} \over {mJy}}\Biggr]
\Biggl[{{T} \over {100~K}}\Biggr]^{-1}
\Biggl[{{D} \over {kpc}}\Biggr]^{2},$$
\normalsize
\noindent where $S_\nu$ is the flux density at 7 mm,
$T$ is the dust temperature and $D$ is the distance
to the source. Assuming
$T$ = 50 K and a distance of 416 pc, we estimate a mass
of $\sim0.21~M_\odot$ for the HH 34 disk.
Almost precisely $\sim0\rlap.{''}4$ north of the HH 34 source there is a weak signal,
which is possibly real since it is detected at a 4-$\sigma$ level, but which will require confirmation from
additional observations. If real, it is located 180~AU in projection
from the main source. Reipurth (2000) has suggested that all giant HH
flows are the result of dynamical interactions in and disintegration
of triple or small multiple systems. In this interpretation, the HH~34
source should be a close binary with separation of order 10~AU. No
existing data of the HH~34 source has sufficient spatial resolution to
detect such a close binary. In addition to the companion possibly
detected here, another very faint companion (``HH~34~X'') was detected
in the K-band with HST by Reipurth et al. (2000), about 5$"$ west of
the jet source. Finally, a faint variable star (``Star~B'') is located
11$"$ SE of the driving source and seen at the left edge of Fig. 4 (Reipurth et al. 1986, 2002b).
Altogether, the HH~34 driving source may have several more distant
companions.
\begin{figure}
\centering
\includegraphics[scale=0.43, angle=0]{HH34Q.PS}
\caption{A 43.3 GHz (7~mm) radio continuum map of a 1.9$"$ $\times$ 1.9$"$ area
around the HH~34 driving source. The elongation of the source is
perpendicular to the jet axis as seen in Figure~\ref{hh34hstfig} and
is likely to represent cool dust in a circumstellar disk. At the
assumed distance of 416~pc, the 0.4$"$ width of the disk corresponds
to about 160~AU. The
contours are -4, -3, 3, 4, 5, 7, and 9 times
77 $\mu$Jy beam$^{-1}$, the rms noise of the image. The half power contour of
the synthesized beam ($0\rlap.{''}14 \times 0\rlap.{''}13$ with a
position angle of $14^\circ$) is shown in the bottom left corner. }
\label{hh34fig}
\end{figure}
\begin{figure}
\includegraphics[scale=0.27, angle=0]{HH34_HST.eps}
\caption{A WFPC2 HST image of the base of the HH 34 jet obtained
through a narrow-band filter transmitting the [SII] 6717/6731
doublet lines. Tick marks are in steps of one arcsec, and the field
of view is about 12$"$ $\times$ 14$"$. The box outlines the 1.9$"$
$\times$ 1.9$"$ field shown in 7~mm radio continuum in
Figure~\ref{hh34fig}. Image from Reipurth et al. (2002b). }
\label{hh34hstfig}
\end{figure}
\section{Conclusions}
We have presented VLA and EVLA observations made at 8.46 GHz (3.6 cm) of the
core of the HH~92 outflow, made at two epochs separated by 8.58 years.
We detect a group of three compact radio sources, of which VLA~1
drives the HH~92 jet. The other two sources, VLA~2 and 3, may be
stationary shocks surrounding either small cloudlets or protostars.
Additional VLA observations obtained at 7~mm of the driving source of
the HH~34 jet have revealed a circumstellar disk perpendicular to the
jet axis and with a radius of $\sim$80~AU and mass of $\sim$0.21 $M_\odot$.
\acknowledgments LFR acknowledges the support of DGAPA, UNAM, and of
CONACyT (M\'exico). BR and HFC acknowledge support by the National
Aeronautics and Space Administration through the NASA Astrobiology
Institute under Cooperative Agreement No. NNA09DA77A issued through
the Office of Space Science. This research has made use of the SIMBAD
database, operated at CDS, Strasbourg, France.
|
\section{Introduction}
The purpose of this paper is to give necessary and sufficient conditions for
the equality of Morse and Oseledets decompositions of a continuous flow on a
flag bundle.
We consider a continuous principal bundle $Q\rightarrow X$ with group $G$,
which is assumed to be semi-simple or reductive. A continuous automorphism $%
\phi \in \mathrm{Aut}\left( Q\right) $ of $Q$ defines a discrete-time flow $%
\phi ^{n}$, $n\in \mathbb{Z}$, on $Q$. For instance $Q\rightarrow X$ could
be the bundle of frames of a $d$-dimensional vector bundle $\mathcal{V}%
\rightarrow X$ over $X$, in which case $G$ is the reductive group $\mathrm{Gl%
}\left( d,\mathbb{R}\right) $. Since a linear flow on a vector bundle lifts
to the bundle of frames our set up includes this classical case.
The flow of automorphisms on $Q$ induces a flow on the base space $X$, also
denoted by $\phi $. It also induces flows on bundles having as typical fiber
a space $F$ acted by $G$. Such bundle is built via the associated bundle
construction and is denoted by $Q\times _{G}F$. If there is no risk of
confusion the flows on the associated bundles are denoted by $\phi $ as well.
When $G$ is a reductive group we are specially interested in their flag
manifolds $\mathbb{F}_{\Theta }$, distinguished by the subindex $\Theta $,
which are compact homogeneous spaces of $G$. We write $\mathbb{E}_{\Theta
}=Q\times _{G}\mathbb{F}_{\Theta }$ for the corresponding flag bundle. For
the flow $\phi $ induced on $\mathbb{E}_{\Theta }$, it was proved in \cite%
{smbflow} and \cite{msm} that it has a finest Morse decomposition (under the
mild assumption that the flow on the base space $X$ is chain transitive).
Each Morse component of this finest decomposition meets a fiber of $\mathbb{E%
}_{\Theta }\rightarrow X$ in a algebraic submanifold of
$\mathbb{F}_{\Theta} $. This submanifold is defined as a set of
fixed points for some $g\in G$ acting on $\mathbb{F}_{\Theta }$.
For instance in a projective bundle the fibers of a Morse
component are subspaces, that can be seen as sets of fixed points
on the projective space of diagonalizable matrices (see also
Selgrade \cite{sel} for the Morse decomposition on a projective
bundle). The Morse decomposition is thus described by a continuous
section $\chi _{\mathrm{Mo}}$
of an associated bundle $Q\times _{G}\left( \mathrm{Ad}\left( G\right) H_{%
\mathrm{Mo}}\right) $, whose typical fiber is an adjoint orbit $\mathrm{Ad}%
\left( G\right) H_{\mathrm{Mo}}$ of $G$. Here $H_{\mathrm{Mo}}$ belongs to
the Lie algebra $\mathfrak{g}$ of $G$ and its adjoint $\mathrm{ad}\left( H_{%
\mathrm{Mo}}\right) $ has real eigenvalues. The Morse components are then
built from the section $\chi _{\mathrm{Mo}}$ and the fixed point sets of $%
\exp H_{\mathrm{Mo}}$ on the flag manifolds. (See \cite{msm}, Theorem 7.5.)
On the other hand we also have the Oseledets decomposition, coming
from the Multiplicative Ergodic Theorem (as proved in
\cite{alvsm}). To consider this decomposition it is required a
$\phi $-invariant measure $\nu $ on the base space. If $\nu $ is
ergodic and $\mathrm{supp}\nu =X$ (which provides chain
transitivity on $X$) then the Multiplicative Ergodic Theorem
yields an analogous decomposition to the Morse decomposition that
describes the level
sets of the $\mathfrak{a}$-Lyapunov exponents (see \cite{alvsm} and \cite%
{smsec}). Again there is an adjoint orbit $\mathrm{Ad}\left( G\right) H_{%
\mathrm{Ly}}$ and a section $\chi _{\mathrm{Ly}}$ of the associated bundle $%
Q\times _{G}\left( \mathrm{Ad}\left( G\right) H_{\mathrm{Ly}}\right) $ such
that the Oseledets decomposition is built from $\chi _{\mathrm{Ly}}$ and the
fixed point sets of $\exp H_{\mathrm{Ly}}$. The section $\chi _{\mathrm{Ly}}$
is now only measurable and defined up to a set of $\nu $-measure $0$.
It turns out that any component of the Oseledets decomposition is contained
in a component of the Morse decomposition (see Section \ref{secversusdec}
below). This means that the multiplicities of the eigenvalues of $\mathrm{ad}%
\left( H_{\mathrm{Mo}}\right) $ are larger than those of $\mathrm{ad}\left(
H_{\mathrm{Ly}}\right) $.
In this paper we write down three conditions that together are necessary and
sufficient for both decompositions to coincide (see Section \ref{threecon}).
In this case the Morse decomposition is a continuous extension of the
Oseledets decomposition.
The first of these conditions requires boundedness of the measurable section
$\chi _{\mathrm{Ly}}$, which means that different components of the
Oseledets decomposition do not approach each other. The other two conditions
are about the Oseledets decomposition for the other ergodic measures on $%
\mathrm{supp}\nu =X$. They can be summarized by saying that if $\rho $ is an
ergodic measure then its Oseledets decomposition is finer than the
decomposition for $\nu $.
It is easy to prove that each of the three conditions is necessary. Our main
result is to prove that together they imply equality of the decompositions.
Now we describe the contents of the paper and say some words about other
results that have independent interest.
Sections \ref{secprelim}, \ref{seclyapmorse} and \ref{secperflow} are
preliminary. Section \ref{secprelim} contains notation \ and general facts
about flag manifolds, while in Section \ref{seclyapmorse} we recall the
results of \cite{alvsm}, \cite{smbflow}, \cite{msm} and \cite{smsec} about
Morse decomposition, Morse and Lyapunov spectra and the Multiplicative
Ergodic Theorem on flag bundles. In Section \ref{secperflow} we discuss
briefly flows over periodic orbits that will be needed later.
Section \ref{secmedinv} is devoted to the analysis of ergodic measures on
the flag bundles. We exploit the Krylov-Bogolyubov technique of occupation
measures to see that any Lyapunov exponent coming from the Multiplicative
Ergodic Theorem is an integral over an ergodic measure and conversely.
Combining this with the fact that an ergodic measure charges just one
Oseledets component allow us to introduce what we called attractor and
repeller measures. Later their supports will provide attractor-repeller
pairs on the flag bundles thus relating them to the finest Morse
decomposition.
In Section \ref{secversusdec} we use the attractor and repeller measures to
check that the components of the Oseledets decomposition are indeed
contained in the finest Morse decomposition.
Another tool is developed in Section \ref{lyextgbun}, namely the Lyapunov
exponents of the derivative flow on the tangent space of the fibers of the
flag bundles. The knowledge of these exponents allow to find $\omega $%
-limits in the bundles themselves.
In Section \ref{teclemma} we prove our main technical lemma that furnishes
attractor-repeller pairs on the flag bundles.
In Section \ref{threecon} we state our conditions and prove that they are
necessary. Their sufficience is proved in Section \ref{sufcon}.
In the next two Sections \ref{secunique} and \ref{seciid} we discuss two
cases that go in opposite directions. Namely flows where the base space is
uniquely ergodic (Section \ref{secunique}) and product of i.i.d. sequences.
For a uniquely ergodic base space the second and third conditions are
vacuous and it follows by previous results that the Morse spectrum is a
polyhedron that degenerates to a point if the first condition is satisfied.
In the other side in the i.i.d. case there are plenty of invariant measures
enabling to find examples that violate our second condition. We do that with
the aid of a result by Guivarch'- Raugi \cite{gr}.
Section \ref{seccont} is independent of the rest of the paper. It contains a
result that motivates the study of the equality between Oseledets and Morse
decompositions. We prove that if both decompositions coincide for $\phi $
then the Lyapunov spectrum is continuous under perturbations $\sigma \phi $
of $\phi $ with $\sigma $ varying in gauge group $\mathcal{G}$ of $Q$. This
continuity is a consequence of the differentiability result of \cite{ferrsm}%
. By that result there exists a subset $\Phi _{\mathrm{Mo}}$ of linear
functionals defined from the finest Morse decomposition such that the map $%
\sigma \mapsto \alpha \left( H_{\mathrm{Ly}}\left( \sigma \phi \right)
\right) $ is differentiable with respect to $\sigma $ (at the identity) if $%
\alpha \in \Phi _{\mathrm{Mo}}$, where $H_{\mathrm{Ly}}\left( \sigma \phi
\right) $ is the vector Lyapunov spectrum of $\sigma \phi $. Having equality
of the decompositions we can exploit upper semi-continuity of the spectrum
to prove continuity of $\beta \left( H_{\mathrm{Ly}}\left( \sigma \phi
\right) \right) $ with $\beta $ in a basis that contains $\Phi _{\mathrm{Mo}%
} $.
Finally, we mention that for a linear flow $\phi $ on a vector bundle $%
\mathcal{V}\rightarrow X$ the topological property given by the finest Morse
decomposition of the flow induced on the projective bundle $\mathbb{P}%
\mathcal{V}\rightarrow X$ can be given an analytic characterization via
exponential separation of vector subbundles of $\mathcal{V}$ (see
Colonius-Kliemann \cite{ck}, Chapter 5, and Bonatti-Diaz-Viana \cite{bodivi}%
, Appendix B). In fact, by a theorem by Bronstein-Chernii \cite{brocher}
(quoted from \cite{ck}) the finest Morse decomposition on $\mathbb{P}%
\mathcal{V}$ corresponds to the finest decomposition of $\mathcal{V}$ into
exponentially separated subbundles (see \cite{ck}, Theorem 5.2.10). Hence
our main result gives, in particular, necessary and sufficient conditions
ensuring that the Oseledets decomposition of a vector bundle is
exponentially separated. Also the result of Section \ref{seccont} shows that
if Oseledets decomposition is exponentially separated then the Lyapunov
spectrum changes continuously when $\phi $ is perturbed in such a way that
the flow on the base $X$ is kept fixed.
\section{Flag manifolds}
\label{secprelim}
We explain here our notation about semi-simple (or reductive) Lie groups and
their flag manifolds. We refer to {Knapp} \cite{knp}, {%
Duistermat-Kolk-Varadarajan} \cite{DKV} and {Warner} \cite{Warner}.
Let $\mathfrak{g}$ be a semi-simple non-compact Lie algebra. In order to
make the paper understandable to readers without acquaintance with Lie
Theory we adopted the strategy of defining the notation by writing
explicitely their meanings for the special linear group $\mathrm{Sl}\left( d,%
\mathbb{R}\right) $ and its Lie algebra $\mathfrak{sl}\left( d,\mathbb{R}%
\right) $ (or $\mathrm{Gl}\left( d,\mathbb{R}\right) $ and $\mathfrak{gl}%
\left( d,\mathbb{R}\right) $ in the reductive case). We hope that the reader
with expertise in semi-simple theory will recognize the notation for the
general objects (e.g. $\mathfrak{k}$ is a maximal compact embedded
subalgebra, etc.)
At the Lie algebra level the Cartan decomposition reads $\mathfrak{g}=%
\mathfrak{k}\oplus \mathfrak{s}$, where $\mathfrak{k}=\mathfrak{so}\left(
d\right) $ is the subalgebra of skew-symmetric matrices and $\mathfrak{s}$
is the space of symmetric matrices. The Iwasawa decomposition of the Lie
algebra is $\mathfrak{g}=\mathfrak{k}\oplus \mathfrak{a}\oplus \mathfrak{n}$
where $\mathfrak{a}$ is the subalgebra of diagonal matrices and $\mathfrak{n}
$ is the subalgebra of upper triangular matrices with zeros on the diagonal.
The set of roots is denoted by $\Pi $. These are linear maps $\alpha
_{ij}\in \mathfrak{a}^{*}$, $i\neq j$, defined by $\alpha _{ij}\left(
\mathrm{diag}\{a_{1},\ldots ,a_{d}\}\right) =a_{i}-a_{j}$. The set of
positive roots is $\Pi ^{+}=\{\alpha _{ij}:i<j\}$ and the set of simple
roots is $\Sigma =\{\alpha _{ij}:j=i+1\}$. The root space is $\mathfrak{g}%
_{\alpha }$ ($\mathfrak{g}_{\alpha _{ij}}$ is spanned by the basic matrix $%
E_{ij}$) and
\[
\mathfrak{g}=\mathfrak{m}\oplus \mathfrak{a}\oplus \sum_{\alpha \in \Pi }%
\mathfrak{g}_{\alpha }
\]
where $\mathfrak{m}=\mathfrak{z}_{\mathfrak{k}}\left( \mathfrak{a}\right) =%
\mathfrak{z}\left( \mathfrak{a}\right) \cap \mathfrak{k}$ is the centralizer
of $\mathfrak{a}$ in $\mathfrak{k}$ ($\mathfrak{m}=0$ in $\mathfrak{sl}%
\left( d,\mathbb{R}\right) $). The basic (positive) Weyl chamber is denoted
by
\[
\mathfrak{a}^{+}=\{H\in \mathfrak{a}:\alpha \left( H\right) >0,\alpha \in
\Sigma \}
\]
(cone of diagonal matrices $\mathrm{diag}\{a_{1},\ldots ,a_{d}\}$ satisfying
$a_{1}>\cdots >a_{d}$). Its closure $\mathrm{cl}\mathfrak{a}^{+}$ is formed
by diagonal matrices with decreasing eigenvalues.
At the Lie group level the Cartan decomposition reads $G=KS$, $K=\exp
\mathfrak{k}$ and $S=\exp \mathfrak{s}$ ($K$ is the group $\mathrm{SO}\left(
d\right) $ and $S$ the space of positive definite symmetric matrices in $%
\mathrm{Sl}\left( d,\mathbb{R}\right) $). The Iwasawa decomposition is $%
G=KAN $, $A=\exp \mathfrak{a}$, $N=\exp \mathfrak{n}$. The Cartan
decomposition splits further into the polar decomposition $G=K\left( \mathrm{%
cl}A^{+}\right) K$, $A^{+}=\exp \mathfrak{a}^{+}$.
$M=\mathrm{Cent}_{K}\left( \mathfrak{a}\right) $ is the centralizer of $%
\mathfrak{a}$ in $K$ (diagonal matrices with entries $\pm 1$), $M^{\ast }=%
\mathrm{Norm}_{K}\left( \mathfrak{a}\right) $ is the normalizer of $%
\mathfrak{a}$ in $K$ (signed permutation matrices) and $\mathcal{W}=M^{\ast
}/M$ is the \textit{Weyl group} (for $\mathrm{Sl}\left( d,\mathbb{R}\right) $
it is the group of permutations in $d$ letters, that acts in $\mathfrak{a}$
by permuting the entries of a diagonal matrix).
The (standard) minimal parabolic subalgebra is $\mathfrak{p}=$ $\mathfrak{m}%
\oplus \mathfrak{a}\oplus \mathfrak{n}$ (= upper triangular matrices), and a
general standard parabolic subalgebra $\mathfrak{p}_{\Theta }$ is defined by
a subset $\Theta \subset \Sigma $ as
\[
\mathfrak{p}_{\Theta }=\mathfrak{m}\oplus \mathfrak{a}\oplus \sum_{\alpha
\in \Pi ^{+}}\mathfrak{g}_{\alpha }\oplus \sum_{\alpha \in \langle \Theta
\rangle ^{+}}\mathfrak{g}_{-\alpha },
\]%
where $\langle \Theta \rangle $ is the set of roots spanned (over $\mathbb{Z}
$) by $\Theta $ and $\langle \Theta \rangle ^{+}=\langle \Theta \rangle \cap
\Pi ^{+}$. That is, $\mathfrak{p}_{\Theta }=\mathfrak{p}\oplus \mathfrak{n}%
^{-}(\Theta )$, where $\mathfrak{n}^{\pm }(\Theta )=\sum_{\alpha \in \langle
\Theta \rangle ^{+}}\mathfrak{g}_{\pm \alpha }$.
Alternatively, given $\Theta $, take $H_{\Theta }\in \mathrm{cl}\mathfrak{a}%
^{+}$ such that $\alpha \left( H_{\Theta }\right) =0$, $\alpha \in \Sigma $,
if and only if $\alpha \in \Theta $. Such $H_{\Theta }$ exists and we call
it a characteristic element of $\Theta $. Then $\mathfrak{p}_{\Theta }$ is
the sum of eigenspaces of $\mathrm{ad}\left( H_{\Theta }\right) $ having
eigenvalues $\geq 0$. In $\mathfrak{sl}\left( d,\mathbb{R}\right) $, $%
H_{\Theta }=\mathrm{diag}\left( a_{1},\ldots ,a_{d}\right) $ with $a_{1}\geq
\cdots \geq a_{d}$, where the multiplicities of the eigenvalues is
prescribed by $a_{i}=a_{i+1}$ if $\alpha _{i,i+1}\in \Theta $, that is, $%
\mathfrak{p}_{\Theta }$ is the subalgebra of matrices that are upper
triangular in blocks, whose sizes are the multiplicities of the eigenvalues
of $H_{\Theta }$. When $\Theta $ is empty, $\mathfrak{p}_{\emptyset }$ boils
down to the minimal parabolic subalgebra $\mathfrak{p}$.
Conversely if $H\in \mathrm{cl}\mathfrak{a}^{+}$ then $\Theta _{H}=\{\alpha
\in \Sigma :\alpha \left( H\right) =0\}$ defines a flag manifold $\mathbb{F}%
_{\Theta _{H}}$ (e.g. the Grassmannian $\mathrm{Gr}_{k}\left( d\right) $ is
a flag manifold of $\mathrm{Sl}\left( d,\mathbb{R}\right) $ defined by $H=%
\mathrm{diag}\{a,\ldots ,a,b\ldots ,b\}$ with $\left( n-k\right) a+kb=0$).
For $H_{1},H_{2}\in \mathrm{cl}\mathfrak{a}^{+}$ we say that $H_{1}$ refines
$H_{2}$ in case $\Theta _{H_{1}}\subset \Theta _{H_{2}}$. In $\mathfrak{sl}%
\left( d,\mathbb{R}\right) $ this means that the blocks determined by the
multiplicities of the eigenvalues of $H_{1}$ is contained in the blocks of $%
H_{2}$.
The \textit{parabolic subgroup} $P_{\Theta }$, associated to $\Theta $, is
defined as the normalizer of $\mathfrak{p}_{\Theta }$ in $G$ (as a group of
matrices it has the same block structure as $\mathfrak{p}_{\Theta }$). It
decomposes as $P_{\Theta }=K_{\Theta }AN$, where $K_{\Theta }=\mathrm{Cent}%
_{K}\left( H_{\Theta }\right) $ is the centralizer of $H_{\Theta }$ in $K$.
We usually omit the subscript when $\Theta =\emptyset $ and $P=P_{\emptyset
} $ is the minimal parabolic subgroup.
The \textit{flag manifold} associated to $\Theta $ is the homogeneous space $%
\mathbb{F}_{\Theta }=G/P_{\Theta }$ (just $\mathbb{F}$ when $\Theta
=\emptyset $). If $\Theta _{1}\subset \Theta _{2}$ then the corresponding
parabolic subgroups satisfy $P_{\Theta _{1}}\subset P_{\Theta _{2}}$, so
that there is a canonical fibration $\pi _{\Theta _{2}}^{\Theta _{1}}:%
\mathbb{F}_{\Theta _{1}}\rightarrow \mathbb{F}_{\Theta _{2}}$, given by $%
gP_{\Theta _{1}}\mapsto gP_{\Theta _{2}}$ (just $\pi _{\Theta _{2}}$ if $%
\Theta _{1}=\emptyset $). For the matrix group the flag manifold $\mathbb{F}%
_{\Theta }$ identifies with the manifold of flags of subspaces $V_{1}\subset
\cdots \subset V_{k}$ where the differences $\dim V_{i+1}-\dim V_{i}$ are
the sizes of the blocks defined by $\Theta $ (or rather the diagonal \
matrix $H_{\Theta }$). The projection $\pi _{\Theta _{2}}^{\Theta _{1}}:%
\mathbb{F}_{\Theta _{1}}\rightarrow \mathbb{F}_{\Theta _{2}}$ is defined by
\textquotedblleft forgetting subspaces\textquotedblright .
The concept of \textit{dual flag manifold} is defined as follows: Let $w_{0}$
be the principal involution of $\mathcal{W}$, that is, the only element of $%
\mathcal{W}$ such that $w_{0}\mathfrak{a}^{+}=-\mathfrak{a}^{+}$, and put $%
\iota =-w_{0}$. Then $\iota (\Sigma )=\Sigma $ and for $\Theta \subset
\Sigma $ write $\Theta ^{\ast }=\iota (\Theta )$. Then $\mathbb{F}_{\Theta
^{\ast }}$ is called the flag manifold dual of $\mathbb{F}_{\Theta }$. For
the matrix group the vector subspaces of the flags in $\mathbb{F}_{\Theta
^{\ast }}$ have complementary dimensions to those in $\mathbb{F}_{\Theta }$
(for instance the dual of a Grassmannian $\mathrm{Gr}_{k}\left( d\right) $
is the Grasmmannian $\mathrm{Gr}_{d-k}\left( d\right) $).
We say that two elements $b_{1}\in \mathbb{F}_{\Theta }$ and $b_{2}\in
\mathbb{F}_{\Theta ^{\ast }}$ are transversal if $(b_{1},b_{2})$ belongs to
the unique open $G$-orbit in $\mathbb{F}_{\Theta }\times \mathbb{F}_{\Theta
^{\ast }}$, by the action $g\left( b_{1},b_{2}\right) =\left(
gb_{1},gb_{2}\right) $. For instance $b_{1}\in \mathrm{Gr}_{k}\left(
d\right) $ and $b_{2}\in \mathrm{Gr}_{d-k}\left( d\right) $ are transversal
if and only if they are transversal as subspaces of $\mathbb{R}^{d}$. In
general transversality can be expressed in terms of transversality of
subalgebras of $\mathfrak{g}$ (see e.g. \cite{smmax}). By the very
definition transversality is an open condition and if $g\in G$ then $gb_{1}$
is transversal to $gb_{2}$ if and only if $b_{1}$ is transversal to $b_{2}$.
The following lemma about transversality will be used afterwards.
\begin{lema}
\label{lemseqtransv}Let $b_{n}^{*}$ be a sequence in $\mathbb{F}_{\Theta
^{*}}$ with $\lim b_{n}^{*}=b^{*}$. Suppose that $b\in \mathbb{F}_{\Theta }$
is not transversal to $b^{*}$. Then there exists a sequence $b_{n}\in
\mathbb{F}_{\Theta }$ with $b_{n}$ not transversal to $b_{n}^{*}$ such that $%
\lim b_{n}=b$.
\end{lema}
\begin{profe}
There exists a sequence $k_{n}\in K$ with $b_{n}^{\ast }=k_{n}b^{\ast }$ and
$k_{n}\rightarrow 1$. Since $b$ is not transversal to $b^{\ast }$ it follows
that $k_{n}b$ is not transversal to $b_{n}^{\ast }$. Hence, $b_{n}=k_{n}b$
is the required sequence.
\end{profe}
We consider now the fixed point set of the action of $h=\exp H$, $H\in
\mathrm{cl}\mathfrak{a}^{+}$, on a flag manifold $\mathbb{F}_{\Theta }$.
Look first at the example of the projective space $\mathbb{R}P^{d-1}$. The
fixed point set is the union of the eigenspaces of $h$. The eigenspace
associated to the biggest eigenvalue is the only attractor (for the
iterations $h^{n}$) that has an open and dense stable manifold. The same way
the eigenspace of the smallest eigenvalue is the unique repeller with open
and dense unstable manifold.
In general the flow defined by $\exp tH$ is gradient in any flag manifold $%
\mathbb{F}_{\Theta }$ (see \cite{DKV}). Its fixed point set is given the
union of the orbits
\[
Z_{H}\cdot wb_{\Theta }=K_{H}\cdot wb_{\Theta }\qquad \,w\in \mathcal{W},
\]%
where $b_{\Theta }=P_{\Theta }$ the origin of $\mathbb{F}_{\Theta
}=G/P_{\Theta }$, $Z_{H}=\left\{ g\in G:\,\mathrm{Ad}(g)H=H\right\} $, $%
K_{H}=Z_{H}\cap K$ and $w$ runs through the Weyl group $\mathcal{W}$. We
write $\mathrm{fix}_{\Theta }(H,w)=Z_{H}\cdot wb_{\Theta }$ and refer to it
as the set of $H$-fixed points of type\textit{\ }$w$. In addition, $\mathrm{%
fix}_{\Theta }(H,1)$ is the only attractor while $\mathrm{fix}_{\Theta
}(H,w_{0})$ is the unique repeller, where $w_{0}$ the principal involution
of $\mathcal{W}$.
For the stable and unstable sets of $\mathrm{fix}_{\Theta }(H,w)$ let $%
\Theta _{H}=\{\alpha \in \Sigma :\alpha \left( H\right) =0\}$ and consider
the nilpotent subalgebras
\[
\mathfrak{n}_{H}^{+}=\sum_{\alpha \in \Pi ^{+}\backslash \langle \Theta
_{H}\rangle }\mathfrak{g}_{\alpha }\qquad \mathfrak{n}_{H}^{-}=\sum_{\alpha
\in \Pi ^{+}\backslash \langle \Theta _{H}\rangle }\mathfrak{g}_{-\alpha }
\]%
and the connected subgroups $N_{H}^{\pm }=\exp \mathfrak{n}_{H}^{\pm }$. Put%
\[
\mathrm{st}_{\Theta }(H,w)=N_{H}^{-}K_{H}\cdot wb_{\Theta }\qquad \mathrm{un}%
_{\Theta }(H,w)=N_{H}^{+}K_{H}\cdot wb_{\Theta }.
\]%
Then $\mathrm{st}_{\Theta }(H,w)$ and $\mathrm{un}_{\Theta }(H,w)$ are the
stable and unstable sets of $\mathrm{fix}_{\Theta }(H,w)$, \ respectively.
More generally if $D=\mathrm{Ad}(g)H$, $g\in G$ and $H\in \mathrm{cl}%
\mathfrak{a}^{+}$, then the dynamics of $\exp tD$ is conjugate under $g$ to
the dynamics of $\exp tH$. Hence $\mathrm{fix}_{\Theta }(D,w)=g\cdot \mathrm{%
fix}_{\Theta }(H,w)$, $\mathrm{st}_{\Theta }(D,w)=g\cdot \mathrm{st}_{\Theta
}(H,w)$ and $\mathrm{un}_{\Theta }(D,w)=g\cdot \mathrm{un}_{\Theta }(H,w)$.
It follows that
\[
\mathrm{st}_{\Theta }(D,w)=P_{D}^{-}\cdot gwb_{\Theta }\qquad \mathrm{un}%
_{\Theta }(H,w)=P_{D}^{+}\cdot gwb_{\Theta },
\]%
where $P_{D}^{\pm }=g{N}_{H}^{\pm }K_{H}g^{-1}=N_{D}^{\pm }K_{D}$, and $%
N_{D}^{\pm }=gN_{H}^{\pm }g^{-1}$ and $K_{D}=gK_{H}g^{-1}$.
If $H_{1}$ refines $H_{2}$ then the centralizers satisfy $Z_{H_{1}}\subset
Z_{H_{2}}$ hence the fixed point set of $\exp H_{1}$ in a flag manifold $%
\mathbb{F}_{\Theta }$ is contained in the fixed point set of $\exp H_{2}$.
The following lemma shows that we can control the inclusion of fixed point
sets for different elements by looking at the attractor and repeller fixed
point sets in the right flag manifolds.
\begin{lema}
\label{lemfixinclu}Suppose $H_{1}$ refines $H_{2}$, take $S\in \mathrm{Ad}%
\left( G\right) H_{1}$ and $T\in \mathrm{Ad}\left( G\right) H_{2}$, and put $%
s=\exp S$, $t=\exp T$. Suppose that $\mathrm{att}_{\Theta \left(
H_{1}\right) }\left( s\right) \subset \mathrm{att}_{\Theta \left(
H_{1}\right) }\left( t\right) $ and $\mathrm{rp}_{\Theta \left( H_{1}\right)
^{*}}\left( s\right) \subset \mathrm{rp}_{\Theta \left( H_{1}\right)
^{*}}\left( t\right) $. Then the fixed point set of $s$ in any flag manifold
is contained in the fixed point set of $t$.\newline
Moreover the fixed points are the same in case these attractor and repeller
fixed points coincide.
\end{lema}
\begin{profe}
If we identify $\mathrm{Ad}\left( G\right) H_{1}$ with the open orbit in $%
\mathbb{F}_{\Theta \left( H_{1}\right) }\times \mathbb{F}_{\Theta \left(
H_{1}\right) ^{\ast }}$ then $S$ is identified to the pair $\left( \mathrm{%
att}_{\Theta \left( H_{1}\right) }\left( s\right) ,\mathrm{rp}_{\Theta
\left( H_{1}\right) }\left( s\right) \right) $. The same way $T$ is
identified to the pair $\left( \mathrm{att}_{\Theta \left( H_{2}\right)
}\left( t\right) ,\mathrm{rp}_{\Theta \left( H_{2}\right) }\left( t\right)
\right) \in \mathbb{F}_{\Theta \left( H_{2}\right) }\times \mathbb{F}%
_{\Theta \left( H_{2}\right) ^{\ast }}$. Now, since $H_{1}$ refines $H_{2}$
there are fibrations $p:\mathrm{Ad}\left( G\right) H_{1}\rightarrow \mathrm{%
Ad}\left( G\right) H_{2}$, $\pi _{1}:\mathbb{F}_{\Theta \left( H_{1}\right)
}\rightarrow \mathbb{F}_{\Theta \left( H_{2}\right) }$ and $\pi _{2}:\mathbb{%
F}_{\Theta \left( H_{1}\right) ^{\ast }}\rightarrow \mathbb{F}_{\Theta
\left( H_{2}\right) ^{\ast }}$ with the equalities $\pi _{1}\left( \mathrm{%
att}_{\Theta \left( H_{1}\right) }\left( t\right) \right) =\mathrm{att}%
_{\Theta \left( H_{2}\right) }\left( t\right) $ and $\pi _{2}\left( \mathrm{%
att}_{\Theta \left( H_{1}\right) }\left( t\right) \right) =\mathrm{att}%
_{\Theta \left( H_{2}\right) }\left( t\right) $, we have $p\left( S\right) =T
$. This means there exists $g\in G$ such that $S=\mathrm{Ad}\left( g\right)
H_{1}$ and $T=\mathrm{Ad}\left( g\right) H_{2}$. Hence the fixed point set
of $s$ (respectively $t$) in a flag manifold $\mathbb{F}_{\Theta }$ is the
image under $g$ of the fixed point set of $\exp H_{1}$ (respectively $\exp
H_{2}$), which implies the lemma.
\end{profe}
\section{Lyapunov and Morse spectra and decompositions\label{seclyapmorse}}
From now on we consider a discrete-time continuous flow $\phi _{n}$ on a
continuous principal bundle $\left( Q,X,G\right) $, where the base space $X$
is a compact metric space endowed with an ergodic invariant measure $\nu $
with $\mathrm{supp}\nu =X$. The structural group $G$ is assumed to be
semi-simple and noncompact, or slightly more generally, $G$ is reductive
with noncompact semi-simple component. We fix once and for all a maximal
compact subgroup $K\subset G$ and a $K$-subbundle $R\subset Q$. (For a
bundle of frames of a vector bundle $\mathcal{V}\rightarrow X$ this amounts
to the choice of a Riemannian metric on $\mathcal{V}$. In case of a trivial
bundle $Q=X\times G$ the reduction is $R=X\times K$.)
The Iwasawa ($G=KAN$) and Cartan ($G=KS$) decompositions of $G$ yield
decompositions of $Q=R\times AN$ and $Q=R\times S$ by writing $q\in Q$ as
\[
q=r\cdot hn\qquad \mathrm{and}\qquad q=r\cdot s
\]%
$r\in R$, $hn\in AN$ and $s\in S$. In what follows we write for $q\in Q$,
\[
\mathsf{a}\left( q\right) =\log \mathsf{A}\left( q\right) \in \mathfrak{a}
\]%
where $\mathsf{A}\left( q\right) $ is the projection onto $A$ against the
Iwasawa decomposition. Also we write $\mathsf{S}:Q\rightarrow S$ as the
projection onto $S$ of $Q=R\times S$. By the polar decomposition $G=K(%
\mathrm{cl}A^{+})K$ we get a map $\mathsf{A}^{+}:Q\rightarrow \mathrm{cl}%
A^{+}$ by $\mathsf{S}\left( q\right) =k\mathsf{A}^{+}\left( q\right) k^{-1}$%
, $k\in K$. We write
\[
\mathsf{a}^{+}(q)=\log \mathsf{A}^{+}(q)\in \mathrm{cl}{\mathfrak{a}}^{+}.
\]
Now the flow $\phi _{n}$ on $Q$ induces a flow $\phi _{n}^{R}$ on $R$ by
declaring $\phi _{n}^{R}\left( r\right) $ to be the projection of $\phi _{n}
$ onto $R$ against the decomposition \ $Q=R\times AN$ ($\phi _{n}^{R}$ is
indeed a flow because $AN$ is a group.) The projections $\mathsf{a}$ and $%
\mathsf{a}^{+}$ define maps (denoted by the same letters) $\mathsf{a}:%
\mathbb{Z}\times R\rightarrow \mathfrak{a}$ and $\mathsf{a}^{+}:\mathbb{Z}%
\times R\rightarrow \mathrm{cl}\mathfrak{a}^{+}$ by
\[
\mathsf{a}\left( n,r\right) =\mathsf{a}\left( \phi _{n}\left( r\right)
\right) \qquad \mathrm{and}\qquad \mathsf{a}^{+}\left( n,r\right) =\mathsf{a}%
^{+}\left( \phi _{n}\left( r\right) \right) .
\]
It turns out that $\mathsf{a}\left( n,r\right) $ is an additive cocycle over
$\phi _{n}^{R}$. This cocycle factors to a cocycle (also denoted by $\mathsf{%
a}$) over the flow induced on $\mathbb{E}=Q\times _{G}\mathbb{F} $, a associated bundle of $Q$ with typical fiber %
the maximal flag manifold $\mathbb{F}$. The $\mathfrak{a}$-Lyapunov
exponent of $\phi _{n}$ in the direction of $\xi \in \mathbb{E}$ is defined
by
\[
\lambda \left( \xi \right) =\lim_{k\rightarrow +\infty }\frac{1}{k}\mathsf{a}%
\left( k,\xi \right) \in \mathfrak{a}\qquad \xi \in \mathbb{E}.
\]%
The polar exponent is defined by
\[
H_{\phi }\left( r\right) =\lim_{k\rightarrow +\infty }\frac{1}{k}\mathsf{a}%
^{+}\left( k,r\right) \in \mathrm{cl}\mathfrak{a}^{+}\qquad r\in R.
\]%
It turns out that $H_{\phi }\left( r\right) $ is constant along the fibers
of $R$ (when it exists) so is written $H_{\phi }\left( x\right) $, $x\in X$.
The existence of these limits is ensured by the
\vspace{12pt}%
\noindent%
\textbf{Multiplicative Ergodic Theorem} (\cite{alvsm}): The polar exponent $%
H_{\phi }\left( x\right) $ exists for $x$ in a set of total measure $\Omega $%
. Assume that $\nu $ is ergodic. Then $H_{\phi }\left( \cdot \right) $ is
almost surely equals to the constant $H_{\mathrm{Ly}}=H_{\mathrm{Ly}}\left(
\nu \right) \in \mathrm{cl}\mathfrak{a}^{+}$. Put $\mathbb{E}_{\Omega }=\pi
^{-1}\left( \Omega \right) $ where $\pi :\mathbb{E}\rightarrow X$ is the
projection. Then,
\begin{enumerate}
\item $\lambda \left( \xi \right) $ exists for every $\xi \in \mathbb{E}%
_{\Omega }$ and the map $\lambda :\mathbb{E}_{\Omega }\rightarrow \mathfrak{a%
}$ assume values in the finite set $\{wH_{\mathrm{Ly}}:w\in \mathcal{W}\}$.
\item There exists a measurable section $\chi _{\mathrm{Ly}}$ of the bundle $%
Q\times _{G}\mathrm{Ad}\left( G\right) \left( H_{\mathrm{Ly}}\right) $,
defined on $\Omega $, such that $\lambda \left( \xi \right) =w^{-1}H_{%
\mathrm{Ly}}$ if $\xi \in \mathrm{st}(\chi _{\mathrm{Ly}}\left( x\right) ,w)$%
, $x=\pi \left( \xi \right) $.
\end{enumerate}
(To be rigorous the stable set \ $\mathrm{st}(\chi _{\mathrm{Ly}}\left(
x\right) ,w)$, simply denoted by $\mathrm{st}(x,w)$, must be defined using the formalism of fiber bundles. If $%
Q=X\times G$ is trivial then $\chi _{\mathrm{Ly}}:X\rightarrow \mathrm{Ad}%
\left( G\right) \left( H_{\mathrm{Ly}}\right) $ and $\mathrm{st}(\chi _{%
\mathrm{Ly}}\left( x\right) ,w)$ is the stable set discussed in the last
section.)
We write $\mathrm{st}(w)$ for the union of the sets $\mathrm{st}(\chi _{%
\mathrm{Ly}}\left( x\right) ,w)$ with $x$ running through $\Omega $. The
same way we let $\mathrm{fix}(w)$ be the union of the fixed point sets $%
\mathrm{fix}(\chi _{\mathrm{Ly}}\left( x\right) ,w)$.
By analogy with the multiplicative ergodic theorem on vector bundles the
union of the sets $\mathrm{fix}(w)$, $w\in \mathcal{W}$, is called the
Oseledets decomposition of $\mathbb{E}$. These sets project to a partial
flag bundle $\mathbb{E}_{\Theta }$ to fixed point sets $\mathrm{fix}_{\Theta
}(w)$ that form the Oseledets decomposition of $\mathbb{E}_{\Theta }$.
To the exponent $H_{\mathrm{Ly}}\left( \nu \right) \in \mathrm{cl}\mathfrak{a%
}^{+}$ we associate the subset of the simple system of roots
\[
\Theta _{\mathrm{Ly}}=\Theta _{\mathrm{Ly}}\left( \nu \right) =\{\alpha \in
\Sigma :\alpha \left( H_{\mathrm{Ly}}\left( \nu \right) \right) =0\}.
\]%
The corresponding flag manifold $\mathbb{F}_{\Theta _{\mathrm{Ly}}}$ and
flag bundle $\mathbb{E}_{\Theta _{\mathrm{Ly}}}=Q\times _{G}\mathbb{F}%
_{\Theta _{\mathrm{Ly}}}$ play a proeminent role in the proofs. (For a
linear flow on a vector bundle $\mathbb{F}_{\Theta _{\mathrm{Ly}}}$ is the
manifold of flags $\left( V_{1}\subset \cdots \subset V_{k}\right) $ of
subspaces of $\mathbb{R}^{d}$ having the same dimensions as the subspaces of
the Oseledets splitting when the Lyapunov spectrum is ordered decreasingly.)
We refer to $\mathbb{F}_{\Theta _{\mathrm{Ly}}}$ as the flag type of $\phi $
with respect to $\nu $.
As another remark we mention that the section $\chi _{\mathrm{Ly}}$ yields
(actually is built from) two sections $\xi $ and $\xi ^{\ast }$ of the flag
bundles $\mathbb{F}_{\Theta _{\mathrm{Ly}} }$ and $\mathbb{%
F}_{\Theta _{\mathrm{Ly}}^{\ast }}$, respectively. Their images are defined
from level sets of Lyapunov exponents and hence are measurable (see \cite%
{alvsm}, Section 7.1).
On the other hand there are continuous decompositions of the flag bundles
(defined the same way as sets of fixed points) obtained by working out the
concept of Morse decomposition of the flows on the bundles (see Conley \cite%
{con} and Colonius-Kliemann \cite{ck}). It was proved in \cite{smbflow} and
\cite{msm} that if the flow on the base space is chain transitive then the
flow on any flag bundle $\mathbb{E}_{\Theta }$ admits a finest Morse
decomposition with Morse sets $\mathcal{M}\left( w\right) $, also
parametrized by $w\in \mathcal{W}$. Analogous to the Oseledets decomposition
the Morse sets are built as fixed point sets defined by a continuous section
of an adjoint bundle $\chi _{\mathrm{Mo}}:X\rightarrow Q\times _{G}\mathrm{Ad%
}\left( G\right) H_{\mathrm{Mo}}$, where $H_{\mathrm{Mo}}\in \mathrm{cl}%
\mathfrak{a}^{+}$ as well. There is just one attractor Morse component which
is given by $\mathcal{M}^{+}=\mathcal{M}\left( 1\right) $. There is a unique repeller
component as well which is $\mathcal{M}^{-}=\mathcal{M}\left( w_{0}\right) $, where $w_{0}$
is the principal involution.
The assumption that the invariant measure $\nu $ is ergodic with support $%
\mathrm{supp}\nu =X$ implies chain transitivity on $X$.
We write
\[
\Theta _{\mathrm{Mo}}=\Theta _{\mathrm{Mo}}\left( \phi \right) =\{\alpha \in
\Sigma :\alpha \left( H_{\mathrm{Mo}}\right) =0\}
\]%
and refer to $\mathbb{F}_{\Theta _{\mathrm{Mo}}}$ as the flag type of $\phi $
(with respect to the Morse decomposition).
The spectral counterpart of the Morse decomposition is the Morse spectrum
associated to the cocycle $\mathsf{a}\left( n,\xi \right) $. This spectrum
was originally defined by Colonius-Kliemann \cite{ck} for a flow on a vector
bundle and extended to flag bundles (and vector valued cocycles) in \cite%
{smsec}. By the results of \cite{smsec}, each Morse set $\mathcal{M}\left(
w\right) $ has a Morse spectrum $\Lambda _{\mathrm{Mo}}\left( w\right) $
which is a compact convex subset of $\mathfrak{a}$ and contains any $%
\mathfrak{a}$-Lyapunov exponent $\lambda \left( \xi \right) $, $\xi \in
\mathcal{M}\left( w\right) $. The attractor Morse component is given by the
identity $1\in \mathcal{W}$ and we write $\Lambda _{\mathrm{Mo}}=\Lambda _{%
\mathrm{Mo}}\left( 1\right) $, which is the only Morse spectrum meeting $%
\mathrm{cl}\mathfrak{a}^{+}$. The Morse spectrum $\Lambda _{\mathrm{Mo}}$
satisfies the following properties:
\begin{enumerate}
\item $\Lambda _{\mathrm{Mo}}$ is invariant under the group $\mathcal{W}%
_{\Theta _{\mathrm{Mo}}}$ generated by reflections with respect to the roots
$\alpha \in \Theta _{\mathrm{Mo}}$. (See \cite{smsec}, Theorem 8.3.)
\item $\alpha \left( H\right) >0$ if $H\in \Lambda _{\mathrm{Mo}}$ and $%
\alpha $ is a positive root that does not belong to the set $\langle \Theta
\rangle ^{+}$ spanned by $\Theta $. (See \cite{smsec}, Corollary 7.4.)
\end{enumerate}
By the last statement $\alpha \left( H_{\mathrm{Ly}}\right) >0$ if $\alpha $
is a simple root outside $\Theta _{\mathrm{Mo}}$ because $H_{\mathrm{Ly}}\in
\Lambda _{\mathrm{Mo}}$. Hence $\alpha \notin \Theta _{\mathrm{Ly}}$ by
definition of $\Theta _{\mathrm{Ly}}$. It follows that $\Theta _{\mathrm{Ly}%
}\subset \Theta _{\mathrm{Mo}}$. Below in Section \ref{secversusdec} we
improve this statement by proving, with the aid of invariant measures on
flag bundles, that the Oseledets decomposition is contained in the Morse
decomposition.
Our objective is to find necessary and sufficient conditions ensuring that $%
\Theta _{\mathrm{Ly}}=\Theta _{\mathrm{Mo}}$, and hence that the Oseledets
decomposition coincides with the Morse decomposition.
\section{Flows over periodic orbits}
\label{secperflow}
Before proceding let us recall the case where the base space is a single
periodic orbit $X=\{x_{0},\ldots ,x_{\omega -1}\}$ of period $\omega $, that
will be used later to reduce some arguments to nonperiodic orbits.
In the periodic case we have $\Theta _{\mathrm{Ly}}=\Theta _{\mathrm{Mo}}$
since, as is well known, the asymptotics depend ultimately on iterations of
a fixed element in the group $G$. Here the principal bundle is $Q=X\times G$
and the flow is given by
\[
\phi \left( x_{i},h\right) =\left( x_{i+1\left( \mathrm{mod}\omega \right)
},A\left( x_{i}\right) h\right)
\]%
for a map $A:X\rightarrow G$, so that
\[
\phi ^{n}\left( x_{i},h\right) =\left( x_{i+n\left( \mathrm{mod}\omega
\right) },g_{n,i}h\right)
\]%
where $g_{n,i}=A\left( x_{i+n-1\left( \mathrm{mod}\omega \right) }\right)
\cdots A\left( x_{i+1}\right) A\left( x_{i}\right) $. We have $%
g_{n+m,i}=g_{n,i+m\left( \mathrm{mod}\omega \right) }g_{m,i}$ so that $%
g_{k\omega ,i}=g_{\omega ,i}^{k}$. Hence the asymptotics of an orbit
starting at a point above $x_{i}$ is dictated by the iterations of the
action of $g_{\omega ,i}$. The iterations for the action of a fixed $g\in G$
on the flag manifolds, as well as the continuous time version of
periodicity, were studied by Ferraiol-Patr\~{a}o-Seco \cite{fepase}. Let $%
g_{n,i}=u_{n,i}h_{n,i}x_{n,i}$ be the Jordan decomposition of $g_{n,i}$ with
$u_{n,i}$, $h_{n,i}$ and $x_{n,i}$ elyptic, hyperbolic and unipotent
respectively. There is a choice of an Iwasawa decomposition $G=KAN$ such
that $u_{n,i}\in K$, $h_{n,i}\in A$ and $x_{n,i}\in N$. It follows that the
Lyapunov spectrum is given by $\log h_{\omega ,i}$, which is the same for
any $i=0,\ldots ,\omega -1$ (because $g_{\omega ,i+1}=A\left( x_{i}\right)
g_{\omega ,i}A\left( x_{i}\right) ^{-1}$). Also, as proved in \cite{fepase}
the Morse decomposition is given by the fixed point sets of $h_{\omega ,i}$.
Hence $\Theta _{\mathrm{Ly}}=\Theta _{\mathrm{Mo}}$.
\section{Invariant measures on the bundles and $\mathfrak{a}$-Lyapunov
exponents}
\label{secmedinv}
Let $\mu $ be an invariant measure for the flow on the maximal flag bundle $%
\pi :\mathbb{E}\rightarrow X$. Then the integral
\[
\int qd\mu \qquad \,q(\xi )=\mathsf{a}(1,\xi )
\]%
is an $\mathfrak{a}$-Lyapunov exponent for the cocycle $\mathsf{a}\left(
n,\xi \right) $ (see \cite{smsec}). On the other hand by applying the
Multiplicative Ergodic Theorem to an invariant measure $\nu $ on the base
space we obtain $\mathfrak{a}$-Lyapunov exponents, which we call regular
Lyapunov exponents with respect to $\nu $ (because they are obtained as
limits of sequences in $\mathfrak{a}$ which in turn comes from regular
sequences in $G$, see \cite{alvsm}).
In this section we show that these Lyapunov exponents coincide. Namely, if $%
\nu $ is ergodic measure on $X$ then any of its $\mathfrak{a}$-Lyapunov
exponents is an integral over an ergodic measure $\mu $ that projects onto $%
\nu $, i.e., $\pi _{\ast }\mu =\nu $, and conversely any such integral is a
regular Lyapunov exponent.
Fix an ergodic invariant measure $\nu $ on the base space and let $\Omega
\subset X$ be the set of $\nu $-total measure given by the Multiplicative
Ergodic Theorem (as proved in \cite{alvsm}). Recall that
\[
\pi ^{-1}\left( \Omega \right) =\dot{\bigcup_{w\in \mathcal{W}_{\Theta _{%
\mathrm{Ly}}}\backslash \mathcal{W}}}\mathrm{st}(w)
\]%
and $\lambda \left( \xi \right) =w^{-1}H_{\mathrm{Ly}}$ if $\xi \in \mathrm{%
st}(w)$, where $H_{\mathrm{Ly}}$ is the polar exponent with respect to $\nu $%
.
\begin{proposicao}
\label{propmedinvconjest}Let $\mu $ be an ergodic measure on $\mathbb{E}$
that projects onto $\nu $. Then there exists $w\in \mathcal{W}$ such that $%
\mu (\mathrm{st}(w))=1$ and $\mu \left( \mathrm{st}(w^{\prime })\right) =0$
if $\mathcal{W}_{\Theta _{\mathrm{Ly}}}w\neq \mathcal{W}_{\Theta _{\mathrm{Ly%
}}}w^{\prime }$. In this case
\[
\int qd\mu =w^{-1}H_{\mathrm{Ly}}.
\]
\end{proposicao}
\begin{profe}
By ergodicity of $\mu $ and the ergodic theorem applied to $\mu $ and $%
q\left( \xi \right) =\mathsf{a}(1,\xi )$, there exists a measurable set $%
\mathcal{I}\subset \mathbb{E}$ with $\mu (\mathcal{I})=1$ and
\[
\lambda (\xi )=\lim_{k\rightarrow \infty }\frac{1}{k}\mathsf{a}(k,\xi )=\int
qd\mu \qquad \xi \in \mathcal{I}.
\]%
Now $\mu (\pi ^{-1}(\Omega )\cap \mathcal{I})=1$ and $\pi ^{-1}(\Omega )\cap
\mathcal{I}$ is the disjoint union of the sets $\mathrm{st}(w)\cap \mathcal{I%
}$. In each $\mathrm{st}(w)\cap \mathcal{I}$, $w\in \mathcal{W}$, $\lambda $
is defined and is a constant equal to $w^{-1}H_{\mathrm{Ly}}$. Since $%
\lambda $ is constant on $\mathcal{I}$, it follows that $\pi ^{-1}(\Omega
)\cap \mathcal{I}\subset \mathrm{st}\left( w\right) $, for some $w\in
\mathcal{W}$. Then for any $\xi \in \mathcal{I}$,
\[
\int qd\mu =\lambda (\xi )=w^{-1}H_{\mathrm{Ly}}.
\]%
Finally, $\mu \left( \mathrm{st}(w)\right) \geq \mu \left( \pi ^{-1}(\Omega
)\cap \mathcal{I}\right) =1$, which implies that $\mu \left( \mathrm{st}%
(w^{\prime })\right) =0$ if $\mathrm{st}(w^{\prime })\neq \mathrm{st}(w)$
that is if $W_{\Theta _{\mathrm{Ly}}}w\neq W_{\Theta _{\mathrm{Ly}%
}}w^{\prime }$.
\end{profe}
\begin{corolario}
Let $\Lambda _{\mathrm{Mo}}\left( w\right) \subset \mathfrak{a}$ be the
Morse spectrum of the Morse set $\mathcal{M}\left( w\right) $. Then the
extremal points of the compact convex set $\Lambda _{\mathrm{Mo}}\left(
w\right) $ are regular Lyapunov exponents for ergodic measures on the base
space.
\end{corolario}
\begin{profe}
In fact it was proved \cite{smsec} (see Theorem 3.2(6)) that any extremal
point of $\Lambda _{\mathrm{Mo}}\left( w\right) $ is an integral $\int qd\mu
$ with respect to an ergodic measure $\mu $ on $\mathbb{E}$. (See also \cite%
{ck}, Lemma 5.4.10.)
\end{profe}
The converse to the above proposition says that any regular Lyapunov
exponent is the integral of $q$ with respect to some ergodic measure
projecting onto $\nu $. In order to prove the converse we recall the
Krylov-Bogolyubov procedure of constructing invariant measures as occupation
measures (see e.g. \cite{ck}). Let $\psi _{n}$, $n\in \mathbb{Z}$, be a flow
on a compact metric space $Y$. Then this means that
\[
\left( L_{n,x}f\right) \left( x\right) =\frac{1}{n}\sum_{k=0}^{n-1}f\left(
\psi _{k}x\right) ,\qquad x\in Y,
\]%
define linear maps on the space $C_{0}\left( Y\right) $ of continuous
functions, and hence Borel probability measures $\rho _{n}$. An accumulation
point $\rho _{x}=\lim_{k}\rho _{n_{k}}$ is called an (invariant) occupation
measure. When the limit $\widetilde{f}\left( x\right) =\lim_{n}\frac{1}{n}%
\sum_{k=0}^{n-1}f\left( \phi _{k}x\right) $ exists it is an integral $%
\widetilde{f}\left( x\right) =\int f\left( y\right) \mu _{x}\left( dy\right)
$ with respect to an occupation measure. The following properties will be
used below:
\begin{enumerate}
\item Let $\rho $ be an ergodic probability measure on $Y$. Then for $\rho $%
-almost every $y\in Y$, any occupation measure $\rho _{y}=\rho $. (This is
an easy consequence of Birkhoff ergodic theorem.)
\item There exists a set $\mathcal{J}\subset Y$ of total probability (that
is $\rho \left( \mathcal{J}\right) =1$ for every invariant measure $\rho $)
such that for all $y\in \mathcal{J}$ there exists an ergodic occupation
measure $\rho _{y}$.
\end{enumerate}
\begin{proposicao}
\label{proplyapreglyapint}Given $w\in \mathcal{W}$ there exists an invariant
ergodic measure $\mu ^{w}$ on $\mathbb{E}$ with $\pi _{*}\mu ^{w}=\nu $ such
that
\[
\int qd\mu ^{w}=w^{-1}H_{\mathrm{Ly}}
\]
and $\mu ^{w}(\mathrm{st}(w))=1$.
\end{proposicao}
\begin{profe}
If $\xi \in \mathrm{st}(w)$ then
\[
\lambda (\xi )=\lim_{k\rightarrow +\infty }\frac{1}{k}\mathsf{a}(k,\xi
)=w^{-1}H_{\mathrm{Ly}}
\]%
and since $\mathsf{a}(k,\xi )$ is a cocycle it follows that there exists an
occupation measure $\mu _{\xi }$ such that
\[
w^{-1}H_{\mathrm{Ly}}=\int qd\mu _{\xi }.
\]%
Note that $\pi _{\ast }(\mu _{\xi })$ is an occupation measure $\rho _{x}$
with $x=\pi (\xi )$. Since $\nu $ is ergodic, for $\nu $-almost all $x$, $%
\rho _{x}=\nu $ and hence we can choose $\xi $ with $\pi _{\ast }\left( \mu
_{\xi }\right) =\nu $.
It is not clear in advance that $\mu _{\xi }$ is ergodic. Nevertheless we
can decompose $\mu _{\xi }$ into ergodic components $\theta _{\eta }$ with $%
\eta $ ranging through a set $\mathcal{A}$ of $\mu _{\xi }$ total
probability, that is,
\[
\mu _{\xi }(\cdot )=\int \theta _{\eta }(\cdot )d\mu _{\xi }(\eta ).
\]
Since $\pi _{*}\mu _{\xi }=\nu $ it follows that $\pi _{*}\theta _{\eta
}=\nu $ for $\mu _{\xi }$-almost all $\eta $.
We claim that
\[
w^{-1}H_{\mathrm{Ly}}=\int qd\theta _{\eta }
\]%
for almost all $\eta \in \mathcal{A}$. In fact,
\[
w^{-1}H_{\mathrm{Ly}}=\int_{\mathbb{E}}\left( \int qd\theta _{\eta }\right)
d\mu _{\xi }(\eta ).
\]%
Hence $w^{-1}H_{\mathrm{Ly}}$ belongs to the convex closure of the set $%
\left\{ \int qd\theta _{\eta }\in \mathfrak{a};\,\eta \in \mathcal{A}%
\right\} $. However, by Proposition \ref{propmedinvconjest}, for any ergodic
$\theta _{\eta }$ there exists $u\in \mathcal{W}$ such that $\int qd\theta
_{\eta }=u^{-1}H_{\mathrm{Ly}}$, so that $w^{-1}H_{\mathrm{Ly}}$ is a convex
combination of points of the orbit $\mathcal{W}\cdot H_{\mathrm{Ly}}$. But
this is possible only if $\int qd\theta _{\eta }=w^{-1}H_{\mathrm{Ly}}$ for
almost all $\eta $, because the convex closure of the orbit $\mathcal{W}%
\cdot H_{\mathrm{Ly}}$ is a polyhedron whose vertices (extremal points) are
the points of the orbit. Hence there exists $\mu ^{w}$ yielding the Lyapunov
exponent $w^{-1}H_{\mathrm{Ly}}$. Finally, the equality $\mu ^{w}(\mathrm{st}%
(w))=1$ follows by the previous proposition.
\end{profe}
Now, we select two special kinds of ergodic measures on the flag bundles.
\begin{defi}
An ergodic measure $\mu $ on the maximal flag bundle $\mathbb{E}$ is said to
be an attractor measure for the flow if $\int qd\mu \in \mathrm{cl}\,%
\mathfrak{a}^{+}$. A measure $\mu _{\Theta }$ in $\mathbb{E}_{\Theta }$ is
an attractor measure if $\mu _{\Theta }={\pi _{\Theta }}_{\ast }\mu $ with $%
\mu $ attractor in $\mathbb{E}$, where $\pi _{\Theta }:\mathbb{E}\rightarrow
\mathbb{E}_{\Theta }$ is the canonical projection. \newline
Similarly, a measure $\mu $ in $\mathbb{E}$ is a repeller measure if $\int
qd\mu ^{w}\in -\mathrm{cl}\,\mathfrak{a}^{+}$, and $\mu _{\Theta }$ in $%
\mathbb{E}_{\Theta }$ is repeller if $\mu _{\Theta }={\pi _{\Theta }}_{\ast
}\mu $ with $\mu $ repeller in $\mathbb{E}$.
\end{defi}
Proposition \ref{proplyapreglyapint} ensures the existence of both attractor
and repeller measures.
\begin{proposicao}
\label{propatracrep}A repeller measure is an attractor measure for the
backward flow.
\end{proposicao}
\begin{profe}
Let $\mu $ be a repeller measure on $\mathbb{E}$ and write $q^{-}\left(
\cdot \right) =\mathsf{a}\left( -1,\cdot \right) $. Then by the cocycle
property $q^{-}(\xi )=-\mathsf{a}(1,\phi _{-1}(\xi ))=-q(\phi _{-1}(\xi ))$,
so that
\[
\int q^{-}d\mu =-\int qd\mu \in \mathrm{cl}\,\mathfrak{a}^{+}
\]
because $\int qd\mu \in -\mathrm{cl}\,\mathfrak{a}^{+}$. Thus $\mu $ is an
attractor measure for the backward flow. This proves the statement on the
maximal flag bundle $\mathbb{E}$. On the other bundles the result follows by
definition.
\end{profe}
Now we relate the supports of the attractor and repeller measures with the
decomposition given by the Multiplicative Ergodic Theorem on the flag
bundles $\mathbb{E}_{\Theta _{\mathrm{Ly}}\left( \nu \right) }$ and in its
dual $\mathbb{E}_{\Theta _{\mathrm{Ly}}^{\ast }\left( \nu \right) }$. This
decomposition is given by sections $\xi $ and $\xi ^{\ast }$ of $\mathbb{E}%
_{\Theta _{\mathrm{Ly}}\left( \nu \right) }$ and $\mathbb{E}_{\Theta _{%
\mathrm{Ly}}^{\ast }\left( \nu \right) }$, respectively.
We write simply $\Theta _{\mathrm{Ly}}=\Theta _{\mathrm{Ly}}\left( \nu
\right) $ and distinguish the several projections as: $\pi :\mathbb{E}%
\rightarrow X$, $\pi _{\Theta _{\mathrm{Ly}}}:\mathbb{E}\rightarrow \mathbb{E%
}_{\Theta _{\mathrm{Ly}}}$, $\pi _{\Theta _{\mathrm{Ly}}^{\ast }}:\mathbb{E}%
\rightarrow \mathbb{E}_{\Theta _{\mathrm{Ly}}^{\ast }}$ and $p$ for either $%
\mathbb{E}_{\Theta _{\mathrm{Ly}}}\rightarrow X$ or $\mathbb{E}_{\Theta _{%
\mathrm{Ly}}^{\ast }}\rightarrow X$.
Let $\mu $ be a repeller measure on $\mathbb{E}$ and put $\mu _{\Theta _{%
\mathrm{Ly}}^{*}}={\pi _{\Theta _{\mathrm{Ly}}^{*}}}_{*}\left( \mu \right) $
for the corresponding repeller measure on $\mathbb{E}_{\Theta _{\mathrm{Ly}%
}^{*}}$. We have $p_{*}(\mu _{\Theta _{\mathrm{Ly}}^{*}})=\nu $ because $%
p\circ \pi _{\Theta _{\mathrm{Ly}}^{*}}=\pi $ and $\pi _{*}\mu =\nu $. Hence
we can desintegrate $\mu _{\Theta _{\mathrm{Ly}}^{*}}$ with respect to $\nu $
to get
\[
\mu _{\Theta _{\mathrm{Ly}}^{*}}(\cdot )=\int_{X}\rho _{x}(\cdot )d\nu (x),
\]
where $x\in X\mapsto \rho _{x}\in \mathbb{M}^{+}\left( \mathbb{E}_{\Theta _{%
\mathrm{Ly}}^{*}}\right) $ is a measurable map into the space of probability
measures on $\mathbb{E}_{\Theta _{\mathrm{Ly}}^{*}}$.
\begin{lema}
For $\nu $-almost all $x\in X$ the component $\rho _{x}$ in the above
desintegration is a Dirac measure at $\xi ^{*}(x)$, that is, $\rho
_{x}=\delta _{\xi ^{*}\left( x\right) }$.
\end{lema}
\begin{profe}
Let $Z$ be the Borel set
\[
Z=\left\{ \mathrm{im}\,\xi ^{\ast }\right\} ^{c}=\mathbb{E}_{\Theta _{%
\mathrm{Ly}}^{\ast }}\backslash \left\{ \mathrm{im}\,\xi ^{\ast }\right\} .
\]%
Then
\[
\mu _{\Theta _{\mathrm{Ly}}^{\ast }}(Z)=\mu (\pi _{\Theta _{\mathrm{Ly}%
}^{\ast }}^{-1}(Z))=\mu (\mathbb{E}\backslash \mathrm{st}(w_{0}))=0,
\]%
because $\mu $ is a repeller measure. However,
\[
0=\mu _{\Theta _{\mathrm{Ly}}^{\ast }}(Z)=\int_{X}\rho _{x}(Z)d\nu (x),
\]%
and since $\rho _{x}$ is supported at $\pi ^{-1}(x)$, it follows that $\rho
_{x}(\mathbb{E}_{\Theta _{\mathrm{Ly}}^{\ast }}\setminus \{\xi ^{\ast
}(x)\})=0$, for $\nu $-almost all $x\in X$.
\end{profe}
This lemma shows also that a repeller measure on the dual flag manifold $%
\mathbb{E}_{\Theta _{\mathrm{Ly}}^{*}}$ is unique. Now we can apply the same
argument for the reverse flow $\phi _{-t}$, and get a similar result now for
an attracting measure on $\mathbb{E}_{\Theta _{\mathrm{Ly}}}$ with $\xi ^{*}$
replaced by $\xi $.
For later reference we summarize these facts in the following proposition.
\begin{proposicao}
\label{propmedatr}There exists a unique attractor measure $\mu _{\Theta _{%
\mathrm{Ly}}}^{+}$ for $\phi _{t}$ in its flag type $\mathbb{E}_{\Theta _{%
\mathrm{Ly}}}$, which is a Dirac measure on $\xi \left( x\right) $, that is,
it desintegrates as
\[
\mu _{\Theta _{\mathrm{Ly}}}^{+}(\cdot )=\int \delta _{\xi (x)}(\cdot )d\nu
(x)
\]
with respect to $\nu $. There exists also a unique repeller measure $\mu
_{\Theta _{\mathrm{Ly}}}^{-}$ on $\mathbb{E}_{\Theta _{\mathrm{Ly}}^{*}}$
which is Dirac at $\xi ^{*}$.
\end{proposicao}
\begin{corolario}
\label{coruniquemedatr}There exists a unique attractor (respectively
repeller) measure in $\mathbb{E}_{\Theta }$ if $\Theta _{\mathrm{Ly}}\subset
\Theta $ (respectively $\Theta _{\mathrm{Ly}}^{*}\subset \Theta $).
\end{corolario}
\begin{profe}
This is because the projection $\mathbb{E}\rightarrow \mathbb{E}_{\Theta }$
factors through $\mathbb{E}_{\Theta _{\mathrm{Ly}}}$ if $\Theta _{\mathrm{Ly}%
}\subset \Theta $: $\mathbb{E}\rightarrow \mathbb{E}_{\Theta _{\mathrm{Ly}%
}}\rightarrow \mathbb{E}_{\Theta }$. Hence a measure in $\mathbb{E}_{\Theta
} $ is attractor if and only if it is the projection of the attractor
measure in $\mathbb{E}_{\Theta _{\mathrm{Ly}}}$.
\end{profe}
\section{Morse decomposition $\times $ Oseledets decomposition\label%
{secversusdec}}
In this section we use the concepts of attractor and repeller measures
developed above to relate the Oseledets decomposition and the Morse
decomposition on a flag bundle $\mathbb{E}_{\Theta }$, as well as the
Lyapunov spectrum \ $H_{\mathrm{Ly}}$ and the Morse spectrum $\Lambda _{%
\mathrm{Mo}}$.
First we have the following consequence of Proposition \ref%
{proplyapreglyapint}.
\begin{proposicao}
\label{proplmbdamocontido}Suppose that $\alpha \left( \Lambda _{\mathrm{Mo}%
}\right) =0$ for all $\alpha \in \Theta _{\mathrm{Mo}}$. Then $\Theta _{%
\mathrm{Ly}}\left( \rho \right) =\Theta _{\mathrm{Mo}}$ for every ergodic
measure $\rho $ on the base space.
\end{proposicao}
\begin{profe}
As checked in Section \ref{seclyapmorse} we have $\Theta _{\mathrm{Ly}%
}\left( \rho \right) \subset \Theta _{\mathrm{Mo}}$. On the other hand by
Proposition \ref{proplyapreglyapint} any regular Lyapunov exponent is a
Morse exponent, that is, $H_{\mathrm{Ly}}\left( \rho \right) \subset \Lambda
_{\mathrm{Mo}}$. So that $\alpha \left( H_{\mathrm{Ly}}\left( \rho \right)
\right) =0$ if $\alpha \in \Theta _{\mathrm{Mo}}$, showing that $\Theta _{%
\mathrm{Mo}}\subset \Theta _{\mathrm{Ly}}\left( \rho \right) $.
\end{profe}
Now we look at the decompositions of the flag bundles.
\begin{proposicao}
Let $\mu $ be an attractor measure on $\mathbb{E}$. Then its support $%
\mathrm{supp}\mu $ is contained in the unique attractor component $\mathcal{M%
}^{+}$ of the finest Morse decomposition.
\end{proposicao}
\begin{profe}
Each point in $\mathrm{supp}\mu $ is recurrent and hence belongs to the set
of chain recurrent points which is the union of the Morse components.
Now since $\mu $ is an attractor measure, by definition the integral
\[
\lambda _{\mu }=\int qd\mu \in \mathrm{cl}\mathfrak{a}^{+}.
\]%
This integral is the $\mathfrak{a}$-Lyapunov exponent of $\mu $-almost all $%
z\in \mathrm{supp}\mu $. Hence is contained in the Morse spectrum. Actually,
$\lambda _{\mu }\in \Lambda _{\mathrm{Mo}}\left( \mathcal{M}^{+}\right) $,
the Morse spectrum of $\mathcal{M}^{+}$, because this is the only Morse
component whose spectrum meets $\mathrm{cl}\mathfrak{a}^{+}$. Therefore for $%
\mu $-almost all $z\in \mathrm{supp}\mu $, $z\in \mathcal{M}^{+}$. Since $%
\mathcal{M}^{+}$ is compact it follows that $\mathrm{supp}\mu \subset
\mathcal{M}^{+}$.
\end{profe}
By taking the backward flow we get a similar result for the repeller
measures.
\begin{proposicao}
Let $\mu $ be a repeller measure on $\mathbb{E}$. Then its support $\mathrm{%
supp}\mu $ is contained in the unique repeller component $\mathcal{M}^{-}$
of the finest Morse decomposition.
\end{proposicao}
\begin{proposicao}
\label{prooselcontmor}Let $\mathcal{O}$ be a component of the Oseledets
decomposition in a flag bundle $\mathbb{E}_{\Theta }$. Then there exists a
component $\mathcal{M}$ of the Morse decomposition of $\mathbb{E}_{\Theta }$
such that $\mathcal{O}\subset \mathcal{M}$.
\end{proposicao}
\begin{profe}
Let $\mu _{\Theta _{\mathrm{Ly}}}$ the only attractor measure in $\mathbb{E}%
_{\Theta _{\mathrm{Ly}}}$ and $\mu _{\Theta _{\mathrm{Ly}}^{\ast }}$ the
repeller measure in $\mathbb{E}_{\Theta _{\mathrm{Ly}}^{\ast }}$. These are
projections of attractor and repeller measures on $\mathbb{E}$. Hence the
above lemmas imply that $\mathrm{supp}\mu _{\Theta _{\mathrm{Ly}}}\subset
\mathcal{M}_{\Theta _{\mathrm{Ly}}}^{+}$ and $\mathrm{supp}\mu _{\Theta _{%
\mathrm{Ly}}^{\ast }}\subset \mathcal{M}_{\Theta _{\mathrm{Ly}}^{\ast }}^{+}$%
. However we checked in Section \ref{seclyapmorse} that $\Theta _{\mathrm{Ly}%
}\subset \Theta _{\mathrm{Mo}}$. Hence by Lemma \ref{lemfixinclu}, we
conclude that the fixed point set -- in any flag bundle -- of the section
Oseledet section $\chi _{\mathrm{Ly}}$ is contained in the fixed point set
of the Morse section $\chi _{\mathrm{Mo}}$. This means that the Oseletet
components are contained in the Morse components.
\end{profe}
\vspace{12pt}%
\noindent \textbf{Remark:} It is proved in \cite{ck}, Corollary 5.5.17, that
the Oseledets decomposition is contained in the Morse decomposition for a
linear flow on a vector bundle.
\section{Lyapunov exponents in the tangent bundle $T^{f}\mathbb{E}_{\Theta _{%
\mathrm{Ly}}}$}
\label{lyextgbun}
A fiber of a flag bundle $\mathbb{E}_{\Theta }$ is differentiable manifold
and hence has a tangent bundle. Gluing together the tangent bundles to the
fibers of $\mathbb{E}_{\Theta }$ we get a vector bundle $T^{f}\mathbb{E}%
_{\Theta }\rightarrow \mathbb{E}_{\Theta }$ over $\mathbb{E}_{\Theta }$. The
flow $\phi _{t}$ on $\mathbb{E}_{\Theta }$ is differentiable along the
fibers with differential map $\psi _{t}$, a linear map of the vector bundle $%
T^{f}\mathbb{E}_{\Theta }$. (See \cite{conleyflags} for a construction of
this vector bundle as an associated bundle $Q\times _{G}V$.)
We look here at the Lyapunov exponents for the linear flow $\psi _{t}$ on $%
T^{f}\mathbb{E}_{\Theta _{\mathrm{Ly}}}$ with respect to an attractor
measure of $\phi _{t}$. These Lyapunov exponents will be used in the proof
of the main technical lemma to describe the $\omega $-limit sets (w.r.t. $%
\phi _{t}$) in the flag bundles.
Equip the bundle $T^{f}\mathbb{E}_{\Theta }$ with a Riemannian metric $%
\langle \cdot ,\cdot \rangle $, which can be built from a $K$-reduction $R$
of the principal bundle $Q$. (Roughly, the metric $\langle \cdot ,\cdot
\rangle $ is constructed by piecing together $K$-invariant metrics on the
fibers. See \cite{conleyflags} for details.)
\begin{proposicao}
Let $\nu $ be an ergodic measure on $X$ and denote by $\mu $ its attractor
measure on the bundle $\mathbb{E}_{\Theta _{\mathrm{Ly}}}$. Let $H\left( \nu
\right) \in \mathrm{cl}\mathfrak{a}^{+}$ be the polar exponent of $\nu $.
Then the Lyapunov spectrum of $\psi $ w.r.t. $\mu $ is $\mathrm{ad}\left(
H\left( \mu \right) \right) _{\left\vert \mathfrak{n}_{\Theta _{\mathrm{Ly}%
}}^{-}\right. }$, which is a diagonal linear map of $\mathfrak{n}_{\Theta _{%
\mathrm{Ly}}}^{-}$ (that is an element of a Weyl chamber $\mathrm{cl}{%
\mathfrak{a}}^{+}$ of $\mathfrak{gl}\left( \mathfrak{n}_{\Theta _{\mathrm{Ly}%
}}^{-}\right) $).
\end{proposicao}
\begin{profe}
Denote by $\mathcal{O}\left( \nu \right) $ the $Z_{H_{\mathrm{Ly}}}$%
-measurable reduction of $Q$, corresponding to the Oseledets section of $\nu
$. This reduction is a principal bundle with structural group $Z_{H_{\mathrm{%
Ly}}}$ over a set $\Omega \subset X$ with $\nu \left( \Omega \right) =1$
(see \cite{alvsm}).
The section $\xi _{\mathrm{Ly}}:\Omega \rightarrow \mathbb{E}_{\Theta _{%
\mathrm{Ly}}}$ gives a desintegration of $\mu $ with respect to $\nu $ by
Dirac measures. Let $\Omega ^{\#}$ be the image of this section. Then the
restriction of $T^{f}\mathbb{E}_{\Theta _{\mathrm{Ly}}}$ to $\Omega ^{\#}$
is a vector bundle $T^{f}\Omega ^{\#}\rightarrow \Omega ^{\#}$, which is
invariant by the differential flow $\psi _{t}$.
We can build the vector bundle $T^{f}\Omega ^{\#}$ as an associated bundle $%
\mathcal{O}\left( \nu \right) $ through the adjoint representation $\theta $
of $Z_{H_{\mathrm{Ly}}}$ in $\mathfrak{n}_{\Theta _{\mathrm{Ly}}}^{-}$.
Then if we take compatible Cartan decompositions of $\theta \left( Z_{H_{%
\mathrm{Ly}}}\right) $ and $\mathrm{Gl}\left( \mathfrak{n}_{\Theta _{\mathrm{%
Ly}}}^{-}\right) $ it follows that the polar exponent of $\psi _{t}$ is
precisely $\theta \left( H\left( \mu \right) \right) $. By the constructions
of Section 8 of \cite{alvsm} it follows that the Lyapunov exponents of $\psi
_{t}$ are the eigenvalues of $\theta \left( H\left( \mu \right) \right) $ as
linear maps of $\mathfrak{n}_{\Theta _{\mathrm{Ly}}}^{-}$.
\end{profe}
\begin{corolario}
\label{corlyapvertneg}Suppose $\Theta _{\mathrm{Ly}}\subset \Theta $. Then
the Lyapunov exponents of $\psi _{t}$ in $T^{f}\mathbb{E}_{\Theta }$ with
respect to the attractor measure $\mu $ are strictly negative.
\end{corolario}
Later on we will combine this corollary with the following general fact
about Lyapunov exponents on vector bundles. Let $p:V\rightarrow X$ be a
continuous vector bundle endowed with a norm $\left\vert \left\vert \cdot
\right\vert \right\vert $. Let $\Phi _{n}$ be a continuous linear flow on $V$%
. If $\nu $ is a $\Phi $-invariant ergodic measure on the base $X$ then $%
\Phi $ has a Lyapunov spectrum $H_{\mathrm{Ly}}\left( \nu \right) =\{\lambda
_{1}\geq \cdots \geq \lambda _{n}\}$ with respect to $\nu $, as ensured by
the Multiplicative Ergodic Theorem. The following lemma may be well known.
For the sake of completeness we prove it here using the Morse spectrum of
the linear flow.
\begin{lema}
\label{lemfibradovetorial}Supppose that for every $\Phi $-invariant ergodic
measure $\nu $ on $X$ the spectrum with respect to $\nu $ is strictly
negative. Then for every $v\in V$,
\[
\lim_{n\rightarrow +\infty }\left| \left| \Phi _{n}v\right| \right| =0.
\]
\end{lema}
\begin{profe}
Let $\mathbb{P}V\rightarrow X$ be the projective bundle of $V$. The cocycle $%
\rho \left( n,v\right) =\frac{\left\vert \left\vert \Phi _{n}v\right\vert
\right\vert }{\left\vert \left\vert v\right\vert \right\vert }$ on $V$
induces the additive cocycle $a\left( n,\eta \right) =\log \rho \left(
n,\eta \right) $, $\eta \in \mathbb{P}V$, whose asymptotics give the
Lyapunov spectrum of $\Phi $. Write $q\left( \cdot \right) =a\left( 1,\cdot
\right) $. Then by general results on Morse spectrum of an additive cocycle
(see \cite{smsec}, Section 3, and references therein), the Morse spectrum of
$a$ is a union of intervals whose extreme points are integrals $\int q\left(
x\right) \mu \left( dx\right) $ with respect to ergodic invariant measures $%
\mu $ for the flow on $\mathbb{P}V$. By the Birkhoff ergodic theorem it
follows that for $\mu $-almost all $\eta \in \mathbb{P}V$,
\[
\lim_{n\rightarrow \infty }\frac{1}{n}a\left( n,\eta \right)
=\lim_{n\rightarrow \infty }\frac{1}{n}\sum_{k=0}^{n-1}q\left( \Phi _{k}\eta
\right) =\int q\left( z\right) \mu \left( dz\right) .
\]%
On the other hand the projection $p_{\ast }\mu =\nu $ is ergodic on the base
$X$. Hence by assumption the spectrum with respect to $\nu $, given by the
multiplicative ergodic theorem is strictly negative. This means that for $%
\nu $-allmost all $x\in X$, $\lim_{n\rightarrow \infty }\frac{1}{n}a\left(
n,\eta \right) $ exists for every $\eta \in p^{-1}\{x\}$ and is strictly
negative. Combining these two facts we conclude that
\[
\int q\left( z\right) \mu \left( dz\right) <0,
\]%
and therefore the Morse spectrum is contained in $\left( -\infty ,0\right) $%
.
Now, for every $\eta \in \mathbb{P}V$, $\lim \sup_{n\rightarrow +\infty }%
\frac{1}{n}a\left( n,\eta \right) $ belongs to the Morse spectrum (see \cite%
{ck}, Theorem 5.3.6). Hence for every $0\neq v\in V$,
\[
\lim \sup_{n\rightarrow +\infty }\frac{1}{n}\log \left\vert \left\vert \Phi
_{n}v\right\vert \right\vert <0.
\]%
This implies that for large $n$, $\left\vert \left\vert \Phi
_{n}v\right\vert \right\vert <e^{cn}$, $c<0$, proving the lemma.
\end{profe}
Applying the lemma for the backward flow we have
\begin{corolario}
\label{corfibradovetorialneg}With the same assumptions of the lemma we have $%
\lim_{n\rightarrow -\infty }\left\vert \left\vert \Phi _{n}v\right\vert
\right\vert =\infty $ se $v\neq 0$.
\end{corolario}
\section{Main technical lemma}
\label{teclemma}
\begin{lema}
\label{lemmaintech}Let $\mathbb{E}_{\Theta }$ be a flag manifold with dual $%
\mathbb{E}_{\Theta ^{\ast }}$. Suppose there are three compact $\phi $%
-invariant subsets $A,B\subset \mathbb{E}_{\Theta }$ and $C\subset \mathbb{E}%
_{\Theta ^{\ast }}$ that project onto $X$ and such that
\begin{enumerate}
\item $A\cap B=\emptyset $.
\item $B^{c}$ is the set of elements transversal to $C$. (That is an element
$v\in \mathbb{E}_{\Theta }$ belongs to $B$ if and only if it is not
transversal to some $w\in C$ in the same fiber as $v$.)
\item For any ergodic measure $\rho $ for the flow on the base space $X$ we
have $\Theta _{\mathrm{Ly}}\left( \rho \right) \subset \Theta $. By
Corollary \ref{coruniquemedatr}, this implies that there is a unique
attractor measure $\mu _{\Theta }^{+}\left( \rho \right) $ for $\rho $ on $%
\mathbb{E}_{\Theta }$ and a unique repeller measure $\mu _{\Theta ^{\ast
}}^{-}\left( \rho \right) $ on $\mathbb{E}_{\Theta ^{\ast }}$.
\item For any ergodic measure $\rho $ on $X$, $\mathrm{supp}\mu _{\Theta
}^{+}\left( \rho \right) \subset A$ and $\mathrm{supp}\mu _{\Theta ^{\ast
}}^{-}\left( \rho \right) \subset C$.
\end{enumerate}
Then $\left( A,B\right) $ is an attractor-repeller pair on $\mathbb{E}%
_{\Theta }$. That is, the $\omega $-limit $\omega \left( v\right) \subset A$
if $v\notin B$ and $\omega ^{\ast }\left( v\right) \subset B$ if $v\notin A$.
\end{lema}
The proof of this lemma will be done in several steps. Before starting we
define a fourth set $D\subset Q\times _{G}\left( \mathbb{F}_{\Theta }\times
\mathbb{F}_{\Theta ^{\ast }}\right) $ by
\[
D=\pi _{1}^{-1}\left( A\right) \cap \pi _{2}^{-1}\left( C\right)
\]%
where $\pi _{1}:Q\times _{G}\left( \mathbb{F}_{\Theta }\times \mathbb{F}%
_{\Theta ^{\ast }}\right) \rightarrow \mathbb{E}_{\Theta }$ and $\pi
_{2}:Q\times _{G}\left( \mathbb{F}_{\Theta }\times \mathbb{F}_{\Theta ^{\ast
}}\right) \rightarrow \mathbb{E}_{\Theta ^{\ast }}$ are the projections.
This set is compact and invariant, and by the transversality given by the
first and second conditions in the lemma we can view $D$ as a compact subset
of the bundle
\[
\mathcal{A}_{\Theta }=Q\times _{G}\mathrm{Ad}\left( G\right) H_{\Theta }
\]%
where $\Theta =\{\alpha \in \Sigma :\alpha \left( H_{\Theta }\right) =0\}$.
Now, to start the proof fix $x\in X$ that has a periodic orbit $\mathcal{O}%
\left( x\right) $. Then the Oseledets decomposition coincides with the Morse
decomposition above $\mathcal{O}\left( x\right) $, which by Section \ref%
{secperflow} is built from the dynamics of the action of a $g_{x}\in G$.
Clearly the homogeneous measure $\theta $ on the periodic orbit is an
ergodic invariant measure on $X$. By the third condition of the lemma $g_{x}
$ has one attractor fixed point at $\mathbb{F}_{\Theta }$, say $b^{+}$, and
a repeller fixed point $b^{-}\in \mathbb{F}_{\Theta ^{\ast }}$. The Morse
decomposition of $g_{x}$ is the union of $\{b^{+}\}$ with subsets whose
elements are not transversal to $b^{-}$. It follows that the attractor $\mu
_{\Theta }^{+}\left( \theta \right) $ and the repeller measures $\mu
_{\Theta ^{\ast }}^{-}\left( \theta \right) $ are homogeneous measures on
periodic orbits. By the fourth condition of the lemma $\mathrm{supp}\mu
_{\Theta }^{+}\left( \theta \right) \subset A$ and $\mathrm{supp}\mu
_{\Theta ^{\ast }}^{-}\left( \theta \right) \subset C$, so that by the
second condition the Morse decomposition above $\mathcal{O}\left( x\right) $
is the union of $\mathrm{supp}\mu _{\Theta }^{+}\left( \theta \right) $ with
subsets contained in $B$. Hence, if $v$ is in the fiber above $x$ then $%
\omega \left( v\right) \subset A$ if $v\notin B$ while $\omega ^{\ast
}\left( v\right) \subset B$ if $v\notin A$.
Therefore, from now on we look at $\omega $-limits $\omega \left( v\right) $
and $\omega ^{\ast }\left( v\right) $ assuming that the orbit $\mathcal{O}%
\left( x\right) $ of $x=\pi \left( v\right) $ is not periodic, that is, the
map $n\in \mathbb{Z}\mapsto x_{n}=\phi _{n}\left( x\right) \in X$ is
injective and $\mathcal{O}\left( x\right) $ is in bijection with $\mathbb{Z}$%
.
To prove that the $\omega $-limits are contained in $A$ we will use the
following consequence of Lemma \ref{lemfibradovetorial}.
\begin{lema}
Given $v\in A$ let $w\in T_{v}^{f}\mathbb{E}_{\Theta _{\mathrm{Ly}}}$ be a
tangent vector at $v$. Then $\lim_{t\rightarrow +\infty }\left| \left| \psi
_{t}w\right| \right| =0$.
\end{lema}
\begin{profe}
Is an immediate consequence of Lemma \ref{lemfibradovetorial}, combined with
the third and fourth conditions of the Lemma.
\end{profe}
Now, above the nonperiodic orbit $\mathcal{O}\left( x\right) $ we reduce the
flow to just a sequence $g_{n}$ of elements of the subgroup $Z_{H_{\Theta }}$%
. The construction is the following: Start with an element $\eta _{0}\in D$
in the fiber over $x$. The orbit $\mathcal{O}\left( \eta \right) $ is the
sequence $\eta _{n}=\phi _{n}\left( \eta _{0}\right) $, $n\in \mathbb{Z}$,
that can be viewed as a section over $\mathcal{O}\left( x\right) $ by $%
x_{n}\mapsto \eta _{n}$. The elements of the associated bundle $\mathcal{A}%
_{\Theta }$ are written as $p\cdot H_{\Theta }$, $p\in R$, where as before $%
R $ is the $K$-reduction of $Q\rightarrow X$. Hence there exists a sequence $%
p_{n}\in R$ such that $\eta _{n}=p_{n}\cdot H_{\Theta }$.
Since $p_{n+m}$ and $\phi _{n}\left( p_{m}\right) $ are in the same fiber,
we have $\phi _{n}\left( p_{m}\right) =p_{n+m}\cdot g_{n,m}$ with $%
g_{n,m}\in G$, $n,m\in \mathbb{Z}$. Actually, $g_{n,m}\in Z_{H_{\Theta }}$
because $\phi _{n}\left( \eta _{m}\right) =\eta _{n+m}$, so that
\[
p_{n+m}\cdot \mathrm{Ad}\left( g_{n,m}\right) H_{\Theta }=\phi _{n}\left(
\eta _{m}\right) =\eta _{n+m}=p_{n+m}\cdot H_{\Theta }.
\]
We write $\xi _{n}$ and $\xi _{n}^{*}$ for the projections of $\eta _{n}$
into $\mathbb{E}_{\Theta }$ and $\mathbb{E}_{\Theta ^{*}}$, respectively. By
definition of $D$ we have $\xi _{n}\in A$ and $\xi _{n}^{*}\in C$. Hence by
the second assumption of the lemma the elements in $\mathbb{E}_{\Theta }$
that are not transversal to $\xi _{n}^{*}$ are contained in $B$. In other
words
\begin{lema}
Take $v\notin B$ in the fiber of $x$. Then $v$ is transversal to $\xi
_{0}^{*}$.
\end{lema}
Now we use Lyapunov exponents of the lifting $\psi _{n}$ of $\phi _{n}$ to $%
T^{f}\mathbb{E}_{\Theta }$ to show that $\omega \left( v\right) \subset A$
if $v\notin B$ is in the fiber of $x$.
To do that we first note that if the starting element $\eta _{0}\in \mathcal{%
A}_{\Theta }$ is written as $\eta _{0}=p\cdot H_{\Theta }$, $p\in R$, then
the set of points that are transversal to $\xi _{0}^{*}$ is given
algebraically by
\[
T=p\cdot \left( N_{\Theta }^{-}\cdot b_{0}\right) =\{p\cdot nb_{0}:n\in
N_{\Theta }^{-}\}
\]
where $b_{0}$ is the origin of the flag $\mathbb{F}_{\Theta }$ and $%
N_{\Theta }^{-}$ is the nilpotent subgroup with Lie algebra $\mathfrak{n}%
_{\Theta }^{-}=\sum_{\alpha \notin \langle \Theta \rangle ,\alpha <0}%
\mathfrak{g}_{\alpha }$ (lower triangular matrices).
Since $\exp :\mathfrak{n}_{\Theta }^{-}\rightarrow N_{\Theta }^{-}$ is a
diffeomorphism we have also $T=\{p\cdot \left( \exp Y\cdot b_{0}\right)
:Y\in \mathfrak{n}_{\Theta }^{-}\}$.
The action of $\phi _{n}$ on an element $p\cdot \left( \exp Y\cdot
b_{0}\right) \in T$ is given as follows: Put $g_{n}=g_{n,0}\in Z_{H_{\Theta
}}$. Then, as remarked above, $\phi _{n}\left( p\right) =p_{n}\cdot g_{n}$,
so that
\[
\phi _{n}\left( p\cdot \left( \exp Y\cdot b_{0}\right) \right) =p_{n}\cdot
\left( g_{n}\exp Yg_{n}^{-1}\cdot g_{n}b_{0}\right) .
\]
But $g_{n}b_{0}=b_{0}$ because $g_{n}\in Z_{H}$, and $g_{n}\exp
Yg_{n}^{-1}=\exp \left( \mathrm{Ad}\left( g_{n}\right) Y\right) $. Hence
\begin{equation}
\phi _{n}\left( p\cdot \left( \exp Y\cdot b_{0}\right) \right) =p_{n}\cdot
\exp \left( \mathrm{Ad}\left( g_{n}\right) Y\right) b_{0}. \label{forfinexp}
\end{equation}
The next lemma relates this action with the lifting $\psi _{n}$ of $\phi
_{n} $ to the tangent space $T^{f}\mathbb{E}_{\Theta }$.
\begin{lema}
\label{lemlimzerotangfibr} Given $Y\in \mathfrak{g}$, denote by $p \cdot Y$
the vertical tangent vector \linebreak $\frac{d}{dt}\left( p\cdot \left(
\exp tY\cdot b_{0}\right) \right) _{t=0}\in T_{p\cdot b_{0}}^{f}\mathbb{E}%
_{\Theta }$. Then $p\cdot Y$, $Y\in \mathfrak{n}_{\Theta }^{-}$, fulfill the
vertical tangent space $T_{p\cdot b_{0}}^{f}\mathbb{E}_{\Theta }$, and the
derivative $\psi _{n}$ of $\phi _{n}$ at $p\cdot b_{0}$ satisfies
\[
\psi _{n}\left( p\cdot Y\right) =p_{n}\cdot \mathrm{Ad}\left( g_{n}\right)
Y.
\]
\end{lema}
\begin{profe}
The fact that any tangent vector in $T_{p\cdot b_{0}}^{f}\mathbb{E}_{\Theta }
$ is given by $p\cdot Y$ for some $Y\in \mathfrak{n}_{\Theta }^{-}$ is due
to the fact that $N_{\Theta }^{-}\cdot b_{0}$ is an open submanifold of $%
\mathbb{F}_{\Theta }$. For the last statement we have
\begin{eqnarray*}
\psi _{n}\left( p\cdot Y\right) &=&\frac{d}{dt}\phi _{n}\left( p\cdot
\left( \exp tY\cdot b_{0}\right) \right) _{t=0} \\
&=&p_{n}\cdot \frac{d}{dt}\left( \exp \left( t\mathrm{Ad}\left( g_{n}\right)
Y\right) b_{0}\right) _{t=0}=p_{n}\cdot \mathrm{Ad}\left( g_{n}\right) Y.
\end{eqnarray*}
\end{profe}
We are now prepared to prove that $\omega \left( v\right) \subset A$ if $%
v\notin B$ is in the fiber of $x$. We have $v=p\cdot \left( \exp Y\cdot
b_{0}\right) $ for some $Y\in \mathfrak{n}_{\Theta }^{-}$, so that by (\ref%
{forfinexp}) $\phi _{n}\left( v\right) =p_{n}\cdot \exp \left( \mathrm{Ad}%
\left( g_{n}\right) Y\right) b_{0}$.
Now by Corollary \ref{corlyapvertneg} and \ Lemma \ref{lemfibradovetorial}
we have $\lim_{n\rightarrow +\infty }\left\vert \left\vert \psi
_{n}w\right\vert \right\vert =0$ if $w\in T_{p\cdot b_{0}}^{f}\mathbb{E}%
_{\Theta }$. Taking $w=p\cdot Y$ we have $\left\vert \left\vert p\cdot
Y\right\vert \right\vert =\left\vert \left\vert Y\right\vert \right\vert $
because $p\in R$ and hence is an isometry between the flag manifold $\mathbb{%
F}_{\Theta }$ and the corresponding fiber of $\mathbb{E}_{\Theta }$ (see
\cite{conleyflags} for the construction of the norm in $T^{f}\mathbb{E}%
_{\Theta }$). Since the same remark holds for $p_{n}\in R$ we have $%
\left\vert \left\vert \psi _{n}w\right\vert \right\vert =\left\vert
\left\vert \mathrm{Ad}\left( g_{n}\right) Y\right\vert \right\vert $, so
that
\[
\mathrm{Ad}\left( g_{n}\right) Y\rightarrow 0\qquad \mathrm{and}\qquad \exp
\mathrm{Ad}\left( g_{n}\right) Y\rightarrow 1.
\]
This implies that if $d$ is the metric on $\mathbb{E}_{\Theta _{\mathrm{Ly}%
}} $ then $d\left( \phi _{n}\left( v\right) ,p_{n}\cdot b_{0}\right)
\rightarrow 0$. But $p_{n}\cdot b_{0}=\xi _{n} \in A$ as well as its limit
points, by invariance and compactness of $A$. Therefore we conclude that $%
\omega \left( v\right) \subset A$, if $v\notin B$.
We turn now to the proof that $\omega ^{*}\left( v\right) \subset B$ if $%
v\notin A$. Again with $v$ above a nonperiodic orbit.
Take a sequence $n_{k}\rightarrow -\infty $ such that $\phi _{n_{k}}v$
converges in $\mathbb{E}_{\Theta }$. Taking subsequences we assume the
convergences $p_{n_{k}}\rightarrow p\in R$, $\eta _{n_{k}}\rightarrow \eta $%
, $\xi _{n_{k}}\rightarrow \xi $ and $\xi _{n_{k}}^{*}\rightarrow \xi ^{*}$.
By invariance it is enough to take $v\notin B$, so that we can write $%
v=p_{0}\cdot \left( \exp Y\right) b_{0}$ with $Y\in \mathfrak{n}_{\Theta
}^{-}$ and $Y\neq 0$ (because $v\notin A$).
Taking subsequences again we assume that $g_{n_{k}}$ $\left( \exp Y\right)
b_{0}$ converges to $b_{1}\in \mathbb{F}_{\Theta }$. Since $g_{n}\in
Z_{H_{\Theta }}$, we have $g_{n}b_{0}=b_{0}$ and hence
\[
\left( \exp \mathrm{Ad}\left( g_{n_{k}}\right) Y\right)
b_{0}=g_{n_{k}}\left( \exp Y\right) b_{0}\rightarrow b_{1}.
\]
Now $\mathrm{Ad}\left( g_{n}\right) Y\rightarrow \infty $ in $\mathfrak{n}%
_{\Theta _{\mathrm{Ly}}}^{-}$, because the Lyapunov exponents for the
backward flow are $>0$. This implies that $b_{1}$ is not transversal to the
origin $b_{0}^{*}$ of $\mathbb{F}_{\Theta ^{*}}$. Hence, $p_{n_{k}}\cdot
b_{1}$ is not transversal to $p_{n_{k}}\cdot b_{0}^{*}$, so that $%
p_{n_{k}}\cdot b_{1}\in B$. But
\[
\phi _{n_k}v=p_{n_{k}}g_{n_{k}}\cdot \left( \exp Y\right)
b_{0}=p_{n_{k}}\cdot \left( \exp \mathrm{Ad}\left( g_{n_{k}}\right) Y\right)
b_{0}
\]
so that $\lim \phi _{n_k}v=p\cdot b_{1}$, showing that $\omega ^{*}\left(
v\right) \subset B$.
In conclusion we have compact invariant sets $A$ and $B$ that satisfy $%
\omega \left( v\right) \subset A$ and $\omega ^{*}\left( v\right) \subset B$
if $v\notin A\cup B$. Hence $A$ and $B$ define a Morse decomposition of $%
\mathbb{E}_{\Theta _{\mathrm{Ly}}}$ with $A$ the attractor component.
\section{Three conditions}
\label{threecon}
In this section we state three conditions that together are necessary and
sufficient to have equality between the Lyapunov and Morse decompositions
over an ergodic invariant measure.
Thus as in Section \ref{seclyapmorse}, let $\phi _{n}$ be a continuous
flow on a continuous principal bundle $\pi :Q\rightarrow X$ whose the
structural group $G$ is reductive and noncompact. We fix once and for all an
ergodic invariant measure $\nu $ on the base space having support $\mathrm{%
supp}\nu =X$. Then the $\mathfrak{a}$-Lyapunov exponents of $\phi _{n}$
select a flag type, which is expressed by a subset $\Theta _{\mathrm{Ly}}$
of simple roots. The flow on $X$ is chain transitive so it also has a flag
type $\Theta _{\mathrm{Mo}}$ coming from the Morse decompositions and $%
\mathfrak{a}$-Morse spectrum.
We start by writting down the three conditions and check that they are
necessary. In the next section we prove that together they are also
sufficient to have $\Theta _{\mathrm{Ly}}=\Theta _{\mathrm{Mo}}$.
\subsection{Bounded section\label{cond1}}
The Oseledets' section is a measurable section $\chi _{\mathrm{Ly}}:\Omega
\rightarrow Q\times _{G}\mathcal{O}_{\mathrm{Ly}}$ of the associated bundle $%
Q\times _{G}\mathcal{O}_{\mathrm{Ly}}\rightarrow X$ above the set of full $%
\nu $-measure $\Omega $. The fiber of this bundle is the adjoint orbit $%
\mathcal{O}_{\mathrm{Ly}}=\mathrm{Ad}\left( G\right) H_{\mathrm{Ly}}$. This
section can be seen as an equivariant map $f_{\mathrm{Ly}}:Q_{\Omega
}\rightarrow \mathcal{O}_{\mathrm{Ly}}$ defined above $\Omega $, where $%
Q_{\Omega }=\pi ^{-1}\left( \Omega \right) $ and $\pi :Q\rightarrow X$ is
the projection.
Let $R\subset Q\rightarrow X$ be a (continuous) $K$-reduction of $Q$. Then
we say that the Oseledets' section is \textit{bounded} if
\begin{itemize}
\item $f_{\mathrm{Ly}}$ is bounded in $R_{\Omega }$.
\end{itemize}
This definition does not depend on the specific $K$-reduction because the
base space $X$ is assumed to be compact.
If $\Theta _{\mathrm{Ly}}=\Theta _{\mathrm{Mo}}$ then we can take $H_{%
\mathrm{Mo}}=H _{\mathrm{Ly}}$ and $\chi _{\mathrm{Mo}}=\chi _{\mathrm{Ly}}$%
. So that $f_{\mathrm{Ly}}$ is continuous and hence bounded by compactness.
Hence boundedness is a necessary condition.
\vspace{12pt}%
\noindent%
\textbf{Example:} Let $\phi $ be a linear flow on a $d$-dimensional trivial
vector bundle $X\times V$ with two Lyapunov exponents $\lambda _{1}>\lambda
_{2}$ whose Oseledets subspaces have dimension $k$ and $d-k$. Then $H_{%
\mathrm{Ly}}$ is the diagonal matrix $\mathrm{diag}\{\lambda _{1},\ldots
,\lambda _{1},\lambda _{2},\ldots ,\lambda _{2}\}$ with $\lambda _{1}$
having multiplicity $k$. The Oseledets section is given by a map $f_{\mathrm{%
Ly}}:X\rightarrow \mathrm{Ad}\left( G\right) H_{\mathrm{Ly}}$ such that the
Oseledets subspaces at $x\in X$ are the eigenspaces $V_{\lambda _{1}}\left(
x\right) $ and $V_{\lambda _{2}}\left( x\right) $ of $f_{\mathrm{Ly}}\left(
x\right) $. To say that $f_{\mathrm{Ly}}$ is bounded means that the
subspaces $V_{\lambda _{1}}\left( x\right) $ and $V_{\lambda _{2}}\left(
x\right) $ have a positive distance.
\subsection{Refinement of the Lyapunov spectrum of other invariant measures
\label{cond2}}
Denote by $\mathcal{P}_{X}\left( \phi \right) $ the set of $\phi $-invariant
probability measures on $X$. For each ergodic measure $\rho \in \mathcal{P}%
_{X}\left( \phi \right) $ we have its Lyapunov spectrum $H_{\mathrm{Ly}%
}\left( \rho \right) $ and the corresponding flag type $\Theta _{\mathrm{Ly}%
}\left( \rho \right) =\{\alpha \in \Sigma :\alpha \left( H_{\mathrm{Ly}%
}\left( \rho \right) \right) =0\}$. Our second condition is
\begin{itemize}
\item $\Theta _{\mathrm{Ly}}\left( \rho \right) \subset \Theta _{\mathrm{Ly}%
}\left( \nu \right) $ for every ergodic $\rho \in \mathcal{P}_{X}\left( \phi
\right) $.
\end{itemize}
This is a necessary condition for $\Theta _{\mathrm{Ly}}\left( \nu \right)
=\Theta _{\mathrm{Mo}}$. To see this let $\rho \in \mathcal{P}_{X}\left(
\phi \right) $ be ergodic and denote by $Y\subset X$ its support. Let $%
\Theta _{\mathrm{Mo}}\left( Y\right) $ be the flag type of the Morse
decomposition of the flow restricted to $Y$ (that is, to the fibers above $Y$%
). We have $\Theta _{\mathrm{Mo}}\left( Y\right) \subset \Theta _{\mathrm{Mo}%
}\left( \nu \right) $ because the Morse components of the flow restricted to
$Y$ are contained in the components over $X$. However, $\Theta _{\mathrm{Ly}%
}\left( \rho \right) \subset \Theta _{\mathrm{Mo}}\left( Y\right) $, so that
the equality $\Theta _{\mathrm{Mo}}=\Theta _{\mathrm{Ly}}\left( \nu \right) $
implies
\[
\Theta _{\mathrm{Ly}}\left( \rho \right) \subset \Theta _{\mathrm{Mo}}\left(
Y\right) \subset \Theta _{\mathrm{Mo}}=\Theta _{\mathrm{Ly}}\left( \nu
\right) .
\]
\vspace{12pt}%
\noindent%
\textbf{Example:} Let $\phi $ be a linear flow on a $d$-dimensional vector
bundle $X\times V$ with Lyapunov spectrum $\lambda _{1}>\cdots >\lambda _{s}$
with multiplicities $k_{1},\ldots ,k_{s}$ with respect to $\nu $. Then this
condition means that the Lyapunov spectrum $\mu _{1}>\cdots >\mu _{t}$ with
respect to another ergodic measure $\rho $ have multiplicities $r_{1},\ldots
,r_{t}$ that satisfy $r_{1}+\cdots +r_{i_{1}}=k_{1}$, $r_{i_{1}}+\cdots
+r_{i_{2}}=k_{2}$, etc.
\subsection{Attracting and repeller measures\label{cond3}}
Recall that we defined an attractor measure on a partial flag manifold $%
\mathbb{E}_{\Theta }$ to be the projection of an ergodic invariant measure $%
\mu $ on $\mathbb{E}$ such that
\[
H_{\mathrm{Ly}}\left( \mu \right) =\int qd\mu \in \mathrm{cl}\mathfrak{a}%
^{+}.
\]
In the specific flag bundle $\mathbb{E}_{\Theta _{\mathrm{Ly}}}$ the
attractor measure $\mu _{\Theta _{\mathrm{Ly}}}^{+}$ is unique and has a
desintegration over $\nu $ by Dirac measures on the fibers above a set $%
\Omega $ of total measure $\nu $. We denote by $\mathrm{att}_{\Theta _{%
\mathrm{Ly}}}\left( \nu \right) $ the support of $\mu _{\Theta _{\mathrm{Ly}%
}}^{+}$.
Analogously in the dual flag bundle $\mathbb{E}_{\Theta _{\mathrm{Ly}}^{*}}$
there is a unique repeller measure $\mu _{\Theta _{\mathrm{Ly}}}^{-}$. We
denote by $\mathrm{rep}_{\Theta _{\mathrm{Ly}}^{*}}\left( \nu \right) $ the
support of $\mu _{\Theta _{\mathrm{Ly}}}^{-}$.
Let $\rho \in \mathcal{P}_{X}\left( \phi \right) $ be an ergodic measure
with support $Y\subset X$, which is an invariant subset. The subset $\pi
_{\Theta _{\mathrm{Ly}}}^{-1}\left( Y\right) \cap \mathrm{att}_{\Theta _{%
\mathrm{Ly}}}\left( \nu \right) $ is invariant as well. We denote by $%
\mathcal{E}_{\Theta _{\mathrm{Ly}}}^{+}\left( \rho \right) $ the set of
ergodic probability measures with support contained in $\pi _{\Theta _{%
\mathrm{Ly}}}^{-1}\left( Y\right) \cap \mathrm{att}_{\Theta _{\mathrm{Ly}%
}}\left( \nu \right) $ that project down to $\rho $. Also, we put $\mathcal{E%
}_{\Theta _{\mathrm{Ly}}^{*}}^{-}\left( \rho \right) $ for the set of
ergodic probability measures with support in $\pi _{\Theta _{\mathrm{Ly}%
}}^{-1}\left( Y\right) \cap \mathrm{rep}_{\Theta^{*}_{\mathrm{Ly}}}\left(
\nu \right) $ that project down to $\rho $. Both sets $\mathcal{E}_{\Theta _{%
\mathrm{Ly}}}^{+}\left( \rho \right) $ and $\mathcal{E}_{\Theta _{\mathrm{Ly}%
}^{*}}^{-}\left( \rho \right) $ are not empty.
Now we can state our third condition.
\begin{itemize}
\item Any $\theta \in \mathcal{E}_{\Theta _{\mathrm{Ly}}}^{+}\left( \rho
\right) $ is an attractor measure and any $\theta \in \mathcal{E}_{\Theta _{%
\mathrm{Ly}}^{*}}^{-}\left( \rho \right) $ is a reppeller measure for $\phi $%
.
\end{itemize}
If $\Theta _{\mathrm{Mo}}=\Theta _{\mathrm{Ly}}\left( \nu \right) $ then the
attractor Morse component $\mathcal{M}_{\Theta _{\mathrm{Ly}}}=\mathcal{M}%
_{\Theta _{\mathrm{Ly}}}\left( 1\right) $ in $\mathbb{E}_{\Theta _{\mathrm{Ly%
}}}$ is the image of a section $\xi :X\rightarrow \mathbb{E}_{\Theta _{%
\mathrm{Ly}}}$ and contains the support \textrm{att}$_{\Theta _{\mathrm{Ly}%
}}\left( \nu \right) $ of the attractor measure. This implies that $\mathcal{%
M}_{\Theta _{\mathrm{Ly}}}=\mathrm{att}_{\Theta _{\mathrm{Ly}}}\left( \nu
\right) $, so that $\mathcal{M}=\pi _{\Theta _{\mathrm{Ly}}}^{-1}\left(
\mathrm{att}_{\Theta _{\mathrm{Ly}}}\left( \nu \right) \right) $ is the
attractor Morse component $\mathcal{M}$ in the maximal flag bundle. Now the
Morse spectrum $\Lambda _{\mathrm{Mo}}\left( \mathcal{M}\right) $ of $%
\mathcal{M}$ is contained in the cone
\[
\mathfrak{a}_{\Theta _{\mathrm{Mo}}}^{+}=\{H\in \mathfrak{a}:\forall \alpha
\notin \langle \Theta _{\mathrm{Mo}}\rangle ,\,\alpha \left( H\right) >0\}
\]%
(see \cite{smsec}). Hence any Lyapunov exponent of $\mathcal{M}$ belongs to $%
\mathfrak{a}_{\Theta _{\mathrm{Mo}}}^{+}$. By projecting down to $\mathbb{E}%
_{\Theta _{\mathrm{Ly}}}$ the measures with support in $\mathcal{M}$ we see
that any $\theta $ with support in $\pi _{\Theta _{\mathrm{Ly}}}^{-1}\left(
Y\right) \cap \mathrm{att}_{\Theta _{\mathrm{Ly}}}\left( \nu \right) $ is an
attractor measure.
The same proof with the backward flow shows that $\theta \in \mathcal{E}%
_{\Theta _{\mathrm{Ly}}^{\ast }}^{-}\left( \rho \right) $ is a reppeller
measure.
\subsection{Oseledets decompositions for other measures}
The second and third conditions above refer to ergodic measures $\rho $ on $%
X $ different from the initial measure $\nu $. These two conditions can be
summarized in just one condition on the Oseledets section for the ergodic
measures $\rho \in \mathcal{P}_{X}\left( \phi \right) $.
Given $\rho \in \mathcal{P}_{X}\left( \phi \right) $, write $\chi ^{\rho }$
for its Oseledets section and $\xi ^{\rho }$ and $\xi ^{\rho *}$ for the
corresponding sections on $\mathbb{E}_{\Theta _{\mathrm{Ly}}\left( \rho
\right) }$ and $\mathbb{E}_{\Theta _{\mathrm{Ly}}^{*}\left( \rho \right) }$,
respectively.
\begin{defi}
We say that $\chi ^{\rho }$ is contained in the Oseledets section of $\nu $
in case the two conditions are satisfied
\begin{enumerate}
\item $\Theta _{\mathrm{Ly}}\left( \rho \right) \subset \Theta _{\mathrm{Ly}%
}\left( \nu \right) $. In this case there is the fibration $p:Q\mathcal{O}%
_{\Theta _{\mathrm{Ly}}\left( \rho \right) }\rightarrow Q\mathcal{O}_{\Theta
_{\mathrm{Ly}}\left( \nu \right) }$.
\item $p\left( \mathrm{im}\chi ^{\rho }\right) \subset \mathrm{cl}\left(
\mathrm{im}\chi \right) $, where $\chi $ is the Oseledets section of $\nu $.
\end{enumerate}
\end{defi}
The second condition implies that the image of the sections $\xi ^{\rho }$
and $\xi ^{\rho *}$ project onto the $\mathrm{cl}\left( \mathrm{im}\xi
\right) $ and $\mathrm{cl}\left( \mathrm{im}\xi ^{*}\right) $, by the
fibrations $\mathbb{E}_{\Theta _{\mathrm{Ly}}\left( \rho \right)
}\rightarrow \mathbb{E}_{\Theta _{\mathrm{Ly}}\left( \nu \right) }$ and $%
\mathbb{E}_{\Theta _{\mathrm{Ly}}^{*}\left( \rho \right) }\rightarrow
\mathbb{E}_{\Theta _{\mathrm{Ly}}^{*}\left( \nu \right) }$, respectively.
Since the attractor and repeller measures for $\rho $ desintegrate according
to the sections $\xi ^{\rho }$ and $\xi ^{\rho *}$, respectively, it follows
that the second and third conditions above is equivalent to have $\chi
^{\rho }$ contained in $\chi $.
\section{Sufficience of the conditions}
\label{sufcon}
We apply here the main Lemma \ref{lemmaintech} to get sufficience of the
conditions of the last section and thus prove the following characterization
for the equality of Morse and Oseledets decompositions.
\begin{teorema}
\label{teomainequal}Suppose the invariant measure on the base space is
ergodic. Then the three conditions together --- bounded section (\ref{cond1}%
), refinement of Lyapunov spectrum (\ref{cond2}) and attractor-repeller
measures (\ref{cond3}) --- are necessary and sufficient to have $\Theta _{%
\mathrm{Ly}}=\Theta _{\mathrm{Mo}}$ and $\chi _{\mathrm{Ly}}=\chi _{\mathrm{%
Mo}}$.
\end{teorema}
As before we have the sections $\xi :\Omega \rightarrow \mathbb{E}_{\Theta _{%
\mathrm{Ly}}}$ and $\xi ^{*}:\Omega \rightarrow \mathbb{E}_{\Theta _{\mathrm{%
Ly}}^{*}}$, respectively, that are combined to give the Oseledets section $%
\chi _{\mathrm{Ly}}:\Omega \subset X\rightarrow \mathcal{A}_{\Theta _{%
\mathrm{Ly}}}$.
We apply Lemma \ref{lemmaintech} with
\begin{enumerate}
\item $A=\mathrm{cl}\left( \mathrm{im}\xi \right) $, which is the support of
the unique attrator measure $\mu _{\Theta _{\mathrm{Ly}}}^{+}$ in $\mathbb{E}%
_{\Theta _{\mathrm{Ly}}}$.
\item $C=\mathrm{cl}\left( \mathrm{im}\xi ^{*}\right) $, which is the
support of the unique repeller measure $\mu _{\Theta _{\mathrm{Ly}}}^{-}$ in
$\mathbb{E}_{\Theta _{\mathrm{Ly}}^{*}}$, and
\item $B=\mathrm{cl}\bigcup_{w\neq 1}\mathrm{st}_{\Theta _{\mathrm{Ly}%
}}\left( x,w\right) $, $x\in \Omega $. That is, $B$ is the closure of the
set of elements that are \textbf{not} transversal to $\xi ^{*}\left(
x\right) $, $x\in \Omega $.
\end{enumerate}
Alternatively we have the following characterization of $B$ in terms of the
closure of the dual section $\xi ^{*}$.
\begin{proposicao}
\label{propbtransv}An element $v\in \mathbb{E}_{\Theta _{\mathrm{Ly}}}$
belongs to $B$ if and only if it is not transversal to some $w\in \mathrm{cl}%
\left( \mathrm{im}\xi ^{*}\right) $ in the same fiber as $v$.
\end{proposicao}
\begin{profe}
Take local trivializations so that locally $\mathbb{E}_{\Theta _{\mathrm{Ly%
}}}\simeq U\times \mathbb{F}_{\Theta _{\mathrm{Ly}}}$, $\mathbb{E}_{\Theta _{%
\mathrm{Ly}}^{*}} \simeq U\times \mathbb{F}_{\Theta _{\mathrm{Ly}}^{*}}$ ($%
U\subset X$ open), $\xi :U\rightarrow \mathbb{F}_{\Theta _{\mathrm{Ly}}}$
and $\xi ^{*}:U\rightarrow \mathbb{F}_{\Theta _{\mathrm{Ly}}^{*}}$. If $%
v=\left( x,b\right) \in B$ then there exists a sequence $\left(
x_{n},b_{n}\right) \rightarrow v$ with $b_{n}$ not transversal to $\xi
^{*}\left( x_{n}\right) $. By taking a subsequence we can assume that $\xi
^{*}\left( x_{n}\right) $ converges to $b^{*}\in \mathbb{F}_{\Theta _{%
\mathrm{Ly}}^{*}}$. Then the pair $\left( b_{n},\xi ^{*}\left( x_{n}\right)
\right) $ converges to $\left( b,b^{*}\right) \in \mathbb{F}_{\Theta _{%
\mathrm{Ly}}}\times \mathbb{F}_{\Theta _{\mathrm{Ly}}^{*}}$. Now, the set of
nontransversal pairs in $\mathbb{F}_{\Theta _{\mathrm{Ly}}}\times \mathbb{F}%
_{\Theta _{\mathrm{Ly}}^{*}}$ is closed. Hence $b$ and $b^{*}$ are not
transversal, showing that $v=\left( x,b\right) $ is not transversal to $%
w=\left( x,b^{*}\right) \in \mathrm{cl}\left( \mathrm{im}\xi ^{*}\right) $.
Conversely, suppose that $v=\left( x,b\right) \in \mathbb{E}_{\Theta _{%
\mathrm{Ly}}}$ is not tranversal to $w=\left( x,b^{*}\right) \in \mathrm{cl}%
\left( \mathrm{im}\xi ^{*}\right) $. Then $b^{*}=\lim \xi ^{*}\left(
x_{n}\right) $ with $\lim x_{n}=x$. By Lemma \ref{lemseqtransv} there exists
a sequence $b_{n}\in \mathbb{F}_{\Theta _{\mathrm{Ly}}}$ such that $b_{n}$
is not transversal to $\xi ^{*}\left( x_{n}\right) $ and $\lim b_{n}=b$.
Hence $\left( x_{n},b_{n}\right) \in B$ and $\lim \left( x_{n},b_{n}\right)
=\left( x,b\right) =v$, showing that $v\in B$.
\end{profe}
Clearly, $A$, $B$ and $C$ are compact sets. Also, $A$ and $C$ are invariant
because the sections $\xi $ and $\xi ^{*}$ are invariant, and since
transversality is preserved by the flow, it follows that $B$ is invariant as
well.
Now we verify that the assumptions of Lemma \ref{lemmaintech} hold in
presence of the three conditions of Theorem \ref{teomainequal}. Statements
(3) and (4) of Lemma \ref{lemmaintech} are the same as the refinement of
Lyapunov spectrum and attractor-repeller measures conditions, respectively.
Item (2) of Lemma \ref{lemmaintech} is the above proposition. So it remains
to prove that $A$ and $B$ are disjoint. This is the only place where the
bounded condition is used.
\begin{proposicao}
$A\cap B=\emptyset $. Precisely, if $v\in A$ and $w\in \mathrm{cl}\left(
\mathrm{im}\xi ^{*}\right) $ then $v$ and $w$ are transversal, and if $v\in
B $ then there exists $w\in \mathrm{cl}\left( \mathrm{im}\xi ^{*}\right) $
in the same fiber which is not transversal to $v$.
\end{proposicao}
\begin{profe}
Since the restriction of $f_{\mathrm{Ly}}$ to $R_{\Omega }$ is bounded its
image in $\mathcal{O}_{\mathrm{Ly}}=\mathrm{Ad}\left( G\right) H_{\mathrm{Ly}%
}$ has compact closure. By the equality $\chi _{\mathrm{Ly}}\left( x\right)
=p\cdot f_{\mathrm{Ly}}\left( p\right) $ ($p\in Q$ with $\pi (p)=x$) it
follows that $\mathrm{cl}\left( \mathrm{im}\chi _{\mathrm{Ly}}\right) $ is a
compact subset of the bundle $Q\times _{G}\mathcal{O}_{\mathrm{Ly}}$. After
identifying $\mathcal{O}_{\mathrm{Ly}}$ with an open subset of $\mathbb{F}%
_{\Theta _{\mathrm{Ly}}}\times \mathbb{F}_{\Theta _{\mathrm{Ly}}^{\ast }}$
we get a section of $Q\times _{G}\left( \mathbb{F}_{\Theta _{\mathrm{Ly}%
}}\times \mathbb{F}_{\Theta _{\mathrm{Ly}}^{\ast }}\right) $ over $\Omega $
also denoted by $\chi _{\mathrm{Ly}}$. The image of this section is
contained in the open subset of those pairs in $Q\times _{G}\left( \mathbb{F}%
_{\Theta _{\mathrm{Ly}}}\times \mathbb{F}_{\Theta _{\mathrm{Ly}}^{\ast
}}\right) $ that are transversal to each other. By compactness the closure
of the image of $\chi _{\mathrm{Ly}}$ contains also only transversal pairs.
Now let $p:Q\times _{G}\left( \mathbb{F}_{\Theta _{\mathrm{Ly}}}\times
\mathbb{F}_{\Theta _{\mathrm{Ly}}^{*}}\right) \rightarrow \mathbb{E}_{\Theta
_{\mathrm{Ly}}}$ and $p^{*}:Q\times _{G}\left( \mathbb{F}_{\Theta _{\mathrm{%
Ly}}}\times \mathbb{F}_{\Theta _{\mathrm{Ly}}^{*}}\right) \rightarrow
\mathbb{E}_{\Theta _{\mathrm{Ly}}^{*}}$ be the canonical projections. Then
\[
\xi =p\circ \chi _{\mathrm{Ly}}\qquad \mathrm{and}\qquad \xi ^{*}=p^{*}\circ
\chi _{\mathrm{Ly}}.
\]
Hence by compactness $p\left( \mathrm{cl}\left( \mathrm{im}\chi _{\mathrm{Ly}%
}\right) \right) =\mathrm{cl}\left( \mathrm{im}\xi \right) $ and $%
p^{*}\left( \mathrm{cl}\left( \mathrm{im}\chi _{\mathrm{Ly}}\right) \right) =%
\mathrm{cl}\left( \mathrm{im}\xi ^{*}\right) $. It follows that two elements
$v\in A=\mathrm{cl}\left( \mathrm{im}\xi \right) $ and $w\in \mathrm{cl}%
\left( \mathrm{im}\xi ^{*}\right) $ are transversal to each other, if they
are in the same fiber.
On the other hand if $v\in B$ then by Proposition \ref{propbtransv}, there
exists $w\in \mathrm{cl}\left( \mathrm{im}\xi ^{*}\right) $ such that $v$
and $w$ are in the same fiber and are not transversal. Hence $v\notin A$,
concluding that $A$ and $B$ are disjoint.
\end{profe}
\vspace{12pt}%
\noindent \textbf{End of proof of Theorem \ref{teomainequal}}: By Lemma \ref%
{lemmaintech}, $A$ and $B$ define a Morse decomposition of $\mathbb{E}%
_{\Theta _{\mathrm{Ly}}}$ with $A$ the attractor component. Hence $A=\mathrm{%
cl}\left( \mathrm{im}\xi \right) $ contains the unique attractor component $%
\mathcal{M}_{\Theta _{\mathrm{Ly}}}^{+}$ of the finest Morse decomposition
of $\mathbb{E}_{\Theta _{\mathrm{Ly}}}$. On the other hand by Proposition %
\ref{prooselcontmor} the Oseledets component $\mathrm{im}\xi \subset
\mathcal{M}_{\Theta _{\mathrm{Ly}}}^{+}$. Therefore $A=\mathrm{cl}\left(
\mathrm{im}\xi \right) \subset \mathcal{M}_{\Theta _{\mathrm{Ly}}}^{+}$, so
that they are equal.
\section{Uniquely ergodic base spaces\label{secunique}}
When the flow on the base space has unique invariant (and hence ergodic)
probability measure $\nu $ the second and third conditions of Section \ref%
{threecon} are meaningless. Hence, in this case, a necessary and sufficient
condition to have equality of Oseledets and Morse decompositions is that the
Oseledets section for $\nu $ is bounded (first condition of Section \ref%
{threecon}).
From another point of view the Morse spectrum $\Lambda _{\mathrm{Mo}}$ of
the attractor component $\mathcal{M}^{+}$ is a compact convex set whose
extremal points are Lyapunov exponents given by integrals with respect to
invariant measures on the maximal flag bundle. By the results of Section \ref%
{secmedinv} any such integral Lyapunov exponent is a regular Lyapunov
exponent of an invariant probability in the base space. Just one of these
Lyapunov exponents belongs to \textrm{cl}$\mathfrak{a}^{+}$, which is the
polar exponent $H_{\mathrm{Ly}}$ associated to the measure.
Hence, $\Lambda _{\mathrm{Mo}}$ has a unique extremal point in \textrm{cl}$%
\mathfrak{a}^{+}$ if the flow on the base space is uniquely ergodic.
\begin{proposicao}
Suppose that the flow on the base space $X$ has a unique invariant
probability measure $\nu $ with $\mathrm{supp}\nu =X$. Let $H_{\mathrm{Ly}%
}=H_{\mathrm{Ly}}\left( \nu \right) $ be its polar exponent. Then $\Lambda _{%
\mathrm{Mo}}$ is the polyhedron whose vertices are $wH_{\mathrm{Ly}}$, $w\in
\mathcal{W}_{\Theta _{\mathrm{Mo}}}$.
\end{proposicao}
\begin{profe}
Since the convex set $\Lambda _{\mathrm{Mo}}$ is invariant by $\mathcal{W}%
_{\Theta _{\mathrm{Mo}}}$ and $H_{\mathrm{Ly}}\in \Lambda _{\mathrm{Mo}}$ we
have that the polyhedron with vertices in $\mathcal{W}_{\Theta _{\mathrm{Mo}%
}}\left( H_{\mathrm{Ly}}\right) $ is contained in $\Lambda _{\mathrm{Mo}}$.
Conversely, suppose that $H$ is an extremal point of $\Lambda _{\mathrm{Mo}}$%
. Then there exists $w\in \mathcal{W}$ such that $wH\in \mathrm{cl}\mathfrak{%
a}^{+}$. We claim that $w\in \mathcal{W}_{\Theta _{\mathrm{Mo}}}$. In fact,
by Weyl group invariance of the Morse spectrum there exists a Morse
component $\mathcal{M}$ such that $wH\in \Lambda _{\mathrm{Mo}}\left(
\mathcal{M}\right) $ (see \cite{smsec}). But the attractor component $%
\mathcal{M}^{+}$ is the only one whose Morse spectrum meets $\mathrm{cl}%
\mathfrak{a}^{+}$. So that $\mathcal{M=M}^{+}$ and $wH\in \Lambda _{\mathrm{%
Mo}}=\Lambda _{\mathrm{Mo}}\left( \mathcal{M}^{+}\right) $ and since the
spectra of distinct Morse components are disjoint we have $w\Lambda _{%
\mathrm{Mo}}=\Lambda _{\mathrm{Mo}}$, implying that $w\in \mathcal{W}%
_{\Theta _{\mathrm{Mo}}}$.
Now $wH\in \mathrm{cl}\mathfrak{a}^{+}$ is an extremal point of $\Lambda _{%
\mathrm{Mo}}=w\Lambda _{\mathrm{Mo}}$ hence $wH=H_{\mathrm{Ly}}$. Therefore,
$\mathcal{W}_{\Theta _{\mathrm{Mo}}}H_{\mathrm{Ly}}$ is the set of extremal
points of $\Lambda _{\mathrm{Mo}}$, concluding the proof.
\end{profe}
\begin{teorema}
Suppose that the flow on the base space $X$ has a unique invariant
probability measure $\nu $ with $\mathrm{supp}\nu =X$. Then the following
conditions are equivalent.
\begin{enumerate}
\item $\Theta _{\mathrm{Ly}}=\Theta _{\mathrm{Mo}}$.
\item The Oseledets section for $\nu $ is bounded.
\item $\alpha \left( \Lambda _{\mathrm{Mo}}\right) =\{0\}$ for all $\alpha
\in \Theta _{\mathrm{Mo}}$.
\end{enumerate}
If these conditions hold then $\Lambda _{\mathrm{Mo}}=\{H_{\mathrm{Ly}}\}$.
\end{teorema}
\begin{profe}
As mentioned above the equivalence between the first two conditions is a
consequence of the main Theorem \ref{teomainequal} and the fact that $\nu $
is the only ergodic measure on $X$. Now, $\Theta _{\mathrm{Ly}}=\Theta _{%
\mathrm{Mo}}$ means that $\alpha \left( H_{\mathrm{Ly}}\right) =0$ for all $%
\alpha \in \Theta _{\mathrm{Mo}}$. If this happens then any $\alpha \in
\Theta _{\mathrm{Mo}}$ annihilates on the polyhedron with vertices $wH_{%
\mathrm{Ly}}$, $w\in \mathcal{W}_{\Theta _{\mathrm{Mo}}}$. Hence $\alpha \in
\Theta _{\mathrm{Mo}}$ is zero on $\Lambda _{\mathrm{Mo}}$ by the above
proposition. Conversely, if (3) holds then $\alpha \left( H_{\mathrm{Ly}%
}\right) =0$ for all $\alpha \in \Theta _{\mathrm{Mo}}$, because $H_{\mathrm{%
Ly}}\in \Lambda _{\mathrm{Mo}}$.
Finally, if $\alpha \left( H_{\mathrm{Ly}}\right) =0$ for all $\alpha \in
\Theta _{\mathrm{Mo}}$ then $wH_{\mathrm{Ly}}=H_{\mathrm{Ly}}$ for every $%
w\in \mathcal{W}_{\Theta _{\mathrm{Mo}}}$ so that $\Lambda _{\mathrm{Mo}%
}=\{H_{\mathrm{Ly}}\}$ by the previous proposition.
\end{profe}
By piecing together known results in the literature we can have examples of
flows over uniquely ergodic systems for which $\Theta _{\mathrm{Ly}}\neq
\Theta _{\mathrm{Mo}}$. Indeed as proved by Furman \cite{fur} the Lyapunov
spectrum is discontinuous at a non-uniform cocycle with values in $\mathrm{Gl%
}\left( d,\mathbb{R}\right) $, that is, at a flow on the trivial bundle $%
X\times \mathrm{Gl}\left( d,\mathbb{R}\right) $. The result of \cite{fur}
(see Theorem 5) makes the assumption that the flow on the base space is
equicontinuous, which is satisfied, e.g., by the translations on compact
groups, like an irrational rotation on the circle $S^{1}$.
Now in Herman \cite{her} there is an example of a non-uniform cocycle with
values in $\mathrm{Sl}\left( 2,\mathbb{R}\right) $ over the irrational
rotation. Thus that example is a discontinuity point of the Lyapunov
spectrum. Finally, in Section \ref{seccont} below we prove continuity of the
whole Lyapunov spectrum if $\Theta _{\mathrm{Ly}}=\Theta _{\mathrm{Mo}}$.
Hence, we get $\Theta _{\mathrm{Ly}}\neq \Theta _{\mathrm{Mo}}$ for the
example in \cite{her}.
\section{Product of i.i.d. sequences\label{seciid}}
The product of independent identically distributed (i.i.d.) random elements
in $G$ yield flows on product spaces $\mathcal{X}\times G$ with plenty of
invariant measures on $\mathcal{X}$. In this section we provide an example
such flow that violates the second condition of Section \ref{threecon} and
hence has distinct Morse and Oseledets decompositions.
Let $C\subset G$ be a compact subset and form the product $\mathcal{X}=C^{%
\mathbb{Z}}$ endowed with the compact product topology. The shift $\tau
\left( \left( x_{n}\right) \right) =\left( x_{n+1}\right) _{n\in \mathbb{Z}}$
is a homeomorphism and hence defines a continuous flow on $\mathcal{X}$.
Now, let $\mu $ be a probability measure with $\mathrm{supp}\mu =C$ and take
the product measure $\mu ^{\times \mathbb{Z}}$ on $\mathcal{X}$. Then $\mu
^{\times \mathbb{Z}}$ is ergodic with respect to the shift $\tau $ and $%
\mathrm{supp}\mu ^{\times \mathbb{Z}}=\mathcal{X}$.
These data defines the continuous flow $\phi _{n}^{\mu }$ on $\mathcal{X}%
\times G$ by $\phi _{n}^{\mu }\left( \mathbf{x},g\right) =\left( \tau
^{n}\left( \mathbf{x}\right) ,\rho ^{\mu }\left( n,\mathbf{x}\right)
g\right) $, where $\mathbf{x}=\left( x_{n}\right) _{n\in \mathbb{Z}}\in C^{%
\mathbb{Z}}\subset G^{\mathbb{Z}}$ and
\[
\rho ^{\mu }\left( n,\mathbf{x}\right) =\left\{
\begin{array}{lll}
x_{n-1}\cdots x_{0} & \mathrm{se} & n\geq 0 \\
x_{1}^{-1}\cdots x_{n}^{-1} & \mathrm{se} & n<0.%
\end{array}%
\right.
\]
The $\mathfrak{a}$-Lyapunov spectrum of $\phi ^{\mu }$ were finded by
Guivarch'-Raugi \cite{gr}. To state their result we recall the following
concepts.
\begin{enumerate}
\item A subgroup $H\subset G$ is \textit{totally irreducible} if it does not
leave invariant a subset which is a finite union of complements of Bruhat
cells in their respective closures (Schubert cells).
\item A sequence $g_{n}\in G$ is said to be \textit{contracting} with
respect to the maximal flag manifold if its polar decomposition $%
g_{n}=u_{n}h_{n}v_{n}\in K\left( \mathrm{cl}A^{+}\right) K$ is such that
\[
\lim_{n\rightarrow \infty }\alpha \left( \log h_{n}\right) =\infty
\]%
for every positive root $\alpha $.
\end{enumerate}
Now denote by $G_{\mu }$ and $S_{\mu }$ the subgroup and semigroup generated
by $\mathrm{supp}\mu =C$, respectively. Then we have the following result of
\cite{gr}, Theorem 2.6.
\begin{teorema}
Let $\mu $ be a probability measure on $G$, and suppose that
\begin{enumerate}
\item the subgroup $G_{\mu }$ is totally irreducible.
\item The semigroup $S_{\mu }$ has a contracting sequence with respect to $%
\mathbb{F}$.
\end{enumerate}
Then the polar exponent of $\phi ^{\mu }$ is regular, that is, belongs to $%
\mathfrak{a}^{+}$. This means that $\Theta _{\mathrm{Ly}}\left( \mu ^{\times
\mathbb{Z}}\right) =\emptyset $.
\end{teorema}
Both conditions of this theorem are satisfied if $S_{\mu }$ has nonempty
interior in $G$:
\begin{enumerate}
\item If $\mathrm{int}S_{\mu }\neq \emptyset $ then $G_{\mu }=G$ because $G$
is assumed to be connected.
\item If $\mathrm{int}S_{\mu }\neq \emptyset $ then there exists a regular $%
h\in \mathrm{int}S_{\mu }$ (see \cite{smics}, Lemma 3.2). Then $h^{n}\in
S_{\mu }$ is a contracting sequence with respect to $\mathbb{F}$.
\end{enumerate}
Hence we get the following consequence of Guivarch'-Raugi \cite{gr} result.
\begin{corolario}
If $\mathrm{int}S_{\mu }\neq \emptyset $ then $\Theta _{\mathrm{Ly}}\left(
\phi ^{\mu }\right) =\emptyset $.
\end{corolario}
Now it is easy to give an exemple that does not satisfy the second condition
of Section \ref{threecon} and hence $\Theta _{\mathrm{Ly}}\left( \phi ^{\mu
}\right) \neq \Theta _{\mathrm{Mo}}\left( \phi ^{\mu }\right) $. In fact,
take a a nonregular element $h=\exp H\in \mathrm{cl}A^{+}$, that is, $\Theta
\left( H\right) =\{\alpha \in \Sigma :\alpha \left( H\right) =0\}\neq
\emptyset $ and a probability $\mu $ whose support $C$ contains $h$ in its
interior. For instance
\[
\mu =\frac{1}{I}f\cdot \eta
\]%
where $\eta $ is Haar measure and $I=\int_{G}f\left( g\right) \eta \left(
dg\right) <\infty $ with $f:G\rightarrow \mathbb{R}$ a nonnegative function
with $\mathrm{supp}f=C$.
Let $\rho =\delta _{\mathbf{x}_{h}}$ be the Dirac measure at\ the constant
sequence $\mathbf{x}_{h}=\left( x_{n}\right) _{n\in \mathbb{Z}}$, $x_{n}=h$.
Clearly $\rho $ is $\tau $-invariant and ergodic. Since
\[
\phi _{n}^{\mu }\left( \mathbf{x}_{h},1\right) =\left( \tau ^{n}\left(
\mathbf{x}_{h}\right) ,\rho \left( n,\mathbf{x}_{h}\right) \right) =\left(
\mathbf{x}_{h},h^{n}\right)
\]%
the polar exponent of $\delta _{\mathbf{x}_{h}}$ is
\[
\lim_{n\rightarrow +\infty }\frac{1}{n}\log h^{n}=\log h=H
\]%
which is not regular. Hence $\Theta _{\mathrm{Ly}}\left( \delta _{\mathbf{x}%
_{h}}\right) =\Theta \left( H\right) \neq \emptyset $ is not contained in $%
\Theta _{\mathrm{Ly}}\left( \mu ^{\times \mathbb{Z}}\right) =\emptyset $.
Therefore, the second condition of Section \ref{threecon} is violated and
the flag types $\Theta _{\mathrm{Ly}}\left( \mu ^{\times \mu }\right) $ and $%
\Theta _{\mathrm{Mo}}\left( \phi ^{\mu }\right) $ are different.
\section{Continuity of the Lyapunov spectrum\label{seccont}}
In this section we apply the differentiability result of \cite{ferrsm} to
show that the equality $\Theta _{\mathrm{Ly}}=\Theta _{\mathrm{Mo}}$ implies
continuity of the Lyapunov spectrum by perturbations of the original $\phi $
that do not change the flow on the base space.
Let $\mathcal{G}=\mathcal{G}\left( Q\right) $ be the gauge group of $%
Q\rightarrow X$, that is, the group of automorphisms of $Q$ that project to
the identity map of $X$. It is well known that $\mathcal{G}$ is a Banach Lie
group.
If $\sigma \in \mathcal{G}$ then $\phi $ and $\sigma \phi $ induce the same
map on $X$ and hence have the same ergodic measure $\nu $. Denote by $H_{%
\mathrm{Ly}}^{\sigma \phi }$ the polar spectrum $\sigma \phi $ with respect
to $\nu $. Assume as before that $\nu $ has full support. Then we have the
following continuity result.
\begin{teorema}
\label{teocont}If $\Theta _{\mathrm{Ly}}\left( \phi \right) =\Theta _{%
\mathrm{Mo}}$ then the map $\sigma \in \mathcal{G}\mapsto H_{\mathrm{Ly}%
}^{\sigma \phi }\in \mathrm{cl}\mathfrak{a}^{+}$ is continuous at $\sigma =%
\mathrm{id}$.
\end{teorema}
We work out separetely the proof for $\mathrm{Sl}\left( n,\mathbb{R}\right) $
in order to explain it in concrete terms. For this group $\mathfrak{a}$ is
the algebra of zero trace diagonal matrices and $\mathfrak{a}^{+}$ are those
with strictly decreasing eigenvalues. The simple set of roots is $\Sigma
=\{\alpha _{1},\ldots ,\alpha _{n-1}\}$ where $\alpha _{i}=\alpha
_{i,i+1}=\lambda _{i}-\lambda _{i+1}$ and $\lambda _{i}\in \mathfrak{a}%
^{\ast }$ maps the diagonal $H\in \mathfrak{a}$ to its $i$-th diagonal
entry. We denote by $\Delta =\{\delta _{1},\ldots ,\delta _{n-1}\}$ the set
of fundamental weights, which is defined by
\[
\frac{2\langle \alpha _{i},\delta _{j}\rangle }{\langle \alpha _{i},\alpha
_{i}\rangle }=\delta _{ij}
\]%
and is given by $\delta _{j}=\lambda _{1}+\cdots +\lambda _{j}$. The Morse
decomposition of $\phi $ on the flag bundles are determined by the subset $%
\Theta _{\mathrm{Mo}}\subset \Sigma $. Alternatively we can look at the
partition
\[
\{1,\ldots ,n\}=\{1,\ldots ,r_{1}\}\cup \{r_{1}+1,\ldots ,r_{2}\}\cup \cdots
\cup \{r_{k}+1,\ldots ,n\}
\]%
where $\Sigma \setminus $ $\Theta _{\mathrm{Mo}}=\{\alpha _{r_{1}},\ldots
,\alpha _{r_{k}}\}$. From the partition we recover $\Theta _{\mathrm{Mo}}$
as the set of $\alpha _{j,j+1}$ such that if $\left[ r,s\right] $ is the
interval of the partition containing $j$ then $r\leq j<s$.
For $H\in \mathrm{cl}\mathfrak{a}^{+}$ with $\alpha \left( H\right) =0$ for
all $\alpha \in \Theta _{\mathrm{Mo}}$ its eigenvalues $a_{i}$ are such that
$a_{i}=a_{j}$ if the indices $i,j$ belong to the same set of the partition.
If furthermore $H$ is such that $\Theta _{\mathrm{Mo}}=\{\alpha \in \Sigma
:\alpha \left( H\right) =0\}$ then the multiplicities of the eigenvalues of $%
H$ are the sizes of the sets of the partition.
Hence $\Theta _{\mathrm{Ly}}=\Theta _{\mathrm{Mo}}$ means that the
multiplicities of the Lyapunov exponents are given by the partition
associated to $\Theta _{\mathrm{Mo}}$.
Now, it was proved in \cite{ferrsm} that the map
\[
\sigma \in \mathcal{G}\mapsto \delta _{j}\left( H_{\mathrm{Ly}}^{\sigma \phi
}\right) \in \mathbb{R}_{+}
\]%
is differentiable at $\sigma =\mathrm{id}$ for any index $j$ such that $%
\alpha _{j,j+1}\notin \Theta _{\mathrm{Mo}}$. Since $\Delta $ is a basis of $%
\mathfrak{a}^{\ast }$ we get continuity of $H_{\mathrm{Ly}}^{\sigma \phi }$
if we prove that $\delta _{j}\left( H_{\mathrm{Ly}}^{\sigma \phi }\right) $
is continuous when $\alpha _{j,j+1}\in \Theta _{\mathrm{Mo}}$.
For this purpose we recall from \cite{alvsm} that $\delta _{j}\left( H_{%
\mathrm{Ly}}^{\sigma \phi }\right) $ is obtained as a limit furnished by the
sub-additive ergodic theorem. Namely,
\begin{equation}
\delta _{j}\left( H_{\mathrm{Ly}}^{\sigma \phi }\right) =\lim \frac{1}{k}%
\delta _{j}\left( \mathsf{a}_{\sigma }^{+}\left( k,x\right) \right)
=\inf_{k\geq 1}\frac{1}{k}\int \delta _{j}\left( \mathsf{a}_{\sigma
}^{+}\left( k,x\right) \right) \nu \left( dx\right) \label{forlimsubad}
\end{equation}%
where $\mathsf{a}_{\sigma }^{+}\left( k,x\right) $, $x\in X$, is the polar
component of the flow define by $\sigma \phi $. (See Section 3.2 in \cite%
{alvsm}. Since $\delta _{j}$ is a fundamental weight $\delta _{j}\left(
\mathsf{a}_{\sigma }^{+}\left( k,x\right) \right) $ is a sub-additive
cocycle on the base space. As showed in \cite{alvsm} this cocycle can be
written as a norm in the space of a representation of $G$, which in this
case is the $j$-fold exterior product of $\mathbb{R}^{n}$.)
By (\ref{forlimsubad}) we have that $\sigma \mapsto \delta _{j}\left( H_{%
\mathrm{Ly}}^{\sigma \phi }\right) $ is upper semi-continuous.
To prove continuity take $j$ with $\alpha _{j,j+1}\in \Theta _{\mathrm{Mo}}$
and let $\left[ r,s\right] $, $r\leq j<s$, be the interval of the partition
that contains $j$. Assume by contradiction that there exist $c>0$ and a
sequence $\sigma _{k}\in \mathcal{G}$ converging to $\mathrm{id}$ such that
\[
\delta _{j}\left( H_{\mathrm{Ly}}^{\sigma _{k}\phi }\right) <\delta
_{j}\left( H_{\mathrm{Ly}}^{\phi }\right) -c.
\]%
Then we have two cases:
\begin{enumerate}
\item $s<n$. Then $\alpha _{s,s+1}\notin \Theta _{\mathrm{Mo}}$, so that $%
\sigma \mapsto \delta _{s}\left( H_{\mathrm{Ly}}^{\sigma \phi }\right) $ is
continuous. The same way $\delta _{r-1}\left( H_{\mathrm{Ly}}^{\sigma \phi
}\right) $ is continuous (where $\delta _{r-1}=0$ if $r=1$). Then for large $%
k$ we have
\[
\delta _{r-1}\left( H_{\mathrm{Ly}}^{\sigma _{k}\phi }\right) >\delta
_{r-1}\left( H_{\mathrm{Ly}}^{\phi }\right) -c/2.
\]%
Since $\lambda _{i_{1}}\geq \lambda _{i_{2}}$ on $\mathrm{cl}\mathfrak{a}%
^{+} $ if $i_{1}\leq i_{2}$ and the polar exponents $H_{\mathrm{Ly}}^{\sigma
_{k}\phi }\in \mathrm{cl}\mathfrak{a}^{+}$ we get
\[
\delta _{j}\left( H_{\mathrm{Ly}}^{\phi }\right) -c>\delta _{r-1}\left( H_{%
\mathrm{Ly}}^{\sigma _{k}\phi }\right) +\left( j-r+1\right) \lambda
_{j}\left( H_{\mathrm{Ly}}^{\sigma _{k}\phi }\right) .
\]%
Hence for large $k$ we have
\[
\delta _{j}\left( H_{\mathrm{Ly}}^{\phi }\right) -c>\delta _{r-1}\left( H_{%
\mathrm{Ly}}^{\phi }\right) -c/2+\left( j-r+1\right) \lambda _{j}\left( H_{%
\mathrm{Ly}}^{\sigma _{k}\phi }\right)
\]%
that is,
\[
\lambda _{j}\left( H_{\mathrm{Ly}}^{\sigma _{k}\phi }\right) <\frac{1}{%
\left( j-r+1\right) }\left( \delta _{j}\left( H_{\mathrm{Ly}}^{\phi }\right)
-\delta _{r-1}\left( H_{\mathrm{Ly}}^{\phi }\right) -c/2\right) .
\]
By the inequality $\delta _{s}=\delta _{j}+\lambda _{j+1}+\cdots +\lambda
_{s}\leq \delta _{j}+\left( s-j\right) \lambda _{j}$ that holds on $\mathrm{%
cl}\mathfrak{a}^{+}$ we get
\begin{equation}
\delta _{s}\left( H_{\mathrm{Ly}}^{\sigma _{k}\phi }\right) \leq \delta
_{j}\left( H_{\mathrm{Ly}}^{\phi }\right) -c+\frac{s-j}{j-r+1}\left( \delta
_{j}\left( H_{\mathrm{Ly}}^{\phi }\right) -\delta _{r-1}\left( H_{\mathrm{Ly}%
}^{\phi }\right) -c/2\right) . \label{fordesigualdade}
\end{equation}%
Now we use the assumption $\Theta _{\mathrm{Ly}}\left( \phi \right) =\Theta
_{\mathrm{Mo}}$ which implies that $\delta _{j}\left( H_{\mathrm{Ly}}^{\phi
}\right) =\delta _{r-1}\left( H_{\mathrm{Ly}}^{\phi }\right) +\left(
j-r+1\right) \lambda _{j}\left( H_{\mathrm{Ly}}^{\phi }\right) $ and $\delta
_{s}\left( H_{\mathrm{Ly}}^{\phi }\right) =\delta _{j}\left( H_{\mathrm{Ly}%
}\left( \phi \right) \right) +\left( s-j\right) \lambda _{j}\left( H_{%
\mathrm{Ly}}^{\phi }\right) $. Hence the last term in (\ref{fordesigualdade}%
) becomes
\[
\frac{s-j}{j-r+1}\left( \left( j-r+1\right) \lambda _{j}\left( H_{\mathrm{Ly}%
}^{\phi }\right) -c/2\right) =\left( s-j\right) \lambda _{j}\left( H_{%
\mathrm{Ly}}^{\phi }\right) -\frac{s-j}{j-r+1}\frac{c}{2},
\]%
so that for large $k$ it holds
\begin{equation}
\delta _{s}\left( H_{\mathrm{Ly}}^{\sigma _{k}\phi }\right) \leq \delta
_{s}\left( H_{\mathrm{Ly}}^{\phi }\right) -\frac{s-j}{j-r+1}\frac{c}{2}
\label{fordesigualfinal}
\end{equation}%
which contradicts the continuity of $\delta _{s}\left( H_{\mathrm{Ly}%
}^{\sigma \phi }\right) $.
\item $s=n$. If $r=1$ then $\Theta _{\mathrm{Ly}}\left( \phi \right) =\Theta
_{\mathrm{Mo}}=\Sigma $ so that $H_{\mathrm{Ly}}^{\phi }=0$ and continuity
follows by upper semi-continuity. When $r\neq 1$ we get continuity of $%
\delta _{r-1}\left( H_{\mathrm{Ly}}^{\sigma \phi }\right) $. By arguing as
in the first case we get the same estimate (\ref{fordesigualfinal}) for $%
0=\delta _{n}=\lambda _{1}+\cdots +\lambda _{n}$, which is a contradiction.
\end{enumerate}
This proves Theorem \ref{teocont} for the group $\mathrm{Sl}\left( d,\mathbb{%
R}\right) $.
Now we consider a general semi-simple group $G$. As before let $\Sigma
=\{\alpha _{1},\ldots ,\alpha _{l}\}$ and $\Delta =\{\delta _{1},\ldots
,\delta _{l}\}$ be the simple system of roots and fundamental weights,
respectively. Given $\Theta \subset \Sigma $ let $\Delta _{\Theta }\subset
\Delta $ be the set of fundamental weights $\delta _{j}$ such that the root
with the same index $\alpha _{j}\in \Theta $. We put%
\[
\mathfrak{a}\left( \Theta \right) =\mathrm{span}\left( \Theta \right) \qquad
\mathfrak{a}_{\Theta }=\mathrm{span}\left( \Delta \setminus \Delta _{\Theta
}\right) .
\]%
These subspaces are orthogonal to each other, and since $\Theta \cup \left(
\Delta \setminus \Delta _{\Theta }\right) $ is a basis of $\mathfrak{a}%
^{\ast }$ \ we have $\mathfrak{a}=\mathfrak{a}\left( \Theta \right) \oplus
\mathfrak{a}_{\Theta }$.
The proof of continuity will be an easy consequence of the following
algebraic lemma.
\begin{lema}
If $\delta \in \Delta _{\Theta }$ then its coordinates with respect to $%
\Theta \cup \left( \Delta \setminus \Delta _{\Theta }\right) $ are
nonnegative.
\end{lema}
\begin{profe}
Let $\gamma _{1}$ and $\gamma _{2}$ be the orthogonal projections of $\delta
$ on $\mathfrak{a}\left( \Theta \right) $ and $\mathfrak{a}_{\Theta }$,
respectively. First we check that the coefficients of $\gamma _{1}$ with
respect to $\Theta $ are nonnegative. By definition there exists just one
root $\alpha \in \Theta $ such that $2\langle \alpha ,\delta \rangle
/\langle \alpha ,\alpha \rangle =1$ and $2\langle \beta ,\delta \rangle
/\langle \beta ,\beta \rangle =0$ if $\beta \neq \alpha $. But if $\beta \in
\Theta $ then $2\langle \beta ,\delta \rangle /\langle \beta ,\beta \rangle
=2\langle \beta ,\gamma _{1}\rangle /\langle \beta ,\beta \rangle $. Hence $%
\gamma _{1}$ \ is a fundamental weight for the root system defined by $%
\Theta $. Its coefficients with respect to $\Theta $ are the entries of the
inverse of the Cartan matrix, which are nonnegative.
As to $\gamma _{2}$ it is given by the mean
\[
\gamma _{2}=\frac{1}{|\mathcal{W}_{\Theta }|}\sum_{w\in \mathcal{W}_{\Theta
}}w\delta ,
\]%
because $\gamma _{2}$ is orthogonal to $\Theta $ and hence $w\gamma
_{2}=\gamma _{2}$ for every $w\in \mathcal{W}_{\Theta }$. Moreover, $\langle
\Theta \rangle $ is a root system in $\mathfrak{a}\left( \Theta \right) $
with Weyl group $\mathcal{W}_{\Theta }$ which has no fixed points in $%
\mathfrak{a}\left( \Theta \right) $ besides $0$. Hence the mean applied to $%
\gamma _{1}$ is $0$ since it is a fixed point.
Now if $\mathfrak{a}^{+}=\{\beta \in \mathfrak{a}^{\ast }:\forall \alpha \in
\Sigma ,~\langle \alpha ,\beta \rangle >0\}$ is the Weyl chamber in $%
\mathfrak{a}^{\ast }$ then the fundamental weight $\delta \in \mathrm{cl}%
\mathfrak{a}^{+}$. Hence
\[
\gamma _{2}\in \mathcal{W}_{\Theta }\left(
\mathrm{cl}\mathfrak{a}^{+}\right) =\bigcup\limits_{w\in
\mathcal{W}_{\Theta }}w\left( \mathrm{cl} \mathfrak{a}^{+}\right)
\]%
because $\mathcal{W}_{\Theta }\left( \mathrm{cl}\mathfrak{a}^{+}\right) $ is
a cone. On the other hand if $\beta \notin \Theta $ and $\gamma \in \mathcal{%
W}_{\Theta }\left( \mathrm{cl}\mathfrak{a}^{+}\right) $ then $\langle \beta
,\gamma \rangle \geq 0$ (see e.g. \cite{smsec}, Lemma 7.5). But
\[
\gamma _{2}=\sum_{\beta \notin \Theta }\frac{2\langle \gamma _{2},\beta
\rangle }{\langle \beta ,\beta \rangle }\delta _{\beta }
\]%
where $\delta _{\beta }$ is the fundamental weight corresponding to $\beta $%
. Hence the coefficients of $\gamma _{2}$ are nonnegative, concluding the
proof.
\end{profe}
\vspace{12pt}%
\noindent%
\textbf{Proof of Theorem \ref{teocont} for }$G$ \textbf{semi-simple:} Let $%
\Theta =\Theta _{\mathrm{Ly}}\left( \phi \right) =\Theta _{\mathrm{Mo}}$ and
take $\delta \in \Delta _{\Theta }$. Then $\sigma \mapsto \delta \left( H_{%
\mathrm{Ly}}^{\sigma \phi }\right) $ is upper semi-continuous. Write
\[
\delta =\sum_{\alpha \in \Theta }a_{\alpha }\alpha +\sum_{\delta \in \Delta
\setminus \Delta _{\Theta }}b_{\delta }\delta .
\]%
with $a_{\alpha }\geq 0$ by the lemma. We have
\[
\delta \left( H_{\mathrm{Ly}}^{\sigma \phi }\right) =\sum_{\alpha \in \Theta
}a_{\alpha }\alpha \left( H_{\mathrm{Ly}}^{\sigma \phi }\right)
+\sum_{\lambda \in \Delta \setminus \Delta _{\Theta }}b_{\lambda }\lambda
\left( H_{\mathrm{Ly}}^{\sigma \phi }\right)
\]%
where the last sum is continuous by the differentiability result of \cite%
{ferrsm}. Hence the first sum is upper semi-continuous as well. The
assumption $\Theta _{\mathrm{Ly}}\left( \phi \right) =\Theta _{\mathrm{Mo}}$
implies that the first sum is zero at $\sigma =\mathrm{id}$. Since it is
nonnegative because $\alpha \left( H_{\mathrm{Ly}}^{\sigma \phi }\right)
\geq 0$ and $a_{\alpha }\geq 0$ we conclude that $\delta $ is continuous at $%
\mathrm{id}$, proving Theorem \ref{teocont}.
It remains to consider the reductive groups, which amounts to check
continuity of the central component defined in Section 3.3 of \cite{alvsm}.
The continuity of this component holds without any further assumption. This
is because this central component is given by an integral
\[
\int \mathfrak{a}_{\sigma }^{+}\left( 1,x\right) \nu \left( dx\right)
\]%
on the base space whose integrand is the time $1$ of a cocycle $\mathfrak{a}%
_{\sigma }^{+}\left( n,x\right) $ that depends continuously of $\sigma \in
\mathcal{G}$ (see \cite{alvsm} for the details).
|
\section{Introduction}
Neutrinoless double beta ($0\nu2\beta$) decay is a crucial process in particle physics since it provides the only experimentally viable method to ascertain the Majorana nature of neutrino, performing in the meantime a sensitive test of the lepton number conservation. In addition, it has the potential to establish the absolute neutrino mass scale and to give information about the hierarchy of the neutrino masses~\cite{Cre14,Sch13,Ell12,Ver12,Giu12,Gom12,Rod11,2bTables}. In particular, one of the main objectives of next-generation $0 \nu 2 \beta$ decay experiments is to explore the inverted hierarchy region of the neutrino mass pattern. This implies reaching a lifetime sensitivity of the order of $10^{26}-10^{27}$ years, which requires a background level close to zero in the ton$\times$year exposure range.
A powerful technology to achieve this outstanding performance is represented by scintillating bolometers, used to study isotopes with a $0\nu2\beta$ Q-value definitely higher than 2.6~MeV, and therefore free from the $\gamma$ background induced by natural radioactivity. In these devices, the dominant background is expected to be given by energy-degraded $\alpha$ particles, which however can be rejected thanks to the different scintillation-to-heat ratio between $\alpha$ and $\beta$ particles~\cite{giusang,Mil-scint,Pirr06}.
The LUMINEU project~\cite{Lumin} (Luminescent Underground Molybdenum Investigation for NEUtrino mass and nature) aims at developing scintillating bolometers based on zinc molybdate (ZnMoO$_4$) crystals to study the isotope $^{100}$Mo (Q-value=3034~keV \cite{Rah08}), capable reaching the performance required to explore the inverted hierarchy region, as shown for the first time in Ref.~\cite{Bee12} and susbsequently confirmed in Ref.~\cite{cuore-luce}. A crytical step in the LUMINEU path is represented by the growth of high-quality radio-pure large-mass (300 -- 500 g) ZnMoO$_4$ monocrystals. As far as non-isotopically enriched material is concerned, the required crystallization technology for a sensitive $0 \nu 2 \beta$ decay experiment has been established, and individual detectors -- prefiguring the single modules of future large arrays -- have already shown to be able to reach the desired bolometric performance and intrinsic radiopurity levels~\cite{Bee12,Bee12a,Bee12b,Bee12c,Che13,Ber14}.
However, the crucial step of reproducing these excellent preliminary results with enriched molybdenum (highly desirable operation for a future large-scale experiment, the $^{100}$Mo natural abundance being only 9.7\% \cite{Ber11}) is not trivial for two main reasons. On one hand, the initial chemical and radioactive purity levels of enriched samples, normally poorer than those characterizing natural material, may conflict with the construction of well performing bolometric detectors. This problem was observed for instance in TeO$_2$ bolometers~\cite{Arn08}. On the other hand, enriched material is very expensive and its procurement represents the highest cost factor in future searches. Therefore, the purification / crystallization chain must imply negligible irrecoverable losses of the $0\nu2\beta$ decay candidate isotope. This strong requirement is not a priori compatible with all the purification and crystal-growth technologies. In this letter, we show for the first time that both potential obstacles are actually overcome in the ZnMoO$_4$ case, adding a further essential element in favor of the use of the LUMINEU technology for future $0\nu2\beta$ decay searches.
\section{Production of a Zn$^{100}$MoO$_4$ crystal boule}
\label{sec:cryst}
Molybdenum enriched in the isotope $^{100}$Mo up to 99.5\% was used to
develop zinc molybdate crystals. The enriched molybdenum was produced at the Kurchatov Institute in the eighties of the last century (former Soviet Union). Approximately one kilogram of the material in form of metal, belonging to the Insitute for Nuclear Research (Kyiv, Ukraine) and now available for the LUMINEU program, was utilized in an experiment at the Modane underground laboratory (France) to search for double beta decay of $^{100}$Mo to excited states of $^{100}$Ru \cite{Blum92}. Afterwards, in order to improve its purity level, the
metallic sample of enriched $^{100}$Mo was dissolved in 20\%
ultrapure nitric acid and transformed into molybdenum acid
($^{100}$MoO$_3 \cdot n$H$_2$O). After rinsing the compound by
nitric acid solution and annealing, a sample of 1199 g of purified molybdenum
oxide ($^{100}$MoO$_3$) was obtained and used in the ARMONIA
experiment \cite{ARMONIA}. The purification procedure has
effectively removed the pollution in $^{40}$K and $^{137}$Cs by
one order of magnitude. The concentrations of thorium and radium
were also decreased by factors 2 and 4, respectively. The
following radioactive contamination of the $^{100}$MoO$_3$ sample
can be derived from the data of the experiment (in
mBq/kg): 2.0(2) for $^{226}$Ra ($^{238}$U chain), 0.8(1) for
$^{228}$Th ($^{232}$Th chain), 5.9(1) for $^{137}$Cs, and 36(2) for
$^{40}$K \cite{ARMONIA} .
The level of impurities in the $^{100}$MoO$_3$ was also measured
by inductively coupled plasma mass-spectrometry (ICP-MS) and
atomic absorption spectroscopy (AAS) methods. The results of these
measurements, presented in Table \ref{100Mo-cont}, as well as
radioactive contamination of the sample, suggest that the material
should be additionally purified to be used for crystal
scintillator production. For example, the concentration of the isotope $^{228}$Th in ZnMoO$_4$ crystals to be used in future $0 \nu 2 \beta$ decay searches must not be higher than 10~$\mu$Bq/kg in order to approach the requested zero background conditions mentioned in the Introduction~\cite{Bee12}.
\begin{table}[htb]
\caption{Contamination of $^{100}$MoO$_3$ measured by inductively
coupled plasma mass-spectrometry (ICP-MS) and atomic absorption
spectroscopy (AAS) methods.}
\begin{center}
\begin{tabular}{p{2cm} p{2cm} p{2cm}}
\hline
\ & \multicolumn{2}{c} {Concentration of element} \\
Element & \multicolumn{2}{c} {in $^{100}$MoO$_3$ (ppm)} \\
\cline {2-3}
~ & ICP-MS & AAS \\
\hline
Na & -- & $<60$ \\
Mg & $<0.5$ & $<4$ \\
Al & 2.4 & -- \\
Si & -- & $<500$ \\
K & $<15$ & $<10$ \\
Ca & -- & $<10$ \\
V & 0.05 & -- \\
Cr & 0.2 & $<5$ \\
Mn & 0.1 & -- \\
Fe & 8 & $<5$ \\
Ni & 0.01 & -- \\
Cu & 0.1 & -- \\
Zn & 0.1 & $<4$ \\
Ag & 0.3 & -- \\
W & 1700 & 550 \\
Pb & 0.008 & -- \\
Th & $<0.0005$ & -- \\
U & 0.001 & -- \\
\hline
\label{100Mo-cont}
\end{tabular}
\end{center}
\end{table}
A two-stage technique of molybdenum purification, consisting of
sublimation of molybdenum oxide in vacuum and recrystallization
from aqueous solutions by co-precipitation of impurities on zinc
molybdate sediment, was applied to purify the enriched molybdenum.
The purification procedure is reported in details in
Ref.~\cite{Ber14}.
Sublimation was carried out with the addition of zinc oxide to
reduce the concentration of tungsten:
\begin{center}
ZnMoO$_4$ + WO$_3$ $\to$ ZnWO$_4$ + MoO$_3$ $\uparrow$.
\end{center}
\noindent The residuals after the sublimation process
were analyzed at the analytical laboratory of the Nikolaev
Institute of Inorganic Chemistry with the help of atomic
emission spectrometry. They contain Fe (0.05 wt\%), Si
($0.6-0.8$ wt\%), Mo ($11-12$ wt\%), W ($22-33$ wt\%) and Zn ($12-16$
wt\%). It should be stressed that the analysis confirmed a contamination
by iron in the initial MoO$_3$. The sublimates were then
annealed in air atmosphere to obtain yellow color stoichiometric
MoO$_3$. Losses of enriched molybdenum at this stage of
purification did not exceed 1.4\%.
After the sublimation process, the molybdenum oxide contained
needlelike granules up to a few millimeter long. The presence of these
granules (which cause difficulties to synthesize zinc
molybdenum compound) is one more argument to apply an additional
stage of purification, consisting of recrystallization of the obtained
molybdenum oxide in aqueous solution. To this purpose, the
molybdenum oxide was dissolved in ammonia solution at room
temperature using zinc molybdate as a collector:
\begin{center}
MoO$_3$ + 2NH$_4$OH = (NH$_4$)$_2$MoO$_4$ + H$_2$O.
\end{center}
The molybdenum losses during recrystallization from
aqueous solutions are expected to be recovered in conditions
of mass production. Therefore, we have estimated the losses at
this stage taking into account a purification process of about 20 kg
of natural molybdenum performed in the framework of the LUMINEU
program and its follow-up. The losses do not exceed 2\%.
Powder of Zn$^{100}$MoO$_4$ (203.98 g) was obtained by solid-phase
synthesis of high-purity zinc oxide (72.23 g, produced by Umicore)
and purified enriched molybdenum oxide $^{100}$MoO$_3$ (131.75 g). The mixture of the two compounds was kept at a temperature of $\approx 680$~$^{\circ}$C in a platinum cup over 12 hours.
A zinc molybdate crystal boule from the enriched $^{100}$Mo compound was grown by the low-thermal-gradient Czo\-ch\-ral\-ski technique
\cite{Pavl92,Boro01,Gala09} in a platinum crucible with size $\oslash 40\times
100$ mm. The crystal was grown at a rotation speed of 20 rotations per
minute at the beginning of the process, decreasing to 4 rotations
per minute at the end. The temperature gradient did not exceed 1
$^{\circ}$C/cm.
The mass of the crystal boule, shown in Fig.~\ref{fig:boule}, is of 170.7~g. Some coloration of the crystal (in contradiction with the practically colorless samples produced from natural molybdenum~\cite{Che13}) can be explained by remaining traces of iron in the powder used for the growth. It should be noted that the initial $^{100}$MoO$_3$ was contaminated by iron at the level of 8~ppm. Besides, the special set of lab-ware used for purification of the small amount of enriched material was not perfectly clean. We hope to improve substantially the optical properties of the enriched crystals in the course of the R\&D in progress now. It is expected that the increase of the purification-cycle scale should improve the quality of the Zn$^{100}$MoO$_4$ compound. Anyway, as it will be demonstrated in the next Section, even the present optical quality of the enriched crystal scintillators is high enough to utilize the material for the development of high-performance scintillating bolometers.
The yield of the crystal boule is 83.7\%. This efficiency is unachievable in the ordinary Czochralski method (maximum $30-45\%$). It should be noted that some amount of Zn$^{100}$MoO$_4$ crystal ($\approx 1.3$ g) remained in the seed. Taking into account the mass of the residual of the melt after the growth process (30.83 g) one can estimate that the losses in the crystal growth process are about 0.6\%.
The data on losses of enriched molybdenum in all the stages of crystal
scintillator production are summarized in Table \ref{100Mo-loss}.
The amount of losses is comparable to that of enriched
cadmium in the production of cadmium tungstate crystal
scintillators from $^{106}$Cd (2.3\%) \cite{Bel10a} and $^{116}$Cd
($\approx 2\%$) \cite{Barabash:2011}. An R\&D program to decrease
the losses of enriched molybdenum by a factor $1.5-2$ by
optimization of all the stages of the production process is in
progress.
\nopagebreak
\begin{figure}[htbp]
\begin{center}
\mbox{\epsfig{figure=fig-boule.eps,height=6.0cm}} \caption{Boule
of Zn$^{100}$MoO$_4$ single crystal with mass of 170.7 g and
length of 95 mm grown by the low-thermal-gradient Czochralski
process.}
\label{fig:boule}
\end{center}
\end{figure}
\begin{table}[htb]
\caption{Irrecoverable losses of enriched molybdenum in all
the stages of Zn$^{100}$MoO$_4$ crystal scintillator production.}
\begin{center}
\begin{tabular}{ll}
\hline
Stage & Loss \\
\hline
Sublimation of $^{100}$MoO$_3$ & 1.4\% \\
Recrystallization from aqueous solutions & 2\% \\
Crystal growth & 0.6\% \\
\hline
Total & 4\% \\
\hline
\label{100Mo-loss}
\end{tabular}
\end{center}
\end{table}
\section{Fabrication and operation of two enriched Zn$^{100}$MoO$_4$ scintillating bolometers}
\label{sec:bolo}
Two samples with a similar size (Zn$^{100}$MoO$_4$-top and Zn$^{100}$MoO$_4$-bottom, with masses of 59.2 and 62.9 g respectively) were produced from the Zn$^{100}$MoO$_4$ boule shown in Fig.~\ref{fig:boule}. The sample Zn$^{100}$MoO$_4$-top corresponds to the top part of the boule, close to its ingot, and characterized by a less intense orange color (corresponding probably to less chemical impurities and defect concentration). A photograph of the produced crystal elements is shown in Fig.~\ref{fig:cryst}. The two samples are only approximately rectangular. Their shape is irregular as it was decided to keep their mass as high as possible in order to make the bolometric tests more significant. We took just the precaution to get two flat parallel bases in order to facilitate holding the crystals by polytetrafluoroethylene (PTFE) elements in the bolometer construction.
\begin{figure}[htb]
\subfigure[]{
\includegraphics[width=0.45\textwidth]{enr-zmo_crystals_top-view.eps}
}
\subfigure[]{
\includegraphics[width=0.45\textwidth]{enr-zmo_crystals_side-view.eps}
}
\caption{Top (a) and side (b) views of the zinc molybdate single crystals obtained from material enriched in $^{100}$Mo to 99.5\%. The Zn$^{100}$MoO$_4$-top sample (see text), at the left, has a mass of 59.2 g and the Zn$^{100}$MoO$_4$-bottom sample, at the right, of 62.9 g.}
\label{fig:cryst}
\end{figure}
The two Zn$^{100}$MoO$_4$ crystals were used to assemble an array of scintillating bolometers. Each sample was equipped with a neutron trasmutation doped (NTD) Ge thermistor for the read-out of the thermal signals. The thermistors have a mass of about 50 mg, a resistance at 20 mK of $\sim 500$ k$\Omega$ and a logarithmic sensitivity $A=-d\log(R)/d\log(T) \simeq 6.5$. They were attached at the crystal surface by using six epoxy glue spots and a 25 $\mu$m thick Mylar spacer (which was removed after the gluing procedure). In addition, each crystal was provided with a heating element glued by means of one epoxy spot, consisting of a resistive meander of heavily-doped silicon with a low-mobility metallic behavior down to very low temperatures. The purpose of this heater is to provide periodically a fixed amount of thermal energy in order to control and stabilize the thermal response of the Zn$^{100}$MoO$_4$ bolometers.
The samples were assembled inside a copper holder by using PTFE elements. A light detector (LD)~\cite{Ten12}, made of $\oslash$50$\times$0.25~mm high-purity Ge disks and instrumented with an NTD Ge thermistor was mounted above $\approx$2 mm from the top face plane of the Zn$^{100}$MoO$_4$ crystals to collect the emitted scintillation light. The inner surface of the copper holder was covered by a reflecting foil (VM2000, VM2002 by 3M) to improve light collection. The Zn$^{100}$MoO$_4$ scintillating bolometer array -- before mounting of the LD -- is shown in Fig. \ref{fig:detector}.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{enr-zmo_bolometer.eps}
\caption{Photograph of the assembled Zn$^{100}$MoO$_4$ bolometer array together with
the photodetector:
(1) Zn$^{100}$MoO$_4$ crystals with mass of 59.2 g (left) and 62.9 g (right);
(2) copper holder of the detectors with the light reflecting foil fixed inside;
(3) PTFE supporting elements;
(4) NTD Ge thermistors;
(5) heating devices;
(6) light detector.}
\label{fig:detector}
\end{figure}
The array was tested at very low temperatures by using a pulse-tube cryostat housing a high-power dilution refrigerator \cite{Man14} installed at the CSNSM (Orsay, France). The cryostat is surrounded by a massive shield made out of low activity lead to minimize pile-up effects, which are particularly disturbing in aboveground measurements with slow detectors such as bolo\-meters based on NTD Ge thermistors.
A room-temperature low-noise electronics, consisting of DC-coupled voltage-sensitive amplifiers~\cite{Arn02} and located inside a Faraday cage, was used for the read-out of the NTD Ge thermistors. Data streams were recorded by a 16 bit commercial ADC with 10 kHz sampling frequency. The test was performed at three base temperatures: 13.7 mK (over 18.3 h), 15 mK (over 4.8 h), and 19 mK (over 24.2 h). The last working point was chosen with the aim to simulate the typical temperature conditions expected in the EDELWEISS set-up \cite{Arm11}, an apparatus to search for dark matter which will be available for our next test deep underground in the Modane underground laboratory (France).
Several experimental parameters were evaluated for each set of measurements in order to extract information about the bolometric performances of the Zn$^{100}$MoO$_4$ array. In particular, we have registered the thermistor resistance ($R_{bol}$) at the working temperature, the signal amplitude ($A_{signal}$) for a unitary deposited energy, the full width at half maximum baseline width (FWHM$_{bsl}$), and pulse rise ($\tau_R$) and decay ($\tau_D$) times. These last two parameters were computed from 10\% to 90\% and from 90\% to 30\% of the signal maximum amplitude respectively. An overview of the detector performance at 13.7 mK and 19 mK in terms of the mentioned parameters is provided in Table \ref{tab:perf}. Similar values were obtained with the Zn$^{100}$MoO$_4$ detectors cooled down to 15 mK, but these results are omitted here due to the very short duration of this measurement.
\begin{table}[!htb]
\caption{Experimental parameters (see text) for the Zn$^{100}$MoO$_4$ scintillating bolometers array registered in aboveground measurements at 13.7 mK (first row) and 19 mK (second row). An event distribution within energy ranges of 500--3000 keV and 10--30 keV for the Zn$^{100}$MoO$_4$ bolometers and for the LD respectively were used to evaluate the parameters $\tau_R$ and $\tau_D$.}
\footnotesize
\centering
\begin{tabular}{cccccc}
\hline
Detector & $R_{bol}$ & $A_{signal}$ & FWHM$_{bsl}$ & $\tau_R$ & $\tau_D$ \\
~ & (M$\Omega$) & ($\mu$V/MeV) & (keV) & (ms) & (ms) \\
\hline
top & 1.54 & 86.8 & 1.4(1) & 9.0 & 46.3 \\
~ & 1.17 & 65.0 & 1.8(1) & 8.9 & 48.4 \\
\hline
bottom & 1.82 & 95.8 & 1.8(1) & 5.5 & 26.2 \\
~ & 1.35 & 84.2 & 2.4(1) & 5.8 & 30.7 \\
\hline
LD & 0.97 & 409 & 0.28(1) & 2.5 & 14.8 \\
~ & 0.81 & 336 & 0.37(2) & 2.5 & 15.5 \\
\hline
\end{tabular}
\label{tab:perf}
\end{table}
The optimum filter procedure~\cite{Gat86}, typically used for the analysis of bolometric data, was applied
to extract the amplitudes of each recorded signal. The calibration of the LD was performed by using a weak $^{55}$Fe source. The energy resolution (FWHM) of the LD at 5.9 keV of $^{55}$Fe was 0.42(2) keV at 13.7 mK and 0.57(4) keV at 19 mK. The energy scale of the Zn$^{100}$MoO$_4$ bolometers was determined by means of a low-activity $^{232}$Th source and $\gamma$ quanta from natural radioactivity (mainly $^{226}$Ra daughters). The energy
spectra accumulated by the Zn$^{100}$MoO$_4$ detectors operated at 13.7 mK are shown in Fig. \ref{fig:bg}. The energy resolution for the detectors at 2614.5 keV of $^{208}$Tl was FWHM = 11(3) keV for the Zn$^{100}$MoO$_4$-top and FWHM = 15(3) keV for the Zn$^{100}$MoO$_4$-bottom in the measurements at 13.7 mK. As one can see from Fig. \ref{fig:bg}, the 2614.5 keV peaks have quite low statistics due to the small mass of the crystals and the short duration of the measurements.
Therefore, it is reasonable to estimate the resolution of the detectors for the more intensive $\gamma$ lines presented in the spectra below 1 MeV. For example, the FWHM at 609.3 keV peak of $^{214}$Bi was measured as 5.0(5) keV and 10(1) keV for the Zn$^{100}$MoO$_4$ top and bottom, respectively. Experience with large-mass slow bolometric detectors shows that the energy resolution on the $\gamma$ lines is significantly worsened by pulse pile-up. We expect therefore that the energy resolution in an underground set-up is much closer to the FWHM baseline width and definitely better than 10~keV, as already observed in underground-operated natural ZnMoO$_4$ bolometers~\cite{Bee12a,Bee12c}.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{enr-zmo_top-bottom_bg_total_13d7mk.eps}
\caption{The energy spectrum accumulated in aboveground measurements over 18.3 h
by two Zn$^{100}$MoO$_4$ scintillating bolometers (top and bottom) mounted in one holder.
The detector was operated at 13.7 mK and irradiated by low-active $^{232}$Th source
and environmental $\gamma$s. The energy of $\gamma$ peaks are in keV.}
\label{fig:bg}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{enr-zmo_top_scatter-plot_13d7mk.eps}
\caption{The scatter plot of light versus detected heat based on the 13.7 mK data accumulated
with the scintillating bolometer Zn$^{100}$MoO$_4$-top during 18.3 h of calibration measurements.
The band populated by ${\gamma(\beta)}$ events (below 2.6 MeV) and cosmic muons is clearly visible.}
\label{fig:scatter}
\end{figure}
Plots reporting the light-to-heat signal amplitude ratio as a function of the heat signal amplitude for the data accumulated over 18.3 h in the aboveground set-up with the enriched Zn$^{100}$MoO$_4$ detectors are presented in Fig.~\ref{fig:scatter}.
The data allow estimating the light yield (the amount of detected light energy per particle energy measured by the deposited heat) related to ${\gamma(\beta)}$ events (LY$_{\gamma(\beta)}$). For instance, the distribution of the LY$_{\gamma(\beta)}$ versus the detected heat accumulated
by the Zn$^{100}$MoO$_4$-top scintillating bolometer is depicted in Fig. \ref{fig:ly}. The LY$_{\gamma(\beta)}$ was estimated by fitting the data in a 600--2700 keV interval. The fit gives similar values for all working temperatures: e.g. 1.01(11) keV/MeV (at 13.7 mK) and 1.02(11) keV/MeV (at 19 mK) for the crystal Zn$^{100}$MoO$_4$-top, and 0.93(11) keV/MeV (at 13.7 mK) and 0.99(12) \allowbreak keV/MeV (at 19 mK) for the crystal Zn$^{100}$MoO$_4$-bottom.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{enr-zmo_top_bottom_ly-vs-e_13d7mk.eps}
\caption{The light yield as a function of the detected heat in aboveground measurements on scintillating bolometers based on Zn$^{100}$MoO$_4$-top (a) and Zn$^{100}$MoO$_4$-bottom (b) crystals. The data were accumulated at 13.7 mK. The events above 2.6 MeV in the ${\gamma(\beta)}$ region are caused by cosmic muons. The position of the $\alpha$ events related to $^{210}$Po is slightly shifted from the nominal value of $Q_{\alpha}$ ($\approx$ 5.4 MeV) due to a different heat response for $\alpha$ particles, already observed in this type of detectors~\cite{Bee12,Bee12a,Bee12b,Bee12c}. The mean value and three $\sigma$ intervals for LY$_{\gamma(\beta)}$ are shown by solid red lines.}
\label{fig:ly}
\end{figure}
In spite of the short duration of the measurements and the aboveground conditions, the data presented
in Fig.~\ref{fig:ly} demonstrate an encouraging internal radiopurity level of the tested Zn$^{100}$MoO$_4$
crystals. There are no remarkable accumulation of counts in the region where events caused by
$\alpha$ radionuclides from the U/Th chains are expected.\footnote{Taking into account the
published values of quenching factors for $\alpha$ particles ($\approx$ 0.12--0.18) derived
in measurements with ZnMoO$_4$-based bolometers~\cite{Bee12,Bee12a,Bee12b,Bee12c}, we consider all the events with energy
in the range 4--9 MeV and light yield below 0.2 keV/MeV as related to potential $\alpha$
contaminants.} A weak event cluster probably related to $^{210}$Po is used to determine the activity of this radionuclide, considering a 200 keV interval around its centroid. The analysis of the data accumulated at 13.7 mK and 19 mK with the Zn$^{100}$MoO$_4$-top and Zn$^{100}$MoO$_4$-bottom bolometers gives an activity of $^{210}$Po in the crystals at the level of 1.1(3) mBq/kg and 1.2(3) mBq/kg, respectively. These activities are similar to those measured precisely in natural ZnMoO$_4$ bolometers~\cite{Bee12c}. It should be noted that the current statistics of the $^{210}$Po counts (4--6 events depending on the set of measurements) is not enough to get precise value of the activities and to distinguish a bulk contamination of $^{210}$Po (or of the progenitor $^{210}$Pb) from surface pollution. The radiopurity of the Zn$^{100}$MoO$_4$ crystals will be precisely determined in next measurements in underground conditions, but the absence of significant $\alpha$ peaks at this level indicates that the contamination of the harmful nuclides $^{228}$Th and $^{214}$Bi~\cite{Bee12} does not exceed a level of a few mBq/kg.
Another important consideration regards the good reproducibility of the behaviour of the two detectors. The slight differences in the operational parameters of the two devices are well within the typical spread observed in this type of detectors. Both detectors, despite some difference in the optical quality, have shown practically identical bolometric and scintillation characteristics.
\section{Conclusions and prospects}
\label{sec:concl}
A zinc molybdate crystal with a mass of 171 g was produced from molybdenum
enriched in $^{100}$Mo to 99.5\% in the framework of the LUMINEU program, after a complex and effective purification method. The output of the crystal boule is 84\%, which demonstrates an important advantage of the low-thermal-gradient Czochralski technique for crystal growing. The irrecoverable losses of enriched molybdenum were found to be of the order of a few \%. The results on the scintillating bolometers fabricated with two enriched crystals obtained from the boule show that the response of these devices meets the requirements of a high-sensitivity double beta decay search and is not distinguishable from the one observed in recent measurements performed with non-enriched detectors \cite{Bee12,Bee12a,Bee12b,Bee12c,Che13,Ber14}. Encouraging, although preliminary, results were also obtained in terms of radiopurity. One can therefore expect a successful operation of scintillating bolometers based on enriched Zn$^{100}$MoO$_4$ crystals in an underground environment. This measurement is presently under preparation.
Recently, significant improvements in the growth technology developed at NIIC (Novosibirsk, Russia) have enabled the synthesis of large regular-shape cylindrical ($\oslash$5$\times$4~cm) natural ZnMoO$_4$ crystals. By using an amount of enriched material larger than that employed in the experiment here described,\footnote{Several kilograms of enriched molybdenum, belonging partly to the Institute of Theoretical and Experimental Physics (Moscow, Russia) and partly to the Institute for Nuclear Research (Kyiv, Ukraine) are available for the development of Zn$^{100}$MoO$_4$ scintillating bolometers.} we foresee to develop Zn$^{100}$MoO$_4$ single detectors with masses in the 300 -- 500 g range in the near future. The size of these devices corresponds to the one envisaged for the single module of large arrays of Zn$^{100}$MoO$_4$ detectors to search for $0 \nu 2 \beta$ decay of $^{100}$Mo~\cite{Bee12}. The currently available enriched material will allow developing an array with a total mass of $\approx 15-20$~kg, corresponding to a sensitivity to the effective Majorana mass in the range 0.05 -- 0.15 eV~\cite{Bee12} and capable therefore to approach the inverted hierarchy region of the neutrino mass pattern.
\section{Acknowledgments}
The development of Zn$^{100}$MoO$_4$ scintillating bolometers is an essential part of the LUMINEU program, receiving funds from the Agence Nationale de la Recherche (France). The work was supported partly by the project ``Cryogenic detectors to search for neutrinoless double beta decay of molybdenum'' in the framework of the Programme ``Dnipro'' based on Ukraine-France Agreement on Cultural, Scientific and Technological Cooperation. The aboveground facility for the test of the scintillating bolometers was realised with the indispensable support of ISOTTA, a project funded by the ASPERA 2nd Common Call dedicated to R\&D activities. The group from the Institute for Nuclear Research (Kyiv, Ukraine) was supported in part by the Space Research Program of the National Academy of Sciences of Ukraine.
It is a pleasure to thank Stefano Nisi from the Gran Sasso laboratory (Italy) for his kind help in the measurements of enriched molybdenum oxide by inductively coupled plasma mass-spectrometry.
|
\section{Introduction}
\begin{figure*}
\begin{center}
\includegraphics[width = \linewidth]{fig1.png}
\end{center}
\caption{Left: sketch of the neutrino-driven wind from the remnant of a BNS merger.
The hot hypermassive neutron star (HMNS) and the accretion
disc emit neutrinos, preferentially along the polar direction and at intermediate latitudes. A fraction
of the neutrinos is absorbed by the disc and can lift matter out of its gravitational potential.
On the viscous time-scale, matter is also ejected along the equatorial direction. Right: sketch of the
isotropised $\nu$ luminosity we are using for our analytical estimates (see the main text for details).}
\label{fig: cartoon}
\end{figure*}
Neutron star mergers play a key role for several branches of modern astrophysics. They are --together
with neutron star-black hole coalescences-- the major astrophysical target of the ground-based gravitational wave detector facilities such as LIGO, VIRGO and KAGRA \citep{Acernese2008,Abbott2009,Harry2010,Somiya2012}.
Moreover, such compact binary mergers have been among the very early suggestions
for the central engines of short gamma-ray bursts (sGRBs) \citep{Paczynski1986,Goodman1986,Eichler1989,Narayan1992}.
While long GRBs (durations $>2$s) very likely have a different origin, compact binary mergers are the
most widely accepted engine for the category of short bursts (sGRBs). Over the years, however, contending models
have emerged and the confrontation of the properties expected from compact binary mergers with those observed
in sGRBs is not completely free of tension \citep[see][for recent reviews]{Piran2004,Lee2007,Nakar2007,Gehrels2009,Berger2011,Berger2013b}.
A binary neutron star merger (hereafter, BNS merger) forms initially a central,
hypermassive neutron star (HMNS) surrounded by a thick accretion disc.
During the merger
process a small fraction of the total mass becomes ejected via gravitational torques and hydrodynamic
processes (``dynamic ejecta''). The decompression of this initially cold and extremely neutron-rich
nuclear matter had long been suspected to provide favourable conditions for the formation
of heavy elements through the rapid neutron capture process (the ``r-process'')
\citep{Lattimer1974,Lattimer1976,Lattimer1977,Symbalisty1982,Eichler1989,Meyer1989,Davies1994}. While initially only considered as an
``exotic'' or second-best model behind core-collapse supernovae, there is nowadays a large literature
that --based on hydrodynamical and nucleosynthetic calculations-- consistently finds that the dynamic
ejecta of a neutron star merger is an extremely promising site for the formation of the heaviest elements
with $A > 130$
\citep[see, e.g., ][]{Rosswog1999,Freiburghaus1999,Oechslin2007a,Metzger2010b,Roberts2011,
Goriely2011a,Goriely2011b,Korobkin2012,Bauswein2013a,
Hotokezaka2013,Kyutoku2013,Wanajo2014}. Core-collapse supernovae, on the contrary, seem seriously challenged
in generating the conditions that are needed to produce elements with $A > 90$
\citep{Arcones2007,Roberts2010,Fischer2010,Huedepohl2010}.
A possible exception, though, may be magnetically driven explosions of rapidly rotating stars \citep{Winteler2012,Moesta2014}.
Such explosions, however, require a combination of rather extreme properties of the pre-explosion star and
are therefore likely rare.\\
Most recently, the idea that compact binary mergers are related to both sGRBs and the nucleosynthesis of the heaviest
elements has gained substantial observational support. In June 2013, the SWIFT satellite detected a
relatively nearby ($z=0.356$) sGRB, GRB130603B, \citep{Melandri2013} for which the Hubble Space Telescope
\citep{Tanvir2013,Berger2013a} detected a nIR point source, 9 days after the burst.
The properties of this second detection are close to model predictions
\citep{Kasen2013,Barnes2013a,Tanaka2013a,Grossman2014,Rosswog2014a,Tanaka2014a}
for the so-called ``macro-'' or ``kilonovae'' \citep{Li1998,Kulkarni2005,Rosswog2005,Metzger2010a,Metzger2010b,Roberts2011},
radioactively powered transients from the decay of freshly produced r-process elements. In particular, the delay of several days
between the sGRB and the nIR detection is consistent with the expanding material having very large opacities, as predicted for
very heavy r-process elements \citep{Kasen2013}.
If this interpretation is correct, GRB130603B would
provide the first observational confirmation of the long-suspected link between compact binary mergers,
heavy elements nucleosynthesis and gamma-ray bursts.\\
There are at least two more channels, apart from the dynamic ejecta, by which a compact binary merger releases matter into space,
and both of them are potentially interesting for nucleosynthesis and --if enough long-lived
radioactive material is produced-- they may also power additional electromagnetic transients.
The first channel is the post-merger accretion disc. As it evolves viscously, expands and cools, the
initially completely dissociated matter recombines into alpha-particles and --together with viscous heating--
releases enough energy to unbind an amount of material that is comparable
to the dynamic ejecta \citep{Metzger2008,Beloborodov2008,Metzger2009b,Lee2009,Fernandez2013}.\\
The second additional channel is related to neutrino-driven winds, the basic mechanisms of which are
sketched in \reffig{fig: cartoon}.
This wind is, in several respects, similar to the one that emerges
from proto-neutron stars. In particular, in both cases a similar amount of gravitational binding energy
is released over a comparable (neutrino diffusion) time-scale, which results in a luminosity of
$L_\nu \sim \Delta E_{\rm grav}/\tau_{\rm diff} \sim 10^{53}$ erg/s and neutrinos
with energies $\sim 10-15$ MeV.
Under these conditions, energy deposition due to neutrino absorption is likely to unbind a fraction of the
merger remnant. In contrast to proto-neutron stars, however, the starting point is extremely neutron-rich
nuclear matter, rather than a deleptonizing stellar core. At remnant temperatures of several MeV,
electron anti-neutrinos dominate over electron neutrinos, contrary to the proto-neutron
star case. Based on scaling relations from the proto-neutron star context \citep{Duncan1986,Qian1996}, early
investigations discussed neutrino-driven winds from merger remnants either in an order-of-magnitude sense
or via parametrized models \citep{Ruffert1997,Rosswog2002b,Rosswog2003,Mclaughlin2005,Surman2006,Surman2008,Metzger2008,Wanajo2012,Caballero2012}.
To date, only one neutrino-hydrodynamics calculation for merger remnants has been published \citep{Dessart2009}.
This study was performed in two dimensions with the code VULCAN/2D and drew its initial conditions from 3D
SPH calculations with similar input physics, but without modelling the heating due to neutrinos \citep{Price2006}.
These calculations confirmed indeed that a neutrino-driven wind develops (with $\dot{M} \sim 10^{-3}$ ${\rm M}_{\sun}$/s),
blown out into the funnel along the binary rotation axis that was previously thought to be practically baryon-free.
By baryon-loading the suspected launch path, this wind could potentially threaten the emergence of
the ultra-relativistic outflow that is needed for a short GRB. \cite{Dessart2009} therefore concluded that the launch
of a sGRB was unlikely to happen in the presence of the HMNS, but could possibly occur after the collapse to a black hole.\\
The aim of this study is to explore further neutrino-driven winds from compact binary mergers remnants.
We focus here on the phase where a HMNS is present in the centre and we assume
that it does not collapse during the time frame of our simulation, as in \cite{Dessart2009}. Given the
various stabilising mechanism such as thermal support, possibly magnetic fields and in particular
the strong differential rotation of the HMNS together with a lower limit on the maximum
mass in excess of $2.0$ ${\rm M}_{\sun}$ \citep{Demorest2010,Antoniadis2013}, we consider this as a very plausible
assumption. We are mainly interested to see how robust the previous 2D results are with respect to a
transition to three spatial dimensions. The questions about
the understanding of the heavy element nucleosynthesis that occurs in compact binary mergers,
the prediction of observable electromagnetic counterparts for the different outflows,
and the emergence of sGRBs are the main drivers behind this work.\\
This paper is organized as follows.
In \refsec{sec: preliminary analysis and estimates},
we estimate the most important disc and wind time-scales.
The details of our numerical model are explained in \refsec{sec: method}. In addition,
we briefly present the merger simulation,
the outcome of which is used as initial condition for our study.
Our results are presented in \refsec{sec: results}.
We briefly discuss in \refsec{sec:discussion} the nucleosynthesis in the neutrino-driven wind
and the properties of the radioactively powered, electromagnetic transients
that result from them. Our major results are finally summarised in \refsec{sec:conclusions}.
\section{Analytical estimates}
\label{sec: preliminary analysis and estimates}
The properties of the remnant of a BNS merger can vary significantly
\citep[see, for example, ][and references therein]{Rosswog2013,Bauswein2013a,Hotokezaka2013,Wanajo2014},
depending on the binary parameters (mass, mass ratio, eccentricity, spins etc.) and on the nuclear
equation of state (hereafter, EoS). For our estimates and scaling relations, we use numerical values
that characterise our initial model, see \refsec{sec: initial conditions} for more details.\\
We consider a central HMNS of mass $M_{\rm ns} \approx 2.5$ ${\rm M}_{\sun}$,
radius $R_{\rm ns} \approx 25 \, {\rm km}$ and temperature $k_{\rm B} T_{\rm ns} \approx 15 \, {\rm MeV} $.
Inside of it, neutrinos are assumed to be in thermal equilibrium with matter. Under these conditions
the typical neutrino energy can be estimated as $E_{\nu,{\rm ns}} \sim \left( F_3(0)/F_2(0) \right) \,
k_{\rm B}T_{\rm ns}\approx 3.15 \, k_{\rm B}T_{\rm ns} \approx 50 \, {\rm MeV}$,
where $F_n(0)$ is the Fermi integral of order $n$, evaluated for a vanishing degeneracy parameter.
The central object is surrounded by a geometrically thick disc of mass $M_{\rm disc} \approx 0.2$ ${\rm M}_{\sun}$,
radius $R_{\rm disc} \approx 100 \, {\rm km}$ and height $H_{\rm disc} \approx 33 \, {\rm km}$.
The aspect ratio of the disc is then $H/R \approx 1/3$.
We assume a neutrino energy in the disc of $E_{\nu,{\rm disc}} \sim 15 \, {\rm MeV}$,
comparable with the mean energy of the ultimately emitted neutrinos \citep[see, for example, ][]{Rosswog2013}.\\
Representative density values in the HMNS and in the disc are $\rho_{\rm ns} \sim 10^{14} {\rm g \, cm^{-3}}$ and
$\rho_{\rm disc}\sim 5 \cdot 10^{11} {\rm g \, cm^{-3}}$, respectively.\\
The \emph{dynamical time-scale} $t_{\rm dyn}$ of the disc is set by the orbital Keplerian
motion around the HMNS,
\begin{equation}
t_{\rm dyn} \sim \frac{2 \pi}{\Omega_K} \approx 0.011 \, {\rm s} \,
\left( \frac{M_{\rm ns}}{2.5 M_{\sun}} \right)^{-1/2}
\left( \frac{R_{\rm disc}}{100 \, {\rm km}} \right)^{3/2},
\label{eqn: keplerian time-scale}
\end{equation}
where $\Omega_K$ is the Keplerian angular velocity.\\
On a time-scale longer than $t_{\rm dyn}$, viscosity drives radial motion.
Assuming it can be described by an $\alpha$-parameter model \citep{Shakura1973},
we estimate the \emph{lifetime of the accretion disc} $t_{\rm disc}$ as
\begin{eqnarray}
\lefteqn{t_{\rm disc} \sim \alpha^{-1} \left( \frac{H}{R} \right)^{-2} \Omega_{K}^{-1} \approx
0.3 \, {\rm s} \, \left( \frac{\alpha}{0.05} \right)^{-1} \left( \frac{H/R}{1/3} \right)^{-2} } \nonumber \\
&& \left( \frac{M_{\rm ns}}{2.5 M_{\sun}} \right)^{-1/2} \left( \frac{R_{\rm disc}}{100 \, {\rm km}} \right)^{3/2}.
\label{eqn: viscosity time}
\end{eqnarray}
The \emph{accretion rate} on the HMNS $\dot{M}$ is then of order
\begin{eqnarray}
\lefteqn{\dot{M} \sim \frac{M_{\rm disc}}{t_{\rm disc}} \approx 0.64 \, \frac{M_{\sun}}{\rm s} \,
\left( \frac{M_{\rm disc}}{0.2 \, M_{\sun}} \right) \left( \frac{\alpha}{0.05} \right) } \nonumber \\
&& \left( \frac{H/R}{1/3} \right)^{2} \left( \frac{M_{\rm ns}}{2.5 M_{\sun}} \right)^{1/2}
\left( \frac{R_{\rm disc}}{100 \, {\rm km}} \right)^{-3/2}.
\label{eqn: mdot}
\end{eqnarray}
Neutrinos are the major cooling agent of the remnant.
Neutrino scattering off nucleons is one of the major sources of {\em opacity} for all neutrino species\footnote{In the
case of $\nu_e$'s, the opacity related with absorption by neutrons is even larger. Nevertheless, it is still comparable
to the scattering off nucleons.}
and the corresponding mean free path can be estimated as
\begin{equation}
\lambda_{N \nu}
\approx 7.44 \cdot 10^3 \, {\rm cm} \, \left( \frac{\rho}{10^{14} \, {\rm g/cm^3}} \right)^{-1}
\left( \frac{E_{\nu}}{10 \, {\rm MeV}} \right)^{-2},
\label{eqn: lambda nu}
\end{equation}
where $\rho$ is the matter density and $E_{\nu}$ is the typical neutrino energy.
The large variation in density between the HMNS and the disc
suggests to treat these two regions separately.\\
For the central compact object, the \emph{cooling time-scale} $t_{\rm cool,ns}$
is governed by neutrino diffusion (see, for example, \cite{Rosswog2003}).
If $\tau_{\nu,{\rm ns}}$ is the neutrino \emph{optical depth} inside the HMNS, then
\begin{equation}
t_{\rm cool, ns} \sim 3 \, \frac{\tau_{\nu,{\rm ns}} \, R_{\rm ns}}{c}.
\label{eqn: diffusion time-scale estimate}
\end{equation}
If we assume $\tau_{\nu,{\rm ns}} \sim R_{\rm ns}/ \lambda_{N \nu}$,
\begin{equation}
t_{\rm cool,ns} \sim 1.88 \, {\rm s} \left( \frac{R_{\rm ns}}{25 \, {\rm km}} \right)^2
\left(\frac{\rho_{\rm ns}}{10^{14}{\rm g/cm^3}} \right)
\left(\frac{k_{\rm B}T_{\rm ns}}{15 \, {\rm MeV}} \right)^2.
\label{eqn: diffusion time HMNS}
\end{equation}
The neutrino luminosity coming from the HMNS is powered by an internal energy reservoir $\Delta E_{\rm ns}$.
We estimate it as the difference between the internal energy of a hot and of a cold HMNS.
For the first one, we consider typical profiles of a HMNS obtained from a BNS merger simulation.
For the second one, we set $T = 0$ everywhere inside it.
Under these assumptions, $\Delta E_{\rm ns} \approx 0.30 \, E_{\rm int, HMNS} \approx 3.4 \cdot 10^{52} {\rm erg} $,
and the associated HMNS neutrino luminosity (integrated over all neutrino species) is approximately
\begin{eqnarray}
\lefteqn{L_{\nu,{\rm ns}} \sim \frac{\Delta E_{\rm ns}}{t_{\rm diff,ns}} \approx
1.86 \cdot 10^{52}\,\frac{\rm erg}{\rm s} \left( \frac{\Delta E_{\rm ns}}{3.5 \cdot 10^{52} {\rm erg} } \right)} \nonumber \\
&& \left( \frac{R_{\rm ns}}{25 \, {\rm km}} \right)^{-2}
\left(\frac{\rho_{\rm ns}}{10^{14}{\rm g/cm^3}} \right)^{-1}
\left(\frac{k_{\rm B}T_{\rm ns}}{15 \, {\rm MeV}} \right)^{-2}.
\label{eqn: HMNS luminosity}
\end{eqnarray}
The disc diffusion time-scale can be estimated using an analogous to
\refeq{eqn: diffusion time-scale estimate}:
\begin{eqnarray}
\lefteqn{t_{\rm cool,disc} \sim 3 \, \frac{\tau_{\nu,{\rm disc}} \, H_{\rm disc}}{c} \approx 1.68 \, {\rm ms} \,
\left( \frac{H_{\rm disc}}{33 \, {\rm km}} \right)^2} \nonumber \\
&& \left(\frac{\rho_{\rm disc}}{5 \cdot 10^{11}{\rm g/cm^3}} \right)
\left(\frac{E_{\nu,{\rm disc}}}{15 \, {\rm MeV}} \right)^2.
\label{eqn: diffusion time disc}
\end{eqnarray}
Due to this fast cooling time-scale, a persistent neutrino luminosity from the disc requires a constant
supply of internal energy. In an accretion disc, this is provided by the \emph{accretion mechanism}:
while matter falls into deeper Keplerian orbits, the released
gravitational energy is partially ($\sim 50$ per cent) converted into
internal energy.
If $R_{\rm disc} \sim 100 \, {\rm km}$ denotes the typical initial distance inside the disc, and
the radius of the HMNS is assumed to be the final one, then
$\Delta E_{\rm grav} \sim \left( G M_{\rm ns} M_{\rm disc} / R_{\rm ns}\right)$,
where we have used $R_{\rm ns}^{-1} \gg R_{\rm disc}^{-1}$.
The neutrino luminosity for the accretion process is approximately
\begin{eqnarray}
\lefteqn{L_{\nu,{\rm disc}} \sim 0.5 \, \frac{\Delta E_{\rm grav}}{t_{\rm disc}} \approx 8.35 \cdot 10^{52} \, \frac{\rm erg}{\rm s}
\left( \frac{M_{\rm ns}}{2.5 \, M_{\sun}} \right)^{3/2}
\, \left( \frac{\alpha}{0.05 } \right)} \nonumber \\
&& \left( \frac{M_{\rm disc}}{0.2 \, M_{\sun}} \right)
\left( \frac{H/R}{1/3} \right)^2
\left( \frac{R_{\rm disc}}{100 \, {\rm km}} \right)^{-3/2}
\left( \frac{R_{\rm ns}}{25 \, {\rm km}} \right)^{-1}.
\label{eqn: disc luminosity}
\end{eqnarray}\\
Note that during the disc accretion phase $L_{\nu,{\rm disc}}$ is larger
than $L_{\nu,{\rm ns}}$. Together, the HMNS and the disc release neutrinos at
a luminosity of $\sim 10^{53}$ erg/s, consistent with the simple estimate from the introduction.\\
Due to the density (opacity) structure of the disc, the neutrino emission is expected to be anisotropic, with a larger
luminosity in the polar directions ($\theta = 0$ and $\theta = \pi$), compared to the one
along the equator ($\theta = \pi/2$), see also \cite{Rosswog2003a,Dessart2009}. For a simple
model of this effect, we assume that the disc creates an axisymmetric shadow area
across the equator, while the emission is uniform outside this area.
The amplitude of the shadow is $2 \, \theta_{\rm disc}$, where
$\tan{\theta_{\rm disc}} = (H/R)$.
Then, we define an isotropised axisymmetric luminosity $L_{\nu,{\rm iso}}(\theta)$ as
(see the sketch on the right in \reffig{fig: cartoon}):
\begin{equation}
L_{\nu,{\rm iso}}(\theta) =
\left\{
\begin{array}{rl}
\xi \, L_{\nu} & \quad \mbox{for } \left| \theta - \pi/2 \right| > \theta_{\rm disc} \\
0 & \quad \mbox{for } \left| \theta - \pi/2 \right| \leq \theta_{\rm disc}.
\end{array}
\right.
\end{equation}
The value of $\xi$ is set by the normalisation of $L_{\nu,{\rm iso}}$ over
the whole solid angle $\Omega$, $\int_{\Omega} L_{\nu,{\rm iso}} \, {\rm d}\Omega = L_{\nu}$:
\begin{equation}
\xi = \frac{1}{1-\sin{\theta_{\rm disc}}}.
\end{equation}
For $(H/R) \approx 1/3$, one finds $\theta_{\rm disc} \approx \pi/10$ and $\xi \approx 1.5$.
After having determined approximate expressions for the neutrino luminosities,
we are ready to estimate
the relevant time-scale for the formation of the $\nu$-driven wind.\\
We define the \emph{wind time-scale} $t_{\rm wind}$
as the time necessary for the matter
to absorb enough energy
to overcome the gravitational well generated by the HMNS.
This energy deposition happens inside the disc and it is due to
the re-absorption of neutrinos emitted
at their last interaction surface.
Thus,
\begin{equation}
t_{\rm wind} \sim e_{\rm grav}/\dot{e}_{\rm heat},
\label{eqn: wind time-scale}
\end{equation}
where $e_{\rm grav} \approx G M_{\rm ns}/R$ is the specific gravitational energy,
and $\dot{e}_{\rm heat}$ is the specific heating rate provided by neutrino absorption
at a radial distance $R$ from the centre:
\begin{equation}
\dot{e}_{\rm heat} \sim k \,\frac{L_{\nu_e,{\rm iso}}(\left| \theta - \pi/2 \right| > \theta_{\rm disc} )}
{4 \, \pi \, R^2} .
\label{eqn: heating rate}
\end{equation}
In the equation above we have assumed that
$L_{\nu_e} \approx L_{\bar{\nu}_e} \sim \left( L_{\nu,{\rm ns}} + L_{\nu,{\rm disc}} \right)/3 $.
If $k \approx 5.65 \cdot 10^{-20} \, {\rm cm^{2}} \, {\rm g^{-1} \, MeV^{-2}} \, E_{\nu}^2 \, $
is the typical absorptivity on nucleons \citep{Bruenn1985},
the heating rate can be re-expressed as
\begin{eqnarray}
\lefteqn{\dot{e}_{\rm heat} \sim 4.6 \cdot 10^{20} \frac{\rm erg}{\rm g \cdot s} \left( \frac{R}{100 \, {\rm km}} \right)^{-2}} \nonumber \\
&&
\left( \frac{L_{\nu_e}}{3 \cdot 10^{52} \, {\rm erg/s}} \right)
\left( \frac{\xi}{1.5} \right)
\left( \frac{E_{\nu,{\rm disc}}}{15 \, {\rm MeV}} \right)^{2}.
\label{eqn: heating rate estimation}
\end{eqnarray}
Finally, the wind time-scale, \refeq{eqn: wind time-scale}, becomes
\begin{eqnarray}
\lefteqn{t_{\rm wind} \sim 0.07 \, {\rm s} \, \left( \frac{M_{\rm ns}}{2.5 \, M_{\sun}} \right)
\left( \frac{R}{100 \, {\rm km}} \right)} \nonumber \\
&& \left( \frac{L_{\nu_e}}{3 \cdot 10^{52} \, {\rm erg/s}} \right)^{-1}
\left( \frac{\xi}{1.5} \right)^{-1}
\left( \frac{E_{\nu,{\rm disc}}}{15 \, {\rm MeV}} \right)^{-2}.
\end{eqnarray}
Since $t_{\rm wind} < t_{\rm disc}$, neutrino heating
can drive a wind within the lifetime of the disc.
Moreover, since the disc provides a substantial fraction of the total neutrino luminosity,
a wind can form also in the absence of the HMNS.
Of course, the neutrino emission processes are much more complicated than what can be captured by
these simple estimates. Nevertheless, they provide a reasonable first guidance for the qualitative understanding
of the remnant evolution.
\section{Numerical model for the remnant evolution}
\label{sec: method}
\subsection{Hydrodynamics}
\label{sec:hydro}
We perform our simulations with the \texttt{FISH} code \citep{Kaeppeli2011}.
\texttt{FISH} is a parallel grid code that solves the equations of ideal, Newtonian
hydrodynamics (HD) \footnote{\texttt{FISH} can actually solve the
equations of ideal magnetohydrodynamics. However, we have not included
magnetic fields in our current setup.}:
\begin{equation}
\label{eq:hd_0010}
\frac{\partial \rho}{\partial t}
+ \nabla \cdot \left( \rho {\bf v} \right)
= 0
\end{equation}
\begin{equation}
\label{eq:hd_0020}
\frac{\partial \rho {\bf v}}{\partial t}
+ \nabla \cdot \left( \rho {\bf v} \otimes {\bf v} \right)
+ \nabla p
= - \rho \nabla \phi + \rho \, \left( \frac{{\rm d}{\bf v}}{{\rm d}t} \right)_{\nu}
\end{equation}
\begin{equation}
\label{eq:hd_0030}
\frac{\partial E}{\partial t}
+ \nabla \cdot \left[ \left( E + p \right) {\bf v} \right]
= - \rho {\bf v} \nabla \phi
+ \rho \, \left( \frac{{\rm d}e}{{\rm d}t} \right)_{\nu}
\end{equation}
\begin{equation}
\label{eq:hd_0040}
\frac{\partial \rho Y_e}{\partial t}
+ \nabla \cdot \left( \rho Y_e {\bf v} \right)
= \rho \, \left( \frac{{\rm d} Y_e}{{\rm d}t} \right)_{\nu}
\end{equation}
Here $\rho$ is the mass density, ${\bf v}$ the velocity,
$E = \rho e + \rho v^2/2$ the total energy density (i.e., the sum of
internal and kinetic energy density), $e$ the specific internal energy,
$p$ the matter pressure and $Y_e$ the electron fraction.
The code solves the HD equations with a second-order accurate finite
volume scheme on a uniform Cartesian grid.
The source terms on the right hand side stem from gravity and from neutrino-matter interactions.
We notice that the viscosity of our code is of numerical nature,
while no physical viscosity is explicitly included.
The neutrino source terms will be discussed in detail in \refsec{sec:neutrino treatment}.
The gravitational potential $\phi$ obeys the Poisson equation
\begin{equation}
\label{eq:hd_0050}
\nabla^2 \phi = 4 \pi G \rho
,
\end{equation}
where $G$ is the gravitational constant.
The merger of two neutron stars with equal masses is expected
to form a highly axisymmetric remnant.
We exploit this approximate invariance by solving the Poisson equation in cylindrical symmetry.
This approximation results in a high gain in computational efficiency, given the
elliptic (and hence global) nature of \refeq{eq:hd_0050}.
To this end, we \emph{conservatively} average the three-dimensional density
distribution onto an axisymmetric grid, having the HMNS rotational axis as the
symmetry axis.
The Poisson equation is then solved with a fast multigrid
algorithm \citep{Press1992}, and
the resulting potential is interpolated back on the
three-dimensional grid.
The HD equations are closed by an
EoS relating the internal energy to the pressure.
In our model, we use the TM1 EoS description of nuclear matter supplemented with
electron-positron and photon contributions, in tabulated form
\citep{Timmes2000,Hempel2012}. This description is equivalent to one provided by the
Shen et al. EoS \citep{Shen1998a,Shen1998b} in the high density part.
\subsection{Neutrino treatment}
\label{sec:neutrino treatment}
In general, the multi-dimensional neutrino transport is described
by the equation of radiative transfer \citep[see, for example, ][]{Mihalas1984}.
Instead of a direct solution of this equation,
which is computationally very expensive in large multi-dimensional simulations, we employ a relatively inexpensive,
effective neutrino treatment. Our goal is to provide expressions for the neutrino source terms,
assuming to know qualitatively the solution of the radiative transfer equation
in different parts of the domain.
Our treatment is a spectral extension of previous grey leakage schemes
\citep{Ruffert1996,Rosswog2003}. However, differently from its predecessors,
it includes also {\em spectral absorption terms in the optically thin regime}.
The treatment has been developed and tested against detailed Boltzmann neutrino transport
for spherically symmetric core collapse supernova models.
For two tested progenitors ($15$ ${\rm M}_{\sun}$~ and $40$ ${\rm M}_{\sun}$~ zero age main sequence stars),
the neutrino luminosities and the shock positions agree within 20 per cent with the corresponding values
obtained by Boltzmann transport, for a few hundreds of milliseconds after core bounce.
A detailed description with
tests will be discussed in a separate paper (Perego et al. 2014, in preparation).
Here we provide a summary of the method and we refer to it as
an Advance Spectral Leakage (ASL) scheme\footnote{The ASL scheme allows also the modelling
of the neutrino trapped component. However, since this component was not included
in the study of the merger process that provided our initial conditions, we neglect it here.}.
The neutrino energy is discredited in 12 geometrically increasing energy bins, chosen in the range
$2 \, {\rm MeV} \leq E_{\nu} \leq 200 \, {\rm MeV}$.
The ASL scheme includes the reactions listed in \reftab{table: nu reactions}.
They correspond to the reactions that we expect to be more relevant in hot and dense matter.
Neutrino pair annihilation is included only as a source of opacity in optically thick conditions.
Due to the geometry of the emission, it is also supposed to be
important in optically thin conditions (see, for example, \cite{Janka1991,Burrows2006}. For
the application to the BNS merger scenario, see \cite{Dessart2009} and references therein).
Therefore, our numbers concerning the mass loss
$\dot{M}$ need to be considered as lower limits on the true value.
The full inclusion of this process in our model will be performed in a future step.
\begin{table}
\begin{center}
\begin{tabular}{| l | c | l |}
\hline
Reaction & Roles & Ref. \\ \hline
$e^- + p \leftrightarrow n + \nu_e $ & O,T,P & a \\
$e^+ + n \leftrightarrow p + \bar{\nu}_e $ & O,T,P & a \\
$e^- + (A,Z) \leftrightarrow \nu_e + (A,Z-1)$ & T,P & a \\
$N + \nu \leftrightarrow N + \nu $ & O & a \\
$(A,Z) + \nu \leftrightarrow (A,Z) + \nu $ & O & a \\
$e^+ + e^- \leftrightarrow \nu + \bar{\nu} $ & T,P & a,b \\
$N + N \leftrightarrow N + N + \nu + \bar{\nu} $ & T,P & c \\ \hline
\end{tabular}
\end{center}
\caption{List of the neutrino reactions included in the simulation (left column;
$\nu \equiv \nu_e,\bar{\nu}_e,\nu_{\mu,\tau}$), of their major effects
(central column; O stands for opacity, P for neutrino
production, T for neutrino thermalisation), and of the references for the implementation (right column):
``a'' corresponds to \protect\cite{Bruenn1985}, ``b'' to \protect\cite{mezzacappa1993},
and ``c'' to \protect\cite{Hannestad1998}.
}
\label{table: nu reactions}
\end{table}
The ASL scheme models explicitly three different neutrino species: $\nu_e$, $\bar{\nu}_e$, and $\nu_{\mu,\tau}$.
The species $\nu_{\mu,\tau}$ is a collective species for $\mu$ and $\tau$ (anti-)neutrinos,
that contributes only as a source of cooling in the energy equation.
As a consequence of the distinction between emission and absorption processes, and between different
neutrino species, the source terms in \refeq{eq:hd_0020}-\refeq{eq:hd_0040}
can be split into different contributions. For the electron fraction,
\begin{equation}
\left( \frac{{\rm d}Y_e}{{\rm d}t} \right)_\nu = - m_b \, \left[ \left( R^0_{\nu_e} - R^0_{\bar{\nu}_e} \right) + \left( H^0_{\nu_e} - H^0_{\bar{\nu}_e} \right) \right],
\label{eqn: yedot contributions}
\end{equation}
where $R^0_{\nu}$ and $H^0_{\nu}$ denote the specific particle emission and absorption rates
for a neutrino type $\nu$ respectively, and $m_b$ is the baryon mass (with $m_b c^2 = 939.021 \, {\rm MeV} $).
For the specific internal energy of the fluid,
\begin{equation}
\left( \frac{{\rm d}e}{{\rm d} t} \right)_\nu = - \left( R^1_{\nu_e} + R^1_{\bar{\nu}_e} + 4 \, R^1_{\nu_{\mu,\tau}} \right) + H^1_{\nu_e} + H^1_{\bar{\nu}_e},
\label{eqn: edot contributions}
\end{equation}
where $R^1_{\nu}$ and $H^1_{\nu}$ indicate the specific energy emission and absorption rates, respectively.
The factor 4 in front of $R^1_{\nu_{\mu,\tau}}$ accounts for the four different species modelled
collectively as $\nu_{\mu,\tau}$.
And, finally, for the fluid velocity,
\begin{equation}
\left( \frac{{\rm d}\mathbf{v}}{{\rm d}t} \right)_{\nu} =
\left( \frac{{\rm d}\mathbf{v}}{{\rm d}t} \right)_{\nu_e} +
\left( \frac{{\rm d}\mathbf{v}}{{\rm d}t} \right)_{\bar{\nu}_e}.
\label{eqn: vdot contributions}
\end{equation}
is the acceleration provided by the momentum transferred by
the absorption of $\nu_e$'s and $\bar{\nu}_e$'s in the optically thin region.
Since the trapped neutrino component is not dynamically modelled, we neglect
the related neutrino stress in optically thick conditions. As a consequence,
$\nu_{\mu,\tau}$'s do not contribute to the acceleration term.\\
For each neutrino $\nu$ species, the \emph{luminosity} ($L_{\nu}$) and
\emph{number luminosity} ($L_{N,\nu}$)
are calculated as:
\begin{equation}
L_{\nu} = \int_V \: \rho \left( R^1_{\nu} - H^1_{\nu} \right) \, {\rm d}V
\label{eqn: nu luminosity}
\end{equation}
and
\begin{equation}
L_{N,\nu} = \int_V \: \rho \left( R^0_{\nu} - H^0_{\nu} \right) \, {\rm d}V.
\label{eqn: nu N luminosity}
\end{equation}
where $V$ is the volume of the domain.
The explicit distinction between the emission and the absorption contributions, as well as
their dependence on the spatial position, allows the introduction of two supplementary
luminosities:\\
1) The \emph{cooling luminosities}, $L_{\nu,{\rm cool}}$ and $L_{N,\nu,{\rm cool}}$,
obtained by neglecting the heating rates $H_{\nu}^1$ and $H_{\nu}^0$ in \refeq{eqn: nu luminosity}
and \refeq{eqn: nu N luminosity}, respectively.\\
2) The \emph{HMNS luminosities}, $L_{\nu,{\rm HMNS}}$ and $L_{N,\nu,{\rm HMNS}}$, obtained by
restricting the volume integral in \refeq{eqn: nu luminosity} and \refeq{eqn: nu N luminosity}
to $V_{\rm HMNS}$, the volume of the central object.
Due to the continuous transition between the HMNS and the disc, the definition of $V_{\rm HMNS}$
is somewhat arbitrary. We decide to include also
the innermost part of the disc, delimited by a density contour of $5 \times 10^{11} {\rm g \, cm^{-3}}$.
This corresponds to the characteristic density close to the innermost stable orbit for a torus
accreting on stellar black holes.
It is also comparable with the surface density of a cooling proto-neutron star.
For each luminosity we associate a \emph{neutrino mean energy},
defined as $\langle E_{\nu} \rangle \equiv L_{\nu}/L_{N,\nu}$.\\
Since the scheme is spectral, all the terms on the right hand side of \refeq{eqn: yedot contributions},
\refeq{eqn: edot contributions} and \refeq{eqn: vdot contributions}
are energy-integrated values of spectral emission ($r_{\nu}$), absorption ($h_{\nu}$) and stress ($\mathbf{a}_{\nu}$) rates:
\begin{equation}
R^n_{\nu} = \int_0^{+\infty} \! {r_{{\nu}}}\, E^{n+2} \, dE , \label{eqn: particle cooling rate}
\end{equation}
\begin{equation}
H^n_{\nu} = \int_0^{+\infty} \! {h_{{\nu}}}\, E^{n+2} \, dE , \label{eqn: particle heating rate}
\end{equation}
\begin{equation}
\left( \frac{{\rm d} \mathbf{v}}{{\rm d} t} \right) _{\nu} = \int_{0}^{+\infty} \mathbf{a}_{\nu} \, E^2 \; {\rm d} E. \label{eqn: stress rate}
\end{equation}
The calculation of $r_{\nu}$, $h_{\nu}$ and $\mathbf{a}_{\nu}$ is the ultimate purpose of the ASL scheme.
\begin{figure}
\begin{center}
\includegraphics[width = 0.8 \linewidth]{fig2.pdf}
\end{center}
\caption{Schematic plot of the seven directions (paths) used to compute the optical depth at each point
of the cylindrical domain.}
\label{fig: tau_paths}
\end{figure}
The neutrino optical depths $\tau_\nu$'s play a central role in our scheme.
We distinguish between the {\em scattering} ($\tau_{\nu,{\rm sc}}$)
and the {\em energy} ($\tau_{\nu,{\rm en}}$) spectral optical depth.
The first one is obtained by summing all the relevant neutrino processes:
\begin{equation}
\label{eqn: tau cl def}
{\rm d} \tau_{\nu,{\rm sc}} = \rho \left( k_{\rm sc} + k_{\rm ab} \right) \, {\rm d}s
\end{equation}
where ${\rm d}s$ is an infinitesimal line element, and $k_{\rm ab}$ and $k_{\rm sc}$ are the neutrino opacities
for absorption and scattering, respectively.
For the second, more emphasis is put on those inelastic processes, that are effective in keeping
neutrinos in thermal equilibrium with matter.
In this case, we have
\begin{equation}
\label{eqn: tau eff def}
{\rm d} \tau_{\nu,{\rm en}} = \rho \sqrt{ k_{\rm ab} \left( k_{\rm sc} + k_{\rm ab} \right)} \, {\rm d} s,
\end{equation}
where we have considered absorption processes as inelastic, and scattering processes as elastic\footnote{This is not true in
general. However, it applies to the set of reactions we have chosen for our model. See \reftab{table: nu reactions}.}.
The values of the spectral $\tau_{\nu}$'s at each point are calculated using a local ray-by-ray method.
It consists of integrating \refeq{eqn: tau cl def} and
\refeq{eqn: tau eff def}
along several predefined paths and taking the minimum values among them.
These paths are straight oriented segments, connecting the considered point with
the edge of the computational domain.
Due to the intrinsically global character of
these integrations, we decided to exploit also here the expected symmetry of
the remnant, and to calculate $\tau_{\nu}$ in axial symmetry.
The seven different paths we explore in the $(R_{\rm cyl}-z)$ plane are shown in \reffig{fig: tau_paths}.
As a future step, we plan to include the more sophisticated and geometrically more flexible
MODA methods \citep[][]{Perego2014} to compute $\tau_{\nu}$.\\
The optical depths vary largely and they decrease, following the density profile,
proceeding from the HMNS to the edge of the remnant. To characterise this behaviour, we define
the unit vector
\begin{equation}
\mathbf{\hat{n}}_{\tau} \equiv - \nabla \tau_{\nu,{\rm sc}} / \left| \nabla \tau_{\nu,{\rm sc}} \right|,
\end{equation}
computed at each point of the domain from finite differences on the grid.
This vector will be crucial later to model the diffusion and the final
emission of the neutrinos.\\
The surfaces where $\tau_{\nu}$ equals 2/3 are defined as {\it neutrino surfaces}.
The neutrino surfaces obtained from
$\tau_{\nu,{\rm sc}}$ can be understood as the \emph{last scattering surfaces}; the ones derived
from $\tau_{\nu,{\rm en}}$ correspond to the surfaces where neutrinos decouple thermally from matter,
and they are often called \emph{energy surfaces} \citep[see, for example,][]{Raffelt2001}.
According to the value of $\tau_{\nu,{\rm sc}}$, we distinguish between three disjoint volumes:
1) $V_{\rm thin}$, for the optically thin region ($\tau_{\nu,{\rm sc}} \ll 2/3$); 2) $V_{\rm surf}$, for the neutrino
surface region\footnote{In principle, the neutrino surfaces should have no volume. However, due
to the discretisation on the (axisymmetric) grid we adopted to calculate $\tau$,
every neutrino surface is replaced by a shell of width
$\sim \Delta x$. This thin layer is formed by the cells $\mathbf{x}$
inside which $\tau$ is expected to become equal to 2/3.} ($\tau_{\nu,{\rm sc}} \sim 2/3$);
3) $V_{\rm thick}$, for the optically thick region ($\tau_{\nu,{\rm sc}} \gg 2/3$).
Obviously, $V = V_{\rm thick} \, \cup \, V_{\rm surf} \, \cup \, V_{\rm thin} $.\\
\begin{table}
\begin{center}
\begin{tabular}{| l | l | l |}
\hline
Quantity & Definition & Related quantities \\ \hline
$j_{\rm em}$ & emissivity & $r_{\nu,{\rm prod}}$ \\
$k_{\rm ab}$ & absorption opacity & $\lambda_{\nu}$, $\tau_{\nu}$, $h_{\nu}$, $\mathbf{a}_{\nu}$ \\
$k_{\rm sc}$ & scattering opacity & $\lambda_{\nu}$, $\tau_{\nu}$ \\ \hline
$\lambda_{\nu}$ & mean free path & $r_{\nu,{\rm diff}}$ \\
$\tau_{\nu}$ & optical depth & $\hat{\mathbf{n}}_{\tau}$, $\hat{\mathbf{n}}_{\rm path}$, $r_{\nu,{\rm diff}}$, $r_{\nu,{\rm ult}}$, \\
& & $\quad h_{\nu} $, $\mathbf{a}_{\nu} $ \\
$\hat{\mathbf{n}}_{\tau}$ & opposite $\tau$ gradient & $\hat{\mathbf{n}}_{\rm path}$, $n_{\nu} $, $\mathbf{s}_{\nu} $ \\
$\hat{\mathbf{n}}_{\rm path}$ & diffusion direction & $r_{\nu,{\rm ult}}$ \\ \hline
$r_{\nu,{\rm prod}}$ & production rate & $r_{\nu}$ \\
$r_{\nu,{\rm diff}}$ & diffusion rate & $r_{\nu}$ \\
$r_{\nu}$ & emission rates & $r_{\nu,{\rm ult}}$, $( {\rm d} Y_e/{\rm d}t)_{\nu}$, $ ({\rm d}e/{\rm d}t)_{\nu}$ \\ \hline
$r_{\nu,{\rm ult}} $ & ultimate emission rates & $n_{\nu}$, $\mathbf{s}_{\nu} $ \\
$n_{\nu} $ & particle density & $h_{\nu} $ \\
$\mathbf{s}_{\nu} $ & momentum density & $\mathbf{a}_{\nu}$ \\
$h_{\nu} $ & absorption rate & $( {\rm d}Y_e/{\rm d}t)_{\nu}$, $ ({\rm d}e/{\rm d}t)_{\nu}$ \\
$\mathbf{a}_{\nu} $ & stress & $( {\rm d} \mathbf{v}/ {\rm d}t)_{\nu}$ \\ \hline
\end{tabular}
\end{center}
\caption{List of the most important spectral quantities appearing in the ASL scheme (left column) and
their definition (central column). In the right column, we list the relevant
quantities (spectral quantities and source terms) that depend directly on each table entry.
See the text for more details.}
\label{table: ASL summary}
\end{table}
After having introduced $\tau_\nu$, we can now explain in which way the neutrino rates
are calculated within the ASL scheme.
In \reftab{table: ASL summary} we have summarised the most important quantities, their
definitions and relations in the context of the ASL scheme.\\
The spectral emission rates
$r_{\nu}$ are calculated as smooth interpolation between diffusion ($r_{\nu,{\rm diff}}$)
and production ($r_{\nu,{\rm prod}}$) spectral rates: the first ones are the relevant rates
in the optically thick regime, the latter in the optically transparent region.\\
We compute $r_{\nu,{\rm prod}}$ and $r_{\nu,{\rm diff}}$ as
\begin{eqnarray}
\lefteqn{r_{\nu,{\rm prod}} = \frac{4 \pi}{\left( hc \right)^3} \frac{ j_{\rm em} }{ \rho },}
\label{eqn: production rates} \\
\lefteqn{r_{\nu,{\rm diff}} = \frac{4 \pi}{ \left( hc \right)^3} \frac{ f_{\nu}^{\rm FD}}{ \rho \, t_{\nu,{\rm diff}}}.}
\label{eqn: diffusion rates}
\end{eqnarray}
$j_{\rm em}$ is the neutrino spectral emissivity, while
$f_{\nu}^{\rm FD}$ is the Fermi-Dirac distribution function
for a neutrino gas in thermal and weak equilibrium with matter.
$t_{\nu,{\rm diff}}$ is the local diffusion time-scale, calculated as
\begin{equation}
t_{\nu,{\rm diff}} = \alpha_{\rm diff} \frac{\tau_{\nu,{\rm sc}}^2 \, \lambda_{\nu,{\rm sc}}}{c}
\label{eqn: diffusion timescale}
\end{equation}
where $\lambda_{\nu,{\rm sc}} = \left( \rho \left( k_{\nu,{\rm ab}} + k_{\nu,{\rm sc}} \right) \right)^{-1}$ is the
total mean free path.
$\alpha_{\rm diff}$ is a constant set to 3.
The interpolation formula for $r_{\nu}$ is provided by half of
the harmonic mean between the production and diffusion rates.\\
We compute the spectral heating rate as the properly normalised product of the
absorption opacity $k_{\nu,{\rm ab}}$ and of the spectral neutrino density $n_{\nu}$:
\begin{eqnarray}
h_{\nu} = c\, k_{\nu,{\rm ab}} \, n_{\nu} \, \mathcal{F}_{e,\nu} \, \mathcal{H}.
\label{eqn: heating term outside}
\end{eqnarray}
$\mathcal{H} \equiv \exp(- \tau_{{\nu},\rm sc})$ is an exponential cut off that ensures
the application of the heating term only outside the neutrino surface, and
$\mathcal{F}_{e,\nu}$ is the Pauli blocking factor for electrons or positrons in the final state.
$n_{\nu}$ is defined so that the energy-integrated particle density $N_{\nu}$ is given by:
\begin{equation}
\label{eqn: spectral neutrino density}
N_{\nu} = \int_{0}^{+ \infty} \, n_{\nu} \, E^2 \, {\rm d} E.
\end{equation}
The stress term is calculated similarly to the neutrino heating rate:
\begin{equation}
\mathbf{a}_{\nu} = c \, k_{\nu,{\rm ab}} \, \mathbf{s}_{\nu} \, \mathcal{F}_{e,\nu} \; \mathcal{H},
\label{eqn: spectral stress rate}
\end{equation}
where $\mathbf{s}_{\nu}$ is the spectral density of linear momentum associated with the streaming neutrinos,
while $\mathcal{H}$ and $\mathcal{F}_{e,\nu}$ are defined as in \refeq{eqn: heating term outside}.\\
\begin{figure}
\begin{center}
\includegraphics[width = \linewidth]{fig3.png}
\end{center}
\caption{Schematic representation of the procedure to calculate the ultimate emission rates at the
neutrino surface and in the optically thin region, $r_{\nu,{\rm ult}}$, from the emission
rates, $r_{\nu}$. The thin black arrows represent the inverse of the gradient of
$\tau_{\rm sc}$ ($\mathbf{\hat{n}}_{\tau}$),
while the thick red arrow is $\mathbf{\hat{n}}_{\rm path}$.
Label $x_{A}$ refers to a point inside the neutrino surface (opaque region),
while $x_{B}$ is a point inside the disc, but in the optically thin zone, for which
$\mathbf{\hat{n}}_{\rm tau} = \mathbf{\hat{n}}_{\rm path}$. See the text for more details.}
\label{fig: path}
\end{figure}
The quantities $n_{\nu}$ and $\mathbf{s}_{\nu}$ are computed using a multidimensional
ray-tracing algorithm.
This algorithm assumes that neutrinos (possibly, after having diffused from the optically thick region)
are \emph{ultimately} emitted isotropically at the neutrino surface and in the
optically transparent region.
If we define $l_{\nu}({\mathbf{x}',\mathbf{\hat{n}}})$
as the specific rate per unit solid angle of the radiation
emitted from a point $\mathbf{x}' \in \left( V_{\rm surf} \cup V_{\rm thin} \right)$, in the direction $\mathbf{\hat{n}}$, then
\begin{equation}
\label{eqn: nu_dens_eq}
n_{\nu}(\mathbf{x}) = \int_{V_{\rm surf} \, \cup \, V_{\rm thin}} \:
\rho \, \frac{ l_{\nu} \left(\mathbf{x}',\mathbf{\hat{n}(\mathbf{x},\mathbf{x}')} \right)}{c \left| \mathbf{x}' - \mathbf{x} \right|^2} {\rm d}^3 \mathbf{x}'
\end{equation}
and
\begin{equation}
\mathbf{s}_{\nu}(\mathbf{x}) = \int_{V_{\rm surf} \, \cup \, V_{\rm thin}} \:
\rho \, \frac{l_{\nu}
\left( \mathbf{x}',\mathbf{\hat{n}(\mathbf{x},\mathbf{x}')} \right)}{c \left| \mathbf{x}' - \mathbf{x} \right|^2}
\, \frac{E}{c} \, \mathbf{\hat{n}(\mathbf{x},\mathbf{x}')} \; {\rm d}^3 \mathbf{x}'. \label{eqn: spectral stress vector}
\end{equation}
where $\mathbf{\hat{n}}(\mathbf{x},\mathbf{x}') =
(\mathbf{x}' - \mathbf{x})/( \left| \mathbf{x}' - \mathbf{x} \right|)$.
The isotropic character of the emission allows us to introduce the angle-integrated \emph{ultimate emission rates}
$r_{\nu,{\rm ult}}$ as:
\begin{equation}
l_{\nu}(\mathbf{\hat{n}}) =
\left\{
\begin{array}{r l}
r_{\nu,{\rm ult}}/\left( 2 \pi \right) & {\rm if} \; \mathbf{\hat{n}} \cdot \mathbf{\hat{n}}_{\tau} \geq 0 \\
0 & \mbox{otherwise}.
\end{array}
\right.
\label{eqn: ultimate rates}
\end{equation}
$r_{\nu,{\rm ult}}$ and $r_{\nu}$ can differ locally, but
they have to provide the same cooling (spectral) luminosities:
\begin{equation}
\int_V \rho \, r_{\nu} \, {\rm d}V = \int_{V} \rho \, r_{\nu,{\rm ult}} \, {\rm d}V .
\label{eqn: r_rad integral 1}
\end{equation}
Since $r_{\nu,{\rm ult}}$ represents the ultimate emission rate, {\it after} the diffusion process has drained
neutrinos from the opaque region to the neutrino surface, $r_{\nu,{\rm ult}} = 0$ inside $V_{\rm thick}$.
On the other hand, inside $V_{\rm thin}$ diffusion does not take place and $r_{\nu,{\rm ult}} = r_{\nu}$.
In light of this, \refeq{eqn: r_rad integral 1} becomes
\begin{equation}
\int_{V_{\rm thick} \, \cup \, V_{\rm surf}} \rho \, r_{\nu} \, {\rm d}^3 \mathbf{x} =
\int_{V_{\rm surf}} \rho \, r_{\nu,{\rm ult}} \, {\rm d}^3 \mathbf{x}.
\label{eqn: r_rad integral 2}
\end{equation}
\refeq{eqn: r_rad integral 2} has a clear physical interpretation: inside $V_{\rm surf}$,
$r_{\nu,{\rm ult}}$ is obtained 1) from the emission rate, $r_{\nu}$, at the neutrino surface and
2) from the re-mapping of the emission rates obtained in the opaque region onto the neutrino surface,
as a consequence of the diffusion process.
A careful answer to this re-mapping problem would rely on the solution of the diffusion equation
in the optically thick regime and of the Boltzmann equation in the semi-transparent region.
The ASL algorithm calculates the amount of neutrinos diffusing from a certain volume element.
But it does not provide information about the angular dependence of their flux, neither
about the point of the neutrino surface where they are ultimately emitted. Thus, a phenomenological model is required.
When the properties of the system under investigation
change on a time-scale larger than (or comparable to)
the relevant diffusion time-scale (see \refsec{sec: preliminary analysis and estimates}),
the neutrino fluxes can be considered as quasi-stationary.
Under these conditions, the statistical interpretation of the optical depth, as the average number
of interactions experienced by a neutrino before escaping, suggests to consider
$\mathbf{\hat{n}}_{\tau}$
as the local preferential direction for neutrino fluxes.
While in the (semi-)transparent regime, this unitary vector provides already the
favourite emission direction (see \refeq{eqn: ultimate rates}),
in the diffusion regime we have to take into account
the spatial variation of $\mathbf{\hat{n}}_{\tau}$.
To this end, at each point $\mathbf{x}$ in $V_{\rm thick}$,
we associate a point $\mathbf{x}_{\rm surf}(\mathbf{x})$ in $V_{\rm surf}$
and a related preferential direction
\begin{equation}
\mathbf{\hat{n}}_{\rm path}\left( \mathbf{x} \right) =
\frac{ \mathbf{x}_{\rm surf}(\mathbf{x}) - \mathbf{x} }{\left| \mathbf{x}_{\rm surf}(\mathbf{x})-\mathbf{x} \right|},
\end{equation}
according to the following prescription:
the points $\mathbf{x}$ and $\mathbf{x}_{\rm surf}$ are connected by a non-straight path $\gamma$
that has $\mathbf{\hat{n}}_{\tau}$ as local gradient:
$\gamma(s):\left[0, 1 \right] \rightarrow \left[\mathbf{x}, \mathbf{x}_{\rm surf} \right]$, $\mathbf{x} \in V_{\rm thick}$,
$\mathbf{x}_{\rm surf} \in V_{\rm surf}$, and ${\rm d} \gamma / {\rm d} s = \mathbf{n}_{\tau}$.
This procedure is sketched in \reffig{fig: path}.\\
Once $\mathbf{\hat{n}}_{\rm path}$ has been calculated everywhere inside $V_{\rm thick}$,
we can re-distribute the neutrinos coming from the optically
thick region on the neutrino surface.
This is done assuming that neutrinos coming from a point $\mathbf{x}$
are emitted preferentially from points of the neutrino surface located around $\mathbf{x}_{\rm surf}(\mathbf{x})$.
More specifically, from points $\mathbf{x}'$ for which 1) $\mathbf{x}' \in V_{\rm surf}$; and
2) $\mu(\mathbf{x},\mathbf{x}') \equiv \mathbf{\hat{n}}(\mathbf{x},\mathbf{x}') \cdot \mathbf{\hat{n}}_{\rm path}(\mathbf{x}) > 0 $, where
$\mathbf{\hat{n}}(\mathbf{x},\mathbf{x}') \equiv \left( \mathbf{x}' - \mathbf{x} \right) / \left| \mathbf{x}' - \mathbf{x} \right|$.
If $\mathbf{\hat{n}}$ and $\mathbf{\hat{n}}_{\rm path}$ are close to the parallel condition (i.e. $\mu \approx 1$) we
expect more neutrinos than in the case of perpendicular directions (i.e. $\mu \approx 0$ ).
We smoothly model this effect assuming a $\mu^2$ dependence. \\
The global character of this re-mapping procedure
represents a severe computational limitation for our large, three dimensional, MPI-parallelised Cartesian
simulation. In order to make the calculation feasible,
we take again advantage of the expected high degree of axial symmetry of remnant
(especially in the innermost part
of it, where the diffusion takes place and most of the neutrino are emitted),
and we compute $r_{\nu,{\rm ult}}$ in axisymmetry.
\subsection{Initial Conditions}
\label{sec: initial conditions}
\begin{figure*}
\begin{center}
\includegraphics[width = \linewidth]{fig4.png}
\end{center}
\caption{Vertical slices of the three dimensional domain (corresponding to the $y=0$ plane), recorded at
the beginning of the simulation. In the left panel, we color coded the logarithm of the matter density (in ${\rm g/cm^3}$, left side)
and the projected fluid velocity (in units of $c$, on the right side);
the arrows indicate the direction of the projected velocity in the plane). On the right panel, we represent the
electron fraction (left side) and the logarithm of the matter temperature (in unit of MeV, right side).}
\label{fig: initial conditions}
\end{figure*}
The current study is based on previous, 3D hydrodynamic studies of the merger of two
non-spinning 1.4 ${\rm M}_{\sun}$ neutron stars. This simulation was performed with a
3D Smoothed Particle Hydrodynamics (SPH) code, the implementation details of which
can be found in the literature \citep{Rosswog2000,Rosswog2003,Rosswog2005,
Rosswog2007c}. For overviews over the SPH method, the interested reader is referred to recent reviews
\citep{Monaghan2005,Rosswog2009b,Springel2010a,Price2012a,Rosswog2014b,Rosswog2014c}.
The neutron star matter is modelled with the Shen et al. EoS \citep{Shen1998a,Shen1998b},
and the profiles of the density and $\beta$-equilibrium electron fraction can be found in fig. 1
of \cite{Rosswog2013}.
During the merger process the debris can cool via neutrino emission, and electron/positron captures
can change the electron fraction. These processes are included via the
opacity-dependent, multi-flavor leakage scheme of \cite{Rosswog2003}. Note, however,
that no heating via neutrino absorption is included. Their effects are
the main topic of the present study.\\
As the starting point of our neutrino-radiation hydrodynamics study, we consider the matter distribution
of the 3D SPH simulation with $10^6$ particles, at 15 ms after the first contact
(corresponding to 18 ms after the simulation start).
Not accounting for the neutrino absorption during this short time, should only have a
small effect, since, according to the estimates from \refsec{sec: preliminary analysis and estimates},
the remnant hardly had time to change.\\
We map the 3D SPH matter distributions of density, temperature, electron fraction
and fluid velocity on the Cartesian, equally spaced
grid of \texttt{FISH}, with a resolution of 1 {\rm km}.
The initial extension of the grid is
$\left( 800 {\rm km} \times 800 {\rm km} \times 640 {\rm km} \right)$.
During the simulation, we increase the domain in all directions to follow the wind expansion,
keeping the HMNS always in the centre.
At the end, the computational box is
$\left( 2240 {\rm km} \times 2240 {\rm km} \times 3360 {\rm km} \right)$ wide.\\
The initial data cover a density range of
$ 10^8 {\rm g \, cm^{-3}} - 3.5 \times 10^{14} {\rm g \, cm^{-3}}$.
Surrounding the remnant, we place an inert atmosphere, characterised
by the following stationary properties:
$\rho_{\rm atm} = 5 \cdot 10^{3}\, {\rm g/cm^3}$, $T_{\rm atm} = 0.1
\, {\rm MeV}$, $Y_{e,{\rm atm}}=0.01$ and $\mathbf{v}_{\rm atm} =
\mathbf{0}$.
The neutrino source terms are set to 0 in this atmosphere. With this treatment, we
minimize the influence of the atmosphere on the disc and on the wind dynamics.
Even though in our model we try to stay as close as possible to the choices adopted in the SPH simulation,
initial transients appear at the start of the simulation.
One of the causes is the difference in the spatial resolutions between the two models.
The resolution we are adopting in \texttt{FISH} is significantly lower than the one
provided by the initial SPH model inside the HMNS, $\sim 0.125 \, {\rm km}$, (which is
necessary to model consistently the central object), while it is comparable or better inside the disc.
Due to this lack of resolution, we decide
to treat the HMNS as a stationary rotating object.
To implement this, we perform axisymmetric averages
of all the hydrodynamical quantities
at the beginning of the simulation.
At the end of each hydrodynamical time step, we re-map these profiles
in cells contained inside an ellipsoid,
with $a_x = a_y = 30 \, {\rm km}$ and $a_z = 23 \, {\rm km}$,
and for which $\rho > 2 \cdot 10^{11} {\rm g/cm^3}$.
For the velocity vector, we consider only the azimuthal component,
since 1) the HMNS is rotating fast around its polar axis (with a period $P \approx 1.4 \, {\rm ms}$)
and 2) the non-azimuthal motion inside it is characterised by much smaller velocities (for example,
$\left| v_R \right| \sim 10^{-3} \left| v_\phi \right| $, where $v_R$ and $ v_\phi $ are the radial and the
azimuthal velocity components).
Concerning the density and the rotational velocity profiles, our treatment is consistent
with the results obtained by \cite{Dessart2009} (fig. 4), who showed that
$\sim 100 \, {\rm ms}$ after the neutron star have collided
those quantities have changed only slightly inside the HMNS.
We expect the electron fraction and the temperature also to stay close to their initial values,
since the most relevant neutrino surfaces for $\nu_e$ and $\bar{\nu}_e$
are placed outside the stationary region and the diffusion time-scale is much longer than the simulated
time (see, for example, \refsec{sec: preliminary analysis and estimates}).
\begin{figure}
\includegraphics[width=\linewidth]{fig5.png}
\caption{Logarithm of the matter density (color coded, in ${\rm g \, cm^{-3}}$)
and isocontours of the gravitational energy (white lines, in ${\rm MeV \, baryon^{-1}}$),
on a vertical slice of the three dimensional domain, at $t=0$.}
\label{fig: rho and grav}
\end{figure}
To give the opportunity to the system to adjust to a more stable configuration on the new grid,
we consider the first $ 10 \, {\rm ms}$ of the simulation as a ``relaxation phase''.
During this phase, we evolve the system considering only neutrino emission.
Its duration is chosen so that the initial transients arrive at
the disc edge, and the profiles inside the disc reach new quasi-stationary
conditions. The ``relaxed'' conditions are visible in \reffig{fig: initial conditions}.
They are considered as the new initial conditions and
we evolved them for $\sim 90 \, {\rm ms}$, including the
effect of neutrino absorption.
In the following, the time $t$ will be measured with respect to this second re-start.
During the relaxation phase, we notice an increase of the electron
fraction, from 0.05 up to 0.1-0.35, for a tiny amount of matter ($\la 10^{-5}$ ${\rm M}_{\sun}$) in the low density region
($\rho \la 10^9 {\rm g/cm^3}$) situated above the
innermost, densest part of the disc ($R_{\rm cyl} \la 50 \, {\rm km}$, $\left| z \right| \ga 20 \, {\rm km}$).
Here, the presence of neutron-rich, hot matter
in optically thin conditions favours the emission of $\bar{\nu}_e$, via
positron absorption on neutrons.
A similar increase of $Y_e$ is also visible in the original SPH simulations,
for times longer than 15 ms after the first collision.\\
In \reffig{fig: rho and grav} we show isocontours of the absolute value of gravitational specific energy,
drawn against the colour-coded matter density, at the beginning of our simulation.
The gravitational energy provides an estimate of the energy that neutrinos have to deposit to unbound matter,
at different locations inside the disc (see \refsec{sec: preliminary analysis and estimates}).
\section{Simulation results}
\label{sec: results}
\subsection{Disc evolution and matter accretion}
\label{sec: disc evolution}
\begin{figure}
\begin{center}
\includegraphics[width = \linewidth]{fig6.png}
\end{center}
\caption{Vertical slice of the inner part of the three dimensional domain ($y=0$ plane),
taken at $41 \, {\rm ms}$ after the beginning
of the simulations. Color coded is the radial component of the fluid velocity.
The two coloured hemispheres in the centre represent the stationary central object for which $v_r \approx 0$
(the two actual colours are very small numbers).}
\label{fig: vrad_acc}
\end{figure}
After the highly dynamical merger phase, the remnant is still dynamically evolving and
not yet in a perfectly stationary state.\\
In \reffig{fig: vrad_acc}, we show the radial component
of the fluid velocity on the $y=0$ plane, at $ 41 \, {\rm ms}$ after the beginning of the simulation.
The central part of the disc, corresponding to a density contour of
$\sim 5 \cdot 10^9 \, {\rm g \, cm^{-3}}$, is slowly being
accreted onto the HMNS ($v_R \sim$ a few $10^{-3}c$), while the outer edge is gradually expanding
along the equatorial direction. The velocity profile shows interesting asymmetries and deviations
from an axisymmetric behaviour.
The surface of the HMNS and the innermost part of the disc are characterised by steep gradients of
density and temperature, and they behave like a pressure wall for the infalling matter. Outgoing
sound waves are then produced and move outwards inside the disc, transporting energy, linear
and angular momentum.
At a cylindrical radius of $R_{\rm cyl} \la 80 \, {\rm km}$, they induce
small scale perturbations in the velocity field, visible as bubbles of slightly positive
radial velocity.
These perturbations dissolve at larger radii, releasing their momentum and energy inside the disc, and
favouring its equatorial expansion.
\begin{figure}
\begin{center}
\includegraphics[width = \linewidth]{fig7.png}
\end{center}
\caption{Temporal evolution of the accretion rate on the HMNS, $\dot{M}$, calculated
as the net flux of matter crossing a cylindrical surface of radius
$R_{\rm cyl}=35 \, {\rm km}$ and axis corresponding to the rotational axis of the disc.}
\label{fig: accretion_eta_time}
\end{figure}
\noindent The temporal evolution of the accretion rate $\dot{M}$,
computed as the net flux of matter crossing a cylindrical surface of radius
$R_{\rm cyl}=35 \, {\rm km}$ and axis corresponding to the rotational axis of the disc,
is plotted in \reffig{fig: accretion_eta_time}.
This accretion rate is compatible with the estimate performed in
\refsec{sec: preliminary analysis and estimates} using an $\alpha$-viscosity disc model.
A direct comparison with \refeq{eqn: mdot} suggests
an effective parameter $\alpha \approx 0.05$ for our disc.
We stress again that no physical viscosity
is included in our model: the accretion is driven by unbalanced pressure gradients,
neutrino cooling (see \refsec{sec: neutrino emission})
and dissipation of numerical origin.
However, the previous estimate is useful to compare our disc with
purely Keplerian discs, in which a physical $\alpha$-viscosity has been included
(usually, with $0.01 \la \alpha \la 0.1$). Our value of $\alpha \approx 0.05$ is close to
what is usually assumed for such discs ($\sim 0.1$). Higher viscosities would enhance
the neutrino emission and probably the mass loss.\\
\begin{figure*}
\begin{center}
\includegraphics[width = \linewidth]{fig8.png}
\end{center}
\caption{Radial (upper row) and vertical (lower row) profiles of the axisymmetric density (solid lines)
and temperature (dashed lines)
inside the disc, recorded at different times during the simulation ($t \approx 2 \, {\rm ms}$ (black-thick lines),
$t \approx 45 \, {\rm ms}$ (blue-normal lines),
$t \approx 80 \, {\rm ms}$ (red-thin lines)).
The different columns correspond to different values of the section coordinates: from left to right,
$ z = 0 \, {\rm km}, \, 20 \, {\rm km}, \, 40 \, {\rm km}$ for the radial profiles;
$R_{\rm cyl} = 35 \, {\rm km}, \, 70 \, {\rm km}, \, 140 \, {\rm km}$ for the vertical ones.}
\label{fig: disc profiles}
\end{figure*}
On a timescale of a few tens of milliseconds, the profiles inside the disc change, as
consequence of the accretion process and of the outer edge expansion.
These effects are visible in the upper row of \reffig{fig: disc profiles}, where
radial profiles of temperature and density are drawn, at different times and heights
inside the disc.
We notice, in particular, that the density decreases in the internal part of the disc
($50 \, {\rm km} \la R_{\rm cyl} \la 200 \, {\rm km}$),
as result of the accretion.
In the same region, the balance between
the increase of internal energy and the efficient cooling provided by neutrino emission
keeps the temperature almost stationary.
At larger radial distances ($R_{\rm cyl} \ga 200 \, {\rm km}$),
the initial accretion of a cold, thin layer of matter (visible in the $t=2 \, {\rm ms}$ profiles) is followed
by the continuous expansion of the outer margin of the hot internal disc.
\subsection{Neutrino emission}
\label{sec: neutrino emission}
\begin{figure*}
\begin{center}
\includegraphics[width = \linewidth]{fig9.png}
\end{center}
\caption{Location of the neutrino surfaces for $\nu_e$ (left column), $\bar{\nu}_e$ (central column) and $\nu_{\mu,\tau}$ (right column),
for the scattering optical depth (upper row) and for the energy optical depth (bottom row), $40 \, {\rm ms}$ after
the beginning of the simulation.
Color coded is the logarithm of cylindrically averaged matter density, $\rho \, [{\rm g/cm^3}]$.
The different lines correspond to the neutrino surfaces for
different values of the neutrino energy: from the innermost line to the outermost one, $E_{\nu} =
4.62 \, {\rm MeV}, \, 10.63 {\rm MeV}, \, 16.22 {\rm MeV}, \, 24.65 {\rm MeV}, \, 56.96 {\rm MeV} $. }
\label{fig: nu_surf_tot}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width = \linewidth]{fig10.pdf}
\end{center}
\caption{Time evolution of the net (solid) and cooling (dashed) luminosities obtained
by the ASL scheme for $\nu_e$ (black-thick), $\bar{\nu}_e$ (blue-normal)
and $\nu_{\mu,\tau}$ (red-thin) neutrino species. The difference between the cooling and the net luminosities is
represented by the re-absorbed luminosity. The contributions to the cooling luminosities
coming from the HMNS, defined as the volume characterised by $\rho > 5 \cdot 10^{11} {\rm g \, cm^{-3}}$,
is also plotted (dot-dashed lines).
Note that for $\nu_{\mu,\tau}$, the net and the cooling luminosities coincide,
and they are almost equal to the HMNS contribution.}
\label{fig: lum with time}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width = \linewidth]{fig11.png}
\end{center}
\caption{Energy-integrated (axisymmetric) neutrino density of the free-streaming neutrinos,
$N_{\nu}$, for $\nu_e$ (left panel) and $\bar{\nu}_e$ (right panel), calculated outside the innermost neutrino surface
(corresponding to $E_{\nu} = 3 \, {\rm MeV}$), at $t \approx 40 \, {\rm ms}$ after the beginning of the simulation.}
\label{fig: nu dens}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width = \linewidth]{fig12.pdf}
\end{center}
\caption{Angular dependence of the isotropised neutrino cooling luminosities (solid line)
and of the neutrino mean energies (dashed lines), as a function of the colatitude.
The black-thick lines correspond to $\nu_e$, while the blue-thin lines to $\bar{\nu}_e$.
As a representative time, we consider $t \approx 40 {\rm ms}$ after the
beginning of the simulation.}
\label{fig: ang lum aven}
\end{figure}
In \reffig{fig: nu_surf_tot}, we show the neutrino
surfaces obtained by the calculation of the spectral neutrino optical depths,
together with the matter density distribution (axisymmetric, color coded).
Different lines correspond to different energy bins.
In the upper panels, we represent the scattering neutrino surfaces, while in the lower panels
the energy ones. Their shapes follow closely the matter density distribution,
due to the explicit dependence appearing in \refeq{eqn: tau cl def} and \refeq{eqn: tau eff def}.
The last scattering surfaces for the
energies that are expected to be more relevant for the neutrino emission
($10 \, {\rm MeV} \la E_{\nu} \la 25 \, {\rm MeV} $, corresponding to the expected range for the mean energies,
as we will discuss below) extend far outside in the disc,
compared with the radius of the central object.
$\nu_e$'s have the largest opacities, due to the extremely neutron rich environment that
favours processes like neutrino absorption on neutrons.
Since the former reaction is also very efficient in thermalising neutrinos,
the scattering and the energy neutrino
surfaces are almost identical for $\nu_e$'s.
In the case of $\bar{\nu}_e$'s, the relatively low density of free protons
determines the reduction of the scattering and, even more, of the energy optical depth.
For $\nu_{\mu,\tau}$'s, neutrino bremsstrahlung and $e^+-e^-$ annihilation freeze out
at relatively high densities
and temperatures ($\rho \sim 10^{13} \, {\rm g/cm^3}$ and $k_{\rm B}T \sim 8 \, {\rm MeV}$),
reducing further the energy neutrino surfaces, while elastic scattering on nucleons
still provides a scattering opacity comparable to the one of $\bar{\nu}_e$'s.
The energy- and volume-integrated luminosities obtained during the simulation are presented
in \reffig{fig: lum with time}.
The cooling luminosities for $\nu_e$'s and $\bar{\nu}_e$'s (dashed lines)
decrease weakly and almost linearly with time.
This behaviour reflects the continuous supply of hot accreting matter.
The faster decrease of $\dot{M}$ (cf. \reffig{fig: accretion_eta_time})
would imply a similar decrease in the luminosities, if the neutrino radiative efficiency
of the disc were constant. However, the latter increases with time due to the decrease of density and
the constancy of temperature characterising the innermost part of the disc
(see \refsec{sec: disc evolution}).
Also the luminosity for the $\nu_{\mu,\tau}$ species is almost constant.
This is a consequence of the stationarity
of the central object, since most of the $\nu_{\mu,\tau}$'s come from there. However, this result
is compatible with the long cooling time-scale of the HMNS, \refeq{eqn: diffusion time HMNS}.
We specify here that the plotted lines for $\nu_{\mu,\tau}$ correspond to one single species. Thus, the
\emph{total} luminosity coming from heavy flavour neutrinos is four times the plotted one, see also
\refeq{eqn: edot contributions}.\\
In the case of $\nu_e$'s and $\bar{\nu}_e$'s, the luminosity provided by $V_{\rm HMNS}$
(defined in \refsec{sec:neutrino treatment} and represented by dot-dashed lines in \reffig{fig: lum with time})
and the luminosity of the accreting disc are comparable. This result is compatible with what is observed
in core collapse supernova simulations \citep[see, for example, fig. 6 of][]{Liebendorfer2005a}, a few tens of milliseconds after bounce:
assuming a density cut of $5 \times 10^{11} {\rm g \, cm^{-3}}$
for the proto-neutron star, its contribution
is roughly half of the total emitted luminosity, for both $\nu_e$ and $\bar{\nu}_e$.
Instead, if we further restrict $V_{\rm HMNS}$ only to the central ellipsoid
(see \refsec{sec: initial conditions} for more details),
we notice that the related luminosity reduces to $\la 10 \times 10^{51} {\rm erg \, s^{-1}}$
for all neutrino species. This is consistent with our preliminary estimate,
\refeq{eqn: HMNS luminosity}.\\
The inclusion of neutrino absorption processes in the optically thin region
reduces the cooling luminosities to the net luminosities
(solid lines in \reffig{fig: lum with time}).
For $\nu_e$'s, the neutron rich environment reduces the number luminosity by $\approx 37$ per cent, while
for $\bar{\nu}_e$'s this fraction drops to $\approx 14$ per cent.
The values of the neutrino mean energies are practically stationary during the simulation:
from the net luminosities at $t \approx 40 {\rm ms}$,
$\langle E_{\nu_e} \rangle \approx 10.6 \, {\rm MeV}$,
$\langle E_{\bar{\nu}_e} \rangle \approx 15.3 \, {\rm MeV}$
and $\langle E_{\nu_{\mu,\tau}} \rangle \approx 17.3 \, {\rm MeV}$.
The mean neutrino energies show the expected hierarchy,
$\langle E_{\nu_e} \rangle < \langle E_{\bar{\nu}_e} \rangle < \langle E_{\nu_{\mu,\tau}} \rangle$,
reflecting the different locations of the thermal decoupling surfaces.
While the values obtained for $\nu_e$'s and $\bar{\nu}_e$'s are consistent with previous calculations,
$\langle E_{\nu_{\mu,\tau}} \rangle$ is smaller than expected \citep[see, for example,][]{Rosswog2013}.
This is due to the lack of resolution at the HMNS surface,
where most of the energy neutrino surfaces for $\nu_{\mu,\tau}$ are located.
This discrepancy has no dynamical effects for us, since most of $\nu_{\mu,\tau}$ come from the
stationary central object.
The ray-tracing algorithm, see \refsec{sec:neutrino treatment}, allows us to compute
1) the neutrino densities outside the neutrino surfaces;
2) the angular distribution of the isotropised neutrino cooling luminosities and
mean neutrino energies, as seen by a far observer.
In \reffig{fig: nu dens}, we represent the energy-integrated
axisymmetric neutrino densities $N_{\nu}$, \refeq{eqn: nu_dens_eq},
for $\nu_e$ (left) and $\bar{\nu}_e$ (right).
These densities reach their maximum in the funnel above the HMNS,
due to the geometry of the emission and to the short distance from the most emitting regions.
At distances much
larger than the dimension of the neutrino surfaces, $N_{\nu}$ shows the expected $R^{-2}$ dependence.\\
The disc geometry introduces a clear anisotropy in the neutrino emission,
visible in \reffig{fig: ang lum aven}.
Due to the larger opacity along the equatorial direction, the isotropic luminosity along
the poles is $\sim 3 - 3.5$ more intense than the one along the equator.
The different temperatures at which neutrinos decouple from matter at different polar angles
determine the angular dependence of the mean energies.
\subsection{Neutrino-driven wind}
\begin{figure*}
\begin{center}
\includegraphics[width = \linewidth]{fig13.png}
\end{center}
\caption{Energy- and species-integrated axisymmetric $\nu$ net rates for energy
(left panel, in units of $10^{20}$ erg/g/s) and $Y_e$ (central panel, in units of 1/baryon/s),
and of the fluid velocity variation provided by neutrino absorption in the optically thin regions
(right panel, in units of $c/{\rm s}$).
As a representative time, we consider $t \approx 40 \, {\rm ms}$ after the beginning of the simulation.
The complex structure of the net $Y_e$ rate in the funnel, above the HMNS poles,
originates from the variety of conditions of $Y_e$, $\rho$ and $\mathbf{v}$ at that specific moment.}
\label{fig: net rates}
\end{figure*}
The evolution of the disc and the formation of a neutrino-driven wind depend crucially on the
competition between neutrino emission and absorption. In \reffig{fig: net rates},
we show axisymmetric averages of the net specific energy rate (left),
of the net electron fraction rate (centre), and of
the acceleration due to neutrino absorption (right), at $t=40 \, {\rm ms}$.\\
Inside the most relevant neutrino surfaces and a few kilometers outside them, neutrino cooling
dominates. Above this region, neutrino heating is always dominant.
The largest neutrino heating rate happens in the funnel,
where the neutrino densities are also larger. However, these regions are characterised
by matter with low density ($\rho \la 10^{7} {\rm g \, cm^{-3}}$) and small specific angular momentum.
Thus, this energy deposition has
a minor dynamical impact on this rapidly accreting matter.
On the other hand, at larger radii
($80 \, {\rm km} \la R_{\rm cyl} \la 120 \, {\rm km}$)
net neutrino heating affects denser matter ($\rho \la 10^{10}{\rm g \, cm^{-3}}$),
rotating inside the disc around the HMNS. This combination provides an efficient net energy deposition.\\
Neutrino diffusion from the optically thick region determines small variations around
the initial weak equilibrium value in the electron fraction.
On the contrary, in optically thin conditions, the initial very low electron fraction
favours reactions like the absorption of $e^+$ and
$\nu_e$ on free neutrons. Both processes
lead to a positive and large $\left({\rm d}Y_e/{\rm d}t \right)_{\nu}$, in association with efficient energy deposition.\\
Due to the geometry of the emission and to the shadow effect
provided by the disc,
the direction of the acceleration provided by neutrino absorption is approximately radial,
but its intensity is much larger in the funnel, where the energy deposition is also more intense.
As a consequence of the continuous neutrino energy and momentum deposition,
the outer layers of the disc start to expand a few milliseconds after the beginning of the simulation,
and they reach an almost stable configuration in a few tens of milliseconds.
Around $t \sim 10 \, {\rm ms}$, also the neutrino-driven wind starts to develop from the expanding disc.
Wind matter moves initially almost vertically (i.e., with
velocities parallel to the rotational axis of the disc),
decreasing its density and temperature during the expansion.
We show the corresponding vertical profiles inside the disc
in the bottom panels of \reffig{fig: disc profiles},
at different times and for three cylindrical radii.
Both the disc and the wind expansions are visible in the rise of the density and
temperature profiles, especially at cylindrical radii of $70 \, {\rm km}$ and $140 \, {\rm km}$.\\
Among the energy and the momentum contributions, the former is the most important one
for the formation of the wind.
To prove this, we repeat our simulation in two cases, starting from
the same initial configuration and relaxation procedure. In a first case, we set the heating
rate $h_{\nu}$ appearing in \refeq{eqn: particle heating rate} and \refeq{eqn: stress rate}
to 0. Under this assumption, we observe neither the disc expansion nor the wind formation.
In a second test, we include the effect of neutrino absorption only in the energy and $Y_e$
equations, but not in the momentum equation. In this case, the wind still develops and
its properties are qualitatively very similar to our reference simulation.\\
\begin{figure*}
\begin{center}
\includegraphics[width = \linewidth]{fig14.png}
\end{center}
\caption{Vertical slices of the three dimensional domain (corresponding to the $y=0$ plane), recorded $20 \, {\rm ms}$ after
the beginning of the simulation.
In the left panel, we represent the logarithm of the matter density (in ${\rm g/cm^3}$, left side)
and the projected fluid velocity (in units of $c$, on the right side);
the arrows indicate the direction of the projected velocity in the plane). On the right panel, we represent the
electron fraction (left side) and the matter entropy (in unit of $k_{\rm B}/{\rm baryon}$, right side).}
\label{fig: wind 20ms}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width = \linewidth]{fig15.png}
\end{center}
\caption{Same as in \reffig{fig: wind 20ms}, but at $\approx 40 \, {\rm ms}$ after the beginning of the simulation.}
\label{fig: wind 40ms}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width = \linewidth]{fig16.png}
\end{center}
\caption{Same as in \reffig{fig: wind 20ms}, but at $\approx 85 \, {\rm ms}$ after the beginning of the simulation.}
\label{fig: wind 85ms}
\end{figure*}
In Fig. \ref{fig: wind 20ms}, \ref{fig: wind 40ms} and \ref{fig: wind 85ms} we present
three different times of the wind expansion,
$t=20,40,85 \, {\rm ms}$.
To characterise them, we have chosen vertical slices of the three dimensional domain,
for the density and the projected velocity (left picture), and for the electron fraction and the matter entropy (right picture).\\
The development of the wind is clearly associated with the progressive increase of the electron fraction.
The resulting $Y_e$ distribution is not uniform,
due to the competition between the wind
expansion time-scale (\refeq{eqn: wind time-scale}) and the time-scale for weak equilibrium to establish.
The latter can be estimated as $t_{\rm weak} \sim Y_{e,{\rm eq}} / \left( {\rm d}{Y_e}/ {\rm d} t \right)_{\nu}$.
Using the values of the neutrino luminosities,
mean energies and net rates for the wind region,
we expect $Y_{e,{\rm eq}} \approx 0.42$ \citep[see, for example, eq. (77) of ][]{Qian1996}
and $0.042 \, {\rm s} \la t_{\rm weak} \la 0.090 \, {\rm s}$.
If we keep in mind that the absorption of neutrinos becomes less efficient
as the distance from the neutrino surfaces increases, we understand the presence of
both radial and vertical gradients for $Y_e$ inside the wind:
the early expanding matter has not enough time to reach $Y_{e,{\rm eq}}$, especially if it is initially
located at large distances from the relevant neutrino surfaces ($R_{\rm cyl} \ga 100 \, {\rm km}$).
On the other hand, matter expanding from the innermost part of the disc and moving in the funnel
(within a polar angle $\la 40^o$), as well as matter that orbits several times around the HMNS before being
accelerated in the wind, increases its $Y_e$ close to the equilibrium value, but on a longer time-scale.\\
Also the matter entropy in the wind rises due to neutrino absorption. Typical initial values in the disc
are $s \sim 5 - 10 \, k_{\rm B} \, {\rm baryon^{-1}}$, while later we observe
$s \sim 15-20 \, k_{\rm B} \, {\rm baryon^{-1}}$. The entropy is usually larger where the
absorption is more intense and $Y_e$ has increased more. However, differently from $Y_e$,
its spatial distribution is more uniform. Once the distance from the HMNS and the disc
has increased above $\sim 400 \, {\rm km}$, neutrino absorption becomes negligible and
the entropy and the electron fraction are simply advected inside the wind.\\
The radial velocity in the wind increases from a few times $10^{-2} \, c$, just above the disc, to
a typical asymptotic expansion velocity of $0.08-0.09 \, c$. This acceleration is caused
by the continuous pressure gradient provided by newly expanding layers of matter
\begin{figure*}
\begin{center}
\includegraphics[width = \linewidth]{fig17.png}
\end{center}
\caption{Occurrence diagrams for $(\rho,Y_e)$ (top panel), $(\rho,s)$ (middle panels)
and $(Y_e,s)$ (bottom panels), for the thermodynamical properties of matter in the whole system,
at $t \approx 0 \, {\rm ms}$ (left column), $t \approx 40 \, {\rm ms}$ (central column) and
$t \approx 85 \, {\rm ms}$ (right column) after the beginning of the simulation.
Colour coded is a measure of the amount of matter experiencing specific thermodynamical conditions
inside the whole system. Occurrence smaller than $10^{-7}$ ${\rm M}_{\sun}$ have been omitted from the plot.}
\label{fig: occurrence}
\end{figure*}
To characterise the matter properties, we plot in \reffig{fig: occurrence} \emph{occurrence diagrams} for couples of quantities, namely
$\rho - Y_e$ (top row), $\rho - s$ (central row)
and $ Y_e - s $ (bottom row), at three different times ($t = 0,40,85$ ms).
Colour coded is a measure of the amount of matter experiencing specific thermodynamical conditions
inside the whole system, at a certain time
\footnote{A formal definition of the plotted quantity can be found in Sec. 2 of \cite{Bacca2012}.
However, in this work we don't calculate the time average.}.\\
We notice that most of the matter is extremely dense
($\rho > 10^{11}\, {\rm g \, cm^{-3}}$), neutron rich ($Y_e < 0.1$)
and, despite the large temperatures ($T > 1 \, {\rm MeV}$), at relatively
low entropy ($s < 7 \, {k_B \, {\rm baryon}^{-1}}$). This matter correspond to the
HMNS and to the innermost part of the disc, where matter conditions
change only on the long neutrino diffusion timescale, \refeq{eqn: diffusion time HMNS}, or
on the disc lifetime, \refeq{eqn: viscosity time}.
In the low density part of the diagrams ($\rho < 10^{11}\, {\rm g \, cm^{-3}}$),
the expansion of the disc and the development of the wind can be traced.
\begin{figure*}
\begin{center}
\includegraphics[width = \linewidth]{fig18.png}
\end{center}
\caption{Nuclear composition provided by the EoS (assuming everywhere NSE) in the disc and in the wind, at $t \approx 40 \, {\rm ms}$.
On the top row, free proton (left), free neutron (centre) and
$\alpha$ particles (right) mass fractions. On the bottom row, heavy nuclei mass fraction (left),
and mass number (centre) and atomic number (right) of the representative heavy nucleus. The black line represents the
$T=0.5 \, {\rm MeV}$ surface.}
\label{fig: composition 40ms}
\end{figure*}
In \reffig{fig: composition 40ms} (a-d), we represent the mass fractions
of the nuclear species provided by the nuclear EoS
inside the disc and the wind, at 40 ms after the beginning of the simulation.
Close to the equatorial plane ($|z| < 100 \, {\rm km}$), the composition
is dominated by free neutrons.
In the wind, the increase of the electron fraction corresponds to the conversion of
neutrons into protons due to $\nu_e$ absorption.
In the early expansion phase, the relatively high temperature
($T \gg 0.6 \, {\rm MeV}$)
favours the presence of free protons.
When the decrease of temperature allows the formation of nuclei,
protons cluster into $\alpha$ particles and,
later, into neutron-rich nuclei.
Then, the composition in the
wind, at large distances from the disc, is distributed between free neutrons
($0.4 \la X_n \la 0.6$) and heavy nuclei
($0.6 \ga X_h \ga 0.4$, respectively).
The heavy nuclei component is described in the EoS
by a representative average nucleus, assuming Nuclear Statistical Equilibrium (NSE) everywhere.
In \reffig{fig: composition 40ms} (e-f), we have represented the values of its
mass and charge number.
The most representative nucleus in the wind corresponds often to ${}^{78}{\rm Ni}$.
The black line defines the surface across which
the freeze-out from NSE is expected to occur
($T = 0.5 \, {\rm MeV}$).
Outside it the actual composition will differ from the NSE prediction
(see \refsec{sec:discussion}).
\subsection{Ejecta}
\label{sec: ejecta}
Matter in the wind can gain enough energy from the neutrino absorption
and from the subsequent disc dynamics to become unbound.
The amount of ejected matter is calculated as
volume integral of the density and fulfils three criteria:
1) has positive radial velocity;
2) has positive specific total energy;
3) lies inside one of the two cones of opening angle $60^\circ$,
vertex in the centre of the HMNS and axes coincident with the disc rotation axes.
The latter geometrical constraint excludes possible contributions coming from equatorial ejecta,
which have not been followed properly during their expansion.
The profile of $Y_e$ at the end of the simulation (see, for example, \reffig{fig: wind 85ms}) suggests
to further distinguish between two zones inside each cone, one at high (H: $0^\circ \leq \theta < 40^\circ$, where $\theta$ is
the polar angle) and one at low (L: $40^\circ \leq \theta < 60^\circ$) latitudes.\\
The specific total internal energy is calculated as:
\begin{equation}
e_{\rm tot} = e_{\rm int} + e_{\rm grav} + e_{\rm kin}.
\label{eqn: total specific energy}
\end{equation}
$e_{\rm grav}$ is the Newtonian gravitational potential,
and $e_{\rm kin}$ is the specific kinetic energy.
The specific internal energy $e_{\rm int}$ takes into account the nuclear recombination energy and,
to compute it, we use the composition provided by the EoS. For the nuclear binding energy
of the representative heavy nucleus, we use the semi-empirical nuclear mass formula
\citep[see, for example, the fitting to experimental nuclei masses reported by][]{Rohlf1994}:
in the wind, for $\langle A \rangle \approx 78$
and $\langle Z \rangle \approx 28$, the nuclear binding
energy is $\sim 8.1 \, {\rm MeV \, baryon^{-1}}$.
\begin{figure*}
\begin{center}
\includegraphics[width = \linewidth]{fig19.pdf}
\end{center}
\caption{Distributions in the $\nu$-driven wind ejecta binned by different physical properties.
The different columns refer to
density ($\rho$, left),
electron fraction ($Y_e$, central-left),
entropy per baryon ($s$, central-right)
and radial velocity ($v_r$, right).
The top (bottom) panels refer to high (low) latitudes.}
\label{fig: histograms}
\end{figure*}
At the end of the simulation, $M_{\rm ej}(t = 91 \, {\rm ms}) \approx 2.12 \times 10^{-3}$ ${\rm M}_{\sun}$,
corresponding to $\sim 1.2$ per cent of the initial disc mass ($M_{\rm disc} \approx 0.17$ ${\rm M}_{\sun}$).
This mass is distributed between $M_{\rm ej,H}(t = 91 \, {\rm ms}) \approx 1.3 \times 10^{-3}$ ${\rm M}_{\sun}$
at high latitudes and $M_{\rm ej,L}(t = 91 \, {\rm ms}) \approx 0.8 \times 10^{-3}$ ${\rm M}_{\sun}$
at low latitudes.
In \reffig{fig: histograms}, we represent the mass distributions of density, electron fraction, entropy
and radial velocity, for the ejecta at the end of our simulation.
At high latitude, the larger $\nu_e$ absorption enhances the electron fraction and the entropy more
than at lower latitudes. The corresponding mass distributions are broader, with peaks at
$Y_e \sim 0.31-0.35$ and $s \sim 15-20 k_{\rm B}/{\rm baryon}$.
At lower latitudes, the electron fraction
presents a relatively uniform distribution between 0.23 and 0.31, while the entropy has a very narrow peak
around 14-15 $k_{\rm B}/{\rm baryon}$. The larger energy and momentum depositions produce a faster
expansion of the wind close to the poles. This effect is visible in the larger average value and in
the broader distribution of the radial velocity that
characterises the high latitude ejecta.\\
To quantify the uncertainties in the determination of the
ejecta mass, we repeat the previous calculation assuming an error of $ 0.5 \, {\rm MeV}$ in the estimate
of the nuclear recombination energy. For $M_{\rm ej,H}$ this translates in an uncertainty of
$\approx 7$ per cent, while in the case of $M_{\rm ej,L}$ the potential error is much larger ($\sim 50$ per cent).
This is a consequence of the different ejecta properties. At high latitudes, most of the free neutrons have
been incorporated into heavy nuclei, releasing the corresponding binding energy. Moreover, the large
radial velocities ($v_r \sim 0.08-0.09 \, c$) provides most of the energy needed to overcome
the gravitational potential.
At lower latitudes, the more abundant free neutrons and the lower radial velocities
($v_r \sim 0.06-0.07 \, c$) translate into a smaller ejecta amount, with a larger dependence on the
nuclear recombination energy. Generally, we consider our numbers for the wind ejecta as lower limits,
since a) we ignore the presence and likely amplification of magnetic fields which could substantially enhance the mass loss
\citep{Thompson2003a}, b) so far, we ignore heating from neutrino-annihilation and c) we do not consider colatitudes $> 60^\circ$.
\section{Discussion}
\label{sec:discussion}
\subsection{Comparison with Previous Works}
The hierarchies we have obtained for the neutrino luminosities and mean energies agree
with previous studies on the neutrino emission from neutron star mergers and their aftermaths.
In the case of Newtonian simulations, the compatibility is good
also from a quantitative point of
view, usually within 25 per cent
\citep[see, for example, the values obtained in ][ for the run H, to be compared with our
cooling luminosities]{Rosswog2013}.
On the other hand, general relativistic simulations
(usually, limited in time to the first tens of milliseconds after the merger)
obtain larger neutrino luminosities (up to a factor 2 or 3), due to larger
matter temperatures and stronger shocks
\citep[see, for example, ][]{Sekiguchi2011,Kiuchi2012,Neilsen2014}.
The higher temperatures reduce also the ratio between $\bar{\nu}_e$ and $\nu_e$ luminosities,
since the difference between charged current reactions on neutrons and protons diminishes ($k_{\rm B}T \gg Q$),
and thermal pair processes are enhanced.
\cite{Dessart2009} studied the formation of the neutrino-driven wind, starting from
initial conditions very similar to ours, in axisymmetric simulations that employ a multi-group
flux limited diffusion scheme for neutrinos.
Our results agree with theirs concerning typical values of the neutrino luminosities and mean energies (with
the exception of $\langle E_{\nu_{\mu,\tau}} \rangle$, see \refsec{sec: neutrino emission}),
as well as their angular distributions.
Also the shape and the extension of the neutrino surfaces inside the
disc are comparable.
There are, however, some differences in the temporal evolution:
while we observe almost stationary profiles, decreasing on a time-scale comparable with the expected disc
lifetime, their luminosities decrease faster. Also the difference between $\nu_e$ and $\bar{\nu}_e$ luminosities decreases,
leading to $L_{\nu_e} \approx L_{\bar{\nu}_e}$.
Both these differences can depend on the different accretion histories: the usage of the three dimensional
initial data (without performing axisymmetric averages) preserves all the initial local perturbations
and favours a substantial $\dot{M}$ inside the disc.\\
The removal and the deposition of energy, operated by neutrinos, is similar in the two cases. As a result,
the subsequent disc and wind dynamics agree well with each other.
The amount of ejecta and its electron fraction, on the other hand, show substantial differences:
at $t \sim 100 \, {\rm ms}$, we observe a larger amount of unbound matter, whose electron fraction has
significantly increased.
The evolution of a purely Keplerian disc around a HMNS, under the influence of $\alpha$-viscosity and
neutrino self-irradiation, as a function of the lifetime of the central object, has been more
recently investigated by \cite{Metzger2014}.
They employ an axisymmetric HD model, coupled with a grey leakage scheme and a light bulb boundary luminosity
for the HMNS. They evolve their system for several seconds
to study the development of the neutrino-driven wind and of the viscous ejecta.
In the case of a long-lived HMNS ($t_{ns} \ga 100 \, {\rm ms}$),
our results for the wind are qualitatively similar to their findings:
we both distinguish between a polar outflow,
characterised by larger electron fractions, entropies and expansion time-scales, and a
more neutron rich equatorial outflow. The polar ejecta, mainly driven by neutrino absorption, represent
a meaningful, but small fraction of the initial mass of the disc (a few percent).
Quantitative differences, connected with the different initial
conditions and the different neutrino treatment, are however present: their entropies and electron fractions
are usually larger, especially at polar latitudes.
The importance of neutrinos in neutron star mergers
has been recently addressed also by \cite{Wanajo2014}. They have shown that
the inclusion of both neutrino emission and
absorption can increase the ejecta $Y_e$ to a wide range of values (0.1-0.4), leading to the production
of all the r-process nuclides from the dynamical ejecta.
However, a direct comparison with our work is difficult since 1) their
simulation employs a softer EoS, that amplifies general relativistic effects, and 2) their analysis is
limited to the dynamical ejecta and the influence of neutrinos on it during the first milliseconds after the
merger.
\subsection{Nucleosynthesis in neutrino-driven winds}
\label{sub:nucleo}
\begin{table}
\begin{center}
\begin{tabular}{l|c|c|c|c|c|c|c}
\hline
Tracer & $Y_e$
& s $[k_{\rm B} \ {\rm baryon}]$
& $\langle A\rangle_{\rm final}$
& $\langle Z\rangle_{\rm final}$
& $X_{\rm La,Ac}$\\
\hline
L1 & 0.213 & 12.46 & 118.0 & 46.2 & $0.04$\\
L2 & 0.232 & 11.84 & 107.1 & 42.5 & $0.009$\\
L3 & 0.253 & 12.68 & 98.0 & 39.2 & $7\cdot10^{-5}$\\
L4 & 0.275 & 12.73 & 90.2 & 36.4 & $1\cdot10^{-7}$\\
L5 & 0.315 & 13.68 & 81.7 & 33.0 & $3\cdot10^{-12}$ \\
\hline
H1 & 0.273 & 13.57 & 93.0 & 37.4 & $8\cdot10^{-7}$\\
H2 & 0.308 & 14.69 & 83.3 & 33.7 & $6\cdot10^{-11}$\\
H3 & 0.338 & 15.36 & 79.4 & 32.1 & $< 10^{-12}$\\
H4 & 0.353 & 16.40 & 78.4 & 31.7 & $< 10^{-12}$\\
H5 & 0.373 & 18.35 & 76.8 & 31.0 & $< 10^{-12}$\\
\hline
\end{tabular}
\end{center}
\caption{Parameters of representative tracers and corresponding
nucleosynthesis: electron fraction $Y_e$, specific entropy per baryon $s$,
average atomic mass $\langle A\rangle_{\rm final}$ and electric charge
$\langle Z\rangle_{\rm final}$ of the resulting nuclei, and the total mass
fractions of Lanthanides and Actinides in the resulting nucleosynthetic mix.
The latter are important for estimating opacities at the location of the
tracers.
}
\label{tab:tracers}
\end{table}
\begin{figure}
\includegraphics[width=0.50 \textwidth]{fig20_top.pdf} \\
\includegraphics[width=0.50 \textwidth]{fig20_bottom.pdf}
\caption{Summed final mass fractions for representative tracers.
Top and bottom panels correspond to high-latitude (H1-H5) and low-latitude
(L1-L5) tracers, respectively (see~\reftab{tab:tracers} for parameters of
individual tracers). Solar r-process abundances (scaled) are also shown for
comparison.}
\label{fig:nucleosynthesis}
\end{figure}
During our simulation, we have computed
trajectories of representative tracer particles (Lagrangian particles, passively
advected in the fluid during the simulation).
The related full nucleosynthesis will be explored in more detail in future work.
To get a first idea about the possible nucleosynthetic signatures, we have selected ten tracers,
extrapolated and post-processed with a nuclear network.
These tracers are equally distributed between the high and the low latitude
region (5+5). Inside each region, we have picked the particles that represent
the most abundant conditions in terms of entropy and electron fraction in the
ejecta at $t \approx 90 \, {\rm ms}$.
\reftab{tab:tracers} lists parameters of the selected tracers.
For the nucleosynthesis calculations we employ the WinNet nuclear reaction
network~\citep{Winteler2012a,Winteler2012}, which represents an update of BasNet
network code~\citep{Thielemann2011}.
The ingredients for the network that we use are the same as described
in~\cite{Korobkin2012}. We have also included the feedback of nuclear heating
on the temperature, but we ignore its impact on the density, since previous
studies have demonstrated that for the purposes of nucleosynthesis this impact
can be neglected~\citep{Rosswog2014a}.
In this exploratory study, we also do not include neutrino irradiation.
Instead we use the final value of electron fraction from the tracer to initialise
the network. In this way, we effectively take into account the final neutrino
absorptions. Our preliminary experiments show that neutrino irradiation has an
effect equivalent to vary $Y_e$ by a few percent, which is a
correction that will be addressed in future work.
It is also worth mentioning that the situation is even less simple if one
takes into account neutrino flavour oscillations, which may alter the
composition of the irradiating fluxes significantly, depending on the
densities and distances involved~\citep{Malkus2014}.
\reffig{fig:nucleosynthesis} shows the resulting nucleosynthetic mass fractions,
summed up for different atomic masses, and \reftab{tab:tracers} lists the
averaged properties of the resulting nuclei.
As expected, lower electron fractions lead to an r-process
with heavier elements, and for the lowest values of $Y_e$ even the elements up to
the third r-process peak ($A\sim190$) can be synthesised.
However, due the high sensitivity to the electron fraction, wind
nucleosynthesis cannot be responsible for the observed astrophysical robust
pattern of abundances of the main r-process elements.
On the other hand, it could successfully contribute to the weak r-process in
the range of atomic masses from the first to second peak
($70 \la A \la 110$).
\reffig{fig:nucleosynthesis} also illustrates that heavier elements tend to
be synthesised at lower latitudes, closer to the equatorial plane. This has
important consequences for directional observability of associated electromagnetic
transients. Material, contaminated with Lanthanides or Actinides is expected
to have opacities that are orders of magnitudes larger than those of iron group
elements. Therefore, the corresponding electromagnetic signal is expected to
peak in the infrared.
\cite{Kasen2013} estimates that as little as $X_{\rm La,Ac} \ga 0.01$ per cent of
these ``opacity polluters'' could be enough to raise the opacities by a factor
of hundred.
\reftab{tab:tracers} lists also the computed mass fraction of the opacity
polluters, which turns out to be negligible for high-latitude tracers, while
being quite significant for low-latitude ones.
We therefore expect that the signal from the wind outflow will look much
redder, dimmer and peak later if the outflow is seen from equatorial rather
than polar direction.
Additionally, if seen from the low latitudes, the signal from the wind outflow
can be further obscured by the dynamical ejecta.
Thus, for the on-axis orientation the signal has better prospects of
detection, therefore making follow-up observations of short GRBs more
promising.
We will discuss these questions in detail in~\refsec{sub:macronova} below.
\subsection{Electromagnetic transients}
\label{sub:macronova}
\begin{figure*}
\begin{tabular}{cc}
\includegraphics[width=0.46\textwidth]{fig21_tl.pdf} &
\includegraphics[width=0.46\textwidth]{fig21_tr.pdf}
\\
\includegraphics[width=0.46\textwidth]{fig21_cl.pdf} &
\includegraphics[width=0.46\textwidth]{fig21_cr.pdf}
\\
\includegraphics[width=0.46\textwidth]{fig21_bl.pdf} &
\includegraphics[width=0.46\textwidth]{fig21_br.pdf}
\end{tabular}
\caption{Electromagnetic transients due to the radioactive material produced in the neutrino-driven wind.
The left column refers to material ejected at high latitudes (H1-H5),
the right column shows the results for the low latitudes (L1-L5).
Top row: predicted macronova lightcurves (bolometric luminosity),
calculated with uniform-composition spherically symmetric Kulkarni-type
models. Model parameters: ejected mass $
m_{\rm ej}=2\cdot10^{-3}~{\rm M_\odot}$,
expansion velocity $v_{\rm e}=0.08\;c$.
Opacity is taken to be $1~{\rm cm}^2/{\rm g}$
and $10~{\rm cm}^2/{\rm g}$ for high- and low-latitude tracers respectively.
Middle row: radioactive heating rate for the representative tracers, normalised
to $\dot{\epsilon}_0=10^{10}t_d^{-1.3}~{\rm erg}/({\rm g}\cdot{\rm s})$.
Bottom row: broadband AB magnitudes in five different bands, calculated for
the case when the wind outflow is viewed from the 'top' (left panel) and
'side' (right panel). For comparison, the J band signal from the dynamic ejecta is
superimposed.}
\label{fig:macronovae}
\end{figure*}
In \refsec{sec: ejecta}, we have estimated the amount of mass ejected
at the end of our simulation ($M_{\rm ej} (t \approx 90 \,{\rm ms}) \approx 2.12 \times 10^{-3}$ ${\rm M}_{\sun}$ ).
As discussed there, it needs to be considered as a lower
limit on the mass loss at that time. The neutrino emission, however, will continue beyond that time and keep driving the
wind outflow. We make here an effort to estimate the total mass loss caused by neutrino-driven winds during the disc lifetime.
During our simulation, the temporal evolution of the accretion rate on the HMNS
(\reffig{fig: accretion_eta_time}) is well described by
\begin{equation}
\dot{M}(t) \approx 0.76 \, \exp{\left[ - t/(0.124 \, {\rm s}) \right]} \, M_{\sun} \, {\rm s^{-1}}.
\end{equation}
We notice that, according to this expression, the total accreted mass is
smaller than the initial mass of the disc:
\begin{equation}
M_{\rm acc} \equiv \int_0^{\infty} \dot{M} \, {\rm d}t \approx 0.095 M_{\sun} < M_{\rm disc}(t=0) \approx 0.17 M_{\sun}.
\end{equation}
This discrepancy can be interpreted as the effect of the wind outflow and of
the disc evaporation. The beginning of the latter process has already been
observed in our model, but not followed properly due to computational
limitations.
At $t \approx 0.285 \, {\rm s}$ the HMNS has accreted 90 per cent of $M_{\rm acc}$. This
agrees well with the viscous lifetime of the disc (\refeq{eqn: viscosity
time}), so we consider $t \approx 0.3 \, {\rm s}$ as a good estimate for the disc lifetime.
Since the wind is powered by neutrino absorption, we assume that the mass of the ejecta
is proportional to the energy emitted in neutrinos during the disc life
time:
\begin{equation}
M_{\rm ej}(t = 0.300 \, {\rm s}) =
\frac{ \int_0^{0.300 \, {\rm s}} L_{\nu,{\rm cool}} \, {\rm d}t }{ \int_0^{0.090 \, {\rm s}} L_{\nu,{\rm cool}} \, {\rm d}t}
\, M_{\rm ej}(t = 0.090 \, {\rm s}).
\end{equation}
To model $L_{\nu}(t)$ for $t>90 \, {\rm ms}$, we consider two possible cases:
\begin{itemize}
\item [A)] the HMNS collapses after the disc has been completely accreted;
\item [B)] it collapses promptly at the end of our simulations.
\end{itemize}
For both cases, we extrapolate linearly the luminosities from \reffig{fig: lum with time}.
But in case B, we decrease the neutrino luminosity by 50 per cent, to account for the
lack of contribution from the HMNS and the innermost part of the disc after
the collapse (see \refsec{sec: neutrino emission}).
Our final mass extrapolations are listed in \reftab{table: ejecta masses}.
So in summary, we find $4.87 \times 10^{-3}$ ${\rm M}_{\sun}$ for case A and
$3.49 \times 10^{-3}$ ${\rm M}_{\sun}$ for case B. Given that we consider
these numbers as lower limits, this implies that the wind would provide a substantial
contribution to the total mass lost in a neutron star merger (and likely similar for a neutron star-black hole
merger; for an overview over the dynamic ejecta masses see \cite{Rosswog2013}).
\begin{table}
\begin{center}
\begin{tabular}{| c | c | c | c | c | c |}
\hline
Case & $t [{\rm s}]$ & $t_{\rm ns}[{\rm s}]$ & $M_{\rm{ej,H}} [M_{\sun}]$ & $M_{\rm{ej,L}} [M_{\sun}]$ & $M_{\rm{ej}} [M_{\sun}]$ \\ \hline \hline
A/B & $0.09$ & $\ge 0.09 \,{\rm s}$ & $1.29 \cdot 10^{-3}$ & $0.82\cdot 10^{-3}$ & $2.11\cdot 10^{-3}$ \\
& & & & \\ \hline
Case & $t [{\rm s}]$ & $t_{\rm ns}[{\rm s}]$& $M_{\rm{ej,H}} [M_{\sun}]$ & $M_{\rm{ej,L}} [M_{\sun}]$ & $M_{\rm{ej}} [M_{\sun}]$ \\ \hline \hline
A & $0.3$ & $> 0.3 \,{\rm s}$ & $2.98 \cdot 10^{-3}$ & $1.89\cdot 10^{-3}$ & $4.87\cdot 10^{-3}$ \\
B & $0.3$ & $\sim 0.09 \,{\rm s}$ & $2.13 \cdot 10^{-3}$ & $1.36\cdot 10^{-3}$ & $3.49\cdot 10^{-3}$ \\
\end{tabular}
\end{center}
\caption{Values of the calculated (top, for $t \approx 90 \, {\rm ms}$) and extrapolated
(bottom for $t \approx 300 \, {\rm ms}$) ejected masses, for the high (H) and low (L) latitude regions, and their sum.
$t_{\rm ns}$ refers to the time-scale for the HMNS to collapse to a black hole.
}
\label{table: ejecta masses}
\end{table}
With these mass estimates, we compute expected lightcurves for each tracer,
using the semi analytic spherically-symmetric models of macronovae by
\cite{Kulkarni2005}, the same as the ones described
in \cite{Grossman2014}.
\reffig{fig:macronovae} shows the resulting lightcurves (top row) for the wind
outflow mass from the case A.
Each lightcurve corresponds to a simplified case when the entire wind ejecta
evolves according to the thermodynamic conditions of one specific tracer.
In this work, we do not take into account spatial or temporal variation of the
electron fraction within the wind outflow, but we assume different opacities
for the high-latitude and low-latitude tracers. Motivated by recent work of
\cite{Kasen2013} and confirmed by \cite{Tanaka2013a}, we
take a uniform grey opacity of $\kappa=10 \, {\rm cm}^2 \, {\rm g}^{-1}$
for the low-latitude tracers that have a low $Y_e$ and produce non-negligible
amounts of Lanthanides and Actinides. For the high-latitude, higher $Y_e$ tracers we use
$1\,{\rm cm}^2 \, {\rm g}^{-1}$.
The tracers result in a wide variety of potential lightcurves, whose shape
reflects individual nuclear heating conditions for a specific tracer.
The middle row of \reffig{fig:macronovae} shows the individual heating rates,
normalised to the power law
$\dot{\epsilon}_0=10^{10}t_d^{-1.3}~{\rm erg} \, {\rm g}^{-1} \, {\rm s}^{-1})$.
Differences in the shapes of the heating rates for different tracers
are due to the dominance of different radioactive elements
at late times \citep{Grossman2014}.
Despite the variety of macronovae for different tracers, the actual lightcurve
will lie somewhere in between, and the individual differences in the heating
rates will be smoothed out.
The bottom row represents averaged broadband lightcurves from the
high-latitude (left panel) and low-latitude (right panel) wind ejecta.
The high-latitude case shows a pronounced peak in the B band at
$t\sim1.3\,{\rm d}$, while the higher opacities
for the low-latitude tracers make the lightcurve
dimmer, redder and cause them to peak later.
An interesting question is whether or not the collapse time of the HMNS could
possibly be inferred from the EM signal, assuming that the collapse happens after
the wind has formed ($t_{\rm ns} \ga 100 \, {\rm ms}$). Therefore we compare in
\reffig{fig:casesAB} (left panel) the averaged bolometric
lightcurves for the cases A and B of long- and short-lived HMNS.
The plot shows the low-latitude and high-latitude components
separately, as well as the lightcurve for the dynamic ejecta for the same merger case.
The two cases differ very little, mainly because the mass of the wind
component changes only by a factor of $\sim1.5$, and the lightcurve is not
very sensitive to this mass.
The long-lived HMNS (case A) is slightly brighter, but it is not likely that
the two cases can be discriminated observationally.
The difference between high- and low-latitude regions shows that perhaps
geometry of the outflow and its orientation relative to the observer plays
much more important role in the brightness and colour of the expected
electromagnetic signal.
Similarly, there is practically no difference in the total summed
nucleosynthetic yields for cases A and B (\reffig{fig:casesAB}, right panel).
Thus it may be difficult to extract the HMNS collapse time-scale from
the macronova signal.
\begin{figure*}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=0.46\textwidth]{fig22_left.pdf}
&
\includegraphics[width=0.46\textwidth]{fig22_right.pdf}
\end{tabular}
\end{center}
\caption{Bolometric lightcurves (left) and summed abundances (right) for the
two cases of a long-lived (case A) and a short-lived (case B) HMNS.
The left panel shows separately bolometric lightcurves of low-latitude and
high-latitude outflows, as well as the lightcurve for the dynamical ejecta
from the same merger simulation. The plot also shows the effective
temperatures of the macronovae at the key points on the curves.
On the right panel, the wind abundances have been added to the abundances from
dynamical ejecta, for which we took the total ejected mass of
$1.3\cdot10^{-3}$ ${\rm M}_{\sun}$ from the merger simulation.
}
\label{fig:casesAB}
\end{figure*}
\section{Conclusions}
\label{sec:conclusions}
We have explored the properties of the neutrino-driven wind
that forms in the aftermath of a BNS merger.
In particular, we have discussed their implications in terms of the r-process nucleosynthesis
and of the electromagnetic counterparts
powered by the decay of radioactive elements in the expanding ejecta.
To model the wind, we have performed for the first time 3D Newtonian hydrodynamics
simulations, covering an interval of $\approx 100 \, {\rm ms}$ after the merger,
and a radial distance of $\ga 1500 \, {\rm km}$ from the HMNS, with high spatial resolution
inside the wind. Neutrino radiation has been treated by a computationally efficient, multi-flavour
Advanced Spectral Leakage scheme, which includes consistent neutrino absorption rates in
optically thin conditions.
Our initial configuration is obtained from the direct re-mapping of the matter distribution
of a 3D SPH simulation of the merger of two non-spinning 1.4 ${\rm M}_{\sun}$ neutron stars
\citep[][and references therein]{Rosswog2007c}, at $\approx 15$~ms after the first contact.
The consistent dimensionality and the high compatibility between the two models
do not require any global average nor any ad hoc assumption for the matter profiles inside the disc.\\
Our major findings are:
\begin{enumerate}
\item the wind provides a substantial contribution to the total
mass lost in a BNS merger.
At the end of our simulation ($\approx 100$ ms after the merger), we compute $2.12 \times 10^{-3}$ ${\rm M}_{\sun}$~
of neutron-rich ($0.2 \la Y_e \la 0.4$) ejected matter, corresponding to 1.2 per cent of the
initial mass of the disc. We distinguish between a high-latitude ($50^{\circ} - 90^{\circ}$)
and a low-latitude ($30^{\circ} - 50^{\circ}$) component
of the ejecta. The former is subject to a more intense neutrino irradiation and
is characterised by larger $Y_e$, entropies and expansion velocities.
We estimate that, on the longer disc lifetime, the ejected mass can increase to
$3.49-4.87 \times 10^{-3}$ ${\rm M}_{\sun}$, where the smaller (larger) value refers to
a quick (late) HMNS collapse after the end of our simulation.
\item The tendency of $Y_e$ to increase with time above 0.3, especially at high latitudes,
suggests a relevant contribution to the nucleosynthesis of the weak r-process elements from the wind,
in the range of atomic masses from the
first to the second peak. Matter ejected closer to the disc plane retains a lower
electron fraction (between 0.2 and 0.3), and produces nuclei from the first to the third peak,
without providing a robust r-process pattern.
\item The geometry of the outflow and its orientation relative to the observer have an important
role for the properties of the electromagnetic transient.
According to our results, the high-latitude outflow can power a
bluer and brighter lightcurve, that peaks within one day after the merger.
Due to the partial contamination of Lanthanides and
Actinides, the low-latitude ejecta is expected to have higher opacity and to peak later,
with a dimmer and redder lightcurve.
\item A significant fraction of the neutrino luminosity is provided by the accretion
process inside the disc. This fraction is expected to power a (less intense) baryonic wind also if
the HMNS collapses to a BH before the disc consumption. According to our calculations, the
collapse time-scale has a minor impact on the possible observables (electromagnetic counterparts and
nucleosynthesis yields), at least if the collapse happens after the wind has formed
and weak equilibrium had time to establish inside it. \cite{Metzger2014} indicate
that more meaningful differences can be potentially seen, in the case of an earlier collapse.
This scenario requires further investigations.
\end{enumerate}
Our 3D results show a good qualitative agreement with the 2D results obtained by \cite{Dessart2009}
for a similar initial configuration, especially for the neutrino emission and the wind dynamics.
Meaningful quantitative differences are still present, probably related to the different accretion
and luminosity histories inside the disc. The distinction between a high-latitude and a low-latitude region in the ejecta
is qualitatively consistent with recent 2D findings of \cite{Metzger2014}.\\
The results we have found for the amount of wind ejecta has to be considered as lower limits,
since in our model we ignore the effects of magnetic fields and neutrino-annihilation in optically thin conditions.
In particular, the latter is expected to deposit energy very efficiently in the funnel above the HMNS poles.
The calculation of this energy deposition rate for our model and its implication
for the sGRB mechanism will be discussed in a future work.\\
The wind ejecta has to be complemented with the dynamical ejecta
and with the outflow coming from the viscous evolution of the disc.
These other channels are expected to provide low-latitude outflows, with an electron fraction similar
or lower than the one obtained by the low-latitude wind component
\citep[see, for example, ][ and references therein]{Rosswog2013,Metzger2014}.
Instead, the high-latitude wind component seems to be peculiar in terms of outflow geometry,
nucleosynthesis yields and related radioactively powered electromagnetic emission.\\
This work represents one of the first steps towards a physically consistent and complete model of the aftermath
of BNS mergers, including the effect of neutrino irradiation.
Our preliminary calculations regarding the nucleosynthesis and the electromagnetic
counterparts motivate further analysis and investigations.
Moreover, additional work has to be done to develop more accurate and complete radiation hydrodynamics treatments,
to include other relevant physical ingredients (like magnetic fields and General Relativity),
and to explore the present uncertainties in terms of nuclear matter properties and neutrino physics.
\section*{Acknowledgements}
The authors thank F.K. Thielemann for useful discussions and for reading the manuscript.
AP and AA were supported by the Helmholtz-University Investigator grant No. VH-NG-825.
The work of SR has been supported by the Swedish Research Council (VR) under grant 621-2012-4870.
RC and ML acknowledge the support from the HP2C Supernova project and the ERC grant FISH.
AP, SR and ML thank the MICRA-2009 Workshop and the Niels Bohr Institute for their hospitality during the summer of
2009, when this project started.
AP thanks the Jacobs University Bremen for its hospitality in February 2010, October 2010 and April 2012,
and the Stockholm University for its hospitality in June 2013.
SR thanks the University of Basel for its hospitality in June 2010.
AP and SR thank COMPSTAR for the Short Visit Grants 3369 and 3536.
AP,RC RK and ML thank the use of computational resources provided by
the Swiss SuperComputing Center (CSCS), under the allocation grants s414.
The SPH simulations for the project have been performed on the facilities
of the H\"{o}chstleistungsrechenzentrum Nord (HLRN).
\bibliographystyle{mn2e}
\input{wind_bib.tex}
\bsp
\label{lastpage}
\end{document}
|
\section{Introduction}
\label{sec:introduction}
Consider two decision-makers (DMs) and that each of them has to select actions or take decisions repeatedly to reach a certain objective say to maximize an average payoff function. Furthermore, assume that there might be an interest for them in exchanging information e.g., about a random system state which can affect their payoff but that no dedicated communication channel is available for this purpose. Therefore, the only way to communicate for a DM is to use his own actions. Although the idea of communicating through actions seems to be quite natural and is in fact used more or less implicitly in real life scenarios e.g., in economics (see \cite{sims-jme-2003}), it appears that, apart from a few exceptions focused on specific problems of control (see e.g., \cite{grover-phd-2011}), it has obviously not penetrated yet engineering problems and definitely not wireless communications. It turns out that important wireless problems such as power control (PC) or radio resource allocation can draw much benefits from being revisited from the new perspective of communication through actions. Because of its importance and ability to easily illustrate the proposed approach, the problem of PC in interference networks has been selected for the application of the main and general result derived in this paper; note that the latter concerns any decision-making problem which has the same structure (see Sec. \ref{sec:problem}) and generalizes \cite{Gossner-2006} and \cite{cuff-itw-2011}.
In the context of PC, the DMs are transmitters (Txs) and the system state is typically given by the state of the communication channel between the Txs and receivers (Rxs); we will use the term DM (resp. Tx) when the general case (resp. the specific case of PC) is concerned. Quite often, each Tx possesses a partial knowledge of the channel state and, in general, there is an incentive for the Txs to exchange the corresponding knowledge between them. Coded PC assumes that this knowledge is transferred from one Tx to another (or others) by encoding the information of the former into a sequence of power levels which are observed by the latter. In this paper, we assume two DMs, that they have a common payoff, and that there is one DM (DM 1) which is informed with the current realization of the state and possibly those associated with the coming stages. DM 2 is only informed of the state in a strictly causal manner and can observe his on power levels. The considered scenario is, in particular, relevant in cognitive radio (CR) settings. In typical CR scenarios, the primary Tx is assumed to be passive and the secondary Tx adapts to what it observes. But, it might be of interest to design primary Txs which coordinate in an active manner the usage of radio resources, which is exactly what coded power control (CPC) allows; one of the salient features of CPC is that interference can be managed directly in the radio-frequency domain and does not require baseband detection or decoding, which is very useful in heterogeneous networks. Another body of works which can be mentioned is given by works on distributed PC and especially those on best response dynamics (BRD) algorithms which include the original iterative water-filling algorithm \cite{yu-jsac-2002}. Existing BRD algorithms implementations for PC (see e.g., \cite{lasaulce-book-2011}\cite{zappone-comlett-2011}\cite{bacci-tsp-2013}) typically assume SINR (signal-to-noise plus interference ratio) feedback and individual channel state information (CSI) and do not exploit the key idea of communicating through the power levels. Encoding power levels allows one to construct PC policies possessing at least three salient features which are generally not available for BRD-based PC: there is no convergence problem and this whatever the payoff functions; efficient solutions (e.g., in terms of sum-payoff) can be obtained; both the cases of discrete and continuous power levels can be easily treated. Since we focus on optimal PC policies and make the choice of an asymmetric information structure whereas BRD algorithms rely on a symmetric one, no explicit comparison with BRD algorithms is conducted but CPC can be applied to symmetric scenarios as well.
\section{Problem statement}
\label{sec:problem}
Consider two DMs which want to coordinate through their actions. Let $\mathcal{X}_j$, $|\mathcal{X}_j| < \infty$, the action alphabet of DM $j \in \{1,2\}$, and $\mathcal{X}_0$, $|\mathcal{X}_0|<\infty$, the random state alphabet. The states are assumed to be i.i.d. and generated from a random variable $X_0$ whose realizations are in $\mathcal{X}_0$ and distribution is denoted by $\rho$. Note that the finiteness assumption is not only realistic (e.g., power levels are discrete in modern cellular systems) but also allows the continuous case to be treated by using classical arguments \cite{Cover:2006:EIT:1146355}. The strategies of DM $1$ and $2$ are sequences of mappings, $(\sigma_i, \tau_i)_{i\geq1}$, which are respectively defined by:
\begin{equation}\label{eq:strategies}
\left\{
\begin{array}{ccccc}
\sigma_i & : & \mathcal{X}_0^T \times \mathcal{X}_1^{i-1} & \rightarrow & \mathcal{X}_1\\
\tau_i & : & \mathcal{X}_0^{i-1} \times \mathcal{Y}^{i-1} \times \mathcal{X}_2^{i-1} & \rightarrow & \mathcal{X}_2
\end{array}
\right.
\end{equation}
where $T$ is the total number of stages, $i\in \{1,...,T\}$, and $\mathcal{Y}$, $|\mathcal{Y}| < \infty$, is the observation alphabet of DM $2$. The definition of the strategy for DM 1 indicates that we assume a non-causal knowledge of the state. The most typical situation in PC is to assume that two phases are available (training phase, action phase) and one state is known in advance to adjust the power level. This special case can be obtained by setting $T=2$ that is, $i\in\{1,2\}$. There are many reasons why we consider here that $T$ might be greater than two. We will only provide three of them, which better explains how the non-causality assumption should be understood. First, the result derived in Sec. \ref{sec:main-analytical-result} can be used for a large variety of settings and not only PC. Second, the proposed approach can be applied to the case where the state is not i.i.d. (e.g., to the $B-$stage block i.i.d. case, $B\geq1$). Indeed, there exist wireless communication standards which assume the channel to be constant over several time-slots and the proposed approach suggests that gains can be obtained by varying the power level from time-slot to time-slot even if the channel is constant. Third, it becomes more and more common to exploit the forecasted trajectory of a mobile user to optimize the system \cite{fourestie-patent-2007}, which makes our approach relevant when the channel state is interpreted as the path loss. Concerning the chosen definition for the strategy of DM 2, several comments are in order. First, note that DM 2 is not assumed to monitor actions of DM 1 perfectly. Rather, they are monitored through an observation channel which is assumed to be discrete, memoryless, and to verify $P(y|x_0,x_1,x_2) = \Gamma(y|x_1)$, where $y \in \mathcal{Y}$ is a realization of the channel output associated with the input $(x_0,x_1,x_2)$. Second, note that, the strategy of DM 2 is defined such that it can choose an action at every stage and not only at the end of a block or sequence of stages as it would be for a classical block decoder. Therefore, contrarily to \cite{Gossner-2006}, DM 1 does not need to observe the actions of DM 2 and DM 2 has only access to imperfect observations of the actions chosen by DM 1. Interestingly, we will see that the fact that DM 1 does not observe DM 2 induces no performance loss in terms of payoff.
The instantaneous or stage payoff function for the DMs is denoted by $w(x_0,x_1,x_2)$. Since the state is random, we will consider as general case the problem of reaching a certain performance level in terms of expected payoff $\mathbb{E}[w] = \sum_{(x_0,x_1,x_2)} P(x_0,x_1,x_2) w(x_0,x_1,x_2)$. Roughly, the task of DM 1 is to maximize the expected payoff by finding the best tradeoff between reaching a good payoff for the current stage and revealing enough information about the future realizations of the state to coordinate for the next stages. The ability for two DMs to coordinate their actions i.e., to reach a certain value for the expected payoff can be translated in terms of joint distribution over $\mathcal{X}_0\times\mathcal{X}_1\times\mathcal{X}_2$, which leads us to the notion of implementable distribution \cite{Gossner-2006}.
\begin{definition}[Implementability] The distribution $\overline{Q}(x_0,x_1,x_2)$ is implementable if there exists a pair of strategies $(\sigma_i, \tau_i)_{i\geq1}$ such that as $t \rightarrow +\infty$ we have for all $(x_0,x_1,x_2)$,
\begin{equation}
\frac{1}{t} \sum_{i=1}^{t} \sum_{y} P_{X_{0,i}, X_{1,i}, X_{2,i}, Y_i}(x_0,x_1,x_2,y) \rightarrow \overline{Q}(x_0,x_1,x_2)
\end{equation}
where $P_{X_{0,i},X_{1,i},X_{2,i},Y_i}$ is the joint distribution induced by $(\sigma_i, \tau_i)_{i\geq1}$ at stage $i$.
\end{definition}
Importantly, note that, since the expectation operator is linear, a certain value for $\mathbb{E}[w]$ is reachable if and only if there exists an implementable distribution. The goal of the next section is precisely to characterize the set of reachable expected payoffs $\mathbb{E}_Q[w] = \sum_{(x_0,x_2,x_2,y)} \overline{Q}(x_0,x_1,x_2) \Gamma(y|x_1) w(x_0,x_1,x_2)$, which thus amounts to characterizing the set of implementable distributions over $\mathcal{X}_0 \times \mathcal{X}_1 \times \mathcal{X}_2$.
\section{Main analytical result}
\label{sec:main-analytical-result}
Notation: $\Delta(\mathcal{A})$ will stand for the set of
distributions over the generic discrete set $\mathcal{A}$. Using this notation, the main analytical result of this paper can be stated.
\begin{thm}\label{theo:info-constraint} Let $\overline{Q} \in \Delta(\mathcal{X}_0 \times \mathcal{X}_1 \times \mathcal{X}_2)$ with $\sum_{(x_1,x_2)}\overline{Q}(x_0,x_1,x_2) = \rho(x_0)$. The distribution $\overline{Q}$ is implementable if and only if there exists $Q \in \Delta(\mathcal{X}_0 \times \mathcal{X}_1 \times \mathcal{X}_2 \times \mathcal{Y})$ which verifies the following information constraint:
\begin{equation}
\label{eq:information-constraint}
I_{Q} (X_0 ; X_2) \leq I_{Q} (X_1;Y |X_0,X_2)
\end{equation}
where the arguments of the mutual information $I_{Q}(.)$ are defined from $Q$ and $Q(x_0,x_1,x_2,y) = \overline{Q}(x_0,x_1,x_2) \Gamma(y|x_1)$.
\end{thm}
\subsection{Proof of Theorem 1}
\label{sec:proof-main-result}
\begin{IEEEproof}[Converse proof] We first start with providing a lemma which is used at the end of the proof and concludes the section.
\begin{lem}\label{lemma:convexity-of-phi} The function $\Phi : Q \mapsto I_{Q} (X_0 ; X_2) - I_{Q} (X_1;Y |X_0,X_2)$ is convex over the set of distributions $Q \in \Delta(\mathcal{X}_0 \times \mathcal{X}_1 \times \mathcal{X}_2 \times \mathcal{Y})$ that verify $\sum_{(x_1,x_2,y) Q(x_0,x_1,x_2,y)} = \rho(x_0)$ and $Q(x_0,x_1,x_2,y) = \Gamma(y|x_1) P(x_0,x_1,x_2)$, with $\rho$ and $\Gamma$ fixed.
\end{lem}
\setlength{\belowdisplayskip}{0pt} \setlength{\belowdisplayshortskip}{0pt}
\setlength{\abovedisplayskip}{0pt} \setlength{\abovedisplayshortskip}{0pt}
\begin{IEEEproof}[Proof of Lemma \ref{lemma:convexity-of-phi}] The function $\Phi$ can be rewritten as $\Phi(Q) = H_Q(X_0) - H_Q(Y,X_0 | X_2) + H_Q(Y|X_0,X_2,X_1)$. The first term $H_Q(X_0) = -\sum_{x_0} \rho(x_0) \log \rho(x_0)$ is a constant w.r.t. $Q$. The third term is linear w.r.t. $Q$ since, with $\Gamma$ fixed,
\begin{multline}
H_Q(Y|X_0,X_2,X_1) = \\
- \sum_{x_0,x_1,x_2,y} Q(x_0,x_1,x_2,y) \log P(y|x_0,x_1,x_2) \\
= - \sum_{x_0,x_1,x_2,y} Q(x_0,x_1,x_2,y) \log \Gamma(y|x_1)
\end{multline}
It is therefore sufficient to prove that $H_Q(Y,X_0 | X_2)$ is concave. Let $\lambda_1 \in [0,1]$, $\lambda_2 = 1 - \lambda_1$, $(Q_1,Q_2) \in \Delta^2(\mathcal{X}_0 \times \mathcal{X}_1 \times \mathcal{X}_2 \times \mathcal{Y})$ and $Q=\lambda_1 Q_1 + \lambda_2 Q_2$. By using the standard notation $A^0 = \emptyset$, $A^n = (A_1, ..., A_n)$, we have that:
\begin{align}
& H_Q(Y,X_0 | X_2) = - \sum_{x_0,x_2,y} \bigg( \sum_{x_1,i} \lambda_i Q_i(x_0,x_1,x_2,y) \bigg) \nonumber \\
& \log \left[ \frac{\sum_{x_1,i} \lambda_i Q_i(x_0,x_1,x_2,y)}{ \sum_{i} \lambda_i P_{X_2}^{Q_i}(x_2)} \right] \\
& = - \sum_{x_0,x_2,y} \bigg(\sum_i \lambda_i \sum_{x_1} Q_i(x_0,x_1,x_2,y) \bigg) \nonumber \\
& \log \left[ \frac{\sum_i \lambda_i \sum_{x_1} Q_i(x_0,x_1,x_2,y)}{ \sum_{i} \lambda_i P_{X_2}^{Q_i}(x_2)} \right] \\
& \geq - \sum_i \lambda_i \sum_{x_0,x_2,y} \bigg( \sum_{x_1} Q_i(x_0,x_1,x_2,y) \bigg) \nonumber \\
& \log \left[ \frac{\lambda_i \sum_{x_1} Q_i(x_0,x_1,x_2,y)}{\lambda_i P_{X_2}^{Q_i}(x_2)} \right] \\
& = - \sum_i \lambda_i \sum_{x_0,x_2,y} \bigg( \sum_{x_1} Q_i(x_0,x_1,x_2,y) \bigg) \nonumber \\
& \log \left[ \frac{\sum_{x_1} Q_i(x_0,x_1,x_2,y)}{ P_{X_2}^{Q_i}(x_2)} \right] \\
& = \lambda_1 H_{Q_1}(Y,X_0 | X_2) + \lambda_2 H_{Q_2}(Y,X_0 | X_2)
\end{align}
where the inequality comes from the log sum inequality \cite{Cover:2006:EIT:1146355}.
\end{IEEEproof}
Now we want to prove that if $\overline{Q}$ is implementable, then $Q$ has to verify the information constraint. Assuming $\overline{Q}$ is implementable means that there exists $(\sigma_i, \tau_i)_{i\geq1}$ such that the empirical distribution $P^{(t)}_{X_0,X_1,X_2,Y}(.) = \frac{1}{t} \sum_{i=1}^{t} P_{X_{0,i}, X_{1,i}, X_{2,i}, Y_i}(.)$ can be made arbitrarily close to $Q$; this argument is used at the end of the proof. We have:
\begin{align}
& \sum_{i=1}^{t} I_{P_{X_{0,i},X_{1,i},X_{2,i},Y_i}} ( X_0 ; X_{2}) = \sum_{i=1}^{t} I ( X_{0,i} ; X_{2,i}) \\
\stackrel{(a)}{=} & H(X_0^t) - \sum_{i=1}^{t} H(X_{0,i}| X_{2,i}) \\
= & H(X_0^t,Y^t,X_2^t) - H(Y^t,X_2^t | X_0^t) - \sum_{i=1}^{t} H(X_{0,i}| X_{2,i}) \\
\leq & H(X_0^t,Y^t,X_2^t) - H(Y^t | X_0^t) - \sum_{i=1}^{t} H(X_{0,i}| X_{2,i})
\end{align}
\begin{align}
\leq & H(X_0^t,Y^t,X_2^t) - H(Y^t | X_0^t,X_{1}^t,X_{2}^t) - \sum_{i=1}^{t} H(X_{0,i}| X_{2,i}) \\
\stackrel{(b)}{=} & H(X_0^t,Y^t,X_2^t) - \sum_{i=1}^{t} H(X_{0,i}| X_{2,i}) \nonumber \\
& - \sum_{i=1}^{t} H(Y_{i} | X_{0,i},X_{1,i},X_{2,i}) \\
\stackrel{(c)}{\leq} & \sum_{i=1}^{t} H(X_{0,i},Y_{i},X_{2,i} | X_{2,i}) - H(X_{0,i}| X_{2,i}) \nonumber \\
&- H(Y_{i} | X_{1,i},X_{0,i},X_{2,i}) \\
= & \sum_{i=1}^{t} H(X_{0,i},Y_{i} | X_{2,i}) - H(X_{0,i}| X_{2,i}) \nonumber \\
&- H(Y_{i} | X_{1,i},X_{0,i},X_{2,i}) \\
= & \sum_{i=1}^{t} I(X_{1,i};Y_{i} | X_{0,i},X_{2,i}) \\
= & \sum_{i=1}^{t} I_{P_{X_{0,i},X_{1,i},X_{2,i},Y_i}} (X_{1} ; Y | X_0,X_2)
\end{align}
where: (a) comes from the fact that $(X_{0,i})_i$ is i.i.d. and the chain rule for entropy; (b) holds because the observation channel from DM 1 to DM 2 is assumed to be discrete and memoryless namely, $P(y^t|x_0^t,x_1^t,x_2^t)=\prod_{i=1}^{t} p(y_{i}|x_{0,i},x_{1,i},x_{2,i})$; (c) holds by the chain rule and because $X_{2,i}$ is a deterministic function of the past: $X_{2,i}= \tau_i \left(X_{0,1},Y_{1},X_{2,1},\dots,X_{0,i-1},Y_{i-1},X_{2,i-1}\right)$. Now, since $\Phi$ is convex (by Lemma \ref{lemma:convexity-of-phi}), we know that
\begin{multline} I_{P^{(t)}_{X_0,X_1,X_2,Y}} (X_{1} ; Y | X_0, X_2)
- I_{P^{(t)}_{X_0,X_1,X_2,Y}} ( X_0 ; X_{2}) \geq \\
\frac{1}{t} \sum_{i=1}^{t} I_{P_{X_{0,i},X_{1,i},X_{2,i},Y_i}} (X_{1} ; Y | X_0, X_2) \\
- I_{P_{X_{0,i},X_{1,i},X_{2,i},Y_i}} ( X_0 ; X_{2})
\end{multline}
The converse follows by observing that the first term of the above inequality can be made arbitrarily close to $I_{Q} (X_1;Y |X_0,X_2) - I_{Q} (X_0 ; X_2)$ and the second term has been proven to be non-negative. \hfill
\end{IEEEproof}
\begin{IEEEproof}[Implementability (sketch)]
The goal here is to prove that if the information constraint is verified for $Q^*$, then an implementable pair of strategies $(\sigma_i, \tau_i)_{i\geq1}$ can be found. Therefore, in contrast with the converse, finding a particular code such as a block code with long codewords is sufficient, which allows one to reuse the standard machinery for the transmission of distorted sources. Assume $T = n B$ large where $n$ is the codeword length and $B$ the number of blocks. Denote $b$ as the block index. The methodology is the following: construct a source codebook and a channel codebook by choosing each symbols of each sequences in the codebooks independently using the same distribution $Q^*$ (more precisely marginal distributions of $Q^*$). DM 1 then uses joint typicality to find in the source codebook the sequence of actions of DM 2 for block $b+1$, and sends the corresponding channel codeword (the one with the same index). DM 2 receives a sequence $y^n[b]$ through the observation channel and decodes the index chosen by DM 1 via joint typicality of the four sequences on block $b$. He uses this index to find in the source codebook his sequence of actions for block $b+1$. At last, for block $b=0$, DM 2 chooses an arbitrary codeword which is known to DM 1. There is an error if we don't have existence and/or unicity of these codewords. The probability of error is made arbitrarily small thanks to the information constraint \eqref{eq:information-constraint} (the analysis, although not trivial, is standard and need the well known Markov Lemma and Packing Lemma). Under this setting, as $Q^*$ meets the information constraint, the empirical distribution $Q^{(T)}$ which is induced by this separate source channel coding i.e.,
\begin{multline} Q^{(T)}(v)=\frac{1}{n B} \big[ \mathcal{N}\big(v \; | \; x_0^n(0),x_1^{n}(0),x_2^n(0),y^{n}(0) \big) \\
+ \sum_{b=1}^{B-1} \mathcal{N}\big(v \; | \; x_0^n(b),x_1^{n}(b),x_2^n(b),y^{n}(b) \big) \big]
\end{multline}
converges to $Q^*$, where $\mathcal{N}(v|v^n)$ is a notation for counting the occurrences of $v$ in $v^n$, $v = (x_0,x_1,x_2,y)$ here. The proof of this involves definitions of typical sets and the triangle inequality.
\end{IEEEproof}
\subsection{Comments on Theorem 1}
\label{sec:comments-main-result}
Theorem 1 can be interpreted as follows. DM 2's actions (represented by $X_2$) correspond to a joint source-channel decoding operation with distortion on the information source (which is represented by $X_0$). To be reachable, the distortion rate has to be less than the transmission rate allowed by the channel whose input and output are respectively represented by $X_1$ and $Y$. Therefore, the pair $S=(X_0,X_2)$ seems to play the same role as the side information in channels with state. Indeed, the implementability proof shows that DM 1 uses in particular $(x_0^n(b), \widehat{x}_2^n(b))$ while DM 2 uses $(x_0^n(b), x_2^n(b))$. Asymptotically, the encoder (DM 1) and decoder (DM 2) have the same side information (which explains by the way the fact DM 1 does not need to observe DM 2 does not induce any performance loss). Furthermore, note that $x_0^n(b+1)$, which plays the role of the message to be encoded, is independent of the side information. Classical coding schemes (such as block Markov coding) can thus be re-exploited. However, the above arguments fails for the converse proof which has to deal with arbitrary coding schemes or strategies. It can no longer be assumed that the side information be independent of the information source vector. This is one of the reasons why the converse proof has to be rethought. Another reason is that classical results (such as Fano's inequality) which rely on block decoding are not exploitable anymore since DM 2 has to be able to act (to decode) at any stage or time instance.
As another type of comments on Theorem \ref{theo:info-constraint}, it can be noted that the information constraint has a very attractive property: the problem of maximizing the expected payoff
takes a particularly simple form. Indeed, by defining a one-to-one mapping between the quadruplets $(x_0,x_1,x_2,y)$ and the finite set $\{1,2,...,L\}$, $L = |\mathcal{X}_0 \times \mathcal{X}_1 \times \mathcal{X}_2 \times \mathcal{Y} |$, the optimization problem of interest can be described as follows:
\begin{equation}
\begin{array}{crcl}
\text{minimize} & -\mathbb{E}_{\underline{q}}[w] = - \displaystyle{\sum_{\ell=1}^L} q_\ell w_\ell & & \\
\text{subject to} & I_{\underline{q}}(X_0;X_2) - I_{\underline{q}}(X_1;Y|X_0,X_2) & \leq & 0 \\
& -q_\ell & \leq & 0 \\
& -1 + \displaystyle{\sum_{\ell=1}^L} q_\ell & =& 0 \\
& \forall x_0, \; \sum_{\ell \in \mathcal{L}_{X_0}(x_0)} q_{\ell} -\rho(x_0)
&=& 0 \\
& \forall (x_1,y), \
\frac{\sum_{\ell \in \mathcal{L}_{X_1,Y}(x_1,y)}
q_{\ell} }{\sum_{\ell \in \mathcal{L}_{X_1}(x_1)} q_{\ell}} - \Gamma(y|x_1) &=& 0
\end{array}
\end{equation}
where $q_\ell$ is the probability of a given quadruplet $(x_0,x_1,x_2,y)$,
$w_\ell$ is the value of the corresponding payoff, the vector $\underline{q} = (q_1, ..., q_L)$ represents the distribution $Q$, and the sets of indices
$\mathcal{L}_{X_0}(x_0)$, $\mathcal{L}_{X_1,Y}(x_1,y)$, $\mathcal{L}_{X_1}(x_1)$ merely translate the marginalization
conditions. By Lemma \ref{lemma:convexity-of-phi}, it follows that the above optimization problem is convex, which makes easy the determination of the information-constrained maximum of the expected payoff. A simple and useful upper bound for this maximum is $\mathbb{E}_{\rho} \max_{(x_1,x_2)} w(x_0,x_1,x_2)$. This bound will be referred to as the costless communication case in Sec. \ref{sec:application-PC-IC-numerical-results}. Indeed, this bound can be attained in the ideal scenario where: given the knowledge of the coming state $x_0$, DM 1 computes an optimal solution for the action pair for the coming stage $(x_1^*,x_2^*) \in \arg \max_{(x_1,x_2)} w(x_0,x_1,x_2)$ and can inform DM 2 of $x_2^*$ without any cost. If the state is stationary for say $S$ stages and $\mathcal{X}_1=\mathcal{X}_2$, a simple strategy for DM 1 can be as follows: $x_1(1) = x_2^*$, $x_1(2)=x_1^*$, ..., $x_1(S) = x_1^*$. This allows DM 2 to choose an optimal action for $i\in\{2,...,S\}$. It can be shown that considering the $S-$stage block i.i.d. case amounts to multiplying the left term of (\ref{eq:information-constraint}) by $\frac{1}{S}$, which makes the info constraint arbitrarily mild as $S$ grows large.
\section{Application to power control over
interference channels}
\label{sec:application-PC-IC-numerical-results}
The main goal is to assess the performance of simple CPC policies and those of good policies, the performance of the latter is obtained by exploiting Theorem 1. A flat-fading interference channel (IC) with two Tx-Rx pairs is considered. Transmissions are assumed to be time-slotted and synchronized. For $j\in\{1,2\}$ and ``$k=-j$'' ($-j$ stands for the Tx other than $j$), the SINR at receiver $j$ at a given stage writes as $\mathrm{SINR}_j= \frac{g_{jj} x_j }{\sigma^2 + g_{kj} x_{k}}$
where $x_j \in \mathcal{X}_j^{\text{IC}} = \left\{0, P_{\max}\right\}$ is the power level chosen by Tx $j$, $g_{jk}$ represents the channel gain of link $jk$, and $\sigma^2$ the noise variance. We assume that: $g_{jk} \in \{g_{\min}, g_{\max}\}$ is i.i.d. and Bernouilli distributed $g_{jk} \sim \mathcal{B}(p_{jk})$ with $P(g_{jk} = g_{\min}) = p_{jk}$. We define $\text{SNR[dB]} = 10\log_{10}\frac{P_{\max}}{\sigma^2}$ and set $g_{\min} = 0.1$, $g_{\max}=1.9$, $\sigma^2=1$. The low and high interference regimes (LIR, HIR) are respectively defined by $(p_{11},p_{12},p_{21},p_{22}) = (0.5,0.9,0.9,0.5)$ and $(p_{11},p_{12},p_{21},p_{22}) = (0.5,0.1,0.1,0.5)$. The assumed payoff is $w^{\text{IC}}(x_0,x_1,x_2) = \sum_{j=1}^2 f( \mathrm{SINR}_j(x_0,x_1,x_2))$ where $f(a) = \log(1+a)$ unless stated otherwise. At last we assume that $Y\equiv X_1$. We consider four CPC policies~:\\
$\blacktriangleright$ the full power control (FPC) policy $x_j=P_{\max}$ for every stage. FPC requires no CSI at all;\\
$\blacktriangleright$ the semi-coordinated PC (SPC) policy $x_2=P_{\max}$, $x_1^{\dag} \in \arg \max_{x_1} w^{\text{IC}}(x_0, x_1, P_{\max})$. SPC requires the knowledge of the current state realization at Tx1;\\
$\blacktriangleright$ the optimal CPC policy (OCPC) whose performance are
obtained, in particular, when the problem has the information structure of Theorem 1;\\
$\blacktriangleright$ the costless communication case (see Sec. \ref{sec:comments-main-result}) for which the maximum of $w^{\text{IC}}$ can be reached at any stage. Fig. \ref{fig1} and \ref{fig2} depict the relative gain in \% in terms of average payoff versus SNR[dB] which is obtained by FPC, SPC, OCPC, and costless case. Compared to FPC, gains are very significant whatever the interference regime and provided the SNR has realistic values. Compared to SPC, the gain is of course less impressive since SPC is precisely a coordinated PC scheme but, in the HIR and when the communication cost is negligible, gains as high as $25\%$ can be obtained with $f(a) = \log(1+ a)$ and $45\%$ with $f(a) = a$.
\begin{figure}[htbp]
\includegraphics[width=0.50\textwidth]{Ref1log.eps}
\caption{Relative gain in terms of expected payoff (``OCPC/FPC - 1'' in [\%]) vs SNR[dB] obtained with CPC (with and without communication cost) when the reference power control policy is to transmit at full power (FPC).}
\label{fig1}
\end{figure}
\begin{figure}[htbp]
\includegraphics[width=0.50\textwidth]{Ref2log.eps}
\caption{The difference with Fig. \ref{fig1} is that the reference power control policy is the SPC policy. Additionally, the top curve is obtained with $f(a) =a$.}
\label{fig2}
\end{figure}
\section{Concluding remarks}
\label{sec:conclusion}
Although some assumptions made in this paper might be too restrictive
in some application scenarios, it is essential to understand that the used methodology to derive the optimal performance is general. It can be applied to analyze the performance of coded power allocation, coded interference alignment, etc, with other information structures and by considering $N\geq 2$ individual payoffs instead of a common one (e.g., in a game-theoretic setting \cite{Gossner-2006}). The methodology to assess the performance of good coded policies consists in deriving the right information constraint(s) by building the proof on Shannon theory for the problem of multi-source coding with distortion over multi-user channels wide side information and then to use this constraint to find an information-constrained maximum of the payoff (common payoff case) or the set of Nash equilibrium points which are compatible with the constraint (non-cooperative game case). Note that assuming i.i.d. from stage to stage the state(s) leads in fact to the worst-case scenario for the information constraint. On the other hand, the costless communication case provides an upper bound for the expected payoff. As a key observation, the communication structure of a multi-person decision-making problem is a multiuser channel, which makes multi-terminal Shannon theory not only relevant for pure communication problems but also for any multi-person decision-making problem. This observation opens new challenges for Shannon-theorists since decision-making problems define new channels for instance. It can also be observed that the need to design payoff-oriented communications urges a rethinking of the problem of coding.
\section*{Acknowledgment}
The authors would like to thank Prof. Olivier Gossner for interesting feedbacks
on his work.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Quantum field theory in curved spacetimes has a broad range of applicability.
It covers not only real gravitation, e.g., the expanding universe and black holes, but also non-inertial reference frames, e.g., accelerated frames and rotating frames.
One of the most prominent quantum phenomena in curved spacetimes is particle creation from the vacuum.
The particle creation occurs on black holes, in accelerated frames, and in the expanding universe \cite{Fulling:1972md}.
The understanding of such a quantum process at the full quantum level is a long standing problem.
Despite enormous efforts, the consistent quantization of gravity is still difficult due to nonrenormalizability.
At energies below the Planck scale, quantum effects of gravity can be neglected and gravity can be treated as a classical field.
Whereas quantum gravity is too difficult, quantum field theory with classical gravity is, at least in principle, tamable.
Practical calculations of interacting quantum field theory are, however, not easy even if gravity is classical,.
In particular, in the strong coupling region of quantum chromodynamics (QCD), a perturbative approach does not work successfully.
To study nonperturbative aspects of QCD, we need the lattice simulation, which is an {\it ab initio} nonperturbative approach in QCD.
In this paper, we formulate lattice QCD with external gravitational fields.
The gravitational fields are classical backgrounds in this framework, unlike in lattice quantum gravity \cite{Hamber:2009mt}.
While the backreaction from QCD to gravity is absent, quantum effects of QCD are exactly taken into account.
Lattice QCD in a curved spacetime was first formulated in a specific case of a rotating frame \cite{Yamamoto:2013zwa}.
We extend this formalism to general curved spacetimes.
There are pioneering works of Abelian gauge theory on curved lattices \cite{Jersak:1996mn}.
In addition, some kinds of simulations in flat spacetimes can be regarded as simulations in curved spacetimes.
The examples are an anisotropic lattice with direction-dependent coupling constant \cite{Hasenfratz:1981tw} and an inhomogeneous lattice with coordinate-dependent coupling constant \cite{Huang:1990jf}.
\section{Formulation}
We consider a four-dimensional Riemannian spacetime with the invariant length
\begin{equation}
ds^2 = g_{\mu\nu}(x) dx^\mu dx^\nu.
\end{equation}
The metric tensor $g_{\mu\nu}(x)$ has positive signatures.
For simplicity, we assume that the spacetime is covered with a single global coordinate patch.
In the continuum, the Yang-Mills action is
\begin{equation}
S_{\rm YM} = \int d^4x \sqrt{\det g} \ \frac{1}{2g_{\rm YM}^2} g^{\mu\nu} g^{\rho\sigma} \mathrm{tr} F_{\mu\rho} F_{\nu\sigma}.
\end{equation}
and the fermion action is
\begin{equation}
S_{\rm F} = \int d^4x \sqrt{\det g} \ \bar{\psi}[\gamma^\mu (\partial_\mu+iA_\mu+i\Gamma_\mu) +m] \psi
\end{equation}
\cite{Misner:1974qy}.
The connection is
\begin{equation}
\Gamma_\mu = \frac{1}{4}\sigma^{ij} \omega_{\mu ij}
\end{equation}
with
\begin{eqnarray}
\sigma^{ij} &=& \frac{i}{2} [\gamma^i , \gamma^j]\\
\omega_{\mu ij} &=& g_{\alpha\beta} e^\alpha_i ( \partial_\mu e^\beta_j + \Gamma^\beta_{\mu\nu} e^\nu_j)\\
\Gamma^\beta_{\mu\nu} &=& \frac{1}{2} g^{\beta\rho} \left( \partial_\mu g_{\rho\nu} + \partial_\nu g_{\mu\rho} - \partial_\rho g_{\mu\nu} \right).
\end{eqnarray}
The Greek and Latin indices refer to the coordinate and tangent spaces, respectively.
They are related through the vierbein $e_i^\mu(x)$, which satisfies $e_i^\mu e_j^\nu g_{\mu\nu} = \delta_{ij}$.
The gamma matrix in curved spacetimes is given
\begin{equation}
\gamma^\mu(x) = \gamma^i e_i^\mu(x) ,
\end{equation}
where $\gamma^i$ is the gamma matrix in the flat Euclid space.
In general, there exists the ambiguity of the choice for the vierbein and the Dirac operator depends on it.
We embed a hypercubic lattice into a spacetime where the gravitational field and the coordinate are fixed.
Thus there is no general covariance, i.e., no local gauge invariance of gravity, unlike the dynamical triangulation in lattice quantum gravity \cite{Hamber:2009mt}.
On the lattice, continuum spacetime symmetry is broken to discrete symmetries.
For example, a hypercubic lattice has the symmetries of reflection, discrete rotation of $\pi/2$, and discrete translation of $a$.
Some of them are further broken by background gravitational fields.
The full spacetime symmetry depends on $g_{\mu\nu}(x)$.
We here consider the hypercubic lattice with a single lattice spacing $a$ that is independent of positions and directions, i.e.,
\begin{equation}
\int dx^\mu = a
\end{equation}
between nearest neighbor sites.
The $SU(N_c)$ link variable is obtained by discretizing the path-ordered product of $A_\mu dx^\mu$,
\begin{equation}
U_\mu(x) = \mathcal{P} e^{i\int dx^\mu A_\mu(x)} = e^{iaA_\mu(x)}.
\end{equation}
In this and the following equations, the contraction of the Lorentz indices is not performed unless otherwise explicitly summed.
The building block of the lattice gauge action is the plaquette
\begin{equation}
U_{\mu\nu}(x) = U_\mu(x) U_\nu(x+\hat{\mu}) U^\dagger_\mu(x+\hat{\nu}) U^\dagger_\nu(x) .
\end{equation}
The shorthand notation $\hat{\mu}$ means the unit lattice vector in the $x^\mu$ direction.
We construct three symmetric combinations of the plaquettes,
\begin{eqnarray}
\bar{U}_{\mu\nu} &=& \frac{1}{4} [ U_{\mu\nu} + U_{-\mu\nu} + U_{\mu-\nu} + U_{-\mu-\nu} ] \\
\bar{V}_{\mu\nu\rho} &=& \frac{1}{8} [ ( U_{\mu\nu} - U_{-\mu\nu} )( U_{\nu\sigma}-U_{\nu-\sigma} ) \nonumber \\
&& + ( U_{\mu-\nu} - U_{-\mu-\nu} )( U_{-\nu\sigma}-U_{-\nu-\sigma} ) ] \\
\bar{W}_{\mu\nu\rho\sigma} &=& \frac{1}{16} [ U_{\mu\nu} - U_{-\mu\nu} - U_{\mu-\nu} + U_{-\mu-\nu} ] \nonumber \\
&& \times [ U_{\rho\sigma} - U_{-\rho\sigma} - U_{\rho-\sigma} + U_{-\rho-\sigma} ].
\end{eqnarray}
The plaquettes with negative directions are defined as $U_{\mu-\nu}(x) = U_\mu(x) U^\dagger_\nu(x+\hat{\mu}-\hat{\nu}) U^\dagger_\mu(x-\hat{\nu}) U_\nu(x-\hat{\nu})$ etc.
In the continuum limit,
\begin{eqnarray}
\mathrm{Retr} \bar{U}_{\mu\nu} &=& N_c - \frac{a^4}{2} \mathrm{Retr} F_{\mu\nu} F_{\mu\nu} \\
\mathrm{Retr} \bar{V}_{\mu\nu\rho} &=& - a^4 \mathrm{Retr} F_{\mu\nu} F_{\nu\rho} \\
\mathrm{Retr} \bar{W}_{\mu\nu\rho\sigma} &=& - a^4 \mathrm{Retr} F_{\mu\nu} F_{\rho\sigma}.
\end{eqnarray}
The lattice gauge action in curved spacetimes is
\begin{equation}
\begin{split}
S_{\rm YM} =& \frac{1}{g_{\rm YM}^2} \sum_{x} \sqrt{\det g(x)} \\
& \times\Bigg[ \sum_{\mu\ne\nu} g^{\mu\mu}(x) g^{\nu\nu}(x) \left( N_c -\mathrm{Retr} \bar{U}_{\mu\nu}(x) \right) \\
& - \frac{1}{2} \sum_{ \{\mu\nu\rho\} } g^{\mu\nu}(x) g^{\nu\rho}(x) \mathrm{Retr} \bar{V}_{\mu\nu\rho}(x) \\
& - \frac{1}{2} \sum_{ \{\mu\nu\rho\sigma\} } g^{\mu\rho}(x) g^{\nu\sigma}(x) \mathrm{Retr} \bar{W}_{\mu\nu\rho\sigma}(x) \Bigg].
\end{split}
\end{equation}
The summation $\sum_{ \{\mu\nu\rho\} }$ is performed so as to satisfy $\mu \ne \nu \ne \rho \ne \mu$ and the summation $\sum_{ \{\mu\nu\rho\sigma\} }$ is performed such that $(\mu\nu\rho\sigma)$ is a permutation of $(1234)$.
In this construction, we adopted the condition that it respects reflection symmetry (apart from explicit symmetry breaking by the metric) and has the smallest number of loops.
As long as the continuum limit is the same, other constructions are possible.
For example, $\bar{V}_{\mu\nu\rho}$ can be replaced by $\bar{W}_{\mu\nu\nu\rho}$ but it has larger number of loops.
Among several choices of lattice fermions, we here consider the simplest one, i.e., the Wilson fermion.
The Wilson fermion action in curved spacetimes is
\begin{equation}
\begin{split}
S_{\rm F} =& \sum_{x_1 x_2} a^3 \bar{\psi}(x_1) \Big[ (am+4) \sqrt{\det g(x_1)}\delta_{x_1,x_2} \\
& - \frac{1}{2} \sum_\mu \big\{(1-\gamma^\mu(x_1)) \sqrt{\det g(x_1)} \\
& \times V_\mu(x_1) U_{\mu}(x_1) \delta_{x_1+\hat{\mu},x_2} + (1+\gamma^\mu(x_2)) \\
& \times \sqrt{\det g(x_2)} V^\dagger_\mu(x_2) U^\dagger_{\mu}(x_2)\delta_{x_1-\hat{\mu},x_2} \big\} \Big] \psi(x_2) .
\end{split}
\label{eqSflat}
\end{equation}
The connection is introduced as the Spin(4) link variable
\begin{equation}
V_\mu(x) = e^{ia\Gamma_\mu(x)} .
\end{equation}
The splitting of the arguments ($x_1$ or $x_2$) in Eq.~(\ref{eqSflat}) between adjacent lattice sites is determined by requiring $\gamma^5$ Hermiticity of the lattice Wilson-Dirac operator $\gamma^5 D \gamma^5 = D^\dagger$.
In the formalism of the Wilson fermion, the so-called Wilson term is added to the naive fermion action to kill the artificial poles of doublers.
Since the detailed form of the Wilson term is irrelevant in the continuum limit, we added the usual Wilson term which is the Laplacian $\Delta_\delta = \delta^{\mu\nu} \partial_\mu \partial_\nu$ in a flat spacetime.
We can use the Laplacian $\Delta_g = (1/\sqrt{\det g})\partial_\mu g^{\mu\nu} \sqrt{\det g} \partial_\nu$ although the lattice action becomes more complicated.
As in lattice QCD in flat spacetimes, improved lattice actions can be constructed by adding higher-order terms of the lattice spacing, which are irrelevant in the continuum limit, to reduce discretization artifacts.
\section{Wick rotation}
In the Euclidean path integral formulation, real time $t$ is transformed to imaginary time $\tau$ by the Wick rotation $\tau = -it$.
In Euclidean quantum gravity, the Wick rotation causes serious problems, such as the conformal instability \cite{Gibbons:1976ue}.
Most of the problems are irrelevant for classical gravity because they originate from the dynamical Einstein-Hilbert action and the diffeomorphism invariance.
However, also in classical gravity, the Wick rotation causes the complex metric problem.
For instance, if $g_{\mu 0}(t)$ ($\mu \ne 0$) is nonzero or if $g_{\mu \nu}(t)$ includes an odd function of $t$, then $g_{\mu\nu}(\tau)$ is complex and thus the Euclidean action is complex.
Also, if the connection is an imaginary number, the fermion action is complex.
Even if all the elements of the metric tensor is real, as in the Minkowski space, the action can be negative and thus non-positive definite.
When the total action is not positive definite, the Monte Carlo simulation does not work because of the sign fluctuation.
This is called the sign problem.
This is a known problem of a quark chemical potential and an external electric field \cite{deForcrand:2010ys}.
The fermion action with a chemical potential or an electric field is complex.
In curved spacetimes, the sign problem is more severe because both of the gauge and fermion actions can be complex.
One approach to avoid the sign problem is to change the real parameter that makes the action complex to the imaginary parameter.
This is the analogy of an imaginary quark chemical potential and a Euclidean electric field \cite{deForcrand:2010ys}.
The real information in the Lorentzian spacetime is obtained by analytic continuation.
Thus this approach is justified only when analytic continuation is validated.
If analyticity is lost, for example in the presence of a phase transition, this approach is not justified.
In numerical simulations, analytic continuation is done as numerical extrapolation along the parameter.
This extrapolation is expected to be reliable for small parameter region, i.e., in weakly curved spacetimes.
Another approach is to utilize special symmetry to cancel the complex phase of the action.
For the sign problem of chemical potentials and electric fields, there are several known symmetries, e.g., isospin symmetry \cite{Son:2000xc}.
For the complex metric problem, such symmetry is not yet known.
In contrast to the first approach, the second approach is applicable to large parameter region, i.e., in strongly curved spacetimes.
An entirely different approach is stochastic quantization \cite{Damgaard:1987rr}.
Although stochastic quantization for complex action is not fully understood, it is developing rapidly.
The direct simulation of complex metric systems might be possible in the future.
\section{Renormalization}
The ultraviolet divergence of quantum field theory comes from infinitely short length scale.
When the gravitational field is classical, it has only fixed intrinsic scales.
Therefore classical gravity does not cause new ultraviolet divergences.
Actually, the continuum theory is known to be renormalizable in general curved spacetimes \cite{Buchbinder:1989zz}.
Although there is no formal proof of the renormalizability by lattice regularization, the lattice theory is expected to be renormalizable.
The lattice spacing is affected by renormalization.
The renormalization of the lattice spacing is troublesome in curved spacetimes.
In flat spacetimes, if the classical lattice spacing is homogeneous and isotropic, the renormalized lattice spacing is homogeneous and isotropic because it is protected by spacetime symmetry.
In the curved spacetimes where spacetime symmetry is explicitly broken, even if the classical lattice spacing is homogeneous and isotropic, the renormalized lattice spacing can be inhomogeneous and anisotropic.
A famous example is the renormalization of the anisotropic lattice action \cite{Hasenfratz:1981tw}.
On the anisotropic lattice, the spatial lattice spacing $a_s$ and the temporal lattice spacing $a_\tau$ are different.
Since they are differently affected by the renormalization, the renormalized anisotropic ratio $\xi_{\rm ren}$ deviates from the classical value $\xi_{\rm cl}=a_s/a_\tau$.
In weakly curved spacetimes, the renormalization correction may approximately be neglected.
In a strongly curved spacetime, we need to find a (perturbative or nonperturbative) scheme to determine the renormalized lattice spacing and the physical unit.
In addition to the quantum correction, the lattice spacing receives a classical gravitational correction.
In a curved spacetime, the invariant length is $ds = \sqrt{g_{\mu\nu} dx^\mu dx^\nu}$.
The distance between nearest neighbor sites changes as $a \to \sqrt{g_{\mu\mu}}a$.
When we calculate a two-point correlator as a function of time or distance, we should use the proper time or the proper length $l = \int ds = \sum \sqrt{g_{\mu\mu}}a$.
\section{Simulation}
We explicitly demonstrate the computational implementation of the above formulation.
For a simple and intuitive example, we consider particle production in an expanding space.
The time evolution of a flat three-dimensional space is described by the Friedmann-Lema\'itre-Robertson-Walker metric
\begin{equation}
ds^2 = d\tau^2 + \alpha(\tau)^2 (dx^2 + dy^2 + dz^2),
\end{equation}
which is used for the cosmological model of the expanding universe \cite{Friedmann:1924bb}.
The functional form of the scale factor $\alpha(\tau)$ depends on the contents of the universe, e.g., nonrelativistic matter, radiation, or the cosmological constant.
The lattice gauge action is
\begin{equation}
\begin{split}
S_{\rm YM} =& \beta \sum_{x} \Bigg[ \sum_{k} \alpha(\tau) \left( 1 - \frac{1}{N_c}\mathrm{Retr} \bar{U}_{4k}(x) \right) \\
&+ \sum_{k > l} \alpha(\tau)^{-1} \left( 1 - \frac{1}{N_c}\mathrm{Retr} \bar{U}_{kl}(x) \right) \Bigg]
\end{split}
\label{eqSFLRW1}
\end{equation}
with $k,l=1,$ 2, and 3.
The lattice coupling constant is defined as $\beta = 2N_c/g_{\rm YM}^2$.
This action is similar to the anisotropic lattice action \cite{Hasenfratz:1981tw} but the coefficient depends on coordinates.
The lattice fermion action is
\begin{equation}
\begin{split}
S_{\rm F} =& \sum_{x_1 x_2} a^3 \bar{\psi}'(x_1) \Big[ \delta_{x_1,x_2} \\
&- \kappa \sum_k \big\{(1-\gamma^k\alpha(\tau_1)^{-1}) V_k (x_1) U_k (x_1) \delta_{x_1+\hat{k},x_2} \\
&+ (1+\gamma^k\alpha(\tau_2)^{-1}) V^\dagger_k (x_2) U^\dagger_k (x_2) \delta_{x_1-\hat{k},x_2} \big\} \\
&- \kappa \big\{(1-\gamma^4) \left( \frac{\alpha(\tau_1)}{\alpha(\tau_2)} \right)^{\frac{3}{2}} U_{4}(x_1) \delta_{x_1+\hat{4},x_2} \\
&+ (1+\gamma^4) \left( \frac{\alpha(\tau_2)}{\alpha(\tau_1)} \right)^{\frac{3}{2}} U^\dagger_{4}(x_2)\delta_{x_1-\hat{4},x_2} \big\} \Big] \psi'(x_2).
\end{split}
\label{eqSFLRW2}
\end{equation}
The spinor fields are rescaled as $\bar{\psi}'(x_1) = \alpha(\tau_1)^{3/2}(am+4)^{1/2}\bar{\psi}(x_1)$ and $\psi'(x_2) = \alpha(\tau_2)^{3/2}(am+4)^{1/2}\psi(x_2)$.
The hopping parameter is defined as $\kappa = 1/(2am+8)$.
The connection is
\begin{equation}
V_k(x) = \exp \left(i \gamma^k \gamma^4 \frac{\partial_4 \alpha(\tau)}{2} \right).
\end{equation}
The vierbein is taken as $e_1^1 = e_2^2 = e_3^3 = 1/\alpha$, $e_4^4 = 1$,
and $ e_i^\mu = 0$ for $\mu \ne i$.
In this study, we consider the expanding universe with the cosmological constant.
The scale factor is $\alpha(t) = \alpha_0 e^{Ht}$ in the Lorentzian spacetime.
The parameter $H$ is the Hubble constant.
The naive Wick rotation to imaginary time provides the complex scale factor $\alpha(\tau) = \alpha_0 e^{iH\tau}$, and causes the sign problem.
To avoid this, we introduce the ``imaginary'' Hubble constant $H_I=iH$ and the Euclidean expansion as
\begin{equation}
\alpha(\tau) = \alpha_0 e^{H_I\tau}.
\end{equation}
The metric tensor is real and the lattice action is positive definite, as seen in Eqs.~(\ref{eqSFLRW1}) and (\ref{eqSFLRW2}).
Note that we cannot directly relate the following result to particle production in the Lorentzian spacetime without analytic continuation.
We here treat the Euclidean expansion itself, as the theoretical study of QCD with an imaginary chemical potential.
\begin{figure}[h]
\includegraphics[scale=1]{fig1.pdf}
\caption{\label{fig1}
Expanding lattice.
}
\end{figure}
The geometry is schematically shown in Fig.~\ref{fig1}.
The three-dimensional space starts to expand at $\tau=0$ and ends at $\tau=a(L_\tau-1)$, where Dirichlet boundary conditions are imposed.
The initial scale factor is set to $\alpha(0)=\alpha_0=1$.
In the three-dimensional space, periodic boundary conditions are imposed.
The lattice size is $L_xL_yL_z \times L_\tau = 10^3 \times 20$.
We performed quenched QCD simulation with $\beta = 5.9$ and $\kappa = 0.154$.
We only consider small parameter region $aH_I \ll 1$.
For particle production in the Euclidean expansion, we computed the imaginary particle number of fermions at fixed time slices
\begin{eqnarray}
N_I(\tau) &=& \int d^3x \sqrt{\det g}\ n_I (x) \\
n_I(x) &=& -i j^4(x) = -i \langle \bar{\psi}(x) \gamma^4 \psi(x) \rangle .
\end{eqnarray}
As shown in Figs.~\ref{fig2} and \ref{fig3}, nonzero positive $N_I$ and $n_I$ are produced in the Euclidean expansion.
The data of $N_I$ is rescaled by multiplying a factor $1/(L_xL_yL_z) = 10^{-3}$.
The inequality $N_I/(L_xL_yL_z) \ge n_I$ holds in this expanding space because of $\sqrt{\det g}=\alpha^3 \ge 1$.
From numerical fitting, we obtained $N_I/(L_xL_yL_z) = C_1 H_I [(\tau/a) + C_2]$ and $n_I = C_1 H_I [(\tau/a) + C_2] \alpha^{-3}$ with $C_1 = 0.012 \pm 0.001$ and $C_2 = 57 \pm 8$.
The best-fit functions are shown in the figures.
\begin{figure}[h]
\includegraphics[scale=1.2]{fig2.pdf}
\caption{\label{fig2}
$\tau$-dependence of the total fermion number $N_I$ and the fermion number density $n_I$ with $aH_I=0.03$.
}
\includegraphics[scale=1.2]{fig3.pdf}
\caption{\label{fig3}
$H_I$-dependence of the total fermion number $N_I$ and the fermion number density $n_I$ at $\tau/a=10$.
}
\end{figure}
\section{Conclusion}
We have formulated lattice QCD in curved spacetimes to study gravitational effects on QCD.
We have performed the first simulation to demonstrate the computational implementation.
For practical applications to phenomenology, there are open issues to be discussed in more detail, in particular, analytic continuation and renormalization.
The full QCD simulation is also necessary for the study of particle production.
By performing quantitative analysis, we can discuss nonperturbative QCD effects on cosmology.
For example, the functional dependence on the Hubble constant is essential for a scenario of dark energy \cite{Zhitnitsky:2013pna}.
There are a large number of future developments of this framework.
On the theoretical side, we can also formulate other kinds of lattice field theory in curved spacetimes, e.g., scalar field theory, electroweak gauge theory, and nonrelativistic field theory, and so on.
In scalar field theory, the action includes the renormalizable term $R \phi^2$, which couples to a scalar curvature $R$.
On the practical side, by applying this framework, we can study nonperturbative phenomena of QCD in various curved spacetimes, e.g., on black holes, in the anti-de Sitter space, and so on.
We have considered only the case that the spacetime is covered with a single regular coordinate patch.
In several physically interesting spacetimes, the metric tensor $g_{\mu\nu}$ is singular, e.g., on black holes, or the inverse metric tensor $g^{\mu\nu}$ is singular, e.g., in polar coordinates.
When such singularities exist, we need the scheme to resolve it: (i) transforming a singular coordinate to a regular one, (ii) cutting the region around the singularities, or (iii) introducing several local coordinate patches and gluing them at boundary regions.
The formulation of this scheme on the lattice is also a future work.
\acknowledgments
The author is grateful to Kenji Fukushima and Yuya Tanizaki for useful discussions.
The numerical simulations were performed by using the RIKEN Integrated Cluster of Clusters (RICC) facility.
|
\section{Introduction}
How to allocate resources among multiple agents in an efficient,
effective, and fair way is one of the most important
sustainability problems. Recently it has become an emerging
research topic in AI. Many centralized approaches to allocating
indivisible goods have been proposed (e.g., in \cite{Cramton}). In
these approaches, agents are required to fully reveal their
preferences to some central authority (which computes the final
allocation) and pay for the resources allocated to them at some
prices. However, there are some drawbacks and limitations of these
approaches:
\begin{itemize}
\item the elicitation process and the winner determination
algorithm can be very expensive;
\item agents have to reveal their full preferences, which they
might be reluctant to do (sometimes an elicitation process is
unwelcome);
\item in many real world situations (e.g., assigning courses to
students \cite{Kalinowski,Budish}, and providing employment
training opportunities to unemployed), resources must be allocated
free and monetary side payments \cite{Chevaleyre1} are impossible
or unwelcome.
\end{itemize}
So it is important to design a decentralized elicitation-free
protocol for allocating indivisible goods. \cite{Brams1} adapted a
cake-cutting protocol (a typical decentralized approach for the
allocation of divisible goods \cite{Chen}) to the allocation of
indivisible goods. However, the protocol is typically designed for
the cases when there are only two agents. \cite{Bouveret} studied
a sequential elicitation-free protocol. By applying this protocol,
any number of objects can be allocated to any number of agents.
The sequential protocol is parameterized by a sequential policy
(i.e., a sequence of agents). Agents take turns to pick objects
according to the sequence when the allocation process begins.
In this paper, we define and study a parallel elicitation-free
protocol for allocating indivisible goods to multiple agents.
According to this protocol, a parallel policy (i.e., an agent
selection policy) has to be defined before the public allocation
process can begin. At each stage of the allocation process, some
agents will be selected (according to the parallel policy) to
publicly report their preferred objects among those that remain,
and every reported object will be allocated to an agent reporting
it. If an object is reported by more than one agent, then the
agents reporting it draw lots and the winner could get it. We give
a general definition of parallel policies, which can consider the
allocation history that had happened; and provide eight different
criteria to measure the social welfare induced by parallel
policies.
In fact, any sequential policy applied in the sequential protocol
is in a specific class of parallel policies that are sensitive to
identities. The social welfare criteria considered in
\cite{Bouveret} and \cite{Kalinowski} are three of the eight
criteria proposed in our paper. We introduce two simple parallel
policies (i.e., $\varpi_A$ and $\varpi_L$), which are insensitive
to identities; and compare $\varpi_A$ and the optimal sequential
policies (for small numbers of objects and agents) with respect to
the three social welfare criteria. The results show that the
parallel protocol is promising because $\varpi_A$ outperforms the
optimal sequential policies in most cases.
We further consider strategical issues under $\varpi_A$. We show
that an agent who knows the preferences of other agents can find
in polynomial time whether she has a strategy for getting a given
set of objects regardless of uncertainty arising from lottery. We
also show that if the scoring function of the manipulator is
lexicographic, computing an optimal strategy in the sense of
pessimism is polynomial.
The remainder of this paper is structured as follows: Section 2
briefly reviews the basics of the sequential protocol. Section 3
presents the parallel protocol and introduces the two specific
parallel policies (i.e., $\varpi_A$ and $\varpi_L$). Section 4
compares $\varpi_A$ and sequential policies with respect to
several social welfare criteria. Section 5 considers strategical
issues under $\varpi_A$. Section 6 summarizes the contributions of
this work and discusses future work.
\section{Preliminaries}
A set of $m$ indivisible objects $\mathcal{O}=\{o_1,\ldots,o_m\}$
need to be allocated free to a set of $n$ agents
$\mathcal{N}=\{1,2,\ldots,n\}$. It is supposed that $m\geq n$ and
all agents have strict preferences. $\succ_i$ denotes agent $i$'s
ordinal preference (which is a total strict order) over
$\mathcal{O}$, and $rank_i(o)\in\{1,\ldots,m\}$ denotes the rank
of object $o$ in $\succ_i$. A \emph{profile} $R$ consists of a
collection of rankings, one for each agent: $R=\langle \succ_{1},
\ldots, \succ_{n} \rangle$; $Prof(\mathcal{O},\mathcal{N})$
denotes the set of possible profiles under $\mathcal{O}$ and
$\mathcal{N}$. In the following discussion, if not specified, we
only consider full independence case, where all preference
orderings are equally probable (i.e., $Pr(R)=\frac{1}{(m!)^n}$ for
every $R\in Prof(\mathcal{O},\mathcal{N})$).
Agent $i$'s value function $u_i:2^{\mathcal{O}}\rightarrow
\mathbb{R}$ specifies her valuation $u_i(B)$ on each bundle $B$
with $u_i(\emptyset)=0$. When $B=\{o\}$, we also write $u_i(B)$ as
$u_i(o)$. For any $i\in \mathcal{N}$, $B\subseteq
2^{\mathcal{O}}$, and $o\in\mathcal{O}$, it is assumed that:
\begin{itemize}
\item $u_i$ is additive, i.e., $u_i(B)=\sum_{o'\in B}u(o')$; and
\item $u_i(o)=g(rank_i(o))$, where $g$ is a non-increasing
function from $\{1,\ldots,m\}$ to $\mathbb{R}^+$.
\end{itemize}
$g$ is called the scoring function. $g$ is convex if
$g(x)-g(x+1)\geq g(y)-g(y+1)$ holds for any $x\leq y$. In this
paper, we focus on two prototypical convex scoring functions (let
$k\in\{1,\ldots,m\}$): (\emph{Borda}) $g_B(k)=m-k+1$, and
(\emph{lexicographic}) $g_L(k)=2^{m-k}$.
In the sequential protocol, agents take turns to pick objects
according to a \emph{sequential policy} $\pi\in\mathcal{N}^m$.
$\pi(i)$ denotes the $i^{th}$ agent designated by $\pi$. Given
$\pi$ and a profile $R=\langle \succ_{1}, \ldots, \succ_{n}
\rangle$, if all the agents act truthfully, then the corresponding
allocation history $h_{R}^{\pi}$ is
$\langle\pi(1),o'_1\rangle,\ldots,\langle\pi(m),o'_m\rangle$
(i.e., agent $\pi(k)$ picks object $o'_k$ at time $k$), where
$o'_k\in \mathcal{O}\setminus\{o'_l|1\leq l<k\}$ and
$o'_k\succ_{\pi(k)}o$ for every $o\in
\mathcal{O}\setminus\{o'_l|1\leq l\leq k\}$. Given a scoring
function $g$, agent $i$'s utility at $\pi$ and $R$ (i.e.,
$u_i(\pi, R)$) and $i$'s expected utility at $\pi$ (i.e.,
$u_i^*(\pi)$) are:
$$u_i(\pi, R)=\sum_{o\in\mathcal{O}_i} g(rank_i(o))$$ where
$\mathcal{O}_i=\{o'_k|1\leq k\leq m\;s.t.\;\pi(k)=i\}$, and
$$u_i^*(\pi)=\frac{\sum_{R\in Prof(\mathcal{O},\mathcal{N})}u_i(\pi, R)}{(m!)^n}$$
Given an aggregation function $F$ (which is a symmetric,
non-decreasing function from $(\mathbb{R}^+)^n$ to
$\mathbb{R}^+$), the expected social welfare of a sequential
policy $\pi$ is defined as:
$$sw_F^*(\pi)=F(u_1^*(\pi),\ldots,u_n^*(\pi)).$$
Sequential policy $\pi$ is optimal for
$\langle\mathcal{O,N},g,F\rangle$ if $sw_F^*(\pi)\geq
sw_F^*(\pi')$ for every $\pi'\in \mathcal{N}^m$.
\cite{Bouveret} considered two typical aggregation functions which
correspond to the utilitarian criterion $F_u(u_1,\ldots,
u_n)=\sum_{i=1}^n u_i$ and the Rawlsian egalitarian criterion
$F_e(u_1,\ldots, u_n)=\min\{u_i|1\leq i\leq n\}$. They also showed
that, strict alternation (i.e., $12\ldots n12\ldots n\ldots$) is
optimal for $\langle\mathcal{O},\mathcal{N},g_B,F_u\rangle$ when
$m\leq 12$ and $n=2$, and $m\leq 10$ and $n=3$. But they did not
know whether this is true for every $m$ and $n$.
The following example is modified from the one given in
\cite{Bouveret}. It illustrates the notions introduced in this
section and will be used throughout the paper.
\begin{Example}
\begin{small}
Let $m=5$, $n=3$, and $\pi=12332$. Then $\langle
u_1^*(\pi),u_2^*(\pi),u_3^*(\pi)\rangle$ is $\langle
5,7.2,7.5\rangle$ under $g_B$, and $\langle 16,17.8667,17\rangle$
under $g_L$. Consequently, $sw_{F_u}^*(\pi)=19.7$ under $g_B$,
$sw_{F_e}^*(\pi)=16$ under $g_L$, etc.
Suppose $R=\langle\succ_1,\succ_2,\succ_3\rangle$ s.t.
$\succ_1=o_1\succ o_2\succ o_3\succ o_4\succ o_5$,
$\succ_2=o_4\succ o_2\succ o_5\succ o_1\succ o_3$, and
$\succ_3=o_1\succ o_3\succ o_5\succ o_4\succ o_2$. Then
$h^\pi_R=\langle 1,o_1\rangle\langle 2,o_4\rangle\langle
3,o_3\rangle\langle 3,o_5\rangle\langle 2,o_2\rangle$. $\langle
u_1(\pi,R),u_2(\pi,R),u_3(\pi,R)\rangle$ is $\langle 5,9,7\rangle$
under $g_B$, and $\langle 16,24,12\rangle$ under $g_L$.
\end{small}
\end{Example}
\section{Parallel Protocol and Policies}
Now we introduce a parallel protocol for allocating indivisible
goods. At each stage $t$ of the allocating process, there is a
designated set of agents $\mathcal{N}_t\subseteq\mathcal{N}$ s.t.
each $i\in\mathcal{N}_t$ reports an object (her
preferred object among those that remain). If object $o$ is
reported by only one agent then it is allocated to the agent,
otherwise the agents demanding $o$ draw lots \footnote{We suppose
the lot is fair, i.e., if there are $k$ agents drawing lots then
each one of these agents has $1/k$ chance of winning the lot.} for
the right to get $o$.
The protocol is parameterized by a parallel policy. Formally, a
parallel policy is a function $\varpi: (2^{\mathcal{N}} \times
2^{\mathcal{N}})^*\rightarrow 2^{\mathcal{N}}$. Given a finite
sequence $\sigma=\langle
\mathcal{N}_1,\mathcal{N'}_1\rangle,\ldots,\langle\mathcal{N}_k,\mathcal{N'}_k\rangle$
(where for every $1\leq l\leq k$, $\mathcal{N}_l$ is the set of
agents reporting at stage $l$, and
$\mathcal{N}'_l\subset\mathcal{N}_l$ is the set of agents losing
some lottery at stage $l$), $\varpi$ designates the set of agents
reporting at stage $k+1$. An allocation history induced by
$\varpi$ is in the form of $\langle\mathcal{O}_1,
D_1\rangle\mathcal{N}'_1\langle\mathcal{O}_2,
D_2\rangle\mathcal{N}'_2\ldots\langle\mathcal{O}_p,
D_p\rangle\mathcal{N}'_p\;\textsc{Stop}$, where (suppose $1\leq
k\leq p$, and $1\leq l<p$):
\begin{itemize}
\item $\mathcal{O}_1=\mathcal{O}$,
$\mathcal{N}_1=\varpi(\epsilon)$\footnote{$\epsilon$ denotes the
empty sequence.};
\item $D_k:\mathcal{N}_k\rightarrow\mathcal{O}_k$,
$\mathcal{O}'_k=\{o\in\mathcal{O}_k|\exists
i\in\mathcal{N}_k.D_k(i)=o\}$,
$\mathcal{N}'_k\subset\mathcal{N}_k$ s.t. $\forall
o\in\mathcal{O}'_k|\{i\in\mathcal{N}_k\setminus\mathcal{N}'_k|D_k(i)=o\}|=1$;
\item $\mathcal{O}_{l+1}=\mathcal{O}_l\setminus\mathcal{O}'_l$,
$\mathcal{N}_{l+1}=\varpi(\langle\mathcal{N}_1,\mathcal{N}'_1\rangle,\ldots,\langle\mathcal{N}_l,\mathcal{N}'_l\rangle)$;
\item $\mathcal{O}_k\neq\emptyset$,
$\emptyset\subset\mathcal{N}_k\subseteq\mathcal{N}$, and
$\mathcal{O}_{p}=\mathcal{O}'_{p}$.
\end{itemize}
Intuitively, at stage $k$, $\mathcal{O}_k$ is the set of objects
remaining, $\mathcal{O}'_k$ is the set of objects reported by some
$i\in\mathcal{N}_k$, and for every $i\in\mathcal{N}_k$, $D_k(i)$
is the object reported by $i$. $\langle\mathcal{O}_k,D_k\rangle$
is called the \emph{demand situation} at $k$, and $\textsc{Stop}$
is called the \emph{termination situation}.
Given a parallel policy $\varpi$ and a profile $R=\langle
\succ_{1}, \ldots, \succ_{n} \rangle$, if all the agents act
truthfully, the set of possible histories can be represented as an
allocation structure $\mathrm{S}_R^\varpi=\langle
\mathcal{V,E}\rangle$ s.t. $\mathcal{V}$ and $\mathcal{E}$ are the
minimal sets satisfying the following rules:
\begin{itemize}
\item $\langle
\mathcal{O},D:\varpi(\epsilon)\rightarrow\mathcal{O}\rangle\in\mathcal{V}$
s.t. $rank_i(D(i))=1$ for every $i\in\varpi(\epsilon)$;
\item if there exists a history $h=\ldots\langle\mathcal{O}_k,
D_k\rangle\mathcal{N}'_k$ $\langle\mathcal{O}_{k+1},
D_{k+1}\rangle\ldots$ induced by $\varpi$ such that:
\begin{itemize}
\item$\langle\mathcal{O}_k, D_k\rangle\in\mathcal{V}$, and
\item$\forall i\in\mathcal{N}_{k+1}\forall
o\in\mathcal{O}_{k+1}\setminus\{D_{k+1}(i)\}.D_{k+1}(i)\succ_i o$,
\end{itemize}
then $\langle\langle\mathcal{O}_k,
D_k\rangle,\mathcal{N}'_k,\langle\mathcal{O}_{k+1},
D_{k+1}\rangle\rangle\in\mathcal{E}$, and
$\langle\mathcal{O}_{k+1}, D_{k+1}\rangle\in\mathcal{V}$;
\item$\textsc{Stop}\in\mathcal{V}$, and if there exists a history
$h=\ldots\langle\mathcal{O}_k,
D_k\rangle\mathcal{N}'_k\;\textsc{Stop}$ induced by $\varpi$ s.t.
$\langle\mathcal{O}_k, D_k\rangle\in\mathcal{V}$ then
$\langle\langle\mathcal{O}_k,
D_k\rangle,\mathcal{N}'_k,\textsc{Stop}\rangle\in\mathcal{E}$.
\end{itemize}
It is easy to find that $\mathrm{S}_R^\varpi$ is acyclic, and
$\langle\mathcal{O},D\rangle$ is the root.
Since the allocation process from some demand situation $v\in
\mathcal{V}$ is nondeterministic in general, each rational agent
$i$ is often concerned with her expected utility $\hat{u_i}(v)$
and the minimal utility $\underline{u_i}(v)$ that she can get
regardless of uncertainty. Formally, given a scoring function
$g$, $\hat{u_i}(v)=\underline{u_i}(v)=0$ if $v=\textsc{Stop}$;
otherwise (suppose $v=\langle
\mathcal{O}',D':\mathcal{N}'\rightarrow\mathcal{O}'\rangle$):
$$\hat{u_i}(v)=w+\frac{\sum_{v'\in\mathcal{V}}\#E_{v'}^v\cdot
\hat{u_i}(v')}{\#out_v}$$
$$\underline{u_i}(v)=\min\{\underline{u_i}(\mathcal{N}'',v')|\langle
v,\mathcal{N}'',v'\rangle\in\mathcal{E}\},\;\;\textrm{where}$$
\begin{itemize}
\item
$w=\frac{g(rank_i(D'(i)))}{|\{j\in\mathcal{N}'|D'(j)=D'(i)\}|}$ if
$i\in\mathcal{N}'$, $w=0$ otherwise;
\item $\#E_{v'}^v=|\{\mathcal{N}''\subset \mathcal{N}|\langle
v,\mathcal{N}'',v'\rangle\in\mathcal{E}\}|$;
\item $\#out_v=|\{\langle\mathcal{N}'',v'\rangle\in
2^{\mathcal{N}}\times \mathcal{V}|\langle
v,\mathcal{N}'',v'\rangle\in\mathcal{E}\}|$;
\item $\underline{u_i}(\mathcal{N}'',v')=\underline{u_i}(v')$ if
$i\in\mathcal{N}''\cup(\mathcal{N}\setminus\mathcal{N}')$,
$\underline{u_i}(\mathcal{N}'',v')=\underline{u_i}(v')+g(rank_i(D'(i)))$
otherwise.
\end{itemize}
$\hat{u_i}(\varpi,R)=\hat{u_i}(v)$ and
$\underline{u_i}(\varpi,R)=\underline{u_i}(v)$ are called agent
$i$'s expected utility and minimum utility at $\varpi$ and $R$,
respectively, where $v$ is the root of $\mathrm{S}_R^\varpi$.
Each agent $i\in\mathcal{N}$ can evaluate a given parallel policy
$\varpi$ according to 4 values, i.e., $v_i(y,z,\varpi)$ where:
\begin{itemize}
\item $y,z\in\{\textsf{u,e}\}$,
\item $v_i(\textsf{u},z,\varpi)=\frac{\sum_{R\in
Prof(\mathcal{O},\mathcal{N})}u_i(z,\varpi, R)}{(m!)^n}$,
\item $v_i(\textsf{e},z,\varpi)=\min\{u_i(z,\varpi, R)|R\in
Prof(\mathcal{O},\mathcal{N})\}$,
\item $u_i(\textsf{u},\varpi, R)=\hat{u_i}(\varpi,R)$, and
$u_i(\textsf{e},\varpi, R)=\underline{u_i}(\varpi,R)$.
\end{itemize}
The social welfare induced by $\varpi$ (i.e., $sw(x,y,z,\varpi)$)
can be measured by the 8 possible orderings over 3 elements taken
from $\{\textsf{u,e}\}$. Formally, $x,y,z\in\{\textsf{u,e}\}$,
$sw(\textsf{u},y,z,\varpi)=\sum_{i=1}^n v_i(y,z,\varpi)$, and
$sw(\textsf{e},y,z,\varpi)=\min\{v_i(y,z,\varpi)|1\leq i\leq
n\}$\footnote{Intuitively, $\textsf{u}$ and $\textsf{e}$ denote
the utilitarian principle and the egalitarian principle in social
welfare aggregation, respectively. }.
Any sequential policy $\pi$ can be seen as a parallel policy
$\varpi_\pi$ s.t. $\varpi_\pi(\epsilon)=\pi(1)$ and
$\varpi_\pi(\sigma_k)=\pi(k+1)$ for every $1\leq k<m$, where
$\sigma_k=\langle\{\pi(1)\},\emptyset\rangle,\ldots,\langle\{\pi(k)\},\emptyset\rangle$.
For every profile $R$, there is only one possible history in
$\mathrm{S}_R^{\varpi_\pi}$. So
$\hat{u_i}(\varpi_\pi,R)=\underline{u_i}(\varpi_\pi,R)=u_i(\pi,R)$,
$v_i(\textsf{u},z,\varpi_\pi)=u_i^*(\pi)$, and
$sw(x,\textsf{u,u},\varpi_\pi)=sw_{F_x}^*(\pi)$.
In this paper, we introduce two specific parallel policies:
\emph{all--reporting} $\varpi_A$, where all the agents report at
every stage, and \emph{loser--reporting} $\varpi_L$, where all the
agents losing some lot at the current stage report at the next
stage. Formally, $\varpi_A(\sigma)=\mathcal{N}$ for any sequence
$\sigma$; $\varpi_L(\epsilon)=\mathcal{N}$, and
\begin{displaymath}
\varpi_L(\ldots,\langle\mathcal{N}_k,\mathcal{N'}_k\rangle)=\left\{\begin{array}{ll} \mathcal{N'}_k & \textrm{if} \: \mathcal{N'}_k\neq\emptyset\\
\mathcal{N} & \mathrm{otherwise}
\end{array} \right.
\end{displaymath}
$\varpi_L$ guarantees that every agent can get $\frac{m}{n}$
objects at least. So in the eyes of pessimists, it may be a better
choice than $\varpi_A$.
Note that neither $\varpi_A$ nor $\varpi_L$ mentions identities of
agents. We called this kind of parallel policies are insensitive
to identities. We can get Lemma 1 directly.
\begin{Lemma}\label{lem:lemma1}
Let parallel policy $\varpi$ be insensitive to identities. Then
for every $y,z\in\{\textsf{u,e}\}$, and $i,j\in \mathcal{N}$, we
have $v_i(y,z,\varpi)=v_j(y,z,\varpi)$, and
$sw(\textsf{u},y,z,\varpi)=n\cdot v_i(y,z,\varpi)=n\cdot
sw(\textsf{e},y,z,\varpi)$.
\end{Lemma}
\begin{figure}
\begin{center}
\includegraphics [width=75mm,height=60mm]{11.eps}
\caption{Allocation structures of $\varpi_A$ and $\varpi_L$}
\end{center}
\end{figure}
\begin{Example}
\begin{small}
Consider the situation depicted in Example 1. Figure 1 shows the
allocation structures of $\varpi_A$ and $\varpi_L$, where (let
$1\leq p\leq 3$, $1\leq q\leq 5$, and $\textsf{ud}$ denote the
undefined value):
\begin{itemize}
\item
$v_p=\langle\mathcal{O}_p,D_p:\mathcal{N}\rightarrow\mathcal{O}_p\rangle$,
$v'_q=\langle\mathcal{O}'_q,D'_q:\mathcal{N}'_q\rightarrow\mathcal{O}'_q\rangle$;
\item $\mathcal{O}_1=\mathcal{O}'_1=\mathcal{O}$,
$\mathcal{O}_2=\mathcal{O}'_2=\mathcal{O}'_3=\{2,3,5\}$,
$\mathcal{O}_3=\{5\}$, $\mathcal{O}'_4=\{3,5\}$, and
$\mathcal{O}'_5=\{2,5\}$;
\item $\mathcal{N}'_1=\mathcal{N}'_4=\mathcal{N}'_5=\mathcal{N}$,
$\mathcal{N}'_2=\{1\}$, and $\mathcal{N}'_3=\{3\}$;
\item $d_p\in (\mathcal{O}_p)^{|\mathcal{N}|}$ s.t.
$d_p[i]=D_p(i)$ for every $i\in \mathcal{N}$, i.e., $d_1=\langle
1,4,1\rangle$, $d_2=\langle 2,2,3\rangle$, and $d_3=\langle
5,5,5\rangle$; and
\item $d'_q\in
(\mathcal{O}'_q\cup\{\textsf{ud}\})^{|\mathcal{N}|}$ s.t.
$d'_q[i]=D'_q(i)$ if $i\in \mathcal{N}'_q$, otherwise
$d'_q[i]=\textsf{ud}$, i.e., $d'_1=\langle 1,4,1\rangle$,
$d'_2=\langle 2,\textsf{nd},\textsf{nd}\rangle$, $d'_3=\langle
\textsf{nd},\textsf{nd},3\rangle$, $d'_4=\langle 3,5,3\rangle$,
and $d'_5=\langle 2,2,5\rangle$.
\end{itemize}
Under $g_B$,
$\langle\hat{u}_1(\varpi_A,R),\hat{u}_2(\varpi_A,R),\hat{u}_3(\varpi_A,R)\rangle=\langle
4.8333,8,$ $7.5\rangle$, and $sw(\textsf{u,u,u,}\varpi_A)=20.382$,
etc. Under $g_L$, $\langle \underline{u_1}(\varpi_L,R),$
$\underline{u_2}(\varpi_L,R),\underline{u_3}(\varpi_L,R)\rangle=\langle
8,16,12\rangle$, and $sw(\textsf{e,e,e,}\varpi_L)=4$.
\end{small}
\end{Example}
\section{Comparison}
\cite{Bouveret} studied what are the sequential policies
maximizing social welfare. They considered a utilitarian principle
and an egalitarian principle, in which the social welfare induced
by a sequential policy $\pi$ is measured by the values of
$sw(\textsf{u,u,u},\varpi_\pi)$ and
$sw(\textsf{e,u,u},\varpi_\pi)$, respectively. They computed the
optimal sequential policies for small numbers of objects and
agents using an exhaustive search algorithm, and further
conjectured that the problem of finding an optimal sequential
policy is \textsf{NP}-hard. It has been proved that the
alternating policy (i.e., $1212\ldots$) maximizes the value of
$sw(\textsf{u,u,u},\varpi_\pi)$ for two agents under Borda scoring
function \cite{Kalinowski1}. However, the general problem (i.e.,
finding a sequential policy maximizing the value of $sw(\textsf{u,u,u},\varpi_\pi)$ and
$sw(\textsf{e,u,u},\varpi_\pi)$, for more than two agents, or
under other scoring functions) is still open.
On another hand, parallel policy $\varpi_A$ (i.e., all--reporting)
is very natural and simple, and does not have costly procedures
like finding optimal sequence in sequential protocol. By applying
$\varpi_A$, every agent has a chance to get a remaining object in
every round of the parallel allocation process. But we don't know
if $sw(\textsf{u,u,u},\varpi_A)$ and $sw(\textsf{e,u,u},\varpi_A)$
can be computed in polynomial time. We conjecture it's much harder
than computing $sw(\textsf{u,u,u},\varpi_\pi)$ and
$sw(\textsf{e,u,u},\varpi_\pi)$, because complications arise not
only from uncertainty over profiles but from uncertainty over
lots.
In this section, we will compare the parallel protocol (applying
$\varpi_A$) and the sequential protocol (applying the optimal
sequential policies) with respect to several social welfare
criteria. We demonstrate experimentally that in most cases,
$\varpi_A$ is better than sequential policies. To sum up, the
parallel protocol is promising.
\begin{table}
\caption{$\pi^*$,
$sw(\textsf{u},\textsf{u},\textsf{u},\varpi_{\pi^*})$, and
$sw(\textsf{u},\textsf{u},\textsf{u},\varpi_A)$ under
$g_B$}\label{tab:table1} \setlength\tabcolsep{3pt}
\begin{center}
\begin{scriptsize}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
& \multicolumn{3}{|c|}{$n=2$} & \multicolumn{3}{|c|}{$n=3$} & \multicolumn{3}{|c|}{$n=4$}\\
\hline
$m$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ \\
\hline
$4$ & \underline{2}\underline{2} & $12.292$ & $12.292$ & \underline{3}1 & $13.083$ & $13.297$ & \underline{4} & $13.583$ & $13.885$ \\
\hline
$5$ & \underline{2}\underline{2}1 & $18.625$ & $18.625$ & \underline{32} & $20.033$ & $20.382$ & \underline{4}1 & $20.800$ & $21.351$ \\
\hline
$6$ & \underline{2}\underline{2}\underline{2} & $26.396$ & $26.396$ & \underline{3}\underline{3} & $28.622$ & $28.840$ & \underline{4}\underline{2} & $29.600$ & $30.377$ \\
\hline
$7$ & \underline{2}\underline{2}\underline{2}1 & $35.396$ & $35.396$ & \underline{3}\underline{3}1 & $38.511$ & $38.864$ & & & \\
\hline
$8$ & \underline{2}\underline{2}\underline{2}\underline{2} & $45.820$ & $45.820$ & \underline{3}\underline{3}\underline{2} & $49.936$ & $50.381$ & & & \\
\hline
$9$ & \underline{2}\underline{2}\underline{2}\underline{2}1 & $57.487$ & $57.487$ & & & & & &\\
\hline
$10$ & \underline{2}\underline{2}\underline{2}\underline{2}\underline{2} & $70.569$ & $70.569$ & & & & & &\\
\hline
\end{tabular}
\end{scriptsize}
\end{center}
\end{table}
\begin{table}
\caption{$\pi^*$,
$sw(\textsf{u},\textsf{u},\textsf{u},\varpi_{\pi^*})$, and
$sw(\textsf{u},\textsf{u},\textsf{u},\varpi_A)$ under $g_L$}
\label{tab:table2} \setlength\tabcolsep{3pt}
\begin{center}
\begin{scriptsize}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
& \multicolumn{3}{|c|}{$n=2$} & \multicolumn{3}{|c|}{$n=3$} & \multicolumn{3}{|c|}{$n=4$} \\
\hline
$m$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ \\
\hline
$4$ & \underline{2}\underline{2} & $20.458$ & $20.458$ & \underline{3}1 & $23.000$ & $23.460$ & \underline{4} & $24.417$ & $25.458$ \\
\hline
$5$ & \underline{2}\underline{2}1 & $44.725$ & $44.725$ & \underline{3}\underline{2} & $51.933$ & $53.028$ & \underline{4}1 & $56.350$ & $58.477$\\
\hline
$6$ & \underline{2}\underline{2}\underline{2} & $95.371$ & $95.371$ & \underline{3}\underline{3} & $114.27$ & $115.63$ & \underline{4}\underline{2} & $125.26$ & $129.80$ \\
\hline
$7$ & \underline{2}\underline{2}\underline{2}1 & $199.49$ & $199.49$ & \underline{3}\underline{3}1 & $244.64$ & $247.13$ & & &\\
\hline
$8$ & \underline{2}\underline{2}\underline{2}\underline{2} & $412.91$ & $412.91$ & \underline{3}\underline{3}\underline{2} & $516.09$ & $520.79$ & & &\\
\hline
$9$ & \underline{2}\underline{2}\underline{2}\underline{2}1 & $847.64$ & $847.64$ & & & & & & \\
\hline
$10$ & \underline{2}\underline{2}\underline{2}\underline{2}\underline{2} & $1731.0$ & $1731.0$ & & & & & &\\
\hline
\end{tabular}
\end{scriptsize}
\end{center}
\end{table}
\begin{table}
\caption{$\pi^*$,
$sw(\textsf{e},\textsf{u},\textsf{u},\varpi_{\pi^*})$, and
$sw(\textsf{e},\textsf{u},\textsf{u},\varpi_A)$ under $g_B$}
\label{tab:table3} \setlength\tabcolsep{2pt}
\begin{center}
\begin{scriptsize}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
& \multicolumn{3}{|c|}{$n=2$} & \multicolumn{3}{|c|}{$n=3$} & \multicolumn{3}{|c|}{$n=4$} \\
\hline
$m$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ \\
\hline
$4$ & \underline{2}21 & $6.000$ & $6.146$ & \underline{3}3 & $3.750$ & $4.432$ & \underline{4} & $2.500$ & $3.471$\\
\hline
$5$ & 1\underline{2}22 & $9.000$ & $9.313$ & \underline{3}32 & $5.000$ & $6.794$ & \underline{4}4 & $4.500$ & $5.338$\\
\hline
$6$ & \underline{2}\underline{2}21 & $13.125$ & $13.198$ & \underline{3}321 & $9.000$ & $9.613$ & \underline{4}43 & $5.833$ & $7.594$\\
\hline
$7$ & 1\underline{2}2\underline{2}2 & $17.333$ & $17.698$ & \underline{3}2133 & $12.250$ & $12.955$ & & & \\
\hline
$8$ & \underline{2}2\underline{2}1\underline{2} & $22.725$ & $22.910$ & 11332232 & $15.000$ & $16.794$ & & & \\
\hline
$9$ & 1\underline{2}\underline{2}22\underline{2} & $28.429$ & $28.744$ & & & & & &\\
\hline
$10$ & \underline{2}21\underline{2}\underline{2}21 & $35.200$ & $35.285$ & & & & & &\\
\hline
\end{tabular}
\end{scriptsize}
\end{center}
\end{table}
\begin{table}
\caption{$\pi^*$,
$sw(\textsf{e},\textsf{u},\textsf{u},\varpi_{\pi^*})$, and
$sw(\textsf{e},\textsf{u},\textsf{u},\varpi_A)$ under $g_L$}
\label{tab:table4} \setlength\tabcolsep{2pt}
\begin{center}
\begin{scriptsize}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
& \multicolumn{3}{|c|}{$n=2$} & \multicolumn{3}{|c|}{$n=3$} & \multicolumn{3}{|c|}{$n=4$} \\
\hline
$m$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ & $\pi^*$ & $sw_{\pi^*}$ & $sw_A$ \\
\hline
$4$ & \underline{2}21 & $10.000$ & $10.229$ & \underline{3}3 & $7.000$ & $7.820$ & \underline{4} & $3.750$ & $6.365$\\
\hline
$5$ & \underline{2}2\underline{2} & $21.667$ & $22.363$ & \underline{3}32 & $16.000$ & $17.676$ & \underline{4}4 & $12.400$ & $14.619$\\
\hline
$6$ & \underline{2}2\underline{2}1 & $47.500$ & $47.686$ & \underline{3}321 & $36.533$ & $38.543$ & \underline{4}43 & $29.333$ & $32.450$\\
\hline
$7$ & \underline{2}2\underline{2}\underline{2} & $98.400$ & $99.745$ & \underline{3}3213 & $80.229$ & $82.377$ & & & \\
\hline
$8$ & \underline{2}2\underline{2}\underline{2}1 & $205.40$ & $206.46$ & \underline{3}32132 & $168.14$ & $173.60$ & & & \\
\hline
$9$ & \underline{2}2\underline{2}\underline{2}11 & $421.59$ & $423.82$ & & & & & &\\
\hline
$10$ & \underline{2}2\underline{2}\underline{2}1\underline{2} & $862.79$ & $865.50$ & & & & & &\\
\hline
\end{tabular}
\end{scriptsize}
\end{center}
\end{table}
We first compare $\varpi_A$ and sequential policies with respect
to the utilitarian criterion considered by Bouveret and Lang, which are $sw(\textsf{u,u,u},\varpi_A)$
and $sw(\textsf{u,u,u},\varpi_\pi)$, respectively.
For small numbers
of objects and agents (i.e., $m$ and $n$), we compute the optimal
sequential policies (denoted by $\pi^*$) and
$sw(\textsf{u,u,u},\varpi_{\pi^*})$ by use of the tool provided by
Bouveret and Lang (http://
recherche.noiraudes.net/en/sequences.php), and compute
$sw(\textsf{u,u,u},\varpi_A)$ by use of an exhaustive method (on a
PC \textsl{Intel(R) Core(TM) i5-3570K @3.4Ghz}). The time required to compute
$sw(\textsf{u,u,u},\varpi_A)$ grows dramatically in the number n
of agents and in the number m of objects. We find that with $n=3$
and $m=8$ the computation of $sw(\textsf{u,u,u},\varpi_A)$
requires about 8 minutes, but $sw(\textsf{u,u,u},\varpi_A)$ can
not be computed in 12 hours with $n=3$ and $m=9$. The results
under Borda and Lexicographic scoring functions (i.e., $g_B$ and
$g_L$) are shown in Table \ref{tab:table1} and Table
\ref{tab:table2}\footnote{Note that in Tables 1 to 4,
$\underline{n}$, $sw_{\pi^*}$, and $sw_A$ denote the sequence
$12\ldots n$, the social welfare induced by $\pi^*$ and
$\varpi_A$.}, respectively. From Table \ref{tab:table1} and Table
\ref{tab:table2}, we can get that when $n=2$ and $10\geq m\geq 4$,
the values of $\pi^*$ and $\varpi_A$ are equal; however, when
$n>2$, the values of $\varpi_A$ are strictly greater than those of
$\pi^*$. These results suggest that, for small numbers of agents
and objects, we could have a better utilitarian social welfare if
we apply $\varpi_A$ rather than $\varpi_{\pi^*}$. We conjecture
that it is not a coincidence, but we could not find a proof.
\begin{Conjecture}\label{con:con1}
Under any convex scoring function $g$ and for any number $m\geq n$
of objects,
$sw(\textsf{u},\textsf{u},\textsf{u},\varpi_{\pi^*})=sw(\textsf{u},\textsf{u},\textsf{u},\varpi_A)$
when $n=2$, and
$sw(\textsf{u},\textsf{u},\textsf{u},\varpi_{\pi^*})<sw(\textsf{u},\textsf{u},\textsf{u},\varpi_A)$
when $n>2$.
\end{Conjecture}
We also compare $\varpi_A$ and sequential policies with respect to
the egalitarian criterion considered by Bouveret and Lang. The
results under $g_B$ and $g_L$ are shown in Table \ref{tab:table3}
and Table \ref{tab:table4}, respectively. We find that the values
of $\varpi_A$ are strictly greater than those of $\pi^*$ in all
the test cases. In fact, $\varpi_A$ is insensitive to identities.
So according to Lemma \ref{lem:lemma1},
$sw(\textsf{e},\textsf{u},\textsf{u},\varpi_A)=1/n\cdot
sw(\textsf{u},\textsf{u},\textsf{u},\varpi_A)$, which is
definitely the fairest way to divide
$sw(\textsf{u},\textsf{u},\textsf{u},\varpi_A)$. However,
sequential policies are sensitive to identities. From Conjecture
\ref{con:con1}, we further conjecture that $\varpi_A$ will always
be a better choice than any sequential policy for the balance of
utilitarianism and egalitarianism.
\begin{Conjecture}\label{con:con2}
Under any convex scoring function $g$, for any number $n$ of
agents and any number $m\geq n$ of objects,
$sw(\textsf{e},\textsf{u},\textsf{u},\varpi_{\pi^*})<sw(\textsf{e},\textsf{u},\textsf{u},\varpi_A)$.
\end{Conjecture}
\cite{Kalinowski} considered a different egalitarian principle in
which the social welfare induced by a sequential policy $\pi$ is
measured by the value of $sw(\textsf{u,e,u},\varpi_\pi)$. They
also computed the optimal sequential policies (denoted by $\pi^*$)
under $g_B$ when $n=2$ and $p\leq 8$. We compute the values of
$sw(\textsf{u},\textsf{e},\textsf{u},\varpi_{\pi^*})$ and
$sw(\textsf{u},\textsf{e},\textsf{u},\varpi_A)$ by use of an
exhaustive method. The result is shown in Table \ref{tab:table5}.
Again, $\varpi_A$ outperforms sequential policies in all the test
cases.
\begin{table}
\caption{$\pi^*$,
$sw(\textsf{u},\textsf{e},\textsf{u},\varpi_{\pi^*})$, and
$sw(\textsf{u},\textsf{e},\textsf{u},\varpi_A)$ under $g_B$ when
$n=2$} \label{tab:table5} \setlength\tabcolsep{6pt}
\begin{center}
\begin{small}
\begin{tabular}{|c|c|c|c|}
\hline
$m$ & $\pi^*$ & $sw(\textsf{u},\textsf{e},\textsf{u},\varpi_{\pi^*})$ & $sw(\textsf{u},\textsf{e},\textsf{u},\varpi_A)$\\
\hline
$2$ & 12 & $1.500$ & $1.750$ \\
\hline
$3$ & 122 & $3.000$ & $3.500$ \\
\hline
$4$ & 1221 & $5.667$ & $5.958$ \\
\hline
$5$ & 12122 & $8.483$ & $8.992$ \\
\hline
$6$ & 121221 & $12.397$ & $12.736$ \\
\hline
$7$ & 1212122 & $16.560$ & $17.082$ \\
\hline
$8$ & 12122121 & $21.738$ & $22.129$\\
\hline
\end{tabular}
\end{small}
\end{center}
\end{table}
\section{Strategical Issues under $\varpi_A$}
In this section, we will discuss strategical issues under
all--reporting policy $\varpi_A$, which is one of the simplest
parallel policies that are insensitive to identities. As most collective decision mechanisms, $\varpi_A$ is not
\emph{strategyproof}. See Example 2. If all the agents play
sincerely, i.e., report their preferred object at each stage, then
$\hat{u}_1(\varpi_A,R)=\frac{1}{2}g(1)+\frac{1}{2}g(2)+\frac{1}{3}g(5)$
and $\underline{u_1}(\varpi_A,R)=0$. Suppose 1 is a pessimist and
believes that she cannot win any lottery. Then she is concerned
only with the utility she can get regardless of uncertainty. If 1
knows other agents' preferences and plays strategically, then she
reports $o_2$ first and she can get $g(2)$ units of utility at
least, which is better than $0=\underline{u_1}(\varpi_A,R)$.
Someone may want to study the impact of strategic behavior on the
complete-information extensive-form game of such parallel
allocation procedures \footnote{In \cite{Kalinowski2}, the
allocation procedure applying the sequential protocol, is viewed
as a finite repeated game with perfect information, where all
agents act strategically. }. However, it is supposed that the
environment matches decentralized elicitation-free protocols'
application motivation. That is to say, we suppose that it is hard
to learn self-interested agents' preferences in advance
\footnote{In the environments where every agent can learn other
agent's preferences in advance, centralized allocation methods
need to be taken into consideration instead. Because in these
cases, the prerequisite to the protection of private preferences
is ruined.}. So we accept the assumption made in \cite{Bouveret},
i.e., all agents but the only one manipulator act truthfully. In
the following discussion, without loss of generality, let 1 be the
manipulator that knows the rankings of the other agents (i.e.,
$\langle\succ_2,\ldots,\succ_n\rangle$), and $o_1\succ_1
o_2\succ_1\ldots\succ_1 o_m$.
Under $\varpi_A$, a \emph{strategy} for agent 1 is a sequence of
objects $\tau=o'_1,\ldots,o'_T$ s.t. $\forall
t,t'\in\{1,\ldots,T\}\:(o'_t\in\mathcal{O}\textrm{ and
}o'_t=o'_{t'}\textrm{ iff }t=t')$ holds. That is to say, $\tau$
specifies which object 1 should report at any stage $1\leq t\leq
T$. Some strategies may fail because some object that 1 intends to
report has already been allocated. We say strategy $\tau$ is
\emph{well--defined} with respect to
$\langle\succ_2,\ldots,\succ_n\rangle$ if at any stage
$t\in\{1,\ldots,T\}$, object $o'_t$ is still available, and there
is no object available after stage $T$.
A manipulation problem $M$ (for agent 1) consists of
$\langle\succ_2,\ldots,\succ_n\rangle$, and a target set of
objects $\mathcal{S}\subseteq\mathcal{O}$. A well-defined strategy
$\tau$ is \emph{successful} for $M$ if, assuming the agents 2 to n
act sincerely, $\tau$ ensures that agent 1 gets all objects in
$\mathcal{S}$. Solving $M$ consists in determining whether there
exists a successful strategy. Below we show that the manipulation
problem $M$ can be solved in polynomial time. First, we define
some notions: for every $i\in\mathcal{N}$,
$A,B\subseteq\mathcal{O}$ s.t. $A\cap B=\emptyset$,
$\textsc{Better}_i(A,B)=\{o\in A|(\forall o'\in B)o\succ_i o'\}$,
and $\textsc{Best}_i(A)=o\in A$ s.t. $o\succ_i o'$ for every
$o'\in A\setminus\{o\}$. We can get Lemma 2 directly.
\begin{Lemma}
Let $A\subseteq C\subseteq\mathcal{O}$, $B\subseteq
D\subseteq\mathcal{O}$, and $C\cap D=\emptyset$. Then for any
$i\in\mathcal{N}$,
$\textsc{Better}_i(A,D)\subseteq\textsc{Better}_i(C,D)\subseteq\textsc{Better}_i(C,B)$,
and if $\textsc{Best}_i(C)\in A$ then
$\textsc{Best}_i(C)=\textsc{Best}_i(A)$, otherwise
$\textsc{Best}_i(C)\succ_i\textsc{Best}_i(A)$.
\end{Lemma}
Second, for a target set $\mathcal{S}\subseteq\mathcal{O}$, we
construct a sequence
$\rho_{\mathcal{S}}=\langle\mathcal{O}'_1,\mathcal{S}_1,\mathcal{O}_1\rangle,\langle\mathcal{O}'_2,\mathcal{S}_2,\mathcal{O}_2\rangle,\ldots$
as follows:
\begin{small}
\begin{itemize}
\item $\mathcal{O}'_1=\mathcal{O}$, $\mathcal{N}'=\{2,\ldots,n\}$;
\item $\mathcal{S}_k=\bigcup_{i\in
\mathcal{N}'}\textsc{Better}_i(\mathcal{O}'_k\cap
\mathcal{S},\mathcal{O}'_k\setminus\mathcal{S})$;
\item
$\mathcal{O}_k=\{o\in\mathcal{O}'_k\setminus\mathcal{S}|(\exists
i\in\mathcal{N}')o=\textsc{Best}_i(\mathcal{O}'_k\setminus\mathcal{S})\}$;
\item
$\mathcal{O}'_{k+1}=\mathcal{O}'_k\setminus(\mathcal{S}_k\cup\mathcal{O}_k)$.
\end{itemize}
\end{small}
Obviously, for every $o\in\mathcal{O}$, there exists one and only
$k\geq 1$ s.t. $o\in \mathcal{S}_k\cup \mathcal{O}_k$. We denote
by $\textsf{app}_{\mathcal{S}}(o)$ the number $k$.
\begin{Lemma}
Let $\mathcal{S}'\subseteq \mathcal{S}\subseteq\mathcal{O}$,
$\rho_{\mathcal{S}'}=\langle\mathcal{O}'_1,\mathcal{S}_1,\mathcal{O}_1\rangle,\langle\mathcal{O}'_2,\mathcal{S}_2,\mathcal{O}_2\rangle,\ldots$,
and
$\rho_{\mathcal{S}}=\langle\mathcal{O}''_1,\mathcal{S}^*_1,\mathcal{O}^*_1\rangle,\langle\mathcal{O}''_2,\mathcal{S}^*_2,\mathcal{O}^*_2\rangle,\ldots$.
Then for every $k\geq 1$ we have $\bigcup_{t\geq 1}^k
\mathcal{S}_{t}\subseteq\bigcup_{t\geq 1}^k \mathcal{S}^*_{t}$ and
$\bigcup_{t\geq
1}^k(\mathcal{S}_t\cup\mathcal{O}_t)\subseteq\bigcup_{t\geq
1}^k(\mathcal{S}^*_t\cup\mathcal{O}^*_t)$.
\end{Lemma}
\textsc{Proof.} (Sketch)
\begin{small}
According to Lemma 2, we have
$\mathcal{S}_1\subseteq\mathcal{S}^*_1$ and
$\mathcal{S}_1\cup\mathcal{O}_1\subseteq\mathcal{S}^*_1\cup\mathcal{O}^*_1$.
Now assume that $\bigcup_{t\geq 1}^k
\mathcal{S}_{t}\subseteq\bigcup_{t\geq 1}^k \mathcal{S}^*_{t}$ and
$\bigcup_{t\geq
1}^k(\mathcal{S}_t\cup\mathcal{O}_t)\subseteq\bigcup_{t\geq
1}^k(\mathcal{S}^*_t\cup\mathcal{O}^*_t)$ for any $k<p$. Then
$\mathcal{O}'_p\supseteq \mathcal{O}''_p$. Let
$\mathcal{N}'=\{2,\ldots,n\}$.
\begin{enumerate}
\item According Lemma 2, $\mathcal{S}_p=\bigcup_{i\in
\mathcal{N}'}\textsc{Better}_i(\mathcal{O}'_p\cap
\mathcal{S}',\mathcal{O}'_p\setminus\mathcal{S}')\subseteq\bigcup_{i\in
\mathcal{N}'}\textsc{Better}_i(\mathcal{O}'_p\cap
\mathcal{S},\mathcal{O}''_p\setminus\mathcal{S})$. Pick an object
$o$ from $\mathcal{S}_p$. If $o\in \mathcal{O}''_p\cap\mathcal{S}$
then there must be $i\in \mathcal{N}'$ s.t.
$o\in\textsc{Better}_i(\mathcal{O}''_p\cap
\mathcal{S},\mathcal{O}''_p\setminus\mathcal{S})$, i.e., $o\in
\mathcal{S}_p^*$. Otherwise
$o\in(\mathcal{O}'_p\setminus\mathcal{O}''_p)\cap\mathcal{S}$,
i.e., there must be some $q<p$ s.t. $o\in\mathcal{S}_q^*$. So
according to the assumption, we have $\bigcup_{t\geq 1}^p
\mathcal{S}_{t}\subseteq\bigcup_{t\geq 1}^p \mathcal{S}^*_{t}$.
\item Pick an object $o'$ from $\mathcal{O}_p$. Then there exists
$i\in \mathcal{N}'$ s.t.
$o'=\textsc{Best}_i(\mathcal{O}'_p\setminus\mathcal{S}')$. It is
easy to find that
$\mathcal{O}''_p\setminus\mathcal{S}\subseteq\mathcal{O}'_p\setminus\mathcal{S}'$.
So if $o'\in \mathcal{O}''_p\setminus\mathcal{S}$ then
$o'=\textsc{Best}_i(\mathcal{O}''_p\setminus\mathcal{S})$, i.e.,
$o'\in\mathcal{O}^*_p$. If $o'\in\mathcal{S}$ then there exists
some $q\leq p$ s.t. $o'\in\mathcal{S}_q^*$. Otherwise (i.e.,
$o'\in \mathcal{O}'_p\setminus(\mathcal{O}''_p\cup\mathcal{S})$)
there exists some $q'<p$ s.t. $o'\in\mathcal{O}_{q'}^*$.
\end{enumerate}
According to items 1 and 2 and the assumption, we have
$\bigcup_{t\geq 1}^p \mathcal{S}_{t}\subseteq\bigcup_{t\geq
1}^p\mathcal{S}^*_{t}$ and $\bigcup_{t\geq
1}^p(\mathcal{S}_t\cup\mathcal{O}_t)\subseteq\bigcup_{t\geq
1}^p(\mathcal{S}^*_t\cup\mathcal{O}^*_t)$. $\quad\square$
\end{small}
Now we can give a simple characterization of successful strategies
in manipulation problems.
\begin{Theorem}
Let
$M=\langle\langle\succ_2,\ldots,\succ_n\rangle,\mathcal{S}\rangle$
be a manipulation problem, and
$\rho_{\mathcal{S}}=\langle\mathcal{O}''_1,\mathcal{S}^*_1,\mathcal{O}^*_1\rangle,\langle\mathcal{O}''_2,\mathcal{S}^*_2,\mathcal{O}^*_2\rangle,\ldots$.
There exists a successful strategy for $M$ iff for any $k\geq 1$
we have $k>|\bigcup_{1\leq t\leq k}\mathcal{S}^*_{t}|$. Moreover,
in this case any strategy $\tau$ starting by reporting the objects
in $\mathcal{S}$, and reporting $o$ before stage
$\textsf{app}_{\mathcal{S}}(o)$ for every $o\in \mathcal{S}$, (and
completed so as to be well-defined) is successful.
\end{Theorem}
\textsc{Proof.} (Sketch)
\begin{small}
We prove the statement by induction on the size of the target set
$\mathcal{S}$. In the case when $\mathcal{S}$ is a singleton
\{o\}, it is easy to find that $|\bigcup_{1\leq t\leq
k}\mathcal{S}^*_{t}|\leq|\{o\}|=1$ for every $k\geq 1$. So
$\mathcal{S}^*_{1}=\emptyset$ (i.e.,
$\textsf{app}_{\mathcal{S}}(o)>1$) iff $k>|\bigcup_{1\leq t\leq
k}\mathcal{S}^*_{t}|$ for any $k\geq 1$. If
$\mathcal{S}^*_1=\emptyset$ (i.e., no agent in $\{2,\ldots,n\}$
reports $o$ at stage 1) then 1 can get $o$ by reporting $o$ first.
If $\mathcal{S}^*_1\neq\emptyset$ (i.e., there must be some agent
in $\{2,\ldots,n\}$ reporting $o$ at stage 1) then there exists no
successful strategy for $M$.
Now assume that the statement holds for any target set whose size
is no more than $p-1$. Consider a target set
$\mathcal{S}=\{o'_1,\ldots,o'_p\}$ s.t.
$\textsf{app}_{\mathcal{S}}(o'_1)\leq\ldots\leq\textsf{app}_{\mathcal{S}}(o'_p)$.
Then $k>|\bigcup_{1\leq t\leq k}\mathcal{S}^*_{t}|$ for any $k\geq
1$ iff $k>|\bigcup_{1\leq t\leq k}\mathcal{S}^*_{t}|$ for any
$p\geq k\geq 1$. Let $\mathcal{S}'=\mathcal{S}\setminus\{o'_p\}$
and
$\rho_{\mathcal{S}'}=\langle\mathcal{O}'_1,\mathcal{S}_1,\mathcal{O}_1\rangle,\langle\mathcal{O}'_2,\mathcal{S}_2,\mathcal{O}_2\rangle,\ldots$.
\begin{itemize}
\item If $k>|\bigcup_{1\leq t\leq k}\mathcal{S}^*_{t}|$ for any
$p\geq k\geq 1$ then:
\begin{enumerate}
\item $\textsf{app}_{\mathcal{S}}(o'_p)>|\bigcup_{1\leq t\leq
\textsf{app}_{\mathcal{S}}(o'_p)}\mathcal{S}^*_{t}|=p$. So
$o'_p\not\in\bigcup_{t\geq
1}^p(\mathcal{S}^*_t\cup\mathcal{O}^*_t)$. From Lemma 3, we have
$o'_p\not\in\bigcup_{t\geq
1}^p(\mathcal{S}_t\cup\mathcal{O}_t)\subseteq\bigcup_{t\geq
1}^p(\mathcal{S}^*_t\cup\mathcal{O}^*_t)$ and $k>|\bigcup_{1\leq
t\leq k}\mathcal{S}^*_{t}|\geq|\bigcup_{1\leq t\leq
k}\mathcal{S}_{t}|$ for any $p\geq k\geq 1$.
\item According to item 1 and the assumption, there exists a
successful strategy $\tau'$ for
$\langle\langle\succ_2,\ldots,\succ_n\rangle,\mathcal{S}'\rangle$
starting by reporting the objects in $\mathcal{S}'$.
\item According to items 1 and 2, if at each stage $k<p$, 1
reports the object specified by $\tau'$, then $o'_p$ is available
and not reported by any $i\in\{2,\ldots,n\}$ at stage $p$. Let
$\tau$ be a well--defined strategy reporting the object specified
by $\tau'$ at any stage $k<p$, and reporting $o'_p$ at stage $p$.
It is easy to find that $\tau$ is successful for $M$.
\end{enumerate}
\item If there exists some $p\geq k\geq 1$ s.t.
$k\leq|\bigcup_{1\leq t\leq k}\mathcal{S}^*_{t}|$, then there must
be some $i\in\{2,\ldots,n\}$ reporting some $o\in\bigcup_{1\leq
t\leq k}\mathcal{S}^*_{t}$ at some stage $k'\leq k$. In this case,
there is no successful strategy for $M$.
\end{itemize}
So the statement holds for any target set whose size is $p$.
$\quad\square$
\end{small}
We develop Algorithm 1 (in which the set of objects
$\mathcal{O}=\{o_1,\ldots,o_m\}$ and the set of agents
$\mathcal{N}=\{1,\ldots,n\}$ are supposed to be global variables)
to find successful strategies. The soundness and and completeness
of the algorithm is from the proof of Theorem 1. It is not hard to
find that Algorithm 1 always terminates and is polynomial in $m$
and $n$.\\
\\
\begin{small}
\begin{tabular}{rl}
\hline\noalign{\smallskip}
& \textbf{Algorithm 1:} Finding a successful strategy $\tau$ for $M$\\
\hline\noalign{\smallskip}
& \textbf{input:} a manipulation
problem
$M=\langle\langle\succ_2,\ldots,\succ_n\rangle,\mathcal{S}\rangle$\\
& \textbf{output:} a successful strategy
$\tau$ for $M$ if it exists,\\
& $\qquad\quad\;\;$otherwise \textsf{failure};\\
1. & $\mathcal{O'}\leftarrow\mathcal{O}$, $\mathcal{S'}\leftarrow\mathcal{S}$, $size\leftarrow 0$, $\tau\leftarrow\varepsilon$, $\tau'\leftarrow\varepsilon$,\\
& $\mathcal{N'}\leftarrow\{2,\ldots,n\}$, $k\leftarrow 1;\quad$ /* Initialization */\\
2. & $\textbf{while}(\mathcal{O'}\neq\emptyset)$\\
3. & $\quad\mathcal{S}^*\leftarrow\bigcup_{i\in\mathcal{N'}}\textsc{Better}_i(\mathcal{S'},\mathcal{O'}\setminus\mathcal{S'})$;\\
4. & $\quad size\leftarrow size+|\mathcal{S}^*|$;\\
5. & $\quad\textbf{if}$ $size\geq k$ $\textbf{then}$ $\textbf{return}$ $\textsf{failure}$;\\
6. & $\quad\textbf{for}$ all $o\in\mathcal{S}^*$\\
7. & $\qquad \tau\leftarrow\tau\bullet o;\quad$ /* $\bullet$ denotes connection*/\\
8. &
$\quad\mathcal{O}^*\leftarrow\{o\in\mathcal{O}'\setminus\mathcal{S}'|(\exists
i\in\mathcal{N}')o=\textsc{Best}_i(\mathcal{O}'\setminus\mathcal{S}')\}$;\\
9. & $\quad\textbf{if}$ $k>|\mathcal{S}|$ and $\mathcal{O}^*\neq\emptyset$\\
10. & $\qquad$randomly pick an object $o$ from $\mathcal{O}^*$;\\
11. & $\qquad \tau'\leftarrow\tau'\bullet o$; /*completed so as to be well--defined*/\\
12. & $\quad\mathcal{O}'\leftarrow\mathcal{O}'\setminus(\mathcal{S}^*\cup\mathcal{O}^*)$, $\mathcal{S}'\leftarrow\mathcal{S}'\setminus\mathcal{S}^*$, $k\leftarrow k+1$;\\
13. & $\textbf{return}$ $\tau\bullet\tau'$;\\
\noalign{\smallskip} \hline
\end{tabular}
\end{small}
We say a well-defined strategy $\tau$ is \emph{optimal} (in the
sense of pessimism) if it maximizes 1's benefit under the
assumption that 1 can not win any lottery. In fact, it is not hard
to find that if agent 1's scoring function is $g_L$ then she can
find an optimal strategy (in the sense of pessimism) in polynomial
time by applying Algorithm 2.\\
\\
\begin{small}
\begin{tabular}{rl}
\hline\noalign{\smallskip}
& \textbf{Algorithm 2:} Finding an optimal strategy\\
\hline\noalign{\smallskip} & \textbf{input:} a profile $R=\langle\succ_1,\ldots,\succ_n\rangle$\\
& \textbf{output:} an optimal strategy $\tau$ in the sense of
pessimism\\
1. & $\mathcal{O'}\leftarrow\mathcal{O}$, $\mathcal{S}\leftarrow\emptyset;\quad$ /* Initialization */\\
2. & $\tau\leftarrow\textbf{Algorithm 1}(\langle\succ_2,\ldots,\succ_n\rangle,\emptyset)$;\\
3. & $\textbf{while}(\mathcal{O'}\neq\emptyset)$\\
4. & $\quad
o\leftarrow\textsc{Best}_1(\mathcal{O}')$, $\mathcal{O}'\leftarrow\mathcal{O}'\setminus\{o\}$;\\
5. & $\quad \tau'\leftarrow\textbf{Algorithm 1}(\langle\succ_2,\ldots,\succ_n\rangle,\mathcal{S}\cup\{o\})$;\\
6. & $\quad\textbf{if}$ $\tau'\neq\textsf{failure}$\\
7. & $\qquad \tau\leftarrow\tau',\mathcal{S}\leftarrow\mathcal{S}\cup\{o\};$\\
8. & $\textbf{return}$ $\tau$;\\
\noalign{\smallskip} \hline
\end{tabular}
\end{small}
\\
\\
Let's run Algorithm 2 on the profile $R$ given in Example 1. Then
$\{o_2\}$, i.e., the best set of objects that 1 can manage to get
is found, and a successful strategy for the set (i.e, $o_2,o_3$ or
$o_2,o_5$) is returned. We conjecture that under the Borda scoring
function $g_B$, the problem of finding an optimal strategy is
\textsf{NP}-hard, but we do not have a proof.
\section{Conclusion}
We have defined and studied a parallel elicitation-free protocol
for allocating indivisible goods. The protocol is parameterized by
a parallel policy (i.e., an agent selection policy), which can
consider the allocation history that had happened. We have
compared a special parallel policy (i.e., $\varpi_A$) with
sequential policies for small numbers of objects and agents with
respect to the three social welfare criteria considered in
\cite{Bouveret} and \cite{Kalinowski}. The results show that
$\varpi_A$ outperforms the optimal sequential policies in most
cases. We have also proved that an agent who knows the preferences
of other agents can find in polynomial time whether she has a
successful strategy for a target set; and that if the scoring
function of the manipulator is $g_L$, she could compute an optimal
strategy (in the sense of pessimism) in polynomial time.
There are several directions for future work. One direction would
be to prove the conjectures about the social welfare induced by
$\varpi_A$, and to design other parallel policies that can
outperform $\varpi_A$ in some social welfare criterion. Another
direction would be to find the missing complexity results for
manipulation under $g_B$, and to consider strategical issues under
the assumption that the manipulator believes any lottery is fair.
Furthermore, we plan to design an elicitation-free protocol for
allocating sharable goods \cite{Airiau}.
\section*{Acknowledgments}
This work is supported by the National Natural Science Foundation
of China under grant No.61070232, No.61105039, No.61272295, and
the Fundamental Research Funds for the Central Universities (Issue
Number WK0110000026). Many thanks to the anonymous reviewers for
their comments.
\bibliographystyle{named}
|
\section{Introduction}
The virial theorem was introduced by Clausius in statistical mechanics in 1870 and since then it became important in many other areas in physics (see \cite{GC78} for an historical account). In the original formulation the virial theorem establishes a relationship between the time averages of the kinetic energy and of the scalar product $\mathbf{r} \cdot \mathbf{F}$ of trajectory by force. In the particular case of a conservative system with homogeneous potential, this amounts to a relation between time averages of the kinetic and the potential energy. The modern approach to the virial theorem given in \cite{CFR12} uses Hamiltonian formalism and establishes that, under the general conditions of application of the theorem, the time average of the Poisson bracket of an observable $G$ with the Hamiltonian vanishes, obtaining as a particular case that of a regular Lagrangian system. The original statement of the virial theorem is recovered when $H$ is a sum of a kinetic and a potential term, $H=T+V$, and the virial function is $G=\mathbf{r}\cdot \mathbf{p}$, so that the Hamiltonian vector field $X_G$ is precisely the generator of dilations in phase space. But virial-like theorems are also available for systems with configuration space different from $\mathbb{R}^N$ \cite{LZC11}.
More concretely, the geometric version of the Virial Theorem (VT) given in~\cite{CFR12} is as follows. On a Poisson manifold $(M,\{\cdot,\cdot\})$, every function $H\in C^\infty(M)$ defines a dynamical system by $\dot{x}=X_H(x)=\{x,H\}$. Then, if a function $G$ remains bounded, the time average of the Poisson bracket $\{G,H\}$ vanishes:
\
\dd<\{G,H\}>=0.
\]
By the time average we mean $\dd<F>=\lim_{T\to\infty}\frac{1}{T}\int_0^T F(\gamma(t))\,dt$, where $\gamma$ is the evolution curve.
In what follows we implicitly assume that the virial function $G$ remains bounded. When
the motion of the dynamical system is periodic with period $\tau$ the time average reduces to $\dd<F>=\frac{1}{\tau}\int_0^\tau F(\gamma(t))\,dt$.
\section{Virial theorem in quasi-coordinates}
In many problems in classical mechanics and control theory it is useful to consider quasi-velocities.
For instance, in studying the rotation of a rigid body, it is traditional to use Euler's angles to parametrize the orientation of the body while using body angular velocities to describe the dynamics. Similarly, for a system with nonholonomic constraints (i.e. constraints on the velocities that are not derivable from position constraints) one can define quasi-velocities in such a way that some of them coincide with the constraints, obtaining in this way fewer equations to solve.
\subsection{Quasi-velocities and quasi-momenta}
We consider a configuration manifold $Q$ where a mechanical system is evolving.
The traditional concept of velocities and momenta are obtained when considering a local chart $(U,q^1,\ldots,q^n)$, and taking the coordinate basis $\{\partial/\partial q^j\}$ and its dual $\{dq^j\}$. Then, $v=v^j\partial/\partial q^j$ and $\zeta=p_j\,dq^j$, with $v^j=\<dq^j,v>$ and $p_j=\<\zeta,\partial/\partial q^j>$ being the usual velocities and momenta.
Alternatively we can chose a local basis of vector fields on $Q$,
$\{X_1,\ldots,X_n\}$, and the dual basis $\{\alpha^1,\ldots,\alpha^n\}$. Any tangent vector $v\in T_qQ$ can be expressed uniquely as $v=w^jX_j(q)$. The real numbers $(w^1,\ldots,w^n)$ are called the \textit{quasi-velocities} of $v$ in the given basis. In terms of the dual basis $w^j=\langle\alpha^j(q),v\rangle$. Similarly a covector $\zeta\in T^*_qQ$ can be expressed as $\zeta=\pi_j\alpha^j(q)$, and then $(\pi_1,\ldots,\pi_n)$ are called the quasi-momenta of $\zeta$ in the given basis, which can be obtained as $\pi_j=\langle{\zeta},X_j(q)\rangle$.
The pair $(q^i,w^i)$ is called the quasi-coordinates of $v\in TQ$ and the pair $(q^i,\pi_k)$ is called the quasi-coordinates of $\zeta\in T^*Q$ (see \cite{CNCS07}).
The relation between standard velocities and quasi-velocities is given by the well-known basis change formulas. If $X_j=\beta_{j}^{k}(q)\partial_{q^k}$ is the coordinate expression of the vector field $X_j$ in the coordinate basis then $dq^j=\beta^j_k(q)\alpha^k$ and it follows that
$v^i=w^j\beta_{j}^{i}(q)$ and $\pi_k=p_i\beta_{k}^{i}(q)$.
A system of quasi-coordinates has an associated set of local functions on $Q$ called Hamel's symbols given by
$
\gamma_{ml}^k=\beta^{j}_{m}\beta^{i}_{l}\bigl(\pd{\alpha^k_{j}}{q^i}-\pd{\alpha^k_{i}}{q^j}\bigr)
$, where $[\alpha^{i}_{m}]$ is the inverse matrix of $[\beta^{m}_{j}]$, i.e.\ $\alpha^{i}_{m}\beta^{m}_{j}=\delta^i_{j}$. They can be defined by means of $d\alpha^k=-\frac{1}{2}\gamma^k_{ml}\alpha^m\wedge\alpha^l$, or alternatively by $[X_m,X_l]=\gamma^k_{ml}X_k$.
\subsection{VT in the Hamiltonian formalism and quasi-momenta}
A function in the phase space, $H\in C^\infty(T^*Q)$, determines an associated Hamiltonian vector field $X_H$ by the dynamical equation $i_{X_H}\omega_0=dH$, where $\omega_0=-d\theta_0$ is the canonical symplectic form on $T^*Q$. The motions of the system are the integral curves of $X_H$. In quasi-coordinates $(q^i,\pi_i)$ on the cotangent bundle $T^*Q$, the differential of an arbitrary function $G\in C^{\infty}(T^*Q)$ is given by
$
dG=X_j(G)\alpha^j+\pd{G}{\pi_j}d\pi^j
$.
The canonical 1-form $\theta_0$ has the expression $\theta_0=\pi_k\alpha^k$ and the canonical symplectic form $\omega_0=-d\theta_0$ is locally given by
$
\omega_0=\alpha^i\wedge d\pi_i+\frac{1}{2}\pi_k\gamma^k_{ij}\alpha^i\wedge\alpha^j
$.
Therefore, the Hamiltonian vector field associated to the function $G$ is given by
\[
X_G=\pd{G}{\pi_i}X_i-\Bigl(\beta_i^j\pd{G}{q^j}+\pi_k\gamma^k_{ij}\pd{G}{\pi_j}\Bigr)\pd{}{\pi_i}.
\]
Given a virial function $G$ on $T^*Q$, the virial theorem in the Hamiltonian formulation written in quasi-coordinates is
\[
\DD<\beta^j_i\pd{G}{\pi_i}\pd{H}{q^j}-\beta^j_i\pd{G}{q^j}\pd{H}{\pi_i}-\pi_k\gamma^k_{ij}\pd{G}{\pi_j}\pd{H}{\pi_i}>=0.
\]
This equation provides a geometric interpretation of the VT as presented
in \cite{QG98} by using the Poincar\'{e}'s formalism. Particularly important are fibrewise linear virial functions. Every vector field $D$ on the base manifold $Q$ is associated with a linear function $G\in C^\infty(T^*Q)$ defined by $G(\zeta)=\langle\zeta,D(q)\rangle$ for $\zeta\in T^*_qQ$. The associated Hamiltonian vector field is the complete lift $D^c$ of $D$ to $T^*Q$. In quasi-coordinates $(q^i,\pi_i)$ on $T^*Q$, if $D$ has the expression $D=f^iX_i$ then $G(q,\pi)=\pi_kf^k(q)$ and the Hamiltonian vector field has the expression
\[
X_G=D^c=f^iX_i-\Bigl(\beta_i^j\pd{f^k}{q^j}+\gamma^k_{ij}f^j\Bigr)\pi_k\pd{}{\pi_i}.
\]
For such a function the virial theorem can be expressed in the form
\begin{equation}\label{VTLFqc}
\DD<\beta^j_if^i\pd{H}{q^j}-\beta^j_i\pd{f^k}{q^j}\pi_k\pd{H}{\pi_i}-\pi_k\gamma^k_{ij}f^j\pd{H}{\pi_i}>=0.
\end{equation}
\subsection{VT in the Lagrangian formalism and quasi-velocities}
Consider now a dynamical system defined by a regular Lagrangian $L\in C^\infty(TQ)$. The dynamical vector field $\Gamma_L\in\mathfrak{X}(TQ)$ is determined by the dynamical equation $i_{\Gamma_L}\omega_L=dE_L$, where $\omega_L=-d\theta_L$ is the Cartan 2-form associated to the Lagrangian and $E_L$ is the energy function defined by $L$. In quasi-coordinates $(q^i,w^i)$ on the tangent bundle $TQ$, the differential of an arbitrary function $G\in C^{\infty}(TQ)$ is given by
$dG=X_j(G)\alpha^j+\pd{G}{w^j}dw^j$,
and the Cartan 2-form $\omega_L=-d\theta_L$ by
$$
\omega_L=\frac{1}{2}\left[\gamma^k_{ml}\pd L{w^k}%
+X_l\left(\pd{L}{w^m}\right)-X_m\left(\pd{L}{w^l}\right)\right]
\alpha^m\land \alpha^l +\pd{^2L}{w^j\partial w^k}\,%
\alpha^k\land dw^j\ .
$$
Therefore, the dynamical vector field $\Gamma_L=X_{E_L}$ is given by
$$
\Gamma_{L}= w^jX_j+
W^{r l}\left[w^m\gamma^k_{m l}\frac{\partial L}{\partial w^k}
-w^m X_m\left(\frac{\partial L}{\partial w^l}\right)+
X_l(L)\right]\pd{}{w^r},
$$
where $[W^{rl}]$ is the inverse matrix of $[\partial^2 L/\partial w^l\partial w^r]$, and the Hamiltonian vector field of the function $G$ is
$$
X_{G}= W^{jl}\pd{G}{w^l} X_j
+W^{jl}\left\{\left[\pd{L}{w^k}\gamma^k_{m l}+X_l\left(\pd{L}{w^m}\right) -X_m\left(\pd{L}{w^l}\right)\right]
W^{m r}\pd{G}{w^r}
- X_l(G)\right\} \pd{}{w^j}.
$$
For a virial function $G$ the virial theorem takes the form
\begin{equation}
\DD< \pd{G}{w^r} W^{r l}\left[w^m X_m\left(\frac{\partial L}{\partial w^l}\right) - X_l(L) - w^m\gamma^k_{m l}\frac{\partial L}{\partial w^k} \right]-w^jX_j(G)>=0
\end{equation}
The above equation provides a geometric interpretation of the Boltzmann's formalism of the VT. An important case is that of the function $G=\langle\theta_L,D^c\rangle$, where $D$ is a vector field on $Q$ and $D^c$ is its complete lift to $TQ$. It was proved in~\cite{CFR12} that $\{G,E_L\}=\Gamma_LG=D^cL$, from where it follows that on the condition of the virial theorem we have $\dd<D^cL>=0$. In quasi-coordinates, if $D=f^iX_i$ then the expression of the complete lift is
\[
D^c= f^iX_i +\left[X_k(f^i)+\gamma^i_{kj}f^j\right]w^k\pd{}{w^i},
\]
and therefore
\[
\DD<
f^iX_i(L) +\left[X_k(f^i)+\gamma^i_{kj}f^j\right]w^k\pd{L}{w^i}
>=0.
\]
If moreover the Lagrangian is of mechanical type, $L=T-V$, then the VT has the form $\dd<D^c(T)>=\dd<D(V)>$. In coordinates, turns out to be
\begin{equation}\label{VTLFqcM}
\DD< f^iX_i(T)+\left[X_k(f^i)+\gamma^i_{kj}f^j\right]w^k\pd{T}{w^i}>=\dd<f^jX_j(V)>.
\end{equation}
\begin{example}[Kepler problem and quasi-velocities]
Let us consider a particle $P$ of mass $m$ moving in a plane under the action of a central force $F(r)=-\gamma mm'/r^2$ on the direction of a fixed point $O$ of mass $m'\gg m$, where $\gamma$ is a positive constant and $r$ represents the distance between $O$ and the particle $P$. The configuration space of the system is $Q=\mathbb{R}^2-\{O\}$. Let $\theta$ be the angle that the line $OP$ makes with a fixed direction on the plane. Consider as quasi-velocities $w^1=\dot r$ and $w^2=r^2 \, \dot \theta$, corresponding to twice the area swept-out per time unit. Then,
$$
\mathbf{L}(r,\theta,w^1,w^2)= \frac{m}{2}\left[(w^1)^2+\frac{1}{r^2}(w^2)^2\right]+\frac{\gamma mm'}{r}.
$$
Let $D=r\partial_r$ be the infinitesimal generator of dilations on the space $\mathbb{R}^2$ written in polar coordinates. The complete lift of $D$ is the vector field $D^c=r\partial_r+w^1\partial_{w^1}+2w^2\partial_{w^2}$ on the tangent bundle $T\mathbb{R}^2$. If the virial function is defined by $G=\langle\theta_L,D^c\rangle$, that is, $G(r,\theta,w^1,w^2)=m rw^1$, then the Hamiltonian vector field of $G$ turns out to be $X_G=r\partial_r-w^1\partial_{w^1}$. Applying formula (\ref{VTLFqcM}), we obtain
$\dd<r\partial_r V> = \dd< m(w^1)^2+m\frac{(w^2)^2}{r^2}>$,
that is, $\dd<-V> = \dd<2T>$ as expected.
\end{example}
\section{Virial theorem for mechanical systems on Lie algebroids}
The geometrical interpretation for quasi-coordinates has been given in \cite{CNCS07} which amounts to forget the tangent structure and to use only the vector bundle structure $\pi:TQ\to Q$ and the Lie algebra structure $[\cdot,\cdot]$ on the set of vector fields on $Q$. This more general framework leads naturally to the use of Lie algebroids. In this formalism we do not have a prefered basis of sections in the bundle induced by the choice of coordinates in the base.
A Lie algebroid is a vector bundle $\tau:A\to M$, together with a Lie algebra structure $[\cdot ,\cdot ]$ in the space of its sections, and a vector bundle map $\rho :A\to TM$ over the identity in the base $M$, called the anchor, that satisfies the Leibniz rule $[\sigma,f \, \eta]=f\,[\sigma,\eta] +(\rho(\sigma)f)\eta,$ for any pair of sections $\sigma,\eta$ of $\tau:A\to M$ and every smooth function $f\in C^\infty(M)$. In the above expression, by $\rho(\sigma)$ we mean the vector field in $M$ given by $\rho(\sigma)(q)=\rho(\sigma(q))$ for every $q\in M$. In this way every section of $A$ defines a dynamical system $\dot{x}=\rho(\sigma)(x)$ on the base manifold $M$, whose orbits are the integral curves of the vector field $\rho(\sigma)$.
The Lie algebroid structure is equivalent to the existence of an exterior differential operator $d:\Omega^\bullet(A)\to \Omega^{\bullet+1}(A)$ on the exterior algebra of $A$-forms, i.e. the algebra of sections of the exterior product $\bigwedge^\bullet A^*\to M$, similar to the de Rham operator on a manifold and also satisfying $d^2=0$.
A choice of local coordinates $(x^1,\ldots,x^n)$ on the base manifold $M$ and a choice of a local basis of sections $\{e_\alpha\}$ of $A$ determines a local coordinate system $(x^i, y^\alpha)$ on the Lie algebroid $A$. The anchor and the bracket are locally determined by functions $\rho^i_{\,j}$ and $C^\gamma_{\alpha\beta}$ on $M$, called structure functions, determined by
$
\rho(e_\alpha)=\rho^i_{\alpha}\pd{}{x^i}\quad\mbox{and}\quad [e_\alpha,e_\beta]=C^\gamma_{\alpha\beta}\,e_\gamma
$.
The structure functions are not arbitrary but satisfy a set of structure equations, that are the local equivalent of the Leibniz identity and the Jacobi identity of the bracket.
\subsection{Symplectic Lie algebroids and general Hamiltonian systems}
By a symplectic Lie algebroid we mean a pair $(A,\omega)$ where $A$ is a Lie algebroid and $\omega\in\Omega^2(A)$ is a symplectic $2$-section, i.e. regular as a bilinear form and closed with respect to the exterior differential operator $d$ of the Lie algebroid $A$ (see \cite{LMM05} for the details). On a symplectic Lie algebroid $(A,\omega)$ every function $H\in C^\infty(M)$ defines a dynamical system on the base manifold $M$ as follows. Given the function $H$, there is a unique section $\sigma_H$ of $A$, called Hamiltonian section of $H$, such that $i_{\sigma_H}\omega=dH$. The vector field $X_H=\rho(\sigma_H)$ is the infinitesimal generator of such a dynamical system.
The Hamiltonian vector field $X_H$ can also be obtained in terms of a Poisson bracket on the base manifold $M$. Indeed, given two function $F,G\in C^\infty(M)$, the bracket defined by $\{F,G\}=\omega(\sigma_F,\sigma_G)$ is a Poisson bracket on $M$. We clearly see the relations $\{F,G\}=i_{\sigma_G}dF=\rho(\sigma_G)F=X_GF=-X_FG$.
For the Poisson structure defined by the symplectic section $\omega$ on the Lie algebroid $A$, the VT implies that $\dd<\rho(\sigma_G)H>=0$. In coordinates, the differential of $G\in C^\infty(M)$ is given by $dG=\rho^i_\alpha\frac{\partial G}{\partial x^i}e^\alpha$, where $\{e^\alpha\}$ is the dual basis of sections of $A^*$. If the symplectic form is locally $\omega=\omega_{\alpha\beta}e^\alpha\wedge e^\beta$ then the VT can be written in the following way
\begin{equation}\label{VTLAF-c}
\DD<\omega^{\alpha\beta}\rho^i_\alpha\rho^j_\beta\pd{H}{x^i}\pd{G}{x^j}>=0,
\end{equation}
where $[\omega^{\alpha\beta}]$ is the inverse matrix of $[\omega_{\alpha\beta}]=[\omega(e_\alpha,e_\beta)]$.
The above construction depends on the availability of a symplectic section on the Lie algebroid $A$. In the next section we will show two important classes of symplectic Lie algebroids which generalize the standard Lagrangian and Hamiltonian approaches of the classical mechanics.
\subsection{Lagrangian and Hamiltonian systems on Lie algebroids}
The Lie algebroid approach to Hamiltonian and Lagrangian mechanics builds on the geometrical structure of the $A$-tangent to a fibre bundle $P$ over the same base, also called the prolongation of $P$ with respect to the Lie algebroid $A$ (see \cite{LMM05,EM01}).
Consider a Lie algebroid $\tau:A\to M$ and let $\nu:P\to M$ be a fibre bundle over the same base manifold $M$. For each point $p\in P$ we consider the vector space
$\mathcal T^A_{p}P =\{(a,v)\in A_q\times T_pP\mid \rho(a)=T_p\nu(v)\}$, where $q=\nu(p)$. The element $(a,v)\in\mathcal{T}_{p}P$ will be denoted by $(p,a,v)$. The manifold $\mathcal T^AP=\cup_{p\in P}\mathcal T^AP$ has a natural vector bundle structure over $P$ with the projection $\nu_1:\mathcal T^A P\to P$ given by $\nu_1(p,a,v)=p$. Moreover, it can be endowed with a natural Lie algebroid structure, as follows. The anchor map $\varrho:{\mathcal T^AP}\to TP$ is the projection onto the third factor and the bracket $[\![\cdot,\cdot]\!]$ is determined by imposing that the bracket of two projectable sections $\mathcal{Z}_i(p)=(p,\sigma_i(\nu(p)), U_i(p))$, $i=1,2$, is $[\![\mathcal{Z}_1,\mathcal{Z}_2]\!](p)=(p,[\sigma_1,\sigma_2](\nu(p)),[U_1,U_2](p))$. Local coordinates $(x^i,u^J)$ on $P$ and a local basis $\{e_\alpha\}$ of sections of $A$ determine a local basis of projectable sections of $\mathcal T^AP$ by $\mathcal{X}_\alpha(p)=(p,e_\alpha(\nu(p)),\rho^i_{\,\alpha}\partial_{x^i}|_p)$ and $\mathcal V_J(p)=(p,0,\partial_{u^J}|_p)$ for all $p\in P$. The structure functions of $\mathcal T^AP$ are $\varrho^i_\alpha=\rho^i_\alpha$, $\varrho^J_K=\delta^J_K$ and $\mathcal{C}^\alpha_{\beta\gamma}=C^\alpha_{\beta\gamma}$, $\mathcal{C}^\alpha_{\beta K}=\mathcal{C}^J_{\beta K}=\mathcal{C}^J_{K L}=0$.
\subsubsection{The Hamiltonian approach}
Let $\tau:A\rightarrow M$ be a Lie algebroid over a manifold $M$, with anchor $\rho$ and bracket $[\cdot,\cdot]$. As the fibre bundle $P$ we may take $\nu:A^*\rightarrow M$, the dual bundle of $A$. Thus we have defined the $A$-tangent to $A^*$. Taking local coordinates $(x^i)$ on $M$ and choosing a basis $\{e_{\alpha}\}$ of sections of $A$ and the dual basis $\{e^{\alpha}\}$, we have the local coordinates $(x^i,\mu_\alpha)$ on the bundle $A^*$, and we can define the local basis $\{\mathcal X_{\alpha},\mathcal P^{\alpha}\}$ of sections of $\mathcal T^AA^*$ as explained above. We will denote by $\{\mathcal X^{\alpha},\mathcal P_{\alpha}\}$ the dual basis. We then have,
$
\varrho(\mathcal X_{\alpha})=\rho^i_{\alpha}\partial_{x^i}\;\mbox{and}\;\varrho(\mathcal P^{\alpha})=\partial_{\mu_{\alpha}},
$
and for a function $f\in\mathcal{C}^\infty(A^*)$ its differential is
$
df=\rho^i_{\alpha}\pd{f}{x^i}\mathcal X^{\alpha}+\pd{f}{\mu_{\alpha}}\mathcal P_{\alpha}
$.
In the $A$-tangent of $A^*$ there is a canonical section $\theta_A$ of $(\mathcal T^A A^*)^*$, called the Liouville section, defined by $\theta_A(a^*)(b,v)=a^*(b)$, for $(b,v)\in \mathcal T_{a^*}A^*$, and a canonical symplectic section $\omega_A=-d\theta_A$. In coordinates, they are given by
$$
\theta_A=\mu_{\alpha}\mathcal X^{\alpha} \quad\mbox{and}\quad\omega_A=\mathcal X^{\alpha}\wedge\mathcal P_{\alpha}+\frac{1}{2}C^{\gamma}_{\alpha\beta}\mu_{\gamma}\mathcal X^{\alpha}\wedge \mathcal X^{\beta}.
$$
The Hamiltonian section $\mathcal X_H\in {\rm Sec}(\mathcal T^AA^*)$ defined by a function $H\in C^\infty(A^*)$ is written in local coordinates
$$
\mathcal X_H=\frac{\partial H}{\partial \mu_{\alpha}}\mathcal X_{\alpha}-\left(C^{\gamma}_{\alpha\beta}\mu_{\gamma}\frac{\partial H}{\partial \mu_{\beta}}+\rho_{\alpha}^{\,i}\frac{\partial H}{\partial x^i}\right)\mathcal P^{\alpha},
$$
and then,
the VT in the Hamiltonian formalism on a Lie algebroid is
\begin{equation}\label{VT-Hf1}
\DD<\rho^{\,i}_{\alpha}\frac{\partial H}{\partial \mu_{\alpha}}\frac{\partial G}{\partial x^i}-\rho_{\alpha}^{\,i}\frac{\partial H}{\partial x^i}\frac{\partial G}{\partial \mu_{\alpha}}-C^{\gamma}_{\alpha\beta}\mu_{\gamma}\frac{\partial H}{\partial \mu_{\beta}}\frac{\partial G}{\partial \mu_{\alpha}}>=0, \quad G\in C^\infty(A^*) .
\end{equation}
\begin{example}
Consider a finite-dimensional Lie algebra $\mathfrak{g}$ as a Lie algebroid over a singleton $M=\{e\}$. For a Hamiltonian $H\in C^\infty(\mathfrak{g}^*)$ and a linear virial function $G(\mu)=\langle a, \mu\rangle$, with $a\in\mathfrak{g}$ a constant vector, the VT becomes $\dd<\mathrm{ad}^*_{\pd{H}{\mu}}\mu>=0$. Taking a local basis on $\mathfrak{g}$ and the corresponding linear coordinates on $\mathfrak{g}^*$ we get $\dd<\mu_\gamma C^{\gamma}_{\alpha\beta}\pd{H}{\mu_\beta}a^\alpha>=0$, where $C_{\alpha\beta}^\gamma$ are the structure constants. Since $a$ is arbitrary we get $\dd<\mu_\gamma C^{\gamma}_{\alpha\beta}\pd{H}{\mu_\beta}>=0$ for every $\alpha=1,\ldots,\dim\mathfrak{g}$.
An important particular case is that of a free rigid body. The Lie algebra is $\mathfrak{g}=\mathfrak{so}(3)$ and the Hamiltonian is defined by $H(\mu)=\frac{1}{2}\mu\cdot I^{-1}\mu$, where $I$ is the inertia tensor. The VT tell us that each component of the cross product $I^{-1}\mu\times\mu$ has vanishing time average.
\end{example}
\subsubsection{The Lagrangian approach}
Similarly, in the general construction above, we can choose $\tau:A\to M$ as the bundle $P$. A regular Lagrangian function $L\in C^\infty(A)$, defines a symplectic section on the $A$-tangent bundle to $A$ as follows. The Cartan 1-section is defined by
$
\theta_L(a,b,v)=\frac{d}{ds}L(a+sb)\Big|_{s=0}$, for $(a,b,v)\in\mathcal T^AA$, and then the Cartan 2-section is $\omega_L=-d\theta_L$, which is symplectic provided that the Lagrangian $L$ is regular. The energy $E_L\in C^{\infty}(A)$ associated to $L$ is given by $E_L=\Delta(L)-L$, where $\Delta$ is the Liouville dilation vector field on the vector bundle~$A$.
In local coordinates, the symplectic form $\omega_L$ is given by
$$
\omega_L =\pd{^2L}{y^\alpha\partial y^\beta}\mathcal{X}^\alpha\land \mathcal{V}^\beta
+\frac{1}{2}\left(\pd{^2L}{x^i\partial y^\alpha}\rho^i_{\,\beta}
-\pd{^2L}{x^i\partial y^\beta}\rho^i_{\,\alpha}+\pd{L}{y^\gamma}C^\gamma_{\alpha\beta}\right)\mathcal{X}^\alpha\land \mathcal{X}^\beta,
$$
The dynamical section $\Gamma_L$, determined by the equation $i_{\Gamma_L}\omega_L=dE_L$, is $\Gamma_L=y^\alpha\mathcal X_\alpha+f^\alpha\mathcal V_\alpha$ with $f^\alpha=W^{\alpha\theta}\Bigl(\rho^i_\theta\pd{L}{x^i}-\rho^i_\beta y^\beta\pd{^2L}{x^i\partial y^\theta}-C^\gamma_{\theta\beta}y^\beta\pd{L}{y^\gamma}\Bigr)$, where $[W^{\alpha\beta}]$ is the inverse matrix of $\bigl[\partial^2L/\partial y^\alpha \partial y^\beta\bigr]$. In the above expressions $\{\mathcal X_\alpha,\mathcal V_\alpha\}$ denotes a basis constructed as in general case and $\{\mathcal X^\alpha,\mathcal V^\alpha\}$ denotes its dual basis.
The differential of a function $G\in C^\infty(A)$ is
$
dG=\rho^i_{\,\alpha} \pd{G}{x^i}\mathcal X^\alpha+\pd{G}{y^\alpha}\mathcal V^\alpha
$ and therefore the virial theorem states that $\dd<\varrho(\Gamma_L)G>=0$, which locally amounts to $\dd<\rho^i_\alpha y^\alpha\pd{G}{x^i}+f^\alpha\pd{G}{y^\alpha}
>=0$, or explicitly
\[
\DD<
\rho^i_\alpha y^\alpha\pd{G}{x^i}+W^{\alpha\theta}\Bigl(\rho^i_\theta\pd{L}{x^i}-\rho^i_\beta y^\beta\pd{^2L}{x^i\partial y^\theta}-C^\gamma_{\theta\beta}y^\beta\pd{L}{y^\gamma}\Bigr)\pd{G}{y^\alpha}
>=0.
\]
\begin{example}
Consider a Lagrangian function $L$ on a finite-dimensional Lie algebra $\mathfrak{g}$, that we consider as a Lie algebroid over a point. For a constant vector $a\in\mathfrak{g}$ we consider the virial function $G(y)=a^\beta\pd{L}{y^\beta}$. The VT becomes $\dd<\mathrm{ad}^*_{y}\pd{L}{y}>=0$, which in quasi-coordinates reads
$
\dd< \pd{L}{y^\gamma}C^{\gamma}_{\alpha\beta}y^\alpha>=0
$,
where we already took into account that $a$ is arbitrary.
In the particular case of a free rigid body, we have $\mathfrak{g}=\mathfrak{so}(3)$ and the Lagrangian is $L(\omega)=\frac{1}{2}\omega\cdot I\omega$, where $I$ are the inertia tensor. It follows that, $\dd<\omega\times I\omega>=0$, in concordance with the result in the Hamiltonian formalism.
\end{example}
Let $\sigma$ be a section of $A$, $\sigma^c$ its complete lift to $\mathcal T^AA$, and take as virial function $G=\langle\theta_L,\sigma^c\rangle$. Then, as it was proved in~\cite{EM01} we have $d_\Gamma G=d_{\sigma^c}L$, or in other words $\{G,E_L\}=d_{\sigma^c}L$. Therefore we have proved the following result.
\begin{theorem
Let $\sigma$ be a section on the Lie algebroid $A$ and let $\sigma^c$ be its complete lift to $\mathcal T^AA$. Assume that
$G=\langle\theta_L,\sigma^c\rangle$ is bounded on its time evolution. Then $\DD<\varrho(\sigma^c)(L)>=0$.
\end{theorem}
\begin{example}
A heavy top can be modeled on the Lie algebroid $S^2\times\mathfrak{so}(3)\to S^2$ with Lagrangian $L=\frac{1}{2}\omega\cdot I\omega - mgl\gamma\cdot e$ (see~\cite{EM01} for the notation and other details). Taking the linear function $G=a\cdot\gamma$ and applying the virial theorem we get $\dd<a\cdot(\gamma\times\omega)>=0$, and since $a$ is arbitrary we arrive to $\dd<\gamma\times\omega>=0$.
On the other hand, we consider a constant vector $a$ on $\mathbb{R}^3\equiv \mathfrak{so}(3)$ and the associated constant section of $A$ given by $\sigma(\gamma)=(\gamma,a)$. The complete lift of $\sigma$ is $\sigma^c=a^i\mathcal X_i-(a\times \omega)^i\mathcal V_i$. Applying preceding theore
we get that $\dd<\rho(\sigma^c)L>=0$, and after an straightforward computation and taking into account that $a$ is arbitrary we arrive at
$
\dd<\omega\times I\omega>=mgl\dd<\gamma\times e>
$.
\end{example}
|
\section{Introduction}
\label{sec:intro}
Ramsey theory refers to a large body of deep results in mathematics whose underlying philosophy is captured succinctly by the statement that ``Every large system contains a large well-organized subsystem.'' This is an area in which a great variety of techniques from many branches of mathematics are used and whose results are important not only to combinatorics but also to logic, analysis, number theory and geometry.
One of the pillars of Ramsey theory, from which many other results follow, is the Hales--Jewett theorem \cite{HJ63}. This theorem may be thought of as a statement about multidimensional multiplayer tic-tac-toe, saying that in a high enough dimension one of the players must win. However, this fails to capture the importance of the result, which easily implies van der Waerden's theorem on arithmetic progressions in colorings of the integers and its multidimensional generalizations. To quote \cite{GrRoSp}, ``The Hales--Jewett theorem strips van der Waerden's theorem of its unessential elements and reveals the heart of Ramsey theory. It provides a focal point from which many results can be derived and acts as a cornerstone for much of the more advanced work."
To state the Hales--Jewett theorem formally requires some notation. Let $[m]$ be the set of integers $\{1, 2, \dots, m\}$. We will refer to elements of $[m]^n$ as points or words and the set $[m]$ as the alphabet. For any $a \in [m]^n$, any $x \in [m]$ and any set $\gamma \subset [n]$, let $a\oplus x\gamma$ be the point of $[m]^n$ whose $j$-th component is $a_j$ if $j \not\in \gamma$ and $x$ if $j \in \gamma$. A combinatorial line is a subset of $[m]^n$ of the form $\{a \oplus x\gamma: 1 \leq x \leq m\}$. The Hales--Jewett theorem is now as follows.
\begin{THM}
For any $m$ and $r$ there exists an $n$ such that any $r$-coloring of the elements of $[m]^n$ contains a monochromatic combinatorial line.
\end{THM}
The original proof of the Hales--Jewett theorem is similar to that of van der Waerden's theorem, using a double induction, where we prove the theorem for alphabet $[m+1]$ and $r$ colors by using the statement for alphabet $[m]$ and a much larger number of colors. This results in bounds of Ackermann type for the dependence of the dimension $n$ on the size of the alphabet $m$. In the late eighties, Shelah \cite{Shelah} made a major breakthrough by finding a way to avoid the double induction and prove the theorem with primitive recursive bounds. This also resulted in the first primitive recursive bounds for van der Waerden's theorem (since drastically improved by Gowers \cite{G01}).
Shelah's proof relied in a crucial way on a lemma now known as the Shelah cube lemma. The simplest case of this lemma says that if we color the edges of the Cartesian product $K_n \times K_n$ in $r$ colors then, for $n$ sufficiently large, there is a rectangle with both pairs of opposite edges receiving the same color. Shelah's proof shows that we may take $n \leq r^{\binom{r+1}{2}} + 1$. In the second edition of their book on Ramsey theory \cite{GrRoSp}, Graham, Rothschild and Spencer asked whether this bound can be improved to a polynomial in $r$. Such an improvement, if it could be generalized, would allow one to improve Shelah's wowzer-type upper bound for the Hales--Jewett theorem to a tower-type bound. The main result of this paper, Theorem~\ref{thm:grid_main} below, answers this question in the negative by providing a superpolynomial lower bound in $r$. We will now discuss this basic case of Shelah's cube lemma, which we refer to as the grid Ramsey problem, in more detail.
\subsection{The grid Ramsey problem}
\label{subsec:intro_grid}
For positive integers $m$ and $n$, let the \emph{grid graph} $\Gamma_{m,n}$
be the graph on vertex set $[m] \times [n]$ where two distinct
vertices $(i,j)$ and $(i', j')$ are adjacent if and only if
either $i=i'$ or $j=j'$. That is, $\Gamma_{m,n}$ is the Cartesian product $K_m \times K_n$.
A \emph{row} of the grid graph $\Gamma_{m,n}$ is a subgraph induced on a
vertex subset of the form $\{i\} \times [n]$ and a \emph{column}
is a subgraph induced on a
vertex subset of the form $[m] \times \{j\}$.
A rectangle in $\Gamma_{m,n}$ is a copy of $K_2 \times K_2$, that is, an induced
subgraph over a vertex subset of the form $\{(i,j), (i',j), (i,j'),(i',j') \}$
for some integers
$1 \le i < i' \le m$ and $1 \le j < j' \le n$. We will usually denote this rectangle
by $(i,j,i',j')$. For an edge-colored grid graph,
an \emph{alternating rectangle} is a rectangle $(i,j,i',j')$
such that the color of the edges $\{(i,j), (i',j)\}$ and
$\{(i,j'), (i',j')\}$ are equal and the color of the edges
$\{(i,j), (i,j')\}$ and $\{(i',j), (i',j')\}$ are equal, that is, opposite sides of the rectangle
receive the same color.
An edge coloring of a grid graph is \emph{alternating-rectangle-free}
(or \emph{alternating-free}, for short) if
it contains no alternating rectangle. The function we will be interested in estimating is the following.
\medskip
\noindent \textbf{Definition.}
(i) For a positive integer $r$, define $G(r)$ as the minimum
integer $n$ for which
every edge coloring of $\Gamma_{n,n}$ with $r$ colors contains an
alternating rectangle.
\noindent (ii) For positive integers $m$ and $n$, define $g(m, n)$ as the
minimum integer $r$ for which there exists an alternating-free
edge coloring of $\Gamma_{m,n}$ with $r$ colors. Define $g(n) = g(n,n)$.
\medskip
Note that the two functions $G$ and $g$ defined above
are in inverse relation to each other, in the sense that
$G(r) = n$ implies $g(n-1) \le r$ and $g(n) \ge r+1$,
while $g(n) = r$ implies $G(r) \ge n+1$ and $G(r-1) \le n$.
We have already mentioned Shelah's bound $G(r) \le r^{r+1 \choose 2}$ + 1.
To prove this, let $n = r^{r+1 \choose 2} + 1$ and suppose that an
$r$-coloring of $\Gamma_{r+1, n}$ is given.
There are at most $r^{r+1 \choose 2}$ ways that one can
color the edges of a fixed column with $r$ colors. Since the number of columns
is $n > r^{r+1 \choose 2}$, the pigeonhole principle implies that
there are two columns which are identically colored. Let
these columns be the $j$-th column and the $j'$-th column and consider the
edges that connect these two columns. Since there are
$r+1$ rows, the pigeonhole principle implies that there are two rows
that have the same color. Let these be the $i$-th row and the
$i'$-th row. Then the rectangle $(i,j, i', j')$ is alternating.
Hence, we see that $G(r) \le n$, as claimed. This argument in fact
establishes the stronger bound
$g(r+1, r^{r+1 \choose 2} + 1) \ge r+1$.
It is surprisingly difficult to improve on this
simple bound. The only known improvement, $G(r) \leq
r^{\binom{r+1}{2}} - r^{\binom{r-1}{2} + 1} + 1$, which improves
Shelah's bound by an additive term, was given by Gy\'arf\'as \cite{Gyarfas}.
Instead, we have focused on the lower bound, proving that $G(r)$ is
superpolynomial in $r$. As already mentioned, this addresses a question
of Graham, Rothschild and Spencer \cite{GrRoSp}. This question was also
reiterated by Heinrich \cite{H90} and by Faudree, Gy\'arf\'as and Sz\H{o}nyi \cite{FaGySz},
who proved a lower bound of $\Omega(r^3)$.
Quantitatively, our main result is the following.
\begin{THM} \label{thm:grid_main}
There exists a positive constant $c$ such that
\[ G(r) > 2^{c (\log r)^{5/2}/\sqrt{\log \log r}}. \]
\end{THM}
We will build up to this theorem, first giving a substantially simpler proof for the weaker bound $G(r) > 2^{c \log^2 r}$. The following theorem, which includes this result, also contains a number of stronger bounds for the off-diagonal case, improving results of Faudree, Gy\'arf\'as and Sz\H{o}nyi \cite{FaGySz}.
\begin{THM} \label{thm:grid_asymmetric}
\begin{enumerate}
\setlength{\itemsep}{1pt} \setlength{\parskip}{0pt}
\setlength{\parsep}{0pt}
\item[(i)] For all $C > e^2$, $g(r^{\log C/2}, r^{r/2C}) \le r$ for large enough $r$.
\item[(ii)] For all positive constants $\varepsilon$, $g(2^{\varepsilon \log^2 r}, 2^{r^{1-\varepsilon}}) \le r$ for large enough $r$.
\item[(iii)] There exists a positive constant $c$ such that $g(cr^2, r^{r^2 /2} / e^{r^2}) \le r$
for large enough $r$.
\end{enumerate}
\end{THM}
Part (i) of this theorem already shows that $G(r)$ is superpolynomial in $r$, while part (ii) implies the more precise bound $G(r) > 2^{c \log^2 r}$ mentioned above, though at the cost of a weaker off-diagonal result. For $(m,n)=(r+1,r^{r+1 \choose 2})$, it is easy to find an alternating-free edge coloring of $\Gamma_{m,n}$ with $r$ colors by reverse engineering the proof of Shelah's bound $G(r) \leq r^{r+1 \choose 2} + 1$. Part (iii) of Theorem~\ref{thm:grid_asymmetric} shows that $n$ can be close to $r^{\binom{r+1}{2}}$ even when $m$ is quadratic in $r$. This goes some way towards explaining why it is so difficult to improve the bound $G(r) \le r^{r+1 \choose 2}$.
\subsection{The Erd\H{o}s--Gy\'arf\'as problem}
\label{subsec:intro_erdos_gyarfas}
The Ramsey-type problem for grid graphs
considered in the previous subsection is closely related to a problem of Erd\H{o}s and
Gy\'arf\'as on generalized Ramsey numbers. To discuss this problem, we need some definitions.
\medskip
\noindent \textbf{Definition.}
Let $k, p$ and $q$ be positive integers satisfying
$p \ge k+1$ and $2 \le q \le {p \choose k}$.
\noindent (i) For each positive integer $r$, define $F_k(r, p, q)$ as the
minimum integer $n$ for which every edge coloring of $K^{(k)}_n$ with $r$ colors
contains a copy of $K^{(k)}_p$ receiving at most $q-1$ distinct colors
on its edges.
\noindent (ii) For each positive integer $n$,
define $f_k(n,p,q)$ as the minimum integer $r$ for which
there exists an edge coloring of $K^{(k)}_n$ with $r$ colors
such that every copy of $K^{(k)}_p$ receives at least $q$ distinct colors
on its edges.
\medskip
For simplicity, we usually write $F(r,p,q) = F_2(r,p,q)$ and $f(n,p,q) = f_2(n,p,q)$.
As in the previous subsection, the two functions are in
inverse relation to each other: $F(r, p, q) = n$ implies $f(n, p, q) \ge r+1$
and $f(n-1, p, q) \le r$,
while $f(n,p,q) = r$ implies $F(r, p, q)\ge n+1$
and $F(r-1,p,q) \le n$.
The function $F(r, p, q)$ generalizes the Ramsey number
since $F(2, p, 2)$ is equal to the Ramsey number $R(p)$.
Call an edge coloring of $K_n$ a
\emph{$(p,q)$-coloring} if every copy
of $K_p$ receives at least $q$ distinct colors on its edges.
The definitions of $F(r,p,q)$ and $f(n,p,q)$ can also be reformulated
in terms of $(p,q)$-colorings. For example, $f(n, p, q)$ asks for the minimum
integer $r$ such that there is a $(p,q)$-coloring of $K_n$ using $r$ colors.
The function $f(n,p,q)$ was first introduced by
Erd\H{o}s and Shelah \cite{E75, E81} and then was
systematically studied by Erd\H{o}s and Gy\'arf\'as \cite{ErGy}.
They studied the asymptotics of $f(n,p,q)$
as $n$ tends to infinity
for various choices of parameters $p$ and $q$. It is clear that $f(n,p,q)$ is increasing
in $q$, but it is interesting to understand the transitions in behavior as $q$ increases.
At one end of the
spectrum, $f(n, p, 2) \le f(n,3,2) \le \lceil \log n \rceil$, while, at the other end, $f(n,p,{p \choose 2}) = {n \choose 2}$
(provided $p \ge 4$). In the middle range, Erd\H{o}s and Gy\'arf\'as proved that
$f(n,p,p) \ge n^{1/(p-2)} - 1$, which in turn implies that $f(n,p,q)$ is polynomial in $n$ for any $q \geq p$.
A problem left open by Erd\H{o}s and Gy\'arf\'as was to determine whether $f(n, p, p-1)$ is also polynomial
in $n$.
For $p = 3$, it is easy to see that $f(n, 3, 2)$ is subpolynomial since it is equivalent to determining
the multicolor Ramsey number of the triangle. For $p = 4$, Mubayi \cite{Mubayi} showed that
$f(n,4,3) \leq 2^{c \sqrt{\log n}}$ and
Eichhorn and Mubayi \cite{EiMu} showed that
$f(n,5,4) \le 2^{c \sqrt{\log n}}$.
Recently, Conlon, Fox, Lee and Sudakov \cite{CoFoLeSu}
resolved the question of Erd\H{o}s and Gy\'arf\'as, proving that
$f(n, p, p-1)$ is subpolynomial for all $p \geq 3$.
Nevertheless, the function $f(n,p,p-1)$ is still far
from being well understood, even for $p=4$, where the
best known lower bound is $f(n,4,3) = \Omega(\log n)$
(see \cite{FoSu, KoMu}).
In this paper, we consider extensions of these problems to hypergraphs.
The main motivation for studying this problem comes from an equivalent formulation
of the grid Ramsey problem (actually, this is Shelah's original formulation).
Let $K^{(3)}(n,n)$ be the $3$-uniform hypergraph with vertex set
$A \cup B$, where $|A| = |B| = n$, and edge set consisting of all those triples which intersect both $A$ and $B$.
We claim that $g(n)$ is within a factor of two of the minimum integer
$r$ for which there exists an $r$-coloring of the edges of $K^{(3)}(n,n)$ such that
any copy of $K_4^{(3)}$ has at least $3$ colors on its edges.
To see the relation, we abuse notation and regard both $A$ and $B$ as copies of the set $[n]$.
For $i \in A$ and $j,j' \in B$, map the edge $\{(i,j), (i, j')\}$ of
$\Gamma_{n,n}$ to the edge $(i,j,j')$ of $K^{(3)}(n,n)$ and,
for $i, i' \in A$ and $j \in B$, map the edge $\{(i,j), (i', j)\}$ of
$\Gamma_{n,n}$ to the edge $(i,i',j)$ of $K^{(3)}(n,n)$.
Note that this defines a bijection between the edges of $\Gamma_{n,n}$
and the edges of $K^{(3)}(n,n)$, where the rectangles of $\Gamma_{n,n}$
are in one-to-one correspondence with
the copies of $K_4^{(3)}$ of $K^{(3)}(n,n)$ intersecting both
sides in two vertices.
Hence, given a desired coloring of $K^{(3)}(n,n)$, we can find a
coloring of $\Gamma_{n,n}$ where all rectangles receive at least
three colors (and are thus alternating-free), showing that $g(n) \le r$.
Similarly, given an alternating-free coloring of $\Gamma_{n,n}$, we may
double the number of colors to ensure that the set of colors used for row edges
and those used for column edges are disjoint. This turns an alternating-free
coloring of $\Gamma_{n,n}$ into a coloring where each $K_4^{(3)}$
receives at least three colors. Hence, as above, we see that $r \le 2 g(n)$.
Therefore, essentially the only difference between $g(n)$ and $f_3(2n, 4, 3)$ is that
the base hypergraph for $g(n)$ is $K^{(3)}(n,n)$ rather than $K_{2n}^{(3)}$. This observation allows
us to establish a close connection between the quantitative estimates for
$f_3(n,4,3)$ and $g(n)$, as exhibited by
the following pair of inequalities (that we will prove in
Proposition \ref{prop:relation_g_f}):
\begin{eqnarray} \label{eq:relation_g_f}
g(n) \le f_3(2n, 4,3) \le 2 \lceil \log n \rceil^2 g(n).
\end{eqnarray}
This implies upper and lower bounds for $f_3(n,4,3)$ and $F_3(r,4,3)$
analogous to those we have established for $g(n)$ and $G(r)$.
More generally, we have the following recursive upper bound for $F_k(r, p, q)$.
\begin{THM} \label{thm:step_down}
For positive integers $r, k$, $p$ and $q$, all greater than $1$
and satisfying $r \ge k$, $p \geq k+1$ and $2 \leq q \leq \binom{p}{k}$,
\[ F_k\left( r, p, q \right) \le r^{{F_{k-1}(r,p-1,q) \choose k-1}}. \]
The above is true even for $q > {p-1 \choose k-1}$, where we trivially
have $F_{k-1}(r,p-1,q) = p-1$.
\end{THM}
By repeatedly applying Theorem \ref{thm:step_down}, we see
that for each fixed $i$ with $0 \le i \leq k$ and large enough $p$,
\[ F_k\left(r, p, {p-i \choose k-i} + 1\right) \le r^{r^{\iddots^{r^{c_{k,p}}} }}, \]
where the number of $r$'s in the tower is $i$.
For $0 < i < k$, it would be interesting to establish a lower bound on $F_k(r, p, {p-i \choose k-i})$
that is significantly larger than this upper bound on $F_k(r,p,{p-i \choose k-i} + 1)$.
This would establish an interesting phenomenon of `sudden jumps'
in the asymptotics of $F_k(r,p,q)$ at the values $q = {p-i \choose k-i}$.
We believe that these jumps indeed occur.
Let us examine some small cases of this problem. For graphs, as mentioned above,
$F(r,p,p)$ is polynomial while $F(r,p,p-1)$ is superpolynomial.
For 3-uniform hypergraphs, $F_3(r, p, {p-1 \choose 2} + 1)$ is polynomial
in terms of $r$. Hence, the first interesting case is to decide
whether the function $F_3(r, p, {p-1 \choose 2})$ is also
polynomial. The fact that $F_3(r,4,3)$ is superpolynomial
follows from Theorem~\ref{thm:grid_main}
and \eqref{eq:relation_g_f},
giving some evidence towards the conjecture that $F_3(r,p,{p-1 \choose 2})$
is superpolynomial. We provide further evidence by establishing
the next case for 3-uniform hypergraphs.
\begin{THM} \label{thm:F_3_5_6}
There exists a positive constant $c$ such that $F_3(r,5,6) \ge 2^{c\log^2 r}$
for all positive integers $r$.
\end{THM}
\subsection{A chromatic number version of the Erd\H{o}s--Gy\'arf\'as problem}
A graph with chromatic number equal to $p$ we call {\it $p$-chromatic}. In the process of studying $G(r)$ (and proving Theorem \ref{thm:grid_main}), we encountered
the following variant of the functions discussed
in the previous subsection, where $K_p$ is replaced by $p$-chromatic subgraphs.
\medskip
\noindent \textbf{Definition.}
Let $p$ and $q$ be positive integers satisfying
$p \ge 3$ and $2 \le q \le {p \choose 2}$.
\noindent (i) For each positive integer $r$, define $F_\chi(r, p, q)$ as the
minimum integer $n$ for which every edge coloring of $K_n$ with $r$ colors
contains a $p$-chromatic subgraph receiving
at most $q-1$ distinct colors on its edges.
\noindent (ii) For each positive integer $n$,
define $f_\chi(n,p,q)$ as the minimum integer $r$ for which
there exists an edge coloring of $K_n$ with $r$ colors
such that every $p$-chromatic subgraph
receives at least $q$ distinct colors on its edges.
\medskip
Call an edge coloring of $K_n$ a \emph{chromatic-$(p,q)$-coloring} if every $p$-chromatic subgraph
receives at least $q$ distinct colors on its edges. As before, the
definitions of $F_{\chi}(r,p,q)$ and $f_{\chi}(n,p,q)$ can be restated in terms of
chromatic-$(p,q)$-colorings.
Also, an edge coloring of $K_n$ is a chromatic-$(p,q)$-coloring if and only
if the union of every $q-1$ color classes induces a graph
of chromatic number at most $p-1$. In some sense,
this looks like the most natural interpretation for the functions
$F_\chi(r,p,q)$ and $f_\chi(n,p,q)$. If we choose to use this definition,
then it is more natural to shift the numbers $p$ and $q$ by $1$. However,
we use the definition above in order to make the connection
between $F_\chi(r,p,q)$ and $F(r,p,q)$ more transparent.
From the definition, we can immediately deduce some simple facts such as
\begin{align} \label{eq:intro_chi_1}
F_{\chi}(r,p,q) \le F(r,p,q), \qquad f_\chi(n,p,q) \ge f(n,p,q)
\end{align}
for all values of $r,p,q, n$ and
\[
f_{\chi}\left(n,p, {p \choose 2}\right) = f\left(n,p, {p \choose 2}\right)
= {n \choose 2}
\]
for all $n \ge p \ge 4$.
Let $n = F_{\chi}(r,p,q)-1$ and
consider a chromatic-$(p,q)$-coloring of $K_n$ that uses $r$
colors in total. Cover the
set of $r$ colors by $\lceil r/(q-1) \rceil$ subsets
of size at most $q-1$. The chromatic number of the graph induced by
each subset is at most $p-1$ and thus, by the product formula
for chromatic number, we see that
\begin{align*} \label{eq:intro_chi_upper}
F_{\chi}(r, p,q) - 1 \le (p-1)^{\lceil r/(q-1) \rceil}.
\end{align*}
This gives a general exponential upper bound.
On the other hand, when $d$ is a positive integer,
$p = 2^d + 1$ and $q=d+1$, a coloring
of the complete graph which we will describe in Section~\ref{sec:preliminaries}
implies that
\[ F_{\chi}(r, 2^d + 1, d + 1) \ge 2^{r} + 1. \]
Whenever $r$ is divisible by $d$, we see that the two bounds
match to give
\begin{align} \label{eq:intro_chi_2}
F_{\chi}(r, 2^d + 1, d + 1) = 2^{r} + 1.
\end{align}
Let us examine the value of $F_\chi(r,p,q)$ for some small values of $p$ and $q$.
By using the observations \eqref{eq:intro_chi_1} and \eqref{eq:intro_chi_2}, we have
\[
F_\chi(r,3,2) = 2^r + 1, \qquad F_\chi(r,3,3) \le F(r, 3, 3) \le r+2 \]
for $p=3$ and
\[
F_\chi(r,4,2) \geq 2^r + 1, \qquad F_\chi(r,4,4) \le F(r, 4, 4) \le r^2 + 2
\]
for $p=4$ and $r \geq 2$. The bound on $F(r,4,4)$ follows since every edge coloring of
the complete graph on $r^2+2$ vertices with $r$ colors contains a vertex $v$ and a set $X$ of size at least
$r+1$ such that all the edges connecting $v$ to $X$ have the same color, say red. If an edge $e$ with vertices in $X$ is red, we get a $K_4$ using at most three colors by taking $v$, the vertices of $e$, and any other vertex in $X$. So we may suppose at most $r-1$ colors are used on the edges with both vertices in $X$. In this case, as $|X| \ge F(r-1,3,3)$, we can find three vertices in $X$ with at most
two colors used on the edges connecting them.
Adding $v$ to this set gives a set of four vertices with
at most three colors used on the edges connecting them.
We show that the asymptotic behavior of $F_\chi(r,4,3)$
is different from both $F_\chi(r,4,2)$ and $F_\chi(r,4,4)$.
\begin{THM} \label{thm:chi_4_3}
There exist positive constants $C$ and $r_0$ such that for all $r \ge r_0$,
\[ 2^{\log^2 r/36} \le F_{\chi}(r, 4, 3) \le C \cdot 2^{130 \sqrt{r\log r}}. \]
\end{THM}
Despite being interesting in its own right, our principal motivation for considering chromatic-$(p,q)$-colorings was to establish the following theorem,
which is an important ingredient in the proof of Theorem \ref{thm:grid_main}.
\begin{THM} \label{thm:chi_slow_grow}
For every fixed positive integer $n$,
there exists an edge coloring of the complete graph $K_n$ with $2^{6\sqrt{\log n}}$ colors
with the following property: for every
subset $X$ of colors with $|X| \ge 2$, the subgraph induced by the edges colored
with a color from $X$ has chromatic number at most $2^{3 \sqrt{|X| \log |X|}}$.
\end{THM}
This theorem has the following immediate corollary.
\begin{COR}
For all positive integers $n, r$ and $s \geq 2$,
\[
f_{\chi}(n, 2^{3\sqrt{s \log s}} + 1, s + 1) \le 2^{6\sqrt{\log n}}
\quad \text{and} \quad
F_{\chi}(r, 2^{3\sqrt{s \log s}} + 1, s + 1) \ge 2^{\log^2 r/36}.
\]
\end{COR}
Our paper is organized as follows.
In Section \ref{sec:preliminaries}, we review two coloring functions that
will be used throughout the paper.
In Section \ref{sec:gridramsey}, we prove Theorems \ref{thm:grid_main} and \ref{thm:grid_asymmetric}.
In Section \ref{sec:eg}, we prove Theorems \ref{thm:step_down} and \ref{thm:F_3_5_6}.
In Section \ref{sec:chi_eg}, we prove Theorems \ref{thm:chi_4_3} and \ref{thm:chi_slow_grow}.
We conclude with some further remarks and open problems in Section~\ref{sec:conclusion}.
\medskip
\noindent \textbf{Notation.}
We use $\log$ for the base 2 logarithm
and $\ln$ for the natural logarithm. For the sake of
clarity of presentation, we systematically omit floor and ceiling signs
whenever they are not essential. We also do not make any serious attempt
to optimize absolute constants in our statements and proofs.
The following standard asymptotic notation will be used throughout.
For two functions $f(n)$ and $g(n)$, we write $f(n) =
o(g(n))$ if $\lim_{n \rightarrow \infty} f(n)/g(n) = 0$ and $f(n) =
O(g(n))$ or $g(n) = \Omega(f(n))$ if there exists a constant $M$ such
that $|f(n)| \leq M|g(n)|$ for all sufficiently large $n$. We also
write $f(n) = \Theta(g(n))$ if both $f(n) = O(g(n))$ and $f(n) =
\Omega(g(n))$ are satisfied.
\section{Preliminaries}
\label{sec:preliminaries}
We begin by defining two edge colorings of the complete graph $K_n$, one a $(3,2)$-coloring
and the other a $(4,3)$-coloring. These will be used throughout the paper.
We denote the $(3,2)$-coloring by $c_B$, where `B' stands
for `binary'. To define this coloring, we let $t$ be the smallest integer such that
$n \leq 2^t$. We consider the vertex set $[n]$ as a subset of $\{0,1\}^t$
by identifying $x$ with $(x_1, \dots, x_t)$, where $\sum_{i=1}^t x_i 2^{i-1}$ is the
binary expansion of $x - 1$. Then, for two vertices $x = (x_1, \dots, x_t)$ and $y = (y_1, \dots, y_t)$,
$c_B(x,y)$ is the minimum $i$ for which $x_i \neq y_i$.
This coloring uses at most $\lceil \log n \rceil$ colors and
is a $(3,2)$-coloring since three vertices cannot all differ in the $i$-th coordinate.
In fact, it is a chromatic-$(2^d + 1, d+1)$-coloring for all integers $d \ge 1$,
since it gives an edge partition
$E = E_1 \cup \dots \cup E_{t}$ of $K_n$ for $t = \lceil \log n \rceil$ such that,
for all $J \subset [t]$, the graph consisting of the edges $\bigcup_{j \in J} E_j$
has chromatic number at most $2^{|J|}$.
The $(4,3)$-coloring, which is a
variant of Mubayi's coloring \cite{Mubayi}, will be denoted by $c_M$. To define this coloring,
we let $t$ be the smallest integer such that $n \leq 2^{t^2}$ and $m = 2^t$. We consider the vertex
set $[n]$ as a subset of $[m]^t$ by identifying $x$ with $(x_1, \dots, x_t)$, this time by examining the
base $m$ expansion of $x-1$. For two vertices
$x = (x_1, \ldots,x_t)$ and $y = (y_1 ,\ldots, y_t)$, let
\[ c_M(x,y) = \Big( \{x_i, y_i\}, a_1, \ldots,a_t \Big), \]
where $i$ is the minimum index in which $x$ and $y$ differ
and $a_j = 0$ or $1$ depending on whether $x_j = y_j$ or not (note that the coloring function
$c_M$ is symmetric in its two variables).
The coloring $c_M$ is a $(3,2)$-coloring since it partitions the edge set of $K_n$ into bipartite graphs
and a simple case analysis similar to that given in \cite{Mubayi} shows that it is a $(4,3)$-coloring (the proof of Theorem \ref{thm:chi_4_3} given in Section \ref{sec:chi_eg} shows that $c_M$ is even a chromatic-$(4,3)$-coloring).
Since $2^{(t-1)^2} < n$, the total number of colors used is at most
\[ m^2 \cdot 2^t = 2^{3 t} < 2^{3 (1 + \sqrt{\log n})} \leq 2^{6 \sqrt{\log n}}. \]
Hence, $c_M$ uses at most $r = 2^{6\sqrt{\log n}}$ colors
to color the edge set of the complete graph on
$n = 2^{\log^2 r/ 36}$ vertices.
\section{The grid Ramsey problem}
\label{sec:gridramsey}
In order to improve the lower bound on $G(r)$, we need to find
an edge coloring of the grid graph which is alternating-free.
The following lemma is the key idea behind our argument.
For two edge-coloring functions $c_1$ and $c_2$ of the complete
graph $K_n$, let $\mathcal{G}(c_1, c_2)$ be the subgraph
of $K_n$ where $e$ is an edge if and only
if $c_1(e) = c_2(e)$.
\begin{LEMMA} \label{lem:row_chromatic}
Let $m, n$ and $r$ be positive integers.
There exists an alternating-rectangle-free edge coloring
of $\Gamma_{m,n}$ with $r$ colors if and only if
there are edge colorings $c_1, \ldots, c_m$ of
the complete graph $K_n$ with $r$ colors satisfying
\[ \chi(\mathcal{G}(c_i, c_j)) \le r \]
for all pairs of indices $i,j$.
\end{LEMMA}
\begin{proof}
We first prove the `if' statement.
Consider the grid graph $\Gamma_{m,n}$. For each $i$, color the edges of
the $i$-th row using the edge coloring $c_i$.
Then, for each distinct pair of indices $i$ and $i'$,
construct auxiliary graphs $H_{i,i'}$ whose vertex
set is the set of edges that connect
the $i$-th row with the $i'$-th row (that is, edges of the form $\{(i,j), (i',j)\}$) and
where two vertices
$\{(i,j), (i',j)\}$ and $\{(i,j'), (i',j')\}$
are adjacent if and only if the two row edges that
connect these two column edges
have the same color.
The fact that $\chi(\mathcal{G}(c_i, c_{i'})) \le r$ implies that
there exists a vertex coloring of $H_{i,i'}$ with $r$ colors. Color the
corresponding edges $\{(i,j), (i',j)\}$ according to this vertex coloring.
Under this coloring, we see that whenever
a pair of edges of the form $\{(i,j), (i,j')\}$ and $\{(i',j), (i',j')\}$
have the same color, the colors of the edges
$\{(i,j), (i',j)\}$ and $\{(i,j'), (i',j')\}$ are distinct.
This gives a coloring of the column edges.
Hence, we found the required alternating-rectangle-free edge
coloring of $\Gamma_{m,n}$ with $r$ colors.
For the `only if' statement, given an alternating-rectangle-free
edge coloring of $\Gamma_{m,n}$ with $r$ colors, define
$c_i$ as the edge-coloring function of the $i$-th row of $\Gamma_{m,n}$,
for each $i \in [m]$. One can easily reverse the
process above to show that the colorings $c_1, \ldots, c_m$ satisfy
the condition. We omit the details.
\end{proof}
To find an alternating-rectangle-free
edge coloring of $\Gamma_{m,n}$, we will find
edge colorings of the rows which satisfy the condition
of Lemma \ref{lem:row_chromatic}.
Suppose that $E(K_n) = E_1 \cup \dots \cup E_t$ is a partition
of the edge set of $K_n$. For an index subset $I \subset [t]$,
we let $\mathcal{G}_I$ be the subgraph of $K_n$ whose edge set is
given by $\bigcup_{i\in I} E_i$.
\begin{LEMMA} \label{lem:grid_coloring}
Let $m, n, r$ and $t$ be positive integers. Suppose that an
edge partition $E(K_n) = E_1 \cup \dots \cup E_t$ of $K_n$ is given.
Let $I$ be a random subset of $[t]$ where
each element in $I$ is chosen independently with probability $1/r$ and suppose that
\[ \mathbb{P}[\chi(\mathcal{G}_I) \ge r+1] \le \frac{1}{2m}. \]
Then $g(m, n) \le r$ and $G(r) \ge \min\{m,n\} + 1$.
\end{LEMMA}
\begin{proof}
For each $i \in [2m]$, choose a vector $v_i \in [r]^t$
independently and uniformly at random.
Let $c_i$ be an edge coloring of $K_n$ with $r$ colors where for
each $t' \in [t]$, we color all the edges in $E_{t'}$ using the
value of the $t'$-th coordinate of $v_i$.
Color the $i$-th row of $\Gamma_{2m,n}$ (which is a copy of $K_n$)
using $c_i$.
For a pair of distinct indices $i,j \in [2m]$, let $I(i,j)$
be the subset of indices $t' \in [t]$ for which $v_i$ and $v_j$
have the same value on their $t'$-th coordinates (thus implying that
$c_i$ and $c_j$ use the same color on $E_{t'}$). Then
$I(i,j)$ has the same distribution as a random subset of
$[t]$ obtained by taking each element independently with probability
$1/r$. Moreover,
\[ \mathcal{G}(c_i, c_j) = \mathcal{G}_{I(i,j)}. \]
Hence,
\[ \mathbb{P}[\chi(\mathcal{G}(c_i, c_j)) \ge r+1] = \mathbb{P}[\chi(\mathcal{G}_{I(i,j)}) \ge r+1] \le \frac{1}{2m}. \]
Therefore, the expected number of pairs $i,j$ with $i<j$ having
$\chi(\mathcal{G}_{I(i,j)}) \ge r+1$ is at most ${2m \choose 2} \frac{1}{2m} \le m$.
Hence, there exists a choice of coloring functions $c_i$ for which this number is
at most $m$. If this event happens, then we can remove one row
from each pair $i,j$ having $\chi(\mathcal{G}_{I(i,j)}) \ge r+1$
to obtain a set $R \subset [2m]$ of size
at least $m$ which has the property that $\chi(\mathcal{G}_{I(i,j)}) \le r$
for all $i,j \in R$. By considering the subgraph of $\Gamma_{2m,n}$
induced on $R \times [n]$ and using Lemma \ref{lem:row_chromatic}, we obtain
an alternating-rectangle-free edge coloring of $\Gamma_{m,n}$ with $r$ colors.
The result follows.
\end{proof}
We prove Theorems \ref{thm:grid_main} and \ref{thm:grid_asymmetric}
in the next two subsections.
We begin with Theorem \ref{thm:grid_asymmetric}, which establishes upper bounds for
$g(m,n)$ in various off-diagonal regimes. As noted in the introduction, parts (i) and (ii) already yield weak versions of
Theorem~\ref{thm:grid_main}. In particular, part (i) implies that $G(r)$ is superpolynomial in $r$,
while part (ii) yields the bound $G(r) > 2^{c \log^2 r}$. We recall the stronger off-diagonal statements below.
\subsection{Proof of Theorem \ref{thm:grid_asymmetric}}
\label{subsec:grid_asymmetric}
\noindent \textbf{Parts (i) and (ii)} : For all $C > e^2$, $\varepsilon > 0$
and large enough $r$, $g(r^{\log C/2}, r^{r/2C}) \le r$ and
$g(2^{\varepsilon \log^2 r}, 2^{r^{1-\varepsilon}}) \le r$.
\medskip
Let $n = 2^t$ for some $t$ to be chosen later. The edge coloring $c_B$
from Section \ref{sec:preliminaries} gives an edge partition
$E = E_1 \cup \dots \cup E_{t}$ of $K_n$ for $t = \log n$ such that,
for all $J \subset [t]$,
\[ \chi(\mathcal{G}_J) = 2^{|J|}. \]
Hence, if we let $I$ be a random subset of $[t]$ obtained by choosing each element
independently with probability $1/r$, then
\begin{align}
\mathbb{P}[\chi(\mathcal{G}_I) \ge r+1]
&= \mathbb{P}\big[|I| \ge \log(r+1)\big] \nonumber \\
&\le {t \choose \log(r+1)} \frac{1}{r^{\log(r+1)}}
\le \left( \frac{et}{r \log (r+1)} \right)^{\log (r+1)}. \label{eq:grid_result1}
\end{align}
For part (i), let $C$ be a given constant and take $t = r\log r /2C$.
Then the right-hand side of \eqref{eq:grid_result1}
is at most $(r+1)^{-\log(C/e)}$. In Lemma \ref{lem:grid_coloring},
we can take $m = \frac{1}{2}(r+1)^{\log (C/e)} \ge r^{\log C/2}$
and $n = 2^{t} = r^{r/2C}$ to get
\[ g(r^{\log C/2}, r^{r/2C}) \le r. \]
For part (ii), let $\varepsilon$ be a given constant
and take $t = r^{1-\varepsilon}$. For large enough $r$,
the right-hand side of \eqref{eq:grid_result1} is at
most $\frac{1}{2}r^{-\varepsilon \log r}$. Hence,
by applying Lemma \ref{lem:grid_coloring} with $m = r^{\varepsilon \log r} =
2^{\varepsilon \log^2 r}$ and $n = 2^t = 2^{r^{1-\varepsilon}}$, we see that
\[ g(2^{\varepsilon \log^2 r}, 2^{r^{1-\varepsilon}}) \le r. \]
\medskip
\noindent \textbf{Part (iii)} : There exists a positive constant $c$ such that $g(cr^2, r^{r^2 /2} / e^{r^2}) \le r$ for large enough $r$.
\medskip
Let $c=e^{-3}$. Let $n = cr^2$ and partition the edge set of $K_n$ into $t = {n \choose 2}$
sets $E_1, \ldots, E_t$, each of size exactly one. As before,
let $I$ be a random subset of $[t]$ obtained by choosing each element
independently with probability $1/r$.
In this case, we get $\mathcal{G}_I = \mathcal{G}(n, \frac{1}{r})$
(where $\mathcal{G}(n, p)$ is the binomial random graph).
Therefore,
\[ \mathbb{P}[\chi(\mathcal{G}_I) \ge r+1] = \mathbb{P}[\chi (\mathcal{G}(cr^2, 1/r)) \ge r+1]. \]
The event $\chi\left(\mathcal{G}(cr^2, \frac{1}{r})\right) \ge r+1$
is contained in the event that
$\mathcal{G}(cr^2, \frac{1}{r})$ contains a subgraph of order $s \ge r+1$
of minimum degree at least $r$.
The latter event has probability at most
\begin{align}
\label{eqn:g_n_p_chromatic}
\sum_{s=r+1}^{cr^2} {cr^2 \choose s} {s^2/2 \choose rs/2} \left(\frac{1}{r}\right)^{rs/2}
&\le
\sum_{s=r+1}^{cr^2} \left( \left( \frac{ecr^2}{s} \right)^{2} \left(\frac{es}{r} \right)^{r} \left(\frac{1}{r} \right)^{r} \right)^{s/2}.
\end{align}
For $s=r$, if $r$ is large enough, then the summand is
\[ \left( (ecr)^{2} e^{r} \left(\frac{1}{r} \right)^{r} \right)^{r/2}
\le \frac{e^{r^2}}{r^{r^2 /2}}\,. \]
We next show that the summands are each at most a quarter of the previous summand. As the series starts at $s=r+1$ and ends at $s=cr^2$, the series is then at most half the summand for $s=r$. The ratio of the summand for $s+1$ to the summand for $s$, where $r+1 \leq s+1 \leq cr^2$, is
$$
\left(\frac{s+1}{s}\right)^{s(r-2)/2} \left(\left(\frac{ecr^2}{s+1}\right)^2\left(\frac{e(s+1)}{r}\right)^r\left(\frac{1}{r}\right)^r\right)^{1/2}$$
which is at most
$$e^{(r-2)/2} \left(\frac{e^{r+2} c^2 (s+1)^{r-2}}{r^{2r-4}}\right)^{1/2} \leq e^{(r-2)/2} (e^{r+2} c^r )^{1/2} = e^{-r/2} < \frac{1}{4},
$$
for $r$ sufficiently large.
Hence, the right-hand side of \eqref{eqn:g_n_p_chromatic} is
at most $e^{r^2}r^{-r^2/2}/2$ and
\[ \mathbb{P}[\chi(\mathcal{G}_I) \ge r+1] \le \frac{e^{r^2}}{2r^{r^2 /2}}. \]
By Lemma \ref{lem:grid_coloring}, we conclude that
$g(cr^2, r^{r^2 / 2} / e^{r^2}) \le r$.
\subsection{Proof of Theorem \ref{thm:grid_main}}
\label{subsec:grid_main}
In the previous subsection we used quite simple edge partitions of the complete graph
as an input to Lemma \ref{lem:grid_coloring} to prove Theorem \ref{thm:grid_asymmetric}.
These partitions were already good enough to give the superpolynomial bound $G(r) > 2^{c \log^2 r}$.
To further improve this bound and prove Theorem \ref{thm:grid_main},
we make use of a slightly more sophisticated edge
partition guaranteed by the following theorem.
\begin{THM} \label{thm:chi_slow_grow_less_colors}
There exists a positive real $r_0$ such that the following holds
for positive integers $r$ and positive reals $\alpha \le 1$
satisfying $(\log r)^\alpha \ge r_0$.
For $n = 2^{(\log r)^{2 + \alpha}/200}$,
there exists a partition $E = E_1 \cup \dots \cup E_{\sqrt{r}}$ of
the edge set of the complete graph $K_n$ such that
\[ \chi(\mathcal{G}_I) \le 2^{3(\log r)^{\alpha/2}\sqrt{|I| \log 2|I|}} \]
for all $I \subset [\sqrt{r}]$.
\end{THM}
The proof of this theorem is based on
Theorem \ref{thm:chi_slow_grow}, which is in turn based on considering the coloring $c_M$, and will be given in Section \ref{sec:chi_eg}.
Now suppose that a positive integer $r$ is given and let
$\alpha \leq 1$ be a real to be chosen later. Let $E_1 \cup \dots \cup E_{\sqrt{r}}$
be the edge partition of $K_n$ for $n = 2^{(\log r)^{2 + \alpha}/200}$
given by Theorem \ref{thm:chi_slow_grow_less_colors}.
Let $I$ be a random subset of $[\sqrt{r}]$ chosen by taking each
element independently with probability $\frac{1}{r}$. Then,
by Theorem \ref{thm:chi_slow_grow_less_colors}, we have
\[ \chi(\mathcal{G}_I) \ge r+1 \quad \Rightarrow \quad |I| \ge c\frac{(\log r)^{2 - \alpha}}{\log \log r}, \]
for some positive constant $c$. Therefore,
\begin{align*}
\mathbb{P}[\chi(\mathcal{G}_I) \ge r+1 ]
&\le \mathbb{P}\left[|I| \ge c \frac{(\log r)^{2 - \alpha}}{\log \log r}\right] \\
&\le {\sqrt{r} \choose c(\log r)^{2 - \alpha}/\log \log r} \left(\frac{1}{r} \right)^{c(\log r)^{2 - \alpha}/ \log \log r}
\le r^{-c'(\log r)^{2 - \alpha}/\log \log r}
\end{align*}
holds for some positive constant $c'$.
By Lemma \ref{lem:grid_coloring}, for $m = 2^{c'(\log r)^{3-\alpha} / \log \log r - 1}$,
we have $g(m, n) \le r$. We may choose $\alpha$ so that
\[ m = n = e^{\Omega((\log r)^{5/2}/ \sqrt{\log \log r})}. \]
This gives $G(r) \ge 2^{\Omega((\log r)^{5/2}/ \sqrt{\log \log r})}$, as required.
\section{The Erd\H{o}s--Gy\'arf\'as problem}
\label{sec:eg}
In the introduction, we discussed how the grid Ramsey problem is connected to a
hypergraph version of the Erd\H{o}s--Gy\'arf\'as problem. We now establish this
correspondence more formally.
\begin{PROP} \label{prop:relation_g_f}
For all positive integers $n$, we have
\[ g(n) \le f_3(2n, 4,3) \le 2 \lceil \log n \rceil ^2 g(n). \]
\end{PROP}
\begin{proof}
Since $K^{(3)}(n,n)$ is a subhypergraph of $K_{2n}^{(3)}$,
a $(4,3)$-coloring of $K_{2n}^{(3)}$ immediately gives
a coloring of $K^{(3)}(n,n)$ such that every copy of $K_{4}^{(3)}$ receives at least three colors.
Hence, by the correspondence between coloring functions for $K^{(3)}(n,n)$ and
coloring functions for $\Gamma_{n,n}$ explained in the introduction, it follows that
$g(n) \le f_3(2n,4,3)$.
We prove the other inequality by showing that for all $m \le n$,
\begin{align}
\label{eq:recursive}
f_3(2m, 4, 3) \le f_3(m, 4, 3) + 2\lceil \log m \rceil g(m).
\end{align}
By repeatedly applying this recursive formula, we obtain the
claimed inequality
\begin{align*}
f_3(2n, 4, 3) \le 2\lceil \log n \rceil^2 g(n).
\end{align*}
Thus it suffices to establish the recursive formula \eqref{eq:recursive}.
We will do this by presenting a $(4,3)$-coloring
of $K^{(3)}_{2m}$. Let $A$ and $B$ be two disjoint vertex subsets of
$K^{(3)}_{2m}$, each of order $m$. Given a
$(4,3)$-coloring of $K_{m}^{(3)}$ with $f_3(m,4,3)$ colors,
color the hyperedges within $A$ using this coloring
and also the hyperedges within $B$ using this coloring. Since
we started with a $(4,3)$-coloring, every
copy of $K_4^{(3)}$ lying inside $A$ or $B$ contains
at least $3$ colors on its edges. This
leaves us with the copies which intersect both
$A$ and $B$.
Let $H$ be the bipartite hypergraph that consists of the edges which
intersect both parts $A$ and $B$.
By definition, we have an alternating-free edge coloring
of the grid graph $\Gamma_{m,m}$ using $g(m)$ colors.
We may assume, by introducing at most $g(m)$ new colors, that the
set of colors used for the row edges and the column edges are disjoint.
This gives an edge coloring of $\Gamma_{m,m}$, where each rectangle
receives at least three colors.
Let $c_1$ be a coloring of $H$ using at most $2g(m)$ colors,
where for an edge $\{i,j,j'\} \in H$ with $i \in A, j,j' \in B$,
we color it with the color of the edge $\{(i,j), (i,j')\}$ in $\Gamma_{m,m}$
and for an edge $\{i,i',j\} \in H$ with $i,i' \in A, j \in B$,
we color it with the color of the edge $\{(i,j), (i',j)\}$ in $\Gamma_{m,m}$.
Let $c_2$ be a coloring of $H$ constructed based on the
coloring $c_B$ given in Section \ref{sec:preliminaries} as follows:
for an edge $\{i,j,j'\} \in H$ with $i \in A, j,j' \in B$,
let $c_2(\{i,j,j'\}) = c_B(\{j,j'\})$ and for an edge
$\{i,i',j\} \in H$ with $i,i' \in A, j \in B$,
let $c_2(\{i,i',j\}) = c_B(\{i,i'\})$. Now color the hypergraph
$H$ using the coloring function $c_1 \times c_2$.
Consider a copy $K$ of $K_4^{(3)}$ which intersects both parts $A$ and $B$.
If $|K \cap A| = |K \cap B| = 2$, then assume that
$K = \{i,i',j,j'\}$ for $i,i' \in A$ and $j,j' \in B$. One can see
that the set of colors used by $c_1$ on $K$ is identical to the
set of colors used on the rectangle $(i,j,i',j')$ in $\Gamma_{m,m}$
considered above. Thus $K$ receives at least three distinct colors.
If $|K \cap A| = 1$ and $|K \cap B| = 3$, then the three hyperedges
in $K$ which intersect $A$ use at least two colors from the coloring $c_2$,
while the unique hyperedge of $K \cap B$ is colored with a different color.
Hence $K$ contains at least three colors. Similarly, $K$ contains
at least three colors if $|K \cap A| = 3$ and $|K \cap B| = 1$.
Since $c_1$ uses at most $2g(m)$ colors and $c_2$ uses at most $\lceil \log m \rceil$ colors, we see that
$c_1 \times c_2$ uses at most $2 \lceil \log m \rceil g(m)$ colors.
Recall that we used at most $f_3(m,4,3)$ colors to color
the edges inside $A$ and $B$. Therefore, we have found a $(4,3)$-coloring
of $K^{(3)}_{2m}$ using at most
\[ f_3(m,4,3) + 2 \lceil \log m \rceil g(m) \]
colors, thereby establishing \eqref{eq:recursive}.
\end{proof}
\subsection{A basic bound on $F_k(r,p,q)$}
Here we prove Theorem \ref{thm:step_down} that provides a basic upper bound on the function $F_k(r,p,q)$. Recall that we are
given positive integers $r, k, p$, and $q$ all greater than $1$ and
satisfying $r \ge k$.
Let $N = r^{{F_{k-1}(r,p-1,q) \choose k-1}}$ and suppose that we are given
an edge coloring of $K_N^{(k)}$ with $r$ colors (denoted by $c$). Let $[N]$ be the vertex set of $K_N^{(k)}$.
For each integer $t$ in the range $1 \le t \le F_{k-1}(r,p-1,q)$,
we will inductively find a pair of disjoint subsets $X_t$ and $Y_t$ of $[N]$
with the following properties:
\begin{enumerate}
\item $|X_t| = t$ and $|Y_t| \ge \min\{N / r^{{t \choose k-1}}, N-t\}$,
\item for all $x \in X_t$ and $y \in Y_t$, $x < y$,
\item for all edges $e \in {X_t \cup Y_t \choose k}$ satisfying $|e \cap X_t| \ge k-1$,
the color of $e$ is determined by the first $k-1$ elements of $e$ (note that the
first $k-1$ elements necessarily belong to $X_t$).
\end{enumerate}
For the base cases $t=1, \ldots, k-2$, the pair of sets $X_{t} = \{1,2,\ldots,t\}$
and $Y_{t} = [N] \setminus X_t$ trivially satisfy the given properties.
Now suppose that for some $t \ge k-2$, we are given pairs $X_{t}$ and $Y_{t}$ and
wish to construct sets $X_{t+1}$ and $Y_{t+1}$.
Since $t < F_{k-1}(r,p-1,q)$, Property 1 implies that $|Y_t| \ge 1$ and in
particular that $Y_t$ is nonempty.
Let $x$ be the minimum element of $Y_{t}$ and let $X_{t+1} = X_{t} \cup \{x\}$.
For each element $y \in Y_{t} \setminus \{x\}$, consider the
vector of colors of length ${|X_{t}| \choose k-2}$ whose coordinates
are $c(e' \cup \{x, y\})$ for each $e' \in {X_{t} \choose k-2}$.
By the pigeonhole principle, there are at least
$\frac{|Y_t| - 1}{ r^{{ |X_t| \choose k-2 }} }$ vertices which
have the same vector. Let $Y_{t+1}$ be these vertices.
This choice immediately implies Properties 2 and 3 above.
To check Property 1, note that
\[
|Y_{t+1}|
\ge \frac{|Y_t| - 1}{ r^{{ |X_t| \choose k-2 }} }
\ge \frac{N / r^{{t \choose k-1}} - t - 1}{ r^{{ |X_t| \choose k-2 }} }
= \frac{N}{ r^{{ t+1 \choose k-1 }} } - \frac{t+1}{r^{{ t \choose k-2 }}}
> \frac{N}{ r^{{ t+1 \choose k-1 }} } - 1,
\]
where the final inequality follows from $t \ge k-2$ and $r \ge k$.
Since $N = r^{{F_{k-1}(r,p-1,q) \choose k-1}}$, $F_{k-1}(r,p-1,q) \ge t+1$
and $|Y_{t+1}|$ is an
integer, this implies that $|Y_{t+1}| \ge \frac{N}{ r^{{ t+1 \choose k-1 }} }$.
Let $T = F_{k-1}(r,p-1,q)$ and note that $|X_T| = F_{k-1}(r,p-1,q)$
and $|Y_T| \ge 1$. Construct an auxiliary complete $(k-1)$-uniform
hypergraph over the vertex set $X_T$ and color each edge
with the color guaranteed by Property 3 above. This gives an edge coloring
of $K^{(k-1)}_T$ with $r$ colors and thus, by definition, we can find a
set $A$ of $p-1$ vertices using fewer than $q$ colors
on its edges in the auxiliary $(k-1)$-uniform hypergraph.
It follows from Property 3 that for an arbitrary $y \in Y_T$,
$A \cup \{y\}$ is a set of $p$ vertices using fewer than $q$ colors
on its edges in the original $k$-uniform hypergraph.
\subsection{A superpolynomial lower bound for $F_3(r, 5, 6)$}
In this subsection, we present a $(5,6)$-coloring of
$K_n^{(3)}$ using $2^{O(\sqrt{\log n})}$ colors.
This shows that $f_3(n,5,6) = 2^{O(\sqrt{\log n})}$
and $F_3(r,5,6) = 2^{\Omega(\log^2 r)}$.
The edge coloring is given as a product $c = c_1 \times c_2 \times c_3 \times c_4$ of
four coloring functions $c_1,c_2,c_3,c_4$. The first coloring $c_1$ is a
$(4,3)$-coloring of $K_n^{(3)}$ using $f_3(n,4,3)$ colors. Combining
Proposition \ref{prop:relation_g_f} and Theorem \ref{thm:grid_main}, we see that
$f_3(n,4,3) = 2^{O((\log n)^{2/5} (\log \log n)^{1/5})}$.
Let $n=2^d$ and write the vertices of $K_n$ as binary strings of length $d$.
To define $c_2, c_3$ and $c_4$,
for three distinct vertices $u,v,w$, assume that the least coordinate
in which not all vertices have the same bit is the $i$-th coordinate
and let $u_i, v_i, w_i$ be the $i$-th coordinate of $u,v,w$, respectively.
Without loss of generality, we may assume that $u_i = v_i \neq w_i$,
i.e. $(u_i, v_i, w_i) = (0,0,1)$ or $(1,1,0)$.
Define the second color $c_2$ of the triple of vertices $\{u,v,w\}$ as $i$.
Thus $c_2$ uses at most $\log n$ colors.
Define the third color $c_3$ as the value of $w_i$, which is either 0 or 1.
Define the fourth color $c_4$ as $c_M(u, v)$, where
$c_M$ is the graph coloring
given in Section \ref{sec:preliminaries}, which is both
a $(3,2)$ and $(4,3)$-coloring. Recall that
$c_M$ uses at most $2^{O(\sqrt{\log n})}$ colors.
The number of colors in the coloring $c$ is
\[ 2^{O((\log n)^{2/5} (\log \log n)^{1/5})} \cdot \log n
\cdot 2 \cdot 2^{O(\sqrt{\log n})} = 2^{O(\sqrt{\log n})}, \]
as desired. Now we show that each set of $5$ vertices receives at
least $6$ colors in the coloring $c$. Let $i$ be the least
coordinate such that the five vertices do not all agree.
\medskip
\noindent \textbf{Case 1}: One of the vertices (call it $v_1$) has one bit at coordinate $i$, while
the other four vertices (call them $v_2, v_3, v_4, v_5$) have the other bit.
The $6$ triples containing $v_1$ are different colors from the other $4$
triples.
Indeed, the triples containing $v_1$ have $c_2
= i$, while the other triples have $c_2$ greater than $i$.
Since $c_M$ is a $(4,3)$-coloring of graphs, $c_4$ tells us that the triples containing $v_1$ have
to use at least $3$ colors. On the other hand,
by the coloring $c_1$, the triples in the 4-set $\{v_2,v_3,v_4,v_5\}$
have to use at least 3 colors. Hence, at least 6 colors have to
be used on the set of five vertices.
\medskip
\noindent \textbf{Case 2}: Two of the vertices
(call them $v_1,v_2$) have one bit at coordinate $i$, while the other three vertices
(call them $v_3, v_4, v_5$) have the other
bit. Let $V_0=\{v_1,v_2\}$ and $V_1=\{v_3,v_4,v_5\}$.
Let $A$ be the set of colors of triples in $\{v_1,...,v_5\}$. We partition $A$
into $A_0, A_1, A_2$ as follows. For each $j \in \{0,1,2\}$,
let $A_j$ be the set consisting of the colors of triples
containing exactly $j$ vertices from $V_0$.
It follows from the colorings $c_2$ and $c_3$ that the three color sets $A_0,
A_1, A_2$ form a partition of $A$. Indeed, the color in $A_0$ has second
coordinate $c_2$ greater than $i$, while the colors in $A_1$ and $A_2$ have
second coordinate $c_2 = i$.
Furthermore, the colors in $A_1$ have third coordinate $c_3$ distinct from the
third coordinate $c_3$ of the colors in $A_2$. Note also that $|A_0| = 1$.
\medskip
\noindent \textbf{Case 2a}: $|A_2|=3$.
Since the coloring
$c_M$ is a $(3,2)$-coloring of graphs, $c_4$ implies that
the triples containing $v_1$ whose other two
vertices are in $V_1$ receive at least $2$ colors. This implies that $|A_1| \geq 2$ and,
therefore, the number of colors used is at least $|A_0|+|A_1|+|A_2| \geq 6$.
\medskip
\noindent \textbf{Case 2b}: $|A_2|=2$.
Suppose without loss of generality that $(v_1,v_2,v_3)$ and $(v_1,v_2,v_4)$
have the same color, which is different from the color of $(v_1,v_2,v_5)$.
As each $K_4^{(3)}$ uses at least $3$ colors in coloring $c_1$, $(v_1,v_3,v_4)$
and $(v_2,v_3,v_4)$ have different colors. Note that
$c_4(v_1, v_3, v_4) = c_4(v_2, v_3, v_4) = c_M(v_3, v_4)$. Since $c_M$
is a $(3,2)$-coloring of graphs, at least one of $c_M(v_3, v_5)$ or
$c_M(v_4, v_5)$ is different from $c_M(v_3, v_4)$. Suppose, without
loss of generality, that $c_M(v_3, v_5) \neq c_M(v_3, v_4)$. Since
$c$ is defined as the product of $c_1, \ldots, c_4$,
we see that the color of $(v_1, v_3, v_5)$ is different from both that of
$(v_1, v_3, v_4)$ and $(v_2, v_3, v_4)$. Thus $|A_1| \geq 3$.
Then the number of colors used is at least $|A_0|+|A_1|+|A_2| \geq 6$.
\medskip
\noindent \textbf{Case 2c}: $|A_2|=1$.
This implies that the three edges $(v_1,v_2,v_j)$ for $j=3,4,5$ are of the same color.
First note that as in the previous case, there are at least two
different colors among $c_M(v_3, v_4)$, $c_M(v_3, v_5)$ and $c_M(v_4, v_5)$.
Without loss of generality, suppose that $c_M(v_3, v_4) \neq c_M(v_3, v_5)$.
Since $c$ is defined as the product of $c_1, \ldots, c_4$,
this implies that the set $A_1' = \{c(v_1, v_3, v_4), c(v_2, v_3, v_4)\}$ is
disjoint from the set $A_1'' = \{c(v_1, v_3, v_5), c(v_2, v_3, v_5)\}$.
Now, by considering the coloring $c_1$, since all
three edges $(v_1,v_2,v_j)$ for $j=3,4,5$ are of the same color,
we see that $|A_1'| = 2$ and $|A_1''| = 2$. Hence
$|A_1| \ge |A_1'| + |A_1''| = 4$.
Then the number of colors used is at least $|A_0|+|A_1|+|A_2| \geq 6$.
\section{A chromatic number version of the Erd\H{o}s--Gy\'arf\'as problem}
\label{sec:chi_eg}
\subsection{Bounds on $F_{\chi}(r,4,3)$}
In this subsection, we prove Theorem \ref{thm:chi_4_3}. This asserts that
\[ 2^{\log^2 r/36} \le F_{\chi}(r,4,3) \le C \cdot 2^{130 \sqrt{r \log r}}. \]
In order to obtain the upper bound, we use the concept of dense pairs.
Suppose that a graph $G$ is given. For positive
reals $\varepsilon$ and $d$, a pair of disjoint vertex subsets
$(V_1, V_2)$ is \emph{($\varepsilon, d$)-dense} if
for every pair of subsets $U_1 \subseteq V_1$ and $U_2 \subseteq V_2$
satisfying $|U_1| \ge \varepsilon |V_1|$ and
$|U_2| \ge \varepsilon |V_2|$, we have
\[ e(U_1, U_2) \ge d |U_1| |U_2|,
\]
where $e(U_1, U_2)$ is the number of edges of $G$ with one
endpoint in $U_1$ and the other in $U_2$.
The following result is due to Peng, R\"odl and Ruci\'nski \cite{PRR}.
Recall that the edge density of a graph $G$ with $m$ edges and $n$ vertices is
$m/\binom{n}{2}$.
\begin{THM} \label{thm:regularity}
For all positive reals $d$ and $\varepsilon$,
every graph on $n$ vertices of edge density at least $d$ contains an
$(\varepsilon, d/2)$-dense pair $(V_1, V_2)$ for which
\[ |V_1| = |V_2| \ge \frac{1}{8}n d^{12/\varepsilon}. \]
\end{THM}
The original theorem of Peng, R\"odl and Ruci\'nski takes a bipartite graph
with $n$ vertices in each part and $dn^2$ edges as input and outputs an
$(\varepsilon,d/2)$-dense pair with
parts of size at least $\frac{1}{2}n d^{12/\varepsilon}$.
The theorem as stated above is an immediate corollary since every $n$-vertex
graph of density $d$ contains a bipartite subgraph with $m = \lfloor \frac{n}{2} \rfloor \ge \frac{n}{4}$ vertices in each part and at least $d m^2$ edges.
\begin{proofof}{upper bound in Theorem \ref{thm:chi_4_3}}
Let $n = F_{\chi}(r,4,3) - 1$ and
suppose that a chromatic-$(4,3)$-coloring of $K_n$ using $r$ colors is given.
Take a densest color, say red, and consider the graph $\mathcal{G}$
induced by the red edges. This graph has density at least $\frac{1}{r}$.
By applying Theorem \ref{thm:regularity} with
$\varepsilon = \left(\frac{\ln r}{r}\right)^{1/2}$,
we obtain an $(\varepsilon, \frac{1}{2r})$-dense pair $(V_1, V_2)$ in $\mathcal{G}$
such that
\[ |V_1| = |V_2| \ge \frac{1}{8}n\left(\frac{1}{r}\right)^{12/\varepsilon}
\ge n e^{-13 \sqrt{r \ln r}}. \]
For a color $c$ which is not red, let $\mathcal{G}_{+c}$ be the graph obtained
by adding all edges of color $c$ to the graph $\mathcal{G}$.
Since the given coloring is a chromatic-$(4,3)$-coloring, we see that
$\mathcal{G}_{+c}$ is $3$-colorable for all $c$.
Consider an arbitrary proper $3$-coloring of $\mathcal{G}_{+c}$.
If there exists a color class in this proper coloring
which intersects both $V_1$ and $V_2$ in
at least $\varepsilon|V_1|$ vertices, then, since $(V_1, V_2)$ is an
$(\varepsilon, \frac{1}{2r})$-dense pair, there exists an edge
between the two intersections, thereby
contradicting the fact that the $3$-coloring is proper.
Hence, $\mathcal{G}_{+c}$ has an independent set $I_c$ of size at least
$(1-2\varepsilon)|V_1|$ in either $V_1$ or $V_2$.
For $i=1,2$, define $C_i$ to be the set of colors $c \in [r]$ for which
this independent set $I_c$ is in $V_i$.
Since $|C_1| + |C_2| \ge r-1$,
we may assume, without loss of generality, that $|C_1| \ge \frac{r-1}{2}$.
For each $v \in V_1$, let $d(v)$ be the number of colors
$c \in C_1$ for which $v \in I_c$. Note that
\begin{align}
\label{eq:degreesum}
\sum_{v \in V_1} d(v) = \sum_{c \in C_1} |I_c|
\ge |C_1| \cdot (1-2\varepsilon)|V_1|.
\end{align}
For each $X \subset C_1$ of size $\frac{r}{4}$, let
$I_X = \bigcap_{c \in X} I_c$. We have
\begin{align*}
\sum_{\substack{X \subset C_1 \\ |X| = r/4}} |I_X|
= \sum_{v \in V_1} {d(v) \choose r/4}
\ge |V_1| \cdot {|C_1| \cdot (1-2\varepsilon) \choose r/4},
\end{align*}
where the inequality follows from \eqref{eq:degreesum} and
convexity. Since $|C_1| \ge \frac{r-1}{2}$, we have
\[
\sum_{\substack{X \subset C_1 \\ |X| = r/4}} |I_X|
\ge (1-8\varepsilon)^{r/4} |V_1| \cdot {|C_1| \choose r/4}.
\]
Thus we can find a set $X \subset C_1$ for which
$|I_X| \ge (1-8\varepsilon)^{r/4}|V_1|$. By
definition, the set $I_X$ does not contain any color from $X$
and hence the original coloring induces a chromatic-$(4,3)$-coloring
of a complete graph on $|I_X|$ vertices using at most $3r/4$ colors.
This gives
\[
F_{\chi}\left(\frac{3r}{4}, 4, 3\right) - 1
\ge |I_X|
\ge (1-8\varepsilon)^{r/4} |V_1|.
\]
For $\varepsilon \le 1/16$, the inequality $1-8\varepsilon \ge e^{-16\varepsilon}$ holds.
Hence, for large enough $r$, the right-hand side above is at least
\[
e^{-4 \varepsilon r} \cdot n e^{-13\sqrt{r \ln r}}
=
e^{-4\sqrt{r\ln r}} \cdot n e^{-13\sqrt{r \ln r}}
= (F_{\chi}(r, 4, 3) - 1) e^{-17 \sqrt{r \ln r}}.
\]
We conclude that there exists $r_0$ such that if $r \ge r_0$, then
\[
F_{\chi}(r,4,3) \le e^{17 \sqrt{r \ln r}} F_{\chi}\left(\frac{3r}{4}, 4, 3\right).
\]
We now prove by induction that there is a constant $C$ such that $F_{\chi}(r,4,3) \leq C e^{130 \sqrt{r \ln r}}$ holds for all $r$. This clearly holds for the base cases $r<r_0$, so suppose $r \geq r_0$. Using the above inequality and the induction hypothesis, we obtain
\[
F_{\chi}(r,4,3) \le e^{17 \sqrt{r \ln r}} F_{\chi}\left(\frac{3r}{4}, 4, 3\right) \le
e^{17 \sqrt{r \ln r}} Ce^{130 \sqrt{(3r/4)\ln (3r/4)}} \le Ce^{130\sqrt{r\ln r}},\]
which completes the proof.
\end{proofof}
We now turn to the proof of the lower bound.
In order to establish the lower bound,
we show that Mubayi's coloring $c_M$ is in fact a chromatic-$(4,3)$-coloring.
This then implies that $F_{\chi}(r,4,3) \ge 2^{\log^2 r/ 36}$, as claimed.
Recall that in the coloring $c_M$, we view the vertex set of $K_n$
as a subset of $[m]^t$ for some integers $m$ and $t$ and, for two vertices
$x, y \in [m]^t$ of the form $x = (x_1, \ldots, x_t)$ and
$y = (y_1, \ldots, y_t)$, we let
\[ c_M(x,y) = \Big(\{x_i, y_i\}, a_1, \ldots, a_t\Big), \]
where $i$ is the minimum index for which $x_i \neq y_i$ and
$a_j = \delta(x_j, y_j)$ is the Dirac delta function.
\begin{proofof}{lower bound in Theorem \ref{thm:chi_4_3}}
Consider the coloring $c_M$ on the vertex set $[m]^t$.
Suppose that two colors $c_{1}$ and $c_{2}$ are given and let
\[ c_1 = \Big(\{x_{1}, y_{1}\}, a_{1,1}, \ldots, a_{1,t} \Big)
\quad \textrm{and} \quad
c_2 = \Big(\{x_{2}, y_{2}\}, a_{2,1}, \ldots, a_{2,t} \Big). \]
Suppose that $a_{1, i_1}$ is the first non-zero $a_{1,j}$ term
and $a_{2, i_2}$ is the first non-zero $a_{2,j}$ term. In other
words, for a pair of vertices which are colored by $c_1$, the
first coordinate in which the pair differ is the $i_1$-th coordinate
(and a similar claim holds for $c_2$).
Let $\mathcal{G}$ be the graph induced by
the edges which are colored by either $c_1$ or $c_2$.
We will prove that $\chi(\mathcal{G}) \le 3$ by presenting
a proper vertex coloring of $\mathcal{G}$ using three colors, red, blue
and green.
\medskip
\noindent \textbf{Case 1}: $i_{1}=i_{2}=i$ for some index $i$.
First, color all the vertices whose $i$-th coordinate is equal
to $x_{1}$ in red. Second, color all the vertices
whose $i$-th coordinate is equal to $x_{2}$ in blue
(if $x_1 = x_2$, there are no vertices of color blue).
Third, color all other vertices in green.
To show that this is a proper coloring, note that
if the color between two vertices $z, w \in [m]^t$ is either
$c_1$ or $c_2$, then the $i$-th coordinate of
$z$ and $w$ must be different.
This shows that the set of red vertices and
the set of blue vertices are both independent sets. It
remains to show that the set of green vertices is an
independent set. To see this, note that if the color between $z$ and
$w$ is either $c_1$ or $c_2$, then the $i$-th coordinates
$z_i$ and $w_i$ must satisfy
\[ \{z_i, w_i\} = \{x_{1}, y_{1}\} \quad \textrm{or} \quad \{x_{2}, y_{2}\}, \]
as this is the only way the first coordinate of $c_M(z, w)$
can match that of $c_1$ or $c_2$. However, all vertices
which have $i$-th coordinate $x_{1}$ or $x_{2}$ are
excluded from the set of green vertices. This shows that our
coloring is proper.
\medskip
\noindent \textbf{Case 2}: $i_{1}\neq i_{2}$.
Without loss of generality, we may assume that $i_1 < i_2$.
We will find a proper coloring by considering only
the $i_1$-th and $i_2$-th coordinates.
For $v \in [m]^t$ of the form $v = (v_1, v_2, \ldots, v_t)$,
let
\[
\pi_{i_1}(v) =
\begin{cases}
0 & \text{if } v_{i_1} = x_1 \\
1 & \text{if } v_{i_1} = y_1 \\
* & \text{otherwise}\\
\end{cases}
\qquad \text{and} \qquad
\pi_{i_2}(v) =
\begin{cases}
0 & \text{if } v_{i_2} = x_2 \\
1 & \text{if } v_{i_2} = y_2 \\
* & \text{otherwise}\\
\end{cases}.
\]
Consider the projection map
\[ \pi \,:\, [m]^t \rightarrow \{0, 1, *\} \times \{0, 1, *\} \]
defined by $\pi(v) = (\pi_{i_1}(v), \pi_{i_2}(v))$ and let
$\mathcal{H} = \pi(\mathcal{G})$ be the graph on $\{0, 1, *\} \times \{0, 1, *\}$
induced by the graph $\mathcal{G}$ and the map $\pi$. More precisely,
a pair of vertices $v, w \in \{0, 1, *\} \times \{0, 1, *\}$
forms an edge if and only if there exists an edge of $\mathcal{G}$
between the two sets $\pi^{-1}(v)$ and $\pi^{-1}(w)$
(see Figure \ref{fig:proj_chi_4_3}). Note that a proper coloring
of $\mathcal{H}$ can be pulled back via $\pi^{-1}$ to give a
proper coloring of $\mathcal{G}$. It therefore suffices to find a proper
$3$-coloring of $\mathcal{H}$.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.75]{coloring_43.eps}
\caption{Graph $\mathcal{H} = \pi(\mathcal{G})$ when $a_{1,i_2} = 1$.}
\label{fig:proj_chi_4_3}
\end{figure}
Consider two vertices $z,w \in [m]^t$. If $c_M(z,w) = c_2$, then
the first coordinate in which $z$ and $w$ differ is the
$i_2$-th coordinate. This implies that $z$ and $w$ have
identical $i_1$-th coordinate. Hence, the set of
possible edges of the form $\{\pi(z), \pi(w)\}$ is
$E_2 = \left\{ \{00, 01\}, \{10, 11\}, \{*0, *1\} \right\}$.
Now suppose that $c_M(z,w) = c_1$. Then the possible edges
of the form $\{\pi(z), \pi(w)\}$ differ according to the
value of $a_{1, i_2}$.
\noindent \textbf{Case 2a}: $a_{1, i_2} = 0$.
In this case, the $i_2$-th coordinate of $z$ and $w$ must be
the same and thus the possible edges of the form $(\pi(z), \pi(w))$
are $E_1 = \left\{ \{00, 10\}, \{01, 11\}, \{0*, 1*\} \right\}$.
One can easily check that the graph with edge set $E_1 \cup E_2$
is bipartite.
\noindent \textbf{Case 2b}: $a_{1, i_2} = 1$.
In this case, the $i_2$-th coordinate of $z$ and $w$ must be
different and thus the possible edges of the form $(\pi(z), \pi(w))$
are $E_1 = \left\{ \{00, 11\}, \{00, 1*\}, \{01, 10\}, \{01, 1*\}
, \{0*, 10\}, \{0*, 11\}, \{0*, 1*\} \right\}$. A $3$-coloring
of the graph with edge set $E_1 \cup E_2$ is given by coloring the set of vertices
$\{00, 10, *0\}$ in red, $\{01, 0*\}$ in blue and $\{11, 1*, *1, **\}$ in green
(see Figure \ref{fig:proj_chi_4_3}).
\end{proofof}
\subsection{An edge partition with slowly growing chromatic number}
In this section, we will prove Theorem \ref{thm:chi_slow_grow}
by showing that $c_M$ has the required property.
\medskip
\noindent {\bf Theorem \ref{thm:chi_slow_grow}}.
The coloring $c_M$ has the following property: for every
subset $X$ of colors with $|X| \geq 2$, the subgraph induced by the edges colored
with a color from $X$ has chromatic number at most $2^{3 \sqrt{|X| \log |X|}}$.
\medskip
\begin{proof}
Consider the coloring $c_M$ on the vertex set $[m]^t$.
For a set of colors $X$, let $\mathcal{G}_X$ be the graph
induced by the edges colored by a color from the set $X$.
Recall that each color $c$ under this coloring is of the form
\begin{align*}
c &= \Big(\{v_i, w_i\}, a_1, a_2, \ldots, a_t \Big).
\end{align*}
Define $\iota(c) = i$ to be the minimum index $i$ for which
$a_i = 1$. For this index $i$, define $\eta_1(c) = v_i$ and $\eta_2(c) = w_i$,
where we break symmetry by imposing $v_i < w_i$. Let
$a_j(c) = a_j$ for $j = 1,\ldots, t$.
Given a set of colors $X$, construct an auxiliary graph $\mathcal{H}$ over the vertex set $X$ whose
edges are defined as follows. For two colors $c_1, c_2 \in X$,
let $i_1 = \iota(c_1)$, $i_2 = \iota(c_2)$ and assume
that $i_1 \le i_2$. Then $c_1$ and $c_2$ are adjacent
if and only if $a_{i_2}(c_1) = 1$ (it is well-defined
since if $i_1 = i_2$, then $a_{i_2}(c_1) = a_{i_1}(c_2) = 1$).
Let $\mathcal{I}$ be the family of all independent sets in $\mathcal{H}$.
We make the following claim, whose proof will be given later.
\begin{CLAIM} \label{claim:indep_sets}
The following holds: \\
(i) For all $I \in \mathcal{I}$, the graph $\mathcal{G}_I$ is bipartite. \\
(ii) $\chi(\mathcal{G}_X) \le |\mathcal{I}|$.
\end{CLAIM}
Suppose that the claim is true. Based on this claim, we will prove by induction on $|X|$ that
$\chi(\mathcal{G}_X) \le 2^{3 \sqrt{|X| \log |X|}}$.
For $|X|=2$, we proved in the previous subsection that $c_M$ is a chromatic-$(4,3)$-coloring, that is,
the union of any two color classes is $3$-colorable. This clearly implies the required result in this case.
Now suppose that the statement has been established for all
sets of size less than $|X|$.
Let $\alpha = \left\lceil \sqrt{\frac{|X|}{\log |X|}} \right\rceil$. If there
exists an independent set $I \in \mathcal{I}$ of
size at least $\alpha$, then, by the fact that
$\mathcal{G}_X = \mathcal{G}_I \cup \mathcal{G}_{X \setminus I}$
and Claim \ref{claim:indep_sets} (i),
we have
\[ \chi(\mathcal{G}_X) \le \chi(\mathcal{G}_I) \cdot \chi(\mathcal{G}_{X \setminus I})
\le 2 \chi(\mathcal{G}_{X \setminus I}).
\]
If $|X \setminus I| \ge 2$, then the right hand side is at most $2 \cdot 2^{3\sqrt{|X\setminus I| \log |X\setminus I|}} < 2^{3\sqrt{|X| \log |X|}}$ (the inequality comes from $|I| \ge \alpha$) by the inductive hypothesis, and if $|X \setminus I| \le 1$, then since $\chi(\mathcal{G}_{X \setminus I}) \leq 2$, the right hand side is at most $4$. Hence the claimed bound holds in both cases.
On the other hand, if the
independence number is less than $\alpha$, then,
by Claim \ref{claim:indep_sets} (ii) and the fact that $|X| \ge 2$, we have
\[ \chi(\mathcal{G}_X) \le \sum_{i=0}^{\alpha - 1} {|X| \choose i}
\le |X|^{2\sqrt{|X|/\log |X|}}
= 2^{2\sqrt{|X|\log |X|}}. \]
This proves the theorem up to Claim \ref{claim:indep_sets}, which we now consider.
\end{proof}
\begin{proofof}{Claim \ref{claim:indep_sets}}
(i) Suppose that $I \in \mathcal{I}$ is given. By definition, for each color $c \in I$,
we have distinct values of $\iota(c)$.
For each $c \in I$, consider the map $\pi_c : [m]^t \rightarrow \{0, 1\}$,
where for $x \in [m]^t$ of the form $x = (x_1, x_2, \ldots, x_t)$, we define
\[
\pi_c(x) =
\begin{cases}
0 & \text{if } x_{\iota(c)} \le \eta_1(c) \\
1 & \text{if } x_{\iota(c)} > \eta_1(c).
\end{cases}
\]
Define the map $\pi : [m]^t \rightarrow \{0, 1\}^{I}$ as
\[ \pi(x) = (\pi_c(x))_{c \in I}. \]
Consider the graph $\pi(\mathcal{G}_I)$ over the vertex set $\{0, 1\}^{I}$.
Let $c$ and $c'$ be two distinct colors in $I$.
If $\iota(c') < \iota(c)$, then
$a_{\iota(c')}(c) = 0$ since $\iota(c)$ is the minimum index $i$
for which $a_i(c) = 1$ and if $\iota(c') > \iota(c)$, then
$a_{\iota(c')}(c) = 0$ since $I$ is an independent set
in the auxiliary graph $\mathcal{H}$ defined above.
Hence, if $e = \{y,z\}$ is an edge of color $c$, then the
two vectors $y$ and $z$ have identical $\iota(c')$-coordinate for all $c' \neq c$, thus implying that $\pi(y)$ and $\pi(z)$ have identical $c'$-coordinate for all $c' \neq c$.
Further note that for $x \in \{0,1\}^I$, we have $\pi_c(x) = 0$ if the $\iota(c)$-th coordinate of $x$ is $\eta_1(c)$ and
$\pi_c(x)=1$ if the $\iota(c)$-th coordinate of $x$ is $\eta_2(c)$.
Since $\{y_{\iota(c)}, z_{\iota(c)}\} = \{\eta_1(c), \eta_2(c)\}$, we see that $\pi_c(y) \neq \pi_c(z)$.
Therefore, two vertices $v,w \in \{0, 1\}^{I}$ can be adjacent in $\pi(\mathcal{G}_I)$
if and only if $v$ and $w$ differ in exactly one coordinate, implying
that $\pi(\mathcal{G}_I)$ is a subgraph of the hypercube, which is clearly bipartite.
A bipartite coloring of this graph can be pulled back to give
a bipartite coloring of $\mathcal{G}_I$.
\medskip
\noindent (ii) We prove this by induction on the size of the set $X$.
The claim is trivially true for $|X| = 0$ and $1$, since $|\mathcal{I}| = 1$ and
$2$, respectively, and the graph $\mathcal{G}_X$ has chromatic number
$1$ and $2$, respectively.
Now suppose that we are given a set $X$ and the family
$\mathcal{I}$ of independent sets in $\mathcal{H}$ (as defined above).
Let $c \in X$ be a color with maximum $\iota(c)$
and let $i = \iota(c)$.
Let $\mathcal{I}_c$ be the family of independent sets containing $c$
and $\mathcal{I}'_c$ be the family of all other independent sets.
Let $A$ be the subset of vertices of $[m]^t$ whose $i$-th
coordinate is $\eta_1(c)$. For two vectors $x, y \in A$,
we have $a_i( c_M(x,y) ) = 0$, since both $x$ and $y$
have $i$-th coordinate $\eta_1(c)$.
Hence, in the subgraph of $\mathcal{G}_X$ induced
on the set $A$, we only see colors $c' \in X$
which have $a_{i}(c') = 0$.
Let $X_c \subseteq X$ be the set of colors $c'$ such that $a_{i}(c') = 0$.
The observation above implies that $\mathcal{G}_X[A]$ is a subgraph
of $\mathcal{G}_{X_c}$. By the inductive hypothesis, $\chi(\mathcal{G}_{X_c})$
is at most the number of independent sets of $\mathcal{H}[X_c]$.
Moreover, by the definitions of $X_c$ and $\mathcal{I}_c$ and the
choice of $c$, the independent sets of
$\mathcal{H}[X_c]$ are in one-to-one correspondence with
the independent sets in $\mathcal{I}_c$.
Thus, we have
\[
\chi(\mathcal{G}_X[A])
\le \chi(\mathcal{G}_{X_c})
\le |\mathcal{I}_c|.
\]
Now consider the set $B = [m]^t \setminus A$. The subgraph
of $\mathcal{G}_X$ induced on $B$ does not contain any edge of color $c$
and therefore $\mathcal{G}_{X}[B]$ is a subgraph of $\mathcal{G}_{X \setminus \{c\}}$.
By the inductive hypothesis, $\chi(\mathcal{G}_{X \setminus \{c\}})$
is at most the number of independent sets of $\mathcal{H}[X \setminus \{c\}]$.
By definition, the independent sets of $\mathcal{H}[X \setminus \{c\}]$ are in
one-to-one correspondence with independent sets in $\mathcal{I}'_c$.
Therefore, we have
\[ \chi(\mathcal{G}_X[B]) \le \chi(\mathcal{G}_{X \setminus \{c\}}) \le |\mathcal{I}'_c|. \]
Hence,
\[ \chi(\mathcal{G}_X) \le \chi(\mathcal{G}_X[A]) + \chi(\mathcal{G}_X[B])
\le |\mathcal{I}_c| + |\mathcal{I}'_c|
= |\mathcal{I}|,
\]
and the claim follows.
\end{proofof}
Using Theorem \ref{thm:chi_slow_grow}, we can now prove Theorem \ref{thm:chi_slow_grow_less_colors}, which
we restate here for the reader's convenience.
Recall that for an edge partition $E_1 \cup \ldots \cup E_t$ of the complete graph $K_n$
and a set $I \subseteq [t]$,
we define $\mathcal{G}_I$ as the subgraph of $K_n$ with edge set $\bigcup_{i \in I} E_i$.
\medskip
\noindent {\bf Theorem \ref{thm:chi_slow_grow_less_colors}}.
There exists a positive real $r_0$ such that the following holds
for every positive integer $r$ and positive real $\alpha \le 1$
satisfying $(\log r)^\alpha \ge r_0$.
For $n = 2^{(\log r)^{2 + \alpha}/200}$,
there exists a partition $E = E_1 \cup \dots \cup E_{\sqrt{r}}$ of
the edge set of the complete graph $K_n$ such that
\[ \chi(\mathcal{G}_I) \le 2^{3(\log r)^{\alpha/2}\sqrt{|I| \log 2 |I|}} \]
for all $I \subset [\sqrt{r}]$.
\medskip
\begin{proof}
Let $N = 2^{\log^2 r/200}$ and $t = (\log r)^{\alpha}$ (since
$(\log r)^{\alpha} \ge r_0$ and $\alpha \le 1$, we can guarantee
that $N$ and $t$ are large enough by asking that $r_0$ be
large enough). Color the
edge set of the complete graph on the vertex set $[N]^t$ as follows.
For two vectors $v, w \in [N]^t$ of the form $v = (v_1 ,\ldots,v_t)$
and $w = (w_1, \ldots, w_t)$, we let
\[ c(v,w) = \Big(i, c_M(v_i, w_i)\Big), \]
where $i$ is the minimum index for which $v_i \neq w_i$. Since
$c_M$ on $K_N$ uses at most
$2^{6\sqrt{\log N}} \le \frac{\sqrt{r}}{\log r}$ colors (see the
discussion in Section \ref{sec:preliminaries}), our coloring
uses at most
\[ t \cdot \frac{\sqrt{r}}{\log r} \le \sqrt{r} \]
colors in total. Since $n = N^t$, this
coloring gives an edge partition $E = E_1 \cup \dots \cup E_{s}$
of the complete graph on $n$ vertices, for some integer $s \le \sqrt{r}$.
Now suppose that a set $I \subset [s]$ is given. The set $I$ can be
partitioned into $t$ sets $I_1 \cup \dots \cup I_t$
according to the value of the first coordinate as follows:
for each $i \in [t]$, define $I_i$ as the set of indices $j \in I$ for which
the color of the edges $E_j$ has $i$ as its first coordinate.
For each $i$, let $\pi_i \,:\, [N]^t \rightarrow [N]$ be the
projection map to the $i$-th coordinate. Then the graph
$\pi_i(\mathcal{G}_{I_i})$ becomes a subgraph of $K_N$
induced by the union of $|I_i|$ colors of $c_M$.
Hence, by Theorem \ref{thm:chi_slow_grow}, we know that
\[ \chi(\mathcal{G}_{I_i}) \le \chi(\pi_i(\mathcal{G}_{I_i})) \le 2^{3\sqrt{|I_i| \log 2 |I_i|}} \]
for each $i \in [t]$, where we introduce the extra $2$ in the logarithm to account for the possibility that $|I_i| = 1$. Therefore, we see that
\[ \chi(\mathcal{G}_I) \le \prod_{i \in [t]} \chi(\mathcal{G}_{I_i}) \le 2^{3\sum_{i \in [t]} \sqrt{|I_i| \log 2 |I_i|}}. \]
Since $\sqrt{x \log 2 x}$ is concave, Jensen's inequality implies that the sum in the exponent satisfies
\[\sum_{i \in [t]} \sqrt{|I_i| \log 2 |I_i|} \leq t \sqrt{(|I|/t) \log (2 |I|/t)}
\le \sqrt{t |I| \log 2 |I|} = (\log r)^{\alpha/2} \sqrt{|I| \log 2 |I|} . \]
This implies the required result.
\end{proof}
\section{Concluding Remarks}
\label{sec:conclusion}
\subsection{The grid Ramsey problem with asymmetric colorings}
One may also consider an asymmetric version of the grid Ramsey problem,
where we color the row edges using $r$ colors but are allowed to
use only two colors on the column edges. Let $G(r,2)$ be the minimum $n$
for which such a coloring is guaranteed to contain an alternating rectangle. One can easily
see that
\[ r \le G(r,2) \le r^3 + 1. \]
The following construction improves the lower bound to
$G(r,2) \ge \frac{1}{4}r^2$. Let $n = \frac{1}{4} r^2$ and $p$ be a prime satisfying
$\frac{r}{2} \le p \le r$ and $n \le 2^p$ (the existence of such a prime follows from
Bertrand's postulate).
Consider the $n \times n$ grid.
For each $i \in [n]$, assign to the $i$-th row a sequence
$(a_{i,1}, \ldots, a_{i, p}) \in [r]^{p}$ so that
for all distinct $i$ and $i'$ there exists at most one coordinate
$j \in [p]$ for which $a_{i,j} = a_{i',j}$ (the construction will be
given below). Given these sequences, for each $i \in [n]$ and distinct
$j, j' \in [n]$, color the edge $\{(i,j), (i,j')\}$ as follows:
examine the binary expansions of $j$ and $j'$ to identify the
first bit $t$ in which the two differ and color the edge with color $a_{i,t}$
(this is possible since $2^p \ge n$).
For two distinct rows, suppose that the sequences corresponding to these rows
coincide in the $k$-th coordinate. Then the intersection of the two rows
is a subgraph of the graph connecting vertices
whose $k$-th bit in the binary expansion is 0 to those whose $k$-th bit
in the binary expansion is 1. Thus, the intersection of any two rows
is a bipartite graph and therefore, by the same argument as in the proof of Lemma \ref{lem:row_chromatic},
we obtain $G(r,2) \ge n$
(note that we in fact obtain a coloring of the $cr^2 \times 2^{c'r}$ grid).
It suffices to construct a collection of sequences with the property claimed above.
For $a,b \in \mathbb{Z}_p$, consider the following sequence
with entries in $\mathbb{Z}_p$:
\[
B_{a,b} = \Big( a, a + b, a + 2b, \ldots, a + (p-1)b \Big).
\]
For two distinct pairs $(a,b)$ and $(a',b')$,
the sequences $B_{a,b}$ and $B_{a',b'}$ can overlap in the $i$-th coordinate
if and only if $a + ib = a' + ib'$, which is
equivalent to $(b-b')i = a' - a$. Since $(a,b) \neq (a',b')$, we see
that there exists at most one index $i$ in the range $0 \le i \le p-1$
for which $a + ib = a' + ib'$. Thus the sequences $B_{a,b}$ have the claimed property.
Note that the total number of sequences is at least $p^2 \ge n$.
Moreover, since $p \le r$, by abusing
notation, we may assume that the sequences are in fact in $[r]^p$ and,
therefore, we can use them in the construction of our coloring.
The following question may be more approachable than the corresponding problem for $G(r)$.
\begin{QUES}
Can we improve the upper bound on $G(r, 2)$?
\end{QUES}
\subsection{The Erd\H{o}s--Gy\'arf\'as problem in hypergraphs}
As mentioned in the introduction,
for each fixed $i$ with $0 \le i \leq k$ and large enough $p$,
\[ F_k\left(r, p, {p-i \choose k-i} + 1\right) \le r^{r^{\iddots^{r^{c_{k,p}}} }}, \]
where the number of $r$'s in the tower is $i$.
It would be interesting to establish a lower bound on $F(r, p, {p-i \choose k-i})$
exhibiting a different behavior.
\begin{PROB}
Let $p, k$ and $i$ be positive integers with $k \ge 3$ and $0 < i < k$.
Establish a lower bound on $F_k(r, p, {p-i \choose k-i})$
that is significantly larger than the upper bound on $F_k(r,p,{p-i \choose k-i} + 1)$
given above.
\end{PROB}
We have principally considered the $i=1$ case of this question. For example, the Erd\H{o}s--Gy\'arf\'as
problem on whether $F(n,p,p-1)$ is superpolynomial for all $p \geq 3$ corresponds to the case where
$k = 2$ and $i = 1$. Theorems~\ref{thm:grid_main} and \ref{thm:F_3_5_6} represent progress on the analogous problem
with $k = 3$. The next open case, showing that $F_3(r, 6, 10)$ is superpolynomial, appears difficult.
For $i \geq 2$, it seems likely that one would have to invoke a variant of the stepping-up technique of Erd\H{o}s and
Hajnal (see, for example, \cite{GrRoSp}). In particular, we would like to know the answer to the following question.
\begin{QUES}
Is $F_3(r, p, p-2)$ larger than $2^{r^c}$ for any fixed $c$?
\end{QUES}
For $p = 4$, a positive solution to this problem follows since we know that the Ramsey number of $K_4^{(3)}$ is
double exponential in the number of colors (see, for example, \cite{AxGyLiMu}). The general case appears to be much more difficult.
Another case of particular importance is $F_{2d-1}(r, 2d, d+1)$, since it is this function (or rather a $d$-partite variant)
which is used by Shelah in his proof of the Hales--Jewett theorem. If the growth rate of this function is a tower of
bounded height for all $d$, then it would be possible to give a tower-type bound for Hales--Jewett numbers. However, we expect
that this is not the case.
\begin{PROB}
Show that for all $s$ there exists $d$ such that
\[ F_{2d-1}\left(r, 2d, d+1\right) \ge 2^{2^{\iddots^{2^{r}}}}, \]
where the number of $2$'s in the tower is at least $s$.
\end{PROB}
\subsection{Studying the chromatic number version of the Erd\H{o}s--Gy\'arf\'as problem}
Since we know that both $F(r,p,p-1)$ and $F_\chi(r,4,3)$ are superpolynomial in $r$, it is natural to ask the following
question (see also \cite{CoFoLeSu}).
\begin{QUES} \label{que:chi_poly}
Is $F_{\chi}(r,p,p-1)$ superpolynomial in $r$?
\end{QUES}
By following a similar line of argument to the lower bound for $F_\chi(r,4,3)$,
we can show that $c_M$ is also a chromatic-$(5,4)$-coloring.
Therefore, $F_{\chi}(r,5,4) = 2^{\Omega(\log^2 r)}$, answering Question~\ref{que:chi_poly} for $p =5$. Since the proof is
based on rather tedious case analysis, we will post a supplementary note rather than including the details here. It would
be interesting to determine whether the $(p, p-1)$-colorings defined in \cite{CoFoLeSu} are also chromatic-$(p,p-1)$-colorings. If so, they
would provide a positive answer to Question~\ref{que:chi_poly}.
In Theorem~\ref{thm:chi_4_3}, we showed that
$2^{\Omega(\log^2 r)} \le F_{\chi}(r, 4, 3) \le 2^{O(\sqrt{r \log r})}$.
It would be interesting to reduce the gap between the lower and upper bounds.
Since $F_{\chi}(r,4,2) \ge 2^{r} + 1$, we see that
$F_{\chi}(r,4,2)$ is exponential in $r$, while $F_{\chi}(r,4,3)$ is subexponential in $r$.
For $p \ge 5$, the value of $q$ for which the transition
from exponential to subexponential happens is not known. However, recall that
$F_{\chi}(r, 2^d + 1, d+1)$ is exponential in $r$ for all $d \ge 1$.
This followed from showing that in the edge coloring $c_B$
the union of every $d$ color classes induces a graph
of chromatic number $2^d$ (see Section~\ref{sec:preliminaries}).
The following question asks whether
a similar edge coloring exists if we want the union of every
$d$ color classes to induce a graph of chromatic number at most $2^d - 1$.
\begin{QUES} \label{ques:ch_3}
Is $F_\chi(r,2^{d}, d+1) = 2^{o(r)}$ for all $d \ge 2$?
\end{QUES}
A positive answer to Question \ref{ques:ch_3} would allow us to determine,
for all $p$, the maximum value of $q$ for which $F_{\chi}(r,p,q)$
is exponential in $r$. Indeed, for $2^{d-1} < p \le 2^{d}$, we have
\[
F_\chi(r, p, d) \ge F_\chi(r, 2^{d-1}+1, d) = 2^{\Omega(r)},
\]
while a positive answer to Question \ref{ques:ch_3} would imply
\[
F_\chi(r, p, d+1) \le F_\chi(r, 2^{d}, d+1) = 2^{o(r)}.
\]
Hence, given a positive answer to Question \ref{ques:ch_3}, the maximum value of $q$ for
which $F_\chi(r, p, q)$ is exponential in $r$ would be $q = \lceil \log p \rceil$.
A key component in our proof of Theorem~\ref{thm:grid_main} was Theorem~\ref{thm:chi_slow_grow}, which
says that in the coloring $c_M$, the chromatic number of the union of any $s$ color classes is not too large.
We suspect that our estimate on the chromatic number is rather weak. It would be interesting to improve it further.
More generally, we have the following rather informal question, progress on which might allow us to improve
the bounds in Theorem~\ref{thm:grid_main}.
\begin{QUES}
Given an edge partition of the complete graph $K_n$, how slowly can the chromatic
number of the graph determined by the union of $s$ color classes grow?
\end{QUES}
Finally, let $\mathcal{F}$ be a family of graphs and define $F(r,q; \mathcal{F})$
to be the minimum integer $n$ for which every edge coloring of $K_n$ with $r$ colors
contains a subgraph $F \in \mathcal{F}$ that contains fewer than
$q$ colors. $F(r,q; \mathcal{F})$ generalizes both $F(r,p,q)$
and $F_\chi(r,p,q)$ since we may take $\mathcal{F}$ to be
$\{K_p\}$ for $F(r,p,q)$ and the family of all $p$-chromatic graphs
for $F_\chi(r,p,q)$. Our results suggest that $F(r,q; \mathcal{F})$ is
closely related to the chromatic number of the graphs in $\mathcal{F}$.
The case where $\mathcal{F}$ consists of a single complete bipartite graph was studied in \cite{AxFuMu}.
\medskip
\noindent
{\bf Acknowledgement.} We would like to thank the anonymous referee for several helpful comments.
|
\section*{Introduction}
The central objects of study in this article are affine difference algebraic groups. Similarly to the case of affine algebraic groups, these groups can all be realized as subgroups of the general linear group defined by algebraic difference equations. The defining equations here are not simply polynomials in the matrix entries but difference polynomials, i.e., the defining equations involve a difference operator $\sigma$, which has to be interpreted as a ring endomorphism. For example, if $G$ is the difference algebraic subgroup of $\operatorname{GL}_n$ defined by the algebraic difference equations $X\sigma(X)^{\operatorname{T}}=\sigma(X)^{\operatorname{T}}X=I_n$, then $G(\mathbb{C})$ is the group of all complex unitary $n\times n$-matrices, where $\sigma\colon\mathbb{C}\to\mathbb{C}$ is the complex conjugation map.
A more classical example of a difference field, i.e., a field equipped with an endomorphism would be $\mathbb{C}(x)$ with $\sigma(f(x))=f(x+1)$.
Alternatively, affine difference algebraic groups may be described as affine group schemes with a certain additional structure (the difference structure). As schemes they are typically not of finite type, but they enjoy a certain finiteness property with respect to the difference structure; they are ``of finite $\sigma$-type''. From an algebraic point of view, an affine difference algebraic group $G$ over a difference field $k$ corresponds to a Hopf algebra $k\{G\}$ over $k$ together with a ring endomorphism $\sigma\colon k\{G\}\to k\{G\}$ which extends $\sigma\colon k\to k$. The Hopf algebra structure maps are required to commute with $\sigma$, and $k\{G\}$ is required to be finitely $\sigma$-generated over $k$, i.e., there exists a finite set $B\subset k\{G\}$ such that $B,\sigma(B),\sigma^2(B),\ldots$ generates $k\{G\}$ as a $k$-algebra.
Difference algebraic groups are the discrete analog of differential algebraic groups, i.e., groups defined by algebraic differential equations. Differential algebraic groups have always played an important role in differential algebra (see, e.g., \cite{Cassidy:differentialalgebraicgroups}, \cite{Sit:DifferentialAlgebraicSubgroupsofSL2}, \cite{Cassidy:TheDifferentialRationalRepresentationAlgebraOnALinearDifferentialAlgebraicGroup}, \cite{Cassidy:UnipotentDifferentialAlgebraicGroups}, \cite{Kolchin:differentialalgebraicgroups}, \cite{Cassidy:TheClassificationOfTheSemisimpleDifferentialAlgebraicGroups} \cite{Buium:DifferntialAlgebraicGroupsOfFiniteDimension}, \cite{Buium:GeometryOfDifferentialPolynomialFunctionsIAgebraicGroups}) and are an active area of research nowadays. See e.g., \cite{Pillay:SomeFoundationalQuestionsConcerningDifferentialAlgebraicGroups}, \cite{KowalskiPillay:ProalgebraicAndDifferentialAlgebraicGroupStructuresOnAffineSpaces}, \cite{CassidySinger:AJordanHoelderTheoremForDifferentialAlgebraicGroups}, \cite{MinchenkiOvchinnikov:ZariskiClosuresOfReductiveLinearDifferentialAlgebraiGroups}, \cite{MinchenkoOvchinnikov:ExtensionsOfDifferentialRepresentationsOfSl2andTori}, \cite{Freitag:IndecomposabilityForDifferentialAlgebraicGroups}, \cite{Minchenko:OnCentralExtensionsOfSimpleDifferentialAlgebraicGroups}. See also \cite{Malgrange:DifferentialAlgebraicGroups} for a more geometric approach to differential algebraic groups and \cite{Buium:DifferentialSubgroupsOfSimpleAlgebraicGroupsOverPadicFields} for an arithmetic analog of difference/differential algebraic groups.
Over the last decade, Galois theories where the Galois groups are differential algebraic groups (\cite{Pillay:DifferentialGaloisTheory1}, \cite{Landesman:GeneralizedDifferentialGaloisTheory}, \cite{CassidySinger:GaloisTheoryofParameterizedDifferentialEquations}, \cite{HardouinSinger:DifferentialGaloisTheoryofLinearDifferenceEquations}) have given a new impetus to the study of differential algebraic groups.
The Galois theories in \cite{CassidySinger:GaloisTheoryofParameterizedDifferentialEquations} and \cite{HardouinSinger:DifferentialGaloisTheoryofLinearDifferenceEquations} are also known as parameterized Picard--Vessiot theories as they generalize the standard Picard-Vessiot theory (\cite{Kolchin:AlgebraicMatricGroupsAndThePicardVessiotTheoryOfHomogeneousLinearOrdinaryDifferentialEquations}, \cite{SingerPut:differential}, \cite{SingerPut:difference}) of linear differential and difference equations to linear differential and difference equations depending on (continuous) parameters. In the standard Picard--Vessiot theory the Galois groups are linear algebraic groups and they measure the algebraic dependencies among the solutions. In the parameterized Picard--Vessiot theory the Galois groups are linear differential algebraic groups and they measure the differential algebraic dependencies with respect to an auxiliary set of derivations. A typical application of the parameterized Picard--Vessiot theory is to prove the differential transcendence of special functions (\cite{Arreche:AGaloisTheoreticProofOfTheDifferentialTranscendenceOfTheIncompleteGammaFunction}, \cite{DiVizio:ApprocheGaloisienneDeLaTranscendanceDifferentielle}).
The tannakian approach to differential algebraic groups (\cite{Ovchinnikov:TannakianApproachToLinearDifferentialAlgebraicGroups}) has found applications in the parameterized Galois theory (\cite{Ovchinnikov:TannakianCategoriesLinearDifferentialAlgebraicGroupsandParametrizedLinearDifferentialEquations}, \cite{GilletGorchinskyOvchinnikov:ParameterizedAndAtiyah}, \cite{GorchinskyOvchinnikov:Isomonodromic}) and a detailed study of representations of linear differential algebraic groups has led to the first algorithms for computing the Galois group of parameterized linear differential equations (\cite{MinchenkoOvchinnikovSinger:UnipotentDifferentialAlgebraicGroupsAsParameterizedDifferentialGaloisGroups}, \cite{MinchenkoOvchinnikovSinger:ReductiveLinearDifferentialAlgebraicGroupsAndTheGaloisGroups}, \cite{Arreche:ComputingTheDifferentialGaloisGroupOfSecondOrderEquations}, \cite{Dreyfus:ComputingTheGaloisGroupOfSomeParameterizedLinearDifferentialEquationOrderTwo}).
\medskip
In contrast to the situation in differential algebra, difference algebraic groups have long played no role at all in difference algebra. The author can only speculate why. Maybe because the traditional definition of a difference variety, which involves a so--called universal system of difference fields (see \cite[Chapter 4]{Cohn:difference}) is not really suitable for studying groups. (For each difference field in the universal system one obtains a group structure but the difference variety itself does not carry a group structure.)
Around the turn of the century a considerable interest in the model theory of difference fields emerged. (See e.g., \cite{Macintyre:GenricAutomorphismsOfFields}, \cite{ChatzidakisHrushovskiPeterzil:ModelTheoryofDifferencefields}, \cite{ChatzidakiHrushovskiPeterzil:ModelTheoryofDifferenceFieldsIIPeriodicIdelas}.) Groups definable in ACFA, the model companion of the theory of difference fields, played a crucial role in remarkable applications of model theory to number theory, especially regarding the Manin--Mumford conjecture. See \cite{Hrushovski:TheManinMumfordConjectureAndTheModelTheorOfDifferenceFields},
\cite{Bouscaren:TheorieModesEtManinMumford}, \cite{Chatzidakis:DifferencefieldsModelTheoryAndApplicationsToNumberTheory},
\cite{Chatzidakis:GroupsDefinableInACFA}, \cite{Scanlon:APositiveCharacteristicManinMumfordTheorem}, \cite{Scanlon:LocalAndreOrtConjectureForTheUniversalAbelianVariety}, \cite{Scanlon:DifferenceAlgebraicSubgroupsOfCommutativeAlgebraicGroups}, \cite{KowalskiPillay:OnAlgebraicSigmaGroups}.
In \cite{KowalskiPillay:ANoteonGroupsDefinableInDifferenceFields} it is shown that every group definable in ACFA is, in a certain sense, close to being a definable (in the language of difference fields) subgroup of an algebraic group. Here we will show that every affine difference algebraic group is isomorphic to a difference closed subgroup of the general linear group. In \cite{KowalskiPillay:ANoteonGroupsDefinableInDifferenceFields} it is also shown that an affine difference algebraic group whose underlying difference variety is an affine space is isomorphic to a difference closed subgroup of a unipotent algebraic group.
Recently, a Galois theory for linear differential equations depending on a discrete parameter has been developed in \cite{diVizioHardouinWibmer:DifferenceGaloisofDifferential}. In \cite{OvchinnikovWibmer:SGaloisTheoryOfLinearDifferenceEquations} a similar Galois theory has been developed for linear difference equations depending on a discrete parameter. The Galois groups in these Galois theories are affine difference algebraic groups and they measure the difference algebraic relations among the solutions. For example, the difference algebraic relation $xJ_{\alpha+2}(x)-2(\alpha+1)J_{\alpha+1}(x)+xJ_\alpha(x)=0$ satisfied by the Bessel function $J_\alpha(x)$, which solves
Bessel's differential equation $x^2\delta^2(y)+x\delta(y)+(x-\alpha^2)y=0$, is witnessed by the associated Galois group.
As illustrated in \cite{diVizioHardouinWibmer:DifferenceAlgebraicRelations} and \cite{OvchinnikovWibmer:SGaloisTheoryOfLinearDifferenceEquations}, these Galois theories make it possible to use structure results about affine difference algebraic groups to analyze and classify the possible difference algebraic relations among the solutions of certain linear differential and difference equations. In this respect the understanding of the Zariski dense difference closed subgroups of a given affine algebraic group is highly relevant. For example, some understanding of the Zariski dense difference closed subgroups of $\operatorname{SL}_2$, originating from \cite{ChatzidakiHrushovskiPeterzil:ModelTheoryofDifferenceFieldsIIPeriodicIdelas}, is a key ingredient to prove that any two linearly independent solution of the Airy equation are difference algebraically independent. More precisely (\cite[Corollary 6.10]{diVizioHardouinWibmer:DifferenceAlgebraicRelations}), if $A(x)$ and $B(x)$ are $\mathbb{C}$-linearly independent solutions of Airy's equation $\delta^2(y)-xy=0$, then $A(x), B(x), A'(x), A(x+1), B(x+1), A'(x+1), A(x+2),\ldots$ are algebraically independent over $\mathbb{C}(x)$.
While the parameterized Picard--Vessiot theory for continuous parameters could draw on a well--established comprehensive theory of differential algebraic groups, the situation for discrete parameters is adverse. But clearly, a well--developed theory of affine difference algebraic groups is indispensable for the parameterized Picard--Vessiot theory with discrete parameters to flourish. Obviously one could not get very far without having fundamental results such as the analogs of the isomorphism theorems for groups at one's disposal. Even though the validity of the isomorphism theorems for affine difference algebraic may not come as a surprise, the proof is far from being obvious. Indeed, even the existence of the quotient of an affine difference algebraic group modulo a normal difference closed subgroup is a highly non--trivial question. A positive answer requires proving that every $k$-$\sigma$-Hopf subalgebra of a finitely $\sigma$-generated $k$-$\sigma$-Hopf algebra is finitely $\sigma$-generated.
While it will for sure be a long way to lift the theory of difference algebraic groups to a level comparable to the contemporary theory of differential algebraic groups, it is the hope of the author that this article may serve as a first step in this direction. Here we are mainly concerned with general properties of affine difference algebraic groups. We plan to study special classes (e.g., \'{e}tale, unipotent, diagonalizable) affine difference algebraic groups next.
A tannakian approach to affine difference algebraic groups has been developed in \cite{OvchinnikovWibmer:TannakianCategoriesWithSemigroupActions}. (See also \cite{Kamensky:TannakianFormalismOverFieldsWithOperators}.)
We expect that the finiteness properties of affine difference algebraic groups proved here will enable us to apply this tannakian approach to the parameterized Picard--Vessiot theory with discrete parameters in a similar fashion as for the case of continuous parameters in \cite{GilletGorchinskyOvchinnikov:ParameterizedAndAtiyah}.
It is well recognized that a functorial--schematic approach to algebraic groups has its benefits. (See \cite{Waterhouse:IntroductiontoAffineGroupSchemes}, \cite{DemazureGabriel:GroupesAlgebriques}, \cite{Jantzen:RepresentationsOfAlgebraicGroups}, \cite{Milne:BasicTheoryOfAffineGroupSchemes}, \cite{Grothendieck:SGA3_1}.)
Here we adopt a similar approach for difference algebraic groups. For example, the general linear group $\operatorname{GL}_n$ over a difference field $k$, considered as a difference algebraic group, is the functor which assigns $\operatorname{GL}_n(R)$ to every $k$-$\sigma$-algebra $R$, that is, $R$ is a $k$-algebra equipped with an endomorphism $\sigma\colon R\to R$ which extends $\sigma\colon k\to k$. In the model theoretic approach to difference equations as well as in classical difference algebra (\cite{Cohn:difference}, \cite{Levin:difference}) one is primarily interested
in solutions in difference field extensions of $k$, i.e., $R$ is required to be a field. An affine difference algebraic group is a group object in the category of affine difference varieties. The definition of an affine difference variety used in this article is more general than the classical definition of a difference variety (as in \cite{Cohn:difference} and \cite{Levin:difference}). Our definition is equivalent to what is called a $\underline{\mathcal{D}}$-subscheme of affine space in \cite{MoosaScanlon:GeneralizedHasseSchmidtVarietiesAndTheirJetSpaces} for a suitable choice of $\underline{\mathcal{D}}$. Also, our notion of affine difference algebraic group is equivalent to what is called a linear $\mathfrak{M}$-group in \cite{Kamensky:TannakianFormalismOverFieldsWithOperators} for a suitable choice of $\mathfrak{M}$.
The classical difference varieties studied in \cite{Cohn:difference} and \cite{Levin:difference} correspond precisely to the affine difference varieties which can be recovered from their points in difference fields. Thus, the relation between our affine difference varieties and the classical difference varieties is similar to the relation between affine schemes of finite type and affine varieties. The affine difference varieties which can be recovered from their points in difference fields are those whose defining ideal is perfect, therefore, we call them perfectly $\sigma$-reduced. Affine difference algebraic groups which are not perfectly $\sigma$-reduced occur quite naturally and frequently as Galois groups of linear differential or difference equations depending on a discrete parameter. There are several examples (Example \ref{ex: no field points}) of difference algebraic groups $G$ with the property that $G(K)$ is the trivial group for every difference field extension $K$ of $k$.
In difference algebraic geometry there is a ``zoo'' of elements playing a role analogous to nilpotent elements in algebraic geometry.
They roughly correspond to the following assertions valid for elements in a difference field extension of $k$ but not necessarily valid for elements in a $k$-$\sigma$-algebra:
\begin{itemize}
\item $a^n=0$ implies $a=0$.
\item $\sigma^n(a)=0$ implies $a=0$.
\item $ab=0$ implies $a\sigma(b)=0$.
\item $a\sigma(a)=0$ implies $a=0$.
\end{itemize}
In this sense we obtain four difference closed subgroups of an affine difference algebraic group which play a role analogous to the maximal reduced subgroup of an affine algebraic group. The theory of affine algebraic groups runs smoother if one allows nilpotent elements in the structure sheaf, and non--reduced algebraic groups play an important role in the representation theory of affine algebraic groups in positive characteristic. The situation for affine difference algebraic groups is similar. The category of affine difference algebraic groups is much better behaved than the category of perfectly $\sigma$-reduced affine difference algebraic groups. This, for example, becomes apparent when dealing with extensions of the base field or when dealing with quotients and the analogs of the isomorphism theorems for groups.
For example, the morphism of affine difference algebraic groups $\phi\colon \operatorname{GL}_n\to\operatorname{GL}_n$ determined by $\phi((g_{ij}))=(\sigma(g_{ij}))$ for $(g_{ij})\in\operatorname{GL}_n(R)$ and $R$ a $k$-$\sigma$-algebra has a non--trivial kernel, even though $\phi\colon \operatorname{GL}_n(K)\to\operatorname{GL}_n(K)$ is injective for every difference field extension of $k$. In particular, if $k$ is a model of ACFA, then $\phi\colon \operatorname{GL}_n(k)\to\operatorname{GL}_n(k)$ is bijective but $\phi$ is not an isomorphism of difference algebraic groups. Here we will show that (with the appropriate conception of injective and surjective) a morphism of affine difference algebraic groups is an isomorphism if and only if it is injective and surjective.
Even though our notion of an affine difference algebraic group as well as the notion of a group definable in ACFA both capture the idea of a group defined by algebraic difference equations, there are several differences between the two notions.
On the one hand groups definable in ACFA are more general since they need not be affine, for example \cite{ChatzidakisHrushovski:OnSubgroupsOfSemiabelianVarietiesDefinedByDifferenceEquations} studies definable subgroups of semi--abelian varieties. Moreover, since ACFA does not (fully) eliminate quantifiers, a group definable in ACFA may not be definable by difference polynomials.
For example, the formula $\exists h: h^2=g,\ \sigma(h)=h$ defines a subgroup of the multiplicative group and the existential quantifier can not be eliminated.
On the other hand, if $k$ is a model of ACFA, the subgroups of $\operatorname{GL}_n(k)$ defined by difference polynomials in the matrix entries only correspond to the perfectly $\sigma$-reduced difference closed subgroups of $\operatorname{GL}_n$.
Let $k$ be a difference field and $K$ a model of ACFA containing $k$. Unfortunately, the functor which associates to a perfectly $\sigma$-reduced affine difference algebraic $G$ over $k$ its $K$-points $G(K)$ is not faithful, as the embedding $k\hookrightarrow K$ involves a non--canonical choice. For example, if $k=\mathbb{Q}$ and $G$ is the difference closed subgroup of the multiplicative group given by $G(R)=\{g\in R^\times|\ g^3=1,\ \sigma(g)=g\}$ for any $k$-$\sigma$-algebra $R$, then $G$ has a non--trivial endomorphism given by $g\mapsto g^2$. But if $K$ is a model of ACFA of characteristic zero such that $\sigma$ permutes the two non--trivial third roots of unity in $K$, then $G(K)$ is the trivial group and the endomorphism of $G$ collapses to the identity on $G(K)$. To obtain a faithful functor, one would need to impose restrictions on the base difference field $k$, but this is something we wish to avoid as it would reduce the applicability of the theory.
\medskip
Let us now describe the content of the article in more detail. The first section contains preliminaries from difference algebraic geometry. We introduce affine difference varieties and some basic constructions with them. In the second section we define affine difference algebraic groups, present several examples and show that every affine difference algebraic group is isomorphic to a difference closed subgroup of the general linear group. In particular, an affine difference algebraic group $G$ can be embedded into an algebraic group (as a difference closed subgroup). In Section 3 we explain how to associate to any such embedding an affine algebraic group called the growth group. We also prove the existence of a Kolchin (or dimension) polynomial for affine difference algebraic groups and define the difference dimension and the order of an affine difference algebraic group. While the growth group depends on the chosen embedding, we will show that it contains some information which only depends on $G$. For example, its dimension is equal to the difference dimension of $G$.
In Section 4 we prove two important finiteness theorems. For clarity we state here the theorems in the language of Hopf algebras: Let $k\{G\}$ be a $k$-$\sigma$-Hopf algebra which is finitely $\sigma$-generated over $k$ and let $\mathbb{I}(H)\subset k\{G\}$ be a $\sigma$-Hopf ideal, i.e., a Hopf ideal such that $\sigma(\mathbb{I}(H))\subset\mathbb{I}(H)$. Then $\mathbb{I}(H)$ is finitely $\sigma$-generated, i.e., there exists a finite set $B\subset \mathbb{I}(H)$ such that $B,\sigma(B),\sigma^2(B),\ldots$ generates $\mathbb{I}(H)$ as an ideal. This result can be seen as a strengthening (for $k$-$\sigma$-Hopf algebras) of the Ritt--Raudenbush basis theorem in difference algebra. The classical Ritt--Raudenbush basis theorem does not apply in our setting as it only applies to perfect difference ideals. This finiteness result turns out to be extremely useful and is used repeatedly in the subsequent developments. For example, it is used in the proof of the dimension theorem which is also proved in Section 4. The second finiteness theorem asserts that every $k$-$\sigma$-Hopf subalgebra of $k\{G\}$ is finitely $\sigma$-generated.
In Section 5 we briefly touch upon representations of affine difference algebraic groups. The main result here is an analog of a theorem of Chevalley: Every difference closed subgroup of $G$ is the stabilizer of a line in a suitable representation of $G$. We also show that the category of representations of a torus (considered as a difference algebraic group) is semi--simple. This is in sharp contrast to what happens in the theory of linear differential algebraic groups. The category of representations of a linear differential algebraic group is semi--simple only for linear differential algebraic groups which are the constant points of a reductive algebraic group (\cite{MinchenkiOvchinnikov:ZariskiClosuresOfReductiveLinearDifferentialAlgebraiGroups}).
In Section 6 we introduce a numerical invariant of affine difference algebraic groups called the limit degree. Its definition is analogous to an important invariant of extensions of difference fields also called the limit degree (\cite[Section 4.3]{Levin:difference}). So--called algebraic $\sigma$-groups have been introduced and studied in \cite{KowalskiPillay:OnAlgebraicSigmaGroups}. We show that the category of affine algebraic $\sigma$-groups is equivalent to the category of affine difference algebraic groups of difference dimension zero and limit degree one.
In Section 7 we establish the existence of quotients and show that the difference dimension, the order and the limit degree behave on quotient in the expectable way.
Section 8 then studies morphisms of affine difference algebraic groups. We characterize injective and surjective morphisms and show that every morphism of affine difference algebraic groups factors uniquely as the composition of a surjective morphism followed by an injective morphism.
In Section 9 we study the components of affine difference algebraic groups. We introduce the connected component and prove a third finiteness theorem: A finitely $\sigma$-generated $k$-$\sigma$-Hopf algebra has only finitely many minimal prime difference ideals. This proves a special case of a reformulation of a question raised by E. Hrushovski.
In Section 10 we deal with the ring elements playing a role analogous to nilpotent elements in algebraic geometry. In particular, we introduce four difference closed subgroups playing a role analogous to the maximal reduced subgroup for algebraic groups.
Sheaves (\cite[Chapter III]{DemazureGabriel:GroupesAlgebriques}) are a useful tool for dealing with quotients of algebraic groups. In Section 11 we adapt the sheaf approach to difference algebraic groups and then use it to prove the analogs of the isomorphism theorems for groups.
Section 12 then further expands on the group theoretic properties of affine difference algebraic groups. We prove an analog of the Jordan--H\"{o}lder decomposition theorem.
Finally, in Section 13 we present an application to the parameterized Picard--Vessiot theory with discrete parameters.
\section{Difference algebraic geometry preliminaries} \label{sec: Difference algebraic geometry preliminaries}
In this section we introduce some basic notions from difference algebraic geometry, e.g., the notion of a difference variety. These definitions and results will then be used in the following sections in the study of difference algebraic groups.
\subsection{Basic definitions}
We start by recalling some basic notions from difference algebra. Standard references are \cite{Levin:difference} and \cite{Cohn:difference}. All rings are assumed to be commutative and unital.
A difference ring, or \emph{$\sigma$-ring} for short, is a ring $R$ together with a ring endomorphism $\sigma\colon R\to R$. Contrary to \cite{Levin:difference} we do not assume that $\sigma\colon R\to R $ is injective. If $R$ is a field, we speak of a \emph{$\sigma$-field}. We usually omit $\sigma$ from the notation, and simply refer to $R$ as a $\sigma$-ring. A morphism between $\sigma$-rings $R$ and $S$ is a morphism $\psi\colon R\to S$ of rings such that $\psi(\sigma(r))=\sigma(\psi(r))$ for all $r\in R$. A $\sigma$-ring $R$ is called \emph{inversive} if $\sigma\colon R\to R$ is an automorphism. A subset $A$ of a $\sigma$-ring $R$ is called \emph{stable under $\sigma$} if $\sigma(A)\subset A$.
Let $k$ be a $\sigma$-ring. A $\sigma$-ring $R$ together with a $k$-algebra structure is called a \emph{$k$-$\sigma$-algebra} if the algebra structure map $k\to R$ is a morphism of $\sigma$-rings. A morphism of $k$-$\sigma$-algebras is a morphism of $k$-algebras which is also a morphism of $\sigma$-rings. The category of $k$-$\sigma$-algebras is denoted by $k$-$\sigma$-$\mathsf{Alg}$. A $k$-subalgebra of a $k$-$\sigma$-algebra is called a \emph{$k$-$\sigma$-subalgebra} if it is stable under $\sigma$. If $k$ is a $\sigma$-field, a $k$-$\sigma$-algebra which is a $\sigma$-field is called a \emph{$\sigma$-field extension} of $k$.
Let $R$ and $S$ be $k$-$\sigma$-algebras. Then $R\otimes_k S$ is naturally a $k$-$\sigma$-algebra by $\sigma(r\otimes s)=\sigma(r)\otimes\sigma(s)$ for $r\in R$ and $s\in S$.
Let $k$ be a $\sigma$-field and $R$ a $k$-$\sigma$-algebra. For a subset $A$ of $R$, the smallest $k$-$\sigma$-subalgebra of $R$ containing $A$ is denoted by $k\{A\}$. If there exists a finite subset $A$ of $R$ such that $R=k\{A\}$, we say that $R$ is finitely $\sigma$-generated over $k$.
The \emph{$\sigma$-polynomial ring} over $k$ in the $\sigma$-variables $y=(y_1,\ldots,y_n)$ is the polynomial ring over $k$ in the variables $y_1,\ldots,y_n,\sigma(y_1),\ldots,\sigma(y_n),\sigma^2(y_1),\ldots$. It is denoted by
$$k\{y\}=k\{y_1,\ldots,y_n\}$$
and has a natural $k$-$\sigma$-algebra structure.
It $R$ is a $k$-$\sigma$-algebra and $F\subset k\{y\}$ a set of $\sigma$-polynomials over $k$, it makes sense to consider the $R$-rational solutions of $F$, that is
$$\mathbb{V}_R(F)=\{ a\in R^n|\ f(a)=0 \text{ for all } f\in F\}.$$ Note that $R\rightsquigarrow\mathbb{V}_R(F)$ is naturally a functor from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$, the category of sets. We denote this functor by $\mathbb{V}(F)$.
\begin{defi} \label{defi: svariety}
Let $k$ be a $\sigma$-field. A \emph{difference variety} (or \emph{$\sigma$-variety} for short) over $k$ is a functor from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$, which is of the form $\mathbb{V}(F)$, for some $n\geq 1$ and $F\subset k\{y_1,\ldots,y_n\}$. A morphism of $\sigma$-varieties is a morphism of functors.
\end{defi}
The above definition does not agree with the traditional definition of a difference variety in \cite{Levin:difference}, since traditionally one is only interested in solutions in $\sigma$-field extensions of $k$. The relation between Definition \ref{defi: svariety} and Definition 2.6.1 in \cite{Levin:difference} is analogous to the relation between the definition of an affine scheme of finite type over a field and the notion of an affine variety. Our definition is equivalent to what is called a $\underline{\mathcal{D}}$-subscheme of affine space in \cite{MoosaScanlon:GeneralizedHasseSchmidtVarietiesAndTheirJetSpaces} for a suitable choice of $\underline{\mathcal{D}}$.
It may seem more accurate to add the word ``affine'' into Definition \ref{defi: svariety}. However, to avoid endless iterations of the word ``affine'' we have chosen not to do so.
By definition, a morphism $\phi\colon X\to Y$ of $\sigma$-varieties consists of maps $\phi_R\colon X(R)\to Y(R)$ for any $k$-$\sigma$-algebra $R$. For convenience, we will sometimes drop the $R$ in $\phi_R$. In particular, if $x\in X(R)$, we may write $\phi(x)$ for $\phi_R(x)\in Y(R)$.
Let $F,G\subset k\{y\}$ and $X=\mathbb{V}(F)$, $Y=\mathbb{V}(G)$. If $F\subset G$ then $Y(R)\subset X(R)$ for every $k$-$\sigma$-algebra $R$ and $Y$ is a subfunctor of $X$. We say that $Y$ is a \emph{$\sigma$-closed $\sigma$-subvariety} of $X$ and write $Y\subset X$.
Let $R$ be a $\sigma$-ring. An ideal $\mathfrak{a}$ of $R$ is called a \emph{$\sigma$-ideal} if $\sigma(\mathfrak{a})\subset\mathfrak{a}$. In this case $R/\mathfrak{a}$ has naturally the structure of a $\sigma$-ring such that the canonical map $R\to R/\mathfrak{a}$ is a morphism of $\sigma$-rings. If $A$ is a subset of $R$, the smallest $\sigma$-ideal of $R$ containing $A$ is denoted by $$[A]\subset R$$ and called the $\sigma$-ideal generated by $A$. A $\sigma$-ideal $\mathfrak{a}\subset R$ is \emph{finitely generated as a $\sigma$-ideal} if there exists a finite set $A\subset\mathfrak{a}$ such that $\mathfrak{a}=[A]$. A $\sigma$-ideal $\mathfrak{p}$ of $R$ which is a prime ideal is called \emph{$\sigma$-prime} if $\sigma^{-1}(\mathfrak{p})=\mathfrak{p}$.
Let $X=\mathbb{V}(F)$, $F\subset k\{y\}=k\{y_1,\ldots,y_n\}$ be a $\sigma$-variety over the $\sigma$-field $k$. Then
\begin{equation} \label{eqn: I(X)}
\mathbb{I}(X)=\{f\in k\{y\}| \ f(a)=0 \ \text{ for all $k$-$\sigma$-algebras $R$ and all } a\in X(R) \}
\end{equation}
is a $\sigma$-ideal of $k\{y\}$. The $k$-$\sigma$-algebra
$$k\{X\}=k\{y\}/\mathbb{I}(X)$$
is called the \emph{coordinate ring} of $X$. As we may choose $R=k\{y\}/[F]$ in (\ref{eqn: I(X)}), we see that $\mathbb{I}(X)=[F]\subset k\{y\}$.
Let $R$ be $k$-$\sigma$-algebra. The bijection
$$\operatorname{Hom}(k\{X\},R)\to X(R)$$
which maps a morphism $\psi\colon k\{X\}\to R$ of $k$-$\sigma$-algebras to $\psi(\overline{y})$ is functorial in $R$. Thus the functor $X$ is represented by $k\{X\}$. Conversely, since every finitely $\sigma$-generated $k$-$\sigma$-algebra can be written in the form $k\{y\}/[F]$, we see that a functor from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$ which is representable by a finitely $\sigma$-generated $k$-$\sigma$-algebra is isomorphic (as a functor) to a $\sigma$-variety.
In the sequel we will allow ourselves the little abuse of notation to also call a functor isomorphic to a $\sigma$-variety a $\sigma$-variety.
In particular, we will often identify $X$ with the functor $\operatorname{Hom}(k\{X\},-)$.
Thus a functor $X$ from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$ is a $\sigma$-variety if and only if it is representable by a finitely $\sigma$-generated $k$-$\sigma$-algebra $k\{X\}$. It is then clear from the Yoneda lemma that:
\begin{rem} \label{rem: equivalence of categories for svarieties}
The category of $\sigma$-varieties over $k$ is anti--equivalent to the category of finitely $\sigma$-generated $k$-$\sigma$-algebras.
\end{rem}
If $\phi\colon X\to Y$ is a morphism of $\sigma$-varieties over $k$, the dual morphism of $k$-$\sigma$-algebras is denoted by
$$\phi^*\colon k\{Y\}\to k\{X\}.$$
Let $R$ and $S$ be $k$-$\sigma$-algebras. Then $R\otimes_k S$ is a $k$-$\sigma$-algebra by $\sigma(r\otimes s)=\sigma(r)\otimes \sigma(s)$. Obviously $R\otimes_k S$ is the coproduct of $R$ and $S$ in the category of $k$-$\sigma$-algebras. From this it follows immediately that:
\begin{rem} \label{rem: products}
The category of $\sigma$-varieties has products. Indeed, if $X$ and $Y$ are $\sigma$-varieties over $k$, then $k\{X\times Y\}=k\{X\}\otimes_k k\{Y\}$. Moreover, there is a terminal object, namely, the functor represented by the $k$-$\sigma$-algebra $k$.
\end{rem}
\subsection{Difference subvarieties and morphisms of difference varieties}
Let $Y$ be a $\sigma$-variety and $f\in k\{Y\}$. Then, for any $k$-$\sigma$-algebra $R$, we have a well--defined map $f\colon Y(R)\to R$ given by evaluating a representative of $f$ in $k\{y_1,\ldots,y_n\}$ at $a\in Y(R)\subset R^n$. In a coordinate free manner $f\colon Y(R)\to R$ can be described as the map that sends $\psi\in Y(R)=\operatorname{Hom}(k\{Y\},R)$ to $\psi(f)\in R$.
Let $X$ be a $\sigma$-closed $\sigma$-subvariety of $Y$. Then
\begin{equation} \label{eqn: I(X)2}
\mathbb{I}(X)=\{f\in k\{Y\}| \ f(a)=0 \ \text{ for all $k$-$\sigma$-algebras $R$ and all } a\in X(R) \}
\end{equation}
is a $\sigma$-ideal of $k\{Y\}$. We call $\mathbb{I}(X)\subset k\{Y\}$ the \emph{defining ideal of $X$} (in $k\{Y\}$).
This notation is consistent with (\ref{eqn: I(X)}) in the sense that $\mathbb{I}(X)$ as defined in (\ref{eqn: I(X)}) is the defining ideal of $X$ in $k\{y\}=k\{y_1,\ldots,y_n\}$. Moreover, $\mathbb{I}(X)\subset k\{Y\}$ agrees with the image in $k\{Y\}=k\{y\}/\mathbb{I}(Y)$ of the defining ideal of $X$ in $k\{y\}$. So $k\{X\}=k\{Y\}/\mathbb{I}(X)$.
Conversely, let $\mathfrak{a}\subset k\{Y\}$ be a $\sigma$-ideal. Then we can define a $\sigma$-closed $\sigma$-subvariety $\mathbb{V}(\mathfrak{a})$ of $Y$ by
$$\mathbb{V}(\mathfrak{a})(R)=\{a\in Y(R)|\ f(a)=0 \text{ for all } f\in\mathfrak{a}\}$$
for any $k$-$\sigma$-algebra $R$.
\begin{lemma} \label{lemma: correspondence sideal ssubvarriety}
Let $Y$ be a $\sigma$-variety. Then $\mathbb{I}$ and $\mathbb{V}$ are mutually inverse bijections between the set of $\sigma$-closed $\sigma$-subvarieties of $Y$ and the set of $\sigma$-ideals of $k\{Y\}$.
\end{lemma}
\begin{proof}
Let $\mathfrak{a}$ be a $\sigma$-ideal of $k\{Y\}$. Clearly $\mathfrak{a}\subset\mathbb{I}(\mathbb{V}(\mathfrak{a}))$. Since we may choose $R=k\{Y\}/\mathfrak{a}$ in (\ref{eqn: I(X)2}) it follows that $\mathfrak{a}=\mathbb{I}(\mathbb{V}(\mathfrak{a}))$.
Let $X$ be $\sigma$-closed $\sigma$-subvariety of $Y$. Then $X=\mathbb{V}(\mathfrak{a})$ for some $\sigma$-ideal $\mathfrak{a}$ of $k\{Y\}$. So $\mathbb{V}(\mathbb{I}(X))=\mathbb{V}(\mathbb{I}(\mathbb{V}(\mathfrak{a})))=\mathbb{V}(\mathfrak{a})=X$.
\end{proof}
Note that if $X$ is a $\sigma$-closed $\sigma$-subvariety of $Y$ and $R$ a $k$-$\sigma$-algebra, then $X(R)\subset Y(R)$ corresponds to $\{\psi\in\operatorname{Hom}(k\{Y\},R)|\ \mathbb{I}(X)\subset\ker(\psi)\}\subset\operatorname{Hom}(k\{Y\},R)$.
\medskip
If $Y$ and $Z$ are $\sigma$-closed $\sigma$-subvarieties of a $\sigma$-variety $X$, then we can define a subfunctor
$$Y\cap Z$$
of $X$ by $R\rightsquigarrow Y(R)\cap Z(R)$. Then $Y\cap Z$ is a $\sigma$-closed $\sigma$-subvariety of $X$, indeed, $\mathbb{I}(Y\cap Z)\subset k\{X\}$ is the ideal generated by $\mathbb{I}(Y)$ and $\mathbb{I}(Z)$.
\medskip
If $\phi\colon X\to Y$ is a morphism of $\sigma$-varieties and $Z\subset Y$ a $\sigma$-closed $\sigma$-subvariety, we can define a subfunctor
$$\phi^{-1}(Z)$$
of $X$ by $R\rightsquigarrow\phi_R^{-1}(Z(R))$. If $Z=\mathbb{V}(\mathfrak{a})$ with $\mathfrak{a}\subset k\{Y\}$, then
\begin{align*}
\phi^{-1}(Z)(R) & =\{\psi\in\operatorname{Hom}(k\{X\},R)|\ \mathfrak{a}\subset\ker(\psi\phi^*)\} \\
& = \{\psi\in\operatorname{Hom}(k\{X\},R)|\ \phi^*(\mathfrak{a})\subset\ker(\psi)\}=\mathbb{V}(\phi^*(\mathfrak{a}))(R).
\end{align*}
Therefore $\phi^{-1}(Z)=\mathbb{V}(\phi^*(\mathfrak{a}))$ is a $\sigma$-closed $\sigma$-subvariety of $X$.
If $\phi\colon X\to Y$ is a morphism of $\sigma$-varieties, we can define a subfunctor
$$\operatorname{Im}(\phi)$$ of $Y$ by $\operatorname{Im}(\phi)(R)=\phi_R(X(R))$ for any $k$-$\sigma$-algebra $R$. In general, $\operatorname{Im}(\phi)$ will not be a $\sigma$-closed $\sigma$-subvariety of $Y$.
It is therefore sometimes helpful to use the following notion, analogous to the scheme theoretic image (\cite[Tag 01R5]{stacks-project}).
\begin{lemma} \label{lemma: f(X)}
Let $\phi\colon X\to Y$ be a morphism of $\sigma$-varieties. There exists a unique $\sigma$-closed $\sigma$-subvariety $$\phi(X)$$ of $Y$ with the following property. The morphism $\phi$ factors through $\phi(X)$ and if $Z\subset Y$ is a $\sigma$-closed $\sigma$-subvariety such that $\phi$ factors through $Z$ then $\phi(X)\subset Z$.
\end{lemma}
\begin{proof}
Let $\mathfrak{a}$ denote the kernel of $\phi^*\colon k\{Y\}\to k\{X\}$ and set $\phi(X)=\mathbb{V}(\mathfrak{a})\subset Y$. As $\phi^*$ factors through $k\{Y\}\to k\{\phi(X)\}=k\{Y\}/\mathfrak{a}$, we see that $\phi$ factors through $\phi(X)$. If $\phi$ factors through $Z$, i.e., $\phi^*$ factors through $k\{Y\}\to k\{Z\}=k\{Y\}/\mathbb{I}(Z)$, then clearly $\mathbb{I}(Z)\subset\mathfrak{a}$. So $\phi(X)\subset Z$.
\end{proof}
Note that $\phi(X)$ is the $\sigma$-closure of $\operatorname{Im}(\phi)$ in the sense that
$$\mathbb{I}(\phi(X))=\{f\in k\{Y\}|\ f(a)=0 \text{ for all $k$-$\sigma$-algebras $R$ and all } a\in\operatorname{Im}(\phi)(R)\}.$$
If $\phi\colon X\to Y$ is a morphism of $\sigma$-varieties and $Z\subset X$ is a $\sigma$-closed $\sigma$-subvariety, then we will write $\phi(Z)$ for the $\sigma$-closed $\sigma$-subvariety $\phi|_Z(Z)$ of $Y$ where $\phi|_Z\colon Z\to X\xrightarrow{\phi}Y$.
A morphism $\phi\colon X\to Y$ of $\sigma$-varieties is called a \emph{$\sigma$-closed embedding} if $\phi$ induces an isomorphism between $X$ and a $\sigma$-closed $\sigma$-subvariety of $Y$, i.e., $X\to \phi(X)$ is an isomorphism.
We write
$$\phi\colon X\hookrightarrow Y$$ to express that $\phi$ is a $\sigma$-closed embedding. In analogy to a well known result in algebraic geometry we have:
\begin{lemma} \label{lemma: sclosed embedding}
A morphism $\phi\colon X\to Y$ of $\sigma$-varieties is a $\sigma$-closed embedding if and only if $\phi^*\colon k\{Y\}\to k\{X\}$ is surjective.
\end{lemma}
\begin{proof}
Let $\mathfrak{a}$ denote the kernel of $\phi^*\colon k\{Y\}\to k\{X\}$. The dual map to $X\to\phi(X)$ is $k\{Y\}/\mathfrak{a}\to k\{X\}$. It is an isomorphism if and only if $\phi^*$ is surjective.
\end{proof}
\subsection{Base extension}
In this subsection we show how to extend the base $\sigma$-field of a $\sigma$-variety. This is already one of the first points where our more general definition of $\sigma$-varieties pays of. For $\sigma$-varieties in the classical sense of \cite{Levin:difference} and \cite{Cohn:difference} base extension is not at all well--behaved. For example, the system of algebraic difference equations
$$y^2=2,\ \sigma(y)=-y$$ clearly has a solution in a $\sigma$-field extension of $\mathbb{Q}$, where we consider $\mathbb{Q}$ as a $\sigma$-field via the identity map.
However, if we consider the system over the $\sigma$-field $\mathbb{C}$, where again $\sigma$ is the identity map on $\mathbb{C}$, then the system has no solution in a $\sigma$-field extension of $\mathbb{C}$. See Chapter 8, Section 6 in \cite{Cohn:difference} for more details on base extension for $\sigma$-varieties in the classical sense. Since we consider solutions in arbitrary $k$-$\sigma$-algebras and not just in $\sigma$-field extensions of $k$ these problems disappear.
\medskip
Let $k$ be a $\sigma$-field and $X$ a $\sigma$-variety over $k$. Moreover, let $k'$ be a $\sigma$-field extension of $k$. We can define a functor
$$X_{k'}$$
from $k'$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$ by $X_{k'}(R')=X(R')$ for any $k'$-$\sigma$-algebra $R'$. Then $X_{k'}$ is a $\sigma$-variety over $k'$. Indeed,
$$k'\{X_{k'}\}=k\{X\}\otimes_k k'.$$
In terms of equations, this of course simply means that a system $F\subset k\{y\}=k\{y_1,\ldots,y_n\}$ of algebraic difference equations over $k$ is considered as a system $F\subset k'\{y\}$ of algebraic difference equations over $k'$. So if $k\{X\}=k\{y\}/[F]$ then $k'\{X_{k'}\}=k'\{y\}/[F]$.
\subsection{Zariski closures} \label{subsec: Zariski closures}
In this subsection we show how a variety over a $\sigma$-field can be considered as a $\sigma$-variety. We also introduce the important Zariski closures of a $\sigma$-subvariety of a variety. Cf. Sections A.4 and A.5 in \cite{diVizioHardouinWibmer:DifferenceGaloisofDifferential}.
Let $k$ be a $\sigma$-field and $\mathcal{X}$ a variety over $k$, where, for our purposes, a variety over $k$ is an affine scheme of finite type over $k$.
For a $k$-$\sigma$-algebra $R$ let
$$R^\sharp$$ denote the $k$-algebra obtained from $R$ by forgetting $\sigma$.
We can define a functor $[\sigma]_k\mathcal{X}$ from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$ by
$$[\sigma]_k\mathcal{X}(R)=\mathcal{X}(R^\sharp)$$
for any $k$-$\sigma$-algebra $R$. Our next goal is to show that $[\sigma]_k\mathcal{X}$ is $\sigma$-variety over $k$. That is, we need to construct $k\{[\sigma]_k\mathcal{X}\}$. By doing so, we also introduce some notations that will be useful later on.
Let $A$ be a $k$-algebra. For every $i\geq 0$ let
$${}^{\sigma^i\!} A=A\otimes_k k,$$ where the tensor product is formed by using $\sigma^i\colon k\to k$ on the right hand side. We consider ${}^{\sigma^i\!}A$ as $k$-algebra via the right factor. We set
$$A[i]=A\otimes_k{}^{\sigma\!}A\otimes_k \cdots\otimes_k{}^{\sigma^i\!}A.$$
We have inclusions $A[i]\hookrightarrow A[i+1]$ of $k$-algebras and the limit, i.e, the union
$$[\sigma]_kA$$
of the $A[i]$'s is a $k$-$\sigma$-algebra where for $(r_0\otimes\lambda_0)\otimes\cdots\otimes(r_i\otimes\lambda_i)\in {}^{\sigma^0\!}A\otimes_k \cdots\otimes_k{}^{\sigma^i\!}A=A[i]$ the map $\sigma\colon [\sigma]_kA\to [\sigma]_kA$ is given by
$$\sigma((r_0\otimes\lambda_0)\otimes\cdots\otimes (r_i\otimes\lambda_i))=(1\otimes 1)\otimes(r_0\otimes\sigma(\lambda_0))\otimes \cdots \otimes(r_i\otimes\sigma(\lambda_i))\in A[i+1].$$
The inclusion $A=A[0]\hookrightarrow [\sigma]_kA$ is characterized by the following universal property.
\begin{lemma} \label{lemma: universal property of sA}
For every $k$-algebra $A$ there exists a $k$-$\sigma$-algebra $[\sigma]_kA$ and a morphism $\psi\colon A\to [\sigma]_kA$ of $k$-algebras such that for every $k$-$\sigma$-algebra $R$ and every morphism $\psi'\colon A\to R$ of $k$-algebras there exists a unique morphism $\varphi\colon [\sigma]_kA\to R$ of $k$-$\sigma$-algebras making
\[
\xymatrix{
A \ar[rr]^-\psi \ar[rd]_{\psi'} & & [\sigma]_kA \ar@{..>}[ld]^\varphi \\
& R &
}
\]
commutative. \qed
\end{lemma}
In other words,
\begin{equation} \label{eqn: adjoin} \operatorname{Hom}([\sigma]_kA,R)\simeq\operatorname{Hom}(A,R^\sharp)\end{equation}
and $[\sigma]_k$ is left adjoint to the forgetful functor $(-)^\sharp$.
\begin{ex}
If $A=k[y_1,\ldots,y_n]$, then $[\sigma]_k A=k\{y_1,\ldots,y_n\}$. More generally, if $A=k[y_1,\ldots,y_n]/(F)$, then $[\sigma]_k A=k\{y_1,\ldots,y_n\}/[F]$.
\end{ex}
Let $k[\mathcal{X}]$ denote the coordinate ring of the variety $\mathcal{X}=\operatorname{Hom}(k[\mathcal{X}],-)$. Then (\ref{eqn: adjoin}) with $A=k[\mathcal{X}]$ shows that $[\sigma]_k\mathcal{X}$ is represented by $[\sigma]_k k[\mathcal{X}]$. If $M\subset k[\mathcal{X}]$ generates $k[\mathcal{X}]$ as a $k$-algebra, then $M\subset [\sigma]_k k[\mathcal{X}]$ generates $[\sigma]_k k[\mathcal{X}]$ as a $k$-$\sigma$-algebra. Therefore $[\sigma]_k\mathcal{X}$ is a $\sigma$-variety over $k$. Indeed, $k\{[\sigma]_k\mathcal{X}\}=[\sigma]_k k[\mathcal{X}]$. In the sequel, if confusion is unlikely we will often write $\mathcal{X}$ instead of $[\sigma]_k\mathcal{X}$. In particular, we will write $k\{\mathcal{X}\}$ instead of $k\{[\sigma]_k\mathcal{X}\}$ and by a $\sigma$-closed $\sigma$-subvariety of a variety $\mathcal{X}$, we mean a $\sigma$-closed $\sigma$-subvariety of $[\sigma]_k\mathcal{X}$.
Of course $[\sigma]_k$ is a functor from the category of varieties over $k$ to the category of $\sigma$-varieties over $k$: If $\phi\colon\mathcal{X}\to\mathcal{Y}$ is a morphism of varieties, then we can define $[\sigma]_k(\phi)\colon[\sigma]_k\mathcal{X}\to[\sigma]_k\mathcal{Y}$ by $$([\sigma]_k(\phi))_R\colon([\sigma]_k\mathcal{X})(R)=\mathcal{X}(R^\sharp)\xrightarrow{\phi_{R^\sharp}}\mathcal{Y}(R^\sharp)=([\sigma]_k\mathcal{Y})(R)$$
for any $k$-$\sigma$-algebra $R$.
\medskip
Next we will introduce the Zariski closures of a $\sigma$-closed $\sigma$-subvariety of a variety. We will use notations for varieties similar to the ones introduced for $k$-algebras above: If $\mathcal{X}$ is a variety over $k$ and $i\geq 1$, then
$${}^{\sigma^i\!}\mathcal{X}$$
denotes the variety over $k$ obtained from $\mathcal{X}$ by base extension via $\sigma^i\colon k\to k$. Similarly, if $\phi\colon \mathcal{X}\to \mathcal{Y}$ is a morphism of varieties over $k$, then ${}^{\sigma^i\!}\phi\colon{}^{\sigma^i\!}X\to{}^{\sigma^i\!}\mathcal{Y}$ denotes the morphism of varieties over $k$ obtained from $\phi$ by base extension via $\sigma^i\colon k\to k$. We also set
$$\mathcal{X}[i]=\mathcal{X}\times^\sigma\!\mathcal{X}\times\cdots\times{}^{\sigma^i\!}\mathcal{X}.$$
Let $Y$ be a $\sigma$-closed $\sigma$-subvariety of $\mathcal{X}$. Then $Y$ is defined by a $\sigma$-ideal $\mathbb{I}(Y)\subset k\{\mathcal{X}\}=\cup_{i\geq 0}k[\mathcal{X}[i]]$. For $i\geq 0$, the closed subvariety
$$Y[i]$$ of $\mathcal{X}[i]$ defined by $\mathbb{I}(Y)\cap k[\mathcal{X}[i]]\subset k[\mathcal{X}[i]]$ is called the \emph{$i$-th order Zariski closure of $Y$ in $\mathcal{X}$}.
The $0$-th order Zariski closure of $Y$ in $\mathcal{X}$ is also \emph{the Zariski closure} of $Y$ in $\mathcal{X}$. We say that $Y$ is \emph{Zariski dense} in $\mathcal{X}$ if the Zariski closure of $Y$ in $\mathcal{X}$ equals $\mathcal{X}$. Note that $Y$ is Zariski dense in $\mathcal{X}$ if and only if $k[\mathcal{X}]\to k\{Y\}$ is injective.
For $i\geq 1$ we have morphisms of varieties
$$\pi_i\colon\mathcal{X}[i]\to\mathcal{X}[i-1],\ (x_0,\ldots,x_i)\mapsto (x_0,\ldots,x_{i-1})$$ and
$$\sigma_i\colon\mathcal{X}[i]\to{^\sigma\!(\mathcal{X}[i-1])}={{}^{\sigma}\!\mathcal{X}}\times\cdots\times{{}^{\sigma^i}\!\mathcal{X}},\ (x_0,\ldots,x_i)\mapsto (x_1,\ldots,x_i)$$ that form a commutative diagram:
\[
\xymatrix{
\mathcal{X}[i] \ar_{\sigma_i}[d] \ar^-{\pi_i}[r] & \mathcal{X}[i-1] \ar^{\sigma_{i-1}}[d] \\
{}^{\sigma\!}(\mathcal{X}[i-1]) \ar^-{{}^{\sigma\!}\pi_{i-1}}[r] & {}^{\sigma\!}(\mathcal{X}[i-2]) \\
}
\]
There are induced morphisms $\pi_i\colon Y[i]\to Y[i-1]$ of varieties over $k$, and since $\mathbb{I}(Y)\subset k\{\mathcal{X}\}$ is a $\sigma$-ideal, we also have induced morphisms $\sigma_i\colon Y[i]\to {}^{\sigma\!}(Y[i-1])$ of varieties over $k$.
\section{Basic definitions}
In this section we introduce $\sigma$-algebraic groups and give some examples. We also show that every $\sigma$-algebraic group is isomorphic to a $\sigma$-closed subgroup of the general linear group.
From now on we will work over a fixed $\sigma$-field $k$. By an algebraic group over $k$, we mean an affine group scheme of finite type over $k$. (So an algebraic group need not be smooth.) Since the category of $\sigma$-varieties has products and a terminal object (Remark \ref{rem: products}) the following definition makes sense.
\begin{defi}
Let $k$ be a $\sigma$-field. A \emph{$\sigma$-algebraic group} over $k$ is a group object in the category of $\sigma$-varieties over $k$.
\end{defi}
Alternatively, we could define a $\sigma$-algebraic group as functor from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Groups}$, the category of groups, which is representable by a finitely $\sigma$-generated $k$-$\sigma$-algebra. A \emph{morphism of $\sigma$-algebraic groups} is of course a morphisms of $\sigma$-varieties that respects the group structure. A \emph{$\sigma$-closed embedding} of $\sigma$-algebraic groups is a morphism of $\sigma$-algebraic groups which is a $\sigma$-closed embedding of $\sigma$-varieties.
\begin{defi}
Let $k$ be a $\sigma$-field. A $k$-$\sigma$-Hopf algebra is a Hopf algebra $A$ over $k$ with the structure of a $k$-$\sigma$-algebra such that the Hopf-algebra structure maps $\Delta\colon A\to A\otimes_k A$, $S\colon A\to A$ and $\varepsilon\colon A\to k$ are morphisms of $k$-$\sigma$-algebras.
\end{defi}
A \emph{$k$-$\sigma$-Hopf subalgebra} of a $k$-$\sigma$-Hopf algebra is a Hopf subalgebra which is a $k$-$\sigma$-subalgebra.
\begin{prop} \label{prop: equivalence sgroups sHopfalgebras}
The category of $\sigma$-algebraic groups over $k$ is anti--equivalent to the category of $k$-$\sigma$-Hopf algebras that are finitely $\sigma$-generated over $k$.
\end{prop}
\begin{proof}
This follows from Remark \ref{rem: equivalence of categories for svarieties} in an analogous fashion to Theorem 1.4 in \cite{Waterhouse:IntroductiontoAffineGroupSchemes}.
\end{proof}
Let $G$ be a $\sigma$-algebraic group. A \emph{$\sigma$-closed subgroup $H$ of $G$} is a $\sigma$-closed $\sigma$-subvariety $H$ of $G$ such that $H(R)$ is a subgroup of $G(R)$ for any $k$-$\sigma$-algebra $R$. Then $H$ itself is a $\sigma$-algebraic group. We write
$$H\leq G$$
to express that $H$ is a $\sigma$-closed subgroup of $G$. If $H_1$ and $H_2$ are $\sigma$-closed subgroups of $G$, then $H_1\cap H_2$ is also a $\sigma$-closed subgroup of $G$.
Let $A$ be $k$-$\sigma$-Hopf algebra. A Hopf ideal of $A$ is called a \emph{$\sigma$-Hopf ideal} if it is a $\sigma$-ideal. In analogy to Section 2.1 in \cite{Waterhouse:IntroductiontoAffineGroupSchemes} it follows from Lemma \ref{lemma: correspondence sideal ssubvarriety} that:
\begin{lemma} \label{lemma: sclosed sungroup corresponds to sHopfideal}
Let $G$ be a $\sigma$-algebraic $G$. There is a one--to--one correspondence between the $\sigma$-closed subgroups of $G$ and the $\sigma$-Hopf ideals in $k\{G\}$. \qed
\end{lemma}
\begin{ex}
Let $V$ be a finite dimensional $k$-vector space. For every $k$-$\sigma$-algebra $R$ let $\operatorname{GL}(V)(R)$ denote the group of $R$-linear automorphisms of $V\otimes_k R$. Then $\operatorname{GL}(V)$ is naturally a functor from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Groups}$. Indeed, $\operatorname{GL}(V)$ is a $\sigma$-algebraic group: By choosing a basis $v_1,\ldots,v_n$ of $V$, we can identify $\operatorname{GL}(V)(R)$ with $\operatorname{GL}_n(R)$ and $\operatorname{GL}(V)$ is represented by
$$k\{\operatorname{GL}_n\}=k\{x_{ij},\tfrac{1}{\det(x)}\}.$$
\end{ex}
More generally:
\begin{ex}
Every algebraic group over $k$ can be interpreted as a $\sigma$-algebraic group over $k$. Here, as throughout the text, by an algebraic group over $k$ we mean an affine group scheme of finite type over $k$. Indeed, let $\mathcal{G}$ be an algebraic group over $k$ and as in Section \ref{subsec: Zariski closures}, let $R^\sharp$ denote the $k$-algebra obtained from the $k$-$\sigma$-algebra $R$ by forgetting $\sigma$. Then
$$R\rightsquigarrow \mathcal{G}(R^\sharp)$$
is a functor from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Groups}$, i.e., $[\sigma]_k\mathcal{G}$ is a $\sigma$-algebraic group.
\end{ex}
We will often write $\mathcal{G}$ instead of $[\sigma]_k\mathcal{G}$. In particular, by a $\sigma$-closed subgroup of $\mathcal{G}$ we mean a $\sigma$-closed subgroup of $[\sigma]_k\mathcal{G}$.
\begin{ex}
Let $0\leq \alpha_1<\cdots<\alpha_n$ and $1\leq\beta_1,\ldots,\beta_n$ be integers. We can define a $\sigma$-closed subgroup $G$ of the multiplicative group $\mathbb{G}_m$ by $$G(R)=\{g\in R^\times|\ \sigma^{\alpha_1}(g)^{\beta_1}\cdots\sigma^{\alpha_n}(g)^{\beta_n}=1\}\leq\mathbb{G}_m(R)$$
for every $k$-$\sigma$-algebra $R$.
\end{ex}
\begin{ex}
Every homogeneous linear difference equation $\sigma^n(y)+\lambda_{n-1}\sigma^{n-1}(y)+\cdots+\lambda_0y=0$ over $k$ defines a $\sigma$-closed subgroup $G$ of the additive group $\mathbb{G}_a$, i.e.,
$$G(R)=\{g\in R| \ \sigma^n(g)+\lambda_{n-1}\sigma^{n-1}(g)+\cdots+\lambda_0g=0\}\leq \mathbb{G}_a(R)$$
for any $k$-$\sigma$-algebra $R$.
\end{ex}
\begin{ex} \label{ex: unitary group}
The equations of the unitary group define a $\sigma$-closed subgroup of the general linear group:
$$G(R)=\{g\in\operatorname{GL}_n(R)|\ g\sigma(g)^{\operatorname{T}}=\sigma(g)^{\operatorname{T}}g=I_n\}\leq\operatorname{GL}_n(R)$$
for any $k$-$\sigma$-algebra $R$.
\end{ex}
Example \ref{ex: unitary group} can be generalized as follows:
\begin{ex}
Let $k$ be a $\sigma$-field, $n\geq 1$ an integer, $\mathcal{G}$ an algebraic group over $k$ and $\phi\colon \mathcal{G}\to{}^{\sigma^n\!}\mathcal{G}$ a morphism of algebraic groups over $k$.
There is a morphism $\sigma^n\colon \mathcal{G}\to{}^{\sigma^n\!}\mathcal{G}$ of $\sigma$-algebraic groups over $k$, which, in terms of equations is given by applying $\sigma^n$ to the coordinates. In terms of $k$-$\sigma$-algebras this morphism can be described as ${}^{\sigma^n\!}(k\{\mathcal{G}\})=k\{\mathcal{G}\}\otimes_k k\to k\{\mathcal{G}\}\ f\otimes\lambda\mapsto\sigma^n(f)\lambda$. We can define a $\sigma$-closed subgroup $G$ of $\mathcal{G}$ by
$$G(R)=\{g\in\mathcal{G}(R)|\ \sigma^n(g)=\psi(g)\}\leq\mathcal{G}(R)$$
for any $k$-$\sigma$-algebra $R$.
\end{ex}
\begin{ex} \label{ex: Frobenius algebraic group}
Let $k$ be a field of positive characteristic $p$ and $q$ be a power of $p$. Consider $k$ as a $\sigma$-field via the Frobenius, i.e., $\sigma(\lambda)=\lambda^q$ for $\lambda\in k$. An algebraic group $\mathcal{G}$ over $k$ can be considered as a $\sigma$-algebraic group over $k$: We turn $k[\mathcal{G}]$ into a $k$-$\sigma$-algebra by setting $\sigma(f)=f^q$ for $f\in k[G]$. Then clearly $k[\mathcal{G}]$ is a $k$-$\sigma$-Hopf algebra.
\end{ex}
A $\sigma$-algebraic group as in Example \ref{ex: Frobenius algebraic group} will be called a \emph{Frobenius algebraic group}.
\begin{ex} \label{ex: split finite}
Let $k$ be a $\sigma$-field, $\mathsf{G}$ a finite group and $\Sigma\colon \mathsf{G}\to \mathsf{G}$ a group endomorphism. Let $A$ be the $k$-vector space with basis $(e_g)_{g\in \mathsf{G}}$. We consider it as $k$-algebra via
$$(\sum \lambda_g e_g)(\sum \mu_g e_g)=\sum \lambda_g\mu_g e_g.$$ We can define a Hopf algebra structure on $A$ by
$$\Delta(e_g)=\sum_{(g_1,g_2)} e_{g_1}\otimes e_{g_2},$$ where the sum ranges over all pairs $(g_1,g_2)\in\mathsf{G}^2$ such that $g_1g_2=g$,
$S(e_g)=e_{g^{-1}}$ and $\varepsilon(e_g)=0$ for $g\neq 1$ and $\varepsilon(e_1)=1$. The algebraic group corresponding to $A$ is usually called the constant group scheme for $\mathsf{G}$. (See \cite[Section 2.3]{Waterhouse:IntroductiontoAffineGroupSchemes}.)
Extend $\sigma\colon k\to k$ to $\sigma\colon A\to A$ by $\sigma(e_g)=\sum_h e_{h}$, where the sum is taken over all $h\in\mathsf{G}$ such that $\Sigma(h)=g$. This defines a $k$-$\sigma$-algebra structure on $A$ and therefore we obtain an associated $\sigma$-algebraic group $G=\operatorname{Hom}(A,-)$.
\end{ex}
Similar to algebraic groups, $\sigma$-algebraic groups can be linearized:
\begin{theo} \label{theo: linearization}
Let $G$ be a $\sigma$-algebraic group over $k$. Then there exists a finite dimensional $k$-vector space $V$ and a $\sigma$-closed embedding
$G\hookrightarrow \operatorname{GL}(V)$. In particular, $G$ is isomorphic to a $\sigma$-closed subgroup of $\operatorname{GL}_{n}$ for some $n\geq 1$.
\end{theo}
\begin{proof}
Assume that $f_1,\ldots,f_m\in k\{G\}$ generate $k\{G\}$ as a $k$-$\sigma$-algebra. By \cite[Section 3.3, p. 24]{Waterhouse:IntroductiontoAffineGroupSchemes} there exists a Hopf subalgebra $A$ of $k\{G\}$ which contains $f_1,\ldots,f_m$ and is finitely generated as a $k$-algebra. By \cite[Section 3.4, p. 25]{Waterhouse:IntroductiontoAffineGroupSchemes} there exists an integer $n\geq 1$ and a surjective morphism $k[\operatorname{GL}_{n}]\to A$ of Hopf algebras. By Lemma \ref{lemma: universal property of sA} the morphism $k[\operatorname{GL}_{n}]\to A\hookrightarrow k\{G\}$ of $k$-algebras extends to a morphism
$k\{\operatorname{GL}_{n}\}\to k\{G\}$ of $k$-$\sigma$-algebras. Indeed, this is a morphism of $k$-$\sigma$-Hopf algebras and since $f_1,\ldots,f_m$ lie in the image, it is surjective. Now the claim follows from Lemma \ref{lemma: sclosed embedding}.
\end{proof}
\section{Zariski closures of difference algebraic groups and the growth group} \label{sec: Zariski closures of difference algebraic groups and the growth group}
In this section we show that the eventual growth of the Zariski closures of a $\sigma$-closed subgroup $G$ of an algebraic group $\mathcal{G}$ is governed by an algebraic group, called the growth group of $G$ (with respect to the $\sigma$-closed embedding $G\hookrightarrow\mathcal{G}$). We also introduce the $\sigma$-dimension and the order of a $\sigma$-algebraic group $G$.
Let $k$ be a $\sigma$-field and $\mathcal{G}$ an algebraic group over $k$. Then, for ever $i\geq 0$, ${^{\sigma^i}\!\mathcal{G}}$ and $\mathcal{G}[i]=\mathcal{G}\times{^{\sigma}\!\mathcal{G}}\times\cdots\times{^{\sigma^i}\!\mathcal{G}}$ (see Section \ref{subsec: Zariski closures}) are naturally algebraic groups over $k$.
For every $i\geq 1$ the maps $\pi_i\colon\mathcal{G}[i]\to\mathcal{G}[i-1]$ and $\sigma_i\colon\mathcal{G}[i]\to{^\sigma\!(\mathcal{G}[i-1])}$ are morphisms of algebraic groups over $k$.
Let $G$ be a $\sigma$-closed subgroup of $\mathcal{G}$. For every $i\geq 0$ we have an inclusion $k[\mathcal{G}[i]]\subset k\{\mathcal{G}\}$ of Hopf algebras. Since $\mathbb{I}(G)\subset k\{\mathcal{G}\}$ is a Hopf ideal, $\mathbb{I}(G)\cap k[\mathcal{G}[i]]$ is a Hopf ideal of $k[\mathcal{G}[i]]$. So the $i$-th order Zariski closure $G[i]$ of $G$ in $\mathcal{G}$ is a closed subgroup of $\mathcal{G}[i]$. The maps $\pi_i\colon G[i]\to G[i-1]$ and $\sigma_i\colon G[i]\to{^\sigma\!(G[i-1])}$ are morphisms of algebraic groups over $k$ and form a commutative diagram:
\begin{equation} \label{eqn: comm diag pi si}
\xymatrix{
G[i] \ar_{\sigma_i}[d] \ar^-{\pi_i}[r] & G[i-1] \ar^{\sigma_{i-1}}[d] \\
{}^{\sigma\!}(G[i-1]) \ar^-{{}^{\sigma\!}\pi_{i-1}}[r] & {}^{\sigma\!}(G[i-2]) \\
}
\end{equation}
For $i\geq 1$ we set
$$\mathcal{G}_i=\ker(\pi_i)\leq G[i].$$ We also set $\mathcal{G}_0=G[0]$. Because of (\ref{eqn: comm diag pi si}) we have induced morphisms $\sigma_i\colon \mathcal{G}_i\to{^\sigma\!(\mathcal{G}_{i-1})}$ of algebraic groups over $k$.
\begin{prop} \label{prop: existence of growth group}
For every $i\geq 1$ the map $\sigma_i\colon \mathcal{G}_i\to{^\sigma\!(\mathcal{G}_{i-1})}$ is a closed embedding and there exists an integer $m\geq 1$ such that
$\sigma_i\colon \mathcal{G}_i\to{^\sigma\!(\mathcal{G}_{i-1})}$ is an isomorphism for every $i\geq m$.
\end{prop}
\begin{proof}
Let us start by describing $\sigma_i$ in algebraic terms. Assume that $a=(a_1,\ldots,a_n)$ generates $k[\mathcal{G}]$ as a $k$-algebra and let $\overline{a}$ denote the image of $a$ in $k\{G\}$. So $$k[\mathcal{G}]=k[a]\subset k[a,\sigma(a),\ldots]=k\{\mathcal{G}\}$$ and $k[G[i]]=k[\overline{a},\ldots,\sigma^i(\overline{a})]$. The morphism $\sigma_i\colon G[i]\to{^\sigma\!(G[i-1])}$ corresponds to the map
$${^\sigma\!\big(k[\overline{a},\ldots,\sigma^{i-1}(\overline{a})]\big)}\longrightarrow k[\overline{a},\ldots,\sigma^i(\overline{a})],\ f\otimes\lambda\mapsto \sigma(f)\lambda$$ and the morphism $\sigma_i\colon \mathcal{G}_i\to{^\sigma\!(\mathcal{G}_{i-1})}$ corresponds to the map
\begin{align*}{^\sigma\!\big(k[\overline{a},\ldots,\sigma^{i-1}(\overline{a})]\otimes_{k[\overline{a},\ldots,\sigma^{i-2}(\overline{a})]} k\big)} & \longrightarrow k[\overline{a},\ldots,\sigma^{i}(\overline{a})]\otimes_{k[\overline{a},\ldots,\sigma^{i-1}(\overline{a})]} k \\ (f\otimes\lambda)\otimes\mu & \longmapsto \sigma(f)\otimes\sigma(\lambda)\mu
\end{align*}
where the tensor products are formed in virtue of the counit $\varepsilon\colon k\{G\}\to k$. Since $k[\overline{a},\ldots,\sigma^{i}(\overline{a})]\otimes_{k[\overline{a},\ldots,\sigma^{i-1}(\overline{a})]} k$ is generated as a $k$-algebra by the image of $\sigma^i(\overline{a})$, the above map is clearly surjective. Thus $\sigma_i\colon \mathcal{G}_i\to{^\sigma\!(\mathcal{G}_{i-1})}$ is a closed embedding.
To prove the second claim of the proposition, let us first assume that $k$ is inversive. The map \begin{align*} \psi_i\colon k[\overline{a},\ldots,\sigma^{i-1}(\overline{a})]\otimes_{k[\overline{a},\ldots,\sigma^{i-2}(\overline{a})]} k & \longrightarrow k[\overline{a},\ldots,\sigma^{i}(\overline{a})]\otimes_{k[\overline{a},\ldots,\sigma^{i-1}(\overline{a})]} k\\ f\otimes\lambda& \longmapsto \sigma(f)\otimes\sigma(\lambda)
\end{align*} does not respect the $k$-algebra structure, but since $k$ is inversive, it is surjective. We thus have a descending chain of closed subschemes
$$\mathcal{G}_0\hookleftarrow\mathcal{G}_1\hookleftarrow\mathcal{G}_2\cdots.$$ Since $\mathcal{G}_0$ is of finite type over $k$, this sequence must stabilize. That is, there exists an integer $m\geq 1$ such that $\psi_i$ is bijective for every $i\geq m$. Since $k$ is inversive, $k[\mathcal{G}_i]\to{^\sigma\!(k[\mathcal{G}_i])},\ f\mapsto f\otimes 1$ is bijective. It follows that for $i\geq m$, the morphism ${^\sigma\!(k[\mathcal{G}_{i-1}])}\to k[\mathcal{G}_i]$ dual to $\sigma_i$ is an isomorphism, since it can be obtained as the composition ${^\sigma\!(k[\mathcal{G}_{i-1}])}\to k[\mathcal{G}_{i-1}]\xrightarrow{\psi} k[\mathcal{G}_i]$ of two bijective maps. This proves the proposition for $k$ inversive.
The general case can be reduced to the inversive case: Let $k^*$ denote the inversive closure of $k$ (\cite[Definition 2.1.6]{Levin:difference}). So, in particular, $k^*$ is an inversive $\sigma$-field extension of $k$. The formation of Zariski--closures and of the $\mathcal{G}_i$ is compatible with base extension. It therefore follows from the inversive case, that there exists an integer $m\geq 1$ such that for every $i\geq m$ the morphism $\sigma_i\colon \mathcal{G}_i\to{^\sigma\!(\mathcal{G}_{i-1})}$ becomes an isomorphism after base extension from $k$ to $k^*$. But then already $\sigma_i$ must be an isomorphism for $i\geq m$.
\end{proof}
Proposition \ref{prop: existence of growth group} allows us to associate to the inclusion $G\leq\mathcal{G}$ an algebraic group, which measures the (eventual) growth of the Zariski closures $G[i]$ of $G$ in $\mathcal{G}$.
\begin{defi}
Let $\mathcal{G}$ be an algebraic group and $G\leq\mathcal{G}$ a $\sigma$-closed subgroup. For $i\geq 1$ let $\mathcal{G}_i$ denote the kernel of the projection $\pi_i\colon G[i]\to G[i-1]$ between the Zariksi closures of $G$ in $\mathcal{G}$. Let $m\geq 0$ denote the smallest integer such that $\sigma_i\colon \mathcal{G}_i\to{^\sigma\!(\mathcal{G}_{i-1})}$ is an isomorphism for every $i>m$. Then $\mathcal{G}_m$ is called the \emph{growth group of $G$} with respect to the $\sigma$-closed embedding $G\hookrightarrow\mathcal{G}$.
\end{defi}
\begin{ex} \label{ex: ssubgroup of Gm}
Let $0\leq \alpha_1<\cdots<\alpha_n$ and $1\leq\beta_1,\ldots,\beta_n$ be integers and $G\leq\mathbb{G}_m$ the $\sigma$-closed subgroup of the multiplicative group $\mathbb{G}_m$ given by $$G(R)=\{g\in R^\times|\ \sigma^{\alpha_1}(g)^{\beta_1}\cdots\sigma^{\alpha_n}(g)^{\beta_n}=1\}$$
for every $k$-$\sigma$-algebra $R$. Then the growth group of $G$ with respect to the given embedding $G\hookrightarrow\mathbb{G}_m$ is $\mu_{\beta_n}$, where $\mu_{\beta_n}(R)=\{g\in R^\times|\ g^{\beta_n}=1\}$ for ever $k$-algebra $R$.
\end{ex}
The following example shows that the growth group does indeed depend on the embedding $G\hookrightarrow\mathcal{G}$ and therefore is not an invariant of $G$.
\begin{ex}
Let $G=\mathbb{G}_m$ (considered as a $\sigma$-algebraic group). For $n\geq 1$ the $\sigma$-closed embedding
$$G\to\mathbb{G}_m^2,\ g\mapsto (g,\sigma(g)^n)$$ identifies $G$ with the $\sigma$-closed subgroup $G\leq\mathbb{G}_m^2$ given by
$$G(R)=\{(g_1,g_2)\in\mathbb{G}_m(R)^2|\ \sigma(g_1)^n=g_2\}$$
for any $k$-$\sigma$-algebra $R$. The growth group of $G$ with respect to the embedding $G\hookrightarrow\mathbb{G}_m^2$ is $\mu_n\times\mathbb{G}_m$.
\end{ex}
Even though the growth group itself does depend on the chosen embedding $G\hookrightarrow\mathcal{G}$, it carries some information which only depends on $G$. We will now show that the dimension of the growth group does not depend on the embedding $G\hookrightarrow\mathcal{G}$. In Section \ref{sec: limit degree} we will see that also the size of the growth group is independent of the chosen embedding.
The following theorem can be seen as a group theoretic analog of the classical theorem on the existence of the so--called dimension polynomial of an extension of $\sigma$-fields (\cite[Theorem 4.2.1]{Levin:difference}). See also \cite[Lemma 4.21]{Hrushovski:elementarytheoryoffrobenius}.
\begin{theo} \label{theo: existence sdim}
Let $\mathcal{G}$ be an algebraic group and $G\leq\mathcal{G}$ a $\sigma$-closed subgroup. For $i\geq 0$ let $d_i=\dim(G[i])$ denote the dimension of the $i$-th order Zariski closure of $G$ in $\mathcal{G}$. Then there exist integers $d,e\geq0$ such that
$$d_i=d(i+1)+e \text{ for } i\gg 0.$$
The integer $d$ only depends on $G$ and not on the choice of $\mathcal{G}$ and the $\sigma$-closed embedding $G\hookrightarrow \mathcal{G}$. If $d=0$, the integer $e$ only depends on $G$ and not on the choice of $\mathcal{G}$ and the $\sigma$-closed embedding $G\hookrightarrow \mathcal{G}$.
\end{theo}
\begin{proof}
Let $\mathcal{G}_i$ denote the kernel of $G[i]\twoheadrightarrow G[i-1]$. Then $\dim(G[i])=\dim(G[i-1])+\dim(\mathcal{G}_i)$. Let $m\geq 0$ denote the smallest integer such that $\sigma_i\colon \mathcal{G}_i\to{^\sigma\!(\mathcal{G}_{i-1})}$ is an isomorphism for every $i>m$. It follows that for $i\gg 0$
\begin{align*}\dim(G[i]) &=\dim(\mathcal{G}_i)+\cdots+\dim(\mathcal{G}_0)=\\
& = (i-m+1)\dim(\mathcal{G}_m)+\dim(\mathcal{G}_{m-1})+\cdots+\dim(\mathcal{G}_0)=\\
& = (i+1)\dim(\mathcal{G}_m)+\dim(\mathcal{G}_{m-1})+\cdots+\dim(\mathcal{G}_0)-m\dim(\mathcal{G}_m)=\\
& = d(i+1)+e.
\end{align*}
It follows from Proposition \ref{prop: existence of growth group} that $e\geq 0$.
Let us now show that $d$ is independent of the chosen embedding $G\hookrightarrow\mathcal{G}$. So let $G\hookrightarrow \mathcal{G}'$ be another $\sigma$-closed embedding. Let $G[i]'$ denote the $i$-th order Zariski closure of $G$ in $\mathcal{G}'$ and let $d_i', d',e'$ have the analogous meaning. We have to show that $d=d'$.
Let $a$ be a finite tuple from $k\{G\}$ which generates $k[G[0]]\subset k\{G\}$ as a $k$-algebra. Similarly, let $a'$ be a finite tuple from $k\{G\}$ which generates $k[G[0]']\subset k\{G\}$ as a $k$-algebra. Then $k[G[i]]=k[a,\ldots,\sigma^i(a)]$ and there exists an integer $n\geq 0$ such that all coordinates of $a'$ lie in $k[G[n]]$. Then $$k[G[i]']=k[a',\ldots,\sigma^i(a')]\subset k[a,\ldots,\sigma^{n+i}(a)]=k[G[n+i]].$$ Therefore $d_i'\leq d_{n+i}$ and for $i\gg 0$ we have
$d'(i+1)+e'\leq d(n+i+1)+e$. Letting $i$ tend to infinity we find $d'\leq d$. By symmetry, $d'=d$.
Let us now assume that $d=0$. We have to show that $e$ does not depend on the choice of the embedding $G\hookrightarrow\mathcal{G}$. Do to this we will show that
\begin{equation} \label{eqn: order}
e=\max\big\{\dim(R)|\ R \text{ is a finitely generated $k$-subalgebra of $k\{G\}$}\big\}.
\end{equation}
For $i\gg0$ the finitely generated $k$-subalgebra $k[G[i]]$ of $k\{G\}$ has dimension $e$. Conversely, if $R$ is a finitely generated $k$-subalgebra of $k\{G\}$, then $R$ is contained in some $k[G[i]]$ and therefore $\dim(R)\leq e$. This proves (\ref{eqn: order}).
\end{proof}
\begin{defi}
Let $G$ be a $\sigma$-algebraic group. The integer $d\geq 0$ defined in Theorem \ref{theo: existence sdim} above is called the \emph{$\sigma$-dimension of $G$} and denoted by $$\sigma\text{-}\operatorname{dim}(G).$$
If $\sigma\text{-}\operatorname{dim}(G)=0$, the integer $e\geq 0$ defined in Theorem \ref{theo: existence sdim} is called the \emph{order of $G$} and denoted by $$\operatorname{ord}(G).$$ If $G$ has positive $\sigma$-dimension the order of $G$ is defined to be infinity.
\end{defi}
\begin{ex}
Let $\mathcal{G}$ be an algebraic group. Then $\sigma\text{-}\operatorname{dim}([\sigma]_k\mathcal{G})=\dim(\mathcal{G})$. Indeed, if we tautologically consider $G=[\sigma]_k\mathcal{G}$ as a $\sigma$-closed subgroup of $\mathcal{G}$, then $G[i]=\mathcal{G}\times{^{\sigma}\!\mathcal{G}}\times\cdots\times{^{\sigma^i}\!\mathcal{G}}$ and so $\dim(G[i])=\dim(\mathcal{G})(i+1)$ for every $i\geq 0$. This also shows that either $\operatorname{ord}(G)=\infty$ (if $\dim(\mathcal{G})>0$) or $\operatorname{ord}(G)=0$ if ($\dim(\mathcal{G})=0$).
\end{ex}
The following example does motivate the naming ``order''.
\begin{ex}
Let $f=\sigma^n(y)+\lambda_{n-1}\sigma^{n-1}(y)+\cdots+\lambda_0y$ be a linear difference equation and $G$ the $\sigma$-closed subgroup of $\mathbb{G}_a$ defined by $f$. Then
$\operatorname{ord}(G)=n$, i.e., the order of $G$ equals the order of $f$.
\end{ex}
A certain numerical invariant analogous to the order and called the total dimension has been introduced in \cite{Hrushovski:elementarytheoryoffrobenius} in a different setting. We have chosen to stick to the more traditional naming from \cite{Levin:difference} and \cite{Cohn:difference}, also for the sake of the beauty of the formulation of Theorem \ref{theo: Galoisapplication}. From the proof of Theorem \ref{theo: existence sdim}, we immediately obtain:
\begin{cor} \label{cor: sdim equals dim of growth group}
Let $G$ be a $\sigma$-algebraic group. Then the dimension of the growth group of $G$ with respect to some $\sigma$-closed embedding $G\hookrightarrow\mathcal{G}$ of $G$ into some algebraic group $\mathcal{G}$ equals the $\sigma$-dimension of $G$. In particular, the dimension of the growth group does not depend on the choice of the $\sigma$-closed embedding $G\hookrightarrow\mathcal{G}$. \qed
\end{cor}
To state the next proposition we need some more definitions. A $\sigma$-ring $R$ is called a \emph{$\sigma$-domain} if $R$ is an integral domain and $\sigma\colon R\to R$ is injective. If $R$ is a $\sigma$-domain, the field of fractions of $R$ is naturally a $\sigma$-field.
A $\sigma$-algebraic group $G$ is called \emph{integral} if $k\{G\}$ is an integral domain. A $\sigma$-algebraic group $G$ is called \emph{$\sigma$-integral} if $k\{G\}$ is a $\sigma$-domain. The \emph{$\sigma$-transcendence degree} of a $\sigma$-field extension $K|k$ is the largest integer $n\geq 1$ such that the $\sigma$-polynomial ring $k\{y_1,\ldots,y_n\}$ may be embedded into $K$. (If no such integer exists the $\sigma$-transcendence degree is infinite.) See Section 4.1 in \cite{Levin:difference} for more details on the $\sigma$-transcendence degree. The following proposition shows that our notions of dimension and order generalize the classical notions. (See page 394 in \cite{Levin:difference}.)
\begin{prop} \label{prop: characterize sdim}
If $G$ is a $\sigma$-integral $\sigma$-algebraic group, then the $\sigma$-dimension of $G$ equals the $\sigma$-transcendence degree of the field of fractions of $k\{G\}$ over $k$.
If $G$ is an integral $\sigma$-algebraic group, then the order of $G$ equals the transcendence degree of the field of fractions of $k\{G\}$ over $k$.
\end{prop}
\begin{proof}
Let us first assume that $G$ is $\sigma$-integral and let us fix a $\sigma$-closed embedding $G\hookrightarrow\mathcal{G}$. Assume that $a=(a_1,\ldots,a_n)$ generates $k[G[0]]\subset k\{G\}$ as a $k$-algebra. Let $K$ denote the field of fractions of $k\{G\}$. Since $G$ is $\sigma$-integral $K$ is naturally a $\sigma$-field extension of $k$ and $a$ generates $K$ as a $\sigma$-field extension of $k$. Since $\operatorname{trdeg}(k(a,\ldots,\sigma^i(a))|k)=\dim(G[i])$, we see that
$d(t+1)+e$ (with $d$ and $e$ as in Theorem \ref{theo: existence sdim}) is really just the difference dimension polynomial (Definition 4.2.2 in \cite{Levin:difference}) of the $\sigma$-field extension $K|k$ associated with $a$. It follows from Theorem 4.2.1 (iii) in \cite{Levin:difference} that $\sigma\text{-}\operatorname{dim}(G)=d=\sigma\text{-}\operatorname{trdeg}(K|k)$.
Now assume that $G$ is integral. Let $K$ denote the field of fractions of $k\{G\}$. If $\sigma\text{-}\operatorname{dim}(G)>0$, then it is clear from Theorem \ref{theo: existence sdim} that the transcendence degree of $K$ over $k$ is infinite. So $\operatorname{ord}(G)=\operatorname{trdeg}(K|k)$ in this case.
If $\sigma\text{-}\operatorname{dim}(G)=0$, the claim follows from equation (\ref{eqn: order}) above.
\end{proof}
The $\sigma$-dimension is invariant under extension of the base $\sigma$-field.
\begin{lemma} \label{lemma: sdim invariant under baseext}
Let $G$ be a $\sigma$-algebraic group and $k'$ a $\sigma$-field extension of $k$. Then $\sigma\text{-}\operatorname{dim}(G_{k'})=\sigma\text{-}\operatorname{dim}(G)$ and
$\operatorname{ord}(G_{k'})=\operatorname{ord}(G)$.
\end{lemma}
\begin{proof}
Let $\mathcal{G}$ be an algebraic group such that $G$ is a $\sigma$-closed subgroup of $\mathcal{G}$. Then $G_{k'}$ is a $\sigma$-closed subgroup of $\mathcal{G}_{k'}$ and for $i\geq 0$ the $i$-th order Zariski closure $G_{k'}[i]$ of $G_{k'}$ in $\mathcal{G}_{k'}$ is obtained from the $i$-th order Zariski closure $G[i]$ of $G$ in $\mathcal{G}$ by base extension, i.e, $G_{k'}[i]=G[i]_{k'}$. Now the claim of the lemma follows from the fact that the usual dimension is invariant under base extension.
\end{proof}
For later use we record:
\begin{lemma} \label{lemma: ord for product}
Let $G$ and $H$ be $\sigma$-algebraic groups. Then $G\times H$ is a $\sigma$-algebraic group with
$$\sigma\text{-}\operatorname{dim}(G\times H)=\sigma\text{-}\operatorname{dim}(G)+\sigma\text{-}\operatorname{dim}(H)$$ and
$$\operatorname{ord}(G\times H)=\operatorname{ord}(G)+\operatorname{ord}(H).$$
\end{lemma}
\begin{proof}
Let $\mathcal{G}$ and $\mathcal{H}$ be algebraic groups containing $G$ and $H$ as $\sigma$-closed subgroups respectively. Then $G\times H$ is a $\sigma$-closed subgroup of $\mathcal{G}\times \mathcal{H}$ and the claim reduces to the similar formula for algebraic groups.
\end{proof}
\section{Finiteness theorems}
There is no direct difference analog of Hilbert's basis theorem: There exist infinite strictly ascending chains of difference ideals in the $\sigma$-polynomial ring $k\{y_1,\ldots,y_n\}$. (See \cite[Example 3, p. 73]{Cohn:difference}.) The Ritt--Raudenbush basis theorem in difference algebra (\cite{RittRaudenbush:IdealTheoryandAlgebraicDifferenceEquations} or \cite[Theorem 2.5.11]{Levin:difference}) only asserts that every ascending chain of perfect difference ideals in $k\{y_1,\ldots,y_n\}$ is finite. However, in our group theoretic setup the situation is better behaved. Indeed, in this section we will prove two important finiteness theorems. Let $A$ be a finitely $\sigma$-generated $k$-$\sigma$-Hopf algebra. The first finiteness theorem (Theorem \ref{theo: finiteness1}) asserts that every $\sigma$-Hopf ideal of $A$ is finitely generated as a $\sigma$-ideal. The second finiteness theorem (Theorem \ref{theo: finiteness2}) asserts that every $k$-$\sigma$-Hopf subalgebra of $A$ is finitely $\sigma$-generated over $k$. In Section \ref{sec: components} we will prove a third finiteness theorem (Theorem \ref{theo: Hrushovskiconj}): The set of minimal prime $\sigma$-ideals of $A$ is finite.
\begin{theo} \label{theo: finiteness1}
Let $H$ be $\sigma$-algebraic group and $G\leq H$ a $\sigma$-closed subgroup. Then the defining ideal $\mathbb{I}(G)\subset k\{H\}$ of $G$ is finitely generated as a $\sigma$-ideal.
\end{theo}
\begin{proof}
We may embed $H$ as a $\sigma$-closed subgroup in some algebraic group $\mathcal{G}$. For example, we may choose $\mathcal{G}=\operatorname{GL}_n$ by Theorem \ref{theo: linearization}.
If the defining ideal of $G$ in $k\{\mathcal{G}\}$ is finitely $\sigma$-generated, then also the defining ideal of $G$ in $k\{H\}=k\{G\}/\mathbb{I}(H)$ is finitely $\sigma$-generated. We can therefore assume that $H=\mathcal{G}$.
As in Section \ref{sec: Zariski closures of difference algebraic groups and the growth group} let $\mathcal{G}_i$ denote the kernel of the projection $\pi_i\colon G[i]\twoheadrightarrow G[i-1]$ between the Zariski closures of $G$ in $\mathcal{G}$. By Proposition \ref{prop: existence of growth group} there exists an integer $m\geq 1$ such that
$\sigma_i\colon \mathcal{G}_i\to {^\sigma\!(\mathcal{G}_{i-1})}$ is an isomorphism for every $i> m$.
To prove the theorem we will show that $\mathbb{I}(G[m])=\mathbb{I}(G)\cap k[\mathcal{G}[m]]$ $\sigma$-generates $\mathbb{I}(G)\subset k\{\mathcal{G}\}=\cup_{i\geq 0}k[\mathcal{G}[i]]$. To do this it is sufficient to show that
\begin{equation} \label{eqn: sideals of Gi}
\mathbb{I}(G[i])=\big(\mathbb{I}(G[i-1]),\sigma(\mathbb{I}(G[i-1]))\big)\subset k[\mathcal{G}[i]]
\end{equation}
for $i> m$.
The ideal to the right--hand side of (\ref{eqn: sideals of Gi}) defines an algebraic group
$$\mathcal{H}_i=(G[i-1]\times{}^{\sigma^i}\!\mathcal{G})\cap (\mathcal{G}\times{^\sigma\!(G[i-1])})\leq \mathcal{G}[i]=\mathcal{G}\times{^{\sigma}\!\mathcal{G}}\times\cdots\times{^{\sigma^i}\!\mathcal{G}}.$$
Clearly $G[i]\leq\mathcal{H}_i$ and the projection $\pi_i\colon \mathcal{H}_i\to G[i-1]$ is surjective\footnote{A morphism $\mathcal{G}\to\mathcal{H}$ of algebraic groups, i.e., affine group schemes of finite type over $k$ is surjective if the dual map $k[\mathcal{H}]\to k[\mathcal{G}]$ is injective. See \cite[Chapter VII, Section 7]{Milne:BasicTheoryOfAffineGroupSchemes} for more details.}.
The kernel of $\pi_i\colon \mathcal{H}_i\twoheadrightarrow G[i-1]$ is $1\times {^\sigma\!(\mathcal{G}_{i-1})}$. Since $i>m$ we have $1\times {^\sigma\!(\mathcal{G}_{i-1})}=\mathcal{G}_i$. Thus the downwards arrows
in the commutative diagram
\[\xymatrix{
G[i] \ar@{^{(}->}[rr] \ar@{>>}_-{\pi_i}[rd] & & \mathcal{H}_i \ar@{>>}^-{\pi_i}[ld] \\
& G[i-1] &
}
\]
have the same kernel. This implies that $G[i]=\mathcal{H}_i$ and identity (\ref{eqn: sideals of Gi}) is proved.
\end{proof}
We have actually proved a slightly stronger statement which we record for later use.
\begin{cor} \label{cor: finiteness1}
Let $\mathcal{G}$ be an algebraic group and $G\leq\mathcal{G}$ a $\sigma$-closed subgroup. For $i\geq 0$ let $G[i]$ denote the $i$-th order Zariski closure of $G$ in $\mathcal{G}$. Then there exists an integer $m\geq 0$ such that $\mathbb{I}(G[i])=(\mathbb{I}(G[i-1],\sigma(\mathbb{I}(G[i-1]))$ for $i>m$, i.e.,
$G[i]=(G[i-1]\times{}^{\sigma^i}\!\mathcal{G})\cap (\mathcal{G}\times{^\sigma\!(G[i-1])}).$ \qed
\end{cor}
\begin{cor}
Every descending chain of $\sigma$-closed subgroups of a $\sigma$-algebraic group is finite.
\end{cor}
\begin{proof}
A descending chain $H_1\geq H_2\geq\cdots$ of $\sigma$-closed subgroups of a $\sigma$-algebraic group $G$ corresponds to an ascending chain
$\mathbb{I}(H_1)\subset\mathbb{I}(H_2)\subset\cdots$ of $\sigma$-Hopf ideals in $k\{G\}$. By Theorem \ref{theo: finiteness1} the union $\bigcup\mathbb{I}(H_i)$ (which corresponds to the intersection $\bigcap H_i$) is finitely generated as a $\sigma$-ideal. Thus there exists an integer $n\geq 1$ such that $\bigcup\mathbb{I}(H_i)=\mathbb{I}(H_n)$. Then $H_n=H_{n+1}=\cdots$.
\end{proof}
To prove the second finiteness theorem we need a simple lemma on $k$-$\sigma$-Hopf algebras.
\begin{lemma} \label{lemma: directed union for ksHopfalgebras}
Let $A$ be a $k$-$\sigma$-Hopf algebra. Then every finite subset of $A$ is contained in a finitely $\sigma$-generated $k$-$\sigma$-Hopf subalgebra.
\end{lemma}
\begin{proof}
By \cite[Section 3.3]{Waterhouse:IntroductiontoAffineGroupSchemes} a finite subset of $A$ is contained in a Hopf subalgebra $B$ which is finitely generated as a $k$-algebra. Then $k\{B\}\subset A$ is finitely $\sigma$-generated over $k$ and since the comultiplication and the antipode are $\sigma$-morphisms, $k\{B\}$ is a Hopf subalgebra.
\end{proof}
\begin{theo} \label{theo: finiteness2}
Let $A$ be a $k$-$\sigma$-Hopf algebra which is finitely $\sigma$-generated over $k$ and $B\subset A$ a $k$-$\sigma$-Hopf subalgebra. Then $B$ is finitely $\sigma$-generated over $k$.
\end{theo}
\begin{proof}
For a Hopf subalgebra $C$ of $A$ let $\mathfrak{m}_C\subset C$ denote the kernel of the counit $\varepsilon\colon C\to k$. The ideal $(\mathfrak{m}_B)\subset A$ is a $\sigma$-Hopf ideal. By Theorem \ref{theo: finiteness1} it is finitely $\sigma$-generated. So there exists a finite set $F\subset \mathfrak{m}_B$ such that $[F]=(\mathfrak{m}_B)$.
By Lemma \ref{lemma: directed union for ksHopfalgebras} there exists a finitely $\sigma$-generated $k$-$\sigma$-Hopf subalgebra $C$ of $B$ containing $F$.
Then $(\mathfrak{m}_C)=(\mathfrak{m}_B)$. By Corollary 3.10 in \cite{Takeuchi:ACorrespondenceBetweenHopfidealsAndSubHopfalgebras} the mapping $C\mapsto (\mathfrak{m}_C)$ from Hopf subalgebras to Hopf ideals is injective. Thus $B=C$ is finitely $\sigma$-generated over $k$.
\end{proof}
As an application of Corollary \ref{cor: finiteness1}, we will prove a dimension theorem, which will then be used in Section \ref{sec: JordanHoelder} in the proof of an analog of the Jordan--H\"{o}lder theorem. A dimension theorem for differential algebraic groups has been proved in \cite{Sit:TypicalDifferentialDimensionOfTheIntersectionOfLinearDifferentialAlgebraicGroups}.
It is interesting to note that the dimension theorem fails for difference varieties. See \cite[Chapter 8, Section 8]{Cohn:difference}. However, it holds for $\sigma$-algebraic groups:
\begin{theo} \label{theo: dimension theorem}
Let $H_1$ and $H_2$ be $\sigma$-closed subgroups of a $\sigma$-algebraic group $G$. Then
\begin{equation} \label{eqn: sdimintersect}
\sigma\text{-}\operatorname{dim}(H_1\cap H_2)+\sigma\text{-}\operatorname{dim}(G)\geq \sigma\text{-}\operatorname{dim}(H_1)+\sigma\text{-}\operatorname{dim}(H_2)\end{equation} and
\begin{equation} \label{eqn: orderintersect} \operatorname{ord}(H_1\cap H_2)+\operatorname{ord}(G) \geq \operatorname{ord}(H_1)+\operatorname{ord}(H_2).\end{equation}
\end{theo}
\begin{proof}
Let $\mathcal{G}$ be an algebraic group containing $G$ as a $\sigma$-closed subgroup. We consider the Zariski closures inside $\mathcal{G}$.
By Corollary \ref{cor: finiteness1} there exists an integer $m\geq 0$ such that
$$\mathbb{I}((H_1\cap H_2)[i])=\big(\mathbb{I}((H_1\cap H_2)[i-1]),\sigma(\mathbb{I}((H_1\cap H_2)[i-1]))\big)$$ for $i>m$. Since $\mathbb{I}(H_1\cap H_2)=\mathbb{I}(H_1)+\mathbb{I}(H_2)$ there exists $n\geq m$ such that $$\mathbb{I}((H_1\cap H_2)[m])\subset \mathbb{I}(H_1[n])+\mathbb{I}(H_2[n])=\mathbb{I}(H_1[n]\cap H_2[n]).$$ But then
\begin{equation} \label{eqn: proof}\mathbb{I}((H_1\cap H_2)[m+i])\subset \mathbb{I}(H_1[n+i]\cap H_2[n+i])\end{equation} for every $i\geq 0$ and so
$$H_1[n+i]\cap H_2[n+i]\leq (H_1\cap H_2)[m+i]\times{{}^{\sigma^{m+i+1}}\!\mathcal{G}}\times\cdots\times {}^{\sigma^{n+i}}\!\mathcal{G}.$$
Therefore
\begin{equation*} \dim(H_1[n+i]\cap H_2[n+i])\leq \dim((H_1\cap H_2)[m+i])+(n-m)\dim(\mathcal{G}) \end{equation*}
and so
\begin{align*}
\dim((H_1\cap & H_2)[m+i]) \geq \dim(H_1[n+i]\cap H_2[n+i])-(n-m)\dim(\mathcal{G}) \\
& \geq \dim(H_1[n+i])+\dim(H_2[n+i])-\dim(G[n+i])-(n-m)\dim(\mathcal{G}).
\end{align*}
Now using Theorem \ref{theo: existence sdim} and comparing the coefficients of $i$ yields identity (\ref{eqn: sdimintersect}).
It remains to prove (\ref{eqn: orderintersect}). Obviously this is true if $\operatorname{ord}(G)=\infty$. So we can assume $\operatorname{ord}(G)<\infty$, i.e., $\sigma\text{-}\operatorname{dim}(G)=0$. Note that (\ref{eqn: proof}) implies that the intersection of $\mathbb{I}(H_1[n+i]\cap H_2[n+i])$ with $k[\mathcal{G}[m+i]]$ equals $\mathbb{I}((H_1\cap H_2)[m+i])$. This means that the morphism of algebraic groups
$$\phi\colon H_1[n+i]\cap H_2[n+i]\to (H_1\cap H_2)[m+i]$$
induced from the projection $\mathcal{G}\times{{}^{\sigma}\!\mathcal{G}}\times\cdots\times {}^{\sigma^{n+i}}\!\mathcal{G}\to \mathcal{G}\times{{}^{\sigma}\!\mathcal{G}}\times\cdots\times {}^{\sigma^{m+i}}\!\mathcal{G}$ is
surjective. Since $\sigma\text{-}\operatorname{dim}(H_1)=0$, the kernel of the projections $H_1[n+i]\twoheadrightarrow H_1[m+i]$ is finite for $i\gg 0$. Therefore also $\phi$ has finite kernel and it follows that for $i\gg 0$
\begin{align*}\operatorname{ord}(H_1\cap H_2)&=\dim((H_1\cap H_2)[m+i]) =\dim(H_1[n+i]\cap H_2[n+i]) \\
& \geq \dim(H_1[n+i])+\dim(H_2[n+i])-\dim(G[n+i])\\
& =\operatorname{ord}(H_1)+\operatorname{ord}(H_2)-\operatorname{ord}(G).
\end{align*}
\end{proof}
\section{Representations of difference algebraic groups}
In this section we present an application of the first finiteness theorem (Theorem \ref{theo: finiteness1}). We prove the difference analog of a theorem of Chevalley for algebraic groups. Namely, we show that every $\sigma$-closed subgroup of a $\sigma$-algebraic group can be realized as the stabilizer of a line. We also show that the category of representations of $\mathbb{G}_m^n$ as a difference algebraic group is semi--simple.
Categories of representations of $\sigma$-algebraic groups have been studied in \cite{OvchinnikovWibmer:TannakianCategoriesWithSemigroupActions}. There the authors introduce $\sigma$-tannakian categories, which provide a purely categorical characterization of those categories, which are the category of representations of a $\sigma$-algebraic group. We plan to explore the relation between properties of a $\sigma$-algebraic group and properties of its category of representations in a future work.
\begin{defi}
Let $G$ be a $\sigma$-algebraic group. A \emph{representation of $G$} is a pair $(V,\phi)$ comprising a finite dimensional $k$-vector space $V$ and a morphism $\phi\colon G\to \operatorname{GL}(V)$ of $\sigma$-algebraic groups. The representation is called \emph{faithful} if $\phi$ is a $\sigma$-closed embedding.
\end{defi} We will often omit $\phi$ from the notation. A \emph{morphism $(V,\phi)\to (V',\phi')$ of representations of $G$} is a $k$-linear map $f\colon V\to V'$ which is $G$-equivariant, i.e.,
\[\xymatrix{
V\otimes_k R \ar^{f\otimes R}[r] \ar_{\phi(g)}[d] & V'\otimes_k R \ar^{\phi'(g)}[d] \\
V\otimes_k R \ar^{f\otimes R}[r] & V'\otimes_k R
}
\]
commutes for every $g\in G(R)$ and any $k$-$\sigma$-algebra $R$.
\begin{lemma}
Let $G$ be a $\sigma$-algebraic group. The category of representations of $G$ is equivalent to the category of finite dimensional comodules over $k\{G\}$.
\end{lemma}
\begin{proof}
This is similar to Section 3.2 in \cite{Waterhouse:IntroductiontoAffineGroupSchemes}.
\end{proof}
\begin{lemma} \label{lemma: stabilizer}
Let $V$ be a representation of the $\sigma$-algebraic group $G$ and $W\leq V$ a linear subspace. Then the stabilizer $G_W$ of $W$, defined by
$$G_W(R)=\{g\in G(R)|\ g(W\otimes_k R)\subset W\otimes_k R\}$$
for any $k$-$\sigma$-algebra $R$, is a $\sigma$-closed subgroup of $G$.
\end{lemma}
\begin{proof}
Let $v_1,\ldots,v_n$ be a basis of $V$ such that $v_1,\ldots,v_m$ is a basis of $W$. Let $\rho\colon V\to V\otimes_k k\{G\}$ denote the comodule structure corresponding to the given representation and write $\rho(v_j)=\sum_{i=1}^n v_i\otimes a_{ij}\in V\otimes_k k\{G\}$ for $j=1,\ldots,n$. So for a $k$-$\sigma$-algebra $R$ and $g\in G(R)=\operatorname{Hom}(k\{G\},R)$ we have $g(v_j\otimes 1)=\sum_{i=1}^n v_i\otimes g(a_{ij})\in V\otimes_k R$.
This shows the $g\in G_W(R)$ if and only if $g(a_{ij})=0$ for $j=1,\ldots,m$ and $i=m+1,\ldots,n$. Thus
$$G_W=\mathbb{V}([a_{ij}|\ 1\leq j\leq m,\ m+1\leq i\leq n ]).$$
\end{proof}
We know from Theorem \ref{theo: linearization} that every $\sigma$-algebraic group admits a faithful representation. The following theorem provides a strengthening of this result:
\begin{theo} \label{theo: Chevalley}
Let $G$ a $\sigma$-algebraic group and $H\leq G$ a $\sigma$-closed subgroup. Then there exists a faithful representation $V$ of $G$ and a line $L\leq V$ (i.e., a one dimensional $k$-subspace) such that $H$ is the stabilizer of $L$, i.e., $$H(R)=\{g\in G(R)|\ g(L\otimes_k R)\subset L\otimes_k R\}$$ for any $k$-$\sigma$-algebra $R$.
\end{theo}
\begin{proof}
We will first construct a representation $V$ of $G$ such that $H$ is the stabilizer of a subspace $W$ of $V$.
By Theorem \ref{theo: finiteness1} there exists a finite set $F\subset\mathbb{I}(H)\subset k\{G\}$ such that $\mathbb{I}(H)=[F]$. By Section 3.3 in \cite{Waterhouse:IntroductiontoAffineGroupSchemes} there exists a finite dimensional subcomodule $V$ of $k\{G\}$ such that $F\subset V$. Let $W=V\cap\mathbb{I}(H)\leq V$. If $v_1,\ldots,v_n$ is a basis of $V$ such that $v_1,\ldots,v_m$ is a basis of $W$ and $\Delta(v_j)=\sum_{i=1}^n v_i\otimes a_{ij}\in V\otimes_k k\{G\}$ for $j=1,\ldots,n$, then
$$\mathbb{I}(G_W)=[a_{ij}|\ 1\leq j\leq m,\ m+1\leq i\leq n ]$$ by the proof of Lemma \ref{lemma: stabilizer}. Since $\mathbb{I}(H)$ and $V$ are $k\{H\}$-comodules also $W=\mathbb{I}(H)\cap V$ is a $k\{H\}$-comodule. So $W$ is stable under $H$, i.e., $H\leq G_W$, or $\mathbb{I}(G_W)\subset\mathbb{I}(H)$.
We have $$v_j=\sum_{i=1}^n\varepsilon(v_i)a_{ij}=\sum_{i=m+1}^n\varepsilon(v_i)a_{ij},$$
since $\varepsilon(\mathbb{I}(H))=0$. This shows that every $f\in F\subset W$ lies in $[a_{ij}|\ 1\leq j\leq m,\ m+1\leq i\leq n ]=\mathbb{I}(G_W)$. Therefore $\mathbb{I}(H)=[F]\subset\mathbb{I}(G_W)$ and consequently $H=G_W$.
Now the $m$-th exterior power $\wedge^m V$ is naturally a representation of $G$ and $L=\wedge^mW\leq \wedge^m V$ is one dimensional. Moreover $G_L=G_W$. (See Section A.2 in \cite{Waterhouse:IntroductiontoAffineGroupSchemes}.)
In case $\wedge^m V$ is not faithful, we can replace it by $\wedge^m V\oplus V'$, where $V'$ is a faithful representation of $G$.
\end{proof}
\medskip
An algebraic group $\mathcal{G}$ over a field of characteristic zero is reductive if and only if it is linearly reductive, i.e., every representation of $\mathcal{G}$ is a direct sum of irreducible representations. In differential algebra, the only linearly reductive linear differential algebraic groups are those which are the constant points of a reductive algebraic group (\cite[Theorem 3.14]{MinchenkiOvchinnikov:ZariskiClosuresOfReductiveLinearDifferentialAlgebraiGroups}). In particular, the linear differential algebraic group $\mathbb{G}_m$ is not linearly reductive. Indeed, the representation $$\mathbb{G}_m\to\operatorname{GL}_2,\quad g\mapsto\begin{pmatrix} g & \delta(g) \\ 0 & g\end{pmatrix}$$ is not semi--simple. For difference algebraic groups the situation is fundamentally different. We show here that tori are linearly reductive difference algebraic goups and leave the characterization of the linearly reductive difference algebraic groups for the future.
\begin{theo}
Every representation of the difference algebraic group $G=\mathbb{G}_m^n$ is a direct sum of one--dimensional representations.
\end{theo}
\begin{proof}
Let us denote by $X$ the set of all elements of $k\{G\}=k\{y_1,y_1^{-1},\ldots,y_n,y_n^{-1}\}$ which are products of elements of the form $\sigma^\alpha(y_i)^\beta$ where $\alpha\in\mathbb{N},\ \beta\in\mathbb{Z}$ and $1\leq i\leq n$. Then $X$ is a $k$-basis of $k\{G\}$ and every element $\chi\in X$ is group--like, i.e., $\Delta(\chi)=\chi\otimes\chi$ and $\varepsilon(\chi)=1$ where $\Delta\colon k\{G\}\to k\{G\}\otimes_k k\{G\}$ is the comultiplication and $\varepsilon\colon k\{G\}\to k$ the counit.
Let $V$ be a representation of $G$, $\rho\colon V\to V\otimes_k k\{G\}$ the corresponding comodule and $v\in V$. Then we can write
$$\rho(v)=\sum_{\chi\in X} a_\chi\otimes \chi,\quad a_\chi\in V.$$
Applying the comodule identities $(\operatorname{id}_V\otimes\Delta)\circ \rho=(\rho\otimes \operatorname{id}_{k\{G\}})\circ \rho$ and $(\operatorname{id}_V\otimes\varepsilon)\circ \rho=\operatorname{id}_V$ (see \cite[Section 3.2]{Waterhouse:IntroductiontoAffineGroupSchemes}) to $v$, we find that
\begin{align}
\sum_{\chi\in X}a_\chi\otimes \chi\otimes \chi & =\sum_{\chi\in X} \rho(a_\chi)\otimes\chi, \quad \text{ and } \label{eqn: chi1}\\
v & = \sum_{\chi\in X} a_\chi. \label{eqn: chi2}
\end{align}
Identity (\ref{eqn: chi1}) implies that $\rho(a_\chi)=a_\chi\otimes \chi$. So $g(a_\chi\otimes 1)=a_\chi\otimes g(\chi)$ for $g\in G(R)=\operatorname{Hom}(k\{G\},R)$ and $R$ a $k$-$\sigma$-algebra. Thus $ka_\chi$ is a subrepresentation of $V$ and it follows form (\ref{eqn: chi2}) that $V$ is spanned by all the one--dimensional subrepresentations arising in this way.
\end{proof}
\section{The limit degree} \label{sec: limit degree}
In this section we introduce an important numerical invariant for $\sigma$-algebraic groups (of $\sigma$-dimension zero) called the limit degree. We also show that the category of algebraic $\sigma$-groups introduced and studied in \cite{KowalskiPillay:OnAlgebraicSigmaGroups} is equivalent to the category of $\sigma$-algebraic groups of $\sigma$-dimension zero and limit degree one.
By the \emph{size} $|\mathcal{G}|$ of an algebraic group $\mathcal{G}$ we mean the dimension of $k[\mathcal{G}]$ as a $k$-vector space. So the size is either a non--negative integer or $\infty$. In the sequel we will employ the usual rules for calculating with the symbol $\infty$.
If $\xymatrix@1{\mathcal{G}_1\ar@{->>}^-{\phi_1}[r] & \mathcal{G}_2 \ar@{->>}^-{\phi_2}[r] & \mathcal{G}_3}$ are surjective morphisms of algebraic groups, then
\begin{equation}\label{eqn: multiplicativity of size}
|\ker(\phi_2\circ\phi_1)|=|\ker(\phi_2)|\cdot |\ker(\phi_1)|.
\end{equation}
\begin{prop} \label{prop: existence of limit degree}
Let $G$ be a $\sigma$-algebraic group and $G\hookrightarrow\mathcal{G}$ a $\sigma$-closed embedding. Then the size of the growth group of $G$ with respect to the embedding $G\hookrightarrow\mathcal{G}$ does not depend on the choice of $\mathcal{G}$ and the $\sigma$-closed embedding.
\end{prop}
\begin{proof}
For $i\geq 0$ let $G[i]$ denote the $i$-th order Zariski closure of $G$ in $\mathcal{G}$ and $\mathcal{G}_i$ the kernel of the projection $\pi_i\colon G[i]\to G[i-1]$. By Proposition \ref{prop: existence of growth group} the integer $d=|\mathcal{G}_i|$ does not depend on $i$ for $i\gg0$. Let $a=(a_1,\ldots,a_n)$ generate $k[G[0]]\subset k\{G\}$ as a $k$-algebra. Then $a$ generates $k\{G\}$ as a $k$-$\sigma$-algebra.
Let $\mathcal{G}'$ be another algebraic group and $G\hookrightarrow\mathcal{G}'$ a $\sigma$-closed embedding. Let $G[i]'$ denote the $i$-th order Zariski closure of $G$ in $\mathcal{G}'$ and let $d'$ and $a'$ be as above. We have to show that $d=d'$.
Since $a$, as well as $a'$, $\sigma$-generate $k\{G\}$, there exists an integer $m\geq 1$ such that
$a'\in k[a,\ldots,\sigma^m(a)]$ and $a\in k[a',\ldots,\sigma^m(a')]$. Then, for $i\geq 0$, we have $k[a',\ldots,\sigma^i(a')]\subset k[a,\ldots,\sigma^{m+i}(a)]$ and $k[a,\ldots,\sigma^i(a)]\subset k[a',\ldots,\sigma^{m+i}(a')]$. So for $j\geq m$:
$$k[a,\ldots,\sigma^i(a)]\subset k[a',\ldots,\sigma^{m+i}(a')]\subset k[a',\ldots,\sigma^{j+i}(a')]\subset k[a,\ldots,\sigma^{m+j+i}(a)].$$
These inclusions of Hopf algebras correspond to surjective morphisms of algebraic groups
$$G[m+j+i]\twoheadrightarrow G[j+i]'\twoheadrightarrow G[m+i]'\twoheadrightarrow G[i].$$
We have $|\ker(G[m+j+i]\twoheadrightarrow G[i])|\geq |\ker(G[j+i]'\twoheadrightarrow G[m+i]')|$ by (\ref{eqn: multiplicativity of size}).
But by (\ref{eqn: multiplicativity of size}) and Proposition \ref{prop: existence of growth group} we also have $|\ker(G[m+j+i]\twoheadrightarrow G[i])|=d^{m+j}$ and $|\ker(G[j+i]'\twoheadrightarrow G[m+i]')|=d'^{j-m}$ for $i\gg 0$. Consequently, $d^{m+j}\geq d'^{j-m}$. Letting $j$ tend to infinity, we find $d\geq d'$. By symmetry, $d=d'$.
\end{proof}
\begin{defi}
Let $G$ be a $\sigma$-algebraic group. Choose an algebraic group $\mathcal{G}$ and a $\sigma$-closed embedding $G\hookrightarrow\mathcal{G}$. The size of the growth group of $G$ with respect to the $\sigma$-closed embedding $G\hookrightarrow\mathcal{G}$ is called the \emph{limit degree} of $G$ and is denoted by
$${\operatorname{ld}}(G).$$
By Proposition \ref{prop: existence of limit degree} the limit degree of $G$ does not depend on the choice of $\mathcal{G}$ and the $\sigma$-closed embedding $G\hookrightarrow\mathcal{G}$.
\end{defi}
The expression ``limit degree'' is motivated by the fact that
$${\operatorname{ld}}(G)=\lim_{i\to\infty}\deg(\pi_i),$$
where $\deg(\pi_i)$ denotes the degree of the projection $\pi_i\colon G[i]\twoheadrightarrow G[i-1]$. The naming is also motivated by Proposition \ref{prop: correspondence of ld} and Theorem \ref{theo: Galoisapplication}.
\begin{rem} \label{rem: ld finite}
By Corollary \ref{cor: sdim equals dim of growth group} the limit degree of a $\sigma$-algebraic group is finite if and only if it has $\sigma$-dimension zero.
\end{rem}
\begin{ex}
Let $0\leq \alpha_1<\cdots<\alpha_n$ and $1\leq\beta_1,\ldots,\beta_n$ be integers and $G\leq\mathbb{G}_m$ the $\sigma$-closed subgroup of the multiplicative group $\mathbb{G}_m$ given by $$G(R)=\{g\in R^\times|\ \sigma^{\alpha_1}(g)^{\beta_1}\cdots\sigma^{\alpha_n}(g)^{\beta_n}=1\}$$
for every $k$-$\sigma$-algebra $R$. Then ${\operatorname{ld}}(G)=\beta_n$ by Example \ref{ex: ssubgroup of Gm}.
\end{ex}
\begin{ex}
Let $\mathcal{G}$ be an algebraic group. Then ${\operatorname{ld}}([\sigma]_k\mathcal{G})=|\mathcal{G}|$. This follows from the fact that the growth group of $[\sigma]_k\mathcal{G}$ with respect to the tautological $\sigma$-closed embedding $[\sigma]_k\mathcal{G}\hookrightarrow\mathcal{G}$ is $\mathcal{G}$.
\end{ex}
Let $K|k$ be an extension of $\sigma$-fields. Assume that there exists a finite set $B\subset K$ such that $B,\sigma(B),\ldots$ generates $K$ as a field extension of $k$, then the limit degree ${\operatorname{ld}}(K|k)$ is the limit $\lim_{i\to\infty} d_i$, where $d_i$ is the degree of the field extension
$k(B,\ldots,\sigma^{i}(B))|k(B,\ldots,\sigma^{i-1}(B))$. The limit exists and does not depend on the choice of $B$ (\cite[Section 4.3]{Levin:difference}).
The following proposition shows that our definition of the limit degree generalizes the classical definition (\cite[p. 394]{Levin:difference}) of the limit degree (for irreducible difference varieties in the sense of \cite{Cohn:difference} and \cite{Levin:difference}).
\begin{prop} \label{prop: correspondence of ld}
Let $G$ be a $\sigma$-algebraic group. If $G$ is $\sigma$-integral (i.e., $k\{G\}$ is an integral domain and $\sigma\colon k\{G\}\to k\{G\}$ is injective), then the limit degree of $G$ equals the limit degree of the field of fractions of $k\{G\}$ over $k$.
\end{prop}
\begin{proof}
Let $\mathcal{G}$ be an algebraic group containing $G$ as a $\sigma$-closed subgroup. For $i\geq 0$ let $G[i]$ denote the $i$-th order Zariski closure of $G$ in $\mathcal{G}$ and let $B\subset k[G[0]]$ be a finite set which generates $k[G[0]]\subset k\{G\}$ as a $k$-algebra. Let $K$ denote the field of fractions of $k\{G\}$. Then $B,\sigma(B),\ldots$ generates $K$ as a field extension of $k$ and $k(B,\sigma(B),\ldots,\sigma^i(B))$ is the field of fractions of $k[G[i]]$. Therefore the degree of the field extension $k(B,\ldots,\sigma^{i}(B))|k(B,\ldots,\sigma^{i-1}(B))$ equals the degree of the projection $G[i]\twoheadrightarrow G[i-1]$.
\end{proof}
Our next aim is to characterize the algebraic $\sigma$-groups introduced in \cite{KowalskiPillay:OnAlgebraicSigmaGroups} within the category of $\sigma$-algebraic groups.
\begin{prop} \label{prop: ld equal 1}
Let $G$ be a $\sigma$-algebraic group. Then ${\operatorname{ld}}(G)=1$ if and only if $k\{G\}$ is finitely generated as a $k$-algebra.
\end{prop}
\begin{proof}
Fix a $\sigma$-closed embedding $G\hookrightarrow\mathcal{G}$ and let $G[i]$ denote the $i$-th order Zariski closure of $G$ in $\mathcal{G}$. So $k\{G\}=\cup_{i\geq 0}k[G[i]]$.
If ${\operatorname{ld}}(G)=1$ there exists an integer $m\geq 0$ such that $\pi_i\colon G[i]\twoheadrightarrow G[i-1]$ has trivial kernel for $i> m$, so $\pi_i$ is an isomorphism and therefore $k[G[i]]=k[G[i-1]]$. Consequently, $k\{G\}=k[G[m]]$ is finitely generated as a $k$-algebra.
Conversely, if $k\{G\}$ is a finitely generated $k$-algebra, we can consider the algebraic group $\mathcal{G}$ associated with the Hopf algebra $k\{G\}^\sharp$.
So $k[\mathcal{G}]=k\{G\}^\sharp$. Let $R$ be a $k$-$\sigma$-algebra. Since $G(R)=\operatorname{Hom}(k\{G\},R)$ is a subgroup of $$\operatorname{Hom}(k\{G\}^\sharp,R^\sharp)=\mathcal{G}(R^\sharp)=([\sigma]_k\mathcal{G})(R)=\operatorname{Hom}(k\{\mathcal{G}\},R)$$ the morphism $k\{\mathcal{G}\}\twoheadrightarrow k\{G\}$ of $k$-$\sigma$-algebras induced by Lemma \ref{lemma: universal property of sA} is a morphism of $k$-$\sigma$-Hopf algebras. With respect to the corresponding $\sigma$-closed embedding $G\hookrightarrow\mathcal{G}$ we have $k\{G\}=k[G[0]]=k[G[1]]=\ldots$ and therefore the associated growth group is trivial. Thus ${\operatorname{ld}}(G)=1$.
\end{proof}
Let us now recall the definition of algebraic $\sigma$-groups from \cite{KowalskiPillay:OnAlgebraicSigmaGroups}. We start by introducing algebraic $\sigma$-varieties. Let $k$ be a $\sigma$-field and $\mathcal{X}$ a variety over $k$, where, for our purposes, a variety over $k$ is an affine scheme of finite type over $k$. As in Section \ref{subsec: Zariski closures} let $^\sigma\! \mathcal{X}$ denote the variety over $k$ obtained from $\mathcal{X}$ by base extension via $\sigma\colon k\to k$. Similarly, if $\phi\colon\mathcal{X}\to\mathcal{Y}$ is a morphism of varieties over $k$, then
$^\sigma\!\phi\colon{^\sigma\!\mathcal{X}}\to{^\sigma\!\mathcal{Y}}$ is the morphism obtained from $\phi$ by base extension via $\sigma\colon k\to k$.
An \emph{algebraic $\sigma$-variety} over $k$ is a variety $\mathcal{X}$ over $k$ together with a morphism $\widetilde{\sigma}\colon\mathcal{X}\to{^\sigma\!\mathcal{X}}$ of varieties over $k$. A morphism between algebraic $\sigma$-varieties is a morphism $\phi\colon\mathcal{X}\to\mathcal{Y}$ of varieties such that
$$
\xymatrix{
\mathcal{X} \ar^-{\widetilde{\sigma}}[r] \ar_\phi[d] & ^\sigma\! \mathcal{X} \ar^-{^\sigma\!\phi}[d]\\
\mathcal{Y} \ar^-{\widetilde{\sigma}}[r] & ^\sigma\! \mathcal{Y}
}
$$
commutes. An \emph{algebraic $\sigma$-group} is a group object in the category of algebraic $\sigma$-varieties. Equivalently, an algebraic $\sigma$-group is an algebraic group $\mathcal{G}$ over $k$ together with a morphism $\widetilde{\sigma}\colon\mathcal{G}\to{^\sigma\!\mathcal{G}}$ of algebraic groups.
\begin{theo}
The category of algebraic $\sigma$-groups is equivalent to the category of $\sigma$-algebraic groups of $\sigma$-dimension zero and limit degree one.
\end{theo}
\begin{proof}
Let $R$ be a $k$-algebra. To define a $k$-$\sigma$-algebra structure on $R$ is equivalent to defining a morphism of $k$-algebras $\overline{\sigma}\colon {^\sigma\! R}\to R$: Given $\sigma\colon R\to R$, we can define $\overline{\sigma}\colon {^\sigma\! R}\to R,\ r\otimes\lambda\mapsto \sigma(r)\lambda$. Conversely, given $\overline{\sigma}\colon {^\sigma\! R}\to R$, we can define $\sigma\colon R\xrightarrow{\operatorname{id}\otimes 1}R\otimes_k k={^\sigma\! R}\xrightarrow{\overline{\sigma}} R$.
Moreover, if $R$ and $S$ are $k$-$\sigma$-algebras and $\psi\colon R\to S$ is a morphism of $k$-algebras, then $\psi$ is a morphism of $k$-$\sigma$-algebras if and only if
$$
\xymatrix{
^\sigma\! R \ar^-{\overline{\sigma}}[r] \ar_-{^\sigma\!\psi}[d] & R \ar^\psi[d]\\
^\sigma\! S \ar^-{\overline{\sigma}}[r] & S
}
$$
commutes.
Dualizing the definition, an algebraic $\sigma$-group $\mathcal{G}$ corresponds to a finitely generated Hopf algebra $k[\mathcal{G}]$ together with a morphism of Hopf algebras $\overline{\sigma}=({\widetilde{\sigma}})^*\colon{^\sigma\! (k[\mathcal{G}])}\to k[\mathcal{G}]$. By the remark from the beginning of the proof, the statement that $\overline{\sigma}$ is a morphism of Hopf algebras corresponds to the statement that $k[\mathcal{G}]$ is a $k$-$\sigma$-Hopf algebra. Thus the category of algebraic $\sigma$-groups is anti--equivalent to the category of $k$-$\sigma$-Hopf algebras, which are finitely generated over $k$. Now the claim follows from Proposition \ref{prop: equivalence sgroups sHopfalgebras} and Proposition \ref{prop: ld equal 1}.
\end{proof}
\section{Quotients}
The first goal in this section is to establish the existence of the quotient $G/N$, where $N$ is a normal $\sigma$-closed subgroup of a $\sigma$-algebraic group $G$. We do not address the (highly non--trivial) question of the existence of the quotient $G/H$ where $H$ is an arbitrary $\sigma$-closed subgroup. As for (affine) algebraic groups, the quotient $G/H$ will in general not be affine (but rather quasi--projective). For algebraic groups, the analog of Theorem \ref{theo: Chevalley} is a crucial ingredient for constructing the quotient as a quasi--projective variety. Therefore Theorem \ref{theo: Chevalley} already gives some indication that ``$G/H$ is a quasi--projective $\sigma$-variety''. But to even make sense of this statement, one would need to introduce a substantially heavier foundation of difference algebraic geometry than we have done in Section \ref{sec: Difference algebraic geometry preliminaries}, where we only deal with the affine setting.
We also show how to compute $\sigma\text{-}\operatorname{dim}(G/N)$, $\operatorname{ord}(G/N)$ and ${\operatorname{ld}}(G/N)$ from the corresponding values for $G$ and $N$.
\medskip
Let $G$ be a $\sigma$-algebraic. A $\sigma$-closed subgroup $N\leq G$ is called \emph{normal} if $N(R)$ is a normal subgroup of $G(R)$ for any $k$-$\sigma$-algebra $R$. We write $N\unlhd G$ to express that $N$ is a normal $\sigma$-closed subgroup of $G$.
If $\phi\colon G\to H$ is a morphism of $\sigma$-algebraic groups, we define the kernel of $\phi$ $$\ker(\phi)$$
to be the subfunctor of $G$ given by $R\rightsquigarrow\ker(\phi_R))$. Then $\ker(\phi)$ is a normal $\sigma$-closed subgroup of $G$. Indeed $\ker(\phi)=\phi^{-1}(1)$, where $1\leq H$ is the trivial $\sigma$-closed subgroup of $H$ defined by the kernel $\mathfrak{m}_{k\{H\}}$ of the counit $k\{H\}\to k$.
\begin{defi}
Let $G$ be a $\sigma$-algebraic group and $N\unlhd G$ a normal $\sigma$-closed subgroup. A morphism of $\sigma$-algebraic groups $\pi\colon G\to G/N$ such that $N\subset\ker(\pi)$ is called a \emph{quotient of $G$ mod $N$} if it universal among such maps, i.e., for every morphism of $\sigma$-algebraic groups $\phi\colon G\to H$ with $N\subset\ker(\phi)$ there exists a unique morphism of $\sigma$-algebraic groups $\phi'\colon G/N\to H$ such that
\[\xymatrix{
G \ar^-{\pi}[rr] \ar_-{\phi}[rd] & & G/N \ar@{..>}^-{\phi'}[ld] \\
& H &
}
\]
commutes.
\end{defi}
Of course, if a quotient of $G$ mod $N$ exists, it is unique up to a unique isomorphism. We will therefore usually speak of \emph{the} quotient of $G$ mod $N$.
For affine group schemes over a field (not necessarily of finite type), the fundamental theorem on quotients can be formulated in a purely Hopf algebraic manner (\cite{Takeuchi:ACorrespondenceBetweenHopfidealsAndSubHopfalgebras}). Recall that a Hopf ideal $\mathfrak{a}$ in a Hopf algebra $A$ over $k$ is called \emph{normal} if, using Sweedler notation,
$$\sum f_{(1)}S(f_{(3)})\otimes f_{(2)}\in A\otimes_k\mathfrak{a}$$ for any $f\in\mathfrak{a}$, where $S$ is the antipode of $A$. Normal Hopf ideals in $A$ correspond to normal closed subgroup schemes (\cite[Lemma 5.1]{Takeuchi:ACorrespondenceBetweenHopfidealsAndSubHopfalgebras}). Similarly, if $G$ is a $\sigma$-algebraic group, then normal $\sigma$-Hopf ideals in $k\{G\}$ correspond to normal $\sigma$-closed subgroups of $G$ (Cf. Lemma \ref{lemma: sclosed sungroup corresponds to sHopfideal}.).
For a Hopf algebra $A$ over $k$ let us the denote the kernel of the counit $\varepsilon\colon A\to k$ by $\mathfrak{m}_A$.
\begin{theo}[M. Takeuchi] \label{theo: takeuchi}
Let $A$ be a Hopf algebra over $k$ and $\mathfrak{a}\subset A$ a Hopf ideal. Then $A(\mathfrak{a})=\{f\in A|\ \Delta(f)-f\otimes 1\in A\otimes_k\mathfrak{a}\}$ is a Hopf subalgebra of $A$ with $(\mathfrak{m}_{A(\mathfrak{a})})=\mathfrak{a}$. Indeed $A(\mathfrak{a})$ is the unique Hopf subalgebra with this property and the largest Hopf subalgebra with the property that $(\mathfrak{m}_{A(\mathfrak{a})})\subset\mathfrak{a}$.
\end{theo}
\begin{proof}
By \cite[Lemma 4.4]{Takeuchi:ACorrespondenceBetweenHopfidealsAndSubHopfalgebras} $A(\mathfrak{a})$ is a Hopf subalgebra. By \cite[Lemma 4.7]{Takeuchi:ACorrespondenceBetweenHopfidealsAndSubHopfalgebras} it is the largest Hopf subalgebra with $(\mathfrak{m}_{A(\mathfrak{a})})\subset\mathfrak{a}$. Finally, by \cite[Theorem 4.3]{Takeuchi:ACorrespondenceBetweenHopfidealsAndSubHopfalgebras} it is the unique Hopf subalgebra with $(\mathfrak{m}_{A(\mathfrak{a})})=\mathfrak{a}$.
\end{proof}
The existence of the quotient of $G$ mod $N$ can be reduced to Theorem \ref{theo: takeuchi}.
A similar approach was taken in \cite[Section A.9]{diVizioHardouinWibmer:DifferenceGaloisofDifferential}. While the result in \cite{diVizioHardouinWibmer:DifferenceGaloisofDifferential} is formulated in a slightly more general setup (there the $k$-$\sigma$-Hopf algebras need not be finitely $\sigma$-generated over $k$) the result we prove here is stronger. Indeed, with the aid of the second finiteness theorem (Theorem \ref{theo: finiteness2}) we show that $G/N$ is $\sigma$-algebraic, i.e., $k\{G/N\}$ is finitely $\sigma$-generated over $k$. This question remained open in \cite{diVizioHardouinWibmer:DifferenceGaloisofDifferential}.
\begin{theo} \label{theo: existence of quotient}
Let $G$ be a $\sigma$-algebraic group and $N\unlhd G$ a $\sigma$-closed subgroup. Then the quotient of $G$ mod $N$ exists. Moreover, a morphism of $\sigma$-algebraic groups $\pi\colon G\to G/N$ is the quotient of $G$ mod $N$ if and only if $\ker(\pi)=N$ and $\pi^*\colon k\{G/N\}\to k\{G\}$ is injective.
\end{theo}
\begin{proof}
By Theorem \ref{theo: takeuchi}
$$k\{G\}(\mathbb{I}(N))=\{f\in k\{G\}|\ \Delta(f)-f\otimes 1\in k\{G\}\otimes_k\mathbb{I}(N)\}$$ is a Hopf subalgebra of $k\{G\}$. Clearly it is also a $k$-$\sigma$-Hopf subalgebra. From Theorem \ref{theo: finiteness2} we know that $k\{G\}(\mathbb{I}(N))$ is finitely $\sigma$-generated over $k$. So we can define $G/N$ as the $\sigma$-algebraic group represented by $k\{G\}(\mathbb{I}(N))$, i.e., $k\{G/N\}=k\{G\}(\mathbb{I}(N))$. Let $\pi\colon G\to G/N$ be the morphism of $\sigma$-algebraic groups corresponding to the inclusion $k\{G/N\}\subset k\{G\}$ of $k$-$\sigma$-Hopf algebras.
Let $\phi\colon G\to H$ be a morphism of $\sigma$-algebraic groups such that $N\subset\ker(\phi)$. As $\ker(\phi)=\mathbb{V}(\phi^*(\mathfrak{m}_{k\{H\}}))$, the Hopf algebraic meaning of $N\subset\ker(\phi)$ is $\phi^*(\mathfrak{m}_{k\{H\}})\subset \mathbb{I}(N)$. To show that $\pi$ has the required universal property, it suffices to show that $\phi^*(k\{H\})\subset k\{G/N\}$. We know from Theorem \ref{theo: takeuchi} that $k\{G/N\}$ is the largest Hopf subalgebra of $k\{G\}$ such that $\mathfrak{m}_{k\{G/N\}}\subset\mathbb{I}(N)$. As
$\mathfrak{m}_{\phi^*(k\{H\})}=\phi^*(\mathfrak{m}_{k\{H\}})\subset\mathbb{I}(N)$ we find $\phi^*(k\{H\})\subset k\{G/N\}$.
Clearly $\pi^*$ is injective. Moreover, $\ker(\pi)=\mathbb{V}(\pi^*(\mathfrak{m}_{k\{G/N\}}))=\mathbb{V}(\mathbb{I}(N))=N$ by Theorem \ref{theo: takeuchi}.
If $\pi\colon G\to G/N$ is a morphism of $\sigma$-algebraic groups such that $N=\ker(\pi)$ and $\pi^*\colon k\{G/N\}\to k\{G\}$ is injective, then
$\pi^*(k\{G/N\})$ is a Hopf subalgebra of $k\{G\}$ such that $(\mathfrak{m}_{\pi^*(k\{G/N\})})=\mathbb{I}(N)$. Therefore $\pi^*(k\{G/N\})=k\{G\}(\mathbb{I}(N))$ by Theorem \ref{theo: takeuchi}.
\end{proof}
\begin{cor} \label{cor: quotient embedding}
Let $\phi\colon G\to H$ be a morphism of $\sigma$-algebraic groups. Then the induced morphism $G/\ker(\phi)\to H$ is a $\sigma$-closed embedding.
\end{cor}
\begin{proof}
The Hopf subalgebra $\phi^*(k\{H\})\subset k\{G\}$ satisfies $(\mathfrak{m}_{\phi^*(k\{H\})})=(\phi^*(\mathfrak{m}_{k\{H\}}))=\mathbb{I}(\ker(\phi))$.
Therefore $\phi^*(k\{H\})=k\{G\}(\mathbb{I}(\ker(f)))=k\{G/\ker(\phi)\}$ by Theorem \ref{theo: takeuchi}. So $k\{H\}\to k\{G/\ker(\phi)\}$ is surjective and
$G/\ker(\phi)\to H$ is a $\sigma$-closed embedding by Lemma \ref{lemma: sclosed embedding}.
\end{proof}
Let $\psi\colon R\to S$ be a morphism of $k$-$\sigma$-algebras. We say that $\psi$ is \emph{faithfully flat} if the underlying morphism
$\psi^\sharp\colon R^\sharp\to S^\sharp$ of $k$-algebras is faithfully flat. In this case, we also say that $S$ is a faithfully flat $R$-$\sigma$-algebra.
Let $R$ be a $k$-$\sigma$-algebra and $\pi\colon G\to G/N$ a quotient. As for algebraic groups, the map $\pi_R\colon G(R)\to (G/N)(R)$ need not be surjective in general. (See Example \ref{ex: surjective morphism}.) This, at least initially, makes it difficult to transfer constructions familiar from the theory of abstract groups which refer to group elements to $\sigma$-algebraic groups. The following lemma is a very useful substitute for the missing surjectivity of $\pi_R$. See Section \ref{sec: Group theory}, in particular Theorem \ref{theo: quotient sheaf} on how this lemma is used.
\begin{lemma} \label{lemma: sheaf surjectivity of quotient}
Let $G$ be a $\sigma$-algebraic group and $N\unlhd G$ a $\sigma$-closed subgroup. Let $R$ be a $k$-$\sigma$-algebra and $\overline{g}\in (G/N)(R)$. Then there exists a faithfully flat morphism $R\to S$ of $k$-$\sigma$-algebras and $g\in G(S)$ such that $G(S)\to (G/N)(S)$ maps $g$ to the image of $\overline{g}$ in $(G/N)(S)$.
\end{lemma}
\begin{proof}
We may use $\overline{g}\in (G/N)(R)=\operatorname{Hom}(k\{G/N\},R)$ to form $S=k\{G\}\otimes_{k\{G/N\}} R$. Since $k\{G\}$ is faithfully flat over $k\{G/N\}$ (See \cite[Chapter 14]{Waterhouse:IntroductiontoAffineGroupSchemes}.) it follows that $R\to S,\ r\mapsto 1\otimes r$ is faithfully flat (\cite[Section 13.3, p. 105]{Waterhouse:IntroductiontoAffineGroupSchemes}). Let $g\colon k\{G\}\to S,\ f\mapsto f\otimes 1$. Then
the maps $k\{G/N\}\xrightarrow{\overline{g}} R\to S$ and $k\{G/N\}\to k\{G\}\xrightarrow{g} S$ are equal. So $g\in G(S)$ has the required property.
\end{proof}
Now that we have established the existence of the quotient $G/N$ we can start to study its properties. To see how the numerical invariants $\sigma$-dimension, order and limit degree behave with respect to quotients, we first need to understand how quotients intertwine with Zariski closures.
\begin{lemma}
Let $\mathcal{G}$ be an algebraic group and $N\leq G\leq \mathcal{G}$ be $\sigma$-closed subgroups. For $i\geq 0$ let $G[i]$ and $N[i]$ denote the $i$-th order Zariski closure of $G$ and $N$ in $\mathcal{G}$ respectively. Then $N$ is normal in $G$ if and only if $N[i]$ is normal in $G[i]$ for every $i\geq 0$.
\end{lemma}
\begin{proof}
As $k\{G\}=\cup_{i\geq 0}k[G[i]]$ is the union of the Hopf subalgebras $k[G[i]]$, we see that $\mathbb{I}(N)$ is a normal Hopf ideal of $k\{G\}$ if and only if $\mathbb{I}(N)\cap k[G[i]]$ is a normal Hopf ideal of $k[G[i]]$ for every $i\geq 0$.
\end{proof}
\begin{prop} \label{prop: Zariskiclosures and quotients}
Let $\mathcal{G}$ be an algebraic group and $N\unlhd G\leq\mathcal{G}$. For $i\geq 0$ let $G[i]$ and $N[i]$ denote the $i$-th order Zariski closure of $G$ and $N$ in $\mathcal{G}$ respectively. Then there exists an integer $m\geq 0$ such that $G/N$ is a $\sigma$-closed subgroup of $G[m]/N[m]$ and for $i\geq 0$ the $i$-th order Zariski closure of $G/N$ in $G[m]/N[m]$ is the quotient of $G[i+m]$ mod $N[i+m]$, i.e.,
$$(G/N)[i]=G[m+i]/N[m+i].$$
\end{prop}
\begin{proof}
By Theorems \ref{theo: takeuchi} and \ref{theo: existence of quotient} we have
\begin{align*}
k\{G/N\} & =\left\{f\in k\{G\}|\ \Delta(f)-f\otimes 1\in k\{G\}\otimes_k\mathbb{I}(N)\right\}\\
& =\bigcup_{i\geq 0}\{f\in k[G[i]]|\ \Delta(f)-f\otimes 1\in k[G[i]]\otimes_k\mathbb{I}(N[i])\}\\
& =\bigcup_{i\geq 0} k[G[i]/N[i]].
\end{align*}
Moreover, $$k[G[i]/N[i]]\subset k[G[i+1]/N[i+1]] \text{ and } \sigma(k[G[i]/N[i]])\subset k[G[i+1]/N[i+1]].$$
By Corollary \ref{cor: finiteness1}, there exists an integer $m\geq 0$ such that $\mathbb{I}(N[j+1])=(\mathbb{I}(N[j]),\sigma(\mathbb{I}(N[j)))$, i.e.,
$N[j+1]=(N[j]\times{{}^{\sigma^j}\!\mathcal{G}})\cap(\mathcal{G}\times{^\sigma\!(N[j])})$ for $j\geq m$.
We claim that
\begin{equation}\label{eqn: good quotients}
k\left[k[G[m]/N[m]],\ldots,\sigma^i(k[G[m]/N[m]])\right]=k[G[m+i]/N[m+i]] \quad \text{ for } i\geq 0.
\end{equation}
The inclusion ``$\subset$'' is obvious. To prove the inclusion ``$\supset$'' it suffices to show that
$$\psi_j\colon k[G[j]/N[j]]\otimes_k{^\sigma\!(k[G[j]/N[j])}\longrightarrow k[G[j+1]/N[j+1]],\ f_1\otimes(\lambda\otimes f_2)\mapsto f_1\lambda\sigma(f_2)$$
is surjective for $j\geq m$. With $\pi_{j+1}$ and $\sigma_{j+1}$ as in (\ref{eqn: comm diag pi si}) the morphisms
$$G[j+1]\xrightarrow{\pi_{j+1}}G[j]\to G[j]/N[j] \text{ and } G[j+1]\xrightarrow{\sigma_{j+1}}{^\sigma\!(G[j])}\to {^\sigma\!(G[j]/N[j])}$$ combine to a morphism
$$G[j+1]\xrightarrow{}(G[j]/N[j])\times {^\sigma\!(G[j]/N[j])}$$
of algebraic groups with kernel $(N[j]\times{{}^{\sigma^{j}}\!\mathcal{G}})\cap(\mathcal{G}\times{^\sigma\!(N[j])})=N[j+1]$.
Therefore $$G[j+1]/N[j+1]\longrightarrow(G[j]/N[j])\times {^\sigma\!(G[j]/N[j])}$$ is a closed embedding and so the dual map is surjective, but the dual map is precisely $\psi_j$. We have thus proved (\ref{eqn: good quotients}). It follows from (\ref{eqn: good quotients}) that $k\{G[m]/N[m]\}\to k\{G/N\}$ is surjective, i.e., $G/N$ is a $\sigma$-closed subgroup of $G[m]/N[m]$. As the ring to the left hand side of (\ref{eqn: good quotients}) is the coordinate ring of the $i$-th order Zariski closure of $G/N$ in $G[m]/N[m]$, we obtain the required equality of the Zariski closures.
\end{proof}
The following example shows that in general one can not take $m=0$ in Proposition \ref{prop: Zariskiclosures and quotients}.
\begin{ex}
Let $G=\mathcal{G}=\mathbb{G}_a$ and $N\unlhd G$ the $\sigma$-closed subgroup given by
$N(R)=\{g\in R|\ \sigma(g)=0\}$ for any $k$-$\sigma$-algebra $R$. Then $N[0]=G[0]=\mathbb{G}_a$ and $G[0]/N[0]$ is the trivial group. Therefore $G/N$ can not be a $\sigma$-closed subgroup of $G[0]/N[0]$.
\end{ex}
\begin{cor} \label{cor: sdim and ord for quotients}
Let $G$ be a $\sigma$-algebraic group and $N\unlhd G$ a normal $\sigma$-closed subgroup. Then
\begin{equation} \label{eqn: sdim} \sigma\text{-}\operatorname{dim}(G)=\sigma\text{-}\operatorname{dim}(N)+\sigma\text{-}\operatorname{dim}(G/N)\end{equation} and \begin{equation} \label{eqn: ord} \operatorname{ord}(G)=\operatorname{ord}(N)+\operatorname{ord}(G/N).\end{equation}
\end{cor}
\begin{proof}
We may assume that $G$ is a $\sigma$-closed subgroup of some algebraic group $\mathcal{G}$ (Theorem \ref{theo: linearization}).
For $i\geq 0$ let $G[i]$ and $N[i]$ denote the $i$-th order Zariski closure of $G$ and $N$ in $\mathcal{G}$ respectively. By Theorem \ref{theo: existence sdim} there exist $e_G,e_N\geq 0$ such that $\dim(G[i])=\sigma\text{-}\operatorname{dim}(G)(i+1)+e_G$ and $\dim(N[i])=\sigma\text{-}\operatorname{dim}(N)(i+1)+e_N$ for $i\gg 0$.
Let $m\geq 0$ be as in Proposition \ref{prop: Zariskiclosures and quotients} and for $i\geq 0$ let $(G/N)[i]$ denote the $i$-th order Zariski closure of $G/N$ in $G[m]/N[m]$. By Theorem \ref{theo: existence sdim} there exist $e_{G/N}\geq 0$ such that $\dim((G/N)[i])=\sigma\text{-}\operatorname{dim}(G/N)(i+1)+e_{G/N}$. For $i\gg 0$
\begin{align*}
\sigma\text{-}\operatorname{dim} & (G/N)(i+1)+e_{G/N} =\dim((G/N)[i])=\dim(G[m+i]/N[m+i])= \\
&=\dim(G[m+i])-\dim(N[m+i])=\\
&=\sigma\text{-}\operatorname{dim}(G)(m+i+1)+e_G-\sigma\text{-}\operatorname{dim}(N)(m+i+1)-e_N=\\
&=(\sigma\text{-}\operatorname{dim}(G)-\sigma\text{-}\operatorname{dim}(N))(i+1)+ (\sigma\text{-}\operatorname{dim}(G)-\sigma\text{-}\operatorname{dim}(N))m +e_G-e_N.
\end{align*}
This proves (\ref{eqn: sdim}). As $\operatorname{ord}(G)<\infty$ if and only if $\sigma\text{-}\operatorname{dim}(G)=0$, it follows from (\ref{eqn: sdim}) that (\ref{eqn: ord}) is
valid if $\sigma\text{-}\operatorname{dim}(G)>0$. We can therefore assume that $\sigma\text{-}\operatorname{dim}(G)=0$, and consequently $\sigma\text{-}\operatorname{dim}(N)=\sigma\text{-}\operatorname{dim}(G/N)=0$ as well. But then
$\operatorname{ord}(G/N)=e_{G/N}=e_G-e_N=\operatorname{ord}(G)-\operatorname{ord}(N)$.
\end{proof}
Next we will show how to compute ${\operatorname{ld}}(G/N)$ from ${\operatorname{ld}}(N)$ and ${\operatorname{ld}}(G)$. For clarity of the exposition, we single out a simple lemma on algebraic groups.
\begin{lemma} \label{lemma: kernel for algebraic groups}
Let $N_1\unlhd G_1$ and $N_2\unlhd G_2$ be algebraic groups and $\phi\colon G_2\twoheadrightarrow G_1$ a surjective morphism of algebraic groups with kernel $\mathcal{G}$. Assume that the restriction of $\phi$ to $N_2$ has kernel $\mathcal{N}$ and maps $N_2$ surjectively onto $N_1$. Then the kernel of the induced map $G_2/N_2\to G_1/N_1$ is isomorphic to $\mathcal{G}/\mathcal{N}$.
\end{lemma}
\begin{proof}
Since $\phi$ is surjective we may identify $G_1$ with $G_2/\mathcal{G}$. Note that the (Noether) isomorphism theorems also hold for algebraic groups. (See e.g., \cite[Chapter IX]{Milne:BasicTheoryOfAffineGroupSchemes}). We have $N_1=N_2/\mathcal{N}=N_2/\mathcal{G}\cap\mathcal{N}_2=N_2\mathcal{G}/\mathcal{G}$ and so $G_1/N_1=(G_2/\mathcal{G})/(N_2\mathcal{G}/\mathcal{G})=G_2/N_2\mathcal{G}$. This shows that the kernel of
$G_2/N_2\to G_1/N_1=G_2/N_2\mathcal{G}$ equals $N_2\mathcal{G}/N_2=\mathcal{G}/N_2\cap\mathcal{G}=\mathcal{G}/\mathcal{N}$.
\end{proof}
\begin{cor} \label{cor: multiplicativity of limit degree}
Let $G$ be a $\sigma$-algebraic group and $N\unlhd G$ a normal $\sigma$-closed subgroup.
Then
$${\operatorname{ld}}(G)={\operatorname{ld}}(G/N)\cdot{\operatorname{ld}}(N).$$
\end{cor}
\begin{proof}
By Remark \ref{rem: ld finite} and Corollary \ref{cor: sdim and ord for quotients} the claim is valid if $\sigma\text{-}\operatorname{dim}(G)>0$. So we may assume that $\sigma\text{-}\operatorname{dim}(G)=0$ and therefore ${\operatorname{ld}}(G),\ {\operatorname{ld}}(G/N)$ and ${\operatorname{ld}}(N)$ are all finite. Let $m\geq 0$ be as in Proposition \ref{prop: Zariskiclosures and quotients}. For $i\geq1$ we have commutative diagrams
$$
\xymatrix{
(G/N)[i] \ar@{->>}^{\pi_i}[r] \ar^-\simeq[d]& (G/N)[i-1] \ar^-\simeq[d] \\
G[m+i]/N[m+i] \ar@{->>}^-{\phi_i}[r] & G[m+i-1]/N[m+i-1]
}
$$
where $\phi_i$ is induced from the projection $G[m+i]\twoheadrightarrow G[m+i-1]$. For $i\gg 0$ we have ${\operatorname{ld}}(G/N)=|\ker(\pi_i)|=|\ker(\phi_i)|$.
Let $\mathcal{G}_{m+i}$ and $\mathcal{N}_{m+i}$ be the kernel of $G[m+i]\twoheadrightarrow G[m+i-1]$ and $N[m+i]\twoheadrightarrow N[m+i-1]$ respectively.
It follows from Lemma \ref{lemma: kernel for algebraic groups} that $\ker(\phi_i)=\mathcal{G}_{m+i}/\mathcal{N}_{m+i}$. Therefore
$${\operatorname{ld}}(G/N)=|\mathcal{G}_{m+i}/\mathcal{N}_{m+i}|=|\mathcal{G}_{m+i}|/|\mathcal{N}_{m+i}|={\operatorname{ld}}(G)/{\operatorname{ld}}(N).$$
\end{proof}
\section{Morphisms}
In this section we characterize the analogs of injective and surjective morphisms in the category of groups. Analogous results for algebraic groups are in \cite[Chapter VII]{Milne:BasicTheoryOfAffineGroupSchemes}.
\begin{theo} \label{theo: injective morphism}
Let $\phi\colon G\to H$ be a morphism of $\sigma$-algebraic groups. Then the following statements are equivalent:
\begin{enumerate}
\item The kernel of $\phi$ is trivial.
\item The map $\phi_R\colon G(R)\to H(R)$ is injective for every $k$-$\sigma$-algebra $R$.
\item The morphism $\phi\colon G\to H$ is a $\sigma$-closed embedding.
\item The dual map $\phi^*\colon k\{H\}\to k\{G\}$ is surjective.
\item The morphism $\phi\colon G\to H$ is a monomorphism in the category of $\sigma$-algebraic groups, i.e., for every pair $\phi_1,\phi_2\colon H'\to G$ of morphisms of $\sigma$-algebraic groups with $\phi\f_1=\phi\f_2$ we have $\phi_1=\phi_2$.
\end{enumerate}
\end{theo}
\begin{proof}
Clearly (i)$\Leftrightarrow$(ii), (iii)$\Rightarrow$(ii) and (ii)$\Rightarrow$(v). Moreover, (iii) and (iv) are equivalent by Lemma \ref{lemma: sclosed embedding}. So it suffices to show that (v) implies (iv). Define $H'=G\times_H G$ by
$$(G\times_H G)(R)=\{(g_1,g_2)\in G(R)\times G(R)|\ \phi(g_1)=\phi(g_2)\}$$
for any $k$-$\sigma$-algebra $R$. This is a $\sigma$-closed subgroup of $G\times G$. Indeed, $G\times_H G$ is represented by $k\{G\}\otimes_{k\{H\}}k\{G\}$.
Let $\phi_1$ and $\phi_2$ denote the projections onto the first and second coordinate respectively. We have $\phi\f_1=\phi\f_2$ and so by (iv) we must have $\phi_1=\phi_2$. This implies that the maps $f\mapsto f\otimes 1$ and $f\mapsto 1\otimes f$ from $k\{G\}\to k\{G\}\otimes_{k\{H\}} k\{G\}$ are equal. As $\phi^*(k\{H\})$ is a Hopf subalgebra of $k\{G\}$ we know that $k\{G\}$ is faithfully flat over $\phi^*(k\{H\})$ (\cite[Chapter 14]{Waterhouse:IntroductiontoAffineGroupSchemes}). Therefore $f\otimes 1=1\otimes f$ in $k\{G\}\otimes_{k\{H\}} k\{G\}= k\{G\}\otimes_{\phi^*(k\{H\})} k\{G\}$ if and only if $f\in \phi^*(k\{H\})$ by \cite[Section 13.1, p. 104]{Waterhouse:IntroductiontoAffineGroupSchemes}. Summarily, we find that $\phi^*\colon k\{H\}\to k\{G\}$ is surjective.
\end{proof}
\begin{defi}
A morphism of $\sigma$-algebraic groups satisfying the equivalent properties of Theorem \ref{theo: injective morphism} is called \emph{injective}.
\end{defi}
\begin{ex}
The morphism $\phi\colon\mathbb{G}_m\to \mathbb{G}_m$ given by $\phi_R(g)=\sigma(g)$ for any $k$-$\sigma$-algebra $R$ and $g\in R^\times$ is not injective. Even though $\phi_R$ is injective for every $\sigma$-field extension $R$ of $k$.
\end{ex}
Recall that in Lemma \ref{lemma: f(X)} we defined $\phi(X)$ for a morphism $\phi\colon X\to Y$ of $\sigma$-varieties.
\begin{lemma} \label{lemma: f(G) subgroup}
Let $\phi\colon G\to H$ be a morphism of $\sigma$-algebraic groups and $G_1$ a $\sigma$-closed subgroup of $G$. Then $\phi(G_1)$ is a $\sigma$-closed subgroup of $H$.
\end{lemma}
\begin{proof}
We may assume that $G_1=G$. It follows from the proof of Lemma \ref{lemma: f(X)} that $\phi(G)=\mathbb{V}(\mathfrak{a})$, where $\mathfrak{a}$ is the kernel of $\phi^*\colon k\{H\}\to k\{G\}$. Since $\phi^*$ is a morphism of $k$-$\sigma$-Hopf algebras, $\mathfrak{a}$ is a $\sigma$-Hopf ideal. So $\phi(G)$ is a $\sigma$-closed subgroup of $H$.
\end{proof}
\begin{theo} \label{theo: surjective morphism}
Let $\phi\colon G\to H$ be a morphism of $\sigma$-algebraic groups. The following statements are equivalent:
\begin{enumerate}
\item $\phi(G)=H$.
\item The morphism $\phi$ is a quotient, i.e., there exists a normal $\sigma$-closed subgroup $N$ of $G$ such $\phi$ is the quotient of $G$ mod $N$.
\item The dual map $\phi^*\colon k\{H\}\to k\{G\}$ is injective.
\item For every $k$-$\sigma$-algebra $R$ and every $h\in H(R)$, there exists a faithfully flat $R$-$\sigma$-algebra $S$ and $g\in G(S)$ such that the image of $h$ in $H(S)$ equals $\phi(g)$.
\end{enumerate}
\begin{proof}
We know from Lemma \ref{lemma: f(X)} that $\phi(G)$ is the $\sigma$-closed $\sigma$-subvariety of $H$ defined by $\ker(\phi^*)$. Therefore (i) and (iii) are equivalent. It is clear from Theorem \ref{theo: existence of quotient} that (iii) and (ii) are equivalent. Moreover (ii) implies (iv) by Lemma \ref{lemma: sheaf surjectivity of quotient}. It thus suffices to show that (iv) implies (iii). Take $R=k\{H\}$ and
$h=\operatorname{id}_{k\{H\}}\in H(R)=\operatorname{Hom}(k\{H\},k\{H\})$. By (iv) there exists a faithfully flat morphism $\psi\colon k\{H\}\to S$ of $k$-$\sigma$-algebras and an element $g\in G(S)=\operatorname{Hom}(k\{G\},S)$ such that the image of $h$ in $H(S)=\operatorname{Hom}(k\{H\},S)$ equals $\phi(g)=g\phi^*$. This means that $\psi=g\phi^*$. As any faithfully flat morphism of rings is injective, $\psi$ is injective. Therefore $\phi^*$ is injective as well.
\end{proof}
\end{theo}
\begin{defi}
A morphism of $\sigma$-algebraic groups satisfying the equivalent properties of Theorem \ref{theo: surjective morphism} is called \emph{surjective}.
\end{defi}
We write $\phi\colon G\twoheadrightarrow H$ to express that $\phi$ is surjective.
\begin{ex} \label{ex: surjective morphism}
The morphism $\phi\colon\mathbb{G}_m\to \mathbb{G}_m$ given by $\phi_R(g)=\sigma(g)$ for any $k$-$\sigma$-algebra $R$ and $g\in R^\times$ is surjective since the dual map $\phi^*\colon k\{y,y^{-1}\}\to k\{y,y^{-1}\},\ y\mapsto \sigma(y)$ is injective. Note that $\phi_R$ need not be surjective.
\end{ex}
\begin{cor}
A morphism of $\sigma$-algebraic groups which is injective and surjective is an isomorphism.
\end{cor}
\begin{proof}
By Theorems \ref{theo: injective morphism} and \ref{theo: surjective morphism}, an injective and surjective morphism of $\sigma$-algebraic groups induces a surjective and injective morphism on the $\sigma$-coordinate rings.
\end{proof}
\begin{cor}
Every morphism of $\sigma$-algebraic groups factors uniquely as a surjective morphism followed by an injective morphism.
\end{cor}
\begin{proof}
Let $\phi\colon G\to H$ be a morphism of $\sigma$-algebraic groups. The uniqueness means that if $G\twoheadrightarrow H_1\hookrightarrow H$ and $G\twoheadrightarrow H_2\hookrightarrow H$ are two factorizations of $\phi$, then there exists an isomorphism $H_1\to H_2$ of $\sigma$-algebraic groups making
$$
\xymatrix{
G \ar@{=}[d] \ar@{->>}[r] & H_1 \ar^{\simeq}[d] \ar@{^(->}[r] & H \ar@{=}[d] \\
G \ar@{->>}[r] & H_2 \ar@{^(->}[r] & H
}
$$
commutative. By Theorem \ref{theo: finiteness2} the $k$-$\sigma$-Hopf subalgebra $\phi^*(k\{H\})$ of $k\{G\}$ is finitely $\sigma$-generated over $k$. So we can define $H_1$ as the $\sigma$-algebraic group represented by $\phi^*(k\{H\})$. The claim of the corollary follows immediately by dualizing.
\end{proof}
Note that $H_1$ has two interpretations, either as $\phi(G)$ or as $G/\ker(\phi)$. See Theorem \ref{theo: isom1}.
As we will see in the sequel, point (iv) of Theorem \ref{theo: surjective morphism} is very helpful to reduce certain questions to computations with group elements. Let $X$ be a $\sigma$-variety. If $R\to S$ is an injective morphism of $k$-$\sigma$-algebras (e.g., $S$ is a faithfully flat $R$-$\sigma$-algebra), then
$$X(R)=\operatorname{Hom}(k\{X\},R)\to\operatorname{Hom}(k\{X\},S)=X(S)$$ is injective. To simplify the notation, we will, in the sequel, often identify $X(R)$ with its image in $X(S)$.
\begin{lemma} \label{lemma: faithfullyflat intersection}
Let $Y$ be a $\sigma$-closed $\sigma$-subvariety of a $\sigma$-variety $X$ and $R\to S$ an injective morphism of $k$-$\sigma$-algebras (e.g., $R\to S$ is faithfully flat). Then
$$
Y(R)=X(R)\cap Y(S),
$$
where, using the above described identification, the intersection is understood to take place in $X(S)$.
\end{lemma}
\begin{proof}
The inclusion ``$\subset$'' is obvious. To prove ``$\supset$'' it suffices to note that for
a morphism $k\{X\}\to S$ with factorizations $k\{X\}\twoheadrightarrow k\{Y\}\to S$ and $k\{X\}\to R\hookrightarrow S$, one has an arrow $k\{Y\}\to R$ such that
$$\xymatrix{ & k\{Y\} \ar@{..>}[dd] \ar[rd] & \\
k\{X\} \ar@{->>}[ur] \ar[rd] & & S \\
& R \ar@{^(->}[ur] &
}
$$
commutes.
\end{proof}
\begin{lemma} \label{lemma: f(G) faithfully flat}
Let $\phi\colon G\to H$ be a morphism of $\sigma$-algebraic groups and $G_1\leq G$ a $\sigma$-closed subgroup. Let $R$ be a $k$-$\sigma$-algebra. Then
$\phi(G_1)(R)$ equals the set of all $h\in H(R)$ such that there exists a faithfully flat $R$-$\sigma$-algebra $S$ and $g_1\in G_1(S)$ with $\phi(g_1)=h$.
\end{lemma}
\begin{proof}
The induced morphism $G_1\to\phi(G_1)$ is surjective. So it follows from Theorem \ref{theo: surjective morphism} (iv) that for $h\in\phi(G_1)(R)$ there exists a faithfully flat $R$-$\sigma$-algebra $S$ and $g_1\in G_1(S)$ with $\phi(g_1)=h$.
Conversely, if $h=\phi(g_1)$, then $h\in\phi(G_1(S))\subset\phi(G_1)(S)$ and it follows from
$\phi(G_1)(S)\cap H(R)=\phi(G_1)(R)$ (Lemma \ref{lemma: faithfullyflat intersection}) that $h\in\phi(G_1)(R)$.
\end{proof}
\begin{lemma} \label{lemma: normal stays normal under surjective map}
Let $\phi\colon G\to H$ be a surjective morphism of $\sigma$-algebraic groups. If $N$ is a normal $\sigma$-closed subgroup of $G$, then $\phi(N)$ is a normal $\sigma$-closed subgroup of $H$.
\end{lemma}
\begin{proof}
Let $R$ be a $k$-$\sigma$-algebra, $h\in\phi(N)(R)$ and $h_1\in H(R)$. We have to show that $h_1hh_1^{-1}\in\phi(N)(R)$. By Theorem \ref{theo: surjective morphism}, there exists a faithfully flat $R$-$\sigma$-algebra $S$ and $g\in N(S)$ with $\phi(g)=h$. Similarly, there exists a faithfully flat $R$-$\sigma$-algebra $S_1$ and $g_1\in G(S_1)$ with $\phi(g_1)=h_1$. Then $S'=S\otimes_R S_1$ is a faithfully flat $R$-$\sigma$-algebra (\cite[Section 13.3, p. 106]{Waterhouse:IntroductiontoAffineGroupSchemes}) and we can consider $G(S)$ and $G(S_1)$ as subgroups of $G(S')$. Since $N(S')\unlhd G(S')$ we see that $g_1gg_1^{-1}\in N(S')$. Therefore
$\phi(g_1gg_1^{-1})=h_1hh_1^{-1}\in\phi(N(S'))\subset\phi(N)(S')$. As $\phi(N)(S')\cap H(R)=\phi(N)(R)$ by Lemma \ref{lemma: faithfullyflat intersection}, this shows that $h_1hh_1^{-1}\in\phi(N)(R)$.
\end{proof}
\section{Components} \label{sec: components}
In \cite[Section 4.6]{Hrushovski:elementarytheoryoffrobenius} E. Hrushovski raised the question whether or not it is possible to strengthen the classical Ritt--Raudenbusch basis theorem (\cite[Theorem 2.5.11]{Levin:difference}). For clarity, let us state the question as a conjecture:
\begin{conj} \label{conj: Hrushovski1}
Let $k$ be a $\sigma$-field and $R$ a finitely $\sigma$-generated $k$-$\sigma$-algebra. Then every ascending chain of radical, mixed $\sigma$-ideals in $R$ is finite.
\end{conj}
Here a $\sigma$-ideal $\mathfrak{a}$ is called mixed, if $ab\in\mathfrak{a}$ implies $a\sigma(b)\in\mathfrak{a}$. E. Hrushovski proved Conjecture \ref{conj: Hrushovski1} under certain additional assumptions on $R$. (See \cite[Lemma 4.35]{Hrushovski:elementarytheoryoffrobenius}.) In \cite{Levin:OnTheAscendingChainCondition} A. Levin showed that the conjecture fails if the assumption that the $\sigma$-ideals are radical is dropped.
A prime $\sigma$-ideal $\mathfrak{p}$ of a $\sigma$-ring $R$ is called minimal if for every prime $\sigma$-ideal $\mathfrak{q}$ of $R$ with $\mathfrak{q}\subset \mathfrak{p}$ we have $\mathfrak{q}=\mathfrak{p}$. Using \cite[Lemma 4.34]{Hrushovski:elementarytheoryoffrobenius} one can show that Conjecture \ref{conj: Hrushovski1} is equivalent to the following:
\begin{conj} \label{conj: Hrushovski2}
Let $k$ be a $\sigma$-field and $R$ a finitely $\sigma$-generated $k$-$\sigma$-algebra. Then the set of minimal prime $\sigma$-ideals of $R$ is finite.
\end{conj}
One of the main results of this section is a special case of the above conjecture:
\begin{theo} \label{theo: Hrushovskiconj}
Conjecture \ref{conj: Hrushovski2} holds under the additional assumption that $R$ can be equipped with the structure of a $k$-$\sigma$-Hopf algebra.
\end{theo}
The proof of Theorem \ref{theo: Hrushovskiconj} is given at the end of this section.
More generally, in this section we study the components of a $\sigma$-algebraic group. The matter is complicated by the fact that, contrary to the case of algebraic groups or differential algebraic groups, a $\sigma$-algebraic group may have infinitely many components. However, as indicated in Theorem \ref{theo: Hrushovskiconj}, a $\sigma$-algebraic group has only finitely many ``$\sigma$-components''.
To be able to speak meaningfully of topological notions, such as connected components, we first need to clarify what is the topological space associated to a $\sigma$-algebraic group. By the underlying topological space of a $\sigma$-algebraic group $G$ we mean the underlying topological space of the scheme $G^\sharp$. In other words, the underlying topological space of $G$ is $\operatorname{Spec}(k\{G\})$. The reader might wonder if not some difference analog of $\operatorname{Spec}(-)$ would provide a more adequate notion. For example (cf. \cite{Hrushovski:elementarytheoryoffrobenius}), one could consider the space of all $\sigma$-prime $\sigma$-ideals of $k\{G\}$ instead of the space of all prime ideals of $k\{G\}$. (Recall that a prime $\sigma$-ideal $\mathfrak{p}$ is called $\sigma$-prime if $\sigma^{-1}(\mathfrak{p})=\mathfrak{p}$.) However, there often are not ``enough'' $\sigma$-prime $\sigma$-ideals. In general, a $\sigma$-ring need not contain a $\sigma$-prime $\sigma$-ideal. This pathology does not occur for $\sigma$-algebraic groups, since the kernel $\mathfrak{m}$ of the counit $k\{G\}\to k$ is a $\sigma$-prime $\sigma$-ideal. But it may happen that $\mathfrak{m}$ is the only $\sigma$-prime $\sigma$-ideal of $k\{G\}$, even if $G$ is quite far from being the trivial group:
\begin{ex} \label{ex: no field points} For all $\sigma$-algebraic groups in the following list, the $\sigma$-ideal $\mathfrak{m}$ is the only $\sigma$-prime $\sigma$-ideal of $k\{G\}$. (Equivalently, $G(K)=1$ for every $\sigma$-field extension $K$ of $k$. Cf. Remark \ref{rem: no field points}.)
\begin{itemize}
\item $G(R)=\{g\in\operatorname{GL}_n(R)|\ \sigma(g_{ij})=\delta_{ij}\}\leq \operatorname{GL}_n(R)$
\item $G(R)=\{g\in R|\ \sigma^n(g)=0\}\leq\mathbb{G}_a(R)$ for some $n\geq 1$.
\item $G(R)=\{g\in R|\ g^n=1,\ \sigma(g)=1\}\leq\mathbb{G}_m(R)$ for some $n\geq 1$.
\item $G(R)=\{g\in R|\ g^3=1,\ \sigma(g)=g\}\leq \mathbb{G}_m(R)$, where $k$ contains two non--trivial third roots of unity which are permuted by $\sigma\colon k\to k$.
\item $G(R)=\{g\in R^\times|\ g^p=1\}\leq\mathbb{G}_m(R)$ where $k$ has characteristic $p>0$.
\item $G(R)=\{g\in R^\times|\ g^p=1,\ \sigma^2(g)=g^3\}\leq\mathbb{G}_m(R)$ where $k$ has characteristic $p>0$.
\item Let $\mathsf{G}$ be a finite group and $\Sigma\colon \mathsf{G}\to \mathsf{G}$ a group endomorphism with only one fixed point (for example, $\Sigma(g)=1$ for $g\in \mathsf{G}$). Let $G$ be the corresponding $\sigma$-algebraic group constructed in Example \ref{ex: split finite}.
\end{itemize}
\end{ex}
As usual, by an \emph{irreducible component} of a topological space we mean a maximal irreducible subset. By a \emph{connected component} of a topological space we mean a maximal connected subset. An irreducible (or connected) component is automatically closed. Every topological space is the disjoint union of its connected components.
By a connected (or irreducible) component of a $\sigma$-algebraic group $G$ we mean a connected (or irreducible) component of the underlying topological space of $G$. If $R$ is a ring and $\mathfrak{a}\subset R$, let us denote by $\mathcal{V}(\mathfrak{a})$ the closed subset of $\operatorname{Spec}(R)$ defined by $\mathfrak{a}$.
\begin{lemma} \label{lemma:connectedcomponentsareirreducible}
Let $G$ be a $\sigma$-algebraic group. The connected components and the irreducible components of $G$ coincide. Moreover, if $\mathfrak{p}$ is a prime ideal of $k\{G\}$, then the connected component of $G$ containing $\mathfrak{p}$ equals $\mathcal{V}(\mathfrak{a})$, where $\mathfrak{a}$ is the ideal of $k\{G\}$ generated by all idempotent elements of $k\{G\}$ contained in $\mathfrak{p}$.
\end{lemma}
\begin{proof}
Let us fix a $\sigma$-closed embedding $G\hookrightarrow\mathcal{G}$ of $G$ into some algebraic group $\mathcal{G}$ and for $i\geq 0$ let $G[i]$ denote the $i$-th order Zariski--closure of $G$ in $\mathcal{G}$. Let $C\subset\operatorname{Spec}(k\{G\})$ denote a connected component of $G$. Then $C=\mathcal{V}(\mathfrak{a})$ for a unique ideal $\mathfrak{a}$ of $k\{G\}$ generated by idempotent elements. (See \cite[Tag 00EB]{stacks-project}.) For every $i\geq 0$, the closure of the image of $C$ under the projection $G^\sharp\to G[i]$ is connected and equal to $\mathcal{V}(\mathfrak{a}\cap k[G[i]])\subset G[i]$. So $\mathcal{V}(\mathfrak{a}\cap k[G[i]])\subset G[i]$ is contained in a connected component of $G[i]$. Assume that $G[i]$ has $n_i$ connected components. Then $$k[G[i]]=e_{i,1}k[G[i]]\oplus\cdots\oplus e_{i,n_i}k[G[i]]$$ for some primitive idempotent elements $e_{i,1},\ldots,e_{i,n_i}\in k[G[i]]$ and $\mathcal{V}(\mathfrak{a}\cap k[G[i]])\subset \mathcal{V}(\mathfrak{b}_i)$ where $\mathfrak{b}_i=(e_{i,1},\ldots e_{i,j_i-1}, e_{i,j_i+1},\ldots,e_{i,n_i})\subset k[G[i]]$ for a unique $j_i\in \{1,\ldots,n_i\}$.
We have $\mathfrak{b}_{i+1}\cap k[G[i]]=\mathfrak{b}_i$ for $i\geq 0$ and $\mathfrak{b}:=\cup_{i\geq 0} \mathfrak{b}_i$ is an ideal of $k\{G\}$.
From $\mathcal{V}(\mathfrak{a}\cap k[G[i]])\subset \mathcal{V}(\mathfrak{b}_i)$ it follows that $\mathfrak{b}_i$ is contained in the radical of $\mathfrak{a}\cap k[G[i]]$. Since the $e_{i,j}$'s are idempotent this shows that $\mathfrak{b}_i\subset\mathfrak{a}$. So $\mathfrak{b}\subset\mathfrak{a}$.
Since $k\{G\}/\mathfrak{b}$ may be interpreted as the directed union of the algebras $k[G[i]]/\mathfrak{b}_i\simeq e_{i,j_i}k[G[i]]$ which have a prime nilradical, it is clear that $\mathcal{V}(\mathfrak{b})$ is irreducible (and a fortiori connected). As $C=\mathcal{V}(\mathfrak{a})\subset \mathcal{V}(\mathfrak{b})$ it follows from the maximality of $C$ that $\mathcal{V}(\mathfrak{a})=\mathcal{V}(\mathfrak{b})$. By the uniqueness of $\mathfrak{a}$ we have $\mathfrak{a}=\mathfrak{b}$.
We have thus shown that every connected component of $G$ is irreducible. So the connected and the irreducible components of $G$ coincide.
The claimed form of the connected component of a prime ideal of $k\{G\}$ follows from the above arguments.
\end{proof}
Since the connected components and the irreducible components of a $\sigma$-algebraic group coincide, we will speak simply of the \emph{components of a $\sigma$-algebraic group} in the sequel. The components of a $\sigma$-algebraic group $G$ are in bijection with the minimal prime ideals of $k\{G\}$. The following simple example shows that a $\sigma$-algebraic group can have infinitely many components and that the components need not be open.
\begin{ex} \label{ex: infinitely many components}
Let $G\leq\mathbb{G}_m$ be the $\sigma$-algebraic group given by
$$G(R)=\{g\in R^\times|\ g^2=1\}\leq\mathbb{G}_m(R)$$
for any $k$-$\sigma$-algebra $R$. We have $$k\{G\}=k[y]/[y^2-1]=k[y,\sigma(y),\ldots]/(y^2-1,\sigma(y)^2-1,\ldots).$$ Let us assume that the characteristic of $k$ is not equal to $2$. Then the prime ideals of $k\{G\}$ are in bijection with the set of all sequences $(a_i)_{i\in\mathbb{N}}$ such that $a_i\in\{1,-1\}$.
Every prime ideal of $k\{G\}$ is its own component. In particular, $G$ has infinitely many components. The open subsets of $\operatorname{Spec}(k\{G\})$ are all infinite, thus the components are not open.
\end{ex}
Allowing ourselves a little abuse of notation we denote the canonical map $$\operatorname{Spec}(k\{G\})\to\operatorname{Spec}(k\{G\}),\ \mathfrak{p}\mapsto\sigma^{-1}(\mathfrak{p})$$ also by $\sigma$.
\begin{defi}
A component $C$ of a $\sigma$-algebraic group is called a \emph{$\sigma$-component} if $\sigma(C)\subset C$.
\end{defi}
\begin{ex}
The $\sigma$-algebraic group from Example \ref{ex: infinitely many components} has two $\sigma$-components, namely the prime ideals corresponding to the sequences $(1,1,\ldots)$ and $(-1,-1,\ldots)$.
\end{ex}
\begin{lemma} \label{lemma: scomponent}
Let $G$ be a $\sigma$-algebraic group and $C\subset\operatorname{Spec}(k\{G\})$ a component of $G$. Let $\mathfrak{p}\subset k\{G\}$ be the prime ideal and $\mathfrak{a}\subset k\{G\}$ the ideal generated by idempotent elements such that $C=\mathcal{V}(\mathfrak{p})=\mathcal{V}(\mathfrak{a})$. Then the following conditions are equivalent:
\begin{enumerate}
\item The component $C$ is a $\sigma$-component.
\item The ideal $\mathfrak{a}$ is a $\sigma$-ideal.
\item The ideal $\mathfrak{p}$ is a $\sigma$-ideal.
\item There exists a prime $\sigma$-ideal in $C$.
\item There exists a $\sigma$-prime $\sigma$-ideal in $C$.
\item The set of all idempotent elements contained in $\mathfrak{a}$ is stable under $\sigma$.
\end{enumerate}
\end{lemma}
\begin{proof}
The equivalence of (i) and (iii) is straightforward.
As $\mathfrak{p}$ is the radical of $\mathfrak{a}$ we see that (ii) implies (iii). Clearly (iii) implies (iv). If $\mathfrak{q}$ is a $\sigma$-ideal in $C$, then its reflexive closure $\mathfrak{q}^*=\{f\in k\{G\}|\ \exists \ n\geq 0 : \sigma^n(f)\in\mathfrak{q}\}$ is a $\sigma$-prime $\sigma$-ideal in $C$ (cf. \cite[p. 107]{Levin:difference} and Section \ref{sec: Subgroups defined by ideal closures} below). So (iv) implies (v).
Let us next show that (v) implies (vi). If $\mathfrak{q}$ is a $\sigma$-prime $\sigma$-ideal in $C$, then by Lemma \ref{lemma:connectedcomponentsareirreducible} we have
$C=\mathcal{V}(\mathfrak{a}')$ where $\mathfrak{a}'\subset k\{G\}$ is the ideal generated by all idempotent elements contained in $\mathfrak{q}$. By \cite[Tag 00EB]{stacks-project} we must have $\mathfrak{a}'=\mathfrak{a}$. Now let $e\in\mathfrak{a}$ be an idempotent. We have to show that $\sigma(e)\in\mathfrak{a}$. But $e\in\mathfrak{q}$ (as $\mathfrak{q}\in C=\mathcal{V}(\mathfrak{a})$) and consequently $\sigma(e)\in\mathfrak{q}$ is also idempotent. So $\sigma(e)\in\mathfrak{a}'=\mathfrak{a}$.
Finally, the implication (vi)$\Rightarrow$(ii) follows from the simple fact that an ideal generated by a $\sigma$-stable set is a $\sigma$-ideal.
\end{proof}
\begin{cor} \label{cor: minsprime is minimal}
Let $k\{G\}$ be a $k$-$\sigma$-Hopf algebra which is finitely $\sigma$-generated over $k$. Then a minimal prime $\sigma$-ideal of $k\{G\}$ is a minimal prime ideal of $k\{G\}$.
\end{cor}
\begin{proof}
Let $\mathfrak{q}\subset k\{G\}$ be a minimal prime $\sigma$-ideal and let $C\subset\operatorname{Spec}(k\{G\})$ be the component which contains $\mathfrak{q}$. Then $C=\mathcal{V}(\mathfrak{p})$ for a minimal prime ideal $\mathfrak{p}$ of $k\{G\}$. Since $\mathfrak{q}\in C$ it follows from Lemma \ref{lemma: scomponent} that $\mathfrak{p}$ is a $\sigma$-ideal. Therefore $\mathfrak{q}=\mathfrak{p}$ by the minimality of $\mathfrak{q}$.
\end{proof}
We will next introduce $\sigma$-\'{e}tale $\sigma$-algebraic groups. The role of these groups in the theory of $\sigma$-algebraic groups is in a certain sense analogous to the role of \'{e}tale algebraic groups in the theory of algebraic groups. We plan to study $\sigma$-\'{e}tale $\sigma$-algebraic groups in more detail in a future paper. In particular, these groups are expected to satisfy a certain decomposition theorem. Here we will only use them to define the group of components and the identity component of a $\sigma$-algebraic group.
\begin{defi}
A finitely $\sigma$-generated $k$-$\sigma$-algebra $R$ is called \emph{$\sigma$-\'{e}tale} (over $k$) if $R$ is integral over $k$ and a separable $k$-algebra.
\end{defi}
Thus a finitely $\sigma$-generated $k$-$\sigma$-algebra $R$ is $\sigma$-\'{e}tale if and only if for every $r\in R$ there exists a separable polynomial $f$ over $k$ with $f(r)=0$. Similar notions of \'{e}taleness in difference algebra occur in \cite{Tomasic:ATwistedTheoremOfChebotarev} and \cite{Tomasic:TwistedGaloisStratification} in a slightly different setting.
\begin{defi}
A $\sigma$-algebraic group $G$ is called \emph{$\sigma$-\'{e}tale} if $k\{G\}$ is $\sigma$-\'{e}tale over $k$.
\end{defi}
\begin{theo} \label{theo: pi0}
Let $G$ be a $\sigma$-algebraic group. There exists a $\sigma$-\'{e}tale $\sigma$-algebraic group $\pi_0(G)$ and a morphism $G\to \pi_0(G)$ of $\sigma$-algebraic groups satisfying the following universal property: If $G\to H$ is a morphism of $\sigma$-algebraic groups with $H$ $\sigma$-\'{e}tale, then there exists a unique morphism $\pi_0(G)\to H$ such that
\[\xymatrix{
G \ar[rr] \ar[rd] & & \pi_0(G) \ar@{..>}[ld] \\
& H &
}
\]
commutes.
\end{theo}
\begin{proof}
Let $R\subset k\{G\}$ denote the set of all elements $r\in k\{G\}$ which annul a separable polynomial over $k$. (So $R$ is the union of all \'{e}tale subalgebras of $k\{G\}$.) Then $R$ is a $k$-subalgebra of $k\{G\}$. Indeed, $R$ is a Hopf subalgebra of $k\{G\}$. (Cf. Section 6.7 in \cite{Waterhouse:IntroductiontoAffineGroupSchemes}.)
Let $r\in R$ and $f$ be a separable polynomial over $k$ with $f(r)=0$. Let ${^\sigma\! f}$ denote the polynomial obtained from $f$ by applying $\sigma$ to the coefficients. Then ${^\sigma\! f}$ is separable and ${^\sigma\! f}(\sigma(r))=0$. This shows that $R$ is a $k$-$\sigma$-Hopf subalgebra of $k\{G\}$.
Let $\pi_0(G)$ denote the $\sigma$-\'{e}tale $\sigma$-algebraic group corresponding to $R$ and $G\to \pi_0(G)$ the morphism corresponding to the inclusion $R\subset k\{G\}$.
If $G\to H$ is a morphism of $\sigma$-algebraic groups with $H$ $\sigma$-\'{e}tale, then the image of the dual map $k\{H\}\to k\{G\}$ consists of elements that annul a separable polynomial. Thus the image lies in $R$ and $k\{H\}\to k\{G\}$ factors uniquely through $R\hookrightarrow k\{G\}$.
\end{proof}
Of course $\pi_0(G)$ is unique up to unique isomorphisms. It is clear from the above proof that $G\to\pi_0(G)$ is surjective.
\begin{defi}
Let $G$ be a $\sigma$-algebraic group. The $\sigma$-\'{e}tale $\sigma$-algebraic group $\pi_0(G)$ defined by the universal property from Theorem \ref{theo: pi0} is called the \emph{group of components} of $G$. The kernel $G^o$ of $G\twoheadrightarrow\pi_0(G)$ is called the \emph{identity component} of $G$.
\end{defi}
So $G/G^o=\pi_0(G)$.
\begin{lemma} \label{lemma: correspondence components}
Let $G$ be a $\sigma$-algebraic group. There is a one--to--one correspondence between the components of $G$ and the components of $\pi_0(G)$. Under this bijection $\sigma$-components correspond to $\sigma$-components. Moreover, every component of $\pi_0(G)$ consists of a single point.
\end{lemma}
\begin{proof}
Every prime ideal of $k\{\pi_0(G)\}$ is maximal and hence also minimal. This shows that the components of $\pi_0(G)$ are points. We identify $k\{\pi_0(G)\}$ with its image in $k\{G\}$. We claim that $\mathfrak{p}\mapsto \mathfrak{p}\cap k\{\pi_0(G)\}$ is a bijection between the minimal prime ideals of $k\{G\}$ and the (minimal) prime ideals of $k\{G\}$. Every (minimal) prime ideal of $k\{\pi_0(G)\}$ is of the form $\mathfrak{p}\cap k\{\pi_0(G)\}$ for some minimal prime ideal of $k\{G\}$ (\cite[Proposition 16, Chapter II, \S 2.6]{Bourbaki:commutativealgebra}). On the other hand, if $\mathfrak{p}$ is a minimal prime ideal of $k\{G\}$, then $\mathfrak{p}=\sqrt{\mathfrak{a}}$ for some ideal $\mathfrak{a}$ of $k\{G\}$ generated by idempotent elements. Since all idempotent elements of $k\{G\}$ lie in $k\{\pi_0(G)\}$, we see that $(\mathfrak{a}\cap k\{\pi_0(G)\})=\mathfrak{a}$. Therefore $\sqrt{(\mathfrak{p}\cap k\{\pi_0(G)\})}=\mathfrak{p}$.
If $\mathfrak{p}$ is a $\sigma$-ideal, then $\mathfrak{p}\cap k\{\pi_0(G)\}$ is a $\sigma$-ideal. Conversely, if $\mathfrak{p}'\subset k\{\pi_0(G)\}$ is a $\sigma$-ideal, then $\sqrt{(\mathfrak{p}')}\subset k\{G\}$ is a $\sigma$-ideal.
\end{proof}
\begin{prop} \label{prop: characterize connected}
The following four conditions on a $\sigma$-algebraic group $G$ are equivalent:
\begin{enumerate}
\item $G^o=G$.
\item $\pi_0(G)=1$.
\item The topological space of $G$ is connected.
\item The nilradical of $k\{G\}$ is a prime ideal.
\end{enumerate}
\end{prop}
\begin{proof}
Clearly, (i)$\Leftrightarrow$(ii). We have (ii)$\Leftrightarrow$(iii) by Lemma \ref{lemma: correspondence components}. Since the connected components are irreducible (Lemma \ref{lemma:connectedcomponentsareirreducible}), it follows from (iii) that $k\{G\}$ has a unique minimal prime ideal, which must then equal the nilradical. Thus (iii)$\Rightarrow$(iv). On the other hand (iv) means that the topological space of $G$ is irreducible. So (iv)$\Rightarrow$(iii).
\end{proof}
\begin{defi}
A $\sigma$-algebraic group satisfying the equivalent conditions of Proposition \ref{prop: characterize connected} is called \emph{connected}.
\end{defi}
Note that the identity component $G^o$ of a $\sigma$-algebraic group $G$ is, strictly speaking, not a component. It carries more structure than a mere component, in particular it has the structure of a $\sigma$-variety. However, as illustrated in the prof of the following lemma, the topological space of $G^o$ can be identified with the component of $G$ which contains the identity $\mathfrak{m}_{k\{G\}}$, i.e., the kernel of the counit $k\{G\}\to k$.
\begin{lemma}
Let $G$ be a $\sigma$-algebraic group. Then $G^o$ is connected.
\end{lemma}
\begin{proof}
Every ideal of $k\{\pi_0(G)\}$ is generated by idempotent elements. It follows that $\mathfrak{m}_{k\{\pi_0(G)\}}=\mathfrak{m}_{k\{G\}}\cap k\{\pi_0(G)\}$ is generated by all idempotent elements contained in $\mathfrak{m}_{k\{G\}}$. Therefore, $\mathbb{I}(G^o)$ is the ideal of $k\{G\}$ generated by all idempotent elements of $k\{G\}$ contained in $\mathfrak{m}_{k\{G\}}$. It follows from Lemma \ref{lemma:connectedcomponentsareirreducible}, that $\mathcal{V}(\mathbb{I}(G^o))\subset\operatorname{Spec}(k\{G\})$ is connected. As $\mathcal{V}(\mathbb{I}(G^o))$ and $\operatorname{Spec}(k\{G^o\})=\operatorname{Spec}(k\{G\}/\mathbb{I}(G^o))$ are homeomorphic, this implies that $G^o$ is connected.
\end{proof}
\begin{lemma} \label{lemma: connected component and zariski closure}
Let $G$ be a $\sigma$-closed subgroup of an algebraic group $\mathcal{G}$ and for $i\geq 0$ let $G[i]$ and $G^o[i]$ denote the $i$-th order Zariski closure of $G$ and $G^o$ in $\mathcal{G}$ respectively. Then
$$G^o[i]=G[i]^o.$$
In particular, $G$ is connected if and only if all its Zariski closures are connected.
\end{lemma}
\begin{proof} Both groups are defined by the ideal of $k[G[i]]\subset k\{G\}$ which is generated by all idempotent elements of $k[G[i]]$ contained in the kernel of the counit $k[G[i]]\to k$.
\end{proof}
\begin{cor} \label{cor: sdim of Go}
Let $G$ be a $\sigma$-algebraic group. Then $\sigma\text{-}\operatorname{dim}(G^o)=\sigma\text{-}\operatorname{dim}(G)$ and $\operatorname{ord}(G^o)=\operatorname{ord}(G)$.
\end{cor}
\begin{proof}
Let $\mathcal{G}$ be a $\sigma$-algebraic group containing $G$ as a $\sigma$-closed subgroup. Then for $i\geq 0$ we have
$\dim(G[i])=\dim(G[i]^o)=\dim(G^o[i])$ by Lemma \ref{lemma: connected component and zariski closure}. Thus the claim follows from Theorem \ref{theo: existence sdim}.
\end{proof}
The limit degree of $G$ and $G^o$ are in general distinct. Indeed ${\operatorname{ld}}(G)={\operatorname{ld}}(\pi_0(G)){\operatorname{ld}}(G^o)$ by Corollary \ref{cor: multiplicativity of limit degree}.
We will next show that a $\sigma$-algebraic group has only finitely man $\sigma$-components.
\begin{lemma} \label{lemma: setalimpliesfinite}
Let $R$ be a finitely $\sigma$-generated $k$-$\sigma$-algebra. If $R$ is $\sigma$-\'{e}tale, then $R$ has only finitely many prime $\sigma$-ideals.
\end{lemma}
\begin{proof}
Since $R$ is $\sigma$-\'{e}tale, every prime ideal of $R$ is maximal and hence also minimal. If $\mathfrak{p}$ is a prime $\sigma$-ideal of $R$, then $\mathfrak{p}\subset\sigma^{-1}(\mathfrak{p})$ and therefore $\mathfrak{p}=\sigma^{-1}(\mathfrak{p})$. So every prime $\sigma$-ideal of $R$ is $\sigma$-prime. It follows from the Ritt--Raudenbush basis theorem (cf. Theorems 2.5.11 and 2.5.7 in \cite{Levin:difference}) that a finitely $\sigma$-generated $k$-$\sigma$-algebra has only finitely many minimal $\sigma$-prime ideals. Since every prime $\sigma$-ideal of $R$ is a minimal $\sigma$-prime ideal of $R$, this implies that $R$ has only finitely many prime $\sigma$-ideals.
\end{proof}
\begin{theo} \label{theo: finiteness of scomponents}
A $\sigma$-algebraic group has only finitely many $\sigma$-components.
\end{theo}
\begin{proof}
Let $G$ be a $\sigma$-algebraic group. By Lemma \ref{lemma: correspondence components}, the $\sigma$-components of $G$ are in bijection with the $\sigma$-components of $\pi_0(G)$ and by Lemma \ref{lemma: setalimpliesfinite} the $\sigma$-algebraic group $\pi_0(G)$ has only finitely many $\sigma$-components.
\end{proof}
\begin{proof}[Proof of Theorem \ref{theo: Hrushovskiconj}]
By assumption $R=k\{G\}$ for a $\sigma$-algebraic group $G$. By Corollary \ref{cor: minsprime is minimal} the set of minimal prime $\sigma$-ideals of $k\{G\}$ equals the set of minimal prime ideals of $k\{G\}$ which are $\sigma$-ideals. The latter set is finite by Theorem \ref{theo: finiteness of scomponents}.
\end{proof}
\section{Subgroups defined by ideal closures} \label{sec: Subgroups defined by ideal closures}
If $\mathcal{G}$ is an algebraic group over a perfect field, then $\mathcal{G}_{\operatorname{red}}$, the associated reduced scheme, is a closed subgroup of $\mathcal{G}$. In difference algebra, there are several closure operations one can define on difference ideals which are in some way similar to taking the radical of an ideal. Therefore we obtain several $\sigma$-closed subgroups of a $\sigma$-algebraic group which are in some way analogous to $\mathcal{G}_{\operatorname{red}}$. Let us introduce now this closure operations on $\sigma$-ideals. (Cf. \cite[Section 2.3]{Levin:difference}.)
\begin{defi}
Let $R$ be a $\sigma$-ring and $\mathfrak{a}\subset R$ a $\sigma$-ideal. Then $\mathfrak{a}$ is called
\begin{itemize}
\item \emph{reflexive} if $\sigma^{-1}(\mathfrak{a})=\mathfrak{a}$, i.e., $\sigma(f)\in\mathfrak{a} $ implies $f\in\mathfrak{a}$.
\item \emph{mixed} if $fg\in\mathfrak{a}$ implies $f\sigma(g)\in\mathfrak{a}$.
\item \emph{perfect} if $\sigma^{\alpha_1}(f)\cdots\sigma^{\alpha_n}(f)\in\mathfrak{a}$ implies $f\in\mathfrak{a}$ for $\alpha_1,\ldots,\alpha_n\geq 0$
\end{itemize}
\end{defi}
A $\sigma$-ring whose zero ideal is reflexive / mixed / perfect is called $\sigma$-reduced / well--mixed / perfectly $\sigma$-reduced.
Note that a $\sigma$-ring which is an integral domain is well--mixed. If it is additionally $\sigma$-reduced, i.e., a $\sigma$-domain, it is perfectly $\sigma$-reduced.
Let $\mathfrak{a}$ be a $\sigma$-ideal of a $\sigma$-ring $R$. Since the intersection of reflexive / radical mixed / perfect $\sigma$-ideals is a reflexive / radical mixed / perfect $\sigma$-ideal there exists a smallest reflexive / radical mixed / perfect $\sigma$-ideal of $R$ containing $\mathfrak{a}$. It is called the
reflexive closure $\mathfrak{a}^*$/ the radical mixed closure $\{\mathfrak{a}\}_{\operatorname{wm}}$/ the perfect closure $\{\mathfrak{a}\}$ of $\mathfrak{a}$.
Let $X$ be a $\sigma$-variety. We say that $X$ is reduced / $\sigma$-reduced / reduced well--mixed / perfectly $\sigma$-reduced if $k\{X\}$ has this property. There exists a unique largest $\sigma$-closed $\sigma$-subvariety
$$X_{\operatorname{red}} \ / \ X_{{\sigma\text{-}\operatorname{red}}}\ / \ X_{\operatorname{wm}} \ / \ X_{\operatorname{per}}$$
of $X$ which is reduced / $\sigma$-reduced / reduced well--mixed / perfectly $\sigma$-reduced. Its defining ideal is the
radical / reflexive closure / radical mixed closure / perfect closure of the zero ideal of $k\{X\}$.
A perfect $\sigma$-ideal is reduced, mixed and reflexive. Therefore we have the following diagram of inclusions of $\sigma$-closed $\sigma$-subvarieties of $X$.
$$
\xymatrix{
& X & \\
X_{\operatorname{red}} \ar@{-}[ur] & & X_{\sigma\text{-}\operatorname{red}} \ar@{-}[ul] \\
X_{\operatorname{wm}} \ar@{-}[u]& & \\
& X_{\operatorname{per}} \ar@{-}[ul] \ar@{-}[uur] &
}
$$
The importance of perfectly $\sigma$-reduced $\sigma$-varieties stems from the fact that they correspond to the classical difference varieties as studied in \cite{Cohn:difference} and \cite{Levin:difference}, where one is only looking for solutions in $\sigma$-field extensions of $k$. Mixed $\sigma$-ideals play a crucial role in the theory of difference schemes as developed by E. Hrushovski in \cite{Hrushovski:elementarytheoryoffrobenius}. Note that for an arbitrary non--empty $\sigma$-variety $X_{\operatorname{per}}$ and $X_{\operatorname{wm}}$ might be empty. Take for example $k\{X\}=k\times k$ with $\sigma((a,b))=(\sigma(b),\sigma(a))$. This pathology does not occur for $\sigma$-algebraic groups since the kernel of the counit $\varepsilon\colon k\{G\}\to k$ is a $\sigma$-prime ideal.
\begin{ex}
Let $k$ be a $\sigma$-field, $\mathsf{G}$ a finite group and $\Sigma\colon \mathsf{G}\to \mathsf{G}$ a group endomorphism. Let $G$ be the $\sigma$-algebraic group from Example \ref{ex: split finite}. Then $G$ is $\sigma$-reduced if and only if $\Sigma$ is an automorphism. Moreover, $G$ is reduced well--mixed if and only if it is perfectly $\sigma$-reduced if and only if $\Sigma$ is the identity map.
\end{ex}
\begin{rem} \label{rem: no field points}
For a $\sigma$-algebraic group $G$ the following statements are equivalent:
\begin{enumerate}
\item $G(K)=1$ for every $\sigma$-field extension $K$ of $k$.
\item $G_{\operatorname{per}}=1$.
\item The kernel of the counit $k\{G\}\to k$ is the only $\sigma$-prime ideal of $k\{G\}$.
\end{enumerate}
\end{rem}
\begin{proof} This follows from the fact that a perfect difference ideal is the intersection of $\sigma$-prime ideals (\cite[Chapter 3, p. 88]{Cohn:difference}).
\end{proof}
\begin{rem}
If $G$ is a $\sigma$-algebraic group such that $k\{G\}$ is well--mixed, then $G$ has only finitely many components and they are all $\sigma$-components.
If $G$ is a $\sigma$-algebraic group such that $G_{\operatorname{wm}}$ is a $\sigma$-closed subgroup of $G$ (e.g., $k$ is algebraically closed, see Corollary \ref{cor: reduced ssubgroups}), then the $\sigma$-components of $G$ are in bijection with the components of $G_{\operatorname{wm}}$.
\end{rem}
\begin{proof}
By \cite[Lemma 2.10]{Hrushovski:elementarytheoryoffrobenius} a radical mixed $\sigma$-ideal is the intersection of prime $\sigma$-ideals.
If $k\{G\}$ is well--mixed, the nilradical of $k\{G\}$ is mixed and therefore it is the intersection of the minimal prime $\sigma$-ideals. Since there are only finitely many minimal prime $\sigma$-ideals in $k\{G\}$ (Theorem \ref{theo: Hrushovskiconj}) we see that there are only finitely many minimal prime ideals in $k\{G\}$ and all of them are $\sigma$-ideals.
Assume that $G$ is a $\sigma$-algebraic group such that $G_{\operatorname{wm}}$ is a $\sigma$-closed subgroup of $G$. The set of minimal prime ideals of $k\{G\}$ which are $\sigma$-ideals equals the set of minimal prime $\sigma$-ideals of $k\{G\}$ (Corollary \ref{cor: minsprime is minimal}) and the latter it the set of all prime ideals of $k\{G\}$ which are minimal above $\{0\}_{\operatorname{wm}}$.
\end{proof}
If $\psi\colon R\to S$ is a morphism of $\sigma$-rings, it is easy to check that $\psi^{-1}(\mathfrak{a})$ is a radical / reflexive / radical mixed / perfect $\sigma$-ideal if $\mathfrak{a}$ has the corresponding property. This shows that $\psi$ maps the radical / reflexive closure / radical mixed closure / perfect closure of the zero ideal of $R$ into the radical / reflexive closure / radical mixed closure / perfect closure of the zero ideal of $S$.
Therefore a morphism of $\sigma$-varieties $X\to Y$ induces a morphism
$$X_{\operatorname{red}}\to Y_{\operatorname{red}} \ / \ X_{{\sigma\text{-}\operatorname{red}}}\to Y_{{\sigma\text{-}\operatorname{red}}} \ / \ X_{\operatorname{wm}}\to Y_{{\operatorname{wm}}} \ / \ X_{\operatorname{per}}\to Y_{\operatorname{per}}.$$
For later use we record a lemma on perfectly $\sigma$-reduced $\sigma$-varieties.
\begin{lemma} \label{lemma: morphism with perfectly sreduced}
Let $\phi\colon X\to Y$ be a morphism of $\sigma$-varieties and let $Z\subset Y$ be a $\sigma$-closed $\sigma$-subvariety. Assume that $X$ is perfectly $\sigma$-reduced. If $\phi_K(X(K))\subset Z(K)$ for every $\sigma$-field extension $K$ of $k$, then $\phi(X)\subset Z$, i.e., $\phi$ factors through $Z\hookrightarrow Y$.
\end{lemma}
\begin{proof}
We have to show that $\mathbb{I}(Z)\subset k\{Y\}$ lies in the kernel of $\phi^*\colon k\{Y\}\to k\{X\}$. So let $f\in\mathbb{I}(X)$. We have to show that $\phi^*(f)=0$. Since the zero ideal of $k\{X\}$ is perfect, it is the intersection of $\sigma$-prime ideals (\cite[Chapter 3, p. 88]{Cohn:difference}). Therefore, it suffices to show that $\phi^*(f)$ lies in every $\sigma$-prime ideal of $k\{X\}$. Let $\mathfrak{p}\subset k\{X\}$ be a $\sigma$-prime ideal, then the field of fractions $K$ of $k\{X\}/\mathfrak{p}$ naturally is a $\sigma$-field and the canonical map $x\colon k\{X\}\to K$ is a morphism of $k$-$\sigma$-algebras. By assumption, $\phi_K(x)\in Z(K)$, i.e., $\mathbb{I}(Z)$ lies in the kernel of $x\circ\phi^*$. So $\phi^*(f)\in\mathfrak{p}$.
\end{proof}
\begin{lemma} \label{lemma: tensor stays}
Let $R$ and $S$ be $k$-$\sigma$-algebras.
\begin{enumerate}
\item If $k$ is perfect and $R$ and $S$ are reduced, then $R\otimes_k S$ is reduced.
\item If $k$ is inversive and $R$ and $S$ are $\sigma$-reduced, then $R\otimes_k S$ is $\sigma$-reduced.
\item If $k$ is algebraically closed and $R$ and $S$ are well--mixed and reduced, then $R\otimes_k S$ is well--mixed and reduced.
\item If $k$ is inversive and algebraically closed and $R$ and $S$ are perfectly $\sigma$-reduced, then $R\otimes_k S$ is perfectly $\sigma$-reduced.
\end{enumerate}
\end{lemma}
\begin{proof}
Of course (i) is well--known. See e.g., \cite[Theorem 3, Chapter V, \S 15.5, A.V.125]{Bourbaki:Algebra2}. Note that (i) is a special case of (ii) as we may take $\sigma$ as the Frobenius endomorphism. For (ii), first note that if $(r_i)_{i\in I}$ is a $k$-basis of $R$, then $(\sigma(r_i))_{i\in I}$ is $k$-linearly independent: If $\sum \lambda_i\sigma(r_i)=0$ we can write $\lambda_i=\sigma(\mu_i)$ as $k$ is inversive and then $0=\sigma(\sum\mu_ir_i)$ implies $\sum\mu_ir_i=0$ as $R$ is $\sigma$-reduced. So $\mu_i=0$ and therefore $\lambda_i=0$ as claimed. Now let $f=\sum r_i\otimes s_i\in R\otimes_k S$ with $\sigma(f)=0$. Then $\sum \sigma(r_i)\otimes\sigma(s_i)=0$. But since $(\sigma(r_i))_{i\in I}$ is $k$-linearly independent this implies $\sigma(s_i)=0$ for all $i\in I$. As $S$ is $\sigma$-reduced it follows that $f=0$.
For (iii), note that the zero ideal of a reduced well-mixed $\sigma$-ring is the intersection of prime $\sigma$-ideals (\cite[Lemma 2.10]{Hrushovski:elementarytheoryoffrobenius}). If $\mathfrak{p}$ is a prime $\sigma$-ideal of $R$ and $\mathfrak{q}$ a prime $\sigma$-ideal of $S$, then $\mathfrak{p}\otimes S+R\otimes \mathfrak{q}$ is a prime $\sigma$-ideal of $R\otimes_k S$ since
$$(R\otimes_k S)/(\mathfrak{p}\otimes S+R\otimes \mathfrak{q})=R/\mathfrak{p}\otimes_k S/\mathfrak{q}$$
and the latter is an integral domain, as the tensor product of integral domains over an algebraically closed field is again an integral domain (\cite[Corollary 3, Chapter V, \S 17.5, A.V.143]{Bourbaki:Algebra2}). We see that the zero ideal of $R\otimes_k S$ is the intersection of prime $\sigma$-ideals of the form $\mathfrak{p}\otimes S+R\otimes \mathfrak{q}$. This shows that $R\otimes_k S$ is well--mixed and reduced.
To prove (iv) we can proceed as in (iii) by noting that a $\sigma$-ideal is perfect if and only if it is the intersection of $\sigma$-prime ideals and that the tensor product of $\sigma$-domains over an inversive algebraically closed $\sigma$-field is again a $\sigma$-domain by (ii).
\end{proof}
There are counterexamples which show that the conditions on the base $\sigma$-field in Lemma \ref{lemma: tensor stays} can not be relaxed. For example, take $k=\mathbb{R}$ with $\sigma$ the identity map, $R=\mathbb{C}$ with the identity map and $S=\mathbb{C}$ with $\sigma$ complex conjugation. Then $R$ and $S$ are perfectly $\sigma$-reduced (hence well--mixed) but $R\otimes_k S$ is not well-mixed (hence not perfectly $\sigma$-reduced).
If $X$ and $Y$ are $\sigma$-varieties, the canonical map $(X\times Y)_{\operatorname{per}}\to X_{\operatorname{per}}\times Y_{\operatorname{per}}$ need not be an isomorphism as $X_{\operatorname{per}}\times Y_{\operatorname{per}}$ need not be perfectly $\sigma$-reduced.
\begin{cor} \label{cor: reduced products}
Let $X$ and $Y$ be $\sigma$-varieties.
\begin{enumerate}
\item If $k$ is perfect, then $(X\times Y)_{\operatorname{red}}\simeq X_{\operatorname{red}}\times Y_{\operatorname{red}}$.
\item If $k$ is inversive, then $(X\times Y)_{\sigma\text{-}\operatorname{red}}\simeq X_{\sigma\text{-}\operatorname{red}}\times Y_{\sigma\text{-}\operatorname{red}}$.
\item If $k$ is algebraically closed, then $(X\times Y)_{\operatorname{wm}}\simeq X_{\operatorname{wm}}\times Y_{\operatorname{wm}}$.
\item If $k$ is inversive and algebraically closed, then $(X\times Y)_{\operatorname{per}}\simeq X_{\operatorname{per}}\times Y_{\operatorname{per}}$.
\end{enumerate}
\end{cor}
\begin{proof}
Exemplarily, let us proof (iv). In terms of $k$-$\sigma$-algebras, we have to show that the canonical map
$$k\{X\}/\{0\}\otimes_k k\{Y\}/\{0\}\to (k\{X\}\otimes_k k\{Y\})/\{0\}$$
is an isomorphism. As the left hand side is perfectly $\sigma$-reduced by Lemma \ref{lemma: tensor stays}, we see that $\{0\}=\{0\}\otimes k\{Y\}+k\{X\}\otimes \{0\}$.
\end{proof}
\begin{cor} \label{cor: reduced ssubgroups}
Let $G$ be a $\sigma$-algebraic group.
\begin{enumerate}
\item If $k$ is perfect, then $G_{\operatorname{red}}$ is a $\sigma$-closed subgroup of $G$.
\item If $k$ is inversive, then $G_{\sigma\text{-}\operatorname{red}}$ is a $\sigma$-closed subgroup of $G$.
\item If $k$ is algebraically closed, then $G_{\operatorname{wm}}$ is a $\sigma$-closed subgroup of $G$.
\item If $k$ is inversive and algebraically closed, then $G_{\operatorname{per}}$ is a $\sigma$-closed subgroup of $G$.
\end{enumerate}
\end{cor}
\begin{proof}
Again, let us restrict to (iv). The other cases are similar. The multiplication morphism $G\times G\to G$ induces a morphism
$(G\times G)_{\operatorname{per}} \to G_{\operatorname{per}}$. But by Corollary \ref{cor: reduced products}, the $\sigma$-closed $\sigma$-subvariety $(G\times G)_{\operatorname{per}}$ of $G\times G$ can be identified with $G_{\operatorname{per}}\times G_{\operatorname{per}}\subset G\times G$. Therefore, the multiplication maps $G_{\operatorname{per}}\times G_{\operatorname{per}}$ into $G_{\operatorname{per}}$.
As the inversion $G\to G,\ g\mapsto g^{-1}$ also passes to $G_{\operatorname{per}}$, we see that $G_{\operatorname{per}}$ is a subgroup of $G$.
\end{proof}
\begin{ex}
For all the $\sigma$-algebraic groups $G$ in the list before Lemma \ref{lemma:connectedcomponentsareirreducible} $G_{\operatorname{per}}$ is the trivial group.
\end{ex}
The following example shows that $G_{\sigma\text{-}\operatorname{red}}$ need to be a subgroup if $k$ is not inversive.
\begin{ex} \label{ex: Gsred not subgroup}
Let $k$ be a $\sigma$-field of characteristic zero which is not inversive. So there exists $\lambda\in k$ with $\lambda\notin \sigma(k)$. Let $G$ be the $\sigma$-closed subgroup of $\mathbb{G}_a$ given by
$$G(R)=\{g\in R|\ \sigma^2(g)+\lambda\sigma(g)=0\}$$
for any $k$-$\sigma$-algebra $R$. We will show that $G$ has no proper, non--trivial $\sigma$-closed subgroup. Suppose that $H$ is a proper, non--trivial $\sigma$-closed subgroup of $G$. By Corollary A.3 in \cite{diVizioHardouinWibmer:DifferenceAlgebraicRelations} every $\sigma$-closed subgroup of $\mathbb{G}_a$ is of the form $\mathbb{V}(f)$, where $f\in k\{y\}$ is the unique monic linear homogeneous difference polynomial of minimal order in $\mathbb{I}(H)\subset k\{\mathbb{G}_a\}=k\{y\}$. As $H$ is non--trivial and properly contained in $G$, $f$ must have order one, i.e., $f=\sigma(y)+\mu y$ for some $\mu\in k$. But then $\sigma^2(h)+\sigma(\mu)\sigma(h)=0$ and therefore $(\lambda-\sigma(\mu))h=0$ for all $h\in H(R)$ for any $k$-$\sigma$-algebra $R$.
Thus $\lambda=\sigma(\mu)$; a contradiction.
Now assume that $\lambda^2\in\sigma(k)$. (For example, we can choose
$k=\mathbb{C}(\sqrt{x},\sqrt{x+1},\ldots)$ with action of $\sigma$ determined by $\sigma(x)=x+1$ and $\lambda=\sqrt{x}$.)
We have $k\{G\}=k[y,\sigma(y)]$ and if we choose $\eta\in k$ such that $\sigma(\eta)=\lambda^2$, then $\sigma(y)^2-\eta y^2$ lies in the reflexive closure of the zero ideal of $k\{G\}$. As $y$ does not lie in the reflexive closure of the zero ideal, $G_{\sigma\text{-}\operatorname{red}}$ is not the trivial group but properly contained in $G$. So, by the above, $G_{\sigma\text{-}\operatorname{red}}$ can not be a subgroup.
\end{ex}
The following example shows that the $\sigma$-closed subgroups constructed in Corollary \ref{cor: reduced ssubgroups} are in general not normal.
\begin{ex} \label{ex: Gsred not normal}
Let $N$ be the $\sigma$-closed subgroup of $\mathbb{G}_a$ given by $N(R)=\{g\in R|\ \sigma(g)=0\}$ for any $k$-$\sigma$-algebra $R$. The $\sigma$-algebraic group $H=\mathbb{G}_m$ acts on $N$ by group automorphisms
$$H(R)\times N(R)\to N(R),\ (h,n)\mapsto hn.$$
So we can form the semidirect product $G=N\rtimes H$ which is the $\sigma$-variety $G\times N$ with group multiplication given by
$$(n_1,h_1)\cdot (n_2,h_2)=(n_1+h_1n_2,h_1h_2).$$
Then $k\{G\}=k\{N\}\otimes_k k\{H\}=k[x]\otimes_k k\{y,y^{-1}\}$ with $\sigma(x)=0$. The reflexive closure of the zero ideal of $k\{G\}$ is the ideal of $k\{G\}$ generated by $x$. Therefore $G_{\sigma\text{-}\operatorname{red}}=H\leq G$. For $h\in H(R)$ and $n\in N(R)$ we have
$$(n,1)(0,h)(n,1)^{-1}=(n-hn,h)$$ which shows that $G_{\sigma\text{-}\operatorname{red}}$ is not normal in $G$.
\end{ex}
In the following lemma we tacitly assume that $k$ has the relevant properties as stated in Corollary \ref{cor: reduced ssubgroups}, so that we are dealing with $\sigma$-closed subgroups.
\begin{lemma} \label{lemma: sdim of sssubgroups}
Let $G$ be a $\sigma$-algebraic group.
Then $\sigma\text{-}\operatorname{dim}(G_{\operatorname{red}})$, $\sigma\text{-}\operatorname{dim}(G_{\sigma\text{-}\operatorname{red}})$, $\sigma\text{-}\operatorname{dim}(G_{\operatorname{wm}})$ and $\sigma\text{-}\operatorname{dim}(G_{\operatorname{per}})$ are all equal to $\sigma\text{-}\operatorname{dim}(G)$.
\end{lemma}
\begin{proof}
As the dimension of a finitely generated $k$-algebra remains invariant if we factor by the nilradical, it follows easily that $\sigma\text{-}\operatorname{dim}(G_{\operatorname{red}})=\sigma\text{-}\operatorname{dim}(G)$.
Since $(G^o)_{\sigma\text{-}\operatorname{red}}\leq G_{\sigma\text{-}\operatorname{red}}$, $(G^o)_{\operatorname{wm}}\leq G_{\operatorname{wm}}$ and $(G^o)_{\operatorname{per}}\leq G_{\operatorname{per}}$ we may assume that $G$ is connected by Lemma \ref{cor: sdim of Go}.
But then the nilradical of $k\{G\}$ is a prime $\sigma$-ideal (Proposition \ref{prop: characterize connected}) and therefore $G_{\operatorname{wm}}=G_{\operatorname{red}}$ and thus $\sigma\text{-}\operatorname{dim}(G_{\operatorname{wm}})=\sigma\text{-}\operatorname{dim}(G)$ also in this case.
To prove $\sigma\text{-}\operatorname{dim}(G_{\operatorname{per}})=\sigma\text{-}\operatorname{dim}(G)$ we may assume that $G$ is reduced. Then the zero ideal of $k\{G\}$ is prime and therefore its reflexive closure $\cup_{i\geq 1}\sigma^{-i}(0)$ is a $\sigma$-prime ideal. This shows that $G_{\operatorname{per}}=G_{\sigma\text{-}\operatorname{red}}$.
It thus suffices to show that $\sigma\text{-}\operatorname{dim}(G_{\sigma\text{-}\operatorname{red}})=\sigma\text{-}\operatorname{dim}(G)$. Let $\mathcal{G}$ be an algebraic group containing $G$ as a $\sigma$-closed subgroup and for $i\geq 0$ let $G[i]$ and $G_{\sigma\text{-}\operatorname{red}}[i]$ denote the $i$-th order Zariski closure of $G$ and $G_{\sigma\text{-}\operatorname{red}}$ in $\mathcal{G}$ respectively. By Corollary \ref{cor: finiteness1} there exists $m\geq 0$ such that $$\mathbb{I}(G_{\sigma\text{-}\operatorname{red}}[i])=\big(\mathbb{I}(G_{\sigma\text{-}\operatorname{red}}[i-1]), \sigma(\mathbb{I}(G_{\sigma\text{-}\operatorname{red}}[i-1]))\big)\subset k[G[i]]\subset k\{G\}$$ for $i>m$. But $\mathbb{I}(G_{\sigma\text{-}\operatorname{red}})\subset k\{G\}$ is the reflexive closure of the zero ideal and so $\mathbb{I}(G_{\sigma\text{-}\operatorname{red}})=\{f\in k\{G\}|\ \exists\ n\geq 1 : \sigma^n(f)=0\}$ (cf. \cite[p. 107]{Levin:difference}). This shows that there exist $f_1,\ldots,f_m$ in $k\{G\}$ such that $\mathbb{I}(G_{\sigma\text{-}\operatorname{red}}[i])=(f_1,\ldots,f_m)\subset k[G[i]]$ for $i\gg 0$. Therefore $\dim(G[i])-\dim(G_{\sigma\text{-}\operatorname{red}}[i])\leq m$ for $i\gg 0$ and consequently $\sigma\text{-}\operatorname{dim}(G)=\sigma\text{-}\operatorname{dim}(G_{\sigma\text{-}\operatorname{red}})$.
\end{proof}
Note that the order of $G_{\sigma\text{-}\operatorname{red}}$ might be strictly smaller than the order of $G$. This for example is the case for $G\leq\mathbb{G}_a$ given by
$G(R)=\{g\in R|\ \sigma(g)=0\}$ for any $k$-$\sigma$-algebra $R$.
\section{Group theory} \label{sec: Group theory}
Ultimately the goal of this section is to prove the analogs of the (Noether) isomorphism theorems for groups. This task is complicated by the fact that for a normal $\sigma$-closed subgroup $N$ of a $\sigma$-algebraic group $G$, the quotient $G/N$ is, at least initially, not easy accessible.
As for algebraic groups, the functor $R\rightsquigarrow G(R)/N(R)$ need not be representable, in particular, it is distinct from $G/N$.
If we could identify $G$ with $G(k)$, and $(G/N)(k)$ with $G(k)/N(k)$, where $k$ is some ``sufficiently large'' difference field, we could apply the isomorphism theorems directly. This approach is not possible for us for several reasons. Firstly, we want to avoid restrictions on the base field. Secondly, the identification of $G$ with $G(k)$ is only feasible for perfectly $\sigma$-reduced $\sigma$-algebraic groups. Thirdly, identifying $G$ with $G(k)$ does introduce some blemishes. For example, the morphism of $\sigma$-algebraic groups $\phi\colon \operatorname{GL}_n\to \operatorname{GL}_n,\ g\mapsto\sigma(g)$ is such that $\phi_k\colon\operatorname{GL}_n(k)\to\operatorname{GL}_n(k)$ has trivial kernel, but $\phi$ is not a $\sigma$-closed embedding. Indeed, the kernel of $\phi$ is non--trivial.
For algebraic groups, the theory of sheaves (see \cite[Chapter III]{DemazureGabriel:GroupesAlgebriques}) provides an elegant and powerful tool to deal with quotients. In the first part of this section we adapt the theory of sheaves to difference algebra. The main result (Theorem \ref{theo: associated sheaf}) provides a canonical way to associate a sheaf to any functor from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$. The relevance of Theorem \ref{theo: associated sheaf} for quotients stems from the fact that the sheaf associated to the functor $R\rightsquigarrow G(R)/N(R)$ equals $G/N$. In the second part we then show how this result can be used to deduce the isomorphism theorems for difference algebraic groups from the isomorphism theorems for (abstract) groups. While it may be possible to prove the isomorphism theorems for difference algebraic groups without explicitly introducing sheaves, we expect that Theorem \ref{theo: associated sheaf} will also be useful in other situations, for example, when considering actions of difference algebraic groups on difference varieties, where the existence of a quotient in the category of difference varieties is problematic.
\subsection{Sheaves}
Our proof of Theorem \ref{theo: associated sheaf} follows the proof of Theorem 1.8, Chapter III, \S 1 \cite{DemazureGabriel:GroupesAlgebriques} rather closely.
Let $k$ be a $\sigma$-field. If $(R_i)_{i\in I}$ is a finite family of $k$-$\sigma$-algebras, then the product $\prod_{i\in I}R_i$ is naturally a $k$-$\sigma$-algebra by $\sigma((r_i)_{i\in I})=(\sigma(r_i))_{i\in I}$. The projections $\prod_{i\in I}R_i\to R_i$ are morphisms of $k$-$\sigma$-algebras.
Recall that a morphism $\psi\colon R\to S$ of $k$-$\sigma$-algebras is called faithfully flat if the corresponding morphism $\psi^\sharp\colon R^\sharp\to S^\sharp$ of $k$-algebras is faithfully flat.
\begin{defi} Let $R$ be a $k$-$\sigma$-algebra. A finite family $(R_i)_{i\in I}$ of $R$-$\sigma$-algebras is called \emph{$R$-covering} if the canonical map $R\to\prod_{i\in I}R_i$ is faithfully flat.
\end{defi}
Note that if $(R_i)_{i\in I}$ and $(S_j)_{j\in J}$ are two $R$-covering families, then $(R_i\otimes_R S_j)_{(i,j)\in I\times J}$ is an $R$-covering family. If $R\to S$ is a morphism of $k$-$\sigma$-algebras and $(R_i)_{i\in I}$ is an $R$-covering family, then $(R_i\otimes_R S)_{i\in I}$ is an $S$-covering family.
Recall that a sequence of sets $\xymatrix@1{A \ar[r]^\alpha & B \ar@<0.5ex>[r]^{\beta_1} \ar@<-0.5ex>[r]_{\beta_2}& C}$ is called exact if $\alpha$ is the equalizer of $\beta_1$ and $\beta_2$, i.e., $\beta_1\alpha=\beta_2\alpha$ and for $b\in B$ with $\beta_1(b)=\beta_2(b)$ there exists a unique $a\in A$ with $\alpha(a)=b$.
\begin{defi}
Let $F$ be a functor from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$. Then $F$ is called a \emph{sheaf}, if for every $k$-$\sigma$-algebra $R$ and every $R$-covering family $(R_i)_{i\in I}$ the sequence
\begin{equation} \label{eqn: sheaf}
\xymatrix{F(R) \ar[r]^-\alpha & \prod_iF(R_i) \ar@<0.5ex>[r]^-{\beta_1} \ar@<-0.5ex>[r]_-{\beta_2} & \prod_{i,j} F(R_i\otimes_R R_j)}
\end{equation}
is exact.
\end{defi}
The maps in the above sequence are the obvious ones: The $i$-th component of $\alpha$ is induced from $R\to R_i$. The $(i,j)$-component of $\beta_1$ (respectively $\beta_2$) is the projection onto $F(R_i)$ (respectively $F(R_j)$) followed by the map induced from $R_i\to R_i\otimes_R R_j,\ f\mapsto f\otimes 1$ (respectively $R_j\to R_i\otimes_R R_j,\ f\mapsto 1\otimes f)$.
A morphism of sheaves is a morphism of functors.
\begin{lemma} \label{lemma: sheaves}
A functor $F$ from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$ is a \emph{sheaf} if and only if
\begin{enumerate}
\item for every finite family of $k$-$\sigma$-algebras $(R)_{i\in I}$, the map $F(\prod_{i\in I}R_i)\to \prod_{i\in I} F(R_i)$ is bijective;
\item for every faithfully flat morphism of $k$-$\sigma$-algebras $R\to S$ the sequence
$$F(R)\to F(S)\rightrightarrows F(S\otimes_R S)$$
is exact.
\end{enumerate}
\end{lemma}
\begin{proof}
Let us first assume that $F$ is a sheaf. If we take $R$ as the zero ring and the $R$-covering family as the empty family, then the exactness of (\ref{eqn: sheaf}) signifies that $F(R)$ consists of one element.
Let $(R_i)_{i\in I}$ be a finite family of $k$-$\sigma$-algebras. If we set $R=\prod_i R_i$ and consider $R_i$ as a $R$-$\sigma$-algebra via the projection $R\to R_i$, then the canonical map $R\to \prod_i R_i$ is the identity and $R_i\otimes_R R_j=0$ for $i\neq j$, whereas $R_i\otimes_R R_j=R_i$ for $i=j$. Thus the exactness of (\ref{eqn: sheaf}) in this case yields condition (i) since $\beta_1$ and $\beta_2$ are reduced to the identity map.
Clearly a sheaf satisfies condition (ii).
Conversely, assume that $F$ satisfies (i) and (ii). Let $(R_i)_{i\in I}$ be an $R$-covering family. Taking $S=\prod R_i$ and applying (i) to (ii) yields the exactness of (\ref{eqn: sheaf}).
\end{proof}
\begin{cor}
Every representable functor from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$ is a sheaf. In particular, every $\sigma$-variety and every $\sigma$-algebraic group is a sheaf.
\end{cor}
\begin{proof}
Let $A$ be a $k$-$\sigma$-algebra and $F=\operatorname{Hom}(A,-)$. Clearly,
$$\operatorname{Hom}(A,\textstyle{\prod_i} R_i)\simeq\textstyle{\prod_i}\operatorname{Hom}(A,R_i),$$
for every finite family $(R_i)_{i\in I}$ of $k$-$\sigma$-algebras. So $F$ satisfies condition (i) of Lemma \ref{lemma: sheaves}.
To verify (ii), let $R\to S$ be a faithfully flat morphism of $k$-$\sigma$-algebras. Then
$$R\to S\rightrightarrows S\otimes_R S$$
is exact. (See e.g. \cite[Section 13.1, p. 104]{Waterhouse:IntroductiontoAffineGroupSchemes}.) Therefore
$$\operatorname{Hom}(A,R)\to\operatorname{Hom}(A,S)\rightrightarrows\operatorname{Hom}(A,S\otimes_R S)$$ is exact.
\end{proof}
Thus the category of $\sigma$-varieties is a full subcategory of the category of sheaves.
\begin{defi} \label{defi: fat subfunctor}
A subfunctor $D$ of a functor $F\colon\text{$k$-$\sigma$-$\mathsf{Alg}$}\to \mathsf{Sets}$ is called \emph{fat} if for every $k$-$\sigma$-algebra $R$ and every $a\in F(R)$ there exists an $R$-covering family $(R_i)_{i\in I}$ such that the image of $a$ in $F(R_i)$ belongs to $D(R_i)$ for every $i\in I$.
\end{defi}
Some of the constructions with $\sigma$-varieties explained in Section \ref{sec: Difference algebraic geometry preliminaries} carry over to arbitrary functors from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$ without difficulty: If $\phi\colon E\to F$ is a morphism of functors and $D\subset F$ a subfunctor, then $\operatorname{Im}(\phi)$ and $\phi^{-1}(D)$ may be defined as in Section \ref{sec: Difference algebraic geometry preliminaries}; similarly for the intersection of subfunctors.
Note that if $\phi\colon E\to F$ is a morphism of functors and $D\subset F$ is fat, then $\phi^{-1}(D)\subset E$ is fat. As in \cite[p. 285]{DemazureGabriel:GroupesAlgebriques} one veryfies: If $D$ is a fat subfunctor of $F$ and $D_1$ is a fat subfunctor of $D$, then $D_1$ is a fat subfunctor of $F$. If $D_1$ and $D_2$ are fat subfunctors of $E$, then $D_1\cap D_2$ is a fat subfunctor of $E$.
Let $R$ be a $k$-$\sigma$-algebra. To simplify the notation we will write
$$\operatorname{Sp}^\sigma(R)$$
for the functor $\operatorname{Hom}(R,-)\colon \text{$k$-$\sigma$-$\mathsf{Alg}$}\to \mathsf{Sets}$. If $\psi\colon R\to S$ is a morphism of $k$-$\sigma$-algebras, there is an induced morphism of functors $\operatorname{Sp}^\sigma(\psi)\colon \operatorname{Sp}^\sigma(S)\to\operatorname{Sp}^\sigma(R)$. Recall (See e.g. \cite[Lemma VI-1 (a)]{EisenbudHarris:TheGeometryOfSchemes}.) that for any functor $F\colon \text{$k$-$\sigma$-$\mathsf{Alg}$}\to \mathsf{Sets}$
and any $k$-$\sigma$-algebra $R$ there is a canonical bijection
$$F(R)\simeq\operatorname{Hom}(\operatorname{Sp}^\sigma(R),F)$$
Indeed, $F$ and $\operatorname{Hom}(\operatorname{Sp}^\sigma(-),F)$ are isomorphic as functors. An element $a\in F(R)$ gives rise to morphism $\operatorname{Sp}^\sigma(R)\to F$ as follows: For a $k$-$\sigma$-algebra $S$ and $\psi\in\operatorname{Hom}(R,S)$ let $\operatorname{Sp}^\sigma(R)(S)\to F(S)$ be defined by sending $\psi$ to $F(\psi)(a)$. Conversely, from a morphism $\operatorname{Sp}^\sigma(R)\to F$ we obtain an element in $F(R)$ by considering the image of the identity under $\operatorname{Sp}^\sigma(R)(R)\to F(R)$.
\begin{lemma} \label{lemma: fat subfunctor of sps}
Let $R$ be a $k$-$\sigma$-algebra. Then a subfunctor $D$ of $\operatorname{Sp}^\sigma(R)$ is fat if and only if there exists an $R$-covering family $(R_i)_{i\in I}$ such that $\operatorname{Im}(\operatorname{Sp}^\sigma(R_i)\to\operatorname{Sp}^\sigma(R))\subset D$ for all $i\in I$.
\end{lemma}
\begin{proof}
If $D\subset\operatorname{Sp}^\sigma(R)$ is fat, we can apply Definition \ref{defi: fat subfunctor} to the element $a=\operatorname{id}_R\in\operatorname{Hom}(R,R)=\operatorname{Sp}^\sigma(R)(R)$ to find an $R$-covering family $(R_i)_{i\in I}$ such that the maps $R\to R_i$ lie in $D(R_i)\subset\operatorname{Sp}^\sigma(R)(R_i)$ for every $i\in I$. Let $\psi\colon R_i\to S$ be a morphism of $k$-$\sigma$-algebras. Since $\operatorname{Sp}^\sigma(R)(\psi)\colon\operatorname{Sp}^\sigma(R)(R_i)\to \operatorname{Sp}^\sigma(R)(S)$ maps $D(R_i)$ into $D(S)$ we see that the composition $R\to R_i\to S$ lies in $D(S)\subset \operatorname{Sp}^\sigma(R)(S)$. Therefore the image of $\operatorname{Sp}^\sigma(R_i)(S)\to\operatorname{Sp}^\sigma(R)(S)$ lies in $D(S)$, i.e., $\operatorname{Im}(\operatorname{Sp}^\sigma(R_i)\to\operatorname{Sp}^\sigma(R))\subset D$.
Conversely, if $(R_i)_{i\in I}$ verifies the condition of the lemma and $a\colon R\to S$ is a morphism of $k$-$\sigma$-algebras, then $(R_i\otimes_R S)_{i\in I}$ is an $S$-covering family. By assumption, $R\to R_i\to R_i\otimes_R S$ lies in $D(R_i\otimes_R S)\subset \operatorname{Sp}^\sigma(R)(R_i\otimes_R S)$. But $R\to R_i\to R_i\otimes_R S$ equals $R\to S\to R_i\otimes_R S$. Therefore $\operatorname{Sp}^\sigma(R)(S)\to \operatorname{Sp}^\sigma(R)(R_i\otimes_R S)$ maps $a$ into $D(R_i\otimes_R S)$.
\end{proof}
\begin{prop} \label{prop: characterize sheaf}
For a functor $F\colon\text{$k$-$\sigma$-$\mathsf{Alg}$}\to\mathsf{Sets}$ the following properties are equivalent:
\begin{enumerate}
\item The functor $F$ is a sheaf.
\item If $D$ is a fat subfunctor of a functor $E\colon\text{$k$-$\sigma$-$\mathsf{Alg}$}\to\mathsf{Sets}$ and $\phi\colon D\to F$ a morphism, then there exists a unique extension of $\phi$ to $E$.
\item If $R$ is a $k$-$\sigma$-algebra, $D$ a fat subfunctor of $\operatorname{Sp}^\sigma(R)$ and $\phi\colon D\to F$ a morphism, then there exists a unique extension of $\phi$ to $\operatorname{Sp}^\sigma(R)$.
\end{enumerate}
\end{prop}
\begin{proof}
Let us start with (i)$\Rightarrow$(ii). Let $R$ be a $k$-$\sigma$-algebra and $a\in E(R)$. We have to construct $\phi(a)\in F(R)$. Since $D$ is fat, there exists an $R$-covering family $(R_i)_{i\in I}$ such that the image $a_i$ of $a$ in $E(R_i)$ belongs to $D(R_i)$. Since $a\in E(R)$ we see that $(a_i)_{i\in I}\in \prod_i D(R_i)\subset \prod_i E(R_i)$ lies in the equalizer of the top row of
$$
\xymatrix{
\prod_{i} D(R_i) \ar@<0.5ex>[r] \ar@<-0.5ex>[r] \ar[d] & \prod_{i,j} D(R_i\otimes_R R_j) \ar[d]\\
\prod_i F(R_i) \ar@<0.5ex>[r] \ar@<-0.5ex>[r] & \prod_{i,j} F(R_i\otimes_R R_j).
}
$$
Therefore $(\phi(a_i))_{i\in I}\in \prod_i F(R_i) $ lies in the equalizer of the bottom row. Since $F$ is a sheaf there exists a unique $b\in F(R)$ mapping to $(\phi(a_i))_{i\in I}$. Set $\phi(a)=b$. We have to show that this definition does not depend on the choice of the $R$-covering family $(R_i)_{i\in I}$. So let $(S_j)_{j\in J}$ be another $R$-covering family such that the image of $a\in E(R)$ lies in $D(S_j)\subset E(S_j)$ for every $j\in J$. Then the $R$-covering family $(R_i\otimes_ R S_j)_{(i,j)\in I\times J}$ also has the property that the image of $a$ in $E(R_i\otimes_ R S_j)$ lies in $D(R_i\otimes_R S_j)$ for $(i,j)\in I\times J$. Performing the above construction of $\phi(a)$ with the $R$-covering family $(R_i\otimes_R S_j)_{(i,j)\in I\times J}$ and considering the commutativity of
$$
\xymatrix{
& \prod_i R_i \ar[d]\\
R \ar[ru] \ar[r] \ar[rd] & \prod_{i,j} R_i\otimes_R S_j \\
& \prod_j S_j \ar[u]
}
$$
we see that $\phi(a)$ is well--defined. To show that the definition of $\phi$ is functorial in $R$ we can consider a morphism $R\to S$ of $k$-$\sigma$-algebras and the $S$-covering family $(R_i\otimes_R S)_{i\in I}$. The fact that $\phi\colon E\to F$ is the unique extension of $\phi\colon D\to F$ is immediate from the construction of $\phi$.
The implication (ii)$\Rightarrow$(iii) is obvious. Thus it only remains to show that (iii)$\Rightarrow$(i).
Let $(R_i)_{i\in I}$ be an $R$-covering family. Let $D$ be the subfunctor of $\operatorname{Sp}^\sigma(R)$ equal to the union $\cup_{i\in I}\operatorname{Im}(\operatorname{Sp}^\sigma(R_i)\to\operatorname{Sp}^\sigma(R))$, i.e., a morphism of $k$-$\sigma$-algebras $R\to S$ lies in $D(S)$ if and only if it factors through $R\to R_i$ for some $i\in I$. Then $D$ is a fat subfunctor of $\operatorname{Sp}^\sigma(R)$ by Lemma \ref{lemma: fat subfunctor of sps}. We claim that the sequence
\begin{equation} \label{eqn: sheaf exactness for Hom} \xymatrix{ \operatorname{Hom}(D,F) \ar^-\alpha[r] & \prod_i\operatorname{Hom}(\operatorname{Sp}^\sigma(R_i),F) \ar@<0.5ex>[r]^-{\beta_1} \ar@<-0.5ex>[r]_-{\beta_2} & \prod_{i,j}\operatorname{Hom}(\operatorname{Sp}^\sigma(R_i\otimes_R R_j),F)}
\end{equation}
is exact. The injectivity of $\alpha$ is clear from the definition of $D$. Let $f=(f_i)_{i\in I}\in \prod_i\operatorname{Hom}(\operatorname{Sp}^\sigma(R_i),F)$ be such that $\beta_1(f)=\beta_2(f)$. Then for every pair $(i,j)\in I\times I$
\begin{equation} \label{eqn: sheaf commutative} \xymatrix{ & \operatorname{Sp}^\sigma(R_i)\times_{\operatorname{Sp}^\sigma(R)}\operatorname{Sp}^\sigma(R_j) \ar^-\simeq[d]& \\
& \operatorname{Sp}^\sigma(R_i\otimes_R R_j) \ar[ld] \ar[rd]& \\
\operatorname{Sp}^\sigma(R_i) \ar_-{f_i}[rd] & & \operatorname{Sp}^\sigma(R_j) \ar^{f_j}[ld] \\
& F(R) & }
\end{equation}
commutes. Let us define a morphism $h\colon D\to F$ as follows: If $S$ is a $k$-$\sigma$-algebra and $a\in D(S)\subset\operatorname{Sp}^\sigma(R)(S)=\operatorname{Hom}(R,S)$ then $a$ factors as $a\colon R\to R_i\xrightarrow{b} S$ for some $i\in I$ and we can define $h(a)=f_i(b)\in F(S)$. If follows from the commutativity of (\ref{eqn: sheaf commutative}) that $h(a)$ does not depend on the choice of $i$ and $b$. Clearly, $\alpha(h)=f$. Therefore (\ref{eqn: sheaf exactness for Hom}) is exact.
By assumption, every morphism $D\to F$ uniquely extends to $\operatorname{Sp}^\sigma(R)\to F$, i.e., $\operatorname{Hom}(\operatorname{Sp}^\sigma(R),F)\to\operatorname{Hom}(D, F)$ is bijective. Thus, applying the canonical bijections $\operatorname{Hom}(\operatorname{Sp}^\sigma(S),F)\simeq F(S)$ to (\ref{eqn: sheaf exactness for Hom}), yields the exactness of (\ref{eqn: sheaf}). So $F$ is a sheaf.
\end{proof}
\begin{theo} \label{theo: associated sheaf}
Let $F$ be a functor from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$. Then there exists a sheaf $\widetilde{F}$ and a morphism $\iota\colon F\to\widetilde{F}$ that is universal among morphisms from $F$ to sheaves, i.e., for every morphism $\phi$ from $F$ to a sheaf $E$ there exists a unique morphism $\widetilde{\phi}\colon\widetilde{F}\to E$ making
\[\xymatrix{
F \ar[rr]^-\iota \ar[rd]_-\phi & & \widetilde{F} \ar@{..>}[ld]^-{\widetilde{\phi}} \\
& E &
}
\]
commutative.
\end{theo}
\begin{proof}
For any $k$-$\sigma$-algebra $R$ we set $(\L F)(R)=\varinjlim\operatorname{Hom}(D,F)$, where the limit is taken over all fat subfunctors $D$ of $\operatorname{Sp}^\sigma(R)$.
An element of $(\L F)(R)$ is thus an equivalence class $(D,\alpha)$, where $D$ is a fat subfunctor of $\operatorname{Sp}^\sigma(R)$ and $\alpha\colon D\to F$ a morphism of functors. We have $(D,\alpha)=(D',\alpha')$ if and only if there exists a fat subfunctor $D''$ of $\operatorname{Sp}^\sigma(R)$ contained in $D$ and $D'$ such that the restrictions of $\alpha$ and $\alpha'$ to $D''$ are equal.
Let $\psi\colon R\to S$ be a morphism of $k$-$\sigma$-algebras. For every fat subfunctor $D$ of $\operatorname{Sp}^\sigma(R)$ the map $\operatorname{Sp}^\sigma(\psi)\colon\operatorname{Sp}^\sigma(\psi)^{-1}(D)\to D$ induces a map $\operatorname{Hom}(D,F)\to\operatorname{Hom}(\operatorname{Sp}^\sigma(\psi)^{-1}(D),F)$. By passing to the limit we obtain a map $\L(\psi)\colon(\L F)(R)\to (\L F)(S)$. Thus $\L F$ is a functor from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$. We have a morphism of functors
$\iota_F\colon F\to \L F$ determined by sending, for each $k$-$\sigma$-algebra $R$, an element of $F(R)\simeq \operatorname{Hom}(\operatorname{Sp}^\sigma(R),F)$ to its canonical image in $\varinjlim\operatorname{Hom}(D,F)$.
Now set $\widetilde{F}=\L(\L F)$ and $\iota=\iota_{\L F}\circ \iota_F$. With the notation of the theorem, we will next show the existence and uniqueness of $\widetilde{\phi}$. Since $\iota=\iota_{\L F}\circ \iota_F$, it suffices to show that for every morphism
$\phi$ from $F$ to a sheaf $E$, there exists a unique morphism $\phi'\colon\L F\to E$ making
\begin{equation} \label{eqn: diag LF}
\xymatrix{
F \ar[rr]^-{\iota_F} \ar[rd]_-\phi & & \L F \ar@{..>}[ld]^-{\phi'} \\
& E &
}
\end{equation}
commutative.
Let us first prove the uniqueness of $\phi'$. Let $R$ be a $k$-$\sigma$-algebra and let $\mu\colon \operatorname{Sp}^\sigma(R)\to \L F$ correspond to $(D,\alpha)\in (\L F)(R)$ under the bijection $(\L F)(R)\simeq\operatorname{Hom}(\operatorname{Sp}^\sigma(R),\L F)$.
It follows from the definitions that
\begin{equation} \label{eqn: diag difficult comm}
\xymatrix{
D \ar@{^(->}[r] \ar_\alpha[d] & \operatorname{Sp}^\sigma(R) \ar^\mu[d] \\
F \ar^{\iota_F}[r] & \L F
}
\end{equation}
commutes. Let $\eta\colon \operatorname{Sp}^\sigma(R)\to E$ correspond to $\phi'((D,\alpha))\in E(R)$ under the bijection $E(R)\simeq\operatorname{Hom}(\operatorname{Sp}^\sigma(R),E)$.
Then $\eta=\phi'\mu$. Combining this with (\ref{eqn: diag difficult comm}) and (\ref{eqn: diag LF}), we see that the restriction of $\eta$ to $D$ equals $\phi\alpha$. By Proposition \ref{prop: characterize sheaf}, this shows that $\eta$ is uniquely determined by $(D,\alpha)$ and $\phi$. Therefore $\phi'((D,\alpha))$ is uniquely determined by $\phi$.
Let us now establish the existence of $\phi'$. Let $R$ be a $k$-$\sigma$-algebra and $(D,\alpha)\in(\L F)(R)$. By Proposition \ref{prop: characterize sheaf}, the composition $D\xrightarrow{\alpha} F\xrightarrow{\phi} E$ has a unique extension $\eta\colon \operatorname{Sp}^\sigma(R)\to E$. Note that $\eta$ only depends on the equivalence class of $D$ and $\alpha$. Now define $\phi'((D,\alpha))\in E(R)$ as the image of $\eta$ under the canonical bijection $\operatorname{Hom}(\operatorname{Sp}^\sigma(R),E)\simeq E(R)$.
This defines a morphism $\phi'\colon \L F\to E$ such that $\phi=\phi'\iota_F$.
\medskip
It remains to show that $\widetilde{F}$ is a sheaf. But let us first show that $\iota_{\L F}\colon (\L F)(R)\to (\L(\L F))(R)$ is injective for every $k$-$\sigma$-algebra $R$. So we have to show that if $\mu,\eta\colon \operatorname{Sp}^\sigma(R)\to \L F$ are two morphisms having the same restriction to a fat subfunctor $D\subset\operatorname{Sp}^\sigma(R)$, then $\mu=\eta$.
Let $\hat{\mu},\hat{\eta}\in(\L F)(R)$ correspond to $\mu$ and $\eta$ under the bijection $(\L F)(R)\simeq\operatorname{Hom}(\operatorname{Sp}^\sigma(R),\L F)$. Replacing $D$ by a smaller fat subfunctor of $\operatorname{Sp}^\sigma(R)$ if necessary, we can assume that $\hat{\mu}=(D,\mu')$ and $\hat{\eta}=(D,\eta')$. As in (\ref{eqn: diag difficult comm}), we have $\iota_F\mu'=\mu|_D$ and $\iota_F\eta'=\eta|_D$. By Lemma \ref{lemma: fat subfunctor of sps}, there exists an $R$-covering family $(R_i)_{i\in I}$ such that $\operatorname{Im}(\operatorname{Sp}^\sigma(R_i)\to\operatorname{Sp}^\sigma(R))\subset D$ for all $i\in I$.
For $i\in I$ let $\mu_i\colon \operatorname{Sp}^\sigma(R_i)\to D\xrightarrow{\mu'}F$ and $\eta_i\colon\operatorname{Sp}^\sigma(R_i)\to D\xrightarrow{\eta'} F$ be the induced morphisms. Then $\iota_F\mu_i=\iota_F\eta_i$. Considering the image of $\operatorname{id}_{R_i}\in\operatorname{Sp}^\sigma(R_i)(R_i)$ in $(\L F)(R_i)$ under $\iota_F\mu_i=\iota_F\eta_i$, shows that there exists a fat subfunctor $D_i$ of $\operatorname{Sp}^\sigma(R_i)$ such that $\mu_i|_{D_i}=\eta_i|_{D_i}$. By Lemma \ref{lemma: fat subfunctor of sps}, there exists for every $i\in I$ and $R_i$-covering family $(S_{ij})_{j\in J_i}$ such that $\operatorname{Im}(\operatorname{Sp}^\sigma(S_{ij})\to\operatorname{Sp}^\sigma(R_i))\subset D_i$ for all $j\in J_i$. The family $(S_{ij})_{i\in I, j\in J_i}$ is $R$-covering and therefore the union $D'$ of all $\operatorname{Im}(\operatorname{Sp}^\sigma(S_{ij})\to R)$ is a fat subfunctor of $\operatorname{Sp}^\sigma(R)$ by Lemma \ref{lemma: fat subfunctor of sps}.
$$
\xymatrix{
\operatorname{Sp}^\sigma(S_{ij}) \ar[r] & D_i \ar@{^(->}[r] &\operatorname{Sp}^\sigma(R_i) \ar@<0.5ex>[rd]^-{\eta_i} \ar@<-0.5ex>[rd]_-{\mu_i} \ar[r] & D \ar@<0.5ex>[d]^-{\eta'} \ar@<-0.5ex>[d]_-{\mu'} \ar@{^(->}[r] & \operatorname{Sp}^\sigma(R) \ar@<0.5ex>[d]^-{\eta} \ar@<-0.5ex>[d]_-{\mu} \\
& & & F \ar^-{\iota_F}[r] & \L F
}
$$
Since $\mu_i$ and $\eta_i$ agree on $D_i$ we see that $\mu'$ and $\eta'$ agree on $D'$. Therefore $\hat{\mu}=\hat{\eta}$ and consequently $\mu=\eta$ as claimed.
To show that $\widetilde{F}$ is a sheaf, it thus suffices to show that $\L F$ is a sheaf whenever $\iota_F\colon F(R)\to(\L F)(R)$ is injective for every $k$-$\sigma$-algebra $R$. Let us identify $F$ with a subfunctor of $\L F$ via $\iota_F$ and let $(D,\alpha)\in (\L F)(R)$.
Since $D$ is fat in $\operatorname{Sp}^\sigma(R)$, by Lemma \ref{lemma: fat subfunctor of sps}, there exists an $R$-covering family $(R_i)_{i\in I}$ such that $\operatorname{Im}(\operatorname{Sp}^\sigma(R_i)\to\operatorname{Sp}^\sigma(R))\subset D$. Let $\phi_i\colon\operatorname{Sp}^\sigma(R_i)\to \operatorname{Sp}^\sigma(R)$ denote the canonical map. The commutativity of
$$
\xymatrix{
{\phi_i}^{-1}(D) \ar[d] \ar@{^(->}[r] & \operatorname{Sp}^\sigma(R_i) \ar[ld] \ar^-{\phi_i}[d] \\
D \ar_\alpha[d] \ar@{^(->}[r] & \operatorname{Sp}^\sigma(R_i) \\
F &
}
$$
shows that $(\L F)(R)\to (\L F)(R_i)$ maps $(D,\alpha)$ into $F(R_i)\subset(\L F)(R_i)$. Thus $F$ is fat in $\L F$.
To show that $\L F$ is a sheaf, it suffices to show (Proposition \ref{prop: characterize sheaf}) that for every $k$-$\sigma$-algebra $R$, every morphism $\mu\colon D\to \L F$ from a fat subfunctor $D\subset\operatorname{Sp}^\sigma(R)$ to $\L F$ has a unique extension $\mu'\colon\operatorname{Sp}^\sigma(R)\to \L F$.
The uniqueness of $\mu'$ follows from the assumed injectivity of $\iota_F\colon F(R)\to (\L F)(R)$. So it remains to prove the existence of $\mu'$. Since $F\subset \L F$ is fat, $\mu^{-1}(F)\subset D$ is fat. Since $D$ is fat in $\operatorname{Sp}^\sigma(R)$ it follows that $\mu^{-1}(F)$ is fat in $\operatorname{Sp}^\sigma(R)$. Let $\mu'\colon \operatorname{Sp}^\sigma(R)\to \L F$ correspond to the image of $\mu\colon \mu^{-1}(F)\to F$ in $(\L F)(R)$ under the bijection $(\L F)(R)\simeq\operatorname{Hom}(\operatorname{Sp}^\sigma(R),\L F)$. Then $\mu'$ is an extension of $\mu\colon \mu^{-1}(F)\to F\subset \L F$. It remains to show that $\mu'$ also extends $\mu\colon D\to \L F$. So we have to show that $\mu(a)=\mu'(a)$ for any $k$-$\sigma$-algebra $S$ and $a\in D(S)$.
Let $\hat{a}\colon\operatorname{Sp}^\sigma(S)\to D$ correspond to $a$ under the bijection $D(S)\simeq\operatorname{Hom}(\operatorname{Sp}^\sigma(S),D)$. Since $\mu\hat{a}$ and $\mu'\hat{a}$ have the same restriction to $\hat{a}^{-1}(\mu^{-1}(F))\subset\operatorname{Sp}^\sigma(S)$ it follows from the assumed injectivity of $\iota_F\colon F(S)\to (\L F)(S)$ that
$\mu\hat{a}=\mu'\hat{a}$. As $\mu(a)$ and $\mu'(a)\in(\L F)(S)$ corresponds to $\operatorname{Sp}^\sigma(S)\xrightarrow{\hat{a}}D\xrightarrow{\mu}\L F$ and $\operatorname{Sp}^\sigma(S)\xrightarrow{\hat{a}}D\xrightarrow{\mu'}\L F$ respectively under the bijection $(\L F)(S)\simeq\operatorname{Hom}(\operatorname{Sp}^\sigma(S),\L F)$ we find that $\mu(a)=\mu'(a)$ as desired.
\end{proof}
The sheaf
$$\widetilde{F}$$ from Theorem \ref{theo: associated sheaf} is called the \emph{sheaf associated with $F$}, or the \emph{sheafification of $F$}.
If $\phi\colon F\to E$ is a morphism of functors from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$, then $F\xrightarrow{\phi} E\to\widetilde{E}$ induces a morphism $\widetilde{\phi}\colon \widetilde{F}\to\widetilde{E}$. So sheafification is a functor. Indeed, as
$$\operatorname{Hom}(\widetilde{F},E)\simeq\operatorname{Hom}(F,E),$$
we see that sheafification is left adjoint to the inclusion of functors into sheaves.
\begin{cor}
Let $D$ be a fat subfunctor of a sheaf $F\colon \text{$k$-$\sigma$-$\mathsf{Alg}$}\to \mathsf{Sets}$, then $\widetilde{D}=F$.
\end{cor}
\begin{proof}
This is clear from Proposition \ref{prop: characterize sheaf}.
\end{proof}
\begin{lemma} \label{lemma: sheaf injective}
Let $\phi\colon F\to E$ be a morphism of functors from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$ such that $\phi_R\colon F(R)\to E(R)$ is injective for any $k$-$\sigma$-algebra $R$. Then $\widetilde{\phi}_R\colon \widetilde{F}(R)\to\widetilde{E}(R)$ is injective for any $k$-$\sigma$-algebra $R$.
\end{lemma}
\begin{proof}
It suffices to show that $(\L F)(R)\to(\L E)(R)$ is injective for any $k$-$\sigma$-algebra $R$. So let $\alpha_1\colon D_1\to F$ and $\alpha_2\colon D_2\to F$ be morphism with $D_1,D_2$ fat subfunctors of $\operatorname{Sp}^\sigma(R)$ such that $\alpha_1\phi$ and $\alpha_2\phi$ have the same image in
$(\L E)(R)=\varinjlim\operatorname{Hom}(D,E)$. Then there exists a fat subfunctor $D$ of $\operatorname{Sp}^\sigma(R)$ with $D\subset D_1\cap D_2$ such that the restriction of
$\alpha_1\phi$ and $\alpha_2\phi$ to $D$ are equal. It follows from the injectivity of $\phi$ that $\alpha_1$ and $\alpha_2$ have the same restriction to $D$.
\end{proof}
For simplicity, let us call a functor $G\colon\text{$k$-$\sigma$-$\mathsf{Alg}$}\to\mathsf{Groups}$ a \emph{group functor}. Equivalently, a group functor is a group object in the category of functors from $k$-$\sigma$-$\mathsf{Alg}$ to $\mathsf{Sets}$.
If $G$ is a group functor, then the associated sheaf $\widetilde{G}$ is naturally a group functor such that the associated morphism of functors $\iota\colon G\to\widetilde{G}$ is a morphism of group functors, i.e., is compatible with the group structure. Indeed, if $G$ is a group functor, then $\operatorname{Hom}(D,G)$ has naturally the structure of a group for any functor $D\colon\text{$k$-$\sigma$-$\mathsf{Alg}$}\to\mathsf{Sets}$. If $R$ is a $k$-$\sigma$-algebra and $D_1\subset D_2$ are fat subfunctors of $\operatorname{Sp}^\sigma(R)$ then $\operatorname{Hom}(D_2,G)\to\operatorname{Hom}(D_1,G)$ is a morphism of groups and we have an induced group structure on the limit $(\L G)(R)=\varinjlim\operatorname{Hom}(D,G)$. This group structure is functorial in $R$ and $\iota_G\colon G\to \L G$ is a morphism of group functors. Applying this construction twice, yields the required group structure on $\widetilde{G}=\L(\L G)$. Alternatively, the group structure on $\widetilde{G}$, can be obtained by applying $\widetilde{()}$ to the multiplication $G\times G\to G$ and using the canonical isomorphism $\widetilde{G\times G}\simeq \widetilde{G}\times\widetilde{G}$.
If $G$ is a group functor and $N$ a normal subgroup functor, i.e., $N(R)$ is a normal subgroup of $G(R)$ for any $k$-$\sigma$-algebra $R$, we can define a group functor
$$G/\!\!/N$$
by $R\rightsquigarrow G(R)/N(R)$. The crux with this functor is that it may not be a sheaf (or $\sigma$-variety) if $G$ and $N$ are sheaves (or $\sigma$-varieties). We therefore need to sheafify $G/\!\!/N$. Sheafification is compatible with quotients in the following sense:
\begin{lemma} \label{lemma: sheafs and quotients}
Let $G$ be a group functor and $N$ a normal subgroup functor. Then $\widetilde{N}$ is a normal subgroup functor of $\widetilde{G}$ and
$$\widetilde{\widetilde{G}/\!\!/\widetilde{N}}\simeq \widetilde{G/\!\!/N}.$$
\end{lemma}
\begin{proof}
By Lemma \ref{lemma: sheaf injective} the map $\widetilde{N}(R)\to\widetilde{G}(R)$ is injective for any $k$-$\sigma$-algebra $R$. Since $N$ is normal in $G$, $\operatorname{Hom}(D,N)$ is normal in $\operatorname{Hom}(D,G)$ for every fat subfunctor $D$ of $\operatorname{Sp}^\sigma(R)$. It follows that $(\L N)(R)$ is normal in $(\L G)(R)$ and therefore $\widetilde{N}$ is normal in $\widetilde{G}$.
Since the morphism $N\to G\to G/\!\!/N$ factors through the sheaf $E$ that maps every $k$-$\sigma$-algebra to a set with one element. Also $\widetilde{N}\to\widetilde{G}\to\widetilde{G/\!\!/N}$ factors through $E$. So $\widetilde{N}$ maps into the kernel of $\widetilde{G}\to\widetilde{G/\!\!/N}$ and we obtain a morphism $\mu\colon\widetilde{G}/\!\!/\widetilde{N}\to \widetilde{G/\!\!/N}$. Since $N$ lies in the kernel of $G\to\widetilde{G}\to \widetilde{G}/\!\!/\widetilde{N}$ we obtain a morphism $\eta\colon G/\!\!/N\to \widetilde{G}/\!\!/\widetilde{N}$.
All the subdiagrams of
$$
\xymatrix{
G \ar_-{\pi_G}[r] \ar_-{i_G}[d] & G/\!\!/N \ar@/_85pt/^\eta[ldd] \ar^-{i_{G/\!\!/N}}[d] \\
\widetilde{G} \ar^-{\widetilde{\pi_G}}[r] \ar_-{\pi_{\widetilde{G}}}[d] & \widetilde{G/\!\!/N} \ar@{..>}_-\beta[d]<-0.5ex>\\
\widetilde{G}/\!\!/\widetilde{N} \ar^-\mu[ur] \ar^-{i_{\widetilde{G}/\!\!/\widetilde{N}}}[r] & \widetilde{\widetilde{G}/\!\!/\widetilde{N}} \ar@{..>}_-\alpha[u]<-0.5ex>\\
}
$$
consisting only of solid arrows commute. By the universal property of $i_{\widetilde{G}/\!\!/\widetilde{N}}$ there exists a morphism $\alpha\colon \widetilde{\widetilde{G}/\!\!/\widetilde{N}}\to \widetilde{G/\!\!/N}$ such that $\alpha i_{\widetilde{G}/\!\!/\widetilde{N}}=\mu$.
By the universal property of $i_{G/\!\!/N}$ there exists a morphism $\beta \colon \widetilde{G/\!\!/N}\to \widetilde{\widetilde{G}/\!\!/\widetilde{N}}$ such that $\beta i_{G/\!\!/N}= i_{\widetilde{G}/\!\!/\widetilde{N}}\eta$. We will show that $\alpha$ and $\beta$ are inverse to each other. We have
$$\beta\widetilde{\pi_G}i_G=\beta i_{G/\!\!/N}\pi_G=i_{\widetilde{G}/\!\!/\widetilde{N}}\eta\pi_G=i_{\widetilde{G}/\!\!/\widetilde{N}}\pi_{\widetilde{G}}i_G.$$
Therefore $\beta\widetilde{\pi_G}=i_{\widetilde{G}/\!\!/\widetilde{N}}\pi_{\widetilde{G}}$. Since, for any $k$-$\sigma$-algebra $R$ and element from
$(\widetilde{G}/\!\!/\widetilde{N})(R)$ lifts to an element of $\widetilde{G}(R)$, this implies that $\beta\mu=i_{\widetilde{G}/\!\!/\widetilde{N}}$.
From $\alpha i_{\widetilde{G}/\!\!/\widetilde{N}}=\mu$ and $\beta\mu=i_{\widetilde{G}/\!\!/\widetilde{N}}$ we deduce that $\beta\alpha i_{\widetilde{G}/\!\!/\widetilde{N}}=i_{\widetilde{G}/\!\!/\widetilde{N}}$. Thus $\beta\alpha$ is the identity on $\widetilde{\widetilde{G}/\!\!/\widetilde{N}}$.
We have $\alpha i_{\widetilde{G}/\!\!/\widetilde{N}}\eta=\mu\eta=i_{G/\!\!/N}$. Together with $i_{\widetilde{G}/\!\!/\widetilde{N}}\eta=\beta i_{G/\!\!/N}$, this implies $\alpha \beta i_{\widetilde{G}/\!\!/\widetilde{N}}=i_{\widetilde{G}/\!\!/\widetilde{N}}$. Therefore $\alpha\beta$ is the identity on $\widetilde{G/\!\!/N}$.
\end{proof}
\begin{theo} \label{theo: quotient sheaf}
Let $G$ be a $\sigma$-algebraic group and $N\unlhd G$ a $\sigma$-closed subgroup. Then $G/\!\!/N$ is a fat subfunctor of $G/N$. In particular, $$\widetilde{G/\!\!/N}=G/N.$$
\end{theo}
\begin{proof}
By Theorem \ref{theo: existence of quotient} the kernel of $G\to G/N$ equals $N$, i.e, for every $k$-$\sigma$-algebra $R$ the kernel of $G(R)\to (G/N)(R)$ is $N(R)$. Thus the canonical map
$G(R)/N(R)\to (G/N)(R)$ is injective. Therefore we can identity $G/\!\!/N$ with a subfunctor of $G/N$. It follows from Lemma \ref{lemma: sheaf surjectivity of quotient} that $G/\!\!/N$ is fat in $G/N$.
\end{proof}
\subsection{The isomorphism theorems}
The following theorem is the analog of the first isomorphism theorem for groups.
\begin{theo} \label{theo: isom1}
Let $\phi\colon G\to H$ be a morphism of $\sigma$-algebraic groups. Then $\phi(G)$ is a $\sigma$-closed subgroup of $H$ and the induced morphism $G/\ker(\phi)\to\phi(G)$ is an isomorphism.
\end{theo}
\begin{proof}
We already observed in Lemma \ref{lemma: f(G) subgroup} that $\phi(G)$ is a $\sigma$-closed subgroup. Since $\ker(\phi)$ is the kernel of $G\to\phi(G)$, the induced morphism $G/\ker(\phi)\to\phi(G)$ is a $\sigma$-closed embedding by Corollary \ref{cor: quotient embedding} and so we can identify $G/\ker(\phi)$ with a $\sigma$-closed $\sigma$-subvariety of $G$. Since $\phi$ factors through $G/\ker(\phi)$ it follows from the definition of $\phi(G)$ that $G/\ker(\phi)=\phi(G)$.
\end{proof}
\begin{cor} \label{cor: image sheaf}
Let $\phi\colon G\to H$ be a morphism of $\sigma$-algebraic groups. Then
$$\phi(G)=\widetilde{\operatorname{Im}(\phi)}.$$
\end{cor}
\begin{proof}
The isomorphism $G/\ker(\phi)\simeq\phi(G)$ identifies $G/\!\!/\ker(\phi)$ with $\operatorname{Im}(\phi)$. Therefore the claim follows from Theorem \ref{theo: quotient sheaf}.
\end{proof}
Let $N$ and $H$ be $\sigma$-closed subgroups of a $\sigma$-algebraic group $G$ such that $H$ normalizes $N$, i.e., $H(R)$ normalizes $N(R)$ for any $k$-$\sigma$-algebra $R$. Then we can form the semidirect product $N\rtimes H$: This is a $\sigma$-algebraic group with underlying $\sigma$-variety $N\times H$ and multiplication given by
$((n_1,h_1),(n_2,h_2))\mapsto(n_1 h_1n_2h_1^{-1}, h_1 h_2)$
for any $k$-$\sigma$-algebra $R$ and $n_1,n_2\in N(R)$, $h_1,h_2\in H(R)$.
The map
$$m\colon N\rtimes H\to G,\ (n,h)\mapsto nh$$
for any $k$-$\sigma$-algebra $R$ and $n\in N(R)$, $h\in H(R)$ is a morphism of $\sigma$-algebraic groups.
We define
$$HN:=NH:=m(N\rtimes H).$$
By construction $HN$ is a $\sigma$-closed subgroup of $G$ (Lemma \ref{lemma: f(G) subgroup}). In fact, $HN$ is the smallest $\sigma$-closed subgroup of $G$ which contains $N$ and $H$.
It follows from Corollary \ref{cor: image sheaf} that $HN$ is the sheaf associated to the functor $R\rightsquigarrow N(R)H(R)$. Moreover, by Lemma \ref{lemma: f(G) faithfully flat} we have
$$(NH)(R)=\{g\in G(R)|\ \exists \text{ faithfully flat $R$-$\sigma$-algebra $S$ such that } g\in N(S)H(S)\}$$
for any $k$-$\sigma$-algebra $R$. It follows from Lemma \ref{lemma: normal stays normal under surjective map} that $N=m(N)$ is normal in $HN$.
The following theorem is the analog of the second isomorphism theorem for groups.
\begin{theo} \label{theo: isom2}
Let $H$ and $N$ be $\sigma$-closed subgroups of a $\sigma$-algebraic group $G$ such that $H$ normalizes $N$. Then the canonical morphism
$$H/(H\cap N)\to HN/N$$
is an isomorphism.
\end{theo}
\begin{proof}
For every $k$-$\sigma$-algebra $R$ we have an isomorphism
$$H(R)/(H(R)\cap N(R))\to H(R)N(R)/N(R)$$ which is functorial in $R$. Passing to the associates sheaves and using Lemma \ref{lemma: sheafs and quotients} we find the required isomorphism.
\end{proof}
The following theorem is the analog of the third isomorphism theorem for groups.
\begin{theo} \label{theo: isom3}
Let $G$ be a $\sigma$-algebraic group, $N\unlhd G$ a normal $\sigma$-closed subgroup and $\pi\colon G\to G/N$ the quotient. The map $H\mapsto \pi(H)=H/N$ defines a bijection between the $\sigma$-closed subgroups $H$ of $G$ containing $N$ and the $\sigma$-closed subgroups $H'$ of $G/N$. The inverse map is $H'\mapsto\pi^{-1}(H')$. A $\sigma$-closed subgroup $H$ of $G$ containing $N$ is normal in $G$ if and only if $H/N$ is normal in $G/N$, in which case the canonical morphism
$$G/H\to (G/N)/(H/N)$$
is an isomorphism.
\end{theo}
\begin{proof}
By Theorem \ref{theo: isom1} and Theorem \ref{theo: existence of quotient} we have $\pi(H)=H/N$.
Let us show that $\pi^{-1}(\pi(H))=H$ for ever $\sigma$-closed subgroup $H$ of $G$ containing $N$.
Let $R$ be a $k$-$\sigma$-algebra and $g\in \pi^{-1}(\pi(H))(R)$, i.e., $\pi(g)\in\pi(H)(R)$. By Lemma \ref{lemma: f(G) faithfully flat} there exists a faithfully flat $R$-$\sigma$-algebra $S$ and $h\in H(S)$ with $\pi(h)=\pi(g)\in (G/N)(S)$. As $\ker(\pi)=N$ by Theorem \ref{theo: existence of quotient}, this implies that $gh^{-1}\in N(S)\leq H(S)$. Therefore $g\in H(S)$ and $g\in H(S)\cap G(R)=H(R)$ by Lemma \ref{lemma: faithfullyflat intersection}. Thus $\pi^{-1}(\pi(H))\subset H$. The reverse inclusion is obvious.
Let us next show that $\pi(\pi^{-1}(H'))=H'$ for a $\sigma$-closed subgroup $H'$ of $G/N$. As $\pi$ maps $\pi^{-1}(H')$ into $H'$, it is clear form the definition of $\pi(\pi^{-1}(H'))$ (Lemma \ref{lemma: f(X)}) that $\pi(\pi^{-1}(H'))\subset H'$.
Let $R$ be a $k$-$\sigma$-algebra and $h'\in H'(R)$. There exists a faithfully flat $R$-$\sigma$-algebra $S$ and $g\in G(S)$ such that $\pi(g)=h'$.
So $g\in\pi^{-1}(H')(S)$ and $h'=\pi(g)\in\pi(\pi^{-1}(H')(S))\subset\pi(\pi^{-1}(H'))(S)$. Thus
$h'\in \pi(\pi^{-1}(H'))(S)\cap H'(R)=\pi(\pi^{-1}(H'))(R)$. Hence $\pi(\pi^{-1}(H'))=H'$.
If $H$ is normal in $G$, then $\pi(H)$ is normal in $G/N$ by Lemma \ref{lemma: normal stays normal under surjective map}. Clearly $\pi^{-1}(H')$ is normal if $H'$ is normal.
For every $k$-$\sigma$-algebra $R$ be have an isomorphism
$$G(R)/H(R)\to (G(R)/N(R))/(H(R)/N(R)).$$ Passing to the associated sheaves and using Lemma \ref{lemma: sheafs and quotients} we obtain the required isomorphism.
\end{proof}
\section{Jordan--H\"{o}lder theorem} \label{sec: JordanHoelder}
In this section we apply the results from the previous sections to prove a Jordan--H\"{o}lder type theorem for $\sigma$-algebraic groups. A Jordan--H\"{o}lder type theorem for algebraic groups can be found in \cite{Rosenlicht:SomeBasicTheoremsOnAlgebraicGroups}, while a Jordan--H\"{o}lder type theorem for differential algebraic groups has been proved in \cite{CassidySinger:AJordanHoelderTheoremForDifferentialAlgebraicGroups}.
As we will show, the Schreier refinement theorem also holds for $\sigma$-algebraic groups (Theorem \ref{theo: schreier refinement}). So any two decomposition series of a $\sigma$-algebraic group have equivalent refinements. However, a $\sigma$-algebraic group rarely has a decomposition series. It is therefore useful to consider more general subnormal series. The basic idea is to consider $\sigma$-algebraic groups up to quotients by zero $\sigma$-dimensional normal subgroups. Formally this is realized by replacing in the uniqueness statement of the classical Jordan--H\"{o}lder theorem the notion of isomorphism by the notion of isogeny.
\medskip
Our first aim is to prove the analog of the Schreier refinement theorem. We follow along the lines of the well--known proof via the Butterly lemma. (Cf. \cite[Section 1.3]{Lang:Algebra} and \cite[Chapter IX, Section 6]{Milne:BasicTheoryOfAffineGroupSchemes}.) We will need two analogs of elementary statements about groups.
\begin{lemma} \label{lemma: Dedekind law}
Let $N$, $G$ and $H$ be $\sigma$-closed subgroups of a $\sigma$-algebraic group $G'$ such that $N\unlhd G$ and $N$ normalizes $H$. Then $G\cap NH=N(G\cap H)$.
\end{lemma}
\begin{proof}
As $N\leq G\cap NH$ and $G\cap H\leq G\cap NH$ it is clear that $N(G\cap H)\subset G\cap NH$.
Conversely, let $R$ be a $k$-$\sigma$-algebra and
$g\in (G\cap NH)(R)$. There exists a faithfully flat $R$-$\sigma$-algebra $S$ and $n\in N(S)$, $h\in H(S)$ such that $g=nh$ in $G'(S)$. But then $h=n^{-1}g\in G(S)$ and therefore $g=nh\in N(S)(G(S)\cap H(S))\subset (N(G\cap H))(S)$. It follows from Lemma \ref{lemma: faithfullyflat intersection} that $g\in (N(G\cap H))(R)$.
\end{proof}
\begin{lemma} \label{lemma: normal normalizes}
Let $H_1\unlhd H_2$ be $\sigma$-closed subgroups of a $\sigma$-algebraic group $G$. Assume that $H_2$ normalizes $N\leq G$. Then $NH_1\unlhd NH_2$.
\end{lemma}
\begin{proof}
Clearly $N\rtimes H_1$ is a normal $\sigma$-closed subgroup of $N\rtimes H_2$. Therefore $NH_1=m(N\rtimes H_1)$ is a normal $\sigma$-closed subgroup of $NH_2=m(N\rtimes H_2)$ by Lemma \ref{lemma: normal stays normal under surjective map}.
\end{proof}
The following lemma is the analog of the Butterfly (or Zassenhaus) lemma.
\begin{lemma} \label{lemma: butterfly}
Let $N_1\unlhd H_1$ and $N_2\unlhd H_2$ be $\sigma$-closed subgroups of a $\sigma$-algebraic group $G$. Then $N_1(H_1\cap N_2)\unlhd N_1(H_1\cap H_2)$,
$N_2(N_1\cap H_2)\unlhd N_2(H_1\cap H_2)$ and
$$\frac{N_1(H_1\cap H_2)}{N_1(H_1\cap N_2)}\simeq \frac{N_2(H_1\cap H_2)}{N_2(N_1\cap H_2)}.$$
\end{lemma}
\begin{proof}
Since $H_1\cap N_2$ is normal in $H_1\cap H_2$ it follows from Lemma \ref{lemma: normal normalizes} that $N_1(H_1\cap N_2)$ is normal in $N_1(H_1\cap H_2)$. Similarly, $N_2(N_1\cap H_2)\unlhd N_2(H_1\cap H_2)$. As $H_1\cap H_2$ normalizes $N_1(H_1\cap N_2)$ it follows from Theorem \ref{theo: isom2} that
\begin{equation} \label{eqn: butterfly}
\frac{H_1\cap H_2}{(H_1\cap H_2)\cap N_1(H_1\cap N_2)}\simeq\frac{(H_1\cap H_2)N_1(H_1\cap N_2)}{N_1(H_1\cap N_2)}.
\end{equation}
Lemma \ref{lemma: Dedekind law} with $N=H_1\cap N_2$, $G=H_1\cap H_2$ and $H=N_1$ shows that
$$(H_1\cap H_2)\cap N_1(H_1\cap N_2)=(H_1\cap N_2)(H_1\cap H_2\cap N_1)= (H_1\cap N_2)(N_1\cap H_2).$$
Because $H_1\cap N_2\subset H_1\cap H_2$ we find $(H_1\cap H_2)N_1(H_1\cap N_2)=N_1(H_1\cap H_2)$. Therefore (\ref{eqn: butterfly}) becomes
$$ \frac{H_1\cap H_2}{(H_1\cap N_2)(N_1\cap H_2)}\simeq\frac{N_1(H_1\cap H_2)}{N_1(H_1\cap N_2)}.$$ By symmetry
$$ \frac{H_1\cap H_2}{(H_1\cap N_2)(N_1\cap H_2)}\simeq\frac{N_2(H_1\cap H_2)}{N_2(N_1\cap H_2)}.$$
\end{proof}
\begin{defi}
Let $G$ be a $\sigma$-algebraic group. A \emph{subnormal series} of $G$ is a sequence
\begin{equation} \label{eqn: subnormal series} G=G_0\supsetneqq G_1\supsetneqq\cdots\supsetneqq G_n=1 \end{equation}
of $\sigma$-closed subgroups of $G$ such that $G_{i+1}$ is normal in $G_i$ for $i=0,\ldots,n-1$. Another subnormal series
\begin{equation} \label{eqn: subnormal series2} G=H_0\supsetneqq H_1\supsetneqq\cdots\supsetneqq H_m=1 \end{equation}
of $G$ is called a \emph{refinement} of (\ref{eqn: subnormal series}) if $\{G_0,\ldots,G_n\}\subset\{H_1,\ldots,H_m\}$. We say that (\ref{eqn: subnormal series}) and (\ref{eqn: subnormal series2}) are \emph{equivalent} if $n=m$ and there exists a permutation $\pi$ such that
the quotient groups $G_i/G_{i+1}$ and $H_{\pi(i)}/H_{\pi(i+1)}$ are isomorphic for $i=0,\ldots,n-1$. A subnormal series is called a \emph{decomposition series} if no quotient group has a proper non--trivial normal $\sigma$-closed subgroup.
\end{defi}
The following theorem is the analog of the Schreier refinement theorem.
\begin{theo} \label{theo: schreier refinement}
Any two subnormal series of a $\sigma$-algebraic group have equivalent refinements.
\end{theo}
\begin{proof}
Let $$G=G_0\supset G_1\supset\cdots\supset G_n=1$$ and
$$G=H_0\supset H_1\supset\cdots\supset H_m=1$$ be subnormal series of a $\sigma$-algebraic group $G$. Set $G_{i,j}=G_{i+1}(H_j\cap G_i)$ for $i=0,\ldots,n-1$ and $j=0,\ldots,m$. Then
$$G=G_0=G_{0,0}\supset G_{0,1}\supset G_{0,2}\supset\cdots\supset G_{0,m}=G_1=G_{1,0}\supset G_{1,1}\supset\cdots\supset G_{n-1,m}=1$$
is a subnormal series for $G$. Similarly, setting $H_{j,i}=H_{j+1}(G_i\cap H_j)$ for $j=0,\ldots,m-1$ and $i=0,\ldots,n$, defines a subnormal series for $G$.
By Lemma \ref{lemma: butterfly}
$$G_{i,j}/G_{i,j+1}\simeq H_{j,i}/H_{j,i+1}.$$
\end{proof}
\begin{cor}
Any two decomposition series of a $\sigma$-algebraic group are equivalent.
\end{cor}
\begin{proof}
By Theorem \ref{theo: schreier refinement} there exist equivalent refinements. But a refinement of a decomposition series is necessarily trivial.
\end{proof}
The above corollary is not that useful as it may seem at first sight, because a $\sigma$-algebraic need not poses a decomposition series. Let us illustrate this with an example.
\begin{ex} \label{ex: no decomposition series for Ga} Let $k$ be a $\sigma$-field of characteristic zero.
The $\sigma$-algebraic group $G=\mathbb{G}_a$ does not have a decomposition series. Indeed, by \cite[Corollary A.3]{diVizioHardouinWibmer:DifferenceAlgebraicRelations}, every proper $\sigma$-closed subgroup $G$ of $\mathbb{G}_a$ is of the form
$G(R)=\{g\in R|\ f(g)=0\}$ for some non--zero homogeneous linear difference equation $f=\sigma^n(y)+\lambda_{n-1}\sigma^{n-1}(y)+\cdots+\lambda_0y$.
If $h$ is another non-trivial linear homogeneous difference equation, then the product $h*f$ in the sense of linear difference operators (See \cite[Section 3.1]{Levin:difference}.) defines a $\sigma$-closed subgroup $H$ of $\mathbb{G}_a$ with $G\subsetneqq H\subsetneqq\mathbb{G}_a$. For example, for $h=\sigma(y)$ we have
$$H(R)=\{g\in R| \ \sigma^{n+1}(g)+\sigma(\lambda_{n-1})\sigma^{n}(y)+\cdots+\sigma(\lambda_0)\sigma(g)=0 \}.$$
\end{ex}
To remedy this shortcoming we need to relax the condition that the quotient groups of a decomposition series are simple.
\begin{defi}
A $\sigma$-algebraic group $G$ with $\sigma\text{-}\operatorname{dim}(G)>0$ is called \emph{almost--simple} if every normal proper $\sigma$-closed subgroup of $G$ has $\sigma$-dimension zero.
\end{defi}
In the differential world, the almost--simple differential algebraic groups are fairly well understood. See \cite{CassidySinger:AJordanHoelderTheoremForDifferentialAlgebraicGroups}, \cite{Minchenko:OnCentralExtensionsOfSimpleDifferentialAlgebraicGroups}, \cite{Freitag:IndecomposabilityForDifferentialAlgebraicGroups}. We hope to elucidate the structure of almost--simple $\sigma$-algebraic groups in the future. Considerable understanding of the Zariski dense $\sigma$-algebraic subgroups of almost--simple algebraic groups has already been obtained in \cite[Proposition 7.10]{ChatzidakiHrushovskiPeterzil:ModelTheoryofDifferenceFieldsIIPeriodicIdelas}. See also \cite[Section A4]{diVizioHardouinWibmer:DifferenceAlgebraicRelations}.
\begin{ex}
The $\sigma$-algebraic group $G=\mathbb{G}_a$ is almost--simple. (Cf. Example \ref{ex: no decomposition series for Ga}).
\end{ex}
\begin{lemma}
Let $G$ be a $\sigma$-algebraic group with $\sigma\text{-}\operatorname{dim}(G)>0$. Among the $\sigma$-closed subgroups $H$ of $G$ with $\sigma\text{-}\operatorname{dim}(H)=\sigma\text{-}\operatorname{dim}(G)$ there exists a unique smallest one.
\end{lemma}
\begin{proof}
Let $H_1$ and $H_2$ be $\sigma$-closed subgroups of $G$ with $\sigma\text{-}\operatorname{dim}(H_1)=\sigma\text{-}\operatorname{dim}(H_2)=\sigma\text{-}\operatorname{dim}(G)$. By Theorem \ref{theo: dimension theorem} we have $\sigma\text{-}\operatorname{dim}(H_1\cap H_2)=\sigma\text{-}\operatorname{dim}(G)$. Thus the claim follows from Corollary \ref{cor: finiteness1}.
\end{proof}
\begin{defi}
Let $G$ be a $\sigma$-algebraic group with $\sigma\text{-}\operatorname{dim}(G)>0$. The smallest $\sigma$-closed subgroup of $G$ with $\sigma$-dimension equal to the $\sigma$-dimension of $G$ is called the \emph{strong identity component} of $G$. It is denoted by
$$G^{so}.$$ A $\sigma$-algebraic group of positive $\sigma$-dimension is called \emph{strongly connected} if it is equal to its strong identity component.
\end{defi}
Thus a $\sigma$-algebraic group $G$ is strongly connected if and only if it has no proper $\sigma$-closed subgroup with $\sigma$-dimension equal to the $\sigma$-dimension of $G$. It is clear from Lemma \ref{cor: sdim of Go} that a strongly connected $\sigma$-algebraic group is connected. The strong identity component of a $\sigma$-algebraic group is strongly connected.
\begin{ex} \label{ex: strong component}
If $\mathcal{G}$ is a smooth connected algebraic group with $\dim(\mathcal{G})>0$, then $G=[\sigma]_k\mathcal{G}$ is strongly connected. Indeed, as $\mathcal{G}$ is smooth and connected $k[\mathcal{G}[i]]$ is an integral domain for every $i\geq 0$. So if $H$ is a proper $\sigma$-closed subgroup of $G$, then $\dim(H[i])<\dim(G[i])$ for $i\gg 0$. But $\dim(G[i])=\dim(\mathcal{G})(i+1)$ and so it follows from Theorem \ref{theo: existence sdim} that $\sigma\text{-}\operatorname{dim}(H)<\sigma\text{-}\operatorname{dim}(G)$.
\end{ex}
The following example shows that a $\sigma$-integral $\sigma$-algebraic group need not be strongly connected.
\begin{ex}
Let $G$ be the $\sigma$-closed subgroup of $\mathbb{G}_a^2$ given by
$$G(R)=\{(g_1, g_2)\in R^2|\ \sigma(g_1)=g_1\}\leq \mathbb{G}_a^2(R)$$ for any $k$-$\sigma$-algebra $R$. As $k\{G\}=k[y_1]\{y_2\}$ with $\sigma(y_1)=y_1$ we see that $G$ is $\sigma$-integral. We have $\sigma\text{-}\operatorname{dim}(G)=1$ (for example by Proposition \ref{prop: characterize sdim}). The $\sigma$-closed subgroup $H$ of $G$ given by
$H(R)=\{(0,g)\in R^2\}$ is isomorphic to $\mathbb{G}_a$ and therefore also has $\sigma$-dimension one. It is clear from Example \ref{ex: strong component} that $G^{so}=H$.
\end{ex}
It is obvious that $G^{so}$ is a characteristic subgroup of $G$ in the weak sense that for every automorphism $\tau$ of $G$ we have $\tau(G^{so})=G^{so}$. However, the following example illustrates the somewhat disturbing fact that $G^{so}$ need not be normal in $G$.
\begin{ex} \label{ex: Gso not normal}
Let $G=N\rtimes H$ be the $\sigma$-algebraic group from Example \ref{ex: Gsred not normal}. Then $\sigma\text{-}\operatorname{dim}(G)=1$.
The $\sigma$-closed subgroup $H=\mathbb{G}_m$ of $G$ has $\sigma$-dimension one. Since $H$ is strongly connected (Example \ref{ex: strong component}) we see that $H=G^{so}$.
We already noted in Example \ref{ex: Gsred not normal} that $H$ is not normal in $G$.
\end{ex}
\begin{lemma} \label{lemma: strongly connected is sintegral}
Assume that $k$ is perfect and inversive. Then a strongly connected $\sigma$-algebraic group is $\sigma$-integral.
\end{lemma}
\begin{proof}
Let $G$ be a strongly connected $\sigma$-algebraic group. Then $G$ is connected and because $\sigma\text{-}\operatorname{dim}(G)=\sigma\text{-}\operatorname{dim}(G_{\operatorname{red}})$ by Lemma \ref{lemma: sdim of sssubgroups}, we must have $G=G_{\operatorname{red}}$. So $G$ is reduced and hence integral. Similarly, $G=G_{\sigma\text{-}\operatorname{red}}$ by Lemma \ref{lemma: sdim of sssubgroups}. Thus $G$ is $\sigma$-integral.
\end{proof}
The following example shows that Lemma \ref{lemma: strongly connected is sintegral} fails over an arbitrary base $\sigma$-field. There exists a strongly connected $\sigma$-algebraic group which is not $\sigma$-reduced.
\begin{ex} \label{ex: strongly connected but not sreduced}
Let $k$ be a non--inversive $\sigma$-field of characteristic zero. So there exists $\lambda\in k$ with $\lambda\notin \sigma(k)$. Let $G$ be the $\sigma$-closed subgroup of $\mathbb{G}_a^2$ given by
$$G(R)=\left\{(g_1,g_2)\in R^2|\ \sigma(g_1)=\lambda\sigma(g_2)\right\}$$
for any $k$-$\sigma$-algebra $R$.
Then $k\{G\}=k[y_1,y_2,\sigma(y_2),\ldots]$ with $\sigma(y_1)=\lambda \sigma(y_2)$. For $i\geq 0$ let $G[i]$ denote the $i$-th order Zariski closure of $G$ in $\mathbb{G}_a^2$. Then $k[G[i]]=k[y_1,y_2,\ldots,\sigma^i(y_2)]$ and therefore $\dim(G[i])=1\cdot(i+1)+1$, in particular $\sigma\text{-}\operatorname{dim}(G)=1$.
We claim that $G$ is strongly connected. Suppose that $H\leq G$ is a proper $\sigma$-closed subgroup with $\sigma\text{-}\operatorname{dim}(H)=\sigma\text{-}\operatorname{dim}(G)$.
Let $a_1$ and $a_2$ denote the image of $y_1$ and $y_2$ in $k\{H\}$ respectively.
By \cite[Corollary A.3]{diVizioHardouinWibmer:DifferenceAlgebraicRelations} the $\sigma$-ideal $\mathbb{I}(H)\subset k\{\mathbb{G}_a^2\}$ is $\sigma$-generated by homogenous linear $\sigma$-polynomials. Thus there exists a non--trivial $k$-linear relation between $a_1, a_2, \sigma(a_2),\ldots$. If that relation would properly involve $\sigma^i(a_2)$ for $i\geq 1$, then $\sigma\text{-}\operatorname{dim}(H)=0$. Thus there exists a non--trivial $k$-linear relation between $a_1$ and $a_2$. We have $a_1\neq 0$ and $a_2\neq 0$ because otherwise $\sigma\text{-}\operatorname{dim}(H)=0$. So there exists $\mu\in k$ with $a_1-\mu a_2=0$. Consequently $$0=\sigma(a_1)-\sigma(\mu)\sigma(a_2)=\lambda\sigma(a_2)-\sigma(\mu)\sigma(a_2)=(\lambda-\sigma(\mu))\sigma(a_2).$$ Since $\lambda\notin\sigma(k)$ this implies $\sigma(a_2)=0$. But then $\sigma\text{-}\operatorname{dim}(H)=0$; a contradiction.
Now assume that $\lambda^2\in \sigma(k)$. (For example, we can choose
$k=\mathbb{C}(\sqrt{x},\sqrt{x+1},\ldots)$ with action of $\sigma$ determined by $\sigma(x)=x+1$ and $\lambda=\sqrt{x}$.) If $\mu\in k$ with $\sigma(\mu)=\lambda^2$ then $\sigma(y_1^2-\mu y_2^2)=0$. Thus $G$ is not $\sigma$-reduced.
\end{ex}
We have seen in Example \ref{ex: Gso not normal} that the strong identity component need not be normal. Our next goal is to reconcile this difficulty:
\begin{prop} \label{prop: normality for strong comp}
Assume that $k$ is algebraically closed and inversive. Let $G$ be a strongly connected $\sigma$-algebraic group and $H$ a normal $\sigma$-closed subgroup of $G$ with $\sigma\text{-}\operatorname{dim}(H)>0$. Then $H^{so}$ is normal in $G$. (In particular, $H^{so}$ is normal in $H$.)
\end{prop}
For the proof of Proposition \ref{prop: normality for strong comp} we need several preparatory results.
\begin{lemma} \label{lemma: exists good extension}
Assume that $k$ is algebraically closed and inversive. Let $K$ be a $\sigma$-field extension of $k$. Then there exists a $\sigma$-field extension $L$ of $K$ such that only the elements of $k$ are fixed by all $\sigma$-field automorphisms of $L|k$, i.e., $L^{\operatorname{Aut}(L|k)}=k$.
\end{lemma}
\begin{proof}
Let us start with proving the following claim: There exists a $\sigma$-field extension $L$ of $K$ such that for all $a\in K\smallsetminus k$ there exists a $\sigma$-field automorphism $\tau$ of $L|k$ with $\tau(a)\neq a$ and such that every $\sigma$-field automorphism of $K|k$ extends to a $\sigma$-field automorphism of $L|k$.
Since $k$ is algebraically closed $K\otimes_k K$ is an integral domain. Since $k$ is inversive $K\otimes_k K$ is $\sigma$-reduced (Lemma \ref{lemma: tensor stays} (ii)). Therefore the quotient field $L$ of $K\otimes_k K$ is naturally a $\sigma$-field. Consider $L$ as a $\sigma$-field extension of $K$ via the embedding $a\mapsto a\otimes 1$. The $\sigma$-field automorphism $\tau$ of $L|k$ determined by $\tau(a\otimes b)=b\otimes a$ moves every element of $K\smallsetminus k$. Moreover, every $\sigma$-field automorphism $\tau'$ of $K|k$ extends to $L|k$, for example by $\tau'(a\otimes b)=\tau'(a)\otimes b$.
Now let us prove the lemma. By the above claim, there exists a $\sigma$-field extension $L_1|K$ such that every element of $K\smallsetminus k$ can be moved by a $\sigma$-field automorphism of $L_1|K$ and every $\sigma$-field automorphism of $K|k$ extends to a $\sigma$-field automorphism of $L_1|k$. Now apply the claim again to $L_1|k$ to find a $\sigma$-field extension $L_2|L_1$ such that every element of $L_1\smallsetminus k$ can be moved by a $\sigma$-field automorphism of $L_2|k$ and every $\sigma$-field automorphism of $L_1|k$ extends to a $\sigma$-field automorphism of $L_2|k$. Continuing like this we obtain a chain of $\sigma$-field extensions $k\subset K\subset L_1\subset L_2\subset\ldots$. The union $L=\cup L_i$ has the required property.
\end{proof}
We need to know if the strong identity component is compatible with base extension.
\begin{lemma} \label{lemma: strong component stable under baseext}
Assume that $k$ is algebraically closed and inversive. Let $G$ be a $\sigma$-algebraic group with $\sigma\text{-}\operatorname{dim}(G)>0$ and $K$ a $\sigma$-field extension of $k$.
Then $$(G_K)^{so}=(G^{so})_K.$$
\end{lemma}
\begin{proof}
As the $\sigma$-dimension is invariant under base extension (Lemma \ref{lemma: sdim invariant under baseext}),
$$\sigma\text{-}\operatorname{dim}((G^{so})_K)=\sigma\text{-}\operatorname{dim}(G^{so})=\sigma\text{-}\operatorname{dim}(G)=\sigma\text{-}\operatorname{dim}(G_K).$$ Therefore $(G_K)^{so}\leq (G^{so})_K$.
Let us now show that $(G_K)^{so}$ descends to $k$, i.e., there exists a $\sigma$-closed subgroup $H$ of $G$ with $(G_K)^{so}=H_K$. By Lemma \ref{lemma: exists good extension} there exists a $\sigma$-field extension $L$ of $K$ such that $L^{\operatorname{Aut}(L|k)}=k$, where $\operatorname{Aut}(L|k)$ is the group of all $\sigma$-field automorphisms of $L|k$. The group $\operatorname{Aut}(L|k)$ acts on $L\{G_L\}=k\{G\}\otimes_k L$ by $k$-$\sigma$-algebra automorphisms via the right factor. Let $H'$ be a $\sigma$-closed subgroup of $G_L$. Since the Hopf algebra structure maps commute with the $\operatorname{Aut}(L|k)$--action, $\tau(\mathbb{I}(H'))$ is a $\sigma$-Hopf ideal of $k\{G\}\otimes_k L$ for every $\tau\in \operatorname{Aut}(L|k)$. Moreover, the $\sigma$-dimension of the $\sigma$-closed subgroup of $G_L$ defined by $\tau(\mathbb{I}(H'))$ is equal to the $\sigma$-dimension of $H'$ (cf. Lemma \ref{lemma: sdim invariant under baseext}). Since $\mathbb{I}((G_L)^{so})$ is the unique maximal $\sigma$-Hopf ideal of $k\{G\}\otimes_k L$ such that $\sigma\text{-}\operatorname{dim}((G_L)^{so})=\sigma\text{-}\operatorname{dim}(G_L)$, we see that $\tau(\mathbb{I}((G_L)^{so}))=\mathbb{I}((G_L)^{so})$ for every $\tau\in \operatorname{Aut}(L|k)$. Let
$$\mathfrak{a}=\{f\in \mathbb{I}((G_L)^{so})|\ \tau(f)=f\ \forall\ \tau\in \operatorname{Aut}(L|k)\}=\mathbb{I}((G_L)^{so})\cap k\{G\}.$$
Since the action of $\operatorname{Aut}(L|K)$ commutes with the Hopf algebra structure maps, $\mathfrak{a}$ is $\sigma$-Hopf ideal of $k\{G\}$ and therefore corresponds to a $\sigma$-closed subgroup $H$ of $k\{G\}$. We have $\mathfrak{a}\otimes_k L=\mathbb{I}((G_L)^{so})$. (See \cite[Corollary to Proposition 6, Chapter V, \S 10.4, A.V.63]{Bourbaki:Algebra2}.) So $H_L=(G_L)^{so}$.
As $\sigma\text{-}\operatorname{dim}(H)=\sigma\text{-}\operatorname{dim}((G_L)^{so})=\sigma\text{-}\operatorname{dim}(G_L)=\sigma\text{-}\operatorname{dim}(G)$ we see that $G^{so}\leq H$, therefore $(G^{so})_L\leq H_L=(G_L)^{so}$. Hence also
$$((G^{so})_K)_L=(G^{so})_L\leq (G_L)^{so}\leq ((G_K)^{so})_L.$$
Thus $(G^{so})_K\leq (G_K)^{so}$.
\end{proof}
The following example shows that the formation of the strongly connected component is in general not compatible with base extension.
\begin{ex}
Let $G$ be the strongly connected $\sigma$-algebraic group from Example \ref{ex: strongly connected but not sreduced}. Let $K=k^*$ be the inversive closure of $k$ (see (\cite[Definition 2.1.6]{Levin:difference})) and let $\mu\in K$ with $\sigma(\mu)=\lambda$. Then $G_K$ is not strongly connected since it has the $\sigma$-closed subgroup $H$ of $\sigma\text{-}\operatorname{dim}(H)=1=\sigma\text{-}\operatorname{dim}(G)$ given by
$$H(R)=\{(g_1,g_2)\in R^2|\ g_1=\mu g_2\}$$ for any $k$-$\sigma$-algebra $R$. So $(G_K)^{so}$ is properly contained in $(G^{so})_K=G_K$.
\end{ex}
Now we are prepared to prove Proposition \ref{prop: normality for strong comp}.
\begin{proof}[Proof of Proposition \ref{prop: normality for strong comp}]
We have to show that the morphism of $\sigma$-varieties
$$\phi\colon G\times H^{so}\to G,\ (g,h)\mapsto ghg^{-1}$$ maps into $H^{so}$.
We know from Lemma \ref{lemma: strongly connected is sintegral} that $H$ and $G$ are $\sigma$-integral. A fortiori $H$ and $G$ are perfectly $\sigma$-reduced. Because of Lemma \ref{lemma: tensor stays} (iv) also the product $G\times H^{so}$ is perfectly $\sigma$-reduced. So by Lemma \ref{lemma: morphism with perfectly sreduced}, it suffices to show that $\phi_K((G\times H^{so})(K))\subset H^{so}(K)$ for every $\sigma$-field extension $K$ of $k$. Let $g\in G(K)$. Then $g$ induces an automorphism of $G_K$ by conjugation. Since $H$ is normal in $G$ we have an induced automorphism on $H_K$. This automorphism maps $(H_K)^{so}$ into $(H_K)^{so}$. But $(H_K)^{so}=(H^{so})_K$ by Lemma \ref{lemma: strong component stable under baseext}. This shows that conjugation by $g$ maps $H^{so}(K)$ into $H^{so}(K)$. Thus $\phi_K((G\times H^{so})(K))\subset H^{so}(K)$ as required.
\end{proof}
\begin{defi}
Let $G$ and $H$ be strongly connected $\sigma$-algebraic groups of positive $\sigma$-dimension. A morphism $\phi\colon G\to H$ is called an \emph{isogeny} if $\phi$ is surjective and $\sigma\text{-}\operatorname{dim}(\ker(\phi))=0$. Two strongly connected $\sigma$-algebraic groups $H_1$, $H_2$ of positive $\sigma$-dimension are called \emph{isogenous} if there exists a strongly connected $\sigma$-algebraic group $G$ and isogenies $G\twoheadrightarrow H_1$, $G\twoheadrightarrow H_2$.
\end{defi}
By Theorem \ref{theo: isom1} and Corollary \ref{cor: sdim and ord for quotients} a surjective morphism $\phi\colon G\twoheadrightarrow H$ is an isogeny, if and only if $\sigma\text{-}\operatorname{dim}(G)=\sigma\text{-}\operatorname{dim}(H)$. In particular, isogenous $\sigma$-algebraic groups have the same $\sigma$-dimension.
\begin{lemma} \label{lemma: composition of isogenies}
The composition of two isogenies is an isogeny.
\end{lemma}
\begin{proof}
Clearly the composition of surjective morphisms is surjective. If $G_1\to G_2$ and $G_2\to G_3$ are isogenies, then $\sigma\text{-}\operatorname{dim}(G_1)=\sigma\text{-}\operatorname{dim}(G_2)$ and $\sigma\text{-}\operatorname{dim}(G_2)=\sigma\text{-}\operatorname{dim}(G_3)$. Therefore $\sigma\text{-}\operatorname{dim}(G_1)=\sigma\text{-}\operatorname{dim}(G_3)$.
\end{proof}
\begin{lemma}
Isogeny is an equivalence relation on the set of strongly connected $\sigma$-algebraic groups of positive $\sigma$-dimension.
\end{lemma}
\begin{proof}
Reflexivity and symmetry are obvious. Let us prove the transitivity. So let $\phi_1\colon G\twoheadrightarrow H_1$, $\phi_2\colon G\twoheadrightarrow H_2$ and $\phi_2'\colon G'\twoheadrightarrow H_2$, $\phi_3'\colon G'\twoheadrightarrow H_3$ be isogenies. The morphism $\phi_2\times\phi_2'\colon G\times G'\to H_2\times H_2$
is surjective with kernel $\ker(\phi_2)\times\ker(\phi_2')$, which has $\sigma$-dimension zero by Lemma \ref{lemma: ord for product}. The diagonal
$D\leq H_2\times H_2$ given by $D(R)=\{(h_2,h_2)|\ h_2\in H_2(R)\}$ for any $k$-$\sigma$-algebra $R$ is a $\sigma$-closed subgroup of $H_2\times H_2$ isomorphic to $H_2$. Therefore $G''=((\phi_2\times\phi_2')^{-1}(D))^{so}$ is a $\sigma$-closed subgroup of $G\times G'$ with $\sigma\text{-}\operatorname{dim}(G'')=\sigma\text{-}\operatorname{dim}(H_2)$. Let $\pi\colon G''\to G$ and $\pi'\colon G''\to G'$ denote the projections onto the first and second factor respectively. We have the following diagram
$$
\xymatrix{
& & G'' \ar_\pi[ld] \ar^{\pi'}[rd] & & \\
& G \ar_-{\phi_1}[ld] \ar^-{\phi_2}[rd] & & G' \ar_-{\phi_2'}[ld] \ar^-{\phi_3'}[rd] & \\
H_1 & & H_2 & & H_3 \\
}
$$
We claim that $\pi$ and $\pi'$ are isogenies. We have $\ker(\pi)\leq 1\times \ker(\phi_2')$. Therefore $\sigma\text{-}\operatorname{dim}(\ker(\pi))=0$ and consequently $$\sigma\text{-}\operatorname{dim}(\pi(G''))=\sigma\text{-}\operatorname{dim}(G'')=\sigma\text{-}\operatorname{dim}(H_2)=\sigma\text{-}\operatorname{dim}(G).$$ Since $G$ is strongly connected, this shows that $\pi(G'')=G$, so $\pi$ is surjective. Hence $\pi$ is an isogeny. Similarly, it follows that $\pi'$ is an isogeny. The isogenies $\phi_1\pi$ and $\phi_3'\pi'$ (Lemma \ref{lemma: composition of isogenies}) now show that $H_1$ and $H_3$ are isogenous.
\end{proof}
\begin{theo} \label{theo: Jordan Hoelder}
Assume that $k$ is algebraically closed and inversive. Let $G$ be a strongly connected $\sigma$-algebraic group with $\sigma\text{-}\operatorname{dim}(G)>0$. Then there exists a subnormal series \begin{equation} \label{eqn: G subnormal series} G=G_0\supset G_1\supset\cdots\supset G_n=1 \end{equation}
such that $\sigma\text{-}\operatorname{dim}(G_i)>0$, $G_i$ is strongly connected, $\sigma\text{-}\operatorname{dim}(G_i/G_{i+1})>0$ and $G_i/G_{i+1}$ is almost--simple for $i=0,\ldots,n-1$. If
\begin{equation} \label{eqn: H subnormal series} G=H_0\supset H_1\supset\cdots\supset H_m=1 \end{equation}
is another such subnormal series, then $m=n$ and there exists a permutation $\pi$ such that
$G_{i}/G_{i+1}$ and $H_{\pi(i)}/H_{\pi(i+1)}$ are isogenous for $i=0,\ldots,n-1$.
\end{theo}
\begin{proof}
Let us first prove the existence statement.
Among all normal proper $\sigma$-closed subgroups of $G$ choose one, say $H$, with maximal $\sigma$-dimension. If $\sigma\text{-}\operatorname{dim}(H)=0$, $G$ is almost--simple and we are done. So let us assume that $\sigma\text{-}\operatorname{dim}(H)>0$. Since $G$ is strongly connected, $\sigma\text{-}\operatorname{dim}(H)<\sigma\text{-}\operatorname{dim}(G)$. By construction $G/H$ is almost--simple (Theorem \ref{theo: isom3}). Let $G_1=H^{so}$. Then $G_1$ is normal in $G$ by Proposition \ref{prop: normality for strong comp} and $\sigma\text{-}\operatorname{dim}(G/G_1)>0$. Let us show that $G/G_1$ is almost--simple. Let $N$ be a proper normal $\sigma$-closed subgroup of $G$ containing $G_1$. By choice of $H$, $\sigma\text{-}\operatorname{dim}(N)\leq\sigma\text{-}\operatorname{dim}(H)$, but since $\sigma\text{-}\operatorname{dim}(H)=\sigma\text{-}\operatorname{dim}(G_1)\leq\sigma\text{-}\operatorname{dim}(N)$ we have $\sigma\text{-}\operatorname{dim}(N)=\sigma\text{-}\operatorname{dim}(G_1)$. So $G/G_1$ is almost--simple (Theorem \ref{theo: isom3}). As $\sigma\text{-}\operatorname{dim}(G_1)<\sigma\text{-}\operatorname{dim}(G)$ the claim follows by induction on $\sigma\text{-}\operatorname{dim}(G)$.
Now let us prove the uniqueness statement. It follows from Theorem \ref{theo: schreier refinement} that (\ref{eqn: G subnormal series}) and (\ref{eqn: H subnormal series}) have equivalent refinements. Let
\begin{equation} \label{eqn: G refinement} G=G_0\supsetneqq G_{0,1}\supsetneqq G_{0,2}\supsetneqq\cdots\supsetneqq G_{0,n_0}\supsetneqq G_1 \supsetneqq G_{1,1}\supsetneqq\cdots\supsetneqq G_{1,n_1}\supsetneqq G_2\supsetneqq\cdots\supsetneqq G_n=1 \end{equation} be such a refinement of (\ref{eqn: G subnormal series}). For $i=0,\ldots,n-1$, as $G_i$ is strongly connected and $G_i/G_{i+1}$ is almost--simple, $\sigma\text{-}\operatorname{dim}(G_i/G_{i,1})=\sigma\text{-}\operatorname{dim}(G_i/G_{i+1})>0$ and
$\sigma\text{-}\operatorname{dim}(G_{i,j}/G_{i,j+1})=0$ for $j=1,\ldots,n_i-1$, also $\sigma\text{-}\operatorname{dim}(G_{i,n_i}/G_{i+1})=0$.
The kernel of $G_i/G_{i+1}\twoheadrightarrow G_i/G_{i,1}$ has $\sigma$-dimension zero, so $G_i/G_{i+1}$ and $G_i/G_{i,1}$ are isogenous. In summary, we find that among the quotient groups of the subnormal series (\ref{eqn: G refinement}), there are precisely $n$ of positive $\sigma$-dimension, (namely $G_i/G_{i,1}$, $i=0,\ldots,n-1$). A similar statement applies to the equivalent refinement of (\ref{eqn: H subnormal series}). Therefore $n=m$ and there is a permutation $\pi$ such that
$G_{i}/G_{i+1}$ and $H_{\pi(i)}/H_{\pi(i+1)}$ are isogenous for $i=0,\ldots,n-1$.
\end{proof}
\begin{rem}
It is clear from the proof that the uniqueness statement in Theorem \ref{theo: Jordan Hoelder} is also valid without any restriction on the base $\sigma$-field $k$.
\end{rem}
\section{An application to $\sigma$-Galois theory}
A Galois theory for linear differential equations depending on a discrete parameter has been developed in \cite{diVizioHardouinWibmer:DifferenceGaloisofDifferential}. In this Galois theory the Galois groups are affine difference algebraic groups.
In this final section we show that under the Galois correspondence the numerical invariants introduced above, namely the difference dimension, the order and the limit degree of the Galois group correspond to well--known invariants of a difference field extension.
Let us recall the basic facts from \cite{diVizioHardouinWibmer:DifferenceGaloisofDifferential}. Let $K$ be $\delta\sigma$-field, i.e., $K$ is a field of characteristic zero equipped with a derivation $\delta\colon K\to K$ and an endomorphism $\sigma\colon K\to K$ which commute up to a factor. Because of this commutativity requirement the field $k=K^\delta=\{a\in K|\ \delta(a)=0\}$ is a $\sigma$-field. For example, $K$ could be $\mathbb{C}(\alpha,x)$ with $\delta$ the derivation with respect to $x$ and action of $\sigma$ given by $\sigma(f(\alpha,x))=f(\alpha+1,x)$, in which case $k=\mathbb{C}(\alpha)$ and $\alpha$ is ``the discrete parameter''.
We are interested in linear differential systems $\delta(y)=Ay$ with $A\in K^{n\times n}$. A field extension $L$ of $K$ equipped with extensions of $\delta$ and $\sigma$ is called a \emph{$\sigma$-Picard-Vessiot extension} for $\delta(y)=Ay$ if
\begin{enumerate}
\item there exists $Y\in\operatorname{GL}_n(L)$ with $\delta(Y)=AY$ such that the entries of $Y,\sigma(Y),\sigma^2(Y),\ldots$ generate $L$ as a field extension of $K$ and
\item $L^\delta=K^\delta$.
\end{enumerate}
Such a $\sigma$-Picard-Vessiot extension exists under rather mild assumptions and it is in a certain sense unique. (See \cite[Section 1]{diVizioHardouinWibmer:DifferenceGaloisofDifferential} and \cite[Section 2]{Wibmer:Chevalley} for more details.)
Let $S$ be the $K$-subalgebra of $L$ generated by $Y,\frac{1}{\det(Y)},\sigma(Y),\frac{1}{\det(\sigma(Y))},\ldots$. Note that $R$ is stable under $\delta$ and $\sigma$. The $\sigma$-Galois group $G$ of $L|K$ is the functor from the category of $k$-$\sigma$-algebras to the category of groups, defined by associating to every $k$-$\sigma$-algebra $R$ the group $\operatorname{Aut}(S\otimes_k R| K\otimes_k R)$ of all automorphisms of $S\otimes_k R$ over $K\otimes_K R$ which commute with $\delta$ and $\sigma$. (The action of $\delta$ on $R$ is understood to be trivial.) One can show that $G$ is a difference algebraic group (\cite[Proposition 2.5]{diVizioHardouinWibmer:DifferenceGaloisofDifferential}). Moreover, the choice of $Y\in\operatorname{GL}_n(L)$ determines a $\sigma$-closed embedding $G\hookrightarrow \operatorname{GL}_n$ of difference algebraic groups.
Let us also recall the definitions of the basic invariants of difference field extensions from \cite{Levin:difference}.
The difference transcendence degree $\sigma\text{-}\operatorname{trdeg}(L|K)$ of a $\sigma$-field extension $L|K$ may be defined as the supremum over all non--negative integers $n$ such that the difference polynomial ring in $n$ difference variables over $K$ can be embedded into $L$. The order $\operatorname{ord}(L|K)$ is the transcendence degree of $L|K$. If there exists a finite set $B\subset L$ such that $B,\sigma(B),\ldots$ generates $L$ as a field extension of $K$, then the limit degree ${\operatorname{ld}}(L|K)$ is the limit $\lim_{i\to\infty} d_i$, where $d_i$ is the degree of the field extension
$K(B,\ldots,\sigma^{i+1}(B))|K(B,\ldots,\sigma^{i}(B))$. The limit exists and does not depend on the choice of $B$ (\cite[Section 4.3]{Levin:difference}).
Note that the equality $\sigma\text{-}\operatorname{trdeg}(L|K)=\sigma\text{-}\operatorname{dim}(G)$ has already been proved in \cite{diVizioHardouinWibmer:DifferenceGaloisofDifferential} but with a different (though equivalent) definition of the difference dimension of $G$.
\begin{theo} \label{theo: Galoisapplication}
Let $L|K$ be a $\sigma$-Picard-Vessiot extension with $\sigma$-Galois group $G$. Then
$$\sigma\text{-}\operatorname{trdeg}(L|K)=\sigma\text{-}\operatorname{dim}(G),$$
$$\operatorname{ord}(L|K)=\operatorname{ord}(G)$$ and
$${\operatorname{ld}}(L|K)={\operatorname{ld}}(G).$$
\end{theo}
\begin{proof}
Let $A$ and $Y$ be as above.
For $i\geq 0$ let $G[i]$ denote the $i$-th order Zariski closure $G$ in $\operatorname{GL}_n$. By Proposition 2.15 in \cite{diVizioHardouinWibmer:DifferenceGaloisofDifferential} the differential field
$$L_i=K\big(Y,\sigma(Y),\sigma^i(Y),\tfrac{1}{\det(Y\cdots\sigma^i(Y))}\big)\subset L$$ is a (standard) Picard--Vessiot extension with Galois group $G[i]$. Therefore $\operatorname{trdeg}(L_i|k)=\dim(G[i])$ for all $i\geq 0$ (\cite[Corollary 1.30]{SingerPut:differential}).
By \cite[Theorem 4.1.17]{Levin:difference} there exists a non--negative integer $e$ such that $\operatorname{trdeg}(L_i|K)=\sigma\text{-}\operatorname{trdeg}(L|K)(i+1)+e$ for $i\gg 0$. Similarly, by Theorem \ref{theo: existence sdim} we have $\dim(G[i])=\sigma\text{-}\operatorname{dim}(G)(i+1)+e$. Therefore $\sigma\text{-}\operatorname{trdeg}(L|K)=\sigma\text{-}\operatorname{dim}(G)$.
The equality $\operatorname{ord}(L|K)=\operatorname{ord}(G)$ also follows from $\operatorname{trdeg}(L_i|k)=\dim(G[i])$.
For $i\geq 1$ let $\mathcal{G}_i$ denote the kernel of the projection $G[i]\twoheadrightarrow G[i-1]$. Then $\mathcal{G}_i$ is the Galois group of the Picard-Vessiot extension $L_i|L_{i-1}$ and therefore the size of $\mathcal{G}_i$ equals the degree of $L_i|L_{i-1}$. This shows that ${\operatorname{ld}}(L|K)={\operatorname{ld}}(G)$.
\end{proof}
\bibliographystyle{alpha}
|
\section{Introduction}\label{section introduction}
A prominent class of random matrix models is the Wigner ensemble, consisting of $N\times N$ real symmetric
or complex Hermitian matrices $W = (w_{ij} )$ whose matrix entries are random variables that are
independent up to the symmetry constraint $W = W^*$. The first rigorous result about the spectrum of random matrices of this type is Wigner's global semicircle law~\cite{Wigner}, which states that the empirical distribution of the rescaled eigenvalues, $(\lambda_i)$, of a Wigner matrix~$W$ is given by
\begin{align}\label{convergence to semicircle law}
\frac{1}{N}\sum_{i=1}^N\delta_{\lambda_i}(E)\longrightarrow \rho_{sc}(E)\mathrel{\mathop:}= \frac{1}{2\pi}\sqrt{(4-E^2)_+}\,,\qquad\qquad (E\in\ensuremath{\mathbb{R}})\,,
\end{align}
as $N\to\infty$, in the weak sense. The distribution $\rho_{sc}$ is called the \emph{semicircle law}.
Let $p_W^N(\lambda_1,\ldots,\lambda_N)$ denote the joint probability density of the (unordered) eigenvalues of $W$.
If the entries of the Wigner matrix $W$ are i.i.d.\ real or complex Gaussian random variables, the joint density of the eigenvalues, $p_W^N\equiv p_{G}^N$, is given by
\begin{align}\label{the GUE}
p_{G}^N(\lambda_1,\ldots\lambda_N)=\frac{1}{Z_{G}^N} \prod_{i<j} |\lambda_i-\lambda_j|^{\beta}\e{-\beta N\sum_{i=1}^N \lambda_{i}^2/4}\,,
\end{align}
with $\beta=1,2$, for the real, complex case, respectively. The normalization $Z_{G}^N\equiv Z_{G}^N(\beta)$ in~\eqref{the GUE} can be computed explicitly. The real and complex Gaussian matrix ensembles so defined are known as the Gaussian orthogonal ensemble (GOE, $\beta=1$) and Gaussian unitary ensemble (GUE, $\beta=2$), respectively, and as noted above we denote the corresponding joint densities as $p_{G}^N$ instead of $p_W^N$.
The \emph{$n$-point correlation functions} are defined by
\begin{align}
\varrho_{W,n}^N(\lambda_1,\ldots,\lambda_n)\mathrel{\mathop:}=\int_{\ensuremath{\mathbb{R}}^{N-n}} p_W^N(\lambda_1,\lambda_2,\ldots,\lambda_N)\,\mathrm{d} \lambda_{n+1}\,\mathrm{d} \lambda_{n+2}\cdots\mathrm{d} \lambda_{N}\,,\qquad\qquad (1\le n\le N)\,.
\end{align}
Using orthogonal polynomials the correlation functions of the GUE and GOE have been explicitly computed by Dyson, Gaudin and Mehta (see, e.g.,~\cite{Mbook}). For the Gaussian unitary ensemble, their results assert that the limiting behavior on small scales at a fixed energy $E$ in the bulk of the spectrum, i.e., for $|E|<2$, satisfies
\begin{align}\label{GUE correlation functions}
\frac{1}{(\rho_{sc}(E))^n} \varrho_{G,n}^N\left( E+\frac{\alpha_1}{\rho_{sc}(E)N},E+\frac{\alpha_2}{\rho_{sc}(E)N},\ldots,E+\frac{\alpha_n}{\rho_{sc}(E)N}\right)\longrightarrow\det\left(K(\alpha_i-\alpha_j)\right)_{i,j=1}^n\,,
\end{align}
as $N \to \infty$, where $K$ is the sine-kernel,
\begin{align}
K(x,y)\mathrel{\mathop:}=\frac{\sin\pi(x-y)}{\pi(x-y)}\,.
\end{align}
Note that the limit in~\eqref{GUE correlation functions} is independent of the energy $E$ as long as $E$ is in the bulk of the spectrum.
The rescaling by a factor $1/N$ of the correlation functions in~\eqref{GUE correlation functions} corresponds to the typical separation of
consecutive eigenvalues and we refer to the law under such a scaling as \emph{local statistics}. Similar but
more complicated formulas were also obtained for the Gaussian orthogonal ensemble; see, e.g.,~\cite{AGZ,Mbook} for reviews. Note that the limiting correlation functions do not factorize, reflecting the fact that the eigenvalues remain strongly correlated in the limit of large $N$.
The Wigner-Dyson-Gaudin-Mehta conjecture, or bulk universality conjecture, states that the local eigenvalue statistics of Wigner matrices are universal in the sense that they depend only on the symmetry class of the matrix, but are otherwise independent of the details of the distribution of the matrix entries. The bulk universality can be formulated in terms of weak convergence of correlation functions or in terms of eigenvalue gap statistics. This conjecture for all symmetry classes has been established in a series of papers~\cite{EPRSY,EKYY2,ESY4,EYY,EKYY4,single_gap}. After this work began, parallel results were obtained in certain cases in~\cite{TV1,TV2}.
In the present paper, we consider deformed Wigner matrices. A deformed Wigner matrix, $H$, is an $N\times N$ random matrix of the form
\begin{align}\label{la matrice}
H = V + W \,,
\end{align}
where $V$ is a real, diagonal, random or deterministic matrix and $W$ is a real symmetric or complex
Hermitian Wigner matrix independent of $V$.
The matrices are normalized so that the eigenvalues
of $V$ and $W$ are order one. If the entries of $V$ are random we may think of $V$ as a ``random potential''; if the entries of $V$ are deterministic, matrices of the form~\eqref{la matrice} are sometimes referred to as ``Wigner matrices with external source''.
Assuming that the empirical eigenvalue distribution of $V=\diag(v_1,\ldots,v_N)$,
\begin{align*}
\widehat\nu\mathrel{\mathop:}=\frac{1}{N}\sum_{i=1}\delta_{v_i}\,,
\end{align*}
converges weakly, respectively weakly in probability, to a non-random measure, $\nu$, it was shown in~\cite{P} that the empirical distribution of the eigenvalues of $H$ converges weakly in probability to a deterministic measure. This measure depends on~$\nu$ and is thus in general distinct from $\rho_{sc}$. We refer to it as the \emph{deformed semicircle law}, henceforth denoted by~$\rho_{fc}$. There is no explicit formula for~$\rho_{fc}$ in terms of~$\nu$. Instead,~$\rho_{fc}$ is obtained as the solution of a functional equation for its Stieltjes transform (see~\eqref{mfc equation} below). It is known that~$\rho_{fc}$ admits a density~\cite{B}. Depending on~$\nu$,~$\rho_{fc}$ may be supported on several disjoint intervals. For simplicity, we assume below that~$\nu$ is such that~$\rho_{fc}$ is supported on a single bounded interval. Further, we choose $\widehat\nu$ such that all eigenvalues of $H$ remain close to the support of $\rho_{fc}$, i.e., there are no ``outliers'' for~$N$ sufficiently large.
If $W$ belongs to the GUE, $H$ is said to belong to the deformed GUE. The deformed GUE for the special case when~$V$ has two eigenvalues~$\pm a$, each with equal multiplicity, has been treated in a series of papers~\cite{BK1,BK2,BK3}. In this setting the local eigenvalue statistics of~$H$ can be obtained via the solution to a Riemann-Hilbert problem; see also~\cite{CW} for the case when~$V$ has equispaced eigenvalues. Bulk universality for correlation functions of the deformed GUE with rather general deterministic or random $V$ has been proved in~\cite{S1} by means of the Brezin-Hikami/Johansson integration formula.
In the present paper, we establish bulk universality of local averages of correlation functions for deformed Wigner matrices of the form $H=V+W$, where $W$ is a real symmetric or complex Hermitian Wigner matrix and~$V$ is a deterministic or random real diagonal matrix. We assume that the entries of~$W$ are centered independent random variables with variance $1/N$ whose distributions decay sub-exponentially; see Definition~\ref{assumption wigner}. If~$V$ is random, we assume for simplicity that its entries $(v_i)$ are i.i.d.\ random variables. We assume that~$\widehat\nu$ converges weakly, respectively weakly in probability, to a non-random measure~$\nu$; see Assumption~\ref{assumption mu_V convergence}. We further assume that the corresponding deformed semicircle law~$\rho_{fc}$ is supported on a single compact interval and has square root decay at both endpoints. Sufficient conditions for these assumptions to hold have appeared in~\cite{S2} and are rephrased in Assumption~\ref{assumption mu_V}. Under these
assumptions, our main results in Theorem~\ref{theorem 1} and in Theorem~\ref{theorem 2} assert that the limiting correlation functions of the deformed Wigner ensemble are universal when averaged over a small energy window. Note that our results hold for complex Hermitian and real symmetric deformed Wigner matrices.
Before we outline our proofs, we recall the notion of {\it $\beta$-ensemble} or {\it log-gas} which generalizes the measures in~\eqref{the GUE}. Let $U$ be a real-valued potential and consider the measure on $\ensuremath{\mathbb{R}}^N$ defined by the density
\begin{align}\label{le beta ensemble}
\mu_{U}^N(\lambda_1,\ldots,\lambda_N)\mathrel{\mathop:}=\frac{1}{Z_{U}^N} \prod_{i<j} |\lambda_i-\lambda_j|^{\beta}\e{-{\beta N}\sum_{i=1}^N \left(\lambda_{i}^2/2+U(\lambda_i)\right)/2}\,,
\end{align}
where $\beta>0$ and $Z_{U}^N\equiv Z_{U}^N(\beta)$ is a normalization. Bulk universality for $\beta$-ensembles asserts that the local correlation functions for measures of the form~\eqref{le beta ensemble} are universal (for sufficiently regular potentials $U$) in the sense that for each value of $\beta>0$ they agree with the local correlation functions of the Gaussian ensemble with $U\equiv 0$.
For the classical values $\beta\in\{1,2,4\}$, the eigenvalue correlation functions of $\mu_U^N$ can be explicitly expressed in terms of polynomials orthogonal to the exponential weight in~\eqref{le beta ensemble}.
Thus the analysis of the correlation functions relies on the asymptotic properties of the
corresponding orthogonal polynomials.
This approach, initiated by Dyson, Gaudin and Mehta (see~\cite{Mbook} for a review), was the starting point for many results on
the universality for $\beta$-ensemble with $\beta\in\{1,2,4\}$~\cite{BI,DKMcLVZ,DKMcLVZ2,DG,KSh,ShMa1}.
For general $\beta>0$, bulk universality of $\beta$-ensembles has been established in~\cite{BEYI,BEYII,BEY} for potentials $U\in C^4$. Recently, alternative approaches to bulk universality for $\beta$-ensembles with general $\beta$ have been presented in~\cite{ShMa} and~\cite{BAG}.
We emphasize at this point that the eigenvalue distributions of the deformed ensemble in~\eqref{la matrice} are in general not of the form~\eqref{le beta ensemble}, even when $W$ belongs to the GUE or the GOE.
Returning to the random matrix setting, we recall that the general approach to bulk universality for (generalized) Wigner matrices in~\cite{EPRSY,ESY4,EYY} consists of three steps:
\begin{itemize}
\item[(1)] establish a local semicircle law for the density of eigenvalues;
\item[(2)] prove universality of Wigner matrices with a small Gaussian component by analyzing the convergence of Dyson Brownian motion to local equilibrium;
\item[(3)] compare the local statistics of Wigner ensembles with Gaussian divisible ensembles to remove the small Gaussian component of Step~(2).
\end{itemize}
For an overview of recent results and this three-step strategy see~\cite{EY}. Note that the ``local equilibrium'' in Step~(2) refers to the measure \eqref{the GUE}, with $\beta=1,2$ respectively in the real symmetric, complex Hermitian case.
For deformed Wigner matrices, the \emph{local deformed semicircle law}, the analogue of Step~(1), was established in~\cite{LS} for random~$V$. However, when~$V$ is random the eigenvalues of~$V+W$ fluctuate on scale~$N^{-1/2}$ in the bulk (see \cite{LS}), but their gaps remain rigid on scale $N^{-1}$. To circumvent the mesoscopic fluctuations of the eigenvalue positions we condition on $V$, considering its entries to be fixed. The methods of \cite{LS} can be extended as outlined in Section~\ref{section local law} to prove a local law on the optimal scale for ``typical'' realizations of random as well as deterministic potentials $V$.
Our corresponding version of Step~(2), a proof of bulk universality for deformed Wigner ensembles with small Gaussian component, is the main novelty of this paper. The local equilibrium of Dyson Brownian motion in the deformed case is unknown but may effectively be approximated by a ``reference'' $\beta$-ensemble that we explicitly construct in Section~\ref{section beta ensemble}. In Section~\ref{section DBM}, we analyze the convergence of the local distribution of the deformed Wigner ensemble under Dyson Brownian motion to the ``reference'' $\beta$-ensemble. However, since the ``reference'' $\beta$-ensemble is not given by the invariant GUE/GOE, it also evolves in time.
Using the rigidity estimates for the deformed ensemble established in Step~(1) and the rigidity estimates for general $\beta$-ensembles established in~\cite{BEY}, we obtain in Section~\ref{section DBM} bounds on the time evolution of the relative entropy between the two measures being compared. The idea to estimate the entropy flow of the Dyson Brownian motion with respect to the ``instantaneous global equilibrium state" was initiated in~\cite{ESY4} and~\cite{ESYY}. On the other hand, the entropy flow with respect to time dependent local equilibrium states was initiated in the work~\cite{Y}. In this paper, we combine both methods to yield an effective estimate on the entropy flow of the Dyson Brownian motion in the deformed case. This global entropy estimate is then used in Section~\ref{local equilibrium measures} to conclude that the local statistics of the locally-constrained deformed ensemble with small Gaussian component agree with those of the locally-constrained reference $\beta$-ensemble. Relying
on the
main technical result of~\cite{single_gap}, we further conclude that the local statistics of the locally-constrained reference $\beta$-ensemble agrees with the local statistics of the GUE/GOE. Once this conclusion is obtained for the locally-constrained ensembles, it can be extended to the non-constrained ensembles. This completes Step~(2) in the deformed case.
In the Sections~\ref{From gap statistics to correlation functions} and~\ref{Proofs of main results}, we outline Step~(3) for deformed Wigner matrices; the proof is similar to the argument for Wigner matrices in \cite{EYY1}. The main technical input is a bound on the resolvent entries of $H$ on scales $N^{-1-\epsilon}$ that can be obtained from the local law in Step~(1). In Section~\ref{Proofs of main results}, we then combine Steps~(1)-(3) to conclude the proof of our main results, Theorem~\ref{theorem 1} and Theorem~\ref{theorem 2}.
In the Appendix, we collect several technical results on the deformed semicircle law and its Stieltjes transform. Some of these results have previously appeared in~\cite{S2} and~\cite{LS, LS2}.
\emph{Acknowledgements:}
We thank Paul Bourgade, L\'aszl\'o Erd\H{o}s and Antti Knowles for helpful comments. We are grateful to Chang-Shou Lin and the
Taida Institute for Mathematical Sciences in Taipei for their hospitality during this research. We thank Thomas Spencer and the Institute for Advanced Study for their hospitality during the academic year 2013-2014. J.\ O.\ Lee is partially supported by National Research Foundation of Korea Grant 2011-0013474 and TJ Park Junior Faculty Fellowship. The stay of K.\ Schnelli at the IAS is supported by the Fund for Math. B.\ Stetler is supported by NSF GRFP Fellowship DGE-1144152. H.-T.\ Yau is partially supported by NSF Grant DMS-1307444 and a Simons investigator fellowship.
\section{Assumptions and main results}\label{section assumptions and main results}
In this section, we list our assumptions and our main results.
\subsection{Definition of the model}
We first introduce real symmetric and complex Hermitian Wigner matrices.
\begin{definition}\label{assumption wigner}
A real symmetric Wigner matrix is an $N\times N$ random matrix, $W$, whose entries, $(w_{ij})$, $(1\le i,j\le N)$, are independent (up to the symmetry constraint $w_{ij}={w}_{ji}$) real centered random variables satisfying
\begin{align}\label{moment estimates}
\E w_{ii}^2 = \frac{2}{N}\,, \qquad\quad \E w_{ij}^2 = \frac{1}{N}\,,\qquad \qquad(i\not=j) \,.
\end{align}
In case $(w_{ij})$ are Gaussian random variables, $W$ belongs to the Gaussian orthogonal ensemble (GOE).
A complex Hermitian Wigner matrix is an $N\times N$ random matrix, $W$, whose entries, $(w_{ij})$, $(1\le i,j\le N)$, are independent (up to the symmetry constraint $w_{ij}=\overline{w}_{ji}$) complex centered random variables satisfying
\begin{align}\label{wigner complex}
\E w_{ii}^2 = \frac{1}{N}\,,\qquad \quad \E |w_{ij}|^2 = \frac{1}{N}\,, \qquad\quad \E w_{ij}^2 = 0\,,\qquad \qquad (i \neq j)\,.
\end{align}
For simplicity, we assume that the real and imaginary parts of $(w_{ij})$ are independent for all $i,j$. This ensures that $\E w_{ij}^2=0$, $(i\not=j)$. In case $(\re w_{ij})$ and $(\im w_{ij})$ are Gaussian random variables, $W$ belongs to the Gaussian unitary ensemble (GUE).
Irrespective of the symmetry class of $W$, we always assume that the entries $(w_{ij})$ have a subexponential decay, i.e.,
\begin{align}\label{eq.C0}
\mathbb{P}\left (\sqrt{N} |w_{ij}|>x\right)\le C_0\,\e{-x^{1/\theta}}\,,
\end{align}
for some positive constants $C_0$ and $\theta>1$. In particular,
\begin{align}
\E|w_{ij}|^p\le C\frac{(\theta p)^{\theta p}}{N^{p/2}} \,,\qquad\qquad (p\ge3)\,.
\end{align}
\end{definition}
Let $V=\mathrm{diag}(v_i)$ be an $N\times N$ diagonal, random or deterministic matrix, whose entries $(v_i)$ are real-valued. We denote by $\widehat\nu$ the empirical eigenvalue distribution of the diagonal matrix~$V=\mathrm{diag}(v_i)$,
\begin{align}\label{widehatmuV}
\widehat\nu\mathrel{\mathop:}=\frac{1}{N}\sum_{i=1}^N\delta_{v_i}\,.
\end{align}
\begin{assumption}\label{assumption mu_V convergence}
There is a (non-random) centered compactly supported probability measure $\nu$ such that the following holds.
\begin{itemize}
\item[$(1)$] If $V$ is a {\it random} matrix, we assume that $(v_i)$ are independent and identically distributed real random variables with law $\nu$. Further, we assume that $(v_i)$ are independent of $(w_{ij})$.
\item[$(2)$] If $V$ is a {\it deterministic} matrix, we assume that there is $\alpha_0>0$, such that for any compact set ${\mathcal D}\subset\ensuremath{\mathbb{C}}^+$ with $\mathrm{dist}({\mathcal D},\supp\nu)>0$, there is $C$ such that
\begin{align}\label{equation assumption mu_V convergence}
\max_{z\in {\mathcal D}}\left|\int\frac{\mathrm{d}\widehat\nu(v)}{v-z}-\int\frac{\mathrm{d}\nu(v)}{v-z} \right|\le CN^{-\alpha_0} \,,
\end{align}
for $N$ sufficiently large.
\end{itemize}
\end{assumption}
Note that~\eqref{equation assumption mu_V convergence} implies that $\widehat\nu$ converges to $\nu$ in the weak sense as $N\to\infty$. Also note that the condition~\eqref{equation assumption mu_V convergence} holds for large $N$ with high probability for $0<\alpha_0<1/2$ if $(v_i)$ are i.i.d.\ random variables.
\subsection{Deformed semicircle law}\label{subsection deforemd semicirle law}
The deformed semicircle can be described in terms of the Stieltjes transform: For a (probability) measure $\omega$ on the real line we define its Stieltjes transform, $m_{\omega}$, by
\begin{align*}
m_{\omega}(z)\mathrel{\mathop:}=\int\frac{\mathrm{d}\omega(v)}{v-z}\,,\qquad\qquad (z\in\ensuremath{\mathbb{C}}^+)\,.
\end{align*}
Note that $m_{\omega}$ is an analytic function in the upper half plane and that $\im m_\omega(z)\ge 0$, $\im z> 0$. Assuming that $\omega$ is absolutely continuous with respect to Lebesgue measure, we can recover the density of $\omega$ from $m_{\omega}$ by the inversion formula
\begin{align}\label{stieltjes inversion formula}
\omega(E)=\lim_{\eta\searrow 0}\frac{1}{\pi}\im m_{\omega}(E+\mathrm{i} \eta)\,,\qquad\qquad (E\in\ensuremath{\mathbb{R}})\,.
\end{align}
We use the same symbols to denote measures and their densities. Moreover, we have
\begin{align*}
\lim_{\eta\searrow 0} \re m_{\omega}(E+\mathrm{i}\eta)=\Xint- \frac{\omega(v)\,\mathrm{d} v}{v-E}\,,\qquad\qquad (E\in\ensuremath{\mathbb{R}})\,,
\end{align*}
whenever the left side exists. Here the integral on the right is understood as principle value integral. We denote in the following by $\re m_{\omega}(E)$ and $\im m_{\omega}(E)$ the limiting quantities
\begin{align}\label{abuse of notation I}
\re m_{\omega}(E)\equiv \lim_{\eta\searrow 0}\re m_{\omega}(E+\mathrm{i}\eta)\,,\qquad\quad \im m_{\omega}(E)\equiv \lim_{\eta\searrow 0}\im m_{\omega}(E+\mathrm{i}\eta)\,,\qquad\qquad (E\in\ensuremath{\mathbb{R}})\,,
\end{align}
whenever the limits exist.
Choosing $\omega$ to be the standard semicircular law $\rho_{sc}$, the Stieltjes transform $m_{\rho_{sc}}\equiv m_{sc}$ can be computed explicitly and one checks that $m_{sc}$ satisfies the relation
\begin{align*}
m_{sc}(z)=\frac{-1}{m_{sc}(z)+z}\,,\qquad\quad \im m_{sc}(z)\ge0\,,\qquad\qquad (z\in\ensuremath{\mathbb{C}}^+)\,.
\end{align*}
The deformed semicircle law is conveniently defined through its Stieltjes transform. Let $\nu$ be the limiting probability measure of Assumption~\ref{assumption mu_V convergence}. Then it is well-known~\cite{P} that the functional equation
\begin{align}\label{mfc equation}
m_{fc}(z)= \int\frac{\mathrm{d}\nu(v)}{v-z-m_{fc}(z)}\,,\qquad\quad \im m_{fc}(z)\ge 0\,,\qquad\qquad (z\in\ensuremath{\mathbb{C}}^+)\,,
\end{align}
has a unique solution, also denoted by $m_{fc}$, that satisfies $\limsup_{\eta\searrow 0}\im m_{fc}(E+\mathrm{i}\eta)<\infty$, for all $E\in\ensuremath{\mathbb{R}}$. Indeed, from~\eqref{mfc equation}, we obtain that
\begin{align}\label{sumrule}
\int\frac{\mathrm{d}\nu(v)}{|v-z-m_{fc}(z)|^2}=\frac{\im m_{fc}(z)}{\im m_{fc}(z)+\eta}\le 1\,,\qquad\qquad (z\in\ensuremath{\mathbb{C}}^+)\,,
\end{align}
thus $|m_{fc}(z)|\le 1$, for all $z\in\ensuremath{\mathbb{C}}^+$.
The deformed semicircle law, denoted by $\rho_{fc}$, is then defined through its density
\begin{align*}
\rho_{fc}(E)\mathrel{\mathop:}=\lim_{\eta\searrow 0}\frac{1}{\pi}\im m_{fc}(E+\mathrm{i}\eta)\,,\qquad \qquad(E\in\ensuremath{\mathbb{R}})\,.
\end{align*}
The measure $\rho_{fc}$ has been studied in detail in~\cite{B}. For example, it was shown there that the density $\rho_{fc}$ is an analytic function inside the support of the measure.
The measure $\rho_{fc}$ is also called the additive free convolution of the semicircular law and the measure~$\nu$. More generally, the additive free convolution of two (probability) measure $\omega_1$ and $\omega_2$, usually denoted by $\omega_1\boxplus\omega_2$, is defined as the distribution of the sum of two freely independent non-commutative random variables, having distributions $\omega_1$,~$\omega_2$ respectively; we refer to, e.g.,~\cite{VDN,AGZ} for reviews. Similarly to~\eqref{mfc equation}, the free convolution measure $\omega_1\boxplus\omega_2$ can be described in terms of a set of functional equations for the Stieltjes transforms; see~\cite{CG, BeB07}.
Our second assumption on $\nu$ guarantees (see Lemma~\ref{lemma mfc} below) that~$\rho_{fc}$ is supported on a single interval and that~$\rho_{fc}$ has a square root behavior at the two endpoints of the support. Sufficient conditions for this behavior have been presented in~\cite{S2}. The assumptions below also rule out the possibility that the matrix $H$ has ``outliers'' in the limit of large $N$.
\begin{assumption}\label{assumption mu_V}
Let $I_{\nu}$ be the smallest closed interval such that $\supp{\nu}\subseteq I_{\nu}$. Then, there exists $\varpi>0$ such that
\begin{align}\label{eq assumption mu_V}
\inf_{x\in I_{\nu}}\int\frac{\mathrm{d}\nu(v)}{(v-x)^2}\ge 1+\varpi\,.
\end{align}
Similarly, let $I_{\widehat\nu}$ be the smallest closed interval such that $\supp{\widehat\nu}\subseteq I_{\widehat\nu}$. Then,
\begin{itemize}
\item[$(1)$] for {\it random} $(v_i)$, there is a constant $\mathfrak{t}>0$, such that
\begin{align}\label{eq assumption mu_V 2}
\mathbb{P}\left(\inf_{x\in I_{\widehat\nu}}\int\frac{\mathrm{d}\widehat\nu(v)}{(v-x)^2}\ge 1+{\varpi} \right)\ge 1-N^{-\mathfrak{t}}\,,
\end{align}
for $N$ sufficiently large;
\item[$(2)$] for {\it deterministic} $(v_i)$,
\begin{align}\label{eq assumption mu_V 3}
\inf_{x\in I_{\widehat\nu}}\int\frac{\mathrm{d}\widehat\nu(v)}{(v-x)^2}\ge 1+{\varpi}\,,
\end{align}
for $N$ sufficiently large.
\end{itemize}
\end{assumption}
\noindent We give two examples for which~\eqref{eq assumption mu_V} is satisfied:
\begin{itemize}
\item [$(1)$] Choosing $\nu=\frac{1}{2}(\mathrm{\delta}_{-a}+\mathrm{\delta}_{a})$, $a\ge 0$, we have $I_{\nu}=[-a,a]$. For $a<1$, one checks that there is a $\varpi=\varpi(a)$ such that~\eqref{eq assumption mu_V} is satisfied and that the deformed semicircle law is supported on a single interval with a square root type behavior at the edges. However, in case $a>1$, the deformed semicircle law is supported on two disjoint intervals. For more details see~\cite{BK1,BK2,BK3}.
\item[$(2)$] Let $\nu$ to be a centered Jacobi measure of the form
\begin{align}\label{Jacobi measure}
\nu(v)=Z^{-1}(v-1)^{\mathrm{a}}(1-v)^{\mathrm{b}} d(v)\lone_{[-1,1]}(v)\,,\end{align}
where $d\in C^{1}([-1,1])$, $d(v)>0$, $-1<\mathrm{a},\mathrm{b}<\infty$, and $Z$ a normalization constant. Then for $\mathrm{a},\mathrm{b}<1$, there is $\varpi>0$ such that~\eqref{eq assumption mu_V} is satisfied with $I_{\nu}=[-1,1]$. However, if $\mathrm{a}>1$ or $\mathrm{b}>1$ then~\eqref{assumption mu_V} may not be satisfied. In this setting the deformed semicircle law is still supported on a single interval, however the square root behavior at the edge may fail. We refer to~\cite{LS,LS2} for a detailed discussion.
\end{itemize}
\begin{lemma}\label{lemma vorbereitung}
Let $\nu$ satisfy~\eqref{eq assumption mu_V} for some $\varpi>0$. Then there are $L_-,L_+\in\ensuremath{\mathbb{R}}$, with $L_-\le -2$, $2\le L_+$, such that $\supp\,\rho_{fc}=[L_-,L_+]$. Moreover, $\rho_{fc}$ has a strictly positive density in $(L_-,L_+)$.
\end{lemma}
Lemma~\ref{lemma vorbereitung} follows directly from Lemma~\ref{lemma mfc} below.
\subsection{Results on bulk universality}
Recall that we denote by $\varrho_{H,n}^N$ the $n$-point correlation function of $H=V+W$, where~$V$ is either a real deterministic or real random diagonal matrix. We denote by $\varrho_{G,n}^N$ the $n$-point correlation function of the GUE, respectively the GOE.
A function $O\,:\,\ensuremath{\mathbb{R}}^{n}\to\ensuremath{\mathbb{R}}$ is called an $n$-particle \emph{observable} if $O$ is symmetric, smooth and compactly supported. Recall from Lemma~\ref{lemma vorbereitung} that we denote by $L_\pm$ the endpoints of the support of the measure $\rho_{fc}$. For deterministic $V$ we have the following result.
\begin{theorem}\label{theorem 1}
Let $W$ be a complex Hermitian or a real symmetric Wigner matrix satisfying the assumptions in Definition~\ref{assumption wigner}. Let $V$ be a deterministic real diagonal matrix satisfying Assumption~\ref{assumption mu_V convergence} and Assumption~\ref{assumption mu_V}. Set $H=V+W$. Let $E,E'$ be two energies satisfying $E\in(L_-,L_+)$, $E'\in(-2,2)$. Fix $n\in\N$ and let $O$ be an $n$-particle observable. Let $\delta>0$ be arbitrary and choose $b\equiv b_N$ such that $N^{-\delta}\ge b_{N}\ge N^{-1+\delta}$. Then,
\begin{align}\label{theorem 1 equation 1}
\lim_{N\to \infty}\int_{{\mathbb R}^n} \mathrm{d}\alpha_1\cdots\mathrm{d}\alpha_n\, O(\alpha_1,\ldots,\alpha_n)\bigg[&\frac{1}{2b}\int_{E-b}^{E+b}\,\frac{\mathrm{d} x}{\rho_{fc}(E)^n}\varrho^N_{H,n}\Big(x+\f{\alpha_1}{\rho_{fc}(E)N},\ldots,x+\f{\alpha_n}{\rho_{fc}(E)N} \Big)\nonumber\\ &-\frac{1}{\rho_{sc}(E')^n}\varrho^N_{G,n}\Big(E'+\f{\alpha_1}{\rho_{sc}(E')N},\ldots,E'+\f{\alpha_n}{\rho_{sc}(E')N} \Big)\bigg] = 0\,,
\end{align}
where $\rho_{fc}$ denotes the density of the deformed semicircle law and $\rho_{sc}$ denotes the density of the standard semicircle law. Here, $\varrho_{G,n}^N$ denotes the $n$-point correlation function of the GUE in case $W$ is a complex Hermitian Wigner matrix, respectively the $n$-point correlation function of the GOE in case $W$ is a real symmetric Wigner matrix.
\end{theorem}
For random $V$ we have the following result.
\begin{theorem}\label{theorem 2}
Let $W$ be a complex Hermitian or a real symmetric Wigner matrix satisfying the assumptions in Definition~\ref{assumption wigner}. Let $V$ be a random real diagonal matrix whose entries are i.i.d.\ random variables that are independent of~$W$ and satisfy Assumption~\ref{assumption mu_V convergence} and Assumption~\ref{assumption mu_V}. Set $H=V+W$. Let $E,E'$ be two energies satisfying $E\in(L_-,L_+)$, $E'\in(-2,2)$. Fix $n\in\N$ and let $O$ be an $n$-particle observable. Let $\delta>0$ be arbitrary and choose $b\equiv b_N$ such that $N^{-\delta}\ge b_{N}\ge N^{-1/2+\delta}$. Then,
\begin{align}\label{theorem 2 equation 1}
\lim_{N\to \infty}\int_{{\mathbb R}^n} \mathrm{d}\alpha_1\cdots\mathrm{d}\alpha_n\, O(\alpha_1,\ldots,\alpha_n)\bigg[&\frac{1}{2b}\int_{E-b}^{E+b}\,\frac{\mathrm{d} x}{\rho_{fc}(E)^n}\varrho^N_{H,n}\Big(x+\f{\alpha_1}{\rho_{fc}(E)N},\ldots,x+\f{\alpha_n}{\rho_{fc}(E)N} \Big)\nonumber\\ &-\frac{1}{\rho_{sc}(E')^n}\varrho^N_{G,n}\Big(E'+\f{\alpha_1}{\rho_{sc}(E')N},\ldots,E'+\f{\alpha_n}{\rho_{sc}(E')N} \Big)\bigg] = 0\,,
\end{align}
where $\rho_{fc}$ denotes the density of the deformed semicircle law and $\rho_{sc}$ denotes the density of the standard semicircle law. Here, $\varrho_{G,n}^N$ denotes the $n$-point correlation function of the GUE in case $W$ is a complex Hermitian Wigner matrix, respectively the $n$-point correlation function of the GOE in case $W$ is a real symmetric Wigner matrix.
\end{theorem}
\begin{remark}
Theorem~\ref{theorem 1} and Theorem~\ref{theorem 2} show that the averaged local correlation functions of $H=V+W$ are universal in the limit of large~$N$ in the sense that they are independent of the diagonal matrix $V$ and also independent of the precise distribution of the entries of $W$. Both theorems hold for real symmetric and complex Hermitian matrices. For the former choice, $\varrho_{G,n}^N$ stands for the $n$-point correlation functions of the GOE. For the latter choice, $\varrho_{G,n}^N$ stands for the $n$-point correlation functions of the GUE.
Note that we can choose $b_N$ of order $N^{-1+\delta}$, $\delta>0$, for deterministic $V$ in Theorem~\ref{theorem 1}, while we have to choose $b_N$ of order $N^{-1/2+\delta}$, $\delta>0$, for random $V$ in Theorem~\ref{theorem 2}. The latter condition is technical and not optimal. It is related to our next comment.
For random $V$ with $(v_i)$ i.i.d.\ bounded random variables, the eigenvalues of $H$ fluctuate on scale $N^{-1/2}$ in the bulk~\cite{LS}. Yet, under the assumptions of Theorem~\ref{theorem 2}, the eigenvalue gaps remain rigid over small scales so that the universality of local correlation functions, a statement about the eigenvalue gaps, is unaffected by these mesoscopic fluctuations. We thus expect Theorem~\ref{theorem 2} to hold with $b_N\gg N^{-1}$. Relying on explicit integration formulas in the complex Hermitian setting, we suppose that the averaging over an energy window can be dropped, c.f., the results for the deformed GUE in~\cite{S1}.
We also remark that the assumption that $(v_i)$ are independent among themselves can be relaxed and our results can be extended to dependent random variables provided that $(v_i)$ satisfy~\eqref{equation assumption mu_V convergence},~\eqref{eq assumption mu_V} and~\eqref{eq assumption mu_V 2} for some constants $\alpha_0,\varpi,\mathfrak{t}>0$, and provided that $(v_i)$ are independent of $(w_{ij})$. In such a setting the required lower bound on~$b_N$ depends on $\alpha_0$.
\end{remark}
\begin{remark} \label{remark for gap universality}
The main ingredient of our proofs of Theorem~\ref{theorem 1} and Theorem~\ref{theorem 2} is an entropy estimate; see Proposition~\ref{the super proposition}. Once such an estimate is obtained, the method in~\cite{single_gap} also implies the single gap universality in the sense that the distribution of any single
gap in the bulk is the same (up to a scaling) as the one from the corresponding Gaussian case. More precisely, fix $\alpha>0$ and let $k\in\N$ be such that $\alpha N\le k\le (1-\alpha)N$. Let $O$ be an $n$-particle observable. Then there are $\chi>0$ and $C$ such that
\begin{align}\label{gap universality for deformed wigner}
&\Big |
\E^{H}
O
\big( (N\rho_{fc,k})(\lambda_k-\lambda_{k+1}),(N\rho_{fc,k})(\lambda_k-\lambda_{k+2}),\ldots,(N\rho_{fc,k})(\lambda_k-\lambda_{k+n})\big)\nonumber\\
&\qquad
- \E^{\mu_{G}}O \big( (N\rho_{sc,k})(\lambda_k-\lambda_{k+1}),(N\rho_{sc,k})(\lambda_k-\lambda_{k+2}),\ldots,(N\rho_{sc,k})(\lambda_k-\lambda_{k+n})\big)
\Big|\leq C N^{-\chi}\,,
\end{align}
for $N$ sufficiently large, where $\mu_G$ is the standard GOE or GUE ensemble, depending
on the symmetry class of $H$. Here $\rho_{fc,k}$ stands for the density of the measure $\rho_{fc}$ at the classical location $\gamma_k$ of the $k$-th eigenvalue defined through
\begin{align}
\int_{-\infty}^{\gamma_k}\rho_{fc}(x)\mathrm{d} x=\frac{k-1/2}{N}\,.
\end{align}
Similarly, $\rho_{sc,k}$ stands for the density of the standard semicircle law $\rho_{sc}$ at the classical location $\gamma_k$ of the $k$-th eigenvalue of the Gaussian ensembles.
\end{remark}
\begin{remark}
In addition to the bulk universality, the edge universality also holds
for our model. More precisely, there are $\varkappa > 0, \chi>0, c_0>0, C>0$ such that the following result holds for any fixed $n\in\N$.
For any $n$-particle observable $O$ and $\Lambda\subset\llbracket 1,N^\varkappa\rrbracket$
with $|\Lambda|= n$, we have
\begin{align}
\left |
\E^{H}
O
\left( \left(c_0N^{ 2/3} j^{1/3}(\lambda_j-\widehat\gamma_j)\right)_{j\in \Lambda}\right)
- \E^{\mu_{G}} O\left(\left(N^{ 2/3} j^{1/3}(\lambda_j-\gamma_j)\right)_{j\in \Lambda}\right)
\right|\leq C N^{-\chi}\,,
\end{align}
for $N$ sufficiently large, where $\mu_G$ is the standard GOE or GUE ensemble, depending
on the symmetry class of~$H$. Here the constant~$c_0$ is a scaling factor so that
the eigenvalue density at the edge of~$H$ can be compared with the Gaussian case. It only depends on~$\lambda$ and~$\nu$. Further,~$\widehat\gamma_j$, $\gamma_j$ denote here the classical locations of the $j$-th eigenvalue with respect to the measure~$\widehat\rho_{fc}$ introduced in~\eqref{stieltjes inversion formula 2} below, respectively with respect to the standard semicircle law $\rho_{sc}$. This result
follows from Proposition~\ref{the super proposition} and~\cite{BEY}. We leave the details to the interested readers.
\end{remark}
\subsection{Notations and Conventions}
In this subsection, we introduce some more notations and conventions used throughout the paper. For high probability estimates we use two parameters $\xi\equiv\xi_N$ and $\varphi\equiv\varphi_N$: We let
\begin{align}\label{eq.xi}
a_0<\xi\le A_0\log\log N\,,\quad\qquad \varphi=(\log N)^{C_1}\,,
\end{align}
for some constants $a_0>2$, $A_0\ge10$, $C_1> 1$.
\begin{definition}
We say an event $\Omega$ has $(\xi,\upsilon)$-high probability, if
\begin{align*}
\mathbb{P}(\Omega^c)\le\e{-\upsilon(\log N)^{\xi}}\,,\qquad\qquad(\upsilon>0)\,,
\end{align*}
for $N$ sufficiently large. We say an event $\Omega$ has $\varsigma$-exponentially high probability, if
\begin{align*}
\mathbb{P}(\Omega^c)\le\e{-N^{\varsigma}}\,,\qquad\qquad(\varsigma>0)\,,
\end{align*}
for $N$ sufficiently large.
Similarly, for a given event $\Omega_0$ we say an event $\Omega$ holds with $(\xi,\upsilon)$-high probability, respectively $\varsigma$-exponentially high probability, on $\Omega_0$, if
\begin{align*}
\mathbb{P}(\Omega^c\cap\Omega_0)\le\e{-\upsilon(\log N)^{\xi}}\,,\qquad\quad (\upsilon>0)\,,\qquad\qquad
\mathbb{P}(\Omega^c\cap\Omega_0)\le\e{-N^{\varsigma}}\,,\qquad\quad (\varsigma>0)\,,
\end{align*}
respectively, for $N$ sufficiently large.
\end{definition}
For brevity, we occasionally say an event holds with exponentially high probability, when we mean $\varsigma$-exponentially high probability. We do not keep track of the explicit value of $\upsilon$ or $\varsigma$ in the following, allowing $\upsilon$ and $\varsigma$ to decrease from line to line such that $\upsilon,\varsigma>0$.
We use the symbols ${\mathcal O}(\,\cdot\,)$ and $o(\,\cdot\,)$ for the standard big-O and little-o notation. The notations ${\mathcal O}\,,\,o$, $\ll$, $\gg$, refer to the limit $N\to \infty$, if not indicated otherwise. Here $a\ll b$ means $a=o(b)$. We use~$c$ and~$C$ to denote positive constants that do not depend on $N$. Their value may change from line to line. We write $a\sim b$, if there is $C\ge1$ such that $C^{-1}|b|\le |a|\le C |b|$, and, occasionally, we write for $N$-dependent quantities $a_N\lesssim b_N$, if there exist constants $C,c>0$ such that $|a_N|\le C(\varphi_N)^{c\xi}|b_N|$.
Finally, we abbreviate
\begin{align*}
\sum_{j}^{(i)}(\,\cdot\,)\equiv \sum_{\substack{j=1\\ j\not=i}}^N(\,\cdot\,)\,\,,
\end{align*}
and we use double brackets to denote index sets, i.e.,
$$
\llbracket n_1, n_2 \rrbracket \mathrel{\mathop:}= [n_1, n_2] \cap \Z\,,
$$
for $n_1, n_2 \in \ensuremath{\mathbb{R}}$.
\section{Local law and rigidity estimates}\label{section local law}
Recall the constant $\varpi>0$ in Assumption~\ref{assumption mu_V}. Set $\varpi'\mathrel{\mathop:}=\varpi/10$. In this section we consider the family of interpolating random matrices
\begin{align}\label{definition Theta}
H^{\vartheta}\mathrel{\mathop:}=\vartheta V+W\,,\qquad\qquad\vartheta\in\Theta_{\varpi}\mathrel{\mathop:}=[0,1+\varpi']\,,
\end{align}
where $V$ and $W$ are chosen to satisfy Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}, respectively the assumptions in Definition~\ref{assumption wigner}. Here~$\vartheta$ has the interpretation of a possibly $N$-dependent positive ``coupling parameter''.
We define the resolvent or \emph{Green function}, $G^{\vartheta}(z)$, and the averaged Green function, $m^{\vartheta}(z)$, of $H^{\vartheta}$ by
\begin{align}\label{definition of the m}
G^{\vartheta}(z)=(G^{\vartheta}_{ij}(z))\mathrel{\mathop:}= \frac{1}{\vartheta V+W-z}\,,\qquad\quad m_N^{\vartheta}(z)\mathrel{\mathop:}=\frac{1}{N}\Tr G^{\vartheta}(z)\,,\qquad\qquad (z\in\mathbb{C}^{+})\,.
\end{align}
Frequently, we abbreviate $G^{\vartheta}\equiv G^{\vartheta}(z)$, $m_N^{\vartheta}\equiv m_N^{\vartheta}(z)$, etc..
To conveniently cope with the cases when $(v_i)$ are random, respectively deterministic, we introduce an event~$\Omega_V$ on which the random variables $(v_i)$ exhibit ``typical'' behavior. Recall that we denote by~$m_{\widehat\nu}$ and~$m_{\nu}$ the Stieltjes transforms of $\widehat\nu$, respectively $\nu$.
\begin{definition}\label{definition of the event omega}
Let $\Omega_V\equiv\Omega_V(N)$ be an event such that the following holds on it:
\begin{itemize}
\item[$(1)$]There is a constant $\alpha_0>0$ such that, for any compact set ${\mathcal D}\subset\ensuremath{\mathbb{C}}^+$ with $\mathrm{dist}({\mathcal D},\supp \nu)>0$, there is~$C$ such that
\begin{align}\label{definition of omega_V}
\left|m_{\widehat\nu}(z)-m_{\nu}(z) \right|\le CN^{-\alpha_0}\,,
\end{align}
for $N$ sufficiently large.
\item[(2)] Recall the constant $\varpi>0$ in Assumption~\ref{assumption mu_V}. We have
\begin{align}\label{definition of omega_V III}
\inf_{x\in I_{\widehat\nu}}\int\frac{\mathrm{d}\widehat\nu(v)}{(v-x)^2}\ge 1+\varpi\,,\qquad\qquad \inf_{x\in I_{\nu}}\int\frac{\mathrm{d}\nu}{(v-x)^2}\ge 1+\varpi\,,
\end{align}
for $N$ sufficiently large.
\end{itemize}
\end{definition}
In case $(v_i)$ are deterministic, $\Omega_V$ has full probability for $N$ sufficiently large by Assumptions~\ref{assumption mu_V convergence}.
Similar to the definition of $m_{fc}$, we define $m_{fc}^{\vartheta}$ and $\widehat{m}^{\vartheta}_{fc}$ as the solutions to the equations
\begin{align}\label{lambda mfc}
m^{\vartheta}_{fc}(z)=\int\frac{\mathrm{d}\nu(v)}{\vartheta v-z-m_{fc}^{\vartheta}(z)}\,,\qquad\quad\im m^{\vartheta}_{fc}(z)\ge 0\,,\qquad\qquad (z\in \ensuremath{\mathbb{C}}^+)\,,
\end{align}
and
\begin{align}\label{hat mfc}
\widehat{m}^{\vartheta}_{fc}(z)=\int\frac{\mathrm{d}\widehat\nu(v)}{\vartheta v-z-\widehat{m}^{\vartheta}_{fc}(z)}\,,\quad\qquad \im \widehat m^{\vartheta}_{fc}(z)\ge 0\,,\qquad\qquad (z\in\ensuremath{\mathbb{C}}^+)\,,
\end{align}
respectively. Following the discussion of Subsection~\ref{subsection deforemd semicirle law}, $m^{\vartheta}_{fc}$ and $\widehat{m}^{\vartheta}_{fc}$ define two probability measures $\rho^{\vartheta}_{fc}$ and $\widehat\rho^{\vartheta}_{fc}$ through the densities
\begin{align}\label{stieltjes inversion formula 2}
\rho^{\vartheta}_{fc}(E)\mathrel{\mathop:}=\lim_{\eta\searrow 0}\frac{1}{\pi}\im m^{\vartheta}_{fc}(E+\mathrm{i}\eta)\,,\qquad\quad \widehat\rho^{\vartheta}_{fc}(E)\mathrel{\mathop:}=\lim_{\eta\searrow 0}\frac{1}{\pi}\im \widehat m^{\vartheta}_{fc}(E+\mathrm{i}\eta)\,,\qquad\qquad (E\in\ensuremath{\mathbb{R}})\,,
\end{align}
c.f.,~\eqref{stieltjes inversion formula}. More precisely, we have the following result which follows directly from the proof of Lemma~\ref{lemma mfc} and of Lemma~\ref{lemma hatmfc}. Recall the definition of $\Theta_{\varpi}$ in~\eqref{definition Theta}.
\begin{lemma}\label{definition hatmufc}
Let $\widehat\nu$ and $\nu$ satisfy the Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}. Then, for any $\vartheta\in\Theta_{\varpi}$ and $N\in\N$, Equations~\eqref{lambda mfc} and~\eqref{hat mfc} define through the inversion formulas in~\eqref{stieltjes inversion formula 2} absolutely continuous measures $\rho^{\vartheta}_{fc}$ and $\widehat\rho^{\vartheta}_{fc}$. Moreover, the measure~$\rho_{fc}^{\vartheta}$ is supported on a single interval with strictly positive density inside this interval. The same holds true on~$\Omega_V$ for the measures $\widehat\rho_{fc}^{\vartheta}$, for $N$ sufficiently large.
\end{lemma}
Note that if $(v_i)$ are random then so are $\widehat{m}_{fc}^{\vartheta}$, respectively~$\widehat{\rho}_{fc}^{\vartheta}$. We use the symbol $\,\,\widehat{\,}\,\,$ to denote quantities that depend on the empirical distribution $\widehat\nu$ of the $(v_i)$ while we drop this symbol for quantities depending on the limiting distribution $\nu$ of $(v_i)$.
We denote by $\widehat L_{\pm}^{\vartheta}$, respectively $L_{\pm}^{\vartheta}$, the endpoints of the support of $\widehat\rho_{fc}^{\vartheta}$, respectively~$\rho_{fc}^{\vartheta}$. Let $E_0\ge 1+\max\{|L_-^1|,L_+^1\}$ and define the domain
\begin{align}\label{eq.DL}
{\mathcal D}_L\mathrel{\mathop:}=\{z=E+\mathrm{i} \eta\in\ensuremath{\mathbb{C}}\,:\,| E|\le E_0\,, (\varphi_N)^{L}\le N\eta\le 3N\}\,,
\end{align}
with $L\equiv L(N)$, such that $L\ge 12\xi$; see~\eqref{eq.xi}.
The following theorem is the main result of this section.
\begin{theorem}\emph{[Strong local deformed semicircle law]}\label{thm.strong}
Let $H^\vartheta=\vartheta V+W$, $\vartheta\in\Theta_{\varpi}$ (see~\eqref{definition Theta}), where $W$ is a real symmetric or complex Hermitian Wigner matrix satisfying the assumptions in Definition~\ref{assumption wigner} and~$V$ is a deterministic or random real diagonal matrix satisfying Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}. Let
\begin{align}\label{eq: strong xi}
\xi=\frac{A_0+o(1)}{2}\log\log N\,.
\end{align}
Then there are constants $\upsilon>0$ and $c_1$, depending on the constants $E_0$ in~\eqref{eq.DL}, $\alpha_0$ in~\eqref{definition of omega_V}, $A_0$, $a_0$, $C_1$ in~\eqref{eq.xi}, $\theta$, $C_0$ in~\eqref{eq.C0} and the measure~$\widehat\nu$ such that the following holds for $L\ge 40\xi$. For any $z\in{\mathcal D}_L$ and any $\vartheta\in\Theta_{\varpi}$, we have
\begin{align}\label{strong1}
\left| m_N^{\vartheta}(z)-\widehat{m}^{\vartheta}_{fc}(z) \right| &\le (\varphi_N)^{c_1\xi}\frac{1}{N\eta}\,,
\end{align}
with $(\xi,\upsilon)$-high probability on $\Omega_V$.
Moreover, we have, for any $z\in{\mathcal D}_L$ and any $\vartheta\in\Theta_{\varpi}$,
\begin{align}\label{strong2}
\left| G_{ij}^{\vartheta}(z)-\delta_{ij} \widehat g_{i}^{\vartheta}(z) \right|\le (\varphi_N)^{c_1\xi}\left(\sqrt{\frac{\im \widehat m^{\vartheta}_{fc}(z)}{N\eta}}+\frac{1}{N\eta}\right)\,,\qquad\qquad (i,j\in\llbracket 1,N\rrbracket)\,,
\end{align}
with $(\xi,\upsilon)$-high probability on $\Omega_V$, where we have set
\begin{align}
\widehat{g}^{\vartheta}_i(z)\mathrel{\mathop:}=\frac{1}{\vartheta v_i-z-\widehat{m}^{\vartheta}_{fc}(z)}\,.
\end{align}
\end{theorem}
The study of local laws for Wigner matrices was initiated in~\cite{ESY1,ESY2,ESY3}. For more recent results, we refer to~\cite{EKYY4}. For deformed Wigner matrices with random potential a local law was obtained in~\cite{LS}.
Denote by $\boldsymbol{\lambda}^{\vartheta}=(\lambda_1^{\vartheta},\lambda_2^{\vartheta},\ldots,\lambda_N^{\vartheta})$ the eigenvalues of the random matrix $H^{\vartheta}=\vartheta V+W$ arranged in ascending order. We define the classical location, $\widehat\gamma_{i}^{\vartheta}$, of the eigenvalue $\lambda_i^\vartheta$ by
\begin{align}\label{classical location}
\int_{-\infty}^{\widehat\gamma_i^{\vartheta}}\widehat{\rho}_{fc}^{\vartheta}(x)\mathrm{d} x=\frac{ i-\frac{1}{2}}{N}\,,\qquad\qquad(1\le i\le N)\,.
\end{align}
Note that $(\widehat\gamma_i^{\vartheta})$ are random in case $(v_i)$ are. We have the following rigidity result on the eigenvalue locations of $H^{\vartheta}$:
\begin{corollary}\label{rigidity of eigenvalues}
Let $H^{\vartheta}=\vartheta V+W$, $\vartheta\in\Theta_{\varpi}$, where $W$ is a real symmetric or complex Hermitian Wigner matrix satisfying the assumptions in Definition~\ref{assumption wigner} and $V$ is a deterministic or random real diagonal matrix satisfying Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}. Let $\xi$ satisfy~\eqref{eq: strong xi}. Then there are constants $\upsilon>0$ and $c_1,c_2$, depending on the constants $E_0$ in~\eqref{eq.DL}, $\alpha_0$ in~\eqref{definition of omega_V}, $A_0$, $a_0$, $C_1$ in~\eqref{eq.xi}, $\theta$, $C_0$ in~\eqref{eq.C0} and the measure~$\widehat\nu$, such that
\begin{align}
|\lambda_{i}^{\vartheta}-\widehat\gamma_{i}^{\vartheta}|&\le (\varphi_N)^{c_1\xi}\frac{1}{N^{2/3}\widehat\alpha_i^{1/3} }\,,\qquad\qquad(1\le i\le N)\,,\\
\sum_{i=1}^N|\lambda_{i}^{\vartheta}-\widehat\gamma_{i}^{\vartheta}|^2&\le (\varphi_N)^{c_2\xi}\frac{1}{N}\,,\label{rigidity of eigenvalues equation}
\end{align}
with $(\xi,\upsilon)$-high probability on $\Omega_V$, for all $\vartheta\in\Theta_{\varpi}$, where we have abbreviated $\widehat\alpha_i\mathrel{\mathop:}=\min\{i,N-i\}$.
\end{corollary}
In the rest of this section we sum up the proofs of Theorem~\ref{thm.strong} and Corollary~\ref{rigidity of eigenvalues}.
\subsection{Properties of $m^{\vartheta}_{fc}$ and $\widehat{m}^{\vartheta}_{fc}$}
In this subsection, we discuss properties of the Stieltjes transforms $m_{fc}^{\vartheta}$ and $\widehat{m}^{\vartheta}_{fc}$. We first derive the desired properties for $m_{fc}^{\vartheta}$ (Lemma~\ref{lemma mfc} and Corollary~\ref{remark about real and imaginary part of mfc} in the Appendix) and then show in a second step that ${m}_{fc}^{\vartheta}$ is a good approximation to $\widehat m_{fc}^{\vartheta}$ so that $\widehat{m}_{fc}^{\vartheta}$ also shares these properties; see Lemma~\ref{lemma hatmfc}.
For $E_0$ as in~\eqref{definition of D'}, we define the domain, ${\mathcal D}'$, of the spectral parameter $z$ by
\begin{align}\label{definition of D'}
{\mathcal D}'\mathrel{\mathop:}=\{z=E+\mathrm{i}\eta\,:\, E\in[-E_0,E_0]\,,\,\eta\in(0,3]\}\,.
\end{align}
The next lemma, whose proof is postponed to the appendix, gives a qualitative description of the deformed semicircle law $\rho^{\vartheta}_{fc}$ and its Stieltjes transform $m^{\vartheta}_{fc}$.
\begin{lemma}\label{lemma mfc}
Let $\nu$ satisfy Assumption~\ref{assumption mu_V}, for some $\varpi>0$. Then the following holds true for any $\vartheta\in\Theta_{\varpi}$. There are $L_-^{\vartheta},L_+^{\vartheta}\in\ensuremath{\mathbb{R}}$, with $L_-^{\vartheta}<0<L_+^{\vartheta}$, such that $\supp \rho_{fc}^{\vartheta}=[L_-^{\vartheta},L_+^{\vartheta}]$ and there exists a constant $C>1$ such that, for all $\vartheta\in\Theta_{\varpi}$,
\begin{align}\label{the square root}
C^{-1}\sqrt{\kappa_E}\le \rho_{fc}^{\vartheta}(E)\le C \sqrt{\kappa_E}\,,\qquad \quad(E\in[L_-^{\vartheta},L_+^{\vartheta}])\,,
\end{align}
where $\kappa_E$ denotes the distance of $E$ to the endpoints of the support of $\rho_{fc}^{\vartheta}$, i.e.,
\begin{align}
\kappa_E\mathrel{\mathop:}=\min\{|E-L_-^{\vartheta}|,\,|E-L_+^{\vartheta}|\}\,.
\end{align}
The Stieltjes transform, $m^{\vartheta}_{fc}$, of $\rho^{\vartheta}_{fc}$ has the following properties,
\begin{itemize}
\item[$(1)$] for all $z=E+\mathrm{i}\eta\in{\mathcal D}'$,
\begin{align}\label{behavior of mfc}
\im m^{\vartheta}_{fc}(z)\sim\begin{cases} \sqrt{\kappa+\eta}\,,\quad& E\in[L_-^{\vartheta},L_+^{\vartheta}]^{\phantom{c}}\,,\\
\frac{\eta}{\sqrt{\kappa+\eta}}\,,\quad &E\in[L_-^{\vartheta},L_+^{\vartheta}]^c\,;
\end{cases}
\end{align}
\item[$(2)$] there exists a constant $C>1$ such that for all $z\in{\mathcal D}'$ and all $x\in I_{\nu}$,
\begin{align}\label{stability bound}
C^{-1}\le |\vartheta x -z-m^{\vartheta}_{fc}(z)|\le C\,.
\end{align}
\end{itemize}
Moreover, the constants in~\eqref{the square root},~\eqref{behavior of mfc} and~\eqref{stability bound} can be chosen uniformly in $\vartheta\in\Theta_{\varpi}$.
\end{lemma}
Next, we argue that $\widehat m_{fc}^{\vartheta}$ behaves qualitatively in the same way as $m_{fc}^{\vartheta}$ on $\Omega_V$ for $N$ sufficiently large. Lemma~\ref{lemma hatmfc} below is proven in the Appendix.
\begin{lemma}\label{lemma hatmfc}
Let $\widehat\nu$ satisfy Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}, for some $\varpi>0$. Then the following holds for all $\vartheta\in\Theta_{\varpi}$ and all sufficiently large $N$ on $\Omega_V$. There are $\widehat L_-^{\vartheta}, \widehat L_+^{\vartheta}\in\ensuremath{\mathbb{R}}$, with $\widehat L_-^{\vartheta}<0<\widehat L_+^{\vartheta}$, such that $\supp\,\widehat\rho_{fc}^{\vartheta}=[\widehat L_-^{\vartheta},\widehat L_+^\vartheta]$. Let $\widehat\kappa_E\mathrel{\mathop:}=\min\{|E-\widehat{L}_-^{\vartheta}|, |E-\widehat{L}_+^{\vartheta}|\}$, then~\eqref{the square root},~\eqref{behavior of mfc} and~\eqref{stability bound} of Lemma~\ref{lemma mfc}, hold true on $\Omega_V$, for $N$ sufficiently large, with $m_{fc}^{\vartheta}$ replaced by $\widehat{m}^{\vartheta}_{fc}$, $\rho_{fc}^\vartheta$ replaced by $\widehat\rho_{fc}^{\vartheta}$, etc.. Moreover, the constants in these inequalities can be chosen uniformly in $\vartheta\in\Theta_{\varpi}$ and $N$, for $N$
sufficiently large.
Further, there is $c>0$ such that for all $z\in{\mathcal D}'$ we have
\begin{align}\label{mini local law}
|\widehat{m}^{\vartheta}_{fc}(z)-m_{fc}^{\vartheta}(z)|\le N^{-c\alpha_0/2}\,,\qquad\qquad |\widehat L_\pm^{\vartheta}-{L}_\pm^{\vartheta}|\le N^{-c\alpha_0}\,,
\end{align}
on $\Omega_V$ for $N$ sufficiently large and all $\vartheta\in\Theta_{\varpi}$.
\end{lemma}
\subsection{Proof of Theorem~\ref{thm.strong} and Corollary~\ref{rigidity of eigenvalues}}
The proof of Theorem~\ref{thm.strong} follows closely the proof of Theorem~2.10 in~\cite{LS}. The difference between Theorem~\ref{thm.strong} of the present paper and Theorem~2.10 in~\cite{LS} is that we presently condition on the diagonal entries $(v_i)$, i.e., we consider the entries of $V$ as fixed. Accordingly, we compare (on the event $\Omega_V$ of typical $(v_i)$) the averaged Green function $m^{\vartheta}$ with $\widehat m^{\vartheta}_{fc}$ (see~\eqref{hat mfc}) instead of $m_{fc}^{\vartheta}$ (see~\eqref{lambda mfc}). For consistency, we momentarily drop the $\vartheta$ dependence form our notation. To establish Theorem~\ref{thm.strong}, we first derive a weak local deformed semicircle law (see Theorem~4.1 in~\cite{LS}) by following the proof in~\cite{LS}. Using the Lemma~\ref{lemma mfc}, Lemma~\ref{lemma hatmfc} and the results in the Appendix, it is then straightforward to obtain the following result.
\begin{lemma}\label{weak local law}
Under the assumption of Theorem~\ref{thm.strong}, there are $c_1$ and $\upsilon>0$ such that
\begin{align*}
|m_N(z)-\widehat m_{fc}(z)|\le (\varphi_N)^{c_1\xi}\frac{1}{(N\eta)^{1/3}}\,,\qquad\qquad |G_{ij}(z)|\le (\varphi_N)^{c_1\xi}\frac{1}{\sqrt{N\eta}}\,,
\end{align*}
with $(\xi,\upsilon)$-high probability on $\Omega_V$, uniformly in $z\in{\mathcal D}_L$ and $\vartheta\in \Theta_{\varpi}$.
\end{lemma}
To prove~Theorem~\ref{thm.strong} we follow mutatis mutandis the proof of Theorem~4.1 in~\cite{LS}. But we note that in the corresponding equation to~(5.25) in~\cite{LS}, we may set $\lambda=0$ in the error term, at the cost of replacing $m_{fc}$ by $\widehat m_{fc}$. In the subsequent analysis, we can simply set $\lambda=0$ in the error terms. In this way, one establishes the proof of Theorem~\ref{thm.strong}. Similarly, Corollary~\ref{rigidity of eigenvalues} can be proven in the same way as is Theorem~2.21 in~\cite{LS}. It suffices to set $\lambda=0$ in the analysis in~\cite{LS}. We leave the details aside.
\section{Reference $\beta$-ensemble}\label{section beta ensemble}
\subsection{Definition of $\beta$-ensemble and known results}\label{subsection beta ensemble intro}
We first recall the notion of $\beta$-ensembles. Let $N\in\N$ and let $\digamma^{(N)}\subset\ensuremath{\mathbb{R}}^N$ denote the set
\begin{align}\label{Weyl chamber}
\digamma^{(N )}\mathrel{\mathop:}=\{ \mathbf{x} = (x_1,x_2 ,\ldots , x_N ) \,:\, x_1\le x_2\le\ldots\le x_N\}\,.
\end{align}
Consider the probability distribution on $\digamma^{(N)}$ given by
\begin{align}\label{eqn:measure}
\mu_{U}^{N} (\mathrm{d}\mathbf{x})\equiv \mu_U (\mathrm{d} \mathbf{x}) \mathrel{\mathop:}=\frac{1}{Z_{U}^{N}}\e{-\beta N {\mathcal H}(\mathbf{x}) }\mathrm{d}\mathbf{x}\,,\qquad \mathrm{d}\mathbf{x}\mathrel{\mathop:}=\lone(\mathbf{x}\in\digamma^{(N)})\mathrm{d} x_1\mathrm{d} x_2\cdots\mathrm{d} x_N\,,
\end{align}
where $\beta>0$,
\begin{align*}
{\mathcal H}(\mathbf{x}) \mathrel{\mathop:}=\sum_{i=1}^N\frac{1}{2}\left(U(x_i)+\frac{x_i^2}{2}\right)-\frac{1}{N}\sum_{1\le i<j\le N}\log (x_j-x_i)
\end{align*}
and $Z_{U}^{N}\equiv Z_{U}^N(\beta)$ is a normalization. Here $U$ is a potential, i.e., a real-valued, sufficiently regular function on $\ensuremath{\mathbb{R}}$. In the following, we often omit the parameters $N$ and $\beta$ from the notation. We use $\mathbb{P}^{\mu_U}$ and $\E^{\mu_U}$ to denote the probability and the expectation with respect to $\mu_U$.
We view $\mu_U$ as a Gibbs measure of~$N$ particles on~$\ensuremath{\mathbb{R}}$ with a logarithmic interaction, where the parameter
$\beta > 0$ may be interpreted as the inverse temperature. (For the results in the present paper, we choose $\beta=2$ in case $W$ is complex hermitian Wigner matrix, or $\beta=1$ in case $W$ is a real symmetric Wigner matrix.) We refer to the variables $(x_i)$ as particles or points and we call the system a log-gas or a $\beta$-ensemble. We assume that the potential $U$ is a $C^4$
function on $\ensuremath{\mathbb{R}}$ such that its second derivative is bounded below, i.e., we have
\begin{align}\label{assumption 1 for beta universality}
\inf_{x\in\ensuremath{\mathbb{R}}} U'' (x)\ge -2C_U\,,
\end{align}
for some constant $C_U\ge 0$, and we further assume that
\begin{align}\label{assumption 2 for beta universality}
U (x)+\frac{x^2}{2} > (2 + \epsilon) \log(1 + |x|)\,, \qquad\qquad(x\in\ensuremath{\mathbb{R}})\,,
\end{align}
for some $\epsilon > 0$, for large enough $|x|$. It is well-known, see, e.g.,~\cite{BPS}, that under these conditions
the measure is normalizable, $Z_{U}^{N} < \infty$. Moreover, the averaged density of the empirical spectral measure, $\rho_U^N$,
defined as
\begin{align}\label{reference section 7 1}
\rho_{U}^{N}\mathrel{\mathop:}= \E^{\mu_{U}}\frac{1}{N}\sum_{i=1}^N\delta_{x_i}\,,
\end{align}
converges weakly in the limit $N\to\infty$ to a continuous function $\rho_U$, the equilibrium density, with compact support. It is
well-known that~$\rho_U$ can be obtained as the unique solution to the variational problem
\begin{align}\label{minimization problem}
\inf\left\{\int_{\ensuremath{\mathbb{R}}}\left(\frac{x^2}{2}+U(x)\right)\mathrm{d}\rho(x)-\int_{\ensuremath{\mathbb{R}}}\log|x-y|\,\mathrm{d}\rho(x)\mathrm{d}\rho(y)\,:\,\rho\textrm{ is a probability measure } \right\}
\end{align}
and that the equilibrium density $\rho=\rho_U$ satisfies
\begin{align}\label{regular potential U}
U'(x)+x=-2\Xint-_\ensuremath{\mathbb{R}}\frac{\rho(y)\mathrm{d} y}{y-x}\,,\qquad\quad (x\in\supp\rho_U)\,.
\end{align}
In fact,~\eqref{regular potential U} holds if and only if $x\in\supp\rho_U$. We will assume in addition that the minimizer $\rho_U$ is supported on a single interval $[A_-,A_+]$, and that $U$ is ``regular''
in the sense of~\cite{KML}, i.e., the equilibrium density of $U$ is positive on $(A_-,A_+)$ and vanishes
like a square root at each of the endpoints of $[A_-, A_+]$. Viewing the points $\mathbf{x}=(x_i)$ as points or particles on $\ensuremath{\mathbb{R}}$, we define the classical location of the $k$-particle, $\gamma_k$, under the $\beta$-ensemble $\mu_U$ by
\begin{align}\label{general beta ensemble classical location}
\int_{-\infty}^{\gamma_k}\rho_U(x)\mathrm{d} x=\frac{k-\frac{1}{2}}{N}\,.
\end{align}
For a detailed discussion of general $\beta$-ensemble we refer, e.g., to~\cite{AGZ,BEY}.
For $U\equiv 0$, we write $\mu_G$ instead of $\mu_0$, since $\mu_0$ is the equilibrium measure for the GUE ($\beta=2$), respectively the GOE ($\beta=1$). More precisely, setting
\begin{align}\label{gaussian hamiltonian}
{\mathcal H}_G(\mathbf{x}) \mathrel{\mathop:}=\sum_{i=1}^N\frac{1}{4}x_i^2-\frac{1}{N}\sum_{1\le i<j\le N}\log (x_j-x_i)\,,
\end{align}
the GUE, respectively GOE, distribution on $\digamma^{(N)}$ are given by
\begin{align}\label{gaussian measures}
\mu_{G}(\mathrm{d}\mathbf{x})=\frac{1}{Z_{G}^N}\e{-\beta N {\mathcal H}_G(\mathbf{x})}\mathrm{d}\mathbf{x}\,,
\end{align}
where $Z_{G}^N\equiv Z_G^N(\beta)$ is a normalization, and we either choose $\beta=2$ or $\beta=1$.
We are interested in the $n$-point correlation functions defined by
\begin{align}\label{eqn:corrFunct}
\varrho^{N}_{U,n}(x_1,\dots,x_n)=
\int_{\ensuremath{\mathbb{R}}^{N-n}} \mu_U^\#(\mathbf{x})\mathrm{d} x_{n+1}\cdots\mathrm{d} x_{N},
\end{align}
where $ \mu_U^\#$ is the symmetrized version of $\mu_U$ given in~\eqref{eqn:measure}
but defined on $\ensuremath{\mathbb{R}}^N$ instead of the simplex~$\digamma^{(N)}$:
\begin{align}\label{symmetrization}
\mu_U^{\# }(\mathrm{d} x)=\frac{1}{N!}\mu_U(\mathrm{d}\mathbf{x}^{(\sigma)})\,,
\end{align}
where
$\mathbf{x}^{(\sigma)}=(x_{\sigma(1)},\dots,x_{\sigma(N)})$, with
$x_{\sigma(1)}<\dots<x_{\sigma(N)}$, and $\mathrm{d} x= \mathrm{d} x_1\cdots\mathrm{d} x_N$.
The following universality result is proven in~\cite{BEY}.
\begin{theorem}{\emph{[Bulk universality for $\beta$-ensembles, Theorem~2.1 in~\cite{BEY}]}}\label{thm: bulk universality beta}
Let $U$ be a ${C}^4$ regular potential with equilibrium density supported on a single interval $[A_-,A_+]$ that satisfies~\eqref{assumption 1 for beta universality} and~\eqref{assumption 2 for beta universality}. Then the following result holds.
For any fixed $\beta>0$, $E\in (A_-,A_+)$, $|E'|<2$, $ n\in \N$, $0<\delta\le \frac{1}{2}$ and any $n$-particle observable $O$
we have with $b\mathrel{\mathop:}= N^{-1+\delta}$,
\begin{align*}
\lim_{N\to\infty}\int_{\ensuremath{\mathbb{R}}^n} \mathrm{d} \alpha_1 \cdots \mathrm{d} \alpha_n\, O(\alpha_1,
\dots, \alpha_n)& \Bigg [
\int_{E - b}^{E + b} \frac{\mathrm{d} x}{2 b} \frac{1}{ \rho_U (E)^n } \varrho_{U,n}^{N} \left ( x +
\frac{\alpha_1}{N\rho_U(E)}, \dots, x + \frac{\alpha_n}{N\rho_U(E)} \right ) \\
&
\qquad\qquad- \frac{1}{\rho_{sc}(E')^n}
\varrho_{{G}, n}^{N}
\left ( E' +
\frac{\alpha_1}{N\rho_{sc}(E')}, \dots, E' + \frac{\alpha_n}{N\rho_{sc}(E')} \right )\Bigg]=0\,.
\end{align*}
Here $\rho_{sc}(E)=\frac{1}{2\pi}\sqrt{4-E^2}$ is Wigner's semicircle law
and $(\varrho_{{G}, n}^{N})$ are the correlation functions of the Gaussian $\beta$-ensemble, i.e.,
with $U\equiv 0$.
\end{theorem}
Theorem~\ref{thm: bulk universality beta} was first proved in~\cite{BEYI} under the assumption that $U$ is analytic,
a hypothesis that was only required for proving rigidity. The analyticity assumption has been removed in~\cite{BEY}. Recently, alternative proofs of bulk universality for $\beta$-ensembles with general $\beta>0$, i.e., results similar to Theorem~\ref{thm: bulk universality beta}, have been obtained in~\cite{ShMa} and~\cite{BAG}. In the present paper, we will not use Theorem~\ref{thm: bulk universality beta}. It is stated here for completeness.
To conclude this subsection, we recall an important tool in the study of $\beta$-ensembles, the ``first order loop'' equation. In the notation above it reads (in the limit $N\to\infty$)
\begin{align}\label{loop equation}
m_U(z)^2=\int\frac{x+ U'(x)}{x-z}\,\rho_U(x)\,\mathrm{d} x\,,\qquad\quad (z\in\ensuremath{\mathbb{C}}^+)\,,
\end{align}
where $m_U$ denotes the Stieltjes transform of the equilibrium measure $\rho_U$, i.e.,
\begin{align*}
m_U(z)=\int\frac{\rho_U(x)}{x-z}\,\mathrm{d} x\,,\qquad\quad (z\in\ensuremath{\mathbb{C}}^+)\,.
\end{align*}
The loop equation~\eqref{loop equation} can be obtained by a change of variables in~\eqref{eqn:measure} (see~\cite{J3}) or
by integration by parts (see~\cite{ShMa1}).
\subsection{Time-dependent modified $\beta$-ensemble}
In this subsection, we introduce a modified $\beta$-ensemble by specifying potentials $\widehat U$ and $ U$ that depend, among other things, on a parameter $t\ge 0$ which has the interpretation of a time. The potential $\widehat U$ also depends on $N$, the size of our original matrix $H=V+W$. Yet, the $N$ dependence in only through the fixed random variables $(v_i)$. Recall that we have defined $\widehat m_{fc}^{\vartheta}$, respectively $ m_{fc}^{\vartheta}$, as the solutions to the equations
\begin{align}\label{nochmals mfc}
\widehat{m}^{\vartheta}_{fc}(z)=\frac{1}{N}\sum_{i=1}^N\frac{1}{\vartheta v_i-z-\widehat{m}^{\vartheta}_{fc}(z)}\,,\qquad\quad m^{\vartheta}_{fc}(z)=\int\frac{\mathrm{d}\nu(v)}{\vartheta v-z-m_{fc}^{\vartheta}(z)} \,,\qquad \qquad(z\in\ensuremath{\mathbb{C}}^+)\,,
\end{align}
subject to the conditions $\im \widehat m^{\vartheta}_{fc}(z),\, \im m^{\vartheta}_{fc}(z)\ge 0$, for $\im z>0$. Recall from~\eqref{definition Theta} that we denote $\Theta_{\varpi}=[0,1+\varpi']$, $\varpi'=\varpi/10$. We then fix some $ t_0\ge 0$ such that $\e{t_0/2}\in\Theta_{\varpi}$ and let
\begin{align}\label{equation for vartheta}
\vartheta\equiv \vartheta(t)\mathrel{\mathop:}=\e{-(t-t_0)/2}\,,\qquad\quad (t\ge 0)\,.
\end{align}
In the following we consider $t\ge 0$ as time and we henceforth abbreviate $m^{\vartheta(t)}_{fc}(z)\equiv m_{fc}(t,z)$, etc.. Equation~\eqref{nochmals mfc} defines time-dependent measures $\widehat\rho_{fc}(t)$, $\rho_{fc}(t)$ respectively, whose densities at the point $x\in\ensuremath{\mathbb{R}}$ are denoted by $\widehat\rho_{fc}(t,x)$, respectively $\rho_{fc}(t,x)$.
We denote by $\widehat U'(t,x)$, $\widehat U^{(n)}(t,x)$ the first, respectively the $n$-th derivative of $\widehat U(t,x)$ with respect to $x$, and we use the same notation for $U$. We define $\widehat U$ and $U$ (up to finite additive constants that enter the formalism only in normalizations) through their derivatives $\widehat U'$ and $U'$. For $t\ge0$, we set
\begin{align}\label{definition widehat U}
\widehat U'(t,x)+x\mathrel{\mathop:}= -2\Xint-_\ensuremath{\mathbb{R}}\frac{\widehat\rho_{fc}(t,y)}{y-x}\,\mathrm{d} y\,,\qquad \quad U'(t,x)+x\mathrel{\mathop:}= -2\Xint-_\ensuremath{\mathbb{R}}\frac{\rho_{fc}(t,y)}{y-x}\,\mathrm{d} y\,,
\end{align}
for $x\in\supp \widehat\rho_{fc}(t)$, respectively $x\in\supp\rho_{fc}(t)$. Outside the support of the measures $\widehat\rho_{fc}(t)$ and $\rho_{fc}(t)$, we define $\widehat U'$ and $U'$ as $C^3$ extensions such that they are ``regular'' potentials satisfying~\eqref{assumption 1 for beta universality} and~\eqref{assumption 2 for beta universality} for all $t\ge 0$. The definitions of such potentials are obviously not unique. One possible construction is outlined in the Appendix in the form of the proof of the next lemma.
\begin{lemma}\label{superlemma}
There exists potentials $\widehat U, U\,:\,\ensuremath{\mathbb{R}}^+\times \ensuremath{\mathbb{R}}\to\ensuremath{\mathbb{R}}$, $(t,x)\mapsto \widehat U(t,x),\,U(t,x)$ such that for $n\in\llbracket 1,4\rrbracket$, $\widehat U^{(n)}(t,x)$, $U^{(n)}(t,x)$, $\partial_t \widehat U^{(n)}(t,x)$, $\partial_t U^{(n)}(t,x)$ are continuous functions of $x\in\ensuremath{\mathbb{R}}$ and $t\in\ensuremath{\mathbb{R}}^+$, which can be uniformly bounded in~$ x$ on compact sets, uniformly in $t\in\ensuremath{\mathbb{R}}^+$ and sufficiently large $N$. Moreover the following holds for all $t\ge0$ on $\Omega_V$ for $N$ sufficiently large:
\begin{itemize}
\item[$(1)$] $\widehat U'(t,x)$ and $ U'(t,x)$ satisfy~\eqref{definition widehat U} for $x\in\supp\widehat\rho_{fc}(t)$, respectively $x\in\supp\rho_{fc}(t)$. For $x\not\in\supp\widehat\rho_{fc}(t)$, respectively $x\not\in\supp\rho_{fc}(t)$ we have
\begin{align}
| \widehat U'(t,x)+x|> 2|\re \widehat m_{fc}(t,x)|\,,\qquad \quad |U'(t,x)+x|> 2|\re m_{fc}(t,x)|\,.
\end{align}
\item[$(2)$] There exists a constant $c>0$ such that for all $x\in\ensuremath{\mathbb{R}}$ and $t\ge0$ we have
\begin{align}
|\widehat U'(t,x)-U'(t,x)|\le N^{-c\alpha_0/2}\,,
\end{align}
where $\alpha_0>0$ is the constant in~\eqref{definition of omega_V}.
\item[$(3)$] The potentials $\widehat U$ and $U$ satisfy~\eqref{assumption 1 for beta universality} and~\eqref{assumption 2 for beta universality}. In particular, there is $C_U\ge 0$ (independent of $N$), such that
\begin{align}\label{uniform convexity bound}
\inf_{x\in\ensuremath{\mathbb{R}}, t\in\ensuremath{\mathbb{R}}^+}\widehat U''(t,x)\ge -2C_U\,,\qquad\quad \inf_{x\in\ensuremath{\mathbb{R}}, t\in\ensuremath{\mathbb{R}}^+}\, U''(t,x)\ge -2C_U\,.
\end{align}
Moreover, $\widehat U$ and $U$ are ``regular'' (see the paragraph below~\eqref{regular potential U} for the definition of ``regular'' potential).
\end{itemize}
\end{lemma}
Below, we are mainly interested in $\beta$-ensembles determined by the potential $\widehat U$. For ease of notation we thus limit the discussion to $\widehat U$.
For $N\in\N$ we define a measure on $\digamma^{(N)}$, by
\begin{align}\label{definition of bt}
{\widehat\psi_t}(\mathbf{x})\,\mu_G(\mathrm{d}\mathbf{x})\mathrel{\mathop:}= \frac{1}{Z_{{\widehat\psi_t}}}\e{-\frac{\beta N}{2}{\sum_{i=1}^N\widehat U(t,x_i)}}\mu_G(\mathrm{d}\mathbf{x})\,,\qquad\quad (\mathbf{x}\in\digamma^{(N)})\,,
\end{align}
where $Z_{{\widehat\psi_t}}\equiv Z_{{\widehat\psi_t}}(\beta)$ is a normalization and we usually choose $\beta=1,2$. By Lemma~\ref{superlemma}, ${\widehat\psi_t}\,\mu_G$ is a well-defined $\beta$-ensemble and from the discussion in Subsection~\ref{subsection beta ensemble intro} we further infer that the equilibrium density of $\widehat\psi_t\,\mu_G$, i.e., the unique measure solving the minimization problem in~\eqref{minimization problem}, is, for any $t\ge0$, $\widehat\rho_{fc}(t)$. Viewing ${\widehat\psi_t}\,\mu_G$ as a Gibbs measure of $N$ (ordered) particles $(x_i)$ on the real line, we define the classical location of the $i$-th particles, $\widehat\gamma_i(t)$, as in~\eqref{general beta ensemble classical location}, i.e.,
\begin{align}\label{nochmals classical locations}
\int_{-\infty}^{\widehat\gamma_i(t)}\widehat\rho_{fc}(t,x)\,\mathrm{d} x=\frac{i-\frac{1}{2}}{N}\,,\qquad\quad (i\in\llbracket 1,N\rrbracket)\,.
\end{align}
From~\cite{BEY} we have the following rigidity result.
\begin{proposition}\label{rigidity for time dependent beta one}
Let $\widehat U(t,\cdot)$, with $t\ge 0$ and $N\in\N$, be given by Lemma~\ref{superlemma}. Then the following holds on $\Omega_V$. For any $\delta>0$, there is $\varsigma>0$, such that for any $t\ge 0$,
\begin{align} \label{rigidity for b ensemble one}
\mathbb{P}^{{\widehat\psi_t}\,\mu_G}\left(|x_i-\widehat \gamma_i(t)|> N^{-\frac{2}{3}+ \delta }\widehat\alpha_i^{-\frac{1}{3}}
\right)\leq \e{- N^\varsigma}\,, \qquad (1\le i\le N)\,,
\end{align}
for $N$ sufficiently large, where $\mathbb{P}^{{\widehat\psi_t}\,\mu_G}$, stands for the probability under ${\widehat\psi_t}\,\mu_{G}$ conditioned on $V$. Here, $\widehat\alpha_i\mathrel{\mathop:}=\min\{i,N-i\}$.
\end{proposition}
\begin{proof}
The rigidity estimate \eqref{rigidity for b ensemble one} is taken from Theorem~2.4 of~\cite{BEY}. To achieve uniformity in $t\ge 0$ and $N$ sufficiently large, we note that the estimate~\eqref{rigidity for b ensemble one} depends on the potential mainly through the convexity bounds~\eqref{assumption 1 for beta universality} and~\eqref{assumption 2 for beta universality}. Starting from the uniform bounds of Lemma~\ref{superlemma}, one checks that Proposition~\ref{rigidity for time dependent beta one} holds
uniformly in $t$ and $N$ large enough.
\end{proof}
In the rest of this section, we derive equations of motion for the potential $\widehat U(t,\cdot)$ and the classical locations $(\widehat\gamma_i(t))$. To derive these equations we observe that the Stieltjes transform $\widehat m_{fc}(t,z)$ can be obtained from $\widehat m_{fc}(t=0,z)$ as the solution to the following complex Burger equation~\cite{P},
\begin{align}\label{burger equation}
\partial_t \widehat m_{fc}(t,z)=\frac{1}{2}\partial_z \left[ \widehat m_{fc}(t,z)\,(\widehat m_{fc}(t,z)+z) \right]\,, \qquad\qquad (z\in\ensuremath{\mathbb{C}}^+\,,t\ge 0)\,.
\end{align}
This can be checked by differentiating~\eqref{nochmals mfc}. Combining the complex Burger equation~\eqref{burger equation} and the loop equation~\eqref{loop equation} we obtain the following result.
\begin{lemma}
Let $N\in\N$. Assume that $\widehat\nu$ satisfies the Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}. Then the following holds on $\Omega_V$ for~$N$ sufficiently large. For $t\ge 0$, we have
\begin{align}
\partial_t\widehat\gamma_i(t)=\frac{1}{2}\widehat U'(t,\widehat\gamma_i(t))\,,\label{EoM classical locations}
\end{align}
respectively,
\begin{align}\label{EoM zwei}
\partial_t\widehat\gamma_i(t)=-\Xint-_\ensuremath{\mathbb{R}}\frac{\widehat\rho_{fc}(t,y)}{y-\widehat\gamma_i(t)}\,\mathrm{d} y-\frac{1}{2}\widehat\gamma_i(t)\,,\qquad\qquad (i\in\llbracket 1,N\rrbracket)\,.
\end{align}
Further, the potential $\widehat U$ satisfies,
\begin{align}\label{EoM drei}
\qquad\partial_t \widehat U(t,x)&=\Xint-_\ensuremath{\mathbb{R}}\frac{\widehat U'(t,y)\widehat\rho_{fc}(t,y)}{y-x}\,\mathrm{d} y\,,\qquad\quad (x\in\supp\widehat\rho_{fc}(t))\,.
\end{align}
Moreover, there exist constants $C,C'$ such that the following bounds hold on $\Omega_V$,
\begin{align}\label{bounds EoM}
|\partial_{t}\widehat\gamma_i(t)|\le C\,,\qquad\quad |\partial_t \widehat U(t,x)|\le C'\,,
\end{align}
for all $i\in\llbracket 1,N\rrbracket$, uniformly in $t\ge 0$, $x\in\supp\,\widehat\rho_{fc}(t)$ and $N$, for $N$ sufficiently large.
Finally, $U(t,\cdot)$ and $(\gamma_i(t))$, share the same properties.
\end{lemma}
\begin{proof}
Combining~\eqref{burger equation} and~\eqref{loop equation}, we find, for $z\in\ensuremath{\mathbb{C}}^+$, $t\ge 0$,
\begin{align*}
\partial_t\widehat m_{fc}(t,z)&=\frac{1}{2}\partial_z\left(-\int\frac{v+\widehat U'(t,v)}{v-z}\widehat\rho_{fc}(t,v)\,\mathrm{d} v+z\int\frac{\widehat\rho_{fc}(t,v)}{v-z}\,\mathrm{d} v\right)\\
&=\frac{1}{2}\partial_z\left(-\int\frac{\widehat U'(t,v)}{v-z}\widehat\rho_{fc}(t,v)\,\mathrm{d} v-1\right)\\
&=-\frac{1}{2}\partial_z\int\frac{\widehat U'(t,v)}{v-z}\widehat\rho_{fc}(t,v)\,\mathrm{d} v\,.
\end{align*}
Hence, for $\im z>0$, we get
\begin{align*}
\partial_t\widehat m_{fc}(t,z)&=-\frac{1}{2}\int\frac{\widehat U'(t,v)}{(v-z)^2}\widehat\rho_{fc}(t,v)\,\mathrm{d} v=-\frac{1}{2}\int\frac{\left(\widehat U'(t,v)\widehat\rho_{fc}(t,v)\right)'}{(v-z)}\,\mathrm{d} v\,.
\end{align*}
Clearly $\widehat U'(t,v)\widehat\rho_{fc}(t,v)$ is a $C^3$ function in the support of $\widehat\rho_{fc}(z)$ that has a square root behavior at the endpoints. Thus we obtain from the Stieltjes inversion formula that
\begin{align}\label{tapir}
\partial_t\widehat\rho_{fc}(t,E)&=\frac{1}{\pi}\lim_{\eta\searrow 0}\im\partial_t \widehat m_{fc}(t,z)=-\frac{1}{2}\big(\widehat U'(t,E)\widehat\rho_{fc}(t,E)\big)'\,,
\end{align}
for all $E\in(\widehat L_-(t),\widehat L_+(t))$, where $\widehat L_{\pm}(t)$ denote the endpoints of the support of $\widehat\rho_{fc}(t)$.
On the other hand, differentiating~\eqref{nochmals classical locations} with respect to time, we obtain
\begin{align*}
\int_{-\infty}^{\widehat\gamma_i(t)}\partial_t\widehat\rho_{fc}(t,v)\,\mathrm{d} v=-\widehat\rho_{fc}(t,\widehat\gamma_i(t))\partial_t\widehat\gamma_i(t)\,.
\end{align*}
Substituting from~\eqref{tapir}, we get
\begin{align*}
\partial_t\widehat\gamma_i(t)=\frac{1}{2}\frac{1}{\widehat\rho_{fc}(t,\widehat\gamma_i(t))}\int_{-\infty}^{\widehat\gamma_i(t)}\,\mathrm{d} v\,\left( \widehat{U}'_{fc}(t,v)\widehat\rho_{fc}(t,v)\right)'\,.\\
\end{align*}
Hence,
\begin{align*}
\partial_t\widehat\gamma_i(t)&=\frac{1}{2}\frac{1}{\widehat\rho_{fc}(t,\widehat\gamma_i(t))}\widehat U'\big(t,\widehat\gamma_i(t)\big)\,\widehat\rho_{fc}\big(t,\widehat\gamma_i(t)\big)\,,
\end{align*}
and~\eqref{EoM classical locations} follows. Using that $\widehat U$ satisfies ~\eqref{definition widehat U}, we can recast this last equation as
\begin{align*}
\partial_t\widehat\gamma_i(t)=-\Xint-_\ensuremath{\mathbb{R}}\frac{\widehat\rho_{fc}(t,y)}{y-\widehat\gamma_i(t)}\,\mathrm{d} y-\frac{1}{2}\widehat\gamma_i(t)
\end{align*}
and we find~\eqref{EoM zwei}. Equation~\eqref{EoM zwei} follows in a similar way by differentiating~\eqref{definition widehat U} with respect to time. By a similar computation we obtain~\eqref{EoM drei}. The bound in~\eqref{bounds EoM} follow from Lemma~\ref{superlemma}.
\end{proof}
Starting from the relations in~\eqref{nochmals mfc}, we derived via the time-dependent potential $\widehat U$, equation of motions for the classical locations $(\widehat\gamma_i(t))$. The points $(\widehat\gamma_i(t))$ may also be viewed as the classical locations of the eigenvalues of a family of random matrices which is parametrized by the times $t_0$ and $t$. This is the subject of the next section.
\section{Dyson Brownian Motion: Evolution of the entropy}\label{section DBM}
\subsection{Dyson Brownian motion}
Let $H_{0}=(h_{ij,0})$ be the matrix
\begin{align*}
H_{0}\mathrel{\mathop:}= \e{t_0/2}\,V+W\,,
\end{align*}
where $V$ satisfies Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}, and $W$ is real symmetric or complex Hermitian satisfying the assumptions in Definition~\ref{assumption wigner}. Here $t_0\ge 0$ is chosen such that $\vartheta=\e{t_0/2}\in\Theta_{\varpi}$ (see~\eqref{definition Theta}) and we consider $\vartheta$ as an a priori free ``coupling parameter'' that we fix in Section~\ref{Proofs of main results} below. Let $B=(b_{ij})\equiv(b_{ij,t})$ be a real symmetric, respectively a complex Hermitian, matrix whose entries are a collection of independent, up to the symmetry constraint, real (complex) Brownian motions, independent of $(h_{ij,0})$. More precisely, in case $W$ is a complex Hermitian Wigner matrix, we choose the entries $(b_{ij,t})$ to have variance $t$; in case $W$ is a real symmetric Wigner matrix, we choose the off-diagonal entries of $(b_{ij,t})$ to have variance $t$, while the diagonal entries are chosen to have variance~$2t$. Let $H_t=(h_{ij,t})$ satisfy the
stochastic differential equation
\begin{align}\label{le sde}
\mathrm{d} h_{ij}=\frac{\mathrm{d} b_{ij}}{\sqrt N}-\frac{1}{2}h_{ij}\mathrm{d} t\,,\qquad \quad(t\ge 0)\,.
\end{align}
It is then easy to check that the distribution of $H_{t}$ agrees with the distribution of the matrix
\begin{align}\label{the new matrix}
\e{-(t-t_0)/2}V+\e{-t/2}W+(1-\e{-t})^{1/2}W'\,,
\end{align}
where $W'$ is, in case $W$ is a complex Hermitian, a GUE matrix, independent of $V$ and $W$, respectively a GOE matrix, independent of $V$ and $W$, in case $W$ is a real symmetric Wigner matrix. The law of the eigenvalues of the matrix $W'$ is explicitly given by~\eqref{gaussian measures} with $\beta=2$, respectively $\beta=1$.
Denote by $\boldsymbol{\lambda}(t)=(\lambda_{1}(t),\lambda_{2}(t),\ldots,\lambda_{N}(t) )$ the ordered eigenvalues of $H_t$. It is well-known that $\boldsymbol{\lambda}(t)$ satisfy the following stochastic differential equation,
\begin{align}\label{dyson brownian motion}
\mathrm{d} \lambda_{i}=\frac{\sqrt{2}}{\sqrt{\beta N}}\,{\mathrm{d} b_{i}}+\left(-\frac{\lambda_{i}}{2}+\frac{1}{N}\sum_{j}^{(i)}\frac{1}{\lambda_{i}-\lambda_{j}}\right)\mathrm{d} t\,,\qquad\quad (i\in\llbracket 1,N\rrbracket)\,,
\end{align}
where $(b_i)$ is a collection of real-valued, independent standard Brownian motions. If the matrix $(b_{ij})$ in~\eqref{le sde} is real symmetric, we have $\beta=1$ in~\eqref{dyson brownian motion}, respectively $\beta=2$ if $(b_{ij})$ is complex Hermitian. The evolution of~$\boldsymbol{\lambda}(t)$ is the celebrated Dyson Brownian motion~\cite{D}.
For $t\ge 0$, we denote by $f_t\,\mu_G$ the distribution of $\boldsymbol{\lambda}(t)$. In particular, \mbox{$\int f_t\,\mathrm{d} \mu_G\equiv\int f_t(\boldsymbol{\lambda})\,\mu_G(\mathrm{d}\boldsymbol{\lambda})=1$}. Note that $f_t\,\mu_G$ depends on $V$ through the initial condition $f_{0}\,\mu_G$. In the following we always keep the $(v_i)$ fixed, i.e., we condition on~$V$. For simplicity, we omit this conditioning from our notation. The density $f_t$ is the solution of the equation
\begin{align*}
\partial_tf_t={\mathcal L} f_t\,,\qquad\quad (t\ge 0)\,,
\end{align*}
where the generator ${\mathcal L}$ is defined via the Dirichlet form
\begin{align}\label{dirichlet form}
D_{\mu_G}(f) = -\int f {\mathcal L} f \mathrm{d}{\mu_G} = \sum_{i=1}^N \frac{1}{\beta N}
\int (\partial_i f)^2 \mathrm{d}{\mu_G}\,, \qquad\quad (\partial_i \equiv\partial_{x_i})\,.
\end{align}
Formally, we have ${\mathcal L}= \frac{1}{\beta N}\Delta - (\nabla {\mathcal H}_G)\cdot\nabla$, i.e.,
\begin{align}\label{le generator}
{\mathcal L}=\sum_{i=1}^N\frac{1}{\beta N}\partial _i^2+\sum_{i=1}^N\Big(-\frac{1}{2}\lambda_i+\frac{1}{N}\sum_{j}^{(i)}\frac{1}{\lambda_i-\lambda_j} \Big)\,\partial_i\,.
\end{align}
We remark that we use a different normalization in the definition of the Dirichlet form $D_{\mu_G}(f)$ in~\eqref{dirichlet form} (and the generator ${\mathcal L}$) than in earlier works, e.g., in~\cite{EY}, where the Dirichlet from was defined as $\sum_{i=1}^N\frac{1}{2N}\int(\partial_i f)^2\mathrm{d}\mu_G$.
\begin{lemma}{\emph{[Dyson Brownian Motion]}}
The equation $\partial_t f_t={\mathcal L} f_t$, with initial data\footnote{ Strictly speaking, the eigenvalue distribution of $H_0$ may not allow
a density $f_0$, but for $t>0$, $H_t$ admits a density $f_t$. Our proofs are not affected by this technicality.} $f_t|_{t=0}=f_{0}$ has a unique solution on $\mathrm{L}^1(\mu_G)\equiv \mathrm{L}^1(\ensuremath{\mathbb{R}}^N,\mu_G)$ for all $t\ge 0$. Moreover, the domain $\digamma^{(N)}$ is invariant under the dynamics, i.e., if $f_{0}$ is supported in $\digamma^{(N)}$ then is $f_t$ for all $t\ge 0$.
\end{lemma}
We refer, e.g., to~\cite{AGZ} for more details and proofs. To conclude, we record one of the technical tools used in the next sections.
\begin{lemma}\label{lemma master rigidity bound}
Denote by $f_t(\boldsymbol{\lambda})\,\mu_G(\mathrm{d}\boldsymbol{\lambda})$ the distribution of the eigenvalues of the matrix~\eqref{the new matrix} with $t\ge 0$. Then, for any $0<\mathfrak{a}_1<1/2$, we have
\begin{align}\label{master rigidity bound}
\sup_{t\ge 0}\int\frac{1}{N}\sum_{i=1}^N(\lambda_i-\widehat\gamma_i(t))^2f_t(\boldsymbol{\lambda})\mathrm{d}\mu(\boldsymbol{\lambda})\le N^{-1-2\mathfrak{a}_1}\,,
\end{align}
on $\Omega_V$ for $N$ sufficiently large, where $(\widehat\gamma_i(t))$ denote the classical locations with respect to the measure $\widehat\rho_{fc}(t)$, i.e., they are defined through the relation
\begin{align}\label{classical locations once more}
\int_{-\infty}^{\widehat\gamma_i(t)}\mathrm{d} x\,\widehat\rho_{fc}(t,x)=\frac{i-\frac{1}{2}}{N}\,,\qquad\qquad (1\le i\le N)\,.
\end{align}
(They agree with the classical locations of~\eqref{nochmals classical locations}.)
\end{lemma}
\begin{proof}
The random matrix $W_t\equiv(w_{ij,t})\mathrel{\mathop:}=\e{-t/2}W+(1-\e{-t})^{1/2}{W'}$, satisfies the assumptions in Definition~\ref{assumption wigner}: The entries are centered and have variance $1/N$. Moreover, since the distributions of $(w_{ij,0})$, satisfies~\eqref{eq.C0} and since $(w'_{ij})$ are real, respectively complex, centered Gaussian random variables with variance $1/N$, respectively $2/N$, the distributions of $(w_{ij,t})$ also satisfy~\eqref{eq.C0}. The claim now follows from~\eqref{rigidity of eigenvalues equation} of Corollary~\ref{rigidity of eigenvalues} and the moment bounds $ \E\Tr W_t^{2p}\le C_p$ (see, e.g.,~\cite{AGZ}), as well as the boundedness of $(v_i)$.
\end{proof}
\subsection{Entropy decay estimates}
Let $\omega$ and $\nu$ be two (probability) measures on $\ensuremath{\mathbb{R}}^N$ that are absolutely continuous with respect to Lebesgue measure. We denote the Radon-Nikodym derivative of $\nu$ with respect to $\omega$ by $\frac{\mathrm{d}\nu}{\mathrm{d}\omega}$, define the relative entropy of $\nu$ with respect to $\omega$ by
\begin{align}\label{definition of relativ entropy}
S(\nu|\omega) \mathrel{\mathop:}= \int_{\ensuremath{\mathbb{R}}^N} \f{\mathrm{d}\nu}{\mathrm{d}\omega} \log \f{\mathrm{d}\nu}{\mathrm{d}\omega}\, \mathrm{d}\omega \,,
\end{align}
and, in case $\nu=f\omega$, $f\in\mathrm{L}^1(\ensuremath{\mathbb{R}}^N)$, abbreviate
\begin{align*}
S_{\omega}(f)=S(f\omega|\omega)\,.
\end{align*}
The entropy $S_{\omega}(f)$ controls the total variation norm of $f$ through the inequality
\begin{align}\label{entropy inequality}
\int |f-1|\,\mathrm{d}\omega\le \sqrt{2 S_{\omega}(f)}\,,
\end{align}
a result we will use repeatedly in the next sections.
Besides the dynamics $(f_t)_{t\ge0}$ generated by ${\mathcal L}$ introduced in the previous subsection, we also consider a (apriori undetermined) time-dependent density, $(\widetilde{\psi}_t)_{t\ge 0}$, with respect to $\mu_G$. We assume that $\widetilde{\psi}_t\not=0$, almost everywhere with respect to $\mu_G$ and abbreviate $\widetilde g_t\mathrel{\mathop:}=\frac{f_t}{\widetilde{\psi}_t}$. Setting $\widetilde\omega_t\mathrel{\mathop:}= \widetilde{\psi}_t\,\mu_G$, we have $f_t(\boldsymbol{\lambda})\,\mu_G(\mathrm{d}\boldsymbol{\lambda})=\widetilde g_t(\boldsymbol{\lambda})\,\widetilde\omega_t(\mathrm{d}\boldsymbol{\lambda})$.
A natural choice for $\widetilde{\psi}_t\,\mu_G$ is the time dependent $\beta$-ensemble, ${\widehat\psi_t}\,\mu_G$, introduced in~\eqref{definition of bt}. Yet, following the arguments of~\cite{ESYY} we make a slightly different choice for $\widetilde{\psi}_t$: For $\tau>0$, we define a measure $\widetilde{\psi}_t\,\mu_G$ on $\digamma^{(N)}$ by setting
\begin{align}\label{definition of psit}
\widetilde{\psi}_t(\boldsymbol{\lambda})\,\mu_G(\mathrm{d}\boldsymbol{\lambda})\mathrel{\mathop:}= \frac{1}{Z_{\widetilde{\psi}_t}'}\e{-N\beta \sum_{i=1}^N\frac{(\lambda_i-\widehat\gamma_i(t))^2}{2\tau}}{\widehat\psi_t}(\boldsymbol{\lambda})\,\mu_G(\mathrm{d}\boldsymbol{\lambda})\,,
\end{align}
where $Z_{\widetilde{\psi}_t}'\equiv Z_{\widetilde{\psi}_t}'(\beta)$ is chosen such that $ \int\widetilde{\psi}_t(\boldsymbol{\lambda})\,\mu_G(\mathrm{d}\boldsymbol{\lambda})=1$. In the following, we mostly choose $\tau$ to be $N$-dependent with $1\gg\tau>0$.
We call the measure $\widetilde{\psi}_t\,\mu_G$ the instantaneous relaxation measure. The density $\widetilde{\psi}_t$ depends on $V=\mathrm{diag}(v_i)$ via the initial condition $\widetilde\psi_{0}$. As for the distribution $f_t$, we condition on~$V$ and omit this from the notation. We may write the measure $\widetilde{\psi}_t\,\mu_G$ in the Gibbs form
\begin{align*}
\widetilde{\psi}_t(\boldsymbol{\lambda})\mu_G(\mathrm{d}\boldsymbol{\lambda})=\frac{1}{Z_{\widetilde{\psi}_t}}\e{-\beta N\widetilde{{\mathcal H}}_t(\boldsymbol{\lambda})}\mathrm{d}\boldsymbol{\lambda}\,,
\end{align*}
with
\begin{align}\label{instantaneous hamiltonian}
\widetilde{{\mathcal H}}_t(\boldsymbol{\lambda})={\mathcal H}_G(\boldsymbol{\lambda})+\sum_{i=1}^N\left(\frac{(\lambda_i-\widehat\gamma_i(t))^2}{2\tau}+\frac{\widehat U(t,\lambda_i)}{2}\right)\,,\qquad\qquad(\boldsymbol{\lambda}\in\digamma^{(N)})\,,
\end{align}
where ${\mathcal H}_G$ is defined in~\eqref{gaussian hamiltonian} and $Z_{\widetilde\psi_t}\equiv Z_{\widetilde\psi_t}(\beta)$ is a normalization. Then we compute
\begin{align}\label{instantaneous convexity bound}
\nabla u\cdot(\nabla^2 \widetilde{{\mathcal H}}_t)\cdot\nabla u&\ge \sum_{i=1}^N(\partial_i u)^2\left(\frac{1}{\tau}+\frac{\widehat U''(t,\lambda_i)}{2}+\frac{1}{2}\right)+\frac{1}{N}\sum_{i=1}^N\sum_{j}^{(i)}\frac{1}{(x_i-x_j)^2}(\partial_i u-\partial_j u)^2\nonumber\\
&\ge \sum_{i=1}^N\frac{(\partial_i u)^2}{2\tau}\,,
\end{align}
for $u\in C^1(\ensuremath{\mathbb{R}}^N)$ and $\tau$ sufficiently small (independent of $N$), where we used that $\widehat U''(t,\cdot)$ is uniformly bounded below by Lemma~\ref{superlemma}. Then, by the Bakry-\'{E}mery criterion~\cite{BE}, there is a constant $C$ such that the following logarithmic Sobolev inequality holds for all sufficiently small $\tau>0$:
\begin{align}\label{log sobolev}
S_{\widetilde\omega_t}(q)\le C \tau D_{\widetilde\omega_t}(\sqrt{q})\,,\qquad\quad (t\ge 0)\,,
\end{align}
where $q\in\mathrm{L}^\infty(\mathrm{d}\widetilde \omega_t)$ is such that $\int q\,\mathrm{d}\widetilde\omega_t=1$. We refer, e.g., to~\cite{ESY4, ESYY,EY,EYY} for more details.
The main result of this section is the following proposition. Recall the definition of ${\widehat\psi_t}\,\mu_G$ in~\eqref{definition of bt}.
\begin{proposition}\label{the super proposition}
Let $\widehat g_t\mathrel{\mathop:}= f_t/\widehat{\psi}_t$ and set $\widehat\omega_t\mathrel{\mathop:}=\widehat{\psi}_t\,\mu_G$ such that $S(f_t\,\mu_G|\widehat{\psi}_t\,\mu_G)= S_{\widehat\omega_t}(\widehat g_t)$. Then, there is a constant~$C$ (independent of $t$) such that
\begin{align}\label{the super proposition bound}
\partial_t S_{\widehat\omega_t}(\widehat g_t)\le - 4 D_{\widehat\omega_t}(\sqrt{\widehat g_t})+C{N^{1-2\mathfrak{a}}}\,,\qquad\qquad (t> 0)\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$.
\end{proposition}
The results of Proposition~\ref{the super proposition} resemble the relative entropy estimate of Theorem~2.5 in~\cite{EY} for Wigner matrices. However, due to the fact that both distributions $f_t\,\mu_G$ and $\widehat\psi_t\,\mu_G$ are not close to the global equilibrium for the Dyson Brownian motion, $\mu_G$, the reference ensemble $\widehat\psi_t\,\mu_G$ changes with time, too. Thus to establish~\eqref{the super proposition bound}, we need to include additional factors coming from time derivatives of $\widehat\psi_t\,\mu_G$. These can be controlled using the definition of the potential~$\widehat U(t)$. The idea of choosing slowly varying time dependent approximation states and controlling the entropy flow goes back to the work
\cite{Y}.
The relative entropy $S_{\widehat\omega_t}$ and the Dirichlet form $D_{\widehat\omega_t}$ do not satisfy the logarithmic Sobolev inequality~\eqref{log sobolev}. However, we have for $t> 0$ the estimates
\begin{align}
D_{\widehat\omega_t}(\sqrt{\widehat g_t})&\le 2D_{{\widetilde\omega_t}}(\sqrt{\widetilde g_t})+C\frac{\beta N^2Q_t}{\tau^2}\,,\qquad\qquad D_{\widetilde\omega_t}(\sqrt{\widetilde g_t})\le 2D_{\widehat\omega_t}(\sqrt{\widehat g_t})+C\frac{\beta N^2Q_t}{\tau^2}\label{dirichlet equivalence}\,,
\end{align}
and
\begin{align}\label{entropy equivalence}
S_{\widetilde\omega_t}(\widetilde g_t)=S_{\widehat\omega_t}(\widehat g_t)+{\mathcal O}\left(\frac{\beta N^2Q_t}{\tau} \right)\,,
\end{align}
where we have set
\begin{align}\label{le Qt}
Q_t\mathrel{\mathop:}=\E^{f_t\,\mu_G}\frac{1}{N}\sum_{i=1}^N(\lambda_i-\widehat\gamma_i(t))^2\,.
\end{align}
The estimates~\eqref{dirichlet equivalence} and~\eqref{entropy equivalence} can be checked by elementary computations which we omit here. In the following we always bound $Q_t\le CN^{-1-2\mathfrak{a}_1}$ ($t\ge 0$, $\mathfrak{a}_1\in(0,1/2)$); see Lemma~\ref{master rigidity bound}. Using~\eqref{dirichlet equivalence} and~\eqref{entropy equivalence} in combination with the logarithmic Sobolev inequality~\eqref{log sobolev} and with Proposition~\ref{the super proposition}, we can follow~\cite{EY} to obtain a bound on the Dirichlet form $D_{\widehat\omega_t}(\sqrt{\widehat g_t})$.
\begin{corollary}\label{lemma bound on entroyp an dirichlet form}\label{beta dirichlet bound}
Under the assumptions of Proposition~\ref{the super proposition} the following holds on $\Omega_V$ for $N$ sufficiently large. For any $\epsilon'>0$ and $t \geq \tau N^{\epsilon'}$ with $1\gg\tau\ge N^{-2\mathfrak{a}}$, we have the entropy and Dirichlet form bounds
\begin{align}\label{bound on entropy and dirichlet form}
S_{\widehat\omega_t}(\widehat g_t)\le C \frac{N^{1-2\mathfrak{a}}}{\tau}\,,\qquad\qquad D_{\widehat\omega_t}(\sqrt{\widehat g_t})\leq C \frac{N^{1-2\mathfrak{a}}}{\tau^2}\,,
\end{align}
where the constants depend on $\epsilon'$.
\end{corollary}
Before we prove Proposition~\ref{the super proposition}, we obtain rigidity estimates for the time-dependent $\beta$-ensemble ${\widehat\psi_t}\,\mu_G$. Recall that we denote by~$\widehat\gamma_i(t)$ the classical locations with respect to the measure~$\widehat\rho_{fc}(t)$. Also recall the notation $\widehat\alpha_i=\min\{i,N-i\}$.
\begin{lemma}\label{rigidity for time dependent beta}
Let $\widehat U(t,\cdot)$, $t\ge 0$, be as in Lemma~\ref{superlemma}. Then the following holds on $\Omega_V$ for $N$ sufficiently large:
For any $\delta>0$, there is $\varsigma>0$ such that
\begin{align} \label{rigidity for b ensemble}
\mathbb{P}^{{\widehat\psi_t}\mu_G}\left(|\lambda_i-\widehat \gamma_i(t)|> N^{-\frac{2}{3}+ \delta }\widehat\alpha_i^{-\frac{1}{3}}
\right)\leq \e{- N^\varsigma}\,, \qquad\quad (1\le i\le N,\,t\ge0)\,,
\end{align}
where $\mathbb{P}^{{\widehat\psi_t}\mu_G}$, stands for the probability under ${\widehat\psi_t}\,\mu_{G}$ conditioned on $\Omega_V$. Moreover, for any $0<\mathfrak{a}_2<1/2$, we have
\begin{align}\label{master rigidity b}
\sup_{t\ge 0}\int\frac{1}{N}\sum_{i=1}^N(\lambda_i-\widehat\gamma_i(t))^2 {\widehat\psi_t}(\boldsymbol{\lambda})\mu_G(\mathrm{d}\boldsymbol{\lambda})\le N^{-1-2\mathfrak{a}_2}\,,
\end{align}
for $N$ sufficiently large.
\end{lemma}
\begin{proof}
The rigidity estimate \eqref{rigidity for b ensemble} follows from Proposition~\ref{rigidity for time dependent beta one} by choosing $N\in\N$ sufficiently large. The estimate~\eqref{master rigidity b} is a direct consequence of~\eqref{rigidity for b ensemble} and the fast decay of the distribution ${\widehat\psi_t}(\boldsymbol{\lambda}) \mu_G(\boldsymbol{\lambda})$.
\end{proof}
Below we choose $\mathfrak{a}\mathrel{\mathop:}=\min\{\mathfrak{a}_1,\mathfrak{a}_2 \}$, where $\mathfrak{a}_1$ is the constant in Lemma~\ref{lemma master rigidity bound}. For brevity, we often drop the $t$-dependence of $\widehat\gamma_i(t)$ from the notation.
\begin{proof}[Proof of Proposition~\ref{the super proposition}]
Recall that we have set $\widehat g_t=f_t/\widehat{\psi}_t$ and $\widehat\omega_t=\widehat{\psi}_t\,\mu_G$. The relative entropy $S(f_t\,\mu_G|\widehat{\psi}_t\,\mu_G)=S_{\widehat\omega_t}(\widehat g_t)$ satisfies~\cite{Y},
\begin{align}\label{landau ginzburg formula}
\ensuremath{\partial_{t}} S(f_t\,\mu_G|\widehat{\psi}_t\,\mu_G) = -\f{1}{\beta N}\int \f{|\nabla \widehat g_t|^2}{g_t} \widehat{\psi}_t\, \mathrm{d}\mu_G
+\int\f{(\mathcal{L} - \partial_t)\widehat{\psi}_t}{\widehat{\psi}_t} f_t \,\mathrm{d}\mu_G\,.
\end{align}
We note that the first term on the right side of~\eqref{landau ginzburg formula} equals
\begin{align}\label{landau ginzburg lemma 0}
-\f{1}{\beta N}\int \f{|\nabla \widehat g_t|^2}{\widehat g_t} \widehat{\psi}_t \,\mathrm{d}\mu_G=-4D_{\widehat\omega_t}(\sqrt{\widehat g_t})\,.
\end{align}
To bound the second term on the right side of~\eqref{landau ginzburg formula}, we write
\begin{align}\label{entropy estimate 1}
\int\f{(\mathcal{L} - \partial_t)\widehat{\psi}_t}{\widehat{\psi}_t} f_t \d\mu_G&
=\int(\widehat{{\mathcal L}}_t \widehat g_t)\mathrm{d}\widehat\omega_t+\frac{1}{2}\int\sum_{i=1}^N\widehat U'(t,\lambda_i)(\partial_i \widehat g_t(\boldsymbol{\lambda}))\mathrm{d}\widehat\omega_t(\boldsymbol{\lambda})-\int \widehat g_t\partial_t\widehat{\psi}_t\d\mu_G\,,
\end{align}
with $\widehat{\mathcal L}_t$ defined through the natural Dirichlet form with respect to $\widehat\omega_t$, i.e.,
\begin{align}
D_{\widehat\omega_t}(q)=\frac{1}{\beta N}\sum_{i=1}^N\int (\partial_i q)^2\mathrm{d}\widehat\omega_t=-\int q\,\widehat{\mathcal L}_t q\,\mathrm{d}\widehat\omega_t\,,\qquad\qquad (t> 0)\,.
\end{align}
Note that the first term on the right side of~\eqref{entropy estimate 1} vanishes since, by construction, $\widehat\omega_t$ is the reversible measure for the instantaneous flow generated by $\widehat{{\mathcal L}}_t$. The last term on the right side of~\eqref{entropy estimate 1} can be computed explicitly as (recall that the normalization $Z_{\widehat{\psi}_t}$ in the definition of $\widehat{\psi}_t\,\mu_G$ also depends on $t$),
\begin{align}\label{neue 1}
-\int \widehat g_t\partial_t\widehat{\psi}_t\d\mu_G&=\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[\frac{\beta N}{2}\sum_{i=1}^N\partial_t \widehat U(t,\lambda_i)\right]\,.
\end{align}
To deal with the second term on the right side of~\eqref{entropy estimate 1}, we integrate by parts to find
\begin{align}\label{freude}
\frac{1}{2}\int\sum_{i=1}^N\widehat U'(t,\lambda_i)(\partial_i\widehat g_t(\boldsymbol{\lambda}))\mathrm{d}\widehat\omega_t(\boldsymbol{\lambda})&=\E^{f_t\,\mu_G}\left[-\frac{1}{2}\sum_{i=1}^N \widehat U''(t,\lambda_i)+\frac{\beta N}{4}\sum_{i=1}^N \widehat U'(t,\lambda_i)\left(\widehat U'(t,\lambda_i)+\lambda_i-\frac{2}{N}\sum_{j}^{(i)}\frac{1}{\lambda_i-\lambda_j}\right)\right]\,.
\end{align}
Setting $\widehat g_t\equiv 1$ in the above computation, we also obtain the identity
\begin{align}\label{loop eq 2}
0&=\E^{\widehat\psi_t\,\mu_G}\left[-\frac{1}{2}\sum_{i=1}^N \widehat U''(t,\lambda_i)+\frac{\beta N}{4}\sum_{i=1}^N \widehat U'(t,\lambda_i)\left(\widehat U'(t,\lambda_i)+\lambda_i-\frac{2}{N}\sum_{j}^{(i)}\frac{1}{\lambda_i-\lambda_j}\right)\right]\,.
\end{align}
Equation~\eqref{loop eq 2} may alternatively be derived from the ``first order loop equation'' for the $\beta$-ensemble $\widehat\psi_t\,\mu_G$. Equation~\eqref{freude} can thus be rewritten as
\begin{align}\label{neue 35}
\frac{1}{2}\int\sum_{i=1}^N\widehat U'(t,\lambda_i)(\partial_i\widehat g_t(\boldsymbol{\lambda}))\mathrm{d}\widehat\omega_t(\boldsymbol{\lambda})&=\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[\frac{\beta N}{4}\sum_{i=1}^N \widehat U'(t,\lambda_i)\left(\widehat U'(t,\lambda_i)+\lambda_i-\frac{2}{N}\sum_{j}^{(i)}\frac{1}{\lambda_i-\lambda_j}\right)\right]\nonumber\\
&\qquad+\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[-\frac{1}{2}\sum_{i=1}^N \widehat U''(t,\lambda_i)\right]\,.
\end{align}
Next, to control the second and third term on the right side of~\eqref{entropy estimate 1}, respectively the right side of~\eqref{neue 35}, we proceed as follows. We expand the potential terms $\widehat U'(t,\lambda_i)$, respectively $\widehat U''(t,\lambda_i)$, in Taylor series in~$\lambda_i$ to second order around the classical location $\widehat\gamma_i$. The resulting zero order terms cancel exactly since the classical locations of the ensembles $f_t\,\mu_G$ and $\widehat\psi_t\,\mu_G$ agree by construction. The first order terms in the Taylor expansion can (1) either be bounded in terms of the expectations of $\sum_{i=1}^N{(\lambda_i-\widehat\gamma_i)^2}$ (which can be controlled with the rigidity estimates in Lemma~\ref{rigidity for time dependent beta} and Lemma~\ref{lemma master rigidity bound}); or (2) they cancel exactly due to the definition of the potential $\widehat U(t,\cdot)$ and the equations of motion for~$\widehat\gamma_i(t)$ in~\eqref{EoM classical locations}. Finally, the second order
terms in the Taylor expansion can be bounded by the rigidity estimates in Lemma~\ref{rigidity for time dependent beta} and Lemma~\ref{lemma master rigidity bound}. The details are as follows.
Expanding $\partial_t \widehat U(t,\lambda_i)$ to second order around $\widehat\gamma_i$, we obtain from~\eqref{neue 1} that
\begin{align}\label{neue 2}
-\int \widehat g_t\partial_t\widehat{\psi}_t\d\mu_G&=\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[\frac{\beta N}{2}\sum_{i=1}^N\partial_t \widehat U(t,\widehat\gamma_i)+\frac{\beta N}{2}\sum_{i=1}^N\partial_t \widehat U'(t,\gamma_i)(\lambda_i-\widehat\gamma_i)\right]+{\mathcal O}(N^{1-2\mathfrak{a}})\,,
\end{align}
on $\Omega_V$, where we used the rigidity estimates in Lemma~\ref{rigidity for time dependent beta} and Lemma~\ref{lemma master rigidity bound}, and that $\partial_t\widehat U''(t,\cdot)$ is uniformly bounded on compact sets by Lemma~\ref{superlemma}.
To save notation, we introduce a function $G\,:\,\ensuremath{\mathbb{R}}^+\times \ensuremath{\mathbb{R}}^2\to \ensuremath{\mathbb{R}}$ by setting
\begin{align}\label{le G}
G(t;x,y)\mathrel{\mathop:}=\frac{\widehat U'(t,x)-\widehat U'(t,y)}{x-y}\,.
\end{align}
Note that $G(t;x,y)=G(t;y,x)$ and that $G$ is $C^2$ in the spatial coordinates by Lemma~\ref{superlemma}. Recalling the equation of motion for $\partial_t \widehat U(t,\cdot)$ in~\eqref{EoM drei} we can write
\begin{align}
\partial_t \widehat U(t,x)=\Xint-\frac{\widehat U'(t,y)\mathrm{d}\widehat\rho_{fc}(t,y)}{y-x}=\int\frac{\widehat U'(t,y)-\widehat U'(t,x)}{y-x}\mathrm{d}\widehat\rho_{fc}(t,y)+ \widehat U'(t,x)\Xint-\frac{\mathrm{d}\widehat\rho_{fc}(t,y)}{y-x}\,,
\end{align}
for $x$ inside the support of the measure $\widehat\rho_{fc}$. Thus, recalling~\eqref{definition widehat U} and~\eqref{le G}, we obtain
\begin{align}
\partial_t \widehat U(t,x)=\int G(t;x,y)\,\mathrm{d}\widehat\rho_{fc}(y)-\frac{1}{2}\widehat U'(t,x)\left(\widehat U'(t,x)+x \right) \,,
\end{align}
for $x$ inside the support of the measure $\widehat\rho_{fc}$.
We hence obtain from~\eqref{neue 2} that
\begin{align}\label{neue 36}
-\int \widehat g_t\partial_t\widehat{\psi}_t\d\mu_G&=\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[\frac{\beta N}{2}\sum_{i=1}^N\int G'(t;\widehat\gamma_i,y)\mathrm{d}\widehat\rho_{fc}(y)(\lambda_i-\widehat\gamma_i)\right]\nonumber\\
&\qquad-\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[\frac{\beta N}{4}\sum_{i=1}^N\widehat U''(t,\widehat\gamma_i)\left(\widehat U'(t,\widehat\gamma_i)+\widehat\gamma_i \right)(\lambda_i-\widehat\gamma_i)\right]\nonumber\\
&\qquad-\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[\frac{\beta N}{4}\sum_{i=1}^N\widehat U'(t,\widehat\gamma_i)\left(\widehat U''(t,\widehat\gamma_i)+1 \right)(\lambda_i-\widehat\gamma_i)\right]\nonumber\\
&\qquad+{\mathcal O}(N^{1-2\mathfrak{a}})\,,
\end{align}
on $\Omega_V$, where we denote by $G'(t;x,y)$ the first derivative of $G(t;x,y)$ with respect to $x$.
Next we return to~\eqref{neue 35}. Using the rigidity estimates of the Lemmas~\ref{rigidity for time dependent beta} and~\ref{lemma master rigidity bound}, we find
\begin{align}\label{laufet}
\frac{1}{2}\int\sum_{i=1}^N\widehat U'(t,\lambda_i)(\partial_i\widehat g_t(\boldsymbol{\lambda}))\mathrm{d}\widehat\omega_t(\boldsymbol{\lambda})&=\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[-\frac{1}{2}\sum_{i=1}^N \widehat U''(t,\widehat\gamma_i(t))\right]\nonumber\\&\qquad+\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[\frac{\beta N}{4}\sum_{i=1}^N \widehat U'(t,\lambda_i)\left(\widehat U'(t,\lambda_i)+\lambda_i-\frac{2}{N}\sum_{j}^{(i)}\frac{1}{\lambda_i-\lambda_j}\right)\right]\nonumber\\
&\qquad+{\mathcal O}\left(N^{1/2-\mathfrak{a}}\right)\,,
\end{align}
on $\Omega_V$, where we used a Taylor expansion of the first term on the right side of~\eqref{neue 35}. Here we also used that $\widehat U^{'}$ is three times continuously differentiable with uniformly bounded derivatives on compact sets. Note that the first term on the right side of~\eqref{laufet} vanishes.
Using the definition of the function $G(t;\cdot,\cdot)$ in~\eqref{le G} we can recast~\eqref{laufet} as
\begin{align}\label{neue}
\frac{1}{2}\int\sum_{i=1}^N\widehat U'(t,\lambda_i)(\partial_i\widehat g_t(\boldsymbol{\lambda}))\mathrm{d}\widehat\omega_t(\boldsymbol{\lambda})&=\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[\frac{\beta N}{4}\sum_{i=1}^N \widehat U'(t,\lambda_i)\left(\widehat U'(t,\lambda_i)+\lambda_i\right)\right]\nonumber\\
&\qquad+\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[-\frac{\beta N}{4}\sum_{i=1}^N\frac{1}{N}\sum_{j}^{(i)}G(t;\lambda_i,\lambda_j)\right]\nonumber\\
&\qquad+{\mathcal O}\left(N^{1/2-\mathfrak{a}}\right)\,,
\end{align}
where we used the symmetry $G(t;x,y)=G(t;y,x)$. Expanding the second term on the right side~\eqref{neue} to second order in $(\lambda_i,\lambda_j)$ around $(\widehat\gamma_i,\widehat\gamma_j)$, we obtain
\begin{align}
\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[-\frac{\beta N}{4}\sum_{i=1}^N\frac{1}{N}\sum_{j}^{(i)}G(t;\lambda_i,\lambda_j)\right]&=
\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[-\frac{\beta N}{2}\sum_{i=1}^N\left(\frac{1}{N}\sum_{j=1}^{N}G'(t;\widehat\gamma_i,\widehat\gamma_i)\right)(\lambda_i-\widehat\gamma_i)\right]\nonumber\\&\qquad+{\mathcal O}(N^{1/2-\mathfrak{a}})+{\mathcal O}(N^{1-2\mathfrak{a}})\,,
\end{align}
on $\Omega_V$, where we used the symmetry $G(t;x,y)=G(t;y,x)$, $G(t;x,x)=\widehat U''(t,x)$ and that $G(t;x,y)$ is $C^2$ in the spatial variables. Thus, also expanding the first term on the right side of~\eqref{neue} in $\lambda_i$ around $\widehat\gamma_i$ we obtain
\begin{align}\label{neue 37}
\frac{1}{2}\int\sum_{i=1}^N\widehat U'(t,\lambda_i)(\partial_i\widehat g_t(\boldsymbol{\lambda}))\mathrm{d}\widehat\omega_t(\boldsymbol{\lambda})&=
\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[-\frac{\beta N}{2}\sum_{i=1}^N\left(\frac{1}{N}\sum_{j}G'(t;\widehat\gamma_i,\widehat\gamma_j)\right)(\lambda_i-\widehat\gamma_i)\right]\nonumber\\
&\qquad+\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[\frac{\beta N}{4}\sum_{i=1}^N \widehat U''(t,\widehat\gamma_i)\left(\widehat U'(t,\widehat\gamma_i)+\widehat\gamma_i\right)(\lambda_i-\widehat\gamma_i)\right]\nonumber\\
&\qquad+\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[\frac{\beta N}{4}\sum_{i=1}^N \widehat U'(t,\widehat\gamma_i)\left(\widehat U''(t,\widehat\gamma_i)+1\right)(\lambda_i-\widehat\gamma_i)\right]\nonumber\\
&\qquad+{\mathcal O}\left(N^{1/2-\mathfrak{a}}\right)+{\mathcal O}\left({N^{1-2\mathfrak{a}}}\right)\,,
\end{align}
on $\Omega_V$, where we used the rigidity estimates in Lemma~\ref{rigidity for time dependent beta} and Lemma~\ref{lemma master rigidity bound}.
Adding up~\eqref{neue 36} and~\eqref{neue 37}, we hence obtain
\begin{align}
\left|\int\f{(\mathcal{L} - \partial_t)\widehat{\psi}_t}{\widehat{\psi}_t} f_t \,\d\mu_G\right|&\le
\left|\left[\E^{f_t\,\mu_G}-\E^{\widehat\psi_t\,\mu_G}\right]\left[-\frac{\beta N}{2}\sum_{i=1}^N\left(\frac{1}{N}\sum_{j=1}^NG'(t;\widehat\gamma_i,\widehat\gamma_j)-\int G'(t;\widehat\gamma_i,y)\,\mathrm{d}\widehat\rho_{fc}(y)\right)(\lambda_i-\widehat\gamma_i)\right]\right|\nonumber\\
&\qquad+{\mathcal O}(N^{1/2-\mathfrak{a}})+{\mathcal O}(N^{1-2\mathfrak{a}})\,,
\end{align}
on $\Omega_V$. To finish the proof we observe that, for all $\widehat\gamma_i$,
\begin{align*}
\frac{1}{N}\sum_{j=1}^{N} G'(t;\widehat\gamma_i,\widehat\gamma_j)=\int G'(t;\widehat\gamma_i,y)\,\mathrm{d}\widehat\rho_{fc}(t,y)+{\mathcal O}(N^{-1})\,,
\end{align*}
on $\Omega_V$, where we used that $\widehat\gamma_{i+1}-\widehat\gamma_{i}\sim N^{-2/3}\widehat\alpha_i^{-1}$, ($\widehat\alpha_i=\min\{i,N-i\}$), and the square root decay of $\widehat\rho_{fc}(t)$ at the edges of the support. Thus,
\begin{align}\label{freudig}
\int\frac{({\mathcal L}-\partial_t)\widehat\psi_t}{\widehat\psi_t}f_t\,\mathrm{d}\mu_G&={\mathcal O}\left(N^{1/2-\mathfrak{a}}\right)+{\mathcal O}(N^{1-\mathfrak{a}})\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$, where we used one last time the rigidity estimates. Using that $N^{1/2-\mathfrak{a}}< N^{1-2\mathfrak{a}}$, $\mathfrak{a}\in(0,1/2)$, we get from~\eqref{landau ginzburg formula},~\eqref{landau ginzburg lemma 0} and~\eqref{freudig} the desired estimate~\eqref{the super proposition bound}.
\end{proof}
Before we move on to the proof of Corollary~\ref{lemma bound on entroyp an dirichlet form}, we give a rough estimate on $S_{\widehat\omega_t}(\widehat g_t)$ for $t> 0$.
\begin{lemma} \label{basic entropy bound}
There is a constant $m$ such that, for $\tau>0$ and $t\ge\tau$, we have
\begin{align}\label{basic entropy bound eq}
S_{\widehat\omega_t}(\widehat g_t)=S(f_t\,\mu_G|\widehat{\psi}_t\,\mu_G) \leq CN^m
\end{align}
on $\Omega_V$, for $N$ sufficiently large. Here the constant $C$ depends on $\tau$.
\end{lemma}
\begin{proof}
From the definition of the relative entropy in~\eqref{definition of relativ entropy}, we have
\begin{align}\label{basic entropy bound I}
S(f_t\,\mu_G|\widehat{\psi}_t\,\mu_G)&\le S(f_t\,\mu_G|\mu_G)+\left|\frac{\beta N}{2}\sum_{i=1}^N\int \widehat U(t,\lambda_i) f_t(\boldsymbol{\lambda})\d\mu_G(\boldsymbol{\lambda})\right|+\log Z_{\widehat\psi_t}\,.
\end{align}
Since the potential $\widehat U(t)$ is bounded below, we have (for $N$ sufficiently large on $\Omega_V$) $\log Z_{{\widehat\psi_t}}\le C\beta N^2$. Similarly, using the rigidity estimate~\eqref{master rigidity bound}, we can bound the second term on the right side of~\eqref{basic entropy bound I} by $CN^2$. To bound the first term on the right of~\eqref{basic entropy bound I}, we use that $S(f_t\,\mu_G|\mu_G)\le S(H_t| W')\le N^2\max S(h_{ij,t}| w'_{ij})+N\max S(h_{ii,t}| w'_{ii})$, where $(h_{ij,t})$ are the entries of the in~\eqref{the new matrix} and~$w'_{ij}$ are the entries of the GOE, respectively GUE, matrix ${W'}$. By explicit calculations, remembering that the diagonal entries $(v_i)$ are fixed, one finds $\max S(h_{ij,t}| g_{ij})\le CN$ for $t\ge\tau$; see e.g.,~\cite{EKYY2}. (Note that we choose $t>0$, otherwise the relative entropy may be ill-defined).
\end{proof}
{
To complete the proof of Corollary~\ref{beta dirichlet bound} we follow the discussion in~\cite{EY}.
\begin{proof}[Proof of Corollary~\ref{beta dirichlet bound}]
Using an approximation argument, we can assume that $\widetilde g_t\in \mathrm{L}^{\infty}(\mathrm{d}\widetilde\omega_t)$. Using first the entropy bound~\eqref{the super proposition bound} and then the second Dirichlet form estimate in~\eqref{dirichlet equivalence}, we obtain
\begin{align}
\partial_t S_{\widehat\omega_t}(\widehat g_t)&\le -4D_{\widehat\omega_t}(\sqrt{\widehat g_t})+C N^{1-2\mathfrak{a}}\nonumber\\
&\le-2D_{\widetilde\omega_t}(\sqrt{\widetilde g_{t}})+C N^{1-2\mathfrak{a}}+C \frac{N^{1-2\mathfrak{a}}}{\tau^2}\nonumber\\
&\le-C\tau^{-1} S_{\widetilde\omega_t}(\widetilde g_t)+C \frac{N^{1-2\mathfrak{a}}}{\tau^2}\,,\nonumber
\end{align}
for $N$ sufficiently large on $\Omega_V$. To get the third line we used the logarithmic Sobolev inequality~\eqref{log sobolev} and that, by assumption, $\tau<1$. Using the entropy estimate~\eqref{entropy equivalence}, we thus obtain
\begin{align}\label{to be integrated}
\partial_t S_{\widehat\omega_t}(\widehat g_t)&\le -C\tau^{-1} S_{\widehat\omega_t}(\widehat g_t)+C \frac{N^{1-2\mathfrak{a}}}{\tau^2}\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$. Integrating~\eqref{to be integrated} from $\tau$ to $t/2$, we infer
\begin{align*}
S_{\widehat\omega_{t/2}}(\widehat g_{t/2})\le \e{-C\tau^{-1}(t/2-\tau)}S_{\widehat\omega_{\tau}}(\widehat g_{\tau})+C\frac{N^{1-2\mathfrak{a}}}{\tau}\,,
\end{align*}
for $N$ sufficiently large on $\Omega_V$. Bounding $S_{\widehat\omega_{\tau}}(\widehat g_{\tau})$ by the rough estimate~\eqref{basic entropy bound I}, we get
\begin{align*}
S_{\widehat\omega_{t/2}}(\widehat g_{t/2})\le CN^m\e{-C\tau^{-1}(t/2-\tau)}+C{N^{1-2\mathfrak{a}}}{\tau}\,,
\end{align*}
for $N$ sufficiently large on $\Omega_V$. Recalling that $t\ge \tau_0=\tau N^{\epsilon'}$, we obtain the first inequality in~\eqref{bound on entropy and dirichlet form} using the monotonicity of the relative entropy.
Integrating~\eqref{the super proposition bound} from $t/2$ to $t$, we obtain
\begin{align*}
\int_{t/2}^{t} D_{\widehat\omega_s}(\sqrt{\widehat g_s})\,\mathrm{d} s &\le- \int_{t/2}^{t}\partial_sS_{\widehat\omega_s}(\widehat g_s)\mathrm{d} s+ Ct{N^{1-2\mathfrak{a}}}\,.
\end{align*}
Thus, using the above estimate on the relative entropy and the monotonicity of the Dirichlet form,
\begin{align*}
D_{\widehat\omega_t} (\sqrt{\widehat g_t})\le C\frac{N^{1-2\mathfrak{a}}}{t \tau}+C{N^{1-2\mathfrak{a}}}\,.
\end{align*}
Recalling that $t\ge\tau_0=N^{\epsilon'}\tau$, we get the second inequality in~\eqref{bound on entropy and dirichlet form}.
\end{proof}
\section{Local equilibrium measures}\label{local equilibrium measures}
The main results of this section, Theorem~\ref{main theorem of section 6} and Theorem~\ref{corollary gaussian}, compare the averaged local gap statistics of $f_t\,\mu_G$ with the averaged local gap statistics of the $\beta$-ensembles ${\widehat\psi_t}\,\mu_G$ and $\mu_G$.
Recall that a symmetric function $O\,:\, \ensuremath{\mathbb{R}}^n\to \ensuremath{\mathbb{R}}$, $n\in\N$, is an $n$-particle observable if $O$ is smooth and compactly supported. For a given observable $O$, a time $t\ge0$, a small constant $\alpha>0$ and $j\in\llbracket \alpha N,(1-\alpha) N\rrbracket$, we define an observable $G_{j,n,t}(\mathbf{x})\equiv G_{j,n}(\mathbf{x})$, by setting
\begin{align}\label{the averaged observable}
G_{j,n}(\mathbf{x})\mathrel{\mathop:}= O(N\rho_{j}(x_{j+1}-x_j),N\rho_{j}(x_{j+2}-x_j),\ldots, N\rho_{j}(x_{j+n+1}-x_{j}))\,,\qquad(\mathbf{x}=(x_k)_{k=1}^N\in\digamma^{(N)})\,,
\end{align}
where we set $G_{j,n}=0$ if $j+n>(1-\alpha) N$. Here $\rho_j$ denotes the density of the measure $\widehat\rho_{fc}(t)$ at the classical location of the $j$-th particle at time $t$, i.e., $\rho_j=\widehat\rho_{fc}(t,\widehat\gamma_j(t))$. We also set
\begin{align}\label{the averaged observable at infty}
G_{j,n,sc}(\mathbf{x})\mathrel{\mathop:}= O(N\rho_{sc,j}(x_{j+1}-x_j),N\rho_{sc,j}(x_{j+2}-x_j),\ldots, N\rho_{sc,j}(x_{j+n+1}-x_{j}))\,,\qquad (\mathbf{x}\in\digamma^{(N)})\,,
\end{align}
where $\rho_{sc,j}$ denotes the density of the semicircle law at the classical location of the $j$-th particle with respect to the semicircle law.
In the following, we denote constants depending on $O$ by $C_O$.
Recall the definition of the density~$\widehat\psi_t$ in~\eqref{definition of bt}. We have the following statement on the averaged local gap statistics.
\begin{theorem}\label{main theorem of section 6}
Let $n\in\N$ be fixed and consider an $n$-particle observable $O$. Fix a small constant $\alpha>0$ and consider an interval of consecutive integers $J\subset\llbracket \alpha N,(1-\alpha)N\rrbracket$ in the bulk. Then, for any small $\delta>0$, there is a constant $\mathfrak{f}>0$ such that, for $t\ge N^{-1/2+\delta}$,
\begin{align}\label{equation in main theorem of sectoin 6}
\left|\int \frac{1}{|J|}\sum_{j\in J}G_{j,n}(\mathbf{x}) \,f_t(\mathbf{x})\mathrm{d}\mu_G(\mathbf{x})-\int \frac{1}{|J|}\sum_{j\in J}G_{j,n}(\mathbf{x})\,{\widehat\psi_t}(\mathbf{x})\mathrm{d}\mu_G(\mathbf{x})\right|\le C_ON^{-\mathfrak{f}}\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$. The constant $C_O$ depends on $\alpha$ and $O$, and the constant $\mathfrak{f}$ depends on $\alpha$ and $\delta$.
\end{theorem}
We can also compare the averaged local gap statistics of $f_t\,\mu_G$, with the averaged local gap statistics of the Gaussian unitary, respectively orthogonal, ensemble.
\begin{theorem}\label{corollary gaussian}
Under the same assumptions as in Theorem~\ref{the averaged observable} and with similar constants, we have
\begin{align}
\left|\int \frac{1}{|J|}\sum_{j\in J}G_{j,n}(\mathbf{x}) f_t(\mathbf{x})\mathrm{d}\mu_G(\mathbf{x})-\int\frac{1}{|J|}\sum_{j\in J}G_{j,n,sc}(\mathbf{x})\mathrm{d}\mu_G(\mathbf{x})\right|\le C_O N^{-\mathfrak{f}}\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$.
\end{theorem}
To prove Theorem~\ref{main theorem of section 6} and Theorem~\ref{corollary gaussian}, we localize the measures $f_t\,\mu_G$ and $\widehat\psi_t\,\mu_G$ as is explained in the next subsections.
We then use Theorem~4.1 of~\cite{single_gap} to compare the gap statistics of these localized measures. The technical input we are using are the Dirichlet form estimates established in Corollary~\ref{lemma bound on entroyp an dirichlet form}. Alternatively, one could combine the approach of~\cite{EY} with Theorem~2.1 in~\cite{BEY} (see Theorem~\ref{thm: bulk universality beta} above), to prove Theorem~\ref{main theorem of section 6} and Theorem~\ref{corollary gaussian}.
To conclude, we remark that once the entropy estimate of Proposition~\ref{the super proposition} has been established, one can apply the methods of~\cite{single_gap} to prove the gap universality in the bulk for deformed Wigner matrices; see Remark~\ref{remark for gap universality} above for an explicit statement; we leave the details to the interested readers.
\subsection{Preliminaries}\label{subsection 1 of section 6}
Let $\alpha,\,\sigma>0$ be two small positive numbers and choose two integer parameters~$L$ and~$K$ such that
\begin{align}\label{condition on K}
L\in\llbracket \alpha N,(1-\alpha)N\rrbracket\,,\qquad\quad K\in\llbracket N^{\sigma}, N^{1/4}\rrbracket\,.
\end{align}
We denote by $ I_{L,K}=\llbracket L-K,L+K\rrbracket$ a set of ${\mathcal K}\mathrel{\mathop:}= 2K+1$ consecutive indices in the bulk of the spectrum. Below we often abbreviate $I\equiv I_{L,K}$. Recall the definition of the set $\digamma^{(N)}\subset\ensuremath{\mathbb{R}}^N$ in~\eqref{Weyl chamber}. For $\boldsymbol{\lambda}\in\digamma^{(N)}$, we write
\begin{align}\label{the bflambda}
\boldsymbol{\lambda}=(\lambda_1,\lambda_2,\ldots,\lambda_N)=(y_1,\ldots, y_{L-K-1},x_{L-K},\ldots, x_{L+K},y_{L+K+1},\ldots ,y_{N})\,,
\end{align}
and we call $\boldsymbol{\lambda}$ a configuration (of $N$ particles or points on the real line). Note that on the right side of~\eqref{the bflambda} the points keep their original indices and are ordered in increasing order,
\begin{align}\label{bfx notation}
\mathbf{x}=(x_{L-K},\ldots, x_{K+L})\in\digamma^{({\mathcal K})}\,,\qquad\mathbf{y}=(y_1,\ldots, y_{L-K-1},y_{L+K+1},\ldots, y_N)\in \digamma^{(N-{\mathcal K})}\,.
\end{align}
We refer to $\mathbf{x}$ as the interior points or particles and to $\mathbf{y}$ as the exterior points or particles.
In the following, we often fix the exterior points and consider the conditional measures on the interior points: Let~$\omega$ be a measure with density on $\digamma^{(N)}$. Then we denote by $\omega^{\mathbf{y}}$ the measure obtained by conditioning on $\mathbf{y}$, i.e., for $\boldsymbol{\lambda}$ of the form~\eqref{the bflambda},
\begin{align*}
\omega^{\mathbf{y}}(\mathrm{d}\mathbf{x})\equiv \omega^{\mathbf{y}}(\mathbf{x})\mathrm{d}\mathbf{x}\mathrel{\mathop:}=\frac{\omega(\boldsymbol{\lambda})\mathrm{d}\mathbf{x}}{\int\omega(\boldsymbol{\lambda})\mathrm{d} \mathbf{x}}=\frac{\omega(\mathbf{x},\mathbf{y})\mathrm{d} \mathbf{x}}{\int \omega(\mathbf{x},\mathbf{y})\mathrm{d}\mathbf{x}}\,,
\end{align*}
where, with slight abuse of notation, $\omega(\mathbf{x},\mathbf{y})$ stands for $\omega(\boldsymbol{\lambda})$. We refer to the fixed exterior points $\mathbf{y}$ as boundary conditions of the measure $\omega^{\mathbf{y}}$. For fixed $\mathbf{y}\in\digamma^{(N-{\mathcal K})}$, all $(x_i)$ lie in the open configuration interval
\begin{align*}
\boldsymbol{I}\equiv\boldsymbol{I}_{L,K}\mathrel{\mathop:}=(y_{L-K-1},y_{L+K+1})\,.
\end{align*}
Set $\bar{y} \mathrel{\mathop:}= (y_{L-K-1}+y_{L+K+1})/2$ and let
\begin{align}\label{equidistant points 1}
\alpha_j\mathrel{\mathop:}=\bar y+\frac{j-L}{{\mathcal K}+1}|\mathbf{y}|\,,\qquad\quad (j\in I_{L,K})\,,
\end{align}
denote ${\mathcal K}$ equidistant points in the interval ${\boldsymbol I}$.
Let $U\in C^4(\ensuremath{\mathbb{R}})$ be a ``regular'' potential satisfying~\eqref{assumption 1 for beta universality} and~\eqref{assumption 2 for beta universality}. We then consider the $\beta$-ensemble
\begin{align}\label{beta ensemble section 6}
\mu(\mathrm{d}\boldsymbol{\lambda})\equiv \mu_U (\mathrm{d} \boldsymbol{\lambda}) \mathrel{\mathop:}=\frac{1}{Z_{U}}\e{-\beta N {\mathcal H}(\boldsymbol{\lambda}) }\mathrm{d}\boldsymbol{\lambda}\,,\qquad\quad ( \beta>0)\,,
\end{align}
with, (c.f.,~\eqref{eqn:measure}),
\begin{align*}
{\mathcal H}(\boldsymbol{\lambda}) \mathrel{\mathop:}=\sum_{i=1}^N\frac{1}{2}\left(U(\lambda_i)+\frac{\lambda_i^2}{2}\right)-\frac{1}{N}\sum_{1\le i<j\le N}\log |\lambda_j-\lambda_i|\,,
\end{align*}
and with $Z_U\equiv Z_U(\beta)$ a normalization. For $K,L$ and $\mathbf{y}$ fixed, we can write $\mu^{\mathbf{y}}$ as the Gibbs measure
\begin{align*}
\mu^{\mathbf{y}}(\mathrm{d}\mathbf{x})=\frac{1}{Z_U^{\mathbf{y}}}\e{-\beta N{\mathcal H}^{\mathbf{y}}(\mathbf{x})}\mathrm{d}\mathbf{x}\,,\qquad\qquad{\mathcal H}^{\mathbf{y}}(\mathbf{x})&=\sum_{i\in I}\frac{1}{2} V^{\mathbf{y}}(x_i)-\frac{1}{N}\sum_{\substack{i,j\in I\\ i<j }}\log |x_j-x_i|\,,
\end{align*}
with
\begin{align}
V^{\mathbf{y}}(x)&= U(x)+\frac{x^2}{2}-\frac{2}{N}\sum_{i\not\in I}\log |x-y_i|\,,\label{equation external potential}
\end{align}
and with $Z_U^{\mathbf{y}}\equiv Z_U^{\mathbf{y}}(\beta)$ a normalization. We refer to $V^{\mathbf{y}}$ as an external potential. Following~\cite{single_gap} we introduce the definition of regular external potential as follows.
\begin{definition}\label{definition regular external potential}
An external potential $V\equiv V^{\mathbf{y}}$ in a $\beta$-ensemble of $K$ points in a configuration interval $\boldsymbol I=(a,b)$ is called $K^{\chi}$-regular, if the following bounds hold:
\begin{align}
|\boldsymbol{I}|&=\frac{{\mathcal K}}{N\rho(\overline y)}+{\mathcal O}\left(\frac{K^{\chi}}{N}\right)\label{external potential 1}\,,\\
V'(x)&=\rho(\overline y)\log\frac{d_+(x)}{d_-(x)}+{\mathcal O}\left(\frac{K^{\chi}}{Nd(x)}\right)\label{external potential 2}\,,\\
V''(x)&\ge 1+\inf U''(x)+\frac{c}{d(x)}\,, \label{external potential 3}
\end{align}
for $x\in\boldsymbol{I}$, with some $c>0$ and for some small $\chi>0$, where
\begin{align*}
d(x)\mathrel{\mathop:}=\min\{|x-a|,|x-b|\}\,
\end{align*}
denotes the distance to the boundary of $\boldsymbol{I}$ and
\begin{align*}
d_-(x)\mathrel{\mathop:}= d(x)+\rho(\overline{y})N^{-1}K^{\chi}\,,\qquad\qquad d_+(x)\mathrel{\mathop:}= \max\{|x-a|,|x-b|\}+\rho(\overline{y})N^{-1}K^{\chi}\,.
\end{align*}
\end{definition}
The main technical result we use in this section is Theorem~4.1 of~\cite{single_gap}; see Theorem~\ref{theorem 3.1} below. It asserts that the local gap statistics of $\mu^{\mathbf{y}}$ are essentially independent of $\mathbf{y}$ and $U$, provided that $V^{\mathbf{y}}$ is $K^{\chi}$-regular for some small $\chi>0$.
\subsection{Comparison of local measures}
Fix small $\alpha,\sigma>0$ and let $K$ and $L$ satisfy~\eqref{condition on K}. Recall that we denote by~$f_t\,\mu_G$ the distribution of the eigenvalues of the matrix in~\eqref{the new matrix} and by~$\widehat\psi_t\,\mu_G$ the reference $\beta$-ensemble defined in~\eqref{definition of psit}. Following the discussion of the previous subsection we introduce the conditioned densities
\begin{align}
f_t^{\mathbf{y}}\,\mu_G^{\mathbf{y}}=(f_t\,\mu_G)^{\mathbf{y}}\,,\qquad\qquad\widehat\psi_t^{\mathbf{y}}\,\mu_G^{\mathbf{y}}=(\widehat\psi_t\,\mu_G)^{\mathbf{y}}\,.
\end{align}
Recall that we denote by $\widehat\rho_{fc}(t)$ the equilibrium density of $\widehat\psi_t\,\mu_G $ and by $\widehat\gamma_k\equiv\widehat\gamma_k(t)$ the classical location of the $k$-th particle with respect to $\widehat\rho_{fc}\equiv\widehat\rho_{fc}(t)$; c.f.,~\eqref{classical location}. Let $\epsilon_0>0$ and define the set of ``good'' boundary conditions
\begin{align}\label{definition of R simple}
{\mathcal R}_{L,K}\equiv{\mathcal R}_{L,K}(\epsilon_0,\alpha)\mathrel{\mathop:}=&\{ {\boldsymbol{\lambda}}\in \digamma^{(N)}\,:\, |\lambda_{k}-\widehat\gamma_{k}| \leq{N^{-1+\epsilon_0}}\,,\forall k\in\llbracket \alpha N,(1-\alpha) N\rrbracket\backslash I_{L,K}\}\nonumber\\ &\cap\{ {\boldsymbol{\lambda}}\in \digamma^{(N)}\,:\, |\lambda_{k}-\widehat\gamma_{k}|\le N^{-2/3+\epsilon_0}\,,\forall k\in\llbracket 1, N\rrbracket\}\,.
\end{align}
The next result compares the local statistics of $ f_t^{\mathbf{y}}\,\mu_G^{\mathbf{y}}$ and $\widehat\psi_t^{\mathbf{y}}\,\mu_G^{\mathbf{y}}$ for $\mathbf{y}\in{\mathcal R}_{L,K}$. Recall that $\mathfrak{a}$ stands for any number in $(0,1/2)$.
\begin{proposition}\label{frozen stats proximity}
Fix small constants $\alpha,\sigma>0$ (see~\eqref{condition on K}) and $\epsilon_0>0$ (see~\eqref{definition of R simple}). Let $K$ satisfy~\eqref{condition on K} and let $O$ be an $n$-particle observable. Let $\epsilon'>0$ and choose $\tau$ satisfying $1\gg\tau>N^{-2\mathfrak{a}}$. Then, for any $t\ge N^{\epsilon'}\tau$ and any constant $\mathfrak{c}\in(0,1)$, there is a set of configurations $\mathcal{G}\equiv \mathcal{G}_{L,K}(\epsilon_0,\alpha) \subset {\mathcal R}_{L,K}(\epsilon_0,\alpha)$, with
\begin{align} \label{probability estimate on good boundary condition}
{\mathbb P}^{f_t \mu_G}(\mathcal{G}) \geq 1 - \frac{N^{-\mathfrak{c}}}{2}\,,
\end{align}
such that
\begin{align} \label{for almost all boxes}
\left| \int O(\mathbf{x}) \left(f_t^{{\mathbf{y}}}(\mathbf{x}) -{\widehat\psi_t^{\mathbf{y}}}(\mathbf{x}) \right) \mu_G^{\mathbf{y}}(\mathrm{d}\mathbf{x}) \right| \leq C_O \sqrt{K} N^{\mathfrak{c}-\mathfrak{a}}\tau^{-1}\,,\qquad\qquad (t\ge N^{\epsilon'}\tau)\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$. The constant~$C_O$, depends only on~$\epsilon'$,~$\alpha$ and $O$.
Moreover, there is $\upsilon>0$, such that
\begin{align} \label{local f_t rigidity}
{\mathbb P}^{f_t^{\mathbf{y}}\mu_G^{\mathbf{y}}}\Big(\left\{\left| x_k - \widehat\gamma_k(t) \right| < N^{-1+\epsilon_0}\,,\,k\in I_{L,K}\right\}\Big) \geq 1 - \e{-\upsilon (\varphi_N)^{\xi}}\,,\qquad\qquad (t\ge N^{\epsilon'}\tau)\,,
\end{align}
for $N$ sufficiently large on~$\Omega_V$, where we have chosen $\xi=A_0\log\log N/2$; see~\eqref{eq.xi}.
\end{proposition}
\begin{proof}
We follow closely the proof of Lemma~6.4 in~\cite{single_gap}. Let $\tau$ satisfy $1\gg\tau> N^{-2\mathfrak{a}}$ and choose $t\ge N^{\epsilon'}\tau$. We estimate
\begin{align}\label{frozen stats bound}
\left|\int O(\mathbf{x}) \left(f_t^{{\mathbf{y}}}(\mathbf{x}) -{\widehat\psi_t^{\mathbf{y}}}(\mathbf{x}) \right) \mu_G^{{\mathbf{y}}}(\mathrm{d} \mathbf{x})\right| & \leq C_O\, \|f_t^{{\mathbf{y}}} \mu_G^{{\mathbf{y}}} - {\widehat\psi_t^{\mathbf{y}}} \mu_G^{{\mathbf{y}}}\|_{1}\nonumber\\ &\leq C_O\sqrt{S_{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}}(\widehat g_t^{{\mathbf{y}}})}\,,
\end{align}
where we used~\eqref{entropy inequality} and set $\widehat g_t\mathrel{\mathop:}= f_t/{\widehat\psi_t}$. For $\mathbf{y}\in{\mathcal R}_{L,K}$, we consider the locally-constrained measure ${\widehat\psi_t^{\mathbf{y}}}\mu_G^{\mathbf{y}}$, explicitly given by
\begin{align*}
{\widehat\psi_t^{\mathbf{y}}}\mu_G^{\mathbf{y}}(\mathrm{d} {\mathbf{x}}) = \frac{1}{Z^{\mathbf{y}}}\e{-N\beta\widehat{\mathcal H}^{\mathbf{y}}(t,\mathbf{x})}\mathrm{d} \mathbf{x}\,,
\end{align*}
with
\begin{align*}
\widehat{\mathcal H}^{\mathbf{y}}(t,\mathbf{x})=\sum_{k\in I}\left(\frac{\widehat U(t,x_k)}{2}+\frac{x_k^2}{4}\right)-\frac{1}{N}\sum_{\substack{k,l\in I\\ k<l}}\log |x_k-x_l|-\frac{1}{N}\sum_{\substack{k\in I\\ l\not\in I}}\log|x_k-y_l|\,.
\end{align*}
Here $I\equiv I_{L,K}$. From~(5.20) of~\cite{single_gap}, we know that
\begin{align}
\nabla^2_{\mathbf{x}} \widehat{\mathcal H}^{\mathbf{y}}(t,\mathbf{x})\ \geq cN/K\,,\qquad \qquad(\mathbf{y}\in{\mathcal R}_{L,K})\,,
\end{align}
for some $c>0$ independent of $t$. Here, $\nabla^2_{\mathbf{x}}$ denotes the Hessian with respect the variables $\mathbf{x}$. Thus the Bakry-\'{E}mery criterion yields the logarithmic Sobolev inequality
\begin{align} \label{frozen log-sobolev}
S_{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{{\mathbf{y}}}}(\widehat g_t^{{\mathbf{y}}}) \leq \f{CK}{N} \, D_{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{{\mathbf{y}}}}\Big(\sqrt{\widehat g_t^{\mathbf{y}}}\Big)\,,\qquad\qquad (\mathbf{y}\in{\mathcal R}_{L,K})\,,
\end{align}
where the constant $C$ can be chosen independent $t$.
For $k\in\llbracket 1,N\rrbracket$, denote by $D_{{\widehat\psi_t} \,\mu_G, k}$ the Dirichlet form of the particle $k$, i.e., $D_{{\widehat\psi_t}\, \mu_G, k} (f) \mathrel{\mathop:}= \f{1}{2N} \int |\ensuremath{\partial}_k f|^2 {\widehat\psi_t}\, \mu_G$, and by $ D_{{\widehat\psi_t^{\mathbf{y}}} \,\mu_G^{{\mathbf{y}}},k}$ its conditioned analogue (with $k\in I_{L,K}$). Using the notation of~\eqref{the bflambda}, we may write
\begin{align*}
\E^{f_t\,\mu_G}D_{{\widehat\psi_t^{\mathbf{y}}}\, \mu_G^{{{\mathbf{y}}}}}\left(\sqrt{\widehat g_t^{{{\mathbf{y}}}}}\right)=\int D_{\widehat\psi_t^{\mathbf{y}} \mu_G^{\mathbf{y}}}\left(\sqrt{\widehat g_t^{\mathbf{y}}}\right)f_t(\boldsymbol{\lambda})\,\mu_G(\mathrm{d}\boldsymbol{\lambda})
\end{align*}
and we can bound
\begin{align*}
\E^{f_t\,\mu_G} D_{{\widehat\psi_t^{\mathbf{y}}}\, \mu_G^{{{\mathbf{y}}}}}\left(\sqrt{\widehat g_t^{{{\mathbf{y}}}}}\right)& =\mathbb{E}^{f_t\mu_G} \sum_{k\in I} D_{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{{\mathbf{y}}},k}\left(\sqrt{\widehat g_t^{{\mathbf{y}}}}\right)
\\& \leq D_{{\widehat\psi_t} \mu_G}(\sqrt{\widehat g_t})
\\& \leq CN^{1-2\mathfrak{a}} \tau^{-2}\,,
\end{align*}
for $N$ sufficiently large, where we used Corollary~\ref{beta dirichlet bound} in the last line. Thus Markov's inequality implies, for $\mathfrak{c}>0$, that there exists a set of configurations $\mathcal{G}^1\subset \mathcal{R}$, with $\mathbb{P}^{f_t\,\mu_G}(\mathcal{G}^1)\ge 1-N^{-\mathfrak{c}}$, such that, for $\mathbf{y}\in{\mathcal G}^1$,
\begin{align} \label{good interval estimate}
D_{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{{\mathbf{y}}}}\left(\sqrt{\widehat g_t^{{\mathbf{y}}}}\right) \leq C N^{2\mathfrak{c}} N^{1-2\mathfrak{a}}\tau^{-2}\,,
\end{align}
holds for $N$ sufficiently large on $\Omega_V$. Substituting~\eqref{good interval estimate} into \eqref{frozen log-sobolev} and then into \eqref{frozen stats bound}, we find that
\begin{align*}
\left| \int O(\mathbf{x}) \left(f_t^{{\mathbf{y}}} -{\widehat\psi_t^{\mathbf{y}}} \right) \mu_G^{{\mathbf{y}}}(\mathrm{d}\mathbf{x})\right|&
\leq C_O \sqrt{K} N^{\mathfrak{c}}N^{-\mathfrak{a}}\tau^{-1}\,,
\end{align*}
on $\Omega_V$ for $N$ sufficiently large. This proves~\eqref{for almost all boxes}.
To prove~\eqref{local f_t rigidity} note that the rigidity estimates of Lemma~\ref{master rigidity bound} imply
\begin{align*}
{\mathbb E}^{f_t \mu_G} \left[ {\mathbb P}^{f_t^{\mathbf{y}} \mu_G^{\mathbf{y}}}\left(\left\{|x_k - \widehat{\gamma}_k(t)| > N^{-1+\epsilon}\,,\,k\in I\right\}\right) \right] = {\mathbb P}^{f_t \mu_G}\left(\left\{|x_k - \widehat{\gamma}_k(t)| > N^{-1+\epsilon}\,,\, k\in I\right\}\right) \leq \e{-\upsilon(\varphi_N)^{\xi}}\,,
\end{align*}
for some $\upsilon>0$, where we have chosen $\xi=A_0\log\log N/2$.
By Markov's inequality we conclude that there is a set of configurations,~${\mathcal G}^{2}$, such that~\eqref{local f_t rigidity} holds with $(\xi,\upsilon)$-high probability. Finally, set ${\mathcal G}\mathrel{\mathop:}= {\mathcal G}^1\cap {\mathcal G}^2$ and note that ${\mathcal G}$ satisfies~\eqref{probability estimate on good boundary condition}.
\end{proof}
\subsection{Gap universality for local measures}\label{gap universality for local measures}
In the previous subsection, we showed that the local gap statistics of the measure $f_t^{\mathbf{y}}\,\mu_G^{\mathbf{y}}$ agree with those of $\widehat\psi_t^{\mathbf{y}}\,\mu_G^{\mathbf{y}}$ for boundary conditions $\mathbf{y}$ in the set ${\mathcal R}_{L,K}$. In this subsection, we are going to show that the local statistics of $\widehat\psi_t^{\mathbf{y}}\,\mu_G^{\mathbf{y}}$ are essentially independent of the precise form of $\mathbf{y}$, as is asserted by the main theorem of~\cite{single_gap}. Recall the notion of external potential introduced in~\eqref{equation external potential}.
\begin{theorem}{\emph {[Gap universality for local measures, Theorem~4.1 in~\cite{single_gap}]}}\label{theorem 3.1}
Let $L,\tilde L$ and ${\mathcal K}=2K+1$ satisfy~\eqref{condition on K} with $\alpha,\sigma>0$. Consider two boundary conditions $\mathbf{y}$, $\tilde\mathbf{y}$ such that the configuration intervals coincide, i.e.,
\begin{align}\label{assumption 1 thm3.1}
{\boldsymbol I}=(y_{L-K-1},y_{L+K+1})=(\tilde{y}_{\tilde L-K-1},\tilde{y}_{\tilde L+K+1})\,.
\end{align}
Consider two measure $\mu$ and $\tilde\mu$ of the form~\eqref{beta ensemble section 6}, with possibly two different potentials $U$ and $\tilde U$, and consider the constrained measures $\mu^{\mathbf{y}}$ and $\tilde\mu^{\tilde\mathbf{y}}$. Let $\chi>0$ and assume that the external potentials $V^{\mathbf{y}}$ and $\tilde V^{\tilde \mathbf{y}}$ (see~\eqref{equation external potential}) are $K^{\chi}$-regular; see Definition~\ref{definition regular external potential}. In particular, assume that ${\boldsymbol I}$ satisfies
\begin{align}\label{assumption 2 thm3.1}
|{\boldsymbol I}|=\frac{{\mathcal K}}{N\varrho_U(\bar y)}+{\mathcal O}\left(\frac{K^{\chi}}{N} \right)\,.
\end{align}
Assume further that
\begin{align}\label{assumption 3 thm3.1}
\max_{j\in I_{L,K}}|\E^{\mu^{\mathbf{y}}} x_j-\alpha_j|+\max_{j\in I_{ \tilde L,K}}|\E^{\tilde\mu^{\tilde{\mathbf{y}}}} x_j-\alpha_j|\le CN^{-1}K^{\chi}\,.
\end{align}
Let $p\in\Z$ satisfy $|p|\le K-K^{1-\chi'}$, for some small $\chi'>0$. Fix $n\in\N$. Then there is a constant $\chi_0$, such that if $\chi,\chi'<\chi_0$, then for any $n$-particle observable $O$, we have
\begin{align*}
\big|\mathbb{E}^{\mu^{\mathbf{y}}}O\left(N(x_{L+p+1}-x_{L+p}),\ldots, N(x_{L+p+n}-x_{L+p})\right) -\mathbb{E}^{\tilde\mu^{\tilde\mathbf{y}}}O\left(N(x_{L+p+1}-x_{L+p}),\ldots, N(x_{L+p+n}-x_{L+p})\right) \big|\le C_O K^{-\mathfrak{b}} \,,
\end{align*}
for some constant $\mathfrak{b}>0$ depending on $\sigma$, $\alpha$, and for some constant $C_O$ depending on $O$. This holds for $N$ sufficiently large (depending on the $\chi,\chi',\alpha$ and $C$ in~\eqref{assumption 3 thm3.1}).
\end{theorem}
Recall that the measure $\widehat\psi_t^{\mathbf{y}}\,\mu_G^{\mathbf{y}}$ can be written as the Gibbs measure
\begin{align}
{\widehat\psi_t^{\mathbf{y}}}\mu_G^{\mathbf{y}}(\mathrm{d} {\mathbf{x}}) = \frac{1}{Z_{\widehat\psi_t}^{\mathbf{y}}}\e{-N\beta{\mathcal H}^{\mathbf{y}}(t,\mathbf{x})}\mathrm{d} \mathbf{x}\,,\qquad\qquad
{\mathcal H}^{\mathbf{y}}(t,\mathbf{x})&=\sum_{i\in I}\frac{1}{2} V^{\mathbf{y}}(t,x_i)-\frac{1}{N}\sum_{\substack{i,j\in I\\ i<j }}\log |x_j-x_i|\,,
\end{align}
with the external potential
\begin{align}\label{le potential exterior}
V^{\mathbf{y}}(t,x)&= \widehat U(t,x)+\frac{x^2}{2}-\frac{2}{N}\sum_{i\not\in I}\log |x-y_i|\,.
\end{align}
Using Theorem~\ref{theorem 3.1} we first show that the local statistics of ${\psi_t^{\mathbf{y}}}\mu_G^{\mathbf{y}}$ are virtually independent of $\mathbf{y}$, i.e., we apply Theorem~\ref{theorem 3.1} with $\mu^{\mathbf{y}}=(\widehat\psi_t\,\mu_G)^{\mathbf{y}}$ and $\tilde\mu^{\tilde\mathbf{y}}=(\widehat\psi_t\,\mu_G)^{\tilde\mathbf{y}}$.
We first check the regularity assumption of the external potential $V^{\mathbf{y}}$. Recall the definition of $K^{\chi}$-regular potential in Definition~\ref{definition regular external potential}.
\begin{lemma}\label{potential is regular}
Fix small constants $\alpha,\sigma>0$ (see~\eqref{condition on K}). Let $\chi>0$ and consider $\mathbf{y}\in{\mathcal R}_{L,K}(\chi\sigma/2,\alpha/2)$. Then, on the event $\Omega_V$, the external potential $V^{\mathbf{y}}(t,x)$ in~\eqref{le potential exterior} is $K^{\chi}$-regular on $\boldsymbol I=(y_{L-K-1},y_{L+K+1})$.
\end{lemma}
The proof of Lemma~\ref{potential is regular} follows almost verbatim the proof of Lemma~4.5 in the Appendix A of~\cite{single_gap} and we therefore omit it here.
To check that assumption~\eqref{assumption 3 thm3.1} of Theorem~\ref{theorem 3.1} holds, we use the following result.
\begin{lemma}\label{frozen beta particle proximity}
Under the assumptions of Proposition~\ref{frozen stats proximity} the following holds. Let $\mathbf{y}\in{\mathcal G}$. Then, for all $k\in I_{L,K}$,
\begin{align} \
\left| {\mathbb E}^{f_t^{\mathbf{y}}\mu_G^{\mathbf{y}}} x_k - {\mathbb E}^{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}} x_k \right|&\leq C \frac{KN^{2\mathfrak{c}}}{N} N^{-\mathfrak{a}}\tau^{-1} \,,\qquad\qquad(t\ge\tau N^{\epsilon'})\,,\label{frozen beta particle proximity equation}
\end{align}
for $N$ sufficiently large on $\Omega_V$.
\end{lemma}
\begin{proof}
We follow the proof of Lemma~6.5 of \cite{single_gap}. Fix $t\ge\tau N^{\epsilon'}$, where $1\gg\tau\ge N^{-2\mathfrak{a}}$. Let $\mathbf{y}\in{\mathcal G}$. Denote by $\mathcal{L}_t^{\mathbf{y}}$ the generator associated to the Dirichlet form $D_{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}}$, i.e.,
\begin{align*}
\int f \mathcal{L}_t^{\mathbf{y}}g\; {\widehat\psi_t^{\mathbf{y}}}\mathrm{d} \mu_G^{\mathbf{y}} = - \f{1}{\beta N} \sum_{i \in I} \int \ensuremath{\partial}_i f \ensuremath{\partial}_i g\; {\widehat\psi_t^{\mathbf{y}}}\mathrm{d}\mu_G^{\mathbf{y}}\,,\qquad \quad(I\equiv I_{L,K})\,.
\end{align*}
Let $q_s$ be the solution of the evolution equation $\ensuremath{\partial}_s q_s = \mathcal{L}_t^{\mathbf{y}} q_s$, $s\ge0$, with initial condition $q_0 \mathrel{\mathop:}= \widehat g_t^{\mathbf{y}} = f_t^{\mathbf{y}} / {\widehat\psi_t^{\mathbf{y}}}$. Note that $q_s$ is a density with respect to the reversible measure, ${\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}$, of this dynamics. Hence, we can write
\begin{align*}
\left| {\mathbb E}^{f_t^{\mathbf{y}}\mu_G^{\mathbf{y}}} x_k - {\mathbb E}^{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}} x_k \right| = \left| \int_0^\infty \mathrm{d} s \int x_k \mathcal{L}_t^{\mathbf{y}} q_s \, {\widehat\psi_t^{\mathbf{y}}}\mathrm{d} \mu_G^{\mathbf{y}} \right| = \left| \f{1}{\beta N} \int_0^{\infty} \mathrm{d} s \int \ensuremath{\partial}_k q_s \, {\widehat\psi_t^{\mathbf{y}}}\mathrm{d} \mu_G^{\mathbf{y}} \right| \,.
\end{align*}
Recall that ${\widehat\psi_t^{\mathbf{y}}}\mu ^{\mathbf{y}}_G$ satisfies the logarithmic Sobolev inequality~\eqref{frozen log-sobolev} with constant $\tau_K\mathrel{\mathop:}= CK/N$, provided that $\mathbf{y}\in{\mathcal R}_{L,K}$. Thus, upon using Cauchy-Schwarz and the exponential decay of the Dirichlet form $D_{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}}(\sqrt{q_s})$, we obtain for some~$\upsilon',c> 0$,
\begin{align*}
\left| {\mathbb E}^{f_t^{\mathbf{y}}\mu_G^{\mathbf{y}}} x_k - {\mathbb E}^{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}} x_k \right| = \left| \f{1}{\beta N} \int_0^{N^{\upsilon'} \tau_K} \mathrm{d} s \int \ensuremath{\partial}_k q_s \, {\widehat\psi_t^{\mathbf{y}}}\mathrm{d} \mu_G^{\mathbf{y}} \right| +{\mathcal O}\left(\e{-cN^{\upsilon'}}\right) \,.
\end{align*}
Using
\begin{align*}
|\ensuremath{\partial}_k q_s| = 2| \sqrt{q_s} \, \ensuremath{\partial}_k\sqrt{q_s}| \leq R (\ensuremath{\partial}_k \sqrt{q_s})^2 + R^{-1} q_s\,,
\end{align*}
where $R>0$ is a free parameter, we obtain
\begin{align*}
\left| \f{1}{\beta N} \int_0^{N^{\upsilon'} \tau_K} \mathrm{d} s \int \ensuremath{\partial}_k q_s \, {\widehat\psi_t^{\mathbf{y}}} \mathrm{d}\mu_G^{\mathbf{y}} \right| &\leq R \left[\int_0^{N^{\upsilon'} \tau_K} \mathrm{d} s \, D_{f_t^{\mathbf{y}}\mu_G}(\sqrt{q_s}) \right] + \f{1}{2}R^{-1}N^{-1+\upsilon'}\tau_K
\\& \leq R \, S_{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}}(\widehat g_t^{\mathbf{y}}) + \f{1}{2}R^{-1}N^{-1+\upsilon'}\tau_K
\\& \leq C R\tau_K D_{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}}(\sqrt{\widehat g_t^{\mathbf{y}}})+ \f{1}{2}R^{-1}N^{-1+\upsilon'}\tau_K\,,
\end{align*}
where in the second line we used that the time integral of the Dirichlet form is bounded by the initial entropy (see, e.g.,~Theorem~2.3 in~\cite{EY}) and in the final line we used the logarithmic Sobolev inequality~\eqref{frozen log-sobolev}. Optimizing over $R$, we get
\begin{align*}
\left| {\mathbb E}^{f_t^{\mathbf{y}}\mu_G^{\mathbf{y}}} x_k - {\mathbb E}^{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}} x_k \right| &\leq C \tau_K\left(N^{-1+\upsilon'}D_{ {\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}}\Big(\sqrt{\widehat g_t^{\mathbf{y}}}\Big) \right)^{1/2} + {\mathcal O}\left(\e{-c N^{\upsilon'}}\right)\,,\nonumber\\
&\le C \frac{KN^{\upsilon'/2}}{N^{3/2}}\left(D_{ {\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}}\Big(\sqrt{\widehat g_t^{\mathbf{y}}}\Big) \right)^{1/2} + {\mathcal O}\left(\e{-c N^{\upsilon'}}\right)\,,
\end{align*}
where we used that $\tau_K= CK/N$.
Using \eqref{good interval estimate} we finally obtain,
\begin{align*}
\left| {\mathbb E}^{f_t^{\mathbf{y}}\mu_G^{\mathbf{y}}} x_k - {\mathbb E}^{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}} x_k \right| \leq C \frac{KN^{2\mathfrak{c}}}{N}N^{-\mathfrak{a}}\tau^{-1} +{\mathcal O}\left(\e{-c N^{\upsilon'}}\right)\,,
\end{align*}
for $N$ sufficiently large on $\Omega_V$.
\end{proof}
\begin{lemma}\label{lemma frozen gap universality}
Fix small constants $\alpha,\sigma>0$. Fix $\epsilon'>0$ and $t\ge\tau N^{\epsilon'}$, where $\tau$ satisfies $1\gg\tau\ge N^{-2\mathfrak{a}}$. Fix $n\in\N$ and consider an $n$-particle observable $O$. Let $\chi',\chi>0$, with $\chi',\chi<\chi_0$, where $\chi_0$ is the constant in Theorem~\ref{theorem 3.1}. Then the following holds.
Assume that $0<\mathfrak{a}<1/2$, $0<\mathfrak{c}<1$, $N^{-2\mathfrak{a}}\le\tau\ll 1$ and $K\in\llbracket N^{\sigma},N^{1/4}\rrbracket$ are chosen such that
\begin{align}\label{choice of the constants}
\frac{KN^{2\mathfrak{c}}}{N} N^{-\mathfrak{a}}\tau^{-1} \le \frac{K^{\chi}}{N} \,.
\end{align}
Let $p$ be an integer satisfying $|p|\le K-K^{1-\chi'}$. Let $\mathbf{y}\in \mathcal{G}_{L,K}(\f{\chi^2 \sigma}{2},\frac{\alpha}{2})$. Then, for the observable $G$, as defined in~\eqref{the averaged observable}, we have
\begin{align}\label{equation frozen gap universality}
\Big| \int G_{L+p,n}(\mathbf{x}) (f_t^{\mathbf{y}}\mathrm{d}\mu_G^{\mathbf{y}}-{\widehat\psi_t}\,\mathrm{d}\mu_G) \Big| \leq C_OK^{-\mathfrak{b}}+C_O \sqrt{K} N^{\mathfrak{c}}N^{-\mathfrak{a}}\tau^{-1}\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$, where the constant $C_O$ depends on $O$ and $\epsilon'$, and the constant $\mathfrak{b}>0$ depends on~$\alpha$ and~$\sigma$.
\end{lemma}
\begin{proof} We follow~\cite{single_gap}. Fix $t\ge \tau N^{\epsilon'}$. Let $\chi>0$ and let $\mathbf{y} \in \mathcal{G}_{L,K}(\f{\chi^2 \sigma}{2}, \alpha) \subset \mathcal{G}_{L,K}( \f{ \chi\sigma}{2}, \alpha)$. Then by Proposition~\ref{frozen stats proximity} and the assumptions in~\eqref{choice of the constants},
\begin{align*}
\left| {\mathbb E}^{f_t^{\mathbf{y}}\mu_G^{\mathbf{y}}} x_k - \widehat{\gamma}_k(t) \right|\le C K^{\chi}N^{-1}\,,
\end{align*}
for all $k\in I\equiv I_{L,K}$. Further, from Lemma~\ref{frozen beta particle proximity} and the assumptions in~\eqref{choice of the constants} we get
\begin{align}\label{super acc}
\left| {\mathbb E}^{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}} x_k - \widehat{\gamma}_k(t) \right| \leq C K^{\chi}N^{-1}\,,
\end{align}
for all $k\in I$. Recall from~\eqref{equidistant points 1} that we denote by $ \bar{y}\mathrel{\mathop:}=\frac{1}{2}(y_{L-K-1}+y_{L+K+1})$ the midpoint of the configuration interval~${\boldsymbol I}$ and that $(\alpha_k)$ denote $2K+1$ equidistant points in the configuration interval~${\boldsymbol I}$. As shown in Lemma~4.5 and Lemma~5.2 of~\cite{single_gap}, we have
\begin{align*}
\left| \widehat{\gamma}_k(t) - \alpha_k \right| \leq CK^{\chi} N^{-1}\,,
\end{align*}
for all $k\in I$, provided that $\mathbf{y}\in{\mathcal G}_{L,K}(\f{\chi \sigma}{2}, \alpha)$. We hence obtain,
\begin{align} \label{improved acc 1}
\left| {\mathbb E}^{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}} x_k - \alpha_k \right| \leq C K^{\chi}N^{-1}\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$.
Proposition \ref{frozen stats proximity} implies that there is $C_O$ such that
\begin{align}\label{what we get from the proposition}
&\left| \int G_{L+p,n}(\mathbf{x}) \left( f_t^{\mathbf{y}}\mathrm{d} \mu_G^{\mathbf{y}} - {\widehat\psi_t^{\mathbf{y}}}\mathrm{d} \mu_G^{\mathbf{y}} \right) \right| \leq C_O \sqrt{K} N^{\mathfrak{c}}N^{-\mathfrak{a}}\tau^{-1}\,,\qquad\qquad (t\ge N^{\epsilon'}\tau)\,,
\end{align}
for $\mathbf{y}\in{\mathcal G}_{L,K}(\f{\chi \sigma}{2}, \alpha)$, $N$ sufficiently large on $\Omega_V$.
For $\alpha,\epsilon_0,\varsigma_1>0$ and a $\beta$-ensemble $\mu$ on $\digamma^{(N)}$, define a set of particle configurations $\mathcal{R}_{\mu}^{*}\equiv\mathcal{R}_{\mu}^{*}(\epsilon_0,\alpha) $ by
\begin{align} \label{definition of R star}
\mathcal{R}^{*}_{\mu}\equiv\mathcal{R}_{\mu}^{*}(\epsilon_0,\alpha) \mathrel{\mathop:}= \left\{\mathbf{y} \in \digamma^{(N-{\mathcal K})}\,: \; {\mathbb P}^{\mu^{\mathbf{y}}}\left(\left| x_k -\gamma_k\right| > N^{-1+\epsilon_0}\right) \le \e{-\f{1}{2} N^{\varsigma_1}}\,,\forall k\in I_{L,K}\right\}\,,
\end{align}
where $\gamma_k$ denotes the classical location of the $k$-th particle with respect to the equilibrium measure of $\mu$.
As in the proof of Theorem \ref{frozen stats proximity}, it follows from Markov's inequality and the rigidity bound for the $\beta$-ensemble ${\widehat\psi_t}\, \mu_G$ in Lemma~\ref{rigidity for time dependent beta} that we can choose $\mathcal{R}^{*}_{{\widehat\psi_t}\,\mu_G}\subset{\mathcal R}_{L,K}$ and that ${\mathbb P}^{{\widehat\psi_t} \mu_G}(\mathcal{R}^{*}_{{\widehat\psi_t}\,\mu_G})\geq 1- c\e{-\f{1}{2} N^{\varsigma_1}}$, for some $c>0$, possibly after decreasing $\varsigma_1$ by a small amount. For ${\tilde\mathbf{y}} \in \mathcal{R}_{{\widehat\psi_t}\,\mu_G}^{*}(\f{\chi^2 \sigma}{2}, \f{\alpha}{2})$, Lemma 5.1 of~\cite{single_gap} implies that
\begin{align} \label{improved acc 2}
\left| {\mathbb E}^{{\widehat\psi_t^{\tilde\mathbf{y}}} \mu_G^{\tilde\mathbf{y}}}x_k- \alpha_k \right| \leq CK^{\chi} N^{-1}\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$. Thus together with~\eqref{improved acc 1}, we have
\begin{align}\label{to check the assumption in theorem}
\left| {\mathbb E}^{{\widehat\psi_t^{\tilde\mathbf{y}}} \mu_G^{\tilde\mathbf{y}}}x_k- \alpha_k \right|+\left| {\mathbb E}^{{\widehat\psi_t^{\mathbf{y}}} \mu_G^{\mathbf{y}}} x_k - \alpha_k \right| \leq C K^{\chi}N^{-1}\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$, for all $\mathbf{y}\in{\mathcal G}(\f{\chi \sigma}{2}, \alpha)$ and all ${\tilde\mathbf{y}} \in \mathcal{R}_{{\widehat\psi_t}\,\mu_G}^{*}(\f{\chi^2 \sigma}{2}, \f{\alpha}{2})$.
We now apply Theorem \ref{theorem 3.1}: Let $\tilde\mathbf{y}$ and $\mathbf{y}$ be as above. By the scaling argument of Lemma~5.3 in~\cite{single_gap}, we can assume that the two configuration intervals $\boldsymbol{\tilde I}$ and $\boldsymbol{I}$ agree, so that assumption~\eqref{assumption 1 thm3.1} of Theorem~\ref{theorem 3.1} holds. Moreover, by Lemma~\ref{potential is regular} we know that $V^\mathbf{y}$ and $V^{\tilde\mathbf{y}}$ are $K^{\chi}$-regular external potentials. By~\eqref{to check the assumption in theorem} assumption~\eqref{assumption 3 thm3.1} of Theorem~\ref{theorem 3.1} is satisfied. Thus Theorem~\ref{theorem 3.1} implies that there is $\mathfrak{b}>0$, depending on~$\sigma$ and~$\alpha$, such that
\begin{align} \label{what we get from 3.1}
\left| \int G_{L+p,n}(\mathbf{x})\left({\widehat\psi_t^{\mathbf{y}}} \mathrm{d}\mu_G^{\mathbf{y}} - {\widehat\psi_t^{\tilde\mathbf{y}}}\mathrm{d}\mu_G^{\tilde\mathbf{y}}\right)\right| \leq C_O K^{-\mathfrak{b}}\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$. Since the estimate~\eqref{what we get from 3.1} holds for all $\tilde\mathbf{y} \in \mathcal{R}_{{\widehat\psi_t}\,\mu_G}^{*}$, and since ${\mathbb P}^{{\widehat\psi_t} \mu_G}(\mathcal{R}_{{\widehat\psi_t}\,\mu_G}^{*})\geq 1- \e{-\f{1}{2} N^{\varsigma_1}}$, we can integrate over $\tilde\mathbf{y}$ to find that
\begin{align*}
\left| \int G_{L+p,n}(\mathbf{x}) \left({\widehat\psi_t^{\mathbf{y}}} \mathrm{d}\mu_G^{\mathbf{y}} - {\widehat\psi_t}\,\mathrm{d} \mu_G\right)\right| \leq C_O K^{-\mathfrak{b}}\,,
\end{align*}
for $N$ sufficiently large on $\Omega_V$. In combination with~\eqref{what we get from the proposition}, this yields~\eqref{equation frozen gap universality}.
\end{proof}
\subsection{Proof of Theorem~\ref{main theorem of section 6} and Theorem~\ref{corollary gaussian}}\label{subsection proof of theorem and corollary section 6}
Lemma~\ref{lemma frozen gap universality} compares the local statistics of the locally-constrained measure $f_t^{\mathbf{y}}\,\mu_G^{\mathbf{y}}$ with the $\beta$-ensemble ${\widehat\psi_t}\,\mu_G$. In order to compare with local statistics of the measure $f_t\,\mu_G$ with $\widehat\psi_t\,\mu_G$, we integrate out the boundary conditions $\mathbf{y}$.
\begin{lemma}\label{lemma final estimate}
Under the assumptions of Lemma~\ref{lemma frozen gap universality} the following holds. Let $J\subset\llbracket \alpha N,(1-\alpha)N\rrbracket$ be an interval of consecutive integers in the bulk. Then,
\begin{align}\label{final estimate}
\left|\int \frac{1}{|J|}\sum_{j\in J}G_{j,n}(\mathbf{x}) (f_t\mathrm{d}\mu_G-{\widehat\psi_t}\mathrm{d}\mu_G)\right|\le C_O\left(N^{-\mathfrak{c}}+
K^{-\mathfrak{b}}+K^{-\chi'/2}\right)+C_O \sqrt{K} N^{\mathfrak{c}}N^{-\mathfrak{a}}\tau^{-1}\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$.
\end{lemma}
\begin{proof}
For small $\chi'>0$ as in Lemma~\ref{lemma frozen gap universality}, set ${\tilde K}\mathrel{\mathop:}= K-K^{1-\chi'/2}$. We first assume that $J$ is such that $|J|\le2\tilde K+1$. We then choose $L$ such that $J\subset I_{L,\tilde K}\subset I_{L,K}$. Recall the set of configurations ${\mathcal G}$ in Lemma~\ref{frozen stats proximity}. Using the conditioned measure $f_t^{\mathbf{y}}\,\mu_G^{\mathbf{y}}$ we estimate
\begin{align}
\E^{f_t\,\mu_G}\left[\frac{1}{|J|}\sum_{j\in J}G_{j,n}\right]=\E^{f_t\,\mu_G}\left[\frac{1}{|J|}\int\sum_{j\in J}G_{j,n}(\mathbf{x})f_t^{\mathbf{y}}\mathrm{d}\mu_G^{\mathbf{y}}\lone({\mathcal G})\right]+{\mathcal O}\left(N^{-\mathfrak{c}} \right)\,,
\end{align}
where we used~\eqref{probability estimate on good boundary condition}. Next, using Lemma~\ref{lemma frozen gap universality} we obtain on $\Omega_V$
\begin{align*}
\frac{1}{|J|}\int\sum_{j\in J}G_{j,n}(\mathbf{x})f_t^{\mathbf{y}}(\mathbf{x})\mathrm{d}\mu_G^{\mathbf{y}}(\mathbf{x})\lone({\mathcal G})=\frac{1}{|J|}\int\sum_{j\in J} G_{j,n}\,\widehat\psi_t\mathrm{d}\mu_G+{\mathcal O}\left( K^{-\mathfrak{b}}\right)+{\mathcal O}\left( \sqrt{K} N^{\mathfrak{c}}N^{-\mathfrak{a}}\tau^{-1}\right)\,,
\end{align*}
on $\Omega_V$. For the special case $|J|\le 2\tilde K+1$, this yields~\eqref{final estimate}.
If $|J|\ge\tilde K+1$, there are $L_a\in\llbracket \alpha N,(1-\alpha)N\rrbracket$, with $a\in\llbracket 1,M_0\rrbracket$, such that the intervals $I_{L_a,K}=\llbracket L_a-K,L_a+K\rrbracket$ are non-intersecting with the properties that $J\subset \bigcup_{a=1}^{M_0} I_{L_a,K}$ and $J\cap I_{L_{a},K}\not=\emptyset$, for all $a\in\llbracket 1,M_0\rrbracket$. Note that $M_0\le \frac{|J|}{K}+2$. For simplicity of notation we abbreviate $I^{(a)}\equiv I_{L_a,K}=\llbracket L_a-K,L_a+K\rrbracket$ and $\tilde I^{(a)}\equiv\llbracket L_a-\tilde K,L_a+\tilde K\rrbracket$. We also label the interior and exterior points of a configuration $\boldsymbol{\lambda}\in\digamma^{(N)}$ accordingly,
\begin{align}
\mathbf{x}^{(a)}=(x_{L_a-K},\ldots, x_{K+L_a})\in\digamma^{({\mathcal K})}\,,\qquad\quad\mathbf{y}^{(a)}=(y_1,\ldots, y_{L_a-K-1},y_{L_a+K+1},\ldots, y_N)\in \digamma^{(N-{\mathcal K})}\,,
\end{align}
c.f.,~\eqref{bfx notation}. We let ${\mathcal G}^{(a)}\equiv {\mathcal G}_{L_a,K}(\epsilon_0,\alpha)\subset{\mathcal R}_{L_a,K}(\epsilon_0,\alpha)$ denote the set of configurations obtained in Lemma~\ref{frozen stats proximity}. Using this notation we can write
\begin{align}\label{eq 6.35}
\E^{f_t\,\mu_G}\left[\frac{1}{|J|}\sum_{j\in J}G_{j,n}\right]=\frac{1}{|J|}\sum_{{ a\,:\,I^{(a)}\cap J\not=\emptyset}} \E^{f_t\,\mu_G}\left[\int\sum_{j\in I^{(a)}\cap J}G_{j,n}(\mathbf{x}^{(a)})f_t^{\mathbf{y}^{(a)}}\mathrm{d}\mu_G^{\mathbf{y}^{(a)}}\lone({\mathcal G}^{(a)})
\right]+{\mathcal O}\left( N^{-\mathfrak{c}}\right)\,,
\end{align}
on $\Omega_V$, where the first summation on the right side is over indices $a\in\llbracket 1,M_0\rrbracket$ such that the intervals $(I^{(a)})$ satisfy $I^{(a)}\cap J\not=\emptyset$. Here, we also used the probability estimate on ${\mathcal G}^{(a)}$ in~\eqref{probability estimate on good boundary condition}. In~\eqref{eq 6.35} we may further restrict, for each $a$, the summation over the index $j$ from $I^{(a)}$ to $\tilde I^{(a)}$ at an expense of an error term of order $ |I^{(a)}\backslash \tilde I^{(a)}|\le K^{1-\chi'/2}$. Then summing over $a\in\llbracket 1,M_0\rrbracket$, with $M_0\sim |J|/K$, we get
\begin{align}\label{on the way to the final estimate 2}
\E^{f_t\,\mu_G}\left[\frac{1}{|J|}\sum_{j\in J}G_{j,n}\right]&=\frac{1}{|J|}\sum_{{ a\,:\,I^{(a)}\cap J\not=\emptyset}} \E^{f_t\,\mu_G}\left[\int\sum_{j\in\tilde I^{(a)}\cap J}G_{j,n}(\mathbf{x}^{(a)})f_t^{\mathbf{y}^{(a)}}\mathrm{d}\mu_G^{\mathbf{y}^{(a)}}\lone({\mathcal G}^{(a)})
\right]\nonumber\\
&\qquad\qquad+{\mathcal O}\left( N^{-\mathfrak{c}}\right)+{\mathcal O}\left( K^{-\chi'/2} \right)\,,
\end{align}
on $\Omega_V$. Since for each choice of the index $a$ the term in the expectation on the right side of~\eqref{on the way to the final estimate 2} can be dealt with as in the case $|J|\le 2\tilde K+1$ above, this concludes the proof of~\eqref{final estimate} for general $J$.
\end{proof}
We can now give the proof of Theorem~\ref{main theorem of section 6}.
\begin{proof}[Proof of Theorem~\ref{main theorem of section 6}]
Let $\alpha>0$. We first choose the constants $\mathfrak{a}\in(0,1/2)$, $\mathfrak{c}\in(0,1)$ and $\epsilon'>0$, and the parameter $K\in\llbracket N^{\sigma},N^{1/4}\rrbracket$ appropriately: Let $\delta>0$ be a small constant. Then we set $\mathfrak{a}\equiv 1/2-\delta$, $\mathfrak{c}\equiv\delta/4$, $K\equiv N^{\delta/4}$, $\epsilon'\equiv\delta$, $\sigma=\delta/8$. Note first that for this choice of $K$ the condition~\eqref{condition on K} is satisfied. Second, for sufficiently small $\delta>0$, we observe that
\begin{align*}
KN^{2\mathfrak{c}} N^{-\mathfrak{a}}\tau^{-1}=N^{{3\delta}/{4}}N^{-\mathfrak{a}}\tau^{-1}\le{K^{\chi}} \,,
\end{align*}
holds, e.g., for $\tau\ge N^{\delta} N^{-\mathfrak{a}}$ and $\chi>0$ (with $\chi\le\chi_0)$. Thus~\eqref{choice of the constants} is satisfied with the above choices.
Hence, for $t\ge N^{2\delta}\tau$, Lemma~\ref{lemma final estimate} yields, for some $\mathfrak{b}>0$
\begin{align*}
\left|\int \frac{1}{|J|}\sum_{j\in J}G_{j,n}(\mathbf{x}) (f_t\mathrm{d}\mu_G-{\widehat\psi_t}\mathrm{d}\mu_G)\right|&\le C_O
K^{-\mathfrak{b}}+C_ON^{-\mathfrak{c}}+C_OK^{-\chi'/2}+C_O \sqrt{K} N^{\mathfrak{c}}N^{-\mathfrak{a}}\tau^{-1}\,,\nonumber\\
\end{align*}
for $N$ sufficiently large on $\Omega_V$. Thus, choosing $\tau\ge N^{\delta}N^{-\mathfrak{a}}$, there is a constant $\mathfrak{f}>0$ such that~\eqref{equation in main theorem of sectoin 6} holds. This completes the proof of Theorem~\ref{main theorem of section 6}.
\end{proof}
Next, we sketch the proof of Theorem~\ref{corollary gaussian}.
\begin{proof}[Proof of Theorem~\ref{corollary gaussian}]
The proof of Theorem~\ref{corollary gaussian} is almost identical to the proof of Theorem~\ref{main theorem of section 6}. In fact, it suffices to establish Lemma~\ref{lemma frozen gap universality} with $\mu_G$ replacing ${\widehat\psi_t}\mu_G$ on the left side of~\eqref{equation frozen gap universality}. This can be accomplished by applying Theorem~\ref{theorem 3.1} with $\mu_G$ instead of ${\widehat\psi_t}\,\mu_G$: Let $\tilde\mathbf{y}\in{\mathcal R}^{*}_{\mu_G}(\chi^2\sigma/2,\alpha/2)$ and let $\mathbf{y}\in{\mathcal G}(\chi^2\sigma/2,\alpha/2)$. Using the arguments of Proposition~5.2 in~\cite{single_gap}, we can rescale $\mu_G$ such that~\eqref{assumption 1 thm3.1} and~\eqref{assumption 2 thm3.1} are satisfied for $\mathbf{y}$ and $\tilde\mathbf{y}$. It is also straightforward to check that the external potentials leading to $\mu_G^{\tilde\mathbf{y}}$, $\tilde\mathbf{y}\in{\mathcal R}^{*}_{\mu_G}(\chi^2\sigma/2,\alpha/2)$, are $K^{\chi}$-regular. By Lemma~5.1 of~\cite{single_gap} we obtain
\begin{align*}
\left| {\mathbb E}^{\mu_G^{\tilde\mathbf{y}}}x_k- \alpha_k \right| \leq CK^{\chi} N^{-1}\,.
\end{align*}
Hence, using the estimate~\eqref{super acc}, we conclude that the assumption~\eqref{assumption 3 thm3.1} is also satisfied. Thus Theorem~\ref{theorem 3.1} yields,
\begin{align}\label{what we get to the second}
\left| \int G_{L+p,n}(\mathbf{x}) {\widehat\psi_t^{\mathbf{y}}} \mathrm{d}\mu_G^{\mathbf{y}} -\int G_{{L}+p,n,sc}(\mathbf{x})\,\mathrm{d}\mu_G^{\tilde\mathbf{y}}\right| \leq C_O K^{-\mathfrak{b}}\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$. We refer to the proof of Proposition~5.2 in~\cite{single_gap} for more details.
Since ${\mathcal R}^{*}_{\mu_G}(\chi^2\sigma/2,\alpha/2)$ has exponentially high probability under $\mu_G$, we can integrate over $\tilde\mathbf{y}$ to find
\begin{align*}
\left| \int G_{L+p,n}(\mathbf{x}) {\widehat\psi_t^{\mathbf{y}}} \mathrm{d}\mu_G^{\mathbf{y}} -\int G_{{L}+p,n,sc}(\mathbf{x})\,\mathrm{d}\mu_G\right| \leq C_O K^{-\mathfrak{b}}\,,
\end{align*}
for $N$ sufficiently large on $\Omega_V$.
The proof of Theorem~\ref{corollary gaussian} is now completed in the same way as the proof of Theorem~\ref{main theorem of section 6}.
\end{proof}
\section{From gap statistics to correlation functions}\label{From gap statistics to correlation functions}
In this section, we translate our results on the averaged local gap statistics into results on averaged correlation functions. Since this procedure is fairly standard, see, e.g.,~\cite{ESYY}, we refrain from stating all proofs in details. We first need to slightly generalize the setup of Section~\ref{local equilibrium measures}.
Fix $n\in\N$, let $O$ be an $n$-particle observable and consider an array of increasing positive integers,
\begin{align}\label{array of interges}
{\boldsymbol m}=(m_1,m_2,\ldots, m_n)\in\N^{n}\,.
\end{align}
Let $\alpha>0$. We define for $j\in\llbracket \alpha N,(1-\alpha)N\rrbracket$ and $t\ge 0$ an observable $G_{j,{\boldsymbol m},t}\equiv G_{j,{\boldsymbol m}}$ by
\begin{align}\label{generalized observable}
G_{j,{\boldsymbol m}}(\mathbf{x})\mathrel{\mathop:}= O(N\rho_j (x_{j+m_1}-x_j), N\rho_j(x_{j+m_2}-x_j),\ldots,N\rho_j(x_{j+m_n}-x_j))\,,
\end{align}
where $\rho_j\equiv\widehat\rho_{fc}(t,\widehat\gamma_j(t))$ denotes the density of the measure $\widehat\rho_{fc}(t)$ at the classical location of the $j$-th~particle,~$\widehat\gamma_i(t)$, with respect to the measure $\widehat\rho_{fc}(t)$. We set $G_{j,{\boldsymbol m}}=0$, if $j+m_n\ge (1-\alpha)N$. Similarly, we define $G_{j,{\boldsymbol m},sc}$ by replacing $\rho_j$ by the density of the standard semicircle law at the classical locations of the $j$-th particle with respect to the semicircle law; c.f.~\eqref{the averaged observable at infty}. The following theorem generalizes Theorem~\ref{corollary gaussian}.
\begin{theorem}\label{generalized gap distribution}
Let $n\in\N$ be fixed and let $O$ be an $n$-particle observable. Fix small constants $\alpha,\delta>0$ and consider an interval of consecutive integers $J\subset\llbracket \alpha N,(1-\alpha)N\rrbracket$ in the bulk. Then there are constants $\mathfrak{f},\delta'>0$ such that the following holds. Let ${\boldsymbol m}\in\N^{n}$ be an array of increasing integers (see~\eqref{array of interges}) such that $m_n\le N^{\delta'}$, and consider the observable $G_{j,{\boldsymbol m}}$ respectively $G_{j,{\boldsymbol m},sc}$ (see~\eqref{generalized observable}). Assume that $t\ge N^{-1/2+\delta}$, then
\begin{align}
\left|\int \frac{1}{|J|}\sum_{j\in J}G_{j,{\boldsymbol m}}(\mathbf{x}) \,f_t(\mathbf{x})\mathrm{d}\mu_G(\mathbf{x})-\int \frac{1}{|J|}\sum_{j\in J}G_{j,{\boldsymbol m},sc}(\mathbf{x})\,\mathrm{d}\mu_G(\mathbf{x})\right|\le C_O N^{-\mathfrak{f}}\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$. The constant $C_O$ depends on $\alpha$ and $O$, and the constants $\mathfrak{f}$ and $\delta'$ depend on $\alpha$ and $\delta$.
\end{theorem}
Theorem~\ref{generalized gap distribution} is proven in the same way as Theorem~\ref{corollary gaussian}. We remark that $\delta'$ is chosen such that $N^{\delta'}\ll K $, i.e., $m_n$ is much smaller than the size of the interval $I_{L,K}$.
For $n\ge 1$, define the $n$-point correlation function, $\varrho_{f_t,n}^N$, by
\begin{align*}
\varrho^N_{f_t,n}(x_1,\ldots,x_n)\mathrel{\mathop:}=\int_{\ensuremath{\mathbb{R}}^{N-n}} (f_t\mu_G)^{\#}\,\mathrm{d} x_{n+1}\cdots\mathrm{d} x_{N}\,,
\end{align*}
where $(f_t\,\mu_G)^{\#}$ denote the symmetrized versions of $f_t\,\mu_G$. Similarly, we denote by
\begin{align*}
\varrho^N_{G,n}(x_1,\ldots,x_n)\mathrel{\mathop:}=\int_{\ensuremath{\mathbb{R}}^{N-n}} \mu_G^{\#}\,\mathrm{d} x_{n+1}\cdots\mathrm{d} x_{N}\,,
\end{align*}
the $n$-point correlation functions of the Gaussian ensembles; see~\eqref{symmetrization}.
Recall that we denote by $\widehat L_{\pm}(t)$, respectively $L_\pm(t)$, the endpoints of the support of the measure~$\widehat\rho_{fc}(t)$, respectively the measure~$\rho_{fc}(t)$.
Recall that the two densities~$f_t$ and ${\widehat\psi_t}$ are both conditioned on $V$, i.e., the entries $(v_i)$ of $V$ are considered fix. We have the following result on the averaged correlation functions of $f_t\,\mu_G$ and ${\widehat\psi_t}\,\mu_G$.
\begin{theorem}\label{theorem translation}
Fix $n\in\N$ and choose an $n$-particle observable $O$. Fix a small $\delta>0$ and let $t \ge N^{-1/2+\delta}$. Let $\widetilde\alpha>0$ be a small constant and consider two energies $E\in[L_-(t)+\widetilde\alpha,L_+(t)-\widetilde\alpha]$ and $E'\in[-2+\widetilde\alpha,2-\widetilde\alpha]$. Then we have, for any $\epsilon>0$ and for $b\equiv b_N$ satisfying $\widetilde\alpha/2\ge b_N>0$,
\begin{align}\label{equation theorem translation}
\bigg|\int_{\ensuremath{\mathbb{R}}^n}&\mathrm{d}\alpha_1\cdots\mathrm{d}\alpha_n\, O(\alpha_1,\ldots,\alpha_n)\bigg[\int_{E-b}^{E+b}\frac{\mathrm{d} x}{2b}\frac{1}{\rho_{fc}(t,E)^n}\,\varrho^N_{f_t,n}\Big(x+\frac{\alpha_1}{N\rho_{fc}(t,E)},\ldots,x+\frac{\alpha_n}{N\rho_{fc}(t,E)} \Big)\nonumber\\
&-\int_{E'-b}^{E'+b}\frac{\mathrm{d} x}{2b}\frac{1}{\rho_{sc}(E')^n}\,\varrho^N_{G,n}\Big(x+\frac{\alpha_1}{N\rho_{sc}(E')},\ldots,x+\frac{\alpha_n}{N\rho_{sc}(E')} \Big)\bigg]\bigg|\le C_ON^{2\epsilon}\left(b^{-1} N^{-1+\epsilon} +N^{-\mathfrak{f}}+N^{-c{\alpha_0}}\right)\,,
\end{align}
for $N$ sufficiently large on $\Omega_V$. Here $\mathfrak{a}$ is the constant in the rigidity estimate~\eqref{master rigidity bound} and $\mathfrak{f}$ is the constant in Theorem~\ref{generalized gap distribution}. Moreover, $\rho_{fc}(t,E)$ stands for the density of the ($N$-independent) measure $\rho_{fc}(t)$ at the energy $E$. The constant $C_O$ depends on $O$ and $\widetilde\alpha$. Further, $\alpha_0$ is the constant appearing in Assumption~\ref{assumption mu_V convergence}. The constant $c$ depends on the measure~$\nu$.
\end{theorem}
Theorem~\ref{theorem translation} follows from Theorem~\ref{corollary gaussian}. This is an application of Section~7 in~\cite{ESYY}. The validity of Assumption~IV in~\cite{ESYY} is a direct consequence of the local law in Theorem~\ref{thm.strong}. Further, we remark that the parameter $b_N$ in Theorem~\ref{theorem translation} and the interval of consecutive integers $J$ in Theorem~\ref{generalized gap distribution} are related by $J=\{i\,:\,\widehat\gamma_i(t)\in[ E-b_N,E+b_N]\}$, where $\widehat\gamma_i(t)$ are the classical locations with respect to the measure $\widehat\rho_{fc}(t)$. This explains, up to minor technicalities, $b_N\gg N^{-1}$. Then Section~7 of~\cite{ESYY} yields~\eqref{equation theorem translation} formulated in terms of $\widehat\rho_{fc}(t)$ instead of $\rho_{fc}(t)$. Using~\eqref{mini local law} and the smoothness of~$O$, we can replace $\widehat\rho_{fc}$ by $\rho_{fc}$ at the expense of an error of size $C_ON^{-c\alpha_0}$. This
eventually
gives~\eqref{equation theorem translation} with $\rho_{fc}(t)$.
\section{Proofs of main results}\label{Proofs of main results}
Theorem~\ref{theorem translation} shows that the averaged local correlation functions of ensembles of the form
\begin{align*}
H_t=\e{-(t-t_0)/2}V+\e{-t/2}W+(1-\e{-t})^{1/2}{W'}\,,
\end{align*}
with some small $t_0\ge 0$, and with $ W'$ a GUE/GOE matrix independent of $W$ and $V$, can be compared with the averaged local correlation functions of the GUE, respectively GOE, for times satisfying $t\gg N^{-1/2}$. In this section, we explain how this can be used to prove the universality at time $t=0$.
\subsection{Green Function Comparison Theorem}
We start with a Green function comparison theorem. Assume that we are given two complex Hermitian or real symmetric Wigner matrices, $X$ and $Y$ both satisfying the assumptions in Definition~\ref{assumption wigner}. Let $V$ be a real random or deterministic diagonal matrix satisfying Assumptions~\ref{assumption mu_V} and~\ref{assumption mu_V convergence}. Consider the deformed Wigner matrices
\begin{align*}
H^X\mathrel{\mathop:}= V+X\,,\qquad\quad H^Y\mathrel{\mathop:}= V+Y\,,
\end{align*}
of size $N$. The main theorem of this subsection, Theorem~\ref{corollary of green function comparison}, states that the correlation functions of the two matrices $H^X$ and $H^Y$, when conditioned on $V$, are identical on scale $1/N$ provided that the first four moments of $X$ and $Y$ almost match. Theorem~\ref{corollary of green function comparison} is a direct consequence of the Green function comparison Theorem~\ref{the green function comparison theorem}.
Denote the Green functions of $H^X$, $H^Y$ respectively, by
\begin{align*}
G^X(z)\mathrel{\mathop:}= \frac{1}{H^X-z}\,,\qquad\quad G^Y(z)\mathrel{\mathop:}= \frac{1}{H^Y-z}\,,\qquad\qquad (z\in\ensuremath{\mathbb{C}}\backslash\ensuremath{\mathbb{R}})\,,
\end{align*}
and set $m_N^X(z)\mathrel{\mathop:}= N^{-1}\Tr G^X(z)$, $m_N^Y(z)\mathrel{\mathop:}= N^{-1}\Tr G^Y(z)$. From Theorem~\ref{thm.strong}, we know that, for all $z\in{\mathcal D}_L$ (see~\eqref{eq.DL}), with $L\ge 40\xi$,
\begin{align}\label{eq of thm.strong}
\left|m_N^X(z)-\widehat m_{fc}(z)\right|\le(\varphi_N)^{c\xi}\frac{1}{N\eta}\,,\qquad |G_{ij}^X(z)-\delta_{ij} \widehat g_{i}(z)|\le (\varphi_N)^{c\xi}\left(\sqrt{\frac{\im\widehat m_{fc}(z)}{N\eta}}+\frac{1}{N\eta} \right)\,,
\end{align}
with $(\xi,\upsilon)$-high probability on $\Omega_V$ for some $\upsilon>0$ and $c>0$, where
\begin{align*}
\widehat g_i(z)\mathrel{\mathop:}=\frac{1}{v_i-z-\widehat m_{fc}(z)}\,,\qquad\qquad (z\in\ensuremath{\mathbb{C}}\backslash\ensuremath{\mathbb{R}})\,.
\end{align*}
Here, $\widehat m_{fc}$, is the Stieltjes transform of the measure $\widehat\rho_{fc}$, which agrees with $\widehat\rho_{fc}^{\vartheta}$ for the choice $\vartheta=1$ and with $\widehat\rho_{fc}(t)$ for the choice $t=t_0$.
The identical estimates hold true when $X$ is replaced by $Y$.
Recall that we denote by $\widehat L_{\pm}$ the endpoints of the support of $\widehat\rho_{fc}$, and that we denote by $\widehat\kappa_E\equiv \widehat\kappa $ the distance of $E\in[\widehat L_-,\widehat L_+]$ to the endpoints $\widehat L_{\pm}$. Adapting the Green function theorem of~\cite{EYY1} we obtain the following theorem.
\begin{theorem}\emph{[Green Function Comparison Theorem]}\label{the green function comparison theorem}
Assume that $X$ and $Y$ satisfy the assumptions in Definition~\ref{assumption wigner} and let~$V$ satisfy the Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}. Assume further that the first two moments of $X=(x_{ij})$ and $Y=(y_{ij})$ agree and that the third and forth moments satisfy,
\begin{align}\label{matching 1}
\left|\E\,\overline{x}_{ij}^{p}\,{x}_{ij}^{3-p}-\E\,\overline{y}_{ij}^{p}\,{y}_{ij}^{3-p}\right|\le N^{-\delta-2}\,,\qquad\qquad (p\in\llbracket 0,3\rrbracket)\,,
\end{align}
respectively,
\begin{align} \label{matching 2}
\left|\E\,\overline{x}_{ij}^{q}\,{x}_{ij}^{4-q}-\E\,\overline{y}_{ij}^{q}\,{y}_{ij}^{4-q}\right|\le N^{-\delta}\,,\qquad\qquad (q\in\llbracket 0,4\rrbracket)\,,
\end{align}
for some given $\delta>0$.
Let $\epsilon > 0$ be arbitrary and let $ N^{-1-\epsilon} \leq \eta \leq N^{-1}$. Fix $N$-independent integers $k_1,\ldots,k_n$ and energies $E^1_j,\ldots,E^{k_j}_j$, $j=1,\ldots,n$, with $\widehat\kappa > \widetilde\alpha$ for all $E^k_j$ with some fixed $\widetilde{\alpha}>0$. Define $z^k_j \mathrel{\mathop:}= E^k_j \pm i \eta$, with the sign arbitrarily chosen. Suppose that $F$ is a smooth function such that for any multi-index ${\sigma} = (\sigma_1,\ldots,\sigma_n)$, with $1\le|{\sigma}|\le 5$, and any $\epsilon'>0$ sufficiently small, there is a $C_0 > 0$ such that
\begin{align*}
\max \{ |\ensuremath{\partial}^{\sigma} F(x_1,\ldots,x_n)|\, : \, \max_j |x_j| \leq N^{\epsilon'} \} & \leq N^{C_0\epsilon'}\,,
\\ \max \{ |\ensuremath{\partial}^{\sigma} F(x_1,\ldots,x_n)| \,: \, \max_j |x_j| \leq N^{2} \} & \leq N^{C_0}\,,
\end{align*}
for some $C_0$.
Then there exists a constant $C_1$, depending on $\sum_m k_m$, $C_0$ and the constants in~\eqref{eq.C0}, such that for any $\eta$ with $N^{-1-\epsilon} \le \eta \le N^{-1}$,
\begin{align} \left| {\mathbb E} F \left(\f{1}{N^{k_1}} \Tr \prod_{j=1}^{k_1} G^X(z^1_j) ,\ldots,\f{1}{N^{k_n}} \Tr \prod_{j=1}^{k_n} G^X(z^n_j) \right) - {\mathbb E} F \left(\f{1}{N^{k_1}} \Tr \prod_{j=1}^{k_1} G^Y(z^1_j) ,\ldots,\f{1}{N^{k_n}} \Tr \prod_{j=1}^{k_n} G^Y(z^n_j) \right) \right| \nonumber \\ \leq C_1 N^{-1/2 + C_1 \epsilon} + C_1 N^{-1/2 + \delta + C_1 \epsilon} \,,
\end{align}
for $N$ sufficiently large on $\Omega_V$.
\end{theorem}
Theorem~\ref{the green function comparison theorem} is proven in the same way as Theorem~2.3 in~\cite{EYY} with the following modifications. Fix some labeling of $\{ (i,j): 1\leq i \leq j \leq N\}$ by $\llbracket 1,\gamma(N)\rrbracket$, with $\gamma(N)\mathrel{\mathop:}= N(N+1)/2$, and write the $\gamma$-th element of this labeling as $(i_\gamma,j_\gamma)$. Starting with $W^{(0)}\equiv X$, inductively define $W^{(\gamma)}$ by replacing the $(i_\gamma,j_\gamma),(j_\gamma,i_\gamma)$ entries of $W^{(\gamma-1)}$ by the corresponding entries of $Y$. Moreover set $H^{(\gamma)}\mathrel{\mathop:}= V+W^{(\gamma)}$. Thus we have $H^{(0)}=H^X$, $H^{(\gamma(N))}=H^Y$, and $H^{(\gamma)} - H^{(\gamma-1)}$ is zero in all but two entries for every $\gamma$. To prove Theorem~\ref{the green function comparison theorem} we use a Lindeberg type replacement strategy: we successively replace the entries of the matrix $X$ by entries of the matrix $Y$. Note, however, that the entries of the matrix $V$ are not changed.
The main technical input in the proof of Theorem~2.3 in~\cite{EYY1} is the estimate~(2.21) in that publication. For the case at hand the corresponding estimate reads as follows: Let $\xi$ satisfy~\eqref{eq.xi}. Then, for all $\delta>0$, and any $N^{-1/2} \gg y\ge N^{-1+\delta}$, we have
\begin{align}\label{the estimate of the green function comparison}
{\mathbb P} \left( \max_{\gamma \leq \gamma(N)} \max_k \sup_{\substack{E\,:\,\widehat\kappa_E \geq \widetilde{\alpha}} }\left| \left( \f{1}{H^{(\gamma)} - E - iy}\right)_{kk} \right| \leq N^{2\delta} \right) \geq 1 - \e{{-\upsilon}(\varphi_N)^{\xi}}\,,
\end{align}
on $\Omega_V$ for $N$ sufficiently large, where $\upsilon>0$ depends only on $\widetilde\alpha$, $\delta$ and the constants in~\eqref{eq.C0}. The estimate~\eqref{the estimate of the green function comparison} follows easily from the local law in~\eqref{eq of thm.strong}, the stability bound~\eqref{stability bound} and Lemma~\ref{lemma hatmfc}. The rest of the proof of Theorem~\ref{the green function comparison theorem} is identical to the proof in~\cite{EYY1}. (The matching conditions~\eqref{matching 1} are weaker than in~\cite{EYY1}, but the proof carries over without any changes.)
Lindeberg's replacement method was applied in random matrix theory in~\cite{C} to compare traces of Green functions. This idea was also used in~\cite{TV1} in the the proof of the ``four moment theorem'' that compares individual eigenvalue distributions. The four-moment matching conditions~\eqref{matching 1} and~\eqref{matching 2} appeared first in~\cite{TV1} with $\delta=0$. The ``Green function comparison theorem'' of~\cite{EYY1} compares Green functions at fixed energies. Since the approach in~\cite{TV1} requires additional difficult estimates due to singularities from neighboring eigenvalues, we follow the method of~\cite{EYY1}, where difficulties stemming from such resonances are absent. For deformed Wigner matrices with deterministic potential the approach of~\cite{TV1} was recently followed in~\cite{OV} where a ``four moment theorem'' has been established. It allows to compare local correlation functions of the matrices $V+W$ and $V+{W'}$ for fixed~$V$, where~$W$ and~$W'$ are real symmetric or
complex Hermitian Wigner matrices, provided that the moments of the off-diagonal entries of~$W$ and~${W'}$ match to fourth
order.
The Green function comparison theorem leads directly to the equivalence of local statistics for the matrices~$H^Y$ and~$H^X$.
\begin{theorem}\label{corollary of green function comparison}
Assume that $X$, $Y$ are two complex Hermitian or two real symmetric Wigner matrices satisfying the assumptions in Definition~\ref{assumption wigner}. Assume further that $X$ and $Y$ satisfy the matching conditions~\eqref{matching 1} and~\eqref{matching 2}, for some $\delta>0$. Let~$V$ be a deterministic real diagonal matrix satisfying the Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}. Denote by $\varrho^N_{{H^X},n}, \varrho^N_{H^Y,n}$ the $n$-point correlation functions of the eigenvalues with respect to the probability laws of the matrices $H^X$, $H^Y$ respectively. Then, for any energy $E$ in the interior of the support of $\rho_{fc}$ and any $n$-particle observable~$O$, we have
\begin{align*}
\lim_{N\to \infty} \int_{{\mathbb R}^k} \mathrm{d}\alpha_1\cdots\mathrm{d}\alpha_n\, O(\alpha_1,\ldots,\alpha_n)(\varrho^N_{H^X,n}-\varrho^N_{H^Y,n} )\left(E+\f{\alpha_1}{N},\ldots,E+\f{\alpha_n}{N} \right) = 0\,,
\end{align*}
for any fixed $n\in\N$.
\end{theorem}
Notice that this comparison theorem holds for any fixed energy $E$ in the bulk. The proof of~\cite{EYY1} applies almost verbatim. The only technical input in the proof are the local law for $m_N^X$, respectively $m_N^Y$, on scales $\eta\sim N^{-1+\epsilon}$, which we have established in Theorem~\ref{thm.strong}; see also~\eqref{eq of thm.strong}.
\subsection{Proof of Theorem~\ref{theorem 1}: }
In the remaining subsections we complete the proofs of our main results in Theorem~\ref{theorem 1} and Theorem~\ref{theorem 2}. The proofs for deterministic and random $V$ differ slightly. We start with the case of deterministic~$V$; the random case is treated in Subsection~\ref{subsection proof random V}.
\begin{proof}[Proof of theorem~\ref{theorem 1}]\label{proof of theorem 1}
Assume that $W=(w_{ij})$ is a complex Hermitian or a real symmetric Wigner matrix satisfying the assumptions in Definition~\ref{assumption wigner}. Let $V=\diag(v_i)$ be a deterministic real diagonal matrix satisfying the Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}. (Note that the event $\Omega_V$ then has full probability). Set $H=(h_{ij})=V+W$. Let $E\in\ensuremath{\mathbb{R}}$ be inside the support of $\rho_{fc}$. Note that by Lemma~\ref{hat mfc}, $E$ is also contained in the support of $\widehat\rho_{fc}$, for $N$ sufficiently large. (Here we have $\widehat\rho_{fc}=\widehat\rho_{fc}^{\vartheta=1}$ and similarly for $\rho_{fc}$.) Fix $\delta'>0$ and set $t\equiv N^{-1/2+\delta'}$. We first claim that there exists an auxiliary complex Hermitian or real symmetric Wigner matrix, $U=(u_{ij})$, satisfying the assumptions in Definition~\ref{assumption wigner} such that the following holds: Set
\begin{align}\label{the matrix Y}
Y\mathrel{\mathop:}= \e{-t/2} U+(1-\e{t/2}){W'}\,,
\end{align}
where $ W'$ is a GUE/GOE matrix independent of $W$, then the moments of the entries of $Y$ satisfy
\begin{align}\label{matching 3}
\E\,\overline{y}_{ij}^{p}\,{y}_{ij}^{3-p}=\E\,\overline{w}_{ij}^{p}\,{w}_{ij}^{3-p}\,,\qquad \qquad
\left|\E\,\overline{y}_{ij}^{q}\,{y}_{ij}^{4-q}-\E\,\overline{w}_{ij}^{q}\,{w}_{ij}^{4-q}\right|\le Ct\,,
\end{align}
for $p\in\llbracket 0,3\rrbracket$, $q\in\llbracket 0,4\rrbracket$, where $(w_{ij})$ are the entries of the Wigner matrix $W$.
Assuming the existence of such a Wigner matrix $U$, we choose $t_0\equiv t$ and set
\begin{align*}
H_t\mathrel{\mathop:}=\e{-(t-t_0)/2}V+\e{-t/2}U+(1-\e{-t})^{1/2}{W'}=V+\e{-t/2}U+(1-\e{-t})^{1/2}{W'}\,.
\end{align*}
Then the matrices $H_t$ and $H=V+W$ satisfy the matching conditions~\eqref{matching 1} and~\eqref{matching 2} of Theorem~\ref{the green function comparison theorem} (with, say, $\delta=1/4-2\delta'$). This follows from~\eqref{matching 3}. Thus Theorem~\ref{corollary of green function comparison} implies that the correlation functions of $H_t$ and $H$ agree in the limit of large $N$, i.e.,
\begin{align}\label{almost at the universality}
\lim_{N\to \infty}\int_{{\mathbb R}^n} \mathrm{d}\alpha_1\cdots\mathrm{d}\alpha_n\, O(\alpha_1,\ldots,\alpha_n)\bigg[&\frac{1}{2b}\int_{E-b}^{E+b}\,\frac{\mathrm{d} x}{\rho_{fc}(E)^n}\varrho^N_{H_t,n}\Big(x+\f{\alpha_1}{\rho_{fc}(E)N},\ldots,x+\f{\alpha_n}{\rho_{fc}(E)N} \Big)\nonumber\\ &-\frac{1}{2b}\int_{E-b}^{E+b}\,\frac{\mathrm{d} x}{\rho_{fc}(E)^n}\varrho^N_{H,n}\Big(x+\f{\alpha_1}{\rho_{fc}(E)N},\ldots,x+\f{\alpha_n}{\rho_{fc}(E)N} \Big)\bigg] = 0\,,
\end{align}
where $(\varrho_{H,n}^{N})$ denote the correlation functions of $H=V+W$ and where $(\varrho_{H_t,n}^N)$ denote the correlation functions of $H_t$. (In fact,~\eqref{almost at the universality} holds even without the averages over $E$.)
On the other hand, for small $\delta>0$, Theorem~\ref{theorem translation} assures that the local correlation functions of the matrix $H_t$ agree with the correlation functions of the GUE (respectively GOE), when averaged over an interval of size $b$, with $1\gg b\ge N^{-\delta}$, i.e., for any $E'$ with $|E'|<2$,
\begin{align}\label{almost at the universality 2}
\lim_{N\to \infty}\int_{{\mathbb R}^n} \mathrm{d}\alpha_1\cdots\mathrm{d}\alpha_n\, O(\alpha_1,\ldots,\alpha_n)\bigg[&\frac{1}{2b}\int_{E-b}^{E+b}\,\frac{\mathrm{d} x}{\rho_{fc}(E)^n}\varrho^N_{H_t,n}\Big(x+\f{\alpha_1}{\rho_{fc}(E)N},\ldots,x+\f{\alpha_n}{\rho_{fc}(E)N} \Big)\nonumber\\ &-\frac{1}{\rho_{sc}(E')^n}\varrho^N_{G,n}\Big(E'+\f{\alpha_1}{\rho_{sc}(E')N},\ldots,E'+\f{\alpha_n}{\rho_{sc}(E')N} \Big)\bigg] = 0\,,
\end{align}
where $(\varrho_{G,n}^N)$ denote the correlations functions of the GUE, respectively GOE. Combining~\eqref{almost at the universality} and~\eqref{almost at the universality 2}, we get~\eqref{theorem 1 equation 1}.
Thus to conclude the proof we need to show the existence of a Wigner matrix $U$ with the properties described above.
For a real random variables $\zeta$, denote by $m_k(\zeta)=\E\, \zeta^k$, $k\in\N$, its moments.
\begin{lemma}\emph{[Lemma 6.5 in~\cite{EYY1}]}\label{existence of matching}
Let $m_3$ and $m_4$ be two real numbers such that
\begin{align*}
m_4 - m_2^2-1\ge 0\,,\qquad m_4\le C_1\,,
\end{align*}
for some constant $C_1$. Let $\zeta_G$ be a Gaussian random variable with mean $0$ and variance $1$. Then
for any sufficient small $\gamma>0$, depending on $C_1$, there exists a real random variable $\zeta_\gamma$ with subexponential
decay and independent of $\zeta_G$, such that the first three moments of
\begin{align*}
\zeta'\mathrel{\mathop:}= (1 - \gamma)^{1/2} \zeta_\gamma + \gamma^{1/2} \zeta_G
\end{align*}
are $m_1 (\zeta' ) = 0$, $m_2 (\zeta') = 1$ , $m_3 (\zeta') = m_3$, and the forth moment $m_4 (\zeta')$ satisfies
\begin{align*}
|m_4 (\zeta' )-m_4 | \le C\gamma\,,
\end{align*}
for some $C$ depending on $C_1$.
\end{lemma}
Since the real and imaginary parts of $W$ are independent, it is sufficient
to match them individually, i.e., we apply Lemma~\ref{existence of matching} separately to the real and imaginary parts of $(w_{ij})$. This concludes the proof of Theorem~\ref{theorem 1} for deterministic~$V$.
\end{proof}
\subsection{Proof of Theorem~\ref{theorem 2}}\label{subsection proof random V}
Next, we prove Theorem~\ref{theorem 2}. Assume that $W=(w_{ij})$ is a complex Hermitian or a real symmetric Wigner matrix satisfying the assumption in Definition~\ref{assumption wigner}. Let $V=\diag(v_i)$ be a random real diagonal matrix satisfying the Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}. Denote by $\widetilde{f}_t\,\mu_G$ the distribution of the eigenvalues of the matrix
\begin{align*}
H_t\mathrel{\mathop:}= V+\e{-t/2}W+(1-\e{t})^{1/2}{W'}\,, \qquad (t\ge 0)\,,
\end{align*}
where $W'$ is a GUE/GOE matrix independent of $V$ and $W$. Let $\E^V$ stand for the expectation with respect to the law of the entries $(v_i)$ of $V$. Recall the definition of the event $\Omega_V$ in Definition~\ref{definition of omega_V}. Following the notation of the previous sections, $f_t\,\mu_G\equiv f_t^V\,\mu_G$ denotes the density conditioned on $V$. For an $n$-particle observable $O$ and for $G_{j,{\boldsymbol m}}$ as in~\eqref{generalized observable}, we may write
\begin{align*}
\int \frac{1}{|J|}\sum_{j\in J}G_{j,{\boldsymbol m}}(\mathbf{x}) \,\widetilde{f}_t(\mathbf{x})\mathrm{d}\mu_G(\mathbf{x})=\E^V\left[\left(\int \frac{1}{|J|}\sum_{j\in J}G_{j,{\boldsymbol m}}(\mathbf{x}) \,f^V_t(\mathbf{x})\mathrm{d}\mu_G(\mathbf{x})\right)\lone(\Omega_V)\right]+{\mathcal O}\left( N^{-\mathfrak{t}}\right)\,,
\end{align*}
where $\mathfrak{t}>0$ is the constant in~\eqref{eq assumption mu_V 2} of Assumptions~\ref{assumption mu_V}. Here we used the definition of $\Omega_V$; see Definition~\ref{definition of omega_V}. Since $(v_i)$ are i.i.d.,~\eqref{definition of omega_V} holds with exponentially high probability. The estimate~\eqref{definition of omega_V III} holds with probability large than $1-N^{-\mathfrak{t}}$ by Assumption~\ref{assumption mu_V}. Hence $\mathbb{P}^V(\Omega_V^c)\le cN^{-\mathfrak{t}}$, for some $c>0$ and $N$ sufficiently large.
Using Theorem~\ref{generalized gap distribution}, we find that
\begin{align}\label{on the way for random V 1}
\int \frac{1}{|J|}\sum_{j\in J}G_{j,{\boldsymbol m}}(\mathbf{x}) \,\widetilde{f}_t(\mathbf{x})\mathrm{d}\mu_G(\mathbf{x})&=\int \frac{1}{|J|}\sum_{j\in J}G_{j,{\boldsymbol m},sc}(\mathbf{x}) \mathrm{d}\mu_G(\mathbf{x})+{\mathcal O}\left( N^{-\mathfrak{f}}\right)+{\mathcal O}\left( N^{-\mathfrak{t}}\right)\,,
\end{align}
where we used once more the estimates on the event $\Omega_V$. Here $\mathfrak{f}>0$ is the constant appearing in Theorem~\ref{generalized gap distribution}.
To establish the equivalent result to Theorem~\ref{theorem translation}, we need a local deformed semicircle law for the setting when the entries $(v_i)$ of $V$ are not fixed. Recall that we denote by $m_{fc}$ the Stieltjes transform of the deformed semicircle law $\rho_{fc}=\rho_{fc}^{\vartheta=1}$.
\begin{lemma}{\emph{[Theorem~2.10 and Theorem~2.21 in~\cite{LS}]}}\label{local law random V}
Let $W$ be a complex Hermitian or a real symmetric Wigner matrix satisfying the assumptions in Definition~\ref{assumption wigner}. Let $V$ be a random real diagonal matrix satisfying Assumptions~\ref{assumption mu_V convergence} and~\ref{assumption mu_V}. Set $H\mathrel{\mathop:}= V+W$, $G(z)\mathrel{\mathop:}= (H-z)^{-1}$ and $m_N(z)\mathrel{\mathop:}= N^{-1}\Tr G(z)$, $(z\in \ensuremath{\mathbb{C}}^+)$. Let $\xi=A_0\log\log N/2$; see~\eqref{eq.xi}. Then there exists $\upsilon>0$ and $c$ (both depending on the constants in~\eqref{eq.C0}, the constants $A_0$, $E_0$ in~\eqref{eq.DL} and the measure~$\nu$), such that for $L\ge 40\xi$, we have
\begin{align}\label{local law random V equation}
|m_N(z)-m_{fc}(z)|\le (\varphi_N)^{c\xi}\left(\min\left\{\frac{1}{N^{1/4}},\frac{1}{\sqrt{\kappa_E+\eta}}\frac{1}{\sqrt{N}}\right\}+\frac{1}{N\eta}\right)\,,
\end{align}
and
\begin{align}\label{lcal law random V 2}
\left| G_{ij}(z)-\delta_{ij} g_{i}(z) \right|\le (\varphi_N)^{c\xi}\left(\sqrt{\frac{\im m_{fc}(z)}{N\eta}}+\frac{1}{N\eta}\right)\,,\qquad\qquad (i,j\in\llbracket 1,N\rrbracket)\,,
\end{align}
with $(\xi,\upsilon)$-high probability, for all $z=E+\mathrm{i}\eta\in{\mathcal D}_L$; (see~\eqref{eq.DL}). Here, we have set
\begin{align*}
{g}_i(z)\mathrel{\mathop:}=\frac{1}{ v_i-z-{m}_{fc}(z)}\,,\qquad \qquad \quad(z\in\ensuremath{\mathbb{C}}^+\,,i\in\llbracket 1,N\rrbracket)\,.
\end{align*}
Moreover, fixing $\alpha>0$, there is $c_1$ (depending on the constants in~\eqref{eq.C0}, the constants $A_0$, $E_0$ in~\eqref{eq.DL}, the measure $\nu$ and $\alpha$), such that
\begin{align}\label{rigidity random V}
|\lambda_i-\gamma_i|\le (\varphi_N)^{c_1\xi}\frac{1}{\sqrt{N}}\,,
\end{align}
with $(\xi,\upsilon)$-high probability, for all $i\in\llbracket \alpha N,(1-\alpha)N\rrbracket$. Here $(\lambda_i)$ denote the eigenvalues of $H=V+W$ and~$(\gamma_i)$ are their classical locations with respect the deformed semicircle law $\rho_{fc}$.
\end{lemma}
Using the local law in~Lemma~\ref{local law random V}, we obtain from~\eqref{on the way for random V 1} the equivalent results to Theorem~\ref{theorem translation}.
\begin{theorem}\label{theorem translation random}
Fix $n\in\N$ and consider an $n$-particle observable $O$. Fix $\delta>0$ and let $t \ge N^{-1/4+\delta}$. Let $\widetilde\alpha>0$ be a small constant and consider two energies $E\in[L_-(t)+\widetilde\alpha,L_+(t)-\widetilde\alpha]$ and $E'\in[-2+\widetilde\alpha,2-\widetilde\alpha]$. Then, for any $\epsilon>0$ and for $b\equiv b_N$ satisfying $\widetilde\alpha/2\ge b_N>0$, we have
\begin{align}\label{equation theorem translation 2}
\bigg|\int_{\ensuremath{\mathbb{R}}^n}\mathrm{d}\alpha_1\cdots\mathrm{d}\alpha_n\, O(\alpha_1,\ldots,\alpha_n)&\bigg[\int_{E-b}^{E+b}\frac{\mathrm{d} x}{2b}\frac{1}{\rho_{fc}(t,E)^n}\,\varrho^N_{\widetilde{f}_t,n}\Big(x+\frac{\alpha_1}{N\rho_{fc}(t,E)},\ldots,x+\frac{\alpha_n}{N\rho_{fc}(t,E)} \Big)\nonumber\\
&-\int_{E'-b}^{E'+b}\frac{\mathrm{d} x}{2b}\frac{1}{\rho_{sc}(E')^n}\,\varrho^N_{G,n}\Big(x+\frac{\alpha_1}{N\rho_{sc}(E')},\ldots,x+\frac{\alpha_n}{N\rho_{sc}(E')} \Big)\bigg]\bigg|\nonumber\\&\qquad\qquad\le C_ON^{2\epsilon}\left(b^{-1} N^{-1/2+\epsilon} +N^{-\mathfrak{f}}+N^{-\mathfrak{t}}+N^{-1/4}\right)\,,
\end{align}
for $N$ sufficiently large. Here $\mathfrak{f}>0$ is the constant in Theorem~\ref{generalized gap distribution}. Moreover, $\rho_{fc}(E)$ stands for the density of the ($N$-independent) measure $\rho_{fc}$ at the energy $E$. The constant $C_O$ depends on $O$, $\widetilde\alpha$ and the measure $\nu$. The constant~$\mathfrak{f}$ depends on $\delta$ and $\widetilde\alpha$.
\end{theorem}
The proof of Theorem~\ref{theorem translation random} is an application of Section~7 in~\cite{ESYY}. The validity of Assumption~IV in~\cite{ESYY} is a direct consequence of the local law in Lemma~\ref{local law random V}. Here and also below, we use that the local laws of~Lemma~\ref{local law random V} are only used on very small scales $\eta\sim N^{-1+\epsilon}$ in the bulk. For such small $\eta$ the first error term in~\eqref{local law random V equation} is negligible compared to the second error term. Also note that the first term on the right side of~\eqref{equation theorem translation 2} is bigger than the corresponding term in~\eqref{equation theorem translation}. This is due to the weaker rigidity bounds in case $V$ is random; see~\eqref{rigidity random V}. We therefore have to impose that $b\gg N^{-1/2}$ in order to have a vanishing error term in the limit of large $N$. Finally, we mention that the error term $C_ON^{2\epsilon}N^{-1/4}$ stems
from replacing $\widehat\rho_{fc}(t,E)$ by $\rho_{fc}(t,E)$; see the comment below Theorem~\ref{theorem translation}.
\begin{proof}[Proof of Theorem~\ref{theorem 2}]
The proof Theorem~\ref{theorem 2} follows now along the lines of the proof of Theorem~\ref{theorem 1}. First, we check that the Green function comparison Theorem~\ref{the green function comparison theorem} holds true for $H^X=V+X$, respectively, $H^Y=V+Y$ with random $V$. This is indeed the case, since the only input we used is the estimate~\eqref{the estimate of the green function comparison}, which also holds for random $V$ by the local laws in Lemma~\ref{local law random V} and the stability estimate~\eqref{stability bound}. Note that we are using that the bound~\eqref{the estimate of the green function comparison} is only required on scales $\eta\ll N^{-1/2}$. Similarly, we can establish Theorem~\ref{corollary of green function comparison} for random $V$ using the Green function comparison theorem for random $V$, the local laws in Lemma~\ref{local law random V} and the stability estimate~\eqref{stability bound}. Finally, we note that the construction of the
matrix $U$ and $Y$ (see~\eqref{the matrix Y}) and the moment matching in~\eqref{matching 3} do not involve $V$. We can thus complete the proof of Theorem~\ref{theorem 2} in the same way as the proof of Theorem~\ref{theorem 1}.
\end{proof}
\begin{appendix}
\section{}
In this appendix we prove the auxiliary results used in Sections~\ref{section local law} and~\ref{section beta ensemble}: Lemma~\ref{lemma mfc}, Lemma~\ref{lemma hatmfc} and Lemma~\ref{superlemma}. We start with a more extended version of Lemma~\ref{lemma mfc}. Recall from~\eqref{definition Theta} that we denote $\Theta_{\varpi}=[0,1+\varpi']$, $\varpi'=\varpi/10$. Also recall the definition of the domain ${\mathcal D}'$ of the spectral parameter $z$ in~\eqref{definition of D'}.
\begin{lemma}\label{lemma mfc A}
Let $\nu$ satisfy Assumption~\ref{assumption mu_V} for some $\varpi>0$. Then the following holds true for any $\vartheta\in\Theta_{\varpi}$. There are $L_-^{\vartheta},L_+^{\vartheta}\in\ensuremath{\mathbb{R}}$, with $L_-^{\vartheta}<0<L_+^{\vartheta}$, such that $\supp \rho_{fc}^{\vartheta}=[L_-^{\vartheta},L_+^{\vartheta}]$ and there is a constant $C>1$ such that, for all $\vartheta\in\Theta_{\varpi}$,
\begin{align}\label{the square root A}
C^{-1}\sqrt{\kappa_E}\le \rho_{fc}^{\vartheta}(E)\le C \sqrt{\kappa_E},\qquad\qquad(E\in[L_-^{\vartheta},L_+^{\vartheta}])\,,
\end{align}
where $\kappa_E$ denotes the distance of $E$ to the endpoints of the support of $\rho_{fc}^{\vartheta}$, i.e.,
\begin{align}\label{definition of kappaE}
\kappa_E\mathrel{\mathop:}=\min\{|E-L_-^{\vartheta}|,\,|E-L_+^{\vartheta}|\}\,.
\end{align}
The Stieltjes transform, $m^{\vartheta}_{fc}$, of $\rho^{\vartheta}_{fc}$ has the following properties,
\begin{itemize}
\item[$(1)$] for all $z=E+\mathrm{i}\eta\in\ensuremath{\mathbb{C}}^+$,
\begin{align}\label{behavior of mfc A}
\im m^{\vartheta}_{fc}(z)\sim\begin{cases} \sqrt{\kappa+\eta}\,,\qquad& E\in[L_-,L_+]^{\phantom{c}}\,,\\
\frac{\eta}{\sqrt{\kappa+\eta}}\,,\quad &E\in[L_-,L_+]^c\,;
\end{cases}
\end{align}
\item[$(2)$] there exists a constant $C>1$ such that for all $z\in{\mathcal D}'$ and for all $x\in I_{\nu}$,
\begin{align}\label{stability bound A'}
C^{-1}\le |\vartheta x -z-m^{\vartheta}_{fc}(z)|\le C\,;
\end{align}
\item[$(3)$] there exists a constant $C>1$ such that for all $z\in{\mathcal D}'$,
\begin{align}
C^{-1}\sqrt{\kappa+\eta}\le\left|1-\int\frac{\mathrm{d}\nu(v)}{(\vartheta v-z-m^{\vartheta}_{fc}(z))^2}\right|\le C\sqrt{\kappa+\eta}\,;
\end{align}
\item[$(4)$] there are constants $C>1$ and $c_0>0$ such for all $z=E+\mathrm{i}\eta\in{\mathcal D}'$ satisfying $\kappa_E+\eta\le c_0$,
\begin{align}
C^{-1}\le\left|\int\frac{\mathrm{d}\nu(v)}{(\vartheta v-z-m^{\vartheta}_{fc}(z))^3}\right|\le C\,,
\end{align}
moreover, there is $C>1$, such that for all $z\in{\mathcal D}'$,
\begin{align}
\left|\int\frac{\mathrm{d}\nu(v)}{(\vartheta v-z-m^{\vartheta}_{fc}(z))^3}\right|\le C\,.
\end{align}
\end{itemize}
The constants in the statements $(1)$-$(4)$ can be chosen uniformly in $\vartheta\in\Theta_{\varpi}$.
\end{lemma}
\begin{proof}
We follow the proofs in~\cite{S1,LS}. Let $\vartheta\in\Theta_{\varpi}$. Set $\zeta=z+m_{fc}^{\vartheta}(z)$ and let
\begin{align}
F(\zeta)\mathrel{\mathop:}=\zeta-\int\frac{\mathrm{d}\nu(v)}{ \vartheta v-\zeta}\,,\qquad\qquad(\zeta\in\ensuremath{\mathbb{C}}^+)\,.
\end{align}
Then the functional equation~\eqref{lambda mfc} is equivalent to $z=F(\zeta)$. As is argued in~\cite{S1}, a point $E\in\ensuremath{\mathbb{R}}$ is inside the support of the measure $\rho_{fc}^{\vartheta}$ iff $\zeta_E=E+m_{fc}^{\vartheta}(E)$ satisfies $\im F(\zeta_E)=0$ and $\im \zeta_E> 0$. Accordingly, the endpoints of the support are characterized as the solutions of
\begin{align}\label{equation for zetapm A}
H(\zeta)\mathrel{\mathop:}= \int\frac{\mathrm{d}\nu(v)}{(\vartheta v-\zeta)^2}=1\,,\qquad\qquad(\zeta\in\ensuremath{\mathbb{R}})\,.
\end{align}
Note that $H(\zeta)$ is a continuous function outside $\vartheta\, I_{\nu}\equiv\{x\,:\,x=\vartheta y, y\in I_{\nu}\}$ which is decreasing as $|\zeta|$ increases. Since $\vartheta\in\Theta_{\varpi}=[0,1+\varpi']$, with $\varpi'=\varpi/10$, we obtain from Assumption~\ref{assumption mu_V} that $H(\zeta)\ge 1+\varpi/2$, for all $\zeta\in \vartheta\, I_{\nu}$. It thus follows that there are only two solutions, $\zeta_{\pm}^{\vartheta}\in \ensuremath{\mathbb{R}}\backslash \vartheta\, I_{\nu}$, to $H(\zeta)=1$, $\zeta\in\ensuremath{\mathbb{R}}$. In particular, $\zeta_-^{\vartheta}<0$, $\zeta_+^{\vartheta} >0$, and there is a constant $\mathfrak{g}>0$, depending only on $\nu$, such that
\begin{align}\label{stability bound A}
\inf_{\vartheta\in\Theta_{\varpi}}\mathrm{dist}(\{\zeta_\pm^{\vartheta}\}, \vartheta\, I_{\nu})\ge \mathfrak{g}\,.
\end{align}
As argued in~\cite{S1,LS}, the set $\gamma\mathrel{\mathop:}=\{\zeta\in \ensuremath{\mathbb{C}}^+\,:\, \im F(\zeta)=0\,,\im\zeta> 0 \}$ is, for each fixed $\vartheta\in\Theta_{\varpi}$, a finite curve in the upper half plane that is the graph of a continuous function which only connects to the real line at $\zeta_\pm^{\vartheta}$.
Since $\mathrm{dist}(\{\zeta_{\pm}^{\vartheta}\}, \vartheta\, I_{\nu}\}\ge\mathfrak{g}>0$, $F(\zeta)$ is analytic in a neighborhood of~$\zeta_{\pm}^{\vartheta}$. Thus for $\zeta$ in a neighborhood of~$\zeta_+^{\vartheta}$, we may write
\begin{align*}
F(\zeta)=F(\zeta_+^{\vartheta})+F'(\zeta_+^{\vartheta})(\zeta-\zeta_+^{\vartheta})+\frac{F''(\zeta_{+}^{\vartheta})}{2}(\zeta-\zeta_{+}^{\vartheta})^2+{\mathcal O}((\zeta-\zeta_{+}^{\vartheta})^3) \,.
\end{align*}
Note that $F'(\zeta_+^{\vartheta})=0$ by the definition of $\zeta_{+}^{\vartheta}$. Moreover, we know that $\im F(\zeta)=0$, for~$\zeta$ in a real neighborhood of~$\zeta_{+}^{\vartheta}$, but we also have $\im F(\zeta)=0$, for~$\zeta\in \gamma\cup\overline\gamma$. Thus $F''(\zeta_{+}^{\vartheta})\not=0$. We can therefore invert $F(\zeta)=z$ in a neighborhood of~$\zeta_{+}$, to obtain
\begin{align}\label{inverted equation}
\zeta(z)=F^{(-1)}(z)=\zeta_{+}^{\vartheta}+c_{+}^{\vartheta}\sqrt{z-L_+^{\vartheta}}\left(1+{\mathcal A}_{+}^{\vartheta}\left(\sqrt{z-L_{+}^{\vartheta}}\right)\right)\,,
\end{align}
(with the convention $\im F^{(-1)}(z)\ge 0$), where $L_+^{\vartheta}$ is defined by $\zeta_{+}^{\vartheta}=L_+^{\vartheta}+m_{fc}(L_+^{\vartheta})$. Here, $c^{\vartheta}_+>0$ is a real constant and ${\mathcal A}_+^{\vartheta}$ is an analytic function that is real-valued on the real line and that satisfies ${\mathcal A}^{\vartheta}_{+}(0)=0$. Recalling that $\zeta(z)=z+m_{fc}^{\vartheta}(z)$ and taking the limit $\eta\to 0$ we obtain~\eqref{the square root A}, for fixed $\vartheta$. To achieve uniformity in $\vartheta$, we use the (uniform) stability bound~\eqref{stability bound A} and the (pointwise) positivity of $|F''(\zeta_+^{\vartheta})|$: We differentiate~\eqref{equation for zetapm A} with respect to $\vartheta$ and observe that $\partial_{\vartheta}H(\zeta,\vartheta)|_{\zeta=\zeta_{+}^{\vartheta}}\not=0$, for all $\vartheta\in\Theta_{\varpi}$, since $F''(\zeta_{+}^{\vartheta})\not=0$. Thus by the implicit function theorem, $\zeta_
{+}^{\vartheta}$ is a $C^1$ function of $\vartheta\in\Theta_{\varpi}$. Next, we observe that $F''(\zeta)$ is an analytic function of~$\zeta$, for~$\zeta$ away from~$\vartheta\, I_{\nu}$, thus, using once more~\eqref{stability bound A}, we can bound $|F''(\zeta_{+}^{\vartheta})|\ge c$, for some $c>0$, uniformly in $\vartheta\in\Theta_{\varpi}$. In fact, $F^{(n)}(\zeta_{+}^{\vartheta})$, $n\in\N$, are all continuous functions of $\vartheta\in\Theta_{\varpi}$, and we can bound them uniformly in $\vartheta$ for each $n\in\N$. Repeating the same argument for $\zeta$ close to $\zeta_-^{\vartheta}$, we conclude the proof of~\eqref{the square root A}.
Statement $(2)$ follows from~\eqref{stability bound A} for $z$ close to the edges. For $z$ away from the edges, Assumption~\ref{assumption mu_V} assures that the curve $\gamma$ stays away from the real line for all $\vartheta\in\Theta_{\varpi}$ as is readily checked. This implies the stability bound for that region.
For the proofs of the remaining statements, we refer to the Appendix of~\cite{LS}.
\end{proof}
Next, we prove Lemma~\ref{lemma hatmfc}.
\begin{proof}[Proof of Lemma~\ref{lemma hatmfc}]
It follows from Assumption~\ref{assumption mu_V} that, on $\Omega_V$ for all $N$ sufficiently large,
\begin{align}\label{assumptions for proof}
\inf_{x\in I_{\widehat\nu}}\frac{1}{N}\sum_{i=1}^N\frac{1}{(\vartheta v_i-x)^2}\ge 1+\varpi/2\,,
\end{align}
for all $\vartheta\in\Theta_{\varpi}=[0,1+\varpi/10]$. The analogous statements of Lemma~\ref{lemma mfc A}, holding on $\Omega_V$ for $N$ sufficiently large, follow in the same way as in the proof of that lemma. To get uniformity in $N$, it suffices to check that the analogous expression to~\eqref{stability bound A} holds uniformly in $N$, for $N$ sufficiency large: By~\eqref{assumptions for proof} there are two real solutions $\widehat\zeta_{\pm}^{\vartheta}$ to $\widehat H(\zeta)\mathrel{\mathop:}=\frac{1}{N}\sum_{i=1}^N\frac{1}{(\vartheta v_i-\zeta)^2}=1$ that both lie outside of the interval~$\vartheta I_{\widehat\nu}$. Thus~\eqref{definition of omega_V} and~\eqref{definition of omega_V III}, imply that
\begin{align}\label{stability bound B}
\inf_{\vartheta\in\Theta_{\varpi}}\mathrm{dist}(\{\widehat\zeta_\pm^{\vartheta}\}, \vartheta\, I_{\widehat\nu})\ge \mathfrak{g}/2\,,
\end{align}
on $\Omega_V$ for all $N$ sufficiently large. Then we can bound~$\widehat F''(\zeta)=-\frac{2}{N}\sum_{i=1}^N \frac{1}{(\vartheta v_i-\zeta)^3}$, evaluated at $\widehat\zeta_{\pm}^{\vartheta}$, uniformly below in~$\vartheta$ and~$N$, for~$N$ sufficiently large, implying the uniformity in $N$ of the constants in the statements $(1)$-$(4)$.
Next we prove~\eqref{mini local law}. For simplicity we drop $\vartheta$ from the notation and work on $\Omega_V$. As above, set $\zeta=z+m_{fc}(z)$ and $\widehat\zeta=z+\widehat m_{fc}(z)$. From the definitions of $F$, $\widehat F$, and the Equations~\eqref{lambda mfc},~\eqref{hat mfc}, we have $\widehat F(\widehat\zeta)=F(\zeta)=z$, for all $z\in{\mathcal D}'$. Using the stability bound~\eqref{stability bound B} and Equation~\eqref{definition of omega_V} in the definition of $\Omega_V$, we get, assuming that $|\widehat\zeta-\zeta|\ll 1$,
\begin{align}\label{mini perturbation expansion}
[F'(\zeta)+{\mathcal O}(N^{-\alpha_0})]\left(\widehat\zeta-\zeta \right)+\frac{F''(\zeta)}{2}\left(\widehat\zeta-\zeta\right)^2=o(1)\left(\widehat\zeta-\zeta\right)^2+{\mathcal O}(N^{-\alpha_0})\,,
\end{align}
uniformly in $\vartheta\in\Theta_{\varpi}$, for all $z\in{\mathcal D}'$. From Lemma~\ref{lemma mfc A}, we get $F'(\zeta)\sim \sqrt{\kappa+\eta}$ and $F''(\zeta)\le C$, for all $z\in{\mathcal D}'$. We abbreviate $\Lambda\mathrel{\mathop:}=|\widehat\zeta-\zeta |$ in the following.
We first consider $z=E+\mathrm{i}\eta\in{\mathcal D}'$, such that $\kappa_E+\eta>N^{-\epsilon}$, for some small $\epsilon>0$ (with $\epsilon<\alpha_0$). Here $\kappa_E$ is defined in~\eqref{definition of kappaE}. For such $z$ we obtain from~\eqref{mini perturbation expansion} that $ \Lambda\le CN^{\epsilon}(\Lambda^2+N^{-\alpha_0})$. Thus either $\Lambda\le C_0 N^{\epsilon} N^{-\alpha_0}$ or $C_0N^{-\epsilon}\le \Lambda$, for some constant $C_0$. We now show that $|\Lambda|\le C_0 N^{\epsilon}N^{-\alpha_0}$, for all $z\in{\mathcal D}'$ such that $\kappa_E+\eta\ge N^{-\epsilon}$. For $z\in{\mathcal D}'$ with $\eta=2$, we have
\begin{align*}
\widehat\zeta(z)-\zeta(z)&=\frac{1}{N}\sum_{i=1}^N\left[\frac{1}{\vartheta v_i-\widehat\zeta(z)}-\frac{1}{\vartheta v_i-\zeta(z)}\right]+\vartheta^{-1}m_{\widehat\nu}\left(\zeta(z))/\vartheta\right)-\vartheta^{-1} m_{\nu}\left(\zeta(z)/\vartheta\right)\\
&=\frac{1}{N}\sum_{i=1}^N\frac{\widehat\zeta(z)-\zeta(z)}{(\vartheta v_i-\widehat\zeta(z) )(\vartheta v_i-\zeta(z))} +{\mathcal O}(N^{-\alpha_0})\,,
\end{align*}
where we used~\eqref{definition of omega_V}. Since $\eta=2$ and $\im\widehat \zeta,\,\im \zeta\ge \eta$, we obtain $\Lambda\le\frac{1}{4}\Lambda+{\mathcal O}(N^{-\alpha_0})$, i.e., $\Lambda(z)\le C N^{-\alpha_0}$, for $\eta=2$. To extend the conclusion to all $\eta$, we use the Lipschitz continuity of $\widehat\zeta (z)$, respectively $\zeta (z)$. Differentiating $z= F(\zeta)$, with respect to $z$ we obtain $\partial_z\zeta=( F'(\zeta))^{-1}$. Thus, using property $(2)$ of Lemma~\ref{lemma mfc A} we infer that the Lipschitz constant of $\zeta(z)$ is, for $z\in{\mathcal D}'$ satisfying $\kappa_E+\eta>N^{-\epsilon}$, bounded above by $N^{\epsilon/2}$. The same conclusion also holds for $\widehat\zeta(z)$. Bootstrapping we obtain
\begin{align}\label{mini local law bis}
|\widehat\zeta(z)-\zeta(z)|\le CN^{\epsilon}N^{-\alpha_0}\,,
\end{align}
on $\Omega_V$ for $N$ sufficiently large, for all $z\in {\mathcal D}'$ satisfying~$\kappa_E+\eta> N^{-\epsilon}$.
In order to control $\widehat\zeta(z)-\zeta(z)$ for $z=E+\mathrm{i}\eta\in{\mathcal D}'$ with $\kappa_E+\eta\le N^{-\epsilon}$, $\epsilon>0$, we first derive the estimate $|\widehat{L}_\pm^{\vartheta}-L_\pm^{\vartheta}|\le C N^{-\alpha_0}$, for some $c>0$, on $\Omega_V$. We recall that $\widehat{L}_\pm$, respectively $L_\pm$, are obtained through the relations
\begin{align*}
\frac{1}{N}\sum_{i=1}^N\frac{1}{(\vartheta v_i-\widehat\zeta_\pm)^2}=1\,,\qquad \int\frac{\mathrm{d} \nu(v)}{(\vartheta v-\zeta_\pm)^2}=1\,.
\end{align*}
Then a similar argument as given above shows that $|\widehat\zeta_\pm-\zeta_\pm|\le C N^{-\alpha_0}$ and $|\widehat L_\pm-L_\pm|\le CN^{-\alpha_0}$ on $\Omega_V$, $N$ sufficiently large. We refer to the Appendix C of~\cite{LS2} for details.
Second, following the arguments in the proof of Lemma~\ref{lemma mfc A}, we may write, for $\widehat\zeta$ and $\zeta$ in a neighborhood of $\zeta_\pm$,
\begin{align}\label{mini local law tres}
\widehat\zeta(z)-\widehat\zeta_\pm=\widehat c_\pm\sqrt{z-\widehat L_\pm}(1+{\mathcal O}(z-\widehat L_\pm))\,,\qquad \zeta(z)-\zeta_\pm= c_+\sqrt{z- L_\pm}(1+{\mathcal O}(z- L_\pm))\,.
\end{align}
We therefore get $|\widehat \zeta(z)-\zeta(z)|\le C\sqrt{\kappa_E+\eta}+CN^{-\alpha_0/2}$. Note that the constants can be chosen uniformly in $\vartheta\in\Theta_{\varpi}$. Choosing, e.g., $\epsilon=\alpha_0/4$, we get from~\eqref{mini local law bis} and~\eqref{mini local law tres} the desired inequality~\eqref{mini local law}.
\end{proof}
We now move on to the construction of the potentials $\widehat U$ and $U$. We first record the following corollary of Lemma~\ref{lemma mfc A}. Set $\mathrm{B}_{r}(p)\mathrel{\mathop:}=\{z\in\ensuremath{\mathbb{C}}\,:\,|z-p|<r\}$. Recall the conventions in~\eqref{abuse of notation I} and the definition of $\kappa_E$ in~\eqref{definition of kappaE}.
\begin{corollary}\label{remark about real and imaginary part of mfc}
Under the assumptions of Lemma~\ref{lemma mfc} (see also Lemma~\ref{lemma mfc A}), there are constants $c_+^{\vartheta}$, $r_+>0$, such that for any $E\in \mathrm{B}_{r_+}(L_+^{\vartheta})\cap\ensuremath{\mathbb{R}}$,
\begin{align}\label{real and imaginary part of mfc I}
\im m_{fc}^{\vartheta}(E)=\begin{cases} \sqrt{\kappa_E}(c_{+}^{\vartheta}+{\mathcal B}_{+}^{\vartheta}({-\kappa_E}))\,, \quad&\, E\le L_+^{\vartheta}\,,\\ 0\,,&E\ge L_+^{\vartheta}\,, \end{cases}
\end{align}
and
\begin{align}\label{real and imaginary part of mfc II}
\re m_{fc}^{\vartheta}(E)=\begin{cases}{\mathcal C}_+^{\vartheta}(-\kappa_E)\,,\quad & E\le L_+^{\vartheta}\,,\\ \sqrt{\kappa_E}(c_{+}^{\vartheta}+{\mathcal B}_{+}^{\vartheta}({\kappa_E}))+{\mathcal C}_+^{\vartheta}(\kappa_E)\,,& E\ge L_+^{\vartheta}\,,\end{cases}
\end{align}
where ${\mathcal B}_+^{\vartheta},{\mathcal C}_+^{\vartheta}$ are analytic functions on $\mathrm{B}_{r_+}(0)$ that are real-valued on $\ensuremath{\mathbb{R}}$, and that satisfy ${\mathcal B}_+^{\vartheta}(0)=0$, $c_+^{\vartheta}+{\mathcal B}_+^{\vartheta}> 0$, respectively ${{\mathcal C}_+^{\vartheta}}<0$, on ${\mathrm{B}_{r_+}(0)}\cap\ensuremath{\mathbb{R}}$. Moreover, for all $z\in \mathrm{B}_{r_+}(L_+^{\vartheta})$, the functions ${\mathcal B}_+^{\vartheta},{\mathcal C}_+^{\vartheta}$, respectively $\im m_{fc}^{\vartheta}$, $\re m_{fc}^{\vartheta}$, are continuous in $\vartheta\in\Theta_{\varpi}$.
Similar statements hold at the lower edge $L_{-}^\vartheta$.
\end{corollary}
\begin{proof}
Fix $\vartheta\in\Theta_{\varpi}$. As argued in the proof of Lemma~\ref{lemma mfc A}, the function $F(\zeta)$ can be inverted locally around $\zeta_{\pm}^{\vartheta}$; see~\eqref{inverted equation} above. Thus for $\zeta$ in a neighborhood of $\zeta_+^{\vartheta}$, we may write
\begin{align*}
m^{\vartheta}_{fc}(z)=F^{-1}(z)-z&=\zeta_{+}^{\vartheta}-z+c_{+}^{\vartheta}\sqrt{z-L_+^{\vartheta}}\left(1+{\mathcal A}_{+}^{\vartheta}\left(\sqrt{z-L_{+}^{\vartheta}}\right)\right)\\
&=c_{+}^{\vartheta}\sqrt{z-L_+^{\vartheta}}\left(1+{\mathcal B}_+^{\vartheta}(z-L_+^{\vartheta})\right)+{\mathcal C}_{+}^{\vartheta}(z-L_{+}^{\vartheta})\,,
\end{align*}
for $z$ in $\mathrm{B}_{r}(L_+^{\vartheta})$, for some $r>0$, where ${\mathcal B}_{+}^{\vartheta}$ and ${\mathcal C}_{+}^{\vartheta}$ are analytic in a neighborhood of zero and real-valued on the real line, since $\im F^{-1}(E)=0$, for $E\in[L_-^{\vartheta},L_+^{\vartheta}]^c$. Equations~\eqref{real and imaginary part of mfc I} and~\eqref{real and imaginary part of mfc II} follow. From the proof of Lemma~\ref{lemma mfc A}, it is immediate that $c^{\vartheta}_+>0$. Thus $c^{\vartheta}_++{\mathcal B}^{\vartheta}_+>0$ in a real neighborhood of zero. Since $x-L_+^{\vartheta}-m_{fc}^{\vartheta}(L_+^{\vartheta})<0 $, for all $x\in\vartheta\,I_{\nu}$, we must have ${\mathcal C}_+^{\vartheta}<0$ in a real neighborhood of zero. Since $F(z)$ is analytic on $\mathrm{B}_{r}(L_+^{\vartheta})$, for all $\vartheta\in\Theta_{\varpi}$, and since $\zeta_{+}^{\vartheta}$ is a $C^1$ function of $\vartheta$, the functions ${\mathcal B}_{+}^{\vartheta}$ and ${\mathcal C}_{+}^{\vartheta}$ are $C^1$ in $\vartheta\in\Theta_{\varpi}$. Then, it
is clear from~\eqref{stability bound A} that we can choose $r>0$ uniformly in $\vartheta\in\Theta_{\varpi}$. The same arguments apply for $\zeta$ close to $\zeta_-^{\vartheta}$.
\end{proof}
The analogous result to Corollary~\ref{remark about real and imaginary part of mfc}, is stated next. Recall the notation $\widehat\kappa_E\mathrel{\mathop:}=\min\{|E-\widehat{L}_-^{\vartheta}|, |E-\widehat{L}_+^{\vartheta}|\}$.
\begin{corollary}\label{remark about real and imaginary part of hatmfc}
Under the assumptions of Lemma~\ref{lemma hatmfc} the following holds on $\Omega_V$, for $N$ sufficiently large. There are constants $\widehat c_+^{\vartheta}$, $r_+'$, with $r_+\ge r_+'>0$, such that for any $E\in \mathrm{B}_{r_+'}(L_+^{\vartheta})\cap\ensuremath{\mathbb{R}}$,
\begin{align}
\im \widehat m_{fc}^{\vartheta}(E)=\begin{cases} \sqrt{ \widehat\kappa_E}\big(\widehat c_{+}^{\vartheta}+\widehat{\mathcal B}_{+}^{\vartheta}({-\widehat \kappa_E})\big)\,, \quad&\, E\le \widehat L_+^{\vartheta}\,,\\ 0\,,&E\ge \widehat L_+^{\vartheta}\,, \end{cases}
\end{align}
and
\begin{align}
\re \widehat m_{fc}^{\vartheta}(E)=\begin{cases} \widehat{\mathcal C}_+^{\vartheta}(-\widehat\kappa_E)\,,\quad & E\le \widehat L_+^{\vartheta}\,,\\ \sqrt{\widehat\kappa_E}\big(\widehat c_+^{\vartheta}+\widehat {\mathcal B}_{+}^{\vartheta}({\widehat\kappa_E})\big)+ \widehat {\mathcal C}_+^{\vartheta}(\widehat\kappa_E)\,,& E\ge \widehat L_+^{\vartheta}\,,\end{cases}
\end{align}
where $ \widehat {\mathcal B}_+^{\vartheta}, \widehat {\mathcal C}_+^{\vartheta}$ are analytic functions on $\mathrm{B}_{r_+'}(0)$ that are real-valued on $\ensuremath{\mathbb{R}}$, and that satisfy
$\widehat{\mathcal B}_+^{\vartheta}(0)=0$ and~$\widehat c_+^{\vartheta}+\widehat{\mathcal B}_+^{\vartheta}> 0$, respectively ${\widehat{\mathcal C}_+^{\vartheta}}<0$, on ${\mathrm{B}_{r'_+}(0)}\cap\ensuremath{\mathbb{R}}$. Moreover, the constant $r_+'$ can be chosen independent of $\vartheta\in\Theta_{\varpi}$ and $N$, for $N$ sufficiently large.
Further, the functions $\widehat {\mathcal B}_+^{\vartheta},\,\widehat {\mathcal C}_+^{\vartheta}$, respectively $\im \widehat m_{fc}^{\vartheta}$, $\re \widehat m_{fc}^{\vartheta}$, are continuous functions in $\vartheta\in\Theta_{\delta}$, for all $z\in \mathrm{B}_{r_+'}(L_+^{\vartheta})$. There is $c>0$, such that
\begin{align}\label{convergence of B and C function}
|\widehat {\mathcal B}_+^{\vartheta}(z)- {\mathcal B}_+^{\vartheta}(z)|\le N^{-c\alpha_0/2}\,,\qquad \qquad| \widehat {\mathcal C}_+^{\vartheta}(z)- {\mathcal C}^{\vartheta}_+(z)|\le N^{-c\alpha_0/2}\,,
\end{align}
for all $ z\in \mathrm{B}_{r_+'}(L_+^{\vartheta})$ and all $\vartheta\in\Theta_{\varpi}$, on $\Omega_V$ for $N$ sufficiently large.
Similar statements hold at the lower edge $\widehat L_{-}^\vartheta$.
\end{corollary}
\begin{proof}
Corollary~\ref{remark about real and imaginary part of hatmfc} is proven in the same way as Corollary~\ref{remark about real and imaginary part of mfc}. The only things to be checked are that $r_\pm'>0$ can be chosen uniformly in $N$, $N$ sufficiently large, and the bounds in~\eqref{convergence of B and C function}. The former statement is an immediate consequence of the stability bound~\eqref{stability bound B}. The latter follows from $z=F(\zeta)=\widehat F(\widehat\zeta)$, with $\zeta=z+m_{fc}^{\vartheta}(z)$ and $\widehat\zeta=z+\widehat m_{fc}^{\vartheta}(z)$. Then using~\eqref{definition of omega_V}, the stability bound~\eqref{stability bound B} and the uniform lower bound on $F''(\zeta_\pm^{\vartheta})$, it is straightforward to derive the estimate~\eqref{convergence of B and C function} from~\eqref{mini local law}.
\end{proof}
Next we prove Lemma~\ref{superlemma}. Recall from~\eqref{equation for vartheta} that we chose $\vartheta\equiv \vartheta(t)=\e{-(t-t_0)/2}$.
\begin{proof}[Proof of Lemma~\ref{superlemma}]
For $c>0$ and a measure $\omega$ on $\ensuremath{\mathbb{R}}$, we set $\supp_c\omega\mathrel{\mathop:}=\supp \omega+[-c,c]$. Recall the constants $r'_\pm>0$ of Corollary~\ref{remark about real and imaginary part of hatmfc}. Set $s\mathrel{\mathop:}= \min \{r'_-,\,r'_+\}/2$.
We specify the potentials $\widehat U$ and $U$ through their spatial derivatives $\widehat U'$ and $U'$. For $t\ge0$, we set
\begin{align}\label{definition widehat U appendix}
\widehat U'(t,x)+x\mathrel{\mathop:}= -2\Xint-_\ensuremath{\mathbb{R}}\frac{\widehat\rho_{fc}(t,y)}{y-x}\,\mathrm{d} y\,,\qquad \quad U'(t,x)+x\mathrel{\mathop:}= -2\Xint-_\ensuremath{\mathbb{R}}\frac{\rho_{fc}(t,y)}{y-x}\,\mathrm{d} y\,,
\end{align}
for $x\in\supp \widehat\rho_{fc}(t)$, respectively $x\in\supp\rho_{fc}(t)$.
For $x\in\ensuremath{\mathbb{R}}$ satisfying $|x- L_\pm(t)|\le s$, where $L_\pm(t)$ denote the endpoints of the support of the measure $\rho_{fc}(t)$, we set
\begin{align}\label{definition widehat U II appendix}
\widehat U'(t,x)+x&\mathrel{\mathop:}=-2\widehat {\mathcal C}^{\vartheta}_\pm(\widehat k_{\pm})\,,\qquad\quad\widehat k_{\pm}\equiv x-\widehat L_\pm(t)\,,\nonumber\\
U'(t,x)+x&\mathrel{\mathop:}=-2 {\mathcal C}^{\vartheta}_\pm(k_{\pm})\,,\qquad\quad k_{\pm}\equiv x-L_\pm(t)\,,
\end{align}
where $\widehat{\mathcal C}^{\vartheta}_\pm$ are the functions appearing in Corollary~\ref{remark about real and imaginary part of hatmfc} with $\vartheta\equiv\vartheta(t)$ and ${\mathcal C}^{\vartheta}_\pm$ are the functions appearing in Corollary~\ref{remark about real and imaginary part of mfc} with $\vartheta\equiv \vartheta(t)$. From Lemma~\ref{lemma mfc A}, Corollary~\ref{remark about real and imaginary part of mfc} and Corollary~\ref{remark about real and imaginary part of hatmfc}, we conclude that $\widehat U'(t,x)$, respectively $U'(t,x)$ are well-defined for $x\in\supp_s\rho_{fc}(t)$, $t\ge0$, $s=\min\{ r'_-,\,r'_+\}/2$.
For $x\not\in\supp_{s}\rho_{fc}(t)$, we define $U'$ as a $C^3$ extension in $x$ such that: (1) $ U^{(n)}(t,x)$, $\partial_t U^{(n)}(t,x)$, $n\in\llbracket 1,3\rrbracket$, are continuous in~$t$; (2) for all $t\ge0$ and for all $x\not\in\supp_s\rho_{fc}(t)$ , $|U'(t,x)+x|>|2\re m_{fc}(t,x)|$ and $\widehat U''(t,x)\ge -C_U$, for some constant $C_U\ge 0$; (3) $U'(t,x)+x\sim x$ for all $t\ge 0$, as $|x|\to\infty$. Similarly, we define $\widehat U(t,x)$ as $C^3$ extensions such that: (1) $\widehat U^{(n)}(t,x)$, $\partial_t\widehat U^{(n)}(t,x)$, $n\in\llbracket 1,3\rrbracket$, are continuous in $t$; (2) there is $c>0$ such that $\sup_{t\ge 0}|\widehat U^{(n)}(t,x)-U^{(n)}(t,x)|\le N^{-c\alpha_0/2}$, $n\in\llbracket 1,3\rrbracket$, for $N$ sufficiently large on $\Omega_V$.
We next show that the potential $ U'(t,x)$ is a $C^3$ function in $x$. For simplicity, we often drop the $t$-dependence from the notation. Let $\zeta=z+ m_{fc}(z)$ and recall from the proof of Lemma~\ref{lemma mfc A} that $\zeta(z)$ satisfies $\zeta(z) = F^{(-1)}(z)$, where $F(\zeta)=\zeta-\int\frac{\mathrm{d}\nu(v)}{(\vartheta v-\zeta)}$. Thus, to prove regularity of $ U'(t,x)$ in $x$ in the support of the measure $\rho_{fc}(t)$ it suffices to show that $F'(\zeta) \neq 0$ on the curve $\gamma\cap \ensuremath{\mathbb{C}}^+$ where $\im F = 0$. Recall that on $\gamma$ we have
\begin{align}\label{appendix A20}
\widetilde{H}(\zeta)\mathrel{\mathop:}=\int\frac{\mathrm{d}\nu(v)}{|\vartheta v- \zeta|^2} = 1\,,
\end{align}
where $\vartheta\equiv \vartheta(t)$. On the other hand, we have
\begin{align*}
\re {F}'(\zeta)=1-\int\frac{(\vartheta v-\re\zeta)^2-(\im \zeta)^2}{|\vartheta v-\zeta|^4}\,\mathrm{d}\nu(v)\,.
\end{align*}
Thus, on the curve $\gamma$,
\begin{align*}
\re {F}'(\zeta)=&\int \frac{\mathrm{d}\nu(v)}{|\vartheta v- \zeta|^2}-\int\frac{(\vartheta v-\re\zeta)^2-(\im \zeta)^2}{|\vartheta v-\zeta|^4}\,\mathrm{d}\nu(v)=\int\frac{2(\im\zeta)^2}{|\vartheta v-\zeta|^4}\,\mathrm{d}\nu(v)\,.
\end{align*}
From~\eqref{sumrule} we get
\begin{align}\label{appendix A24}
\int\frac{\mathrm{d}\nu(v)}{|\vartheta v-\zeta|^4}\ge \left(\int\frac{\mathrm{d}\nu(v)}{|\vartheta v-\zeta|^2}\right)^2=1\,,
\end{align}
on $\gamma$. Since $F'\not=0$ on $\gamma$, the inverse function theorem implies that $\re {m}_{fc}(t,x)$ is a smooth function in the interior of $\supp\rho_{fc}(t)$, whose derivatives are continuous in $t$. For $x\in\mathrm{B}_s(L_{\pm}^\vartheta)$, we already showed in Lemma~\ref{remark about real and imaginary part of mfc} that ${\mathcal C}_{\pm}^{\vartheta}(x)$ is a smooth function, whose derivatives are continuous in $t$. Thus, we have shown that $U'(t,x)$ is smooth in $\supp_s\rho_{fc}(t)$. Outside $\supp_s\rho_{fc}(t)$, $U'(t,x)$ is manifestly $C^3$ by definition: it is a $C^3$ extension of the functions ${\mathcal C}_\pm(t)$. Thus $\ensuremath{\mathbb{R}}\ni x\mapsto U'(t,x),\,\partial_t U'(t,x)$ are $C^3$ functions for all $t\ge0$.
Clearly, we can bound the derivatives $U^{(n)}(t,x)$, $\partial_t U^{(n)}(t,x)$, $n\in\llbracket 1,3\rrbracket$, uniformly on compact sets. It is also immediate that $U^{(n)}(t,x)$ are continuous function in $t\ge 0$. Thus we can bound $U^{(n)}$ uniformly in $t$ and uniformly in $x$ on compact sets, for $n\in\llbracket 1,3\rrbracket$. For $x\in\supp_s\rho_{fc}(t)$, we have $U''(t,x)\ge -C$, for some $C\ge 0$. For $x\not\in\supp_s\rho_{fc}(t)$, a similar bound holds true by construction. Thus $U'(t,x)$ satisfies~\eqref{assumption 1 for beta universality} uniformly in $t\ge0$. Further, since $U'(t,x)+x\sim x$, as $|x|\to\infty$, ~\eqref{assumption 2 for beta universality} also holds uniformly in $t\ge0$.
On $\Omega_V$, we can extend the reasoning above to $\widehat U'(t,x)$, $\partial_t U'(t,x)$, for $N$ sufficiently large. For example, the arguments in~\eqref{appendix A20}--\eqref{appendix A24} can be extended to the finite $N$ case by using~\eqref{definition of omega_V} and Lemma~\ref{lemma hatmfc}. Let again $s\equiv\min\{ r'_-,\,r'_+\}/2$. Then for~$x\in\supp_s\rho_{fc}(t)$ we have by Lemma~\ref{lemma hatmfc} that $|\widehat m_{fc}(t,x+\mathrm{i}\eta)-m_{fc}(t,x+\mathrm{i}\eta)|\le N^{-c\alpha_0}$, for some $c>0$, on~$\Omega_V$ for all $\eta\ge 0$ and all $t\ge 0$. Together with~\eqref{convergence of B and C function} we can conclude that $|\widehat U'(t,x)-U'(t,x)|\le N^{-c\alpha_0/2}$ on~$\Omega_V$, for $x\in\supp_s\rho_{fc}(t)$. We also have $|\partial_x\widehat m_{fc}(t,x+\mathrm{i}\eta)-\partial_x m_{fc}(t,x+\mathrm{i}\eta)|\le CN^{-c\alpha_0}$, for $x$ satisfying $\min\{|x-L_+|,|x-L_-|\}\ge s$, as can be checked as in the proof of Lemma~\ref{lemma hatmfc}. Hence, combining this last statement with the regularity of~$\widehat
C_\pm^{\vartheta}$ claimed in
Lemma~\ref{remark about real and imaginary part of hatmfc}, we have $|\widehat U''(t,x)-U''(t,x)|\le N^{-c\alpha_0}$, for~$x\in \supp_s\rho_{fc}(t)$, $t\ge0$, on~$\Omega_V$ for~$N$ sufficiently large. This conclusion can be extended to arbitrary~$\widehat U^{(n)}$. Similarly, one checks that~$U^{(n)}(t,x)$, $n\in\llbracket1,3\rrbracket$, are continuous functions of~$t\ge 0$. For $x\not\in\supp_s\rho_{fc}(t)$, these properties follow directly from the definition of $\widehat U'$ above. Thus~$\widehat U'(t,x)$ satisfies~\eqref{assumption 1 for beta universality} and~\eqref{assumption 2 for beta universality} with uniform constants for all~$t\ge 0$ and~$N$ sufficiently large on~$\Omega_V$.
Finally, the potentials $\widehat U(t)$ and $U(t)$ are ``regular'' as follows from Lemmas~\ref{lemma mfc} and~\ref{lemma hatmfc}.
\end{proof}
\end{appendix}
|
\section{Introduction}
The sectional category of a fibration is the least number of open sets needed to cover the base, on each of which the fibration admits a continuous local section. This concept, originally studied by A.\ S.\ Schwarz \cite{Schwarz} under the name \emph{genus}, has found applications in diverse areas. Notable special cases include the Lusternik--Schnirelmann category (for which the standard reference has become the monograph \cite{CLOT} by Cornea, Lupton, Oprea and Tanr\'e) and Farber's topological complexity \cite{Far}, both of which are homotopy invariants of spaces which arise as the sectional category of associated path fibrations. The LS-category is classical and related to critical point theory, while topological complexity was conceived in the early part of the twenty-first century as part of a topological approach to the motion planning problem in Robotics. Further details and definitions will be given in Section \ref{secat} below. We remark that in the modern literature it is common to normalise these invariants so that the sectional category of a fibration with section is zero, a convention which we will adopt in this paper.
Most of the existing estimates for sectional category are cohomological in nature and are based on obstruction theory. The objective of the current paper is to produce more refined estimates using methods from unstable homotopy theory. There is an extensive literature on the application of generalized Hopf invariants to LS-category, originating with Berstein and Hilton \cite{B-H} and including spectacular applications by Iwase \cite{Iwase}, Stanley \cite{Stanley}, Strom \cite{Strom} and others (a nice summary can be found in Chapter 6 of \cite{CLOT}). Building on and generalizing the work of these authors, we develop a theory of generalized Hopf invariants in the setting of sectional category. We then apply our theory to give new computations of topological complexity which we believe would not be possible using obstruction-theoretic arguments.
Our first application is to the computation of the topological complexity of two-cell complexes $X=S^p\cup_\alpha e^{q+1}$. The LS-category of such a space $X$ is determined by the Berstein--Hilton--Hopf invariant
\[
H(\alpha)\in \pi_q(\Sigma\Omega S^p \wedge \Omega S^p) \cong \pi_q (S^{2p-1}\vee S^{3p-2}\vee\cdots )
\]
of the attaching map $\alpha: S^q\to S^p$ \cite{B-H}. When $p\ge2$, we have
\[
{\rm{cat}\hskip1pt}(X) = \left\{\begin{array}{ll} 1 & \mbox{if }H(\alpha)=0, \\ 2 & \mbox{if }H(\alpha)\neq 0.
\end{array}\right.
\]
Note that in the metastable range $2p-1\le q\le 3p-3$ we may identify $H(\alpha)$ with its projection onto the bottom cell $H_0(\alpha) \in \pi_q(S^{2p-1})$. If $H_0(\alpha)\neq 0$ then ${\rm{cat}\hskip1pt}(X) =2$, which by standard inequalities implies that $2\le {\rm{TC}\hskip1pt}(X)\le 4$. In almost all cases, the usual cohomological bounds fail to determine the exact value of ${\rm{TC}\hskip1pt}(X)$, for reasons of dimension. Using Hopf invariants, however, we are able to identify many cases with ${\rm{TC}\hskip1pt}(X)\le 3$ (in Theorem~\ref{le3bis} below, where we use the symbol $\circledast$ to denote the join functor) as well as many cases with ${\rm{TC}\hskip1pt}(X)\ge 3$ (in Theorem~\ref{ge3bis} below).
\begin{thm}[Theorem \ref{le3}]\label{le3bis}
Let $X=S^p\cup_\alpha e^{q+1}$, where $\alpha: S^q\to S^p$ is in the metastable range $2p-1< q\le 3p-3$ and $H_0(\alpha)\neq 0$.
Then ${\rm{TC}\hskip1pt}(X)\le 3$ in any of the following cases:
\begin{enumerate}
\item $2 H_0(\alpha)\circledast H_0(\alpha)=0$ (e.g.~when $2H_0(\alpha)=0$);
\item $6 H_0(\alpha)\circledast H_0(\alpha)=0$ (e.g.~when $\,6 H_0(\alpha)= 0$) and $p$ is even;
\item $q$ is even.
\end{enumerate}
\end{thm}
\begin{thm}[Theorem \ref{ge3}]\label{ge3bis}
Let $X=S^p\cup_\alpha e^{q+1}$, where $\alpha: S^q\to S^p$ is in the metastable range $2p-1< q\le 3p-3$ and $H_0(\alpha)\neq 0$.
Then ${\rm{TC}\hskip1pt}(X)\ge 3$ in any of the following cases:
\begin{enumerate}
\item $p$ is odd;
\item $p$ is even and $3H_0(\alpha)\neq 0$.
\end{enumerate}
\end{thm}
Combining these two theorems gives the precise value ${\rm{TC}\hskip1pt}(X)=3$ for large classes of two-cell complexes (see for instance Corollaries~\ref{partiuno} and~\ref{partidos}).
We are also able to draw conclusions about ${\rm{TC}\hskip1pt}(X)$ outside of the metastable range, under the additional assumption $H(\alpha) = H_0(\alpha)$.
\begin{rem}
We also get a full description of ${\rm{TC}\hskip1pt}(X)$ for $X=S^p\cup_\alpha e^{2p}$ (Theorems~\ref{grados} and~\ref{ClassicHopf} below). The proofs are, however, much more elementary than those in the cases of Theorems~\ref{le3bis} and~\ref{ge3bis}.
\end{rem}
Our second application is to the analogue of Ganea's conjecture for topological complexity. Recall that the product inequality ${\rm{cat}\hskip1pt}(X\times Y)\le {\rm{cat}\hskip1pt}(X) + {\rm{cat}\hskip1pt}(Y)$ is satisfied by LS-category. Examples of strict inequality were given by Fox \cite{Fox}, involving Moore spaces with torsion at different primes. Ganea asked, in his famous list of problems \cite{Ganea1}, if we always get equality when one of the spaces involved is a sphere. That is, if $X$ is a finite complex, is it true that
\[
{\rm{cat}\hskip1pt}(X\times S^k) = {\rm{cat}\hskip1pt}(X) + 1 \qquad \mbox{for all $k\ge1$?}
\]
A positive answer became known as \emph{Ganea's conjecture}. The conjecture remained open for nearly $30$ years, shaping research in the subject. It was shown to hold for simply-connected rational spaces by work of Jessup~\cite{Jessup} and Hess~\cite{Hess}, and for large classes of manifolds by Singhof \cite{Singhof} and Rudyak \cite{Rudyak}, until eventually proven to be false in general by Iwase \cite{Iwase,IwaseAinfty}. Iwase's counter-examples are two-cell complexes $X$ outside of the metastable range, whose Berstein--Hilton--Hopf invariants are essential but stably inessential, from which it follows that ${\rm{cat}\hskip1pt}(X)={\rm{cat}\hskip1pt}(X\times S^k) =2$.
The analogous question for topological complexity (which also satisfies the product inequality)
asks whether, for any finite complex $X$ and $k\ge 1$, we always have an equality
\begin{equation}\label{TCGanea}
{\rm{TC}\hskip1pt}(X\times S^k) = {\rm{TC}\hskip1pt}(X) + {\rm{TC}\hskip1pt}(S^k) = \left\{\begin{array}{ll} {\rm{TC}\hskip1pt}(X) + 1 & \mbox{if $k$ odd,}\\ {\rm{TC}\hskip1pt}(X) + 2 &\mbox{if $k$ even.} \end{array}\right.
\end{equation}
This question was raised by Jessup, Murillo and Parent \cite{JMP}, who proved that equation (\ref{TCGanea}) holds when $k\ge2$ for any formal, simply-connected rational complex $X$ of finite type. In this paper, we give a counter-example to (\ref{TCGanea}) for all even $k$, using Hopf invariant techniques.
\begin{thm}[Theorem \ref{noGaneaTC}]\label{noGaneaTCbis}
Let $Y$ be the stunted real projective space ${\mathbb{R}} P^6/{\mathbb{R}} P^2$, and let $X=Y\vee Y$. Then for all $k\ge2$ even,
\[
{\rm{TC}\hskip1pt}(X)=4\quad\mbox{and}\quad {\rm{TC}\hskip1pt}(X\times S^{k}) = 5.
\]
\end{thm}
We now briefly outline the contents of each section, and in doing so indicate the method of proof of the above results. Section~\ref{seccionpreliminar} is preliminary, and establishes our conventions and notations regarding base points, cones, suspensions and joins. In Section~\ref{secat} we give the necessary background on sectional category and relative sectional category, working in the context of fibred joins. The main new result here is Proposition \ref{secatupbyone}, which shows that the sectional category of a fibration relative to a subspace increases by at most one on attaching a cone, and moreover the section over the cone can be controlled in a certain sense. Section~\ref{hifscsec} is split into two subsections. In Subsection~\ref{HopfDefns} we define the Hopf invariants for sectional category, which determine the behaviour of relative sectional category under cone attachments. We also recall the definition of the Berstein--Hilton--Hopf invariants mentioned above and show how they fit into our framework. In Subsection~\ref{HopfProducts} we investigate Hopf invariants for cartesian products of fibrations. Using naturality of the exterior join construction, we prove our Theorem \ref{Hopfproducts} which states that Hopf invariants of a product can be obtained as joins of Hopf invariants of the factors, composed with a topological shuffle map (see below). This result is germane to the proofs of Theorems~\ref{le3bis} and~\ref{ge3bis} and Theorem~\ref{noGaneaTCbis}, which we give in Sections~\ref{secaptc2cw} and~\ref{secapgc4tc} respectively. Finally, Section~\ref{secsiete} is the technical heart of the paper. Using the description of the join in terms of simplicial (barycentric) parameters, together with the standard decomposition of the product of simplices $\Delta^n\times \Delta^m$ into simplices $\Delta^{n+m}$, we construct \emph{topological shuffle maps}
\[
\Phi^{A,B}_{n,m}: J^n(A)\circledast J^m(B)\to J^{n+m+1}(A\times B),
\]
which map from the join of the $(n+1)$-fold join of a space $A$ with the $(m+1)$-fold join of a space $B$ to the $(n+m+2)$-fold join of their product $A\times B$. We then describe (in Proposition~\ref{shuffleinhomology}) the effect of this map in homology, in terms of algebraic shuffles. This result is used in all of our calculations of Hopf invariants of products.
The idea of applying Hopf invariant techniques to obtain estimates for sectional category has been around for some time. For example, the upper bounds for the (higher) topological complexity of configuration spaces given in \cite{FG2} and \cite{GonzalezGrant}, and proved using obstruction theory, were originally obtained using Hopf invariants. In developing the ideas in this paper, we have benefitted from discussions with many people. In this regard, the first author would like to thank Michael Farber, Hugo Rodr\'iguez-Ord\'o\~nez, and Enrique Torres-Giese, the second author would like to thank Michael Farber, and the third author would like to thank Pierre Ghienne. The three authors are grateful to the organizers of the ESF ACAT conference and workshop ``Applied Algebraic Topology'' held in Castro Urdiales, Spain, from June~26 to July~5, 2014, where the final form of this project was shaped.
We conclude this introductory section with a list of open problems, some of which may be accessible by extending the techniques in this paper.
\begin{probs}
\begin{enumerate}[(a)]
\item Do there exist two-cell complexes $X=S^p\cup_\alpha e^{q+1}$ with $q>2p-1$ for which ${\rm{cat}\hskip1pt}(X)={\rm{TC}\hskip1pt}(X)=2$? (The smallest such open case is that of $\alpha\in\pi_4(S^2)=\mathbb{Z}/2$ the generator, see Example~\ref{contieneposiblegap}(a).) Or for which ${\rm{TC}\hskip1pt}(X)=4$?
\item Does every two-cell complex satisfy equation (\ref{TCGanea})?
\item Do there exist finite complexes $X$ and \emph{odd} $k$ for which ${\rm{TC}\hskip1pt}(X\times S^k) = {\rm{TC}\hskip1pt}(X)$?
\end{enumerate}
\end{probs}
\section{Preliminaries}\label{seccionpreliminar}
We work in the category of well-pointed\footnote{Base points will generically be denoted by an asterisk $\ast$.} compactly generated spaces having the homotopy type of CW-complexes. Thus, all maps, diagrams, and homotopies will be pointed, unless explicitly noticed otherwise. For instance, a homotopy section of a map $p\colon \mathcal{A}\to X$ is a pointed map $s\colon X \to\mathcal{A}$ with a pointed homotopy between $p\circ s$ and the identity on $X$. Products and mapping spaces are topologized in such a way that the product of two proclusions is a proclusion and evaluation maps are continuous. Fibrations are assumed to be pointed fibrations in the sense that they lift pointed homotopies or, equivalently, that they admit a pointed lifting function. Likewise, cofibrations are assumed to be pointed cofibrations\footnote{These are not actual restrictions in view of~\cite[Theorem~5.95]{MR2839990}.}.
We let $I$ stand for the unit interval $[0,1]$ with base point 0. For a pointed space $(A,\ast)$, we denote by $CA$ the cone $A\times I/A\times 1$ {with base point so that the projection $A\times I\to CA$ and the inclusion} $A\hookrightarrow CA$, $a\mapsto [a,0]$ {are pointed} (in general, the class of $(a,u)$ will be denoted by $[a,u]$). Note that the inclusion $A\hookrightarrow CA$ is a cofibration. The suspension of $A$ is defined by $\Sigma A:=CA/A$. We will also use the reduced suspension of $A$ given by the quotient $\widetilde{\Sigma} A:=\Sigma A/\ast\times I=A\wedge (I/\partial I)$. In both cases we take the obvious base points. The class of $[a,u]$ in both $\Sigma A$ and $\widetilde{\Sigma} A$ will also be denoted by $[a,u]$.
We denote by $\Delta^n$ the standard $n$-simplex of ${\mathbb{R}}^{n+1}$, given by
$$\Delta^n=\{(t_0,\ldots,t_{n})\in [0,1]^{n+1}\mid t_0+\cdots + t_{n}=1\}$$
with base point $(1,0,\ldots,0)$. The iterated $n$-fold reduced suspension of $A$ is homotopically equivalent to the quotient
$$ A\times \Delta^n/(A\times \partial \Delta^n \cup \ast \times \Delta^n)= A \wedge (\Delta^n/\partial \Delta^n)$$
that we will denote by $\widetilde{\Sigma}^nA$.
If $(B,\ast)$ is another pointed space we denote by $A\circledast B$ the join $CA\times B\;\cup\, A\times CB$ with base point $(\ast,\ast)\in A\times B\subset CA\times B\;\cap\, A \times CB$.
\begin{rems} \label{rmkpinchmap} We collect here some well-known facts which will be used in this work:
\begin{itemize}
\item[(a)] There is a canonical map given by the composition
$$\begin{array}{rcccl}
\zeta:A\circledast B &\to& \Sigma (A\times B) &\to & \widetilde{\Sigma} (A\times B)\\
\end{array}
$$
in which the second arrow is the identification map and the first arrow is the (non-pointed) map induced by the (non-pointed) maps
$$\begin{array}{rclcrcl}
A \times CB & \to& \Sigma (A\times B) & CA \times B & \to & \Sigma (A\times B)\\
(a,[b,u])&\mapsto &[(a,b),\frac{1-u}{2}] & ([a,u],b)&\mapsto &[(a,b),\frac{1+u}{2}]
\end{array}$$
Although the first map in the composition defining $\zeta$ is non-pointed, the composite $\zeta$ is pointed and, for this reason, we will mainly consider reduced suspensions in the sequel.
\item[(b)] The composition
$
A\circledast B \stackrel{\zeta}{\longrightarrow} \widetilde{\Sigma} (A\times B) \longrightarrow \widetilde{\Sigma} (A\wedge B),
$$
where the second arrow is the identification map $[(a,b),u] \mapsto [a\wedge b,u]$, is a homotopy equivalence.
\item[(c)] The canonical (pointed) identification maps
$$\begin{array}{rclcrcl}
A \times CB & \to& \widetilde{\Sigma} (A\times B) & CA \times B & \to &\widetilde{\Sigma} (A\times B)\\
(a,[b,u])&\mapsto &[(a,b),u] & ([a,u],b)&\mapsto &[(a,b),u]
\end{array}$$
induce the following map
$$\begin{array}{rcl}
\nu:A\circledast B &\to& \widetilde{\Sigma} (A\times B) \vee \widetilde{\Sigma} (A\times B)\\
(a,[b,u]) &\mapsto & ([(a,b),u],*)\\
([a,u],b) &\mapsto & (*,[(a,b),u])\\
\end{array}
$$
which we call the {\em difference pinch map}. Indeed $\nu$ fits in the following commutative diagram
\begin{equation}\label{difzetadif}\xymatrix{
A\circledast B \ar[d]^{\zeta} \ar[rrd]^-{\nu}\\
\widetilde{\Sigma} (A\times B)\ar[d] \ar[rr]^-{\tilde{\nu}} &&\widetilde{\Sigma} (A\times B) \vee \widetilde{\Sigma} (A\times B)\ar[d]\\
\widetilde{\Sigma} (A\wedge B) \ar[rr]^-{\tilde{\nu}} &&\widetilde{\Sigma} (A\wedge B) \vee \widetilde{\Sigma} (A\wedge B)\\
}\end{equation}
where $\tilde \nu$ is the standard difference pinch map, that is the map
$$\begin{array}{rcl}
\tilde{\nu}:\widetilde{\Sigma} Z&\to& \widetilde{\Sigma} Z \vee \widetilde{\Sigma} Z\\
\left[z,u\right] &\mapsto & \left\{\begin{array}{lr}
([z,1-2u],*) & 0\leq u\leq 1/2\\
(*,[z,2u-1]) & 1/2\leq u\leq 1
\end{array}\right.
\end{array}
$$
In particular, if $f,g: \widetilde{\Sigma}Z \to X$ are two maps then the composite $\nabla \circ (f\vee g)\circ\tilde{\nu}$ is the difference $g-f$.
\end{itemize}
\end{rems}
\section{Sectional category and relative sectional category}\label{secat}
In this section we will review some known facts about sectional category, most of which can be found in the references \cite{CLOT}, \cite{James}, and \cite{Schwarz}. We will pay particular attention to the two most well-studied examples, namely Lusternik--Schnirelmann category and Farber's topological complexity.
In Definitions~\ref{definicionLScattegoria}--\ref{definicionsecat} below we do not require that subspaces are pointed. Consequently, the nullhomotopies in Definition~\ref{definicionLScattegoria}, and the partial sections in Definitions~\ref{definicionTC} and~\ref{definicionsecat}, are not assumed to be pointed.
\begin{defn}\label{definicionLScattegoria} The \emph{(Lusternik--Schnirelmann) category} of a space $X$, denoted ${\rm{cat}\hskip1pt}(X)$, is the least non-negative integer $n$ such that $X$ admits a cover by $n+1$ open sets $U_0, \ldots , U_n$ such that each inclusion $U_i\hookrightarrow X$ is null-homotopic. If no such integer exists, we set ${\rm{cat}\hskip1pt}(X)=\infty$.
\end{defn}
\begin{defn} \label{definicionTC}The \emph{topological complexity} of a space $X$, denoted ${\rm{TC}\hskip1pt}(X)$, is the least non-negative integer $n$ such that $X\times X$ admits a cover by $n+1$ open sets $U_0, \ldots , U_n$, on each of which there exists a continuous partial section of the evaluation fibration
\[
\pi_X:X^I\rightarrow X\times X, \qquad \alpha \mapsto \big(\alpha (0),\alpha (1)\big).
\]
Here $X^I$ denotes the space of paths in $X$ with base point the constant path in $\ast\in X$. If no such integer exists, we set ${\rm{TC}\hskip1pt}(X)=\infty$.
\end{defn}
Both of these concepts are special cases of the sectional category of a fibration, first studied by A.\ S.\ Schwarz under the name \emph{genus}.
\begin{defn}\label{definicionsecat}
Let $p: \mathcal{A}\to X$ be a (surjective) fibration. The \emph{sectional category} of $p$, denoted ${\rm{secat}\hskip1pt}(p)$, is defined to be the least non-negative integer $n$ such that $X$ admits a cover by $n+1$ open sets $U_0,\ldots , U_n$ on each of which there exists a continuous partial section of $p$. If no such integer exists, we set ${\rm{secat}\hskip1pt}(p)=\infty$.
\end{defn}
One of the key results in this area is a characterization of sectional category in terms of fibred joins, which we now recall within our base-point setting. For $n\geq 0$ we denote by $p^{n+1}_X:\mathcal{A}^{n+1}_X\to X$ the fibred product of $n+1$ copies of $p$. The total space of the fibred join of $n+1$ copies of $p$ is the quotient space
$$J^n_X(\mathcal{A}) =\mathcal{A}_X^{n+1} \times \Delta^ n/\sim$$
where $\sim$ is the equivalence relation generated by
\[
(a_0,\ldots,a_i,\ldots, a_{n},t_0,\ldots,t_i,\ldots,
t_{n})\sim(a_0,\ldots,a'_i,\ldots,a_{n},t_0,\ldots,t_i,\ldots,t_{n})
\]
if $t_i=0$. We denote a general element of $\mathcal{A}_X^{n+1}\times \Delta^n$ by $(\mathbf{a},\mathbf{t})$ and its class in $J^n_X(\mathcal{A})$ by $\langle\mathbf{a}\mid \mathbf{t}\rangle$. In these terms, we choose
\begin{equation}\label{puntobasedeljoin}
\ast=\left\langle(\ast,\ldots,\ast) \;\left\rvert \,\left(1,0,\ldots,0\right)\right.\right\rangle
\end{equation}
as the base point in $J^n_X(\mathcal{A})$---naturally induced by the base points of $\mathcal{A}$ and $\Delta^n$. The map $p_n:J^n_X(\mathcal{A})\rightarrow X$ given by $\langle\mathbf{a}\mid \mathbf{t}\rangle\mapsto p_X^{n+1}(\mathbf{a})$ is a (pointed) fibration, called the {\em $(n+1)$-fold fibred join of $p$}. If $A=p^ {-1}(\ast)$ is the fibre of $p$ over $\ast\in X$, then $p_n^{-1}(\ast)$ is the quotient $J^ n(A)=A^{n+1} \times \Delta^ n/\sim$ where the relation is the same as above. The latter space is homotopically equivalent to the $n$-fold suspension of the $(n+1)$-fold smash product of $A$ with itself. More precisely, with the notation introduced before, the following composite of identification maps
\begin{equation}\label{lasidentificacionesobvias}
J^n(A) \stackrel{r}{\longrightarrow}\widetilde{\Sigma}^nA^{n+1} \longrightarrow \widetilde{\Sigma}^nA^{\wedge n+1}
\end{equation}
is a homotopy equivalence.
\begin{thm}\label{puntodepartida}
Let $p:\mathcal{A}\to X$ be a (surjective) fibration with $X$ paracompact. If $n\ge1$ or $\mathcal{A}$ is path-connected, then ${\rm{secat}\hskip1pt}(p)\leq n$ if and only
if $p_n:J^n_X(\mathcal{A})\rightarrow X$ admits a (pointed) homotopy section.
\end{thm}
Most of the standard formulations of Theorem~\ref{puntodepartida} in the literature (e.g.~\cite[Theorem 3]{Schwarz}) are base-point free. In our context, the hypothesis ``$n\ge1$ or $\mathcal{A}$ path-connected'' in Theorem~\ref{puntodepartida} assures that the fiber of $p_n$ is path-connected. The pointed homotopy section (and even a pointed section) is then warranted since spaces are well pointed and $p_n$ is a pointed fibration.
\begin{rem}\label{fibredjoins}
For any $n\ge 0$ there is a commutative diagram
\begin{equation}\label{ijkappa}\xymatrix{
J^n(A) \ar[r]^{i_n}\ar@{^{(}->}[d] & J^n_X(\mathcal{A}) \ar@{^{(}->}[d]^{\jmath_n} \\
CJ^n(A) \ar[r]^{\kappa_n} &J^{n+1}_X(\mathcal{A}).
}\end{equation}
The inclusions $\jmath_n: J^n_X(\mathcal{A})\to J^{n+1}_X(\mathcal{A})$ are given by
$$\langle\mathbf{a}\mid \mathbf{t}\rangle\mapsto \langle\mathbf{a},a_{n+1}\mid \mathbf{t},0\rangle$$
where $\mathbf{a}\in \mathcal{A}_X^{n+1}$, $\mathbf{t}\in \Delta^n$ and $a_{n+1}$ is any element of $\mathcal{A}$
with $(\mathbf{a},a_{n+1})\in\mathcal{A}_X^{n+1}$.
They are compatible with the maps $p_n$ and $p_{n+1}$. The maps $\kappa_n: CJ^n(A) \to J^{n+1}_X(\mathcal{A})$ are given by $[\langle \mathbf{a}\mid \mathbf{t}\rangle,u]\mapsto \langle \mathbf{a},*\mid (1-u)\mathbf{t},u\rangle$. Notice that this map factors through the inclusion $J^{n+1}(A)\to J^{n+1}_X(\mathcal{A})$.
\end{rem}
Coming back to topological complexity and category, we suppose that $X$ is path-connected and paracompact with base point $\ast \in X$. Then we have
$${\rm{TC}\hskip1pt}(X)={\rm{secat}\hskip1pt}(\pi_X:X^I\rightarrow X\times X) \mbox{ \ \ and \ \ } {\rm{cat}\hskip1pt}(X)={\rm{secat}\hskip1pt}(p_X:PX\to X)$$ where $PX\subset
X^I$ is the space of paths beginning at the base point $*\in X$ and $p_X(\gamma)=\gamma(1)$.
The fibred join of $n+1$ copies of $p_X$ will be denoted by $g_n(X):G_n(X)\to X$ and referred as the {\em $n$-th Ganea fibration of $X$}. Thus ${\rm{cat}\hskip1pt}(X)\leq n$ if an only if $g_n(X):G_n(X)\to X$ admits a (pointed) homotopy section. It is well-known that $G_1(X)\simeq \widetilde{\Sigma}\Omega X$ {with $g_1(X)$ homotopic to the adjoint of the identity map on $\Omega X$,} and that when $X={\widetilde{\Sigma}} A$ is a suspension {and $n\ge1$}, $g_n({\widetilde{\Sigma}} A)$ admits a canonical section given by the composition
\[
s_0: {\widetilde{\Sigma}} A \to {\widetilde{\Sigma}}\Omega{\widetilde{\Sigma}} A \simeq G_1({\widetilde{\Sigma}} A)\hookrightarrow G_n({\widetilde{\Sigma}} A)
\]
{where $\widetilde{\Sigma}A\to\widetilde{\Sigma}\Omega\widetilde{\Sigma}A$ is the suspension of the adjoint of the identity map on $\widetilde{\Sigma} A$.}
Likewise, the fibred join of $n+1$ copies of $\pi_X: X^I\to X\times X$ will be denoted by $g^{{\rm{TC}\hskip1pt}}_n(X):G^{{\rm{TC}\hskip1pt}}_n(X) \to X\times X$ and referred as the \emph{$n$-th ${\rm{TC}\hskip1pt}$-Ganea fibration of $X$.} Thus ${\rm{TC}\hskip1pt}(X)\leq n$ if and only if $g^{{\rm{TC}\hskip1pt}}_n(X): G^ {{\rm{TC}\hskip1pt}}_n(X) \to X\times X$ admits
a (pointed) homotopy section.
Both $g_n(X)$ and $g_n^{{\rm{TC}\hskip1pt}}(X)$ have as fibre (over the {corresponding} base points $\ast$ {and $(\ast,\ast)$}) the join $J^n(\Omega X)$ of $n+1$ copies of the based loop space $\Omega X$. When considering the Ganea fibrations, we denote this space by $F_n(X)$. The inclusions will be denoted {by}
$$i_n(X):F_n(X)\to G_n(X) \quad\mbox{and}\quad i^{{\rm{TC}\hskip1pt}}_n(X):F_n(X)\to G^{{\rm{TC}\hskip1pt}}_n(X).$$
As mentioned before the inclusions
$$G_n(X)\hookrightarrow G_{n+1}(X) \quad\mbox{and}\quad G^{{\rm{TC}\hskip1pt}}_n(X)\hookrightarrow G^{{\rm{TC}\hskip1pt}}_{n+1}(X)$$
correspond to the inclusion on the first $n+1$ factors. The constructions $G_n$, $G_n^{\rm{TC}\hskip1pt}$ and $F_n$ are homotopy functors.
\begin{prop}\label{TCcatHopf}
The map $\chi:P(X\times X)=PX\times PX \to X^I$ given by $(\alpha, \beta)\mapsto \alpha^ {-1}\beta$ induces a map $\bar{\chi}: \Omega(X\times X)=\Omega X\times \Omega X \to \Omega X$ and commutative diagrams for any $n$:
\[
\xymatrix{
F_n (X\times X) \ar[d]_{i_n(X\times X)} \ar[rr]^{\bar\chi_n} & &F_n( X) \ar[d]^{i^{{\rm{TC}\hskip1pt}}_n(X)}\\
G_n(X\times X)\ar[rd]_{g_n(X\times X)} \ar[rr]^{\chi_n} &&G^{{\rm{TC}\hskip1pt}}_n(X) \ar[ld]^{g^{{\rm{TC}\hskip1pt}}_n(X)}\\
&X\times X.
}
\]
\end{prop}
This fact will be important in later sections, as it will allow us to construct Hopf invariants for ${\rm{TC}\hskip1pt}(X)$ from Hopf invariants for ${\rm{cat}\hskip1pt}(X\times X)$.
Next we define relative sectional category and give some of its properties.
\begin{defn}\label{secatreldef}
Let $p:\mathcal{A}\to X$ be a fibration and let $\varphi: K\to X$ be any map. The {\em sectional category of $p: \mathcal{A}\to X$ relative to $\varphi\colon\thinspace K\to X$,} denoted by ${\rm{secat}\hskip1pt}_\varphi(p)$, is the sectional category of $\varphi^*p$, the pullback of $p$ under $\varphi$. If $K\subseteq X$ and $\varphi$ is the inclusion, we denote ${\rm{secat}\hskip1pt}_\varphi(p)=:{\rm{secat}\hskip1pt}_K(p)$. In particular, ${\rm{secat}\hskip1pt}_X(p)={\rm{secat}\hskip1pt}(p)$.
\end{defn}
\begin{prop}\label{secatprops}
Let $p:\mathcal{A}\to X$ be a (surjective) fibration with fibre $A$, and let $\varphi: K\to X$ be any map. The relative sectional category satisfies the following properties:
\begin{enumerate}
\item ${\rm{secat}\hskip1pt}_\varphi(p)$ depends only on the homotopy class of $\varphi$.
\item\label{itemdosdel38} ${\rm{secat}\hskip1pt}_\varphi(p)\le {\rm{secat}\hskip1pt}(p)$.
\item ${\rm{secat}\hskip1pt}_\varphi(p)\le {\rm{cat}\hskip1pt}(K)$.
\item If $\pi_i(A)=0$ for $i<r$ then\footnote{We write ${\rm{hdim}\hskip1pt}(K)$ for the \emph{homotopy dimension of $K$,} i.e.~the smallest dimension of CW complexes having the homotopy type of $K$.} ${\rm{secat}\hskip1pt}_\varphi(p) \le \frac{{\rm{hdim}\hskip1pt}(K)}{r+1}$.
\item Suppose there are cohomology classes $x_1,\ldots , x_k\in H^*(X)$ with any coefficients such that $p^*(x_1) = \cdots = p^*(x_k)=0$ and $\varphi^*(x_1\cdots x_k)\neq 0$. Then ${\rm{secat}\hskip1pt}_\varphi(p)\ge k$.
\end{enumerate}
In addition, if either $n\geq1$ or $\mathcal{A}$ is path-connected, then
\begin{enumerate}\addtocounter{enumi}{5}
\item ${\rm{secat}\hskip1pt}_\varphi(p)$ equals the smallest $n$ such that the map $\varphi: K\to X$ admits a (pointed) lift through $p_n: J^n_X(\mathcal{A})\to X$.
\end{enumerate}
\end{prop}
If $\varphi: K\to X$ we denote ${\rm{secat}\hskip1pt}_\varphi(p_X)$ by ${\rm{cat}\hskip1pt}_\varphi(X)$, or by ${\rm{cat}\hskip1pt}_K(X)$ when $\varphi$ is an inclusion. Similarly if $\varphi: K\to X\times X$ we denote ${\rm{secat}\hskip1pt}_\varphi(\pi_X)$ by ${\rm{TC}\hskip1pt}_\varphi(X)$, or by ${\rm{TC}\hskip1pt}_K(X)$ when $\varphi$ is an inclusion. Note that this notation differs from the notation for subspace category and subspace topological complexity used in \cite{CLOT}, \cite{FarInv} and \cite{Grant}, for example.
The main advantage of relative sectional category over its absolute counterpart is monotonicity: if $K\subseteq K'\subseteq X$ and $p:{\mathcal{A}}\to X$ is a fibration, then ${\rm{secat}\hskip1pt}_K(p)\le {\rm{secat}\hskip1pt}_{K'}(p)$. Moreover, the relative sectional category either remains the same or increases by one on attaching a cell, or more generally on attaching a cone along a map. The following result and its proof are integral to the results in this paper.
\begin{prop}\label{secatupbyone}
Let $p: \mathcal{A} \to X$ be a fibration. Suppose that $X=K\cup_\alpha CS$ is the mapping cone of a map $\alpha: S\to K$. Then
\[
{\rm{secat}\hskip1pt}_K(p)\le {\rm{secat}\hskip1pt}(p) \le {\rm{secat}\hskip1pt}_K(p)+1.
\]
\end{prop}
\begin{proof}
We may assume from the outset that $\alpha$ is a closed cofibration, and therefore an inclusion. This follows from Proposition \ref{secatprops} (1) and the standard construction, replacing $K$ with the mapping cylinder of $\alpha$. In particular, $S$, $CS$, $K$, and $X$ all share the base point $\ast=[\ast,0]$.
Assume ${\rm{secat}\hskip1pt}_K(p)=n$, and choose a (pointed) lift $\phi: K\to J^{n}_X(\mathcal{A})$ of the inclusion $\iota: K\hookrightarrow X$ through the $(n+1)$-fold fibred join $p_n:J^{n}_X(\mathcal{A})\to X$. Let
\[
h_t : CS\to X = K\cup_S CS
\]
be a (pointed) homotopy which contracts the cone to its base point (indicated above). Then $h_t\circ\alpha :S\to X$ is a (pointed) null-homotopy of $\iota\circ\alpha = p_n\circ \phi\circ \alpha$. Since $p_n$ is a (pointed) fibration, we can lift {$h_t\circ\alpha$} to a (pointed) homotopy from $\phi\circ \alpha$ to a map with values in $J^n(A)=p_n^{-1}(\ast)$. Since $\alpha$ is a {(pointed)} cofibration, we can extend {the latter} homotopy to a (pointed) homotopy $k_t:K\to J^n_X({\mathcal{A}})$ from $\phi$ to a map $\phi':K\to J^n_X({\mathcal{A}})$ such that $\phi'\circ \alpha$ takes values in $J^n(A)$. We therefore have a strictly commuting square
\[
\xymatrix{
S \ar[r]^\alpha \ar[d]^H & K \ar[d]^{\phi'} \\
J^n(A) \ar[r]^{i_n} & J^n_X({\mathcal{A}})
}
\]
where the map $H$ is obtained by restriction of domain and codomain. Note that $p_n\circ k_t$ is a homotopy from $\iota$ to $p_n\circ \phi'$.
The maps
\[
CS \stackrel{CH}{\longrightarrow} CJ^n(A) \stackrel{\kappa_n}\longrightarrow J^{n+1}_X({\mathcal{A}})\qquad\mbox{and}\qquad K \stackrel{\phi'}{\longrightarrow} J^n_X({\mathcal{A}}) \stackrel{\jmath_n}{\longrightarrow} J^{n+1}_X(\mathcal{A})
\]
agree on $S$ and so together define a map $\sigma: X\to J^{n+1}_X({\mathcal{A}})$. We claim that $\sigma$ is a homotopy section of $p_{n+1}:J^{n+1}_X({\mathcal{A}})\to X$, and therefore ${\rm{secat}\hskip1pt}(p)\le n+1 = {\rm{secat}\hskip1pt}_K(p)+1$.
To prove the claim we exhibit an explicit homotopy from the identity map of $X=K\cup_S CS$ to $p_{n+1}\circ \sigma$. {By construction,} the homotopies $h_t: CS\to X$ and $p_n\circ k_t: K\to X$ agree on $S$, and therefore glue together to give {a homotopy $H_t\colon X\to X$. It is clear that $H_0$ is the identity. On the other hand, $H_1=p_{n+1}\circ \sigma$ because, again by construction, both maps send $CS$ to the basepoint $\ast$ and are given by $p_{n+1}\circ\jmath_n\circ\phi' = p_n\circ \phi'$ on $K$.}
The inequality ${\rm{secat}\hskip1pt}_K(p)\le {\rm{secat}\hskip1pt}(p)$ comes directly from item~(\ref{itemdosdel38}) in Proposition~\ref{secatprops}.
\end{proof}
\begin{cor}\label{relsecatupbyone}
Let $p:\mathcal{A}\to X$ be a fibration, and let $\varphi:X'\to X$ be any map. Suppose that $X'=K\cup_\alpha CS$ is the mapping cone of a map $\alpha: S\to K$. Then
\[
{\rm{secat}\hskip1pt}_{\varphi|_K}(p)\le {\rm{secat}\hskip1pt}_\varphi(p) \le {\rm{secat}\hskip1pt}_{\varphi|_K}(p)+1.
\]
\end{cor}
\begin{proof} This is Proposition \ref{secatupbyone} applied to $\varphi^* p$.
\end{proof}
\begin{cor}\label{amejorarse}
For any fibration $p: {\mathcal{A}}\to X$ we have ${\rm{secat}\hskip1pt}(p)\le \operatorname{cl}(X)$ where $\operatorname{cl}(X)$ denotes the cone length of $X$.
\end{cor}
Corollary~\ref{amejorarse} is of course improved by the standard estimate ${\rm{secat}\hskip1pt}(p)\le{\rm{cat}\hskip1pt}(X)$. The real strength of Proposition~\ref{secatupbyone} will become apparent with the constructions in the next section.
\begin{rem}\label{neuralgico}
We could have given a simpler proof of Proposition~\ref{secatupbyone} using Lemma \ref{extendsection} in the next section. Note however that the above proof furnishes a strictly commuting cubical diagram
$$
\xymatrix{
& S \ar[rr]^{\alpha} \ar[dl]_{H} \ar@{^{(}->}'[d][dd] & & K \ar[dd]^\iota \ar[dl]_{\phi'} \\
J^n(A)\ar[rr] \ar@{^{(}->}[dd] & & J^n_X({\mathcal A}) \ar[dd] & \\
& CS \ar'[r][rr] \ar[ld]^{CH} & & X \ar[ld]_{\sigma} \\
CJ^n(A)\ar[rr]^{\kappa_n} & & J^{n+1}_X({\mathcal A}) &}
$$
in which $\phi'$ is a (pointed) homotopy lifting of $\iota$ through $p_n$ and $\sigma$ is a (pointed) homotopy section of $p_{n+1}$. This diagram will be especially important in what follows.
\end{rem}
\section{Hopf invariants for sectional category}\label{hifscsec}
In this section we introduce the Hopf invariants which determine whether the relative sectional category increases on attaching a cone. We give the main definitions in {Subsection} \ref{HopfDefns}, and then in {Subsection} \ref{HopfProducts} prove a fundamental result about Hopf invariants of product fibrations.
\subsection{Definitions}\label{HopfDefns}
Before giving the definition of the Hopf invariants considered in this paper, we record a couple of technical lemmas.
\begin{lem}\label{fibjoinssplit}
Given any fibration $p:\mathcal{A}\to X$ and $n\geq 1$, the $(n+1)$-fold fibred join $p_n:J^n_X(\mathcal{A})\rightarrow X$ splits after looping once. Consequently, if $Y=\widetilde{\Sigma}S$ is a suspension, then the induced maps of (pointed) homotopy groups $(p_n)_*:[Y,J^n_X(\mathcal{A})]\to [Y,X]$ and $(i_n)_*:[Y,J^n(A)]\to [Y,J^n_X(\mathcal{A})]$ are split surjective and split injective, respectively.
\end{lem}
\begin{proof}
Using a (pointed) lifting function for $p$ we may construct a map $\chi: PX\to \mathcal{A}$ rendering the following diagram commutative:
\[
\xymatrix{
PX \ar[rr]^{\chi} \ar[rd]_{p_X} & & \mathcal{A} \ar[ld]^p \\
& X &
}
\]
The fibred join construction is functorial for fibrewise maps, and so we obtain diagrams
\[
\xymatrix{
G_n(X) \ar[rr] \ar[rd]_{g_n(X)} & & J^n_X(\mathcal{A}) \ar[ld]^{p_n} \\
& X &
},
\qquad
\xymatrix{
\Omega G_n(X) \ar[rr] \ar[rd]_{\Omega g_n(X)} & & \Omega J^n_X(\mathcal{A}) \ar[ld]^{\Omega p_n} \\
& \Omega X &
}
\]
for each $n\ge 0$. Since $\Omega g_n(X)$ admits a homotopy section for $n\ge1$ (see \cite[Exercise 2.1]{CLOT}, for instance) so does $\Omega p_n$, and the result follows.
\end{proof}
\begin{lem}\label{extendsection}
Let $S\stackrel{\alpha}{\to} K\stackrel{\iota}{\to} X=K\cup_\alpha CS$ be a cofibration sequence, and let $\rho: Z\to X$ {and $\phi: K\to Z$ be maps with homotopies} $\phi\circ \alpha\simeq *$ and $\rho\circ \phi \simeq \iota$. If $\rho_\ast:[{\widetilde{\Sigma}} S,Z]\to [{\widetilde{\Sigma}} S,X]$ is surjective, then there exists a map $\sigma: X\to Z$ and pointed homotopies $\sigma\circ \iota \simeq \phi$ and $\rho\circ \sigma \simeq \mathrm{Id}_X$.
\end{lem}
\begin{proof}
This is a slight generalization of \cite[Lemma 6.28]{CLOT}, with the same proof.
\end{proof}
\begin{defn}\label{Hopfinv}
Let $p:\mathcal{A}\to X$ be a fibration. Suppose that $X=K\cup_\alpha CS$ is the mapping cone of a map $\alpha: S\to K$. Suppose also that ${\rm{secat}\hskip1pt}_K(p)\le n$, and let $\phi: K\to J^n_X(\mathcal{A})$ be a (pointed) homotopy lifting of the inclusion $\iota: K\hookrightarrow X$ through $p_n: J^n_X({\mathcal{A}})\to X$. As in the proof of Proposition \ref{secatupbyone}, {consider a pointed-}homotopy commutative diagram
\[
\xymatrix{
S \ar[r]^\alpha \ar[d]^H & K \ar[d]^{\phi} \\
J^n(A) \ar[r]^{i_n} & J^n_X({\mathcal{A}}).
}
\]
If $n\ge 1$ and $S$ is a {reduced} suspension, then by Lemma \ref{fibjoinssplit} the pointed-homotopy class of the map $H: S \to J^n(A)$ depends only on the pointed-homotopy classes of $\alpha$ and~$\phi$. Any representative of this class will be denoted $H^n_{\phi,\alpha}(p)$ and called the {\em Hopf invariant} associated to the data $(p,n,\phi,\alpha)$. The set of all such Hopf invariants as $\phi$ ranges over all possible (pointed) lifts is denoted $\mathcal{H}_{\alpha}^n(p)$ and called the {\em Hopf set} associated to $(p,n,\alpha)$.
\end{defn}
\begin{prop}\label{caracterizacionshida}
Under the conditions of Definition \ref{Hopfinv}, we have ${\rm{secat}\hskip1pt}(p)\le n$ if and only if the Hopf set $\mathcal{H}^n_\alpha(p)$ contains the trivial element.
\end{prop}
\begin{proof}
Suppose there is a lift $\phi: K\to J^n_X(\mathcal{A})$ such that the associated Hopf invariant $H^n_{\phi,\alpha}(p):S\to J^n(A)$ is null-homotopic. Then $\phi\circ\alpha\simeq *$ and so Lemma \ref{extendsection} gives a homotopy section $\sigma$ of $p_n$ extending $\phi$ {up to pointed homotopy.}
Conversely, suppose that ${\rm{secat}\hskip1pt}(p)\le n$ and let $\sigma: X\to J^n_X(\mathcal{A})$ be a homotopy section of $p_n$. Then $\phi:=\sigma\circ\iota: K\to J^n_X(\mathcal{A})$ is a homotopy lifting of $\iota$ through $p_n$ satisfying $\phi\circ\alpha\simeq \ast$, whose associated Hopf invariant is therefore trivial.
\end{proof}
\begin{prop}\label{uniqueness}
Under the conditions of Definition \ref{Hopfinv}, suppose that the fibre $A$ is $(r-1)$-connected with $r\ge1$. If ${\rm{hdim}\hskip1pt}(K)<(n+1)r+n$, then the Hopf set $\mathcal{H}^n_\alpha(p)$ consists of a single element.
\end{prop}
\begin{proof}
Note that we are assuming ${\rm{secat}\hskip1pt}_K(p)\le n$, so that the Hopf set is non-empty. If $A$ is $(r-1)$-connected then $J^n(A)$ is $\big((n+1)r + n-1\big)$-connected. Then $p_n$ is an $\big((n+1)r + n\big)$-equivalence, and it follows that the induced map $(p_n)_*:[K,J^n_X(\mathcal{A})]\to [K,X]$ is bijective when
${\rm{hdim}\hskip1pt}(K)<(n+1)r + n$ (see~\cite[Corollary 7.6.23]{Spanier}). Thus the lifting $\phi$ of $\iota$ is unique up to homotopy.
\end{proof}
\begin{ex}[Berstein--Hilton--Hopf invariants \cite{B-H,CLOT,Iwase,IwaseAinfty,Stanley}]\label{ejemploBHH}
Let $K$ be a path-connected space with ${\rm{cat}\hskip1pt}(K)\le n\ge 1$, and let $\alpha: S^q\to K$ be a map with $q\ge 1$. The cofiber $X=K\cup_\alpha CS$ of $\alpha$ satisfies ${\rm{cat}\hskip1pt}(X)\le n+1$. Berstein and Hilton introduced in~\cite{B-H} generalized Hopf invariants to detect whether ${\rm{cat}\hskip1pt}(X)\le n$. Here we give the modification of their definition used by Iwase in \cite{Iwase}.
Let $s: K\to G_n(K)$ be a (pointed) section of $g_n(K):G_n(K)\to K$, the $n$-th Ganea fibration of $K$. Then we define:
\begin{itemize}
\item $H_s'(\alpha):=s\circ \alpha -G_n(\alpha)\circ s_0 \in \pi_{q}(G_n(K))$ where $s_0$ is the canonical section of $g_n(S^q)$.
\item $H_s(\alpha)\in \pi_q(F_n(K))$ the unique (up to homotopy) map satisfying $H_s'(\alpha)= i_n(K)\circ H_s(\alpha)$.
\end{itemize}
Both of these elements will be called the \emph{Berstein--Hilton--Hopf invariant of $\alpha$ associated to $s$}. The set of such elements as $s$ ranges over all (homotopy classes of) such sections is denoted $\mathcal{H}'(\alpha)\subseteq \pi_q(G_n(K))$ or $\mathcal{H}(\alpha)\subseteq \pi_q(F_n(K))$, and called the \emph{Berstein--Hilton--Hopf set of $\alpha$}. If $K$ is a CW-complex, it is shown in~\cite[Section~6.4]{CLOT} that
\begin{equation}\label{recuperarclot}
\mbox{{\it ${\rm{cat}\hskip1pt}(X)\le n$ if and only if $\,0\in \mathcal{H}(\alpha)$, provided $\max\{\dim(K),2\}\le q$.}}
\end{equation}
We now make explicit the relationship of these Berstein--Hilton--Hopf invariants with the Hopf invariants discussed in this section. Let $\iota: K\to X$ denote the inclusion into the cofiber. By Proposition \ref{secatprops} we have ${\rm{cat}\hskip1pt}_K(X)\le{\rm{cat}\hskip1pt}(K)\le n$, and indeed
\begin{equation}\label{primerafactorizacion}
\phi=G_n(\iota)\circ s:K\to G_n(X)
\end{equation}
is a (pointed) lifting of $\iota$ through $g_n(X)$. Then the associated Hopf invariant $H^n_{\phi,\alpha}(g_0(X)):S^q\to F_n(X)$ satisfies $H^n_{\phi,\alpha}(g_0(X)) = F_n(\iota)\circ H_s(\alpha)$. This follows from the definitions, together with the diagram
\[
\xymatrix{
F_n(K) \ar[r]^{i_n(K)} \ar[d]_{F_n(\iota)} & G_n(K) \ar[d]^{G_n(\iota)} \\
F_n(X) \ar[r]^{i_n(X)} & G_n(X)
}
\]
and the observation that $G_n(\iota)\circ H_s'(\alpha)\simeq G_n(\iota)\circ s\circ\alpha$ since $G_n$ is a homotopy functor and $\iota\circ\alpha$ is null-homotopic. {Thus}
\begin{equation}\label{lainclusion}
F_n(\iota)_\ast\big( \mathcal{H}(\alpha)\big)\subseteq \mathcal{H}^n_{\alpha}(g_0(X)).
\end{equation}
Finally, we note that {(\ref{lainclusion}) can be improved to an equality under the hypothesis in~(\ref{recuperarclot}). Namely, if} $K$ is a connected complex {and $q\ge\max\{\dim(K),2\}$, then} the maps $F_n(\iota):F_n(K)\to F_n(X)$ and $G_n(\iota):G_n(K)\to G_n(X)$ are $(q+1)$-equivalences for all $n\ge 1$, by \cite[Lemma 6.26]{CLOT}. Consequently, any (pointed) homotopy lift $\phi$ of $\iota$ through $g_n(X)$ arises as in~(\ref{primerafactorizacion}) for some $s$, so that in fact $F_n(\iota)_\ast\big( \mathcal{H}(\alpha)\big)= \mathcal{H}^n_{\alpha}(g_0(X))$. Furthermore, the triviality of $H^n_{\phi,\alpha}(g_0(X))$ is equivalent to the triviality of $H_s(\alpha)$, so that Proposition~\ref{caracterizacionshida} recovers~(\ref{recuperarclot}).
\end{ex}
\begin{ex}[cat-Hopf invariants of spheres]\label{cathopfsphere}
Let $q\ge 2$. We may regard the $q$-sphere $S^q$ as the cofiber $C^-S^{q-1} \cup_\alpha CS^{q-1}$ of the inclusion $\alpha: S^{q-1}\hookrightarrow C^-S^{q-1}$ of the base of the cone $C^-S^{q-1}=S^{q-1}\times [-1,0]/S^{q-1}\times\{-1\}$. The base point of this cone is $[\ast,0]$, where $\ast$ is the base point of $S^{q-1}$, and $\alpha$ is a pointed cofibration. Here, ${\rm{cat}\hskip1pt}_{C^-S^{q-1}}(S^q)=0$ and ${\rm{cat}\hskip1pt}(S^q)=1$. Since $n=0$ and the fibration $g_0(S^q): G_0(S^q)\to S^q$ does not split after looping, the uniqueness statement in Definition~\ref{Hopfinv} breaks down. We can, however, define a Hopf invariant $H^0(S^q)$ as follows.
Fix a pointed homotopy equivalence $\xi: S^q\to \widetilde \Sigma S^{q-1}$ between $S^q$ defined as above and the reduced suspension of $S^{q-1}$, and let $\xi^{-1}$ denote a {pointed} homotopy inverse of $\xi$. Denote by $\eta: S^{q-1}\to \Omega\widetilde\Sigma S^{q-1}$ the standard adjunction, given by $\eta(x)(t) = [x,t]$. We define $\tilde\eta: S^{q-1}\to F_0(S^q)$ to be the composition
$$S^{q-1} \stackrel{\eta}{\to} \Omega \widetilde \Sigma S^{q-1} \stackrel{\Omega \xi^{-1}}{\to} \Omega S^q = F_0(S^q),$$
and note that $\tilde\eta$ is a pointed map.
With these preliminaries, we can construct a commuting diagram
\[
\xymatrix{
& S^{q-1} \ar[rr] \ar[ld]_{\tilde\eta} \ar@{^{(}->}'[d][dd] & & C^-S^{q-1} \ar[dd] \ar[ld]_{\phi'} \\
F_0(S^q)\ar[rr] \ar@{^{(}->}[dd] & & G_0(S^q) \ar[dd] & \\
& CS^{q-1} \ar'[r][rr] \ar[ld]_{C\tilde\eta} & & S^q \ar[ld]_{\sigma}\\
CF_0(S^q) \ar[rr] & & G_1(S^q) &
}
\]
where $\phi'([x,u])(t) = \tilde\eta(x)\big( (1+u)t\big)$, {which is of course a homotopy lifting of the inclusion $C^-S^{q-1}\hookrightarrow S^q$ through $g_0(S^q)$,} and
\[
\sigma([x,u]) =\left\{
\begin{array}{lr}
\left\langle \hspace{.1mm}{\phi'([x,u])
\hspace{.3mm}, \hspace{.2mm}
\phi'([x,u])}
\mid 1,0\rule{0mm}{3.2mm}\right\rangle, & -1\leq u\leq 0; \\
\left\langle \tilde\eta(x), *\mid 1-u,u\rule{0mm}{3.2mm}\right\rangle, & 0\leq u\leq 1.
\end{array}
\right.
\]
It is straightforward to check that the diagram commutes and $\sigma: S^q\to G_1(S^q)$ is a homotopy section of $g_1(S^q)$. {The situation is now analogous to that in Remark~\ref{neuralgico} and we} may therefore consider the homotopy class of $\tilde\eta$ as the Hopf invariant $H^0(S^q)$. {Note that, by construction, $H^0(S^q)$ is homotopic to the adjoint of the identity on $\widetilde{\Sigma}S^{q-1}$ and, therefore, can be identified up to homotopy to the inclusion of the bottom cell in $F_0(S^q)=\Omega S^q\simeq S^{q-1}\cup e^{2(q-1)}\cup\cdots$.}
\end{ex}
\subsection{Products} \label{HopfProducts}
Let $p:{\mathcal A}\to X$ and $q:{\mathcal B}\to Y$ be fibrations with respective fibres $A$ and $B$. Our main result in this section shows how to construct Hopf invariants of $p\times q: \mathcal{A}\times\mathcal{B}\to X\times Y$ in terms of Hopf invariants of the factors. We begin by recalling the exterior join construction \cite{Bau,Stanley}. The following is a slight generalization of \cite[Proposition 2.9]{Stanley}.
\begin{prop}\label{exteriorjoin}
{Given commutative} diagrams
\[
\xymatrix{
S(i) \ar[r]^{f(i)} \ar@{^{(}->}[d] & K(i) \ar[d]^{g(i)}\\
CS(i) \ar[r]_{F(i)} & X(i)
}
\]
for $i=1,2$, consider the commutative cube
\[
\xymatrix{
& S(1)\times S(2) \ar[ld]_{f(1)\times f(2)\hspace{3mm}} \ar@{^{(}->}'[d][dd] \ar@{^{(}->}[rr] & & CS(1)\times S(2) \ar[ld]^{\hspace{3mm}F(1)\times f(2)} \ar@{^{(}->}[dd] \\
K(1)\times K(2) \ar[rr]^{\hspace{2cm}g(1)\times1} \ar[dd]_{1\times g(2)} &
& X(1)\times K(2) \ar[dd]_>>>>>>{1\times g(2)} & \\
& S(1)\times CS(2) \ar@{^{(}->}'[r][rr] \ar[ld]_{f(1)\times F(2)\hspace{2mm}} & & CS(1)\times CS(2) \ar[dl]^{\hspace{3mm}F(1)\times F(2)} \\
K(1) \times X(2) \ar[rr]_{g(1)\times1} & & X(1)\times X(2). &
}
\]
Recall $S(1)\circledast S(2)$ is the push-out of the ``back'' face of the above cube with whisker map given by the obvious inclusion $S(1)\circledast S(2)\hookrightarrow CS(1)\times CS(2)$. Likewise, let $X(1)\times K(2)\cup K(1)\times X(2)$ stand for the push-out of the ``front'' face of the above cube, and let
\begin{equation}\label{doswhiskersobvios}
S(1)\circledast S(2) \stackrel{W}{\longrightarrow} X(1)\times K(2)\cup K(1)\times X(2) \stackrel{W'}{\longrightarrow} X(1)\times X(2)
\end{equation}
be the obvious whisker maps.
We obtain a commutative diagram
\begin{equation}\label{diagresultante}\xymatrix{
S(1)\circledast S(2) \ar[rr]^-{W}\ar@{^{(}->}[d] & & X(1)\times K(2) \cup K(1) \times X(2) \ar[d]^{{W'}} \\
CS(1)\times CS(2) \ar[rr]^{{F(1)\times F(2)}} & & X(1)\times X(2).
}\end{equation}
If the two initial diagrams are homotopy push-outs {(respectively, strict push-outs),} then so is~{(\ref{diagresultante}).} Furthermore, this construction is natural in the following sense. Suppose for $i=1,2$ there are maps $H(i): S(i)\to S'(i)$, $\phi(i): K(i)\to K'(i)$ and $\sigma(i): X(i)\to X'(i)$ rendering the following cube commutative:
\[
\xymatrix{
& S(i) \ar[ld]^{H(i)} \ar@{^{(}->}'[d][dd] \ar[rr] & & K(i) \ar[ld]^{\phi(i)} \ar[dd] \\
S'(i) \ar[rr] \ar@{^{(}->}[dd] & & K'(i) \ar[dd] & \\
& CS(i) \ar'[r][rr] \ar[ld]^{CH(i)} & & X(i) \ar[dl]^{\sigma(i)} \\
CS'(i) \ar[rr] & & X'(i) &
}
\]
Then there results a commutative cube
\[
\resizebox{1\textwidth}{!}{\xymatrix{
& S(1)\circledast S(2) \ar[dl]_{H(1)\circledast H(2)\hspace{3mm}} \ar@{^{(}->}'[d][dd] \ar[rr] & & X(1)\times K(2) \cup K(1)\times X(2) \ar[dl]^{\hspace{3mm}{W''}} \ar[dd] \\
S'(1)\circledast S'(2) \ar[rr] \ar@{^{(}->}[dd] & & X'(1)\times K'(2) \cup K'(1)\times X'(2) \ar[dd] & \\
& CS(1)\times CS(2) \ar'[r][rr] \ar[dl]^>>>>>>>>>>>>>>{\hspace{6mm}CH(1)\times CH(2)} & & X(1)\times X(2) \ar[dl]^{\hspace{3mm}\sigma(1)\times\sigma(2)} \\
CS'(1)\times CS'(2) \ar[rr] & & X'(1)\times X'(2) &
}}
\]
{where $W''$ is the obvious whisker map.}
\end{prop}
\begin{rem}\label{conosenelextjoicon}
As shown in~\cite[Lemma 2.8]{Stanley}, $CS(1) \times CS(2)$ is homeomorphic to $C(S(1) \circledast S(2))$ in such a way that the inclusion $S(1) \circledast S(2) \hookrightarrow CS(1) \times CS(2)$ corresponds to the inclusion of the base of the cone.
\end{rem}
{For fibrations $p\colon\mathcal{A}\to X$ and $q\colon\mathcal{B}\to Y$ with respective fibers $A$ and $B$, and} for each pair of non-negative integers $n$ and $m$, we will construct in Section~\ref{secsiete} {maps $\Phi^{A,B}_{n,m}$ and $\Psi^{\mathcal{A},\mathcal{B}}_{n,m}$ fitting in the} commutative diagram
\begin{equation}\label{shufflewjoins}
\xymatrix{
J^ n(A)\circledast J^m(B)\ar[rr]^-{{\Phi_{n,m}^{A,B}}} \ar[d]_{{W}}&& J^{n+m+1}(A\times B)\ar[d]^{{i_{n+m+1}}} \\
J^{n+1}_X({\mathcal A})\times J_Y^{m}({\mathcal B})\cup J_X^{n}({\mathcal A})\times J_Y^{m+1}({\mathcal B})\ar[rr]^-{{\Psi_{n,m}^{{\mathcal A},{\mathcal B}}}} \ar[d]_{{W'}}&& J_{X\times Y}^{n+m+1}({\mathcal A}\times {\mathcal B})\ar[d]^{{(p\times q)_{n+m+1}}} \\
J^{n+1}_X({\mathcal A})\times J_Y^{m+1}({\mathcal B})\ar[rr]^{{p_{n+1}\times q_{m+1}}} &&X\times Y}
\end{equation}
where $W$ and $W'$ are as in~(\ref{doswhiskersobvios}), given the diagram~(\ref{ijkappa}) and the analogous square for $\mathcal{B}$ (replacing $n$ by $m$). The map $\Phi_{n,m}^{A,B}$ is a sort of topological shuffle map, and is natural in $A$ and $B$. We postpone further details to Section 7.
The following is a generalisation to arbitrary fibrations of a result due to Iwase (\cite[Proposition 5.8]{Iwase}, \cite[Theorem 5.5]{IwaseAinfty}, see also \cite{Harper}).
\begin{thm}\label{Hopfproducts}
Let $p:{\mathcal A}\to X$ and $q:{\mathcal B}\to Y$ be fibrations with respective fibres $A$ and $B$. Suppose that $X=K\cup_\alpha CS$ is the cofibre of a map $\alpha: S\to K$, and $Y=L\cup_\beta CT$ is the cofibre of a map $\beta: T\to L$, where $S$ and $T$ are reduced suspensions. If ${\rm{secat}\hskip1pt}_K(p)\le n$ with Hopf invariant $H^n_{\phi,\alpha}(p):S\to J^n(A)$ {associated to a (pointed) homotopy lifting $\phi \colon K \to J^n_X(\mathcal{A})$ of the inclusion $K\hookrightarrow X$ through $p_n$,} and ${\rm{secat}\hskip1pt}_L(q)\le m$ with Hopf invariant $H^m_{\psi,\beta}(q): T\to J^m(B)$ {associated to a (pointed) homotopy lifting $\psi \colon L \to J^m_Y(\mathcal{B})$ of the inclusion $L\hookrightarrow Y$ through $q_m$ (as in Definition~\ref{Hopfinv}),} then ${\rm{secat}\hskip1pt}_{X\times L \cup K\times Y}(p\times q)\le n+m+1$ with Hopf invariant the composition
\[
\xymatrix{
S\circledast T \ar[rrr]^-{H^n_{\phi,\alpha}(p)\circledast H^m_{\psi,\beta}(q)} &&& J^n(A)\circledast J^m(B) \ar[rr]^-{\Phi_{n,m}^{A,B}} & & J^{n+m+1}(A\times B).
}
\]
\end{thm}
\begin{proof}
We have commutative cubes
\[
\resizebox{1\textwidth}{!}{
\xymatrix{
& S \ar[rr]^{\alpha} \ar[dl]_{H^n_{\phi,\alpha}(p)} \ar@{^{(}->}'[d][dd] & & K \ar[dd] \ar[dl]_{\phi'} & & T \ar[rr]^{\beta} \ar[ld]_{H^m_{\psi,\beta}(q)} \ar@{^{(}->}'[d][dd] & & L \ar[dd] \ar[ld]_{\psi'} \\
J^n(A)\ar[rr] \ar@{^{(}->}[dd] & & J^n_X({\mathcal A}) \ar[dd] & & J^m(B)\ar[rr] \ar@{^{(}->}[dd] & & J^m_Y({\mathcal B}) \ar[dd] & \\
& CS \ar'[r][rr] \ar[ld]^>>>>>>>>>>{\;CH^n_{\phi,\alpha}(p)} & & X \ar[ld]_{\sigma} && CT \ar'[r][rr] \ar[ld]^>>>>>>>>>>{\;CH^m_{\psi,\beta}(q)} & & Y \ar[ld]_{\tau}\\
CJ^n(A)\ar[rr] & & J^{n+1}_X({\mathcal A}) && CJ^m(B) \ar[rr] & & J^{m+1}_Y({\mathcal B}) &
}}
\]
constructed as in Proposition \ref{secatupbyone}, where $\sigma$ and $\tau$ are (pointed) homotopy sections of $p_{n+1}$ and $q_{m+1}$ respectively. Applying the naturality statement of Proposition \ref{exteriorjoin}, and splicing the top and right faces of the resulting cube with diagram (\ref{shufflewjoins}), yields a large diagram
\[
\resizebox{1\textwidth}{!}{\xymatrix{
S\circledast T \ar[r]^-{{W}} \ar[d]_{H^n_{\phi,\alpha}(p)\circledast H^m_{\psi,\beta}(q)} & X\times L \cup K \times Y \ar[d] \ar[r]^{{W'}} & X\times Y \ar[d]_{\sigma\times\tau} \\
J^n(A)\circledast J^m(B) \ar[r] \ar[d]_{\Phi_{n,m}^{A,B}} & J^{n+1}_X({\mathcal A})\times J_Y^{m}({\mathcal B})\cup J_X^{n}({\mathcal A})\times J_Y^{m+1}({\mathcal B}) \ar[r] \ar[d]_{\Psi_{n,m}^{{\mathcal A},{\mathcal B}}} & J^{n+1}_X({\mathcal A})\times J_Y^{m+1}({\mathcal B}) \ar[d]_{p_{n+1}\times q_{m+1}} \\
J^{n+m+1}(A\times B) \ar[r] & J_{X\times Y}^{n+m+1}({\mathcal A}\times {\mathcal B}) \ar[r]^{(p\times q)_{n+m+1}} & X\times Y.
}}
\]
One easily sees that the middle vertical composition is a (pointed) homotopy lifting of the inclusion $X\times L \cup K \times Y\hookrightarrow X\times Y$ through $(p\times q)_{n+m+1} :J_{X\times Y}^{n+m+1}({\mathcal A}\times {\mathcal B})\to X\times Y$, hence ${\rm{secat}\hskip1pt}_{X\times L \cup K\times Y}(p\times q)\le n+m+1$. The left-hand vertical composition is the Hopf invariant associated to this lifting.
\end{proof}
\begin{cor}\label{forTC2cell}
Let $K$ and $L$ be path-connected spaces, and let
$X=K\cup_\alpha e^{q+1}$ and $Y=L\cup_\beta e^{r+1}$ for maps $\alpha: S^q\to K$ and $\beta: S^r\to L$ {with $q,r\ge1$.} Suppose that ${\rm{cat}\hskip1pt}(K)\le n$ and ${\rm{cat}\hskip1pt}(L)\le m$ {with $n,m\ge1$,} and let $s: K\to G_n(K)$ and $\sigma: L\to G_m(L)$ be (pointed) sections of the respective Ganea fibrations. Then:
\begin{enumerate}[(a)]
\item ${\rm{cat}\hskip1pt}(X\times Y)\le n+m+1$ if the composition
\[
\xymatrix{
S^q\circledast S^r \ar[rr]^-{H_s(\alpha)\circledast H_\sigma(\beta)} && F_n(K)\circledast F_m(L) \ar[r]^-{\Phi_{n,m}^{\Omega K,\Omega L}} & F_{n+m+1}(K\times L)
}
\]
is null-homotopic.
\item ${\rm{TC}\hskip1pt}(X)\le 2n+1$ if the composition
\[
\xymatrix{
S^q\circledast S^q \ar[rr]^-{H_s(\alpha)\circledast H_s(\alpha)} && F_n(K)\circledast F_n(K) \ar[r]^-{\Phi^{\Omega K,\Omega K}_{n,n}} & F_{2n+1}(K\times K) \ar[r]^>>>>>{\bar\chi_{2n+1}} & F_{2n+1}(K)
}
\]
is null-homotopic.
\item If $K=S^p$, where $p\ge 2$ and $q\le 3p-3$, then ${\rm{TC}\hskip1pt}_{X\times S^p}(X)\le 2$ if and only if the composition
\[
\xymatrix{
S^q\circledast S^{p-1} \ar[rr]^-{H_s(\alpha)\circledast H^0(S^p)} && F_1(S^p)\circledast F_0(S^p) \ar[r]^>>>>{\Phi^{\Omega S^p,\Omega S^p}_{1,0}} & F_2(S^p\times S^p) \ar[r]^>>>>>{\bar\chi_2} & F_2(S^p)
}
\]
is null-homotopic.
\end{enumerate}
\end{cor}
\begin{proof}
\begin{enumerate}[(a)]
\item Naturality of the maps $\Phi_{n,m}$, together with Theorem \ref{Hopfproducts} applied to the product of the fibrations $g_0(X):G_0(X)\to X$ and $g_0(Y):G_0(Y)\to Y$, yield ${\rm{cat}\hskip1pt}_{X\times K\cup L\times Y}(X\times Y)\le n+m+1$ with Hopf invariant the lower composition in the diagram
\[\xymatrix{
S^q\circledast S^r \ar[rr]^-{H_s(\alpha)\circledast H_\sigma(\beta)} \ar[drr] && F_n(K)\circledast F_m(L) \ar[d] \ar[rr]^-{\Phi_{n,m}^{\Omega K,\Omega L}} && F_{n+m+1}(K\times L)\ar[d] \\
&& F_n(X)\circledast F_m(Y) \ar[rr]^-{\Phi_{n,m}^{\Omega X,\Omega Y}} && F_{n+m+1}(X\times Y).
}\]
Here the {vertical} maps are induced by inclusions, {whereas the slanted map is $H^n_{G_n(\iota)\circ s,\alpha}(g_0(X))\circledast H^m_{G_m(\iota)\circ \sigma,\beta}(g_0(Y))$}. The result follows.
\item By naturality of the maps $\Phi_{n,n}$ and $\bar\chi_{2n+1}$ we obtain a diagram
\[
\xymatrix{
S^q\circledast S^q \ar[rr]^-{H_s(\alpha)\circledast H_s(\alpha)} \ar[drr] && F_n(K)\circledast F_n(K) \ar[d] \ar[r]^-{\Phi_{n,n}^{\Omega K,\Omega K}} & F_{2n+1}(K\times K)\ar[d]\ar[r]^-{\bar\chi_{2n+1}} & F_{2n+1}(K) \ar[d] \\
&& F_n(X)\circledast F_n(X) \ar[r]^-{\Phi_{n,n}^{\Omega X,\Omega X}} & F_{2n+1}(X\times X) \ar[r]^-{\bar\chi_{2n+1}} & F_{2n+1}(X).
}
\]
The diagram of Proposition \ref{TCcatHopf} shows that composition with $\bar\chi_{2n+1}$ takes Hopf invariants for ${\rm{cat}\hskip1pt}(X\times X)$ to Hopf invariants for ${\rm{TC}\hskip1pt}(X)$. Therefore ${\rm{TC}\hskip1pt}_{X\times K\cup K\times X}(X)\le 2n+1$ with Hopf invariant the lower composition. The result follows.
\item {We can safely assume $p\le q$ because if $\alpha$ is nullhomotopic, then in fact $H_s(\alpha)$ is nullhomotopic and ${\rm{TC}\hskip1pt}(X)\le2$. In particular, Proposition~\ref{uniqueness} and the discussion in Example~\ref{ejemploBHH} imply that $\mathcal{H}(\alpha)$ is a singleton. Think of $S^p$ with the cell structure in Example~\ref{cathopfsphere}, so} $X\times S^p\,=\,X\times C^-S^{p-1} \,\cup\, S^p\times S^p\,\cup\, e^{p+q+1}$. {The usual deformation retraction of the south hemisphere of $S^p$ to its south pole shows that the inclusion $X\times \ast \,\cup\, S^{p}\times S^p\hookrightarrow X\times C^-S^{p-1} \,\cup\, S^p\times S^p$ is a homotopy equivalence. Thus} ${\rm{TC}\hskip1pt}_{X\times CS^{p-1} \cup S^p\times S^p}(X)= {\rm{TC}\hskip1pt}_{X\times \ast \cup S^p\times S^p}(X)\le{\mathrm{cl}(X\times \ast \cup S^p\times S^p)}\le2$. {Further,} since $\Omega X$ is $(p-2)$-connected and $$\dim(X\times \ast \cup S^p\times S^p)=\max\{q+1,2p\}<3(p-1)+2 = 3p-1,$$ the Hopf set under consideration is a singleton and is given by the lower composition in the diagram
\[
\xymatrix{
S^q\circledast S^{p-1} \ar[rr]^-{H_s(\alpha)\circledast H^0(S^p)} \ar[rrd] && F_1(S^p)\circledast F_0(S^p) \ar[d] \ar[r]^-{\hspace{2mm}\Phi^{\Omega S^p,\Omega S^p}_{1,0}} & F_2(S^p\times S^p)\ar[d] \ar[r]^{\hspace{3mm}\bar\chi_2} & F_2(S^p) \ar[d]\\
&& F_1(X)\circledast F_0(X) \ar[r]^-{\Phi_{1,0}^{\Omega X,\Omega X}} & F_{2}(X\times X) \ar[r]^-{\bar\chi_{2}} & F_{2}(X).
}
\]
This proves the {``if''} statement. A homological argument shows that the map $F_2(S^p)\to F_2(X)$ is a $(2p+q-1)$-equivalence (compare \cite[proof of Lemma~6.26]{CLOT}), and so the lower composition is essential if and only if the upper composition is. This completes the proof.
\end{enumerate}
\end{proof}
\section{Application: The topological complexity of $2$-cell complexes}\label{secaptc2cw}
A $2$-cell complex is a finite complex $X = S^p \cup_\alpha e^{q+1}$ presented as the mapping cone of a map of spheres $\alpha: S^{q}\to S^p$, where $q\ge p\ge 1$. In this section we investigate the topological complexity of $2$-cell complexes, using the results of the previous section together with homological results proved in Section 7 below.
Recall that the Lusternik--Schnirelmann category of $X$ is determined as follows.
For $p=q$ and
\begin{itemize}
\item $\deg(\alpha)=\pm1$, $X$ is contractible and ${\rm{cat}\hskip1pt}(X)={\rm{TC}\hskip1pt}(X)=0$.
\item $\deg(\alpha)=0$, $X\simeq S^p\vee S^{p+1}$ and ${\rm{cat}\hskip1pt}(X)=1$ while ${\rm{TC}\hskip1pt}(X)=2$.
\item $|\deg(\alpha)|>1$, ${\rm{cat}\hskip1pt}(X)=2$ if $p=1$, whereas ${\rm{cat}\hskip1pt}(X)=1$ if $p>1$.
\end{itemize}
(The behavior of ${\rm{TC}\hskip1pt}(X)$ in the latter case is discussed below.)
For $p<q$, we can safely assume $p\ge2$---for otherwise $\alpha$ is nullhomotopic, in which case ${\rm{cat}\hskip1pt}(X)={\rm{cat}\hskip1pt}(S^p\vee S^{q+1})=1$. Then the Berstein--Hilton--Hopf set $\mathcal{H}(\alpha)$ consists of a single element represented by a map $H(\alpha): S^q\to F_1(S^p)$, and we have
\[
{\rm{cat}\hskip1pt}(X) = \left\{\begin{array}{ll} 1 & \mbox{if }H(\alpha)=0, \\ 2 & \mbox{if }H(\alpha)\neq 0.
\end{array}\right.
\]
Methods for computing $H(\alpha)$ for various $\alpha$ are given in \cite[Chapter 6]{CLOT}.
{As for ${\rm{TC}\hskip1pt}(X)$, we start by noticing that $1\le {\rm{TC}\hskip1pt}(X)\le 2$ whenever ${\rm{cat}\hskip1pt}(X)=1$ (e.g.~if $q>p\ge2$ and $H(\alpha)=0$). Actually,} since $X$ cannot be homotopy equivalent to an odd-dimensional sphere, the main result of \cite{GLO} implies that {in fact} ${\rm{TC}\hskip1pt}(X)=2$. This happens whenever $\alpha$ is null-homotopic, or more generally a suspension. Therefore in what follows we will assume we are outside of the stable range, i.e.~we assume $q\ge 2p-1$. Likewise, proof details will be limited to the case ${\rm{cat}\hskip1pt}(X)=2$, where $2\le{\rm{TC}\hskip1pt}(X)\leq4$.
When {$q=2p-1$} it is possible to give a complete computation of ${\rm{TC}\hskip1pt}(X)$ using mainly cohomological arguments. We first address the case $p=1$.
\begin{thm}\label{grados}
Let $X$ be the mapping cone of a map $\alpha: S^1\to S^1$, whose degree we denote by $d_\alpha$. Then
\[
{\rm{TC}\hskip1pt}(X)=\left\{\begin{array}{ll} 2 & \mbox{if }d_\alpha=0, \\ 0 & \mbox{if }d_\alpha=\pm 1, \\ 3 & \mbox{if }d_\alpha=\pm 2, \\ 4 & \mbox{otherwise.} \end{array} \right.
\]
\end{thm}
\begin{proof}
The first two cases have been dealt with above. If $d_\alpha=\pm 2$ then $X\simeq {\mathbb{R}} P^2$, and the result follows from \cite{FTY} since the immersion dimension of ${\mathbb{R}} P^2$ is $3$.
The remaining {case} can be dealt with using ${\rm{TC}\hskip1pt}$-weights of cohomology classes (see \cite{FG}, \cite[Section 4.5]{FarInv}, \cite[Section 4]{FarCosta}). Here $X\simeq M({\mathbb{Z}}/k,1)$ is a mod $k$ Moore space, where $k=|d_\alpha|>2$. Let $x\in H^1(X;{\mathbb{Z}}/k)\cong{\mathbb{Z}}/k$ and $y\in H^2(X;{\mathbb{Z}}/k)\cong{\mathbb{Z}}/k$ be generators. Then $y = \beta(x)$ where $\beta$ is the mod $k$ Bockstein operator. The class
\[
\overline{y} = 1\times y - y\times 1 \in H^2(X\times X;{\mathbb{Z}}/k)
\]
therefore has ${\rm{TC}\hskip1pt}$-weight $2$, by \cite[Theorem 6]{FG}. An easy calculation gives
\[
0\neq \overline{y}^2 = -2 y\times y \in H^4(X\times X;{\mathbb{Z}}/k)\cong{\mathbb{Z}}/k,
\]
and so ${\rm{TC}\hskip1pt}(X)\ge 2+2=4$ by \cite[Proposition 2]{FG}. Since ${\rm{TC}\hskip1pt}(X)\le 2\,{\rm{cat}\hskip1pt}(X)\le 4$, this completes the proof.
\end{proof}
In the case of a map $\alpha: S^{2p-1}\to S^p$ with $p\ge 2$, the Berstein--Hilton--Hopf invariant $H(\alpha)\in \pi_{2p-1}(F_1(S^p))$ agrees with the classical Hopf invariant $h(\alpha)\in{\mathbb{Z}}$ up to a sign. To be more explicit, projection onto the bottom cell of
\[
F_1(S^p) = \Omega S^p\circledast \Omega S^p \simeq \Sigma(\Omega S^p\wedge \Omega S^p)\simeq S^{2p-1}\vee S^{3p-2}\vee\cdots
\]
induces an isomorphism
\[
\pi_{2p-1}(F_1(S^p))\cong \pi_{2p-1}(S^{2p-1})\cong {\mathbb{Z}},
\]
which sends $H(\alpha)$ to $\pm h(\alpha)$, see \cite[Section 6.2]{CLOT}.
\begin{thm}\label{ClassicHopf}
Let $X$ be the mapping cone of a map $\alpha: S^{2p-1}\to S^p$ with {$p\geq2$ and} classical Hopf invariant $h(\alpha)\in{\mathbb{Z}}$. Then
\[
{\rm{TC}\hskip1pt}(X) = \left\{\begin{array}{ll} 2 & \mbox{if }h(\alpha)=0, \\ 4 & \mbox{if }h(\alpha)\neq 0. \end{array}\right.
\]
\end{thm}
\begin{proof}
If $h(\alpha)=0$ then $H(\alpha)=0$ and ${\rm{TC}\hskip1pt}(X)=2$, as already noted above.
Let $u\in H^p(X;{\mathbb{Z}})$ and $v\in H^{2p}(X;{\mathbb{Z}})$ be generators of the integral cohomology groups of $X$. The cup product structure is given by $uv=v^2=0$ and $u^2 = h(\alpha) v$. Therefore, if $h(\alpha)\neq 0$ {(so $p$ is even),} we calculate that
\[
0\neq (1\times u - u\times 1)^4 = 6\, h(\alpha)^2\, v\times v \in H^{4p}(X\times X;{\mathbb{Z}})\cong {\mathbb{Z}}.
\]
Hence ${\rm{TC}\hskip1pt}(X)\ge 4$ by the usual zero-divisors cup-length lower bound. Since ${\rm{TC}\hskip1pt}(X)\le 4$ this completes the proof.
\end{proof}
We now suppose that $\alpha: S^q\to S^p$ is a map of spheres with
$2p-1< q \le 3p-3$ {(so $q-1>p\geq3$).} Such a map is said to be in the \emph{metastable range}. In this range, projection onto the bottom cell induces an isomorphism
$
\pi_q\big(F_1(S^p)\big) \cong \pi_q(S^{2p-1}).
$
The image of $H(\alpha)$ under this isomorphism is a map $H_0(\alpha):S^q\to S^{2p-1}$. Note that this map is in the stable range, and is therefore a suspension. {As noted above,} we will assume $H_0(\alpha)\neq 0$, so that ${\rm{cat}\hskip1pt}(X)=2$ and $2\le {\rm{TC}\hskip1pt}(X)\le 4$. Using Corollary \ref{forTC2cell}, we give conditions under which ${\rm{TC}\hskip1pt}(X)\ge 3$ holds, {and conditions under which} ${\rm{TC}\hskip1pt}(X)\le 3$ holds. In many cases we will be able to conclude that ${\rm{TC}\hskip1pt}(X)=3$.
\begin{thm} \label{le3}
Let $X=S^p\cup_\alpha e^{q+1}$, where $\alpha: S^q\to S^p$ is in the metastable range $2p-1< q\le 3p-3$ and $H_0(\alpha)\neq 0$.
Then ${\rm{TC}\hskip1pt}(X)\le 3$ in any of the following cases:
\begin{enumerate}
\item {$2 H_0(\alpha)\circledast H_0(\alpha)=0$ (e.g.~when $2H_0(\alpha)=0$);}
\item {$6 H_0(\alpha)\circledast H_0(\alpha)=0$ (e.g.~when $\,6 H_0(\alpha)= 0$) and $p$ is even;}
\item $q$ {is} even.
\end{enumerate}
\end{thm}
\begin{proof}
By Corollary \ref{forTC2cell}(b) it suffices to check that the composition
\[
\xymatrix{
S^q \circledast S^q \ar[rr]^-{H(\alpha)\circledast H(\alpha)} && F_1(S^p)\circledast F_1(S^p) \ar[rr]^-{\Phi_{1,1}^{\Omega S^p, \Omega S^p}} && F_3(S^p\times S^p) \ar[r]^{\bar{\chi}_3} & F_3(S^p)
}
\]
is trivial in each of the stated cases. Up to homotopy, the first map factors as
\[
\xymatrix{
S^q \circledast S^q \ar[rr]^-{H(\alpha)\circledast H(\alpha)} \ar[rrd]_{H_0(\alpha)\circledast H_0(\alpha)\hspace{7mm}} && F_1(S^p)\circledast F_1(S^p) \\
&& S^{2p-1}\circledast S^{2p-1} \ar@{^{(}->}[u] .
}
\]
By Example \ref{degree-2cells}(a) below, the degree of the composition $\bar{\chi}_3\circ\Phi_{1,1}^{\Omega S^p, \Omega S^p}$ on the bottom cell $S^{4p-1}$ is $\pm(4+2(-1)^p)$. The claim follows in the first two cases. For the third case, since $H_0(\alpha):S^q\to S^{2p-1}$ is in the stable stem $\pi^S_{q-2p+1}$, the join product (which is the suspension of the smash product) is anti-commutative. Therefore $H_0(\alpha)\circledast H_0(\alpha) = (-1)^{q-2p+1}H_0(\alpha)\circledast H_0(\alpha)$ is of order $2$ if $q$ is even. This completes the proof.
\end{proof}
\begin{thm} \label{ge3}
Let $X=S^p\cup_\alpha e^{q+1}$, where $\alpha: S^q\to S^p$ is in the metastable range $2p-1< q\le 3p-3$ and $H_0(\alpha)\neq 0$.
Then ${\rm{TC}\hskip1pt}(X)\ge 3$ in any of the following cases:
\begin{enumerate}
\item $p$ {is} odd;
\item $p$ {is} even and $3H_0(\alpha)\neq 0$.
\end{enumerate}
\end{thm}
\begin{proof}
Since ${\rm{TC}\hskip1pt}(X)\ge {\rm{TC}\hskip1pt}_{X\times S^p}(X)$, it suffices by Corollary \ref{forTC2cell}(c) to check that the composition
\[
\xymatrix{
S^q \circledast S^{p-1} \ar[rr]^-{H(\alpha)\circledast H^0(S^p) } && F_1(S^p)\circledast F_0(S^p) \ar[rr]^-{\Phi_{1,0}^{\Omega S^p, \Omega S^p}} && F_2(S^p\times S^p) \ar[r]^{\hspace{3.5mm}\bar{\chi}_2} & F_2(S^p)
}
\]
is essential. {As noted in Example~\ref{cathopfsphere}, the Hopf invariant}
\[
H^0(S^p):S^{p-1}\to \Omega S^p \simeq {S^{p-1}\cup e^{2p-2}\cup \cdots}
\]
is homotopic to the inclusion of the bottom cell. Therefore the first map in the composition factors as
\[
\xymatrix{
S^q \circledast S^{p-1} \ar[rr]^-{H(\alpha)\circledast H^0(S^p)\hspace{1.5mm} } \ar[rrd]_{H_0(\alpha)\circledast\mathrm{Id}_{S^{p-1}}\hspace{3mm}} && F_1(S^p)\circledast F_0(S^p) \\
&& S^{2p-1}\circledast S^{p-1} \ar@{^{(}->}[u].
}
\]
The diagonal arrow can be identified with the $p$-th {reduced} suspension ${\widetilde{\Sigma}}^p H_0(\alpha): S^{q+p}\to S^{3p-1}$, which is essential since $H_0(\alpha)$ is stable.
By Example \ref{degree-2cells}(b) below, the degree of the composition $\bar{\chi}_2\circ \Phi_{1,0}^{\Omega S^p, \Omega S^p}$ on the bottom cell $S^{3p-1}$ is $\pm(2+(-1)^p)$. This yields the result.
\end{proof}
\begin{ex}
Let $X=S^3\cup_\alpha e^7$ where $\alpha: S^6\to S^3$ is the Blakers--Massey element, a generator of $\pi_6(S^3)={\mathbb{Z}}/12$. Then $0\neq H(\alpha)=H_0(\alpha)\in \pi_6(S^5)={\mathbb{Z}}/2$, and the above results imply that ${\rm{TC}\hskip1pt}(X)=3$. We remark that the lower bound ${\rm{TC}\hskip1pt}(X)\ge 3$ was obtained in \cite[Proposition 30]{Weaksecat} using the weak sectional category, and the upper bound ${\rm{TC}\hskip1pt}(X)\le 3$ was obtained in \cite[Example 6]{GC-V} using the fact that $X$ is the $9$-skeleton of the group $Sp(2)$.
\end{ex}
More generally, for $q\in\{2p,2p+1\}$ with $q\leq 3p-3$---i.e.~the first two cases of the metastable range---we have:
\begin{cor}\label{partiuno}
Let $\delta\in\{0,1\}$ and set $p\geq3+\delta$. Then
$${\rm{TC}\hskip1pt}(S^p\cup_\alpha e^{2p+\delta+1})=\begin{cases}
2, & \mbox{ if }H_0(\alpha)=0;\\
3, & \mbox{ if }H_0(\alpha)\neq0.
\end{cases}$$
\end{cor}
\begin{proof}
Multiplication by 3 yields an isomorphism in $\pi_{2p+\delta}(S^{2p-1})={\mathbb{Z}}/2$.
\end{proof}
{Also worth mentioning is:}
\begin{cor}\label{partidos}
{For $p$ odd and $q$ even with $2p-1<q\leq3p-3$,}
$${{\rm{TC}\hskip1pt}(S^p\cup_\alpha e^{2p+\delta+1})=}\begin{cases}
{2,} & {\mbox{ if }H_0(\alpha)=0;}\\
{3,} & {\mbox{ if }H_0(\alpha)\neq0.}
\end{cases}$$
\end{cor}
When {$q>3p-3\geq3$} and we are outside of the metastable range, it is still possible to draw conclusions about ${\rm{TC}\hskip1pt}(X)$. In particular, the conclusion of Theorem \ref{le3} still holds under the weaker assumption that $H(\alpha)=H_0(\alpha)$ {(and the additional requirement $q<4p-3$ in the case of \ref{le3}(3), which assures the stability of $H_0(\alpha)$ needed in the proof of Theorem~\ref{le3}).} To obtain maps $\alpha: S^q\to S^p$ satisfying {$H(\alpha)=H_0(\alpha)$,} observe that if $\alpha = \gamma\circ\beta$ is a composition
\[
\xymatrix{ S^q\ar[r]^\beta & S^r \ar[r]^\gamma & S^p},
\]
then $H(\alpha) = H(\gamma)\circ \beta$ whenever $H(\beta)=0$ \cite[Proposition 6.18(2)]{CLOT}. Therefore if $H(\beta)=0$ and $\gamma$ is in the stable or metastable range, then $H(\alpha) = H_0(\alpha) = H_0(\gamma)\circ \beta$.
\begin{exs}\label{contieneposiblegap}
\begin{enumerate}[(a)]
\item {The case $q=2p$ fails to lie in the metastable range only for $p=2$ (so $q=4$). The generator of $\pi_4(S^2)=\mathbb{Z}/2$ is represented by the composition $\alpha = \eta\circ \Sigma\eta: S^4\to S^2$ where $\eta: S^3\to S^2$ is the Hopf map.} Then $H(\alpha)=H_0(\alpha)$ is the nonzero element $\Sigma\eta\in \pi_4(S^3)={\mathbb{Z}}/2$. For $X=S^2\cup_\alpha e^5$, we conclude as in Theorem \ref{le3} that ${\rm{TC}\hskip1pt}(X)\le 3$. {In this case, however, we cannot use Theorem \ref{ge3} to get in fact that ${\rm{TC}\hskip1pt}(X)=3$, as the relevant Hopf set fails to be a singleton.}
\item Let $X= S^2\cup_\alpha e^{10}$ where $\alpha = \eta\circ\beta$ and $\beta\in \pi_9(S^3)={\mathbb{Z}}/3$ is the generator. This is one of the spaces considered by Iwase in \cite{Iwase}. Here we have $H(\alpha)=H_0(\alpha) = \beta$, and we can conclude as in Theorem \ref{le3} that ${\rm{TC}\hskip1pt}(X)\le 3$. We cannot {use} Theorem \ref{ge3} {to conclude} that ${\rm{TC}\hskip1pt}(X)\ge 3$ {not only because of the situation noted in item (a) above, but} since {this time} the map $H_0(\alpha)$ is no longer stable (this being the reason why $X$ is a counter-example to Ganea's conjecture). In fact $\Sigma^2\beta = 0$ which, {by the argument in the proof of Theorem~\ref{ge3},} shows that ${\rm{TC}\hskip1pt}_{X\times S^2}(X)\le 2$ {(the latter is in fact an equality since $2={\rm{cat}\hskip1pt}(X)={\rm{TC}\hskip1pt}_{X\times\ast}(X)\leq{\rm{TC}\hskip1pt}_{X\times S^2}(X)$).}
\end{enumerate}
\end{exs}
\section{Application: Ganea's condition for topological complexity}\label{secapgc4tc}
The analogue of Ganea's conjecture for topological complexity asks whether, for any finite complex $X$ and $k\ge 1$, we have
\begin{equation}\label{GaneaTC}
{\rm{TC}\hskip1pt}(X\times S^k) = {\rm{TC}\hskip1pt}(X) + {\rm{TC}\hskip1pt}(S^k) = \left\{\begin{array}{ll} {\rm{TC}\hskip1pt}(X) + 1 & \mbox{if $k$ odd,}\\ {\rm{TC}\hskip1pt}(X) + 2 &\mbox{if $k$ even.} \end{array}\right.
\end{equation}
By \cite[Corollary 1.7]{JMP}, the answer is known to be positive when $k\ge 2$ and $X$ is a simply-connected, rational, formal complex of finite type. The answer is also known to be positive, without the formality assumption, if TC is replaced by either of the related rational invariants ${\rm mtc}$ (\cite[Theorem~1.6]{JMP}) or MTC (\cite[Theorem~12]{carrasquel}), the former introduced in~\cite{JMP} and the latter in~\cite{FGKV}. {Theorem~\ref{noGaneaTC} below gives} a counter-example to equation (\ref{GaneaTC}) for all $k\ge 2$ even.
We first observe that the topological complexity of a product of spaces can be described as the sectional category of a product of fibrations.
\begin{lem}\label{TCproduct}
Let $X$ and $Y$ be spaces. Then ${\rm{TC}\hskip1pt}(X\times Y) = {\rm{secat}\hskip1pt}(\pi_X\times \pi_Y)$.
\end{lem}
\begin{proof}
This follows from the commutative diagram
\[
\xymatrix{
(X\times Y)^I \ar[r]^{T} \ar[d]^{\pi_{X\times Y}} & X^I\times Y^I \ar[d]^{\pi_X\times \pi_Y} \\
X\times Y \times X \times Y \ar[r]^{\tau} & X\times X \times Y \times Y.
}
\]
Here $T$ sends a path in $X\times Y$ to the pair of paths consisting of its projections onto $X$ and $Y$, and $\tau$ transposes the middle two factors. Both of these maps are homeomorphisms, so it follows that ${\rm{secat}\hskip1pt}(\pi_{X\times Y}) = {\rm{secat}\hskip1pt}(\pi_X\times \pi_Y)$.
\end{proof}
Next we investigate the ${\rm{TC}\hskip1pt}$-Hopf invariants for spheres. Consider the cofibration sequence
\[
\xymatrix{
S^{2k-1} \ar[r]^{\beta} & S^k\vee S^k \ar[r] & S^k\times S^k,
}
\]
where the attaching map $\beta=[\iota_1,\iota_2]:S^{2k-1}\to S^k\vee S^k$ is the Whitehead product of the inclusions of the two wedge factors. Note that the subcomplex $S^k\vee S^k\subseteq S^k\times S^k$ satisfies ${\rm{TC}\hskip1pt}_{S^k\vee S^k}(S^k)= 1$ by Proposition \ref{secatprops}. If $k\ge2$ then by Proposition \ref{uniqueness} the Hopf set
\[
\mathcal{H}^1_{\beta}(\pi_{S^k})\subseteq \pi_{2k-1}(F_1(S^k))
\]
consists of a single element represented by a map
\[
H^1_\beta(\pi_{S^k}): S^{2k-1}\to F_1(S^k)\simeq S^{2k-1}\vee S^{3k-2}\vee\cdots.
\]
This map is determined up to homotopy by the degree $d_k\in{\mathbb{Z}}$ of its projection on to the bottom cell $S^{2k-1}$. The following lemma gives a purely homotopy-theoretic proof of \cite[Theorem 8]{Far}.
\begin{lem}\label{TCHopfspheres}
We have $\pm d_k = 1 + (-1)^k$.
\end{lem}
\begin{proof}
Arguing as in Corollary \ref{forTC2cell}(c), the Hopf invariant $H^1_\beta(\pi_{S^k})$ is given by the composition
\[
\resizebox{1\textwidth}{!}{\xymatrix{
S^{k-1}\circledast S^{k-1} \ar[rr]^-{H^0(S^k)\circledast H^0(S^k)} && F_0(S^k)\circledast F_0(S^k) \ar[rr]^-{\Phi^{\Omega S^k,\Omega S^k}_{0,0}} && F_1(S^k\times S^k) \ar[r]^-{\bar\chi_1} & F_1(S^k)
}}
\]
where $H^0(S^k):S^{k-1}\to \Omega S^k$ can be identified with inclusion of the bottom cell. By Example \ref{degree-2cells}(c) the degree of the map $\bar\chi_1\circ \Phi^{\Omega S^k,\Omega S^k}_{0,0}$ on the bottom cell is $\pm(1+(-1)^k)$, and the result follows.
\end{proof}
\begin{thm}\label{noGaneaTC}
Let $Y$ be the stunted real projective space ${\mathbb{R}} P^6/{\mathbb{R}} P^2$, and let $X=Y\vee Y$. Then {${\rm{TC}\hskip1pt}(X)=4$ and, for all $k\ge2$ even, ${\rm{TC}\hskip1pt}(X\times S^{k}) = 5$.}
\end{thm}
\begin{proof}
We first show that ${\rm{TC}\hskip1pt}(X)=4$. Since $X$ is $6$-dimensional and $2$-connected, the standard dimension/connectivity upper bound gives
\[
{\rm{TC}\hskip1pt}(X) \le \frac{2\cdot 6}{3} = 4.
\]
One checks using the cofibration ${\mathbb{R}} P^2\to {\mathbb{R}} P^6 \to Y$ that {there is a ring monomorphism} $${\mathbb{Z}}/2[u]/(u^3)\hookrightarrow H^*(Y;{\mathbb{Z}}/2)$$ where $|u|=3$. Therefore the cup-length of $H^*(Y\times Y;{\mathbb{Z}}/2)$ is $4$, and we have
\[
4\le {\rm{cat}\hskip1pt}(Y\times Y)\le {\rm{TC}\hskip1pt}(Y\vee Y) = {\rm{TC}\hskip1pt}(X)
\]
on applying the result of \cite[Theorem 3.6]{Dran} or \cite[Corollary 2.9]{GLO2}. (Alternatively, the inequality ${\rm{TC}\hskip1pt}(X)\ge 4$ follows directly from a zero-divisors calculation in mod $2$ cohomology.) This completes the proof of the first claim.
The lower bound ${\rm{TC}\hskip1pt}(X\times S^k)\ge5$ is a quick calculation using zero-divisors in mod $2$ cohomology, and is omitted. We prove that ${\rm{TC}\hskip1pt}(X\times S^k)\le 5$ using Hopf invariants.
By Lemma \ref{TCproduct} we have ${\rm{TC}\hskip1pt}(X\times S^{k}) = {\rm{secat}\hskip1pt}(\pi_X\times \pi_{S^k})$. Let $K$ denote the $11$-skeleton of the standard product cell structure on $X\times X$. Then $X\times X \simeq K\cup_\alpha CS$ where $S = \vee_{i=1}^4 S^{11}$ is a wedge of spheres and $\alpha: S\to K$ is the attaching map of the four $12$-cells. Similarly we have $S^k\times S^k = L\cup_{\beta} e^{2k}$, where $L = S^k\vee S^k$ and $\beta: S^{2k-1}\to L$ is the attaching map of the top cell.
By Proposition \ref{secatprops} we have ${\rm{TC}\hskip1pt}_K(X)\le \frac{11}{3}<4$, and since relative topological complexity can increase by at most one on attaching the top cells, ${\rm{TC}\hskip1pt}_K(X)=3$. Let $\phi: K\to G_3^{{\rm{TC}\hskip1pt}}(X)$ be any (pointed) lift of the inclusion $K\hookrightarrow X\times X$ through $g_3^{{\rm{TC}\hskip1pt}}(X)$. The associated Hopf invariant $H^3_{\phi,\alpha}(\pi_X): S\to F_3(X)$ is non-trivial of order $2$, since
\begin{align*}
[S,F_3(X)] & \;\cong\; \bigoplus \pi_{11}(F_3(X))
\;\cong\; \bigoplus H_{11}(F_3(X))
\;\cong\; \bigoplus H_{8}((\Omega X)^{\wedge 4})\\
& \;\cong\; \bigoplus H_2(\Omega X)^{\otimes 4}
\;\cong\; \bigoplus H_3(X)^{\otimes 4}
\;\cong\; \bigoplus {\mathbb{Z}}/2.
\end{align*}
By Theorem \ref{Hopfproducts} we have
$
{\rm{secat}\hskip1pt}_{X\times X\times L \cup K\times S^{k}\times S^{k}}(\pi_X\times \pi_{S^k}) \le 5
$
with an element of the Hopf set $\mathcal{H}_{W}^5(\pi_X\times \pi_{S^k})$ given by the composition
\[
\xymatrix{
S\circledast S^{2k-1} \ar[rrr]^-{H^3_{\phi,\alpha}(\pi_X)\circledast H^1_\beta(\pi_{S^k})} &&& F_3(X)\circledast F_1(S^{k}) \ar[rr]^-{\Phi^{\Omega X,\Omega S^k}_{3,1}} && F_5(X\times S^{k}).
}
\]
The first map in this composition is null-homotopic. To see this, note that by Lemma \ref{TCHopfspheres} it factors as
\[
\xymatrix{
S\circledast S^{2k-1} \ar[rrr]^-{H^3_{\phi,\alpha}(\pi_X)\circledast H^1_\beta(\pi_{S^k})} \ar[rrrd]_{H^3_{\phi,\alpha}(\pi_X)\circledast (\pm 2)} &&& F_3(X)\circledast F_1(S^{k})\\
&&& F_3(X)\circledast S^{2k-1} \ar@{^{(}->}[u]
}
\]
where the diagonal map is the join of a map of order $2$ with a map of even degree. Therefore $0\in \mathcal{H}_{W}^5(\pi_X\times \pi_{S^k})$ and ${\rm{TC}\hskip1pt}(X\times S^k)\le 5$, as claimed.
\end{proof}
\section{Topological shuffle maps}
\label{secsiete}
Let $p:{\mathcal A}\to X$ and $q:{\mathcal B}\to Y$ be fibrations with respective fibres $A$ and $B$. In this section we construct, for any non negative integers $n$ and $m$, a commutative diagram of the following form:
$$
\xymatrix{
J^{n+1}(A)\times J^{m}(B)\cup J^{n}(A)\times J^{m+1}(B)\ar[rr]^-{{\Psi_{n,m}^{A,B}}} \ar[d]&& J^{n+m+1}(A\times B)\ar[d] \\
J^{n+1}_X({\mathcal A})\times J_Y^{m}({\mathcal B})\cup J_X^{n}({\mathcal A})\times J_Y^{m+1}({\mathcal B})\ar[rr]^-{{\Psi_{n,m}^{{\mathcal A},{\mathcal B}}}} \ar[d]&& J_{X\times Y}^{n+m+1}({\mathcal A}\times {\mathcal B})\ar[d] \\
X\times Y \ar@{=}[rr] &&X\times Y.}
$$
The construction of $\Psi_{n,m}^{{\mathcal A},{\mathcal B}}$ is given in Subsection \ref{shufflemap}
based on the standard decomposition of a product of simplices into simplices, which is recalled in Subsection \ref{standarddecomp}.
{The needed diagram~(\ref{shufflewjoins}) in Subsection~\ref{HopfProducts} is then obtained by setting $\Phi_{n,m}^{A,B}:=\Psi_{n,m}^{A,B}\circ \xi$ where $$\xi: J^ n(A)\circledast J^m(B) \to J^{n+1}(A)\times J^{m}(B)\cup J^{n}(A)\times J^{m+1}(B)$$
is induced by the maps $\kappa_n\colon CJ^n(A)\to J^{n+1}(A)$ and $\kappa_n\colon CJ^m(B)\to J^{m+1}(B)$ introduced in Remark~\ref{fibredjoins}.}
{In Subsection~\ref{homotopycharacterization} we characterize the homotopy type of $\Phi_{n,m}^{A,B}$ in terms of $(n,m)$ shuffles, and}
in Subsection \ref{homologyComp} we compute the morphism induced by $\Phi_{n,m}^{A,B}$ in homology.
\subsection{Standard decomposition of the product $\Delta^n\times \Delta^m$}
\label{standarddecomp}
We recall here the standard subdivision of the product of two simplices into simplices as described in \cite[p.68]{EilenbergSteenrod}, giving a different presentation which will be more convenient for our future
computations in homology (see also \cite[pp. 277-278]{Hatcher}).
Let $n$ and $m$ two non-negative integers. Let us denote by $\mathcal{S}_{n+m}$ the set of permutations of $\{1,\ldots, n+m\}$ and by $\mathcal{S}_{n,m}\subset \mathcal{S}_{n+m}$ the subset of $(n,m)$ shuffles. A shuffle $\theta\in \mathcal{S}_{n,m}$ will be indicated by
$\theta=(\theta(1)\cdots \theta(n)\,||\,\theta(n+1) \cdots \theta(n+m))$
{(so $1\leq\theta(1)<\cdots<\theta(n)\leq n+m$, $1\leq\theta(n+1)<\cdots<\theta(n+m)\leq n+m$, and $\{\theta(1),\ldots,\theta(n)\}\cap\{\theta(n+1),\ldots,\theta(n+m)\}=\varnothing$)} and can be represented in a $n\times m$ rectangular grid with vertices $(i,j)$ ($0\leq i\leq n$, $0\leq j\leq m$) by a continuous path going from $(0,0)$ to $(n,m)$ formed by $n$ horizontal edges and $m$ vertical edges. More precisely, the $k$-th edge of the edgepath $\theta$ is horizontal if $1\leq \theta^{-1}(k)\leq n$ and vertical if $n+1\leq \theta^{-1}(k)\leq n+m$. For instance, the edgepath representing the $(5,3)$ shuffle $(1\,2\,4\,6\,7\,||\,3\,5\,8)$ is shown in Figure 1 below. The signature of the shuffle $\theta$ is denoted by $(-1)^{|\theta|}$ where $|\theta|$ can be interpreted as the number of squares in the grid lying below the path $\theta$.
\vspace{-.45cm
\psset{xunit=1.0cm,yunit=1.0cm}
\begin{pspicture}(-3,-1)(7.5,4)
\psgrid[subgriddiv=0,gridlabels=0,gridcolor=lightgray](0,0)(0,0)(5,3)
\psline(0,0)(1,0)
\psline(1,0)(2,0)
\psline(2,0)(2,1)
\psline(2,1)(3,1)
\psline(3,1)(3,2)
\psline(3,2)(5,2)
\psline(5,2)(5,3)
\begin{scriptsize}
\psdots[dotstyle=*](0,0)
\psdots[dotstyle=*](1,0)
\psdots[dotstyle=*](2,0)
\psdots[dotstyle=*](2,1)
\psdots[dotstyle=*](3,1)
\psdots[dotstyle=*](3,2)
\psdots[dotstyle=*](4,2)
\psdots[dotstyle=*](5,2)
\psdots[dotstyle=*](5,3)
\rput[bl](-1,-0.2){$(0,0)$}
\rput[bl](5.5,3){$(n,m)$}
\rput[bl](-.7,-.8){\normalsize Figure 1. The $(5,3)$ shuffle $(1\,2\,4\,6\,7\,||\,3\,5\,8)$}
\end{scriptsize}
\end{pspicture}
\vspace{2mm}
The product of the standard simplices $\Delta^n=[e_0,\ldots ,e_{n}]$ and $\Delta^m=[e'_0,\ldots ,e'_{m}]$
(where $(e_i)_{0\leq i\leq n}$ and $(e'_j)_{0\leq j\leq m}$ are the canonical bases of ${\mathbb{R}}^ {n+1}$ and ${\mathbb{R}}^ {m+1}$)
is the union of $\binom{n+m}{n}$ simplices of dimension $(n+m)$, each of which is determined by a $(n,m)$ shuffle. Explicitly, let $\theta$ be a $(n,m)$ shuffle.
Then the (ordered) $(n+m)$-simplex $\Delta_{\theta}\subset \Delta^ n\times \Delta^m$ associated to $\theta$ is given by
$\Delta_{\theta}=[v_1,\ldots , v_{n+m+1}]$
with $v_k=(e_{i_k},e'_{j_k})$ where $(i_k,j_k)$ is the $k$-th vertex of the edgepath $\theta$.
Observe that $v_1=(e_0,e_0')$ and $v_{n+m+1}=(e_{n},e_{m}')$ for any $\theta$.
Taken together, the simplices $\Delta_{\theta}$ form a chain
\begin{equation}\label{lachain}
\sum\limits_{\theta\in {\mathcal S}_{n,m}}(-1)^{|\theta|}\Delta_{\theta}
\end{equation}
\noindent which represents a generator $\iota_{n,m}$ of
\[
H_{n+m}(\Delta^n\times \Delta^m, \partial(\Delta^n \times \Delta^m);{\mathbb{Z}})\cong\tilde{H}_{n+m}((\Delta^n\times \Delta^m)/ \partial(\Delta^n \times \Delta^m);{\mathbb{Z}})\cong{\mathbb{Z}}.
\]
We denote by $\psi_{\theta}:\Delta_{\theta}\to \Delta^{n+m}$ the affine map that
sends the $k$-th vertex of $\Delta_{\theta}$ to the $k$-th vertex of $\Delta^{n+m}$. This gives a homeomorphism from $(\Delta_{\theta},\partial\Delta_{\theta})$ to $(\Delta^{n+m},\partial\Delta^{n+m})$ which induces a map $$\bar{\psi}_{\theta}:\Delta_{\theta}/\partial\Delta_{\theta}\to\Delta^{n+m}/\partial\Delta^{n+m}$$
of degree $1$.
When $\theta$ runs over the set ${\mathcal S}_{n,m}$ of $(n,m)$ shuffles, the maps $\psi_{\theta}$ glue together. In order to see that, it suffices to check that $\psi_{\theta}$ and $\psi_{\theta'}$ agree on $\Delta_{\theta}\cap \Delta_{\theta'}$ when this intersection is a simplex of dimension exactly $n+m-1$. This happens when the edgepaths $\theta$ and $\theta'$ differ in only one vertex, say the $k$-th vertex. In that case $2\leq k\leq n+m$, {and $\Delta_{\theta}\cap \Delta_{\theta'}$ is the $(n+m-1)$-simplex determined by the (ordered) common vertices in the edge paths of $\theta$ and $\theta'$. But those vertices are sent preserving order, both by $\psi_{\theta}$ and $\psi_{\theta'}$, into the ordered vertices of the face of $\Delta^{n+m}$ opposite to its $k$-th vertex. (In the situation above, $|\theta|$ and $|\theta'|$ differ by one, which explains why~(\ref{lachain}) is a relative cycle, obviously representing a generator $\iota_{n,m}$.)}
As a consequence, the maps $\psi_{\theta}$ produce
a map
\begin{equation}\label{primerfinm}
\psi_{n,m}:\Delta^n \times \Delta^m \to \Delta^{n+m}
\end{equation}
{which} sends the boundary of each $\Delta_{\theta}$ {and, in particular,} the boundary of $\Delta^n\times \Delta^m$ to the boundary of $\Delta^{n+m}$.
When passing to the quotient of $\Delta^ n\times \Delta^m$ first by the boundary $\partial(\Delta^n \times \Delta^m)$ and secondly by the boundaries of the simplices $\Delta_{\theta}$ we get the \textit{shuffle pinch map}
$$\nu_{n,m}:(\Delta^ n\times \Delta^m)/\partial(\Delta^n \times \Delta^m)\to \bigvee\limits_{\theta\in {\mathcal S}_{n,m}}\Delta_{\theta}/\partial \Delta_{\theta}$$
which gives the following morphism in homology:
\begin{equation}\label{inducidoabajo}
\iota_{n,m}\mapsto \bigoplus\limits_{\theta\in {\mathcal S}_{n,m}}(-1)^{|\theta|}\iota_{n+m}^{\theta}.
\end{equation}
Here, as before, $\iota_{n,m}$ is the generator of $H_{n+m}(\Delta^n\times \Delta^m/ \partial(\Delta^n \times \Delta^m);{\mathbb{Z}})$ corresponding to the chain~(\ref{lachain}), and $\iota_{n+m}^{\theta}\in H_{n+m}(\Delta_{\theta}/\partial \Delta_{\theta};{\mathbb{Z}})={\mathbb{Z}}$ is the generator corresponding to the ordered simplex $\Delta_{\theta}$.
The map $\psi_{n,m}:\Delta^n \times \Delta^m \to \Delta^{n+m}$ then fits into the commutative diagram
\begin{equation}\label{diagramPinchmap}
\xymatrix{
\Delta^n\times \Delta^m \ar[rr]^{\psi_{n,m}} \ar[d]&& \Delta^{m+n} \ar[d]\\
\disfrac{\Delta^n\times \Delta^m}{\partial (\Delta^n\times \Delta^m)}\ar[r]_{\nu_{n,m}} &\bigvee\limits_{\theta\in {\mathcal S}_{n,m}}\Delta_{\theta}/\partial \Delta_{\theta}\ar[r]^-{(\bar{\psi}_{\theta})}&\disfrac{\Delta^{n+m}}{\partial \Delta^{n+m}}.
}
\end{equation}
Since the maps $\bar{\psi}_{\theta}$ are of degree $1$, the bottom line induces in homology the morphism $$\iota_{n,m}\mapsto \sum\limits_{\theta\in {\mathcal S}_{n,m}}(-1)^{|\theta|}\iota_{n+m}
$$ where $\iota_{n+m}$ corresponds to the standard generator of $H_{n+m}(\Delta^{n+m}/\partial \Delta^{n+m};{\mathbb{Z}})$.
\subsection{The maps $\Psi_{n,m}^{{\mathcal A},{\mathcal B}}$}
\label{shufflemap}
For any integer $k$ we shall denote by $T_k$ the map
$$\begin{array}{rcl}
T_k: {\mathcal A}^k_X\times {\mathcal B}^k_Y & \to & ({\mathcal A}\times {\mathcal B})^k_{X\times Y}\\
\big((a_1,\ldots,a_k), (b_1,\ldots,b_k)\big)&\mapsto & \big((a_1,b_1),\ldots,(a_k,b_k)\big),
\end{array}$$
and by $\Delta_k^{\mathcal A}:{\mathcal A}\to {\mathcal A}^k_X$ the $k$-th diagonal, $a\mapsto (a,a,\ldots,a)$. In addition, for an $(n,m)$ shuffle $\theta$, we shall define two sequences of integers $\alpha_0,\ldots ,\alpha_n$ and $\beta_0,\ldots ,\beta_m$.
In the grid illustrated in Figure~1, $\alpha_i$ will correspond to the number of vertices of the edgepath $\theta$ belonging to the column $i$, while $\beta_j$ will correspond to the number of vertices of the edgepath $\theta$ belonging to the row $j$. Explicitly, if $n\neq 0$, we set
$
\begin{array}{lr}
\alpha_0=\theta(1),\\
\alpha_i=\theta(i+1)- \theta(i),& 1\leq i\leq n-1,\\
\alpha_{n}=n+m+1-\theta(n),
\end{array
$$
and, if $n=0$, we set $\alpha_0=m+1$. Similarly, if $m\neq 0$, we set
$
\begin{array}{lr}
\beta_0=\theta(n+1),\\
\beta_j=\theta(n+j+1)-\theta(n+j),& 1\leq j\leq m-1,\\
\beta_{m}=n+m+1-\theta(n+m),
\end{array
$$
and, if $m=0$, we set $\beta_0=n+1$.
We thus define the map
\[
\delta_{\theta}:{\mathcal A}_X^{n+1}\times {\mathcal B}_Y^{m+1}\to ({\mathcal A}\times {\mathcal B})_{X\times Y}^{n+m+1}
\]
by $\delta_{\theta}=T_{n+m+1}\circ\big((\Delta^{\mathcal A}_{\alpha_0}\times \cdots \times \Delta^{\mathcal A}_{\alpha_{n}})\times (\Delta^{\mathcal B}_{\beta_0}\times \cdots \times \Delta^{\mathcal B}_{\beta_{m}})\big)$.
If two edgepaths $\theta$ and $\theta'$ differ in only one vertex, say the $k$-th vertex, the maps $\delta_{\theta}$ and $\delta_{\theta'}$ differ in only their $k$-th component. Since, in that case, the $k$-th {(barycentric)} components of $\psi_{\theta}$ and $\psi_{\theta'}$ vanish on $\Delta_{\theta}\cap \Delta_{\theta'}$ we obtain that the maps
\[
\xymatrix{
{\mathcal A}_X^{n+1}\times {\mathcal B}_Y^{m+1}\times \Delta_{\theta}\ar[r]^-{\delta_{\theta}\times \psi_{\theta}} & ({\mathcal A}\times {\mathcal B})_{X\times Y}^{n+m+1}\times \Delta^{n+m}\ar[r] & J^{n+m}_{X\times Y}({\mathcal A}\times {\mathcal B})
}
\]
glue together to give a map
$${\mathcal A}^{n+1}_X\times {\mathcal B}^{m+1}_Y\times \Delta^ n\times \Delta^m\to J^{n+m}_{X\times Y}({\mathcal A}\times {\mathcal B}).$$
This map induces
$$\begin{array}{rcl}
\psi_{n,m}^{{\mathcal A},{\mathcal B}}:J^n_X({\mathcal A})\times J^m_Y({\mathcal B})& \to& J^{n+m}_{X\times Y}({\mathcal A}\times {\mathcal B})\\
\left(\langle \mathbf{a} \mid \mathbf{t}\rangle,\langle \mathbf{b} \mid \mathbf{s}\rangle\right)& \mapsto &
\langle \delta_{\theta}(\mathbf{a},\mathbf{b})\mid \psi_{\theta}(\mathbf{t},\mathbf{s})\rangle, \quad (\mathbf{t},\mathbf{s})\in \Delta_{\theta}.
\end{array}$$
The map $\psi_{n,m}^{{\mathcal A},{\mathcal B}}$ is over $X\times Y$ and {restricts to} a map between fibres
\begin{equation}\label{segundafinm}
\psi_{n,m}^{{ A},{ B}}:J^n({ A})\times J^m({ B}) \to J^{n+m}({A}\times { B})
\end{equation}
which is given by the construction above for trivial fibrations $A\to *$ and $B\to *$. That is, $\psi_{n,m}^{{ A},{ B}}$ comes from the gluing of maps $\delta_{\theta}\times \psi_{\theta}$ where $\delta_{\theta}:{ A}^{n+1}\times { B}^{m+1}\to ({ A}\times { B})^{n+m+1}$ is given by
$$\delta_{\theta}=T_{n+m+1}\circ((\Delta^{A}_{\alpha_0}\times \cdots \times \Delta^{ A}_{\alpha_{n}})\times (\Delta^{B}_{\beta_0}\times \cdots \times \Delta^{B}_{\beta_{m}})).$$
{In turn,~(\ref{segundafinm}) recovers~(\ref{primerfinm}) for $A=B=\ast$.}
\begin{lem} The following diagram
$$\xymatrix{
J_X^{n+1}({\mathcal A}) \times J_Y^{m}({\mathcal B}) \ar[ddr]_{\psi_{n+1,m}^{{\mathcal A},{\mathcal B}}} &
J_X^{n}({\mathcal A})\times J_Y^{m}({\mathcal B})\ar[l]_{\hspace{3mm}\jmath_n\times \mathrm{Id}}\ar[r]^{\mathrm{Id}\times \jmath_m\hspace{3mm}}
\ar[d]^{{\psi_{n,m}^{{\mathcal A},{\mathcal B}}}}
& J_X^{n}({\mathcal A}) \times J_Y^{m+1}({\mathcal B})\ar[ddl]^{\psi_{n,m+1}^{{\mathcal A},{\mathcal B}}}\\ &
{J^{n+m}_{X\times Y}({\mathcal A}\times {\mathcal B})}
\ar[d]|-{\raisebox{1.5mm}{\scriptsize${\rule{0mm}{2mm}\jmath_{n+m}}$}} & \\
& J^{n+m+1}_{X\times Y}({\mathcal A}\times {\mathcal B})&
}$$
is commutative.
\end{lem}
\begin{proof} Let $\langle \mathbf{a} \mid \mathbf{t}\rangle \in J_X^{n}({\mathcal A})$ and $\langle \mathbf{b} \mid \mathbf{s}\rangle \in J_Y^{m}({\mathcal B})$ with $(\mathbf{t},\mathbf{s})\in \Delta_{\theta}\subset \Delta^n\times \Delta^m$ and let $(\alpha_i)$ and $(\beta_j)$ the sequences of integers associated to the $(n,m)$ shuffle $\theta$.
Recall from Remark~\ref{fibredjoins} that
$$(\jmath_n\times \mathrm{Id})(\langle \mathbf{a} \mid \mathbf{t}\rangle,\langle \mathbf{b} \mid \mathbf{s}\rangle)=(\langle \mathbf{a},a_{n+1} \mid \mathbf{t},0\rangle,\langle \mathbf{b} \mid \mathbf{s}\rangle).$$
The inclusion $\Delta^n\times \Delta^m\hookrightarrow \Delta^{n+1}\times \Delta^m$, $(\mathbf{t},\mathbf{s})\mapsto
(\mathbf{t},0,\mathbf{s})$, restricts to the inclusion of $\Delta_{\theta}$ into $\Delta_{\theta'}\subset \Delta^{n+1}\times \Delta^m$ where $\theta'$ is the $(n+1,m)$ shuffle given by
$$(\theta(1)\cdots \theta(n) \,\,\,n+m+1\,||\,\theta(n+1) \cdots \theta(n+m)).$$
Through this inclusion $\Delta_{\theta}$ can be seen as the face of $\Delta_{\theta'}$ opposite to its vertex $(e_{n+1},e'_m)$. The sequences $(\alpha'_i)$ and $(\beta'_j)$ associated to $\theta'$ are given by
$$\alpha_0,\ldots ,\alpha_n,1 \qquad \beta_0,\ldots ,\beta_m+1$$
As a consequence we obtain
$${\psi_{n+1,m}^{{\mathcal A},{\mathcal B}}}(\jmath_n\times \mathrm{Id})(\langle \mathbf{a} \mid \mathbf{t}\rangle,\langle \mathbf{b} \mid \mathbf{s}\rangle)=\langle \delta_{\theta}(\mathbf{a},\mathbf{b}),(a_{n+1},b_m) \mid \psi_{\theta}(\mathbf{t},\mathbf{s}),0\rangle.$$
Similarly we see
$${\psi_{n,m+1}^{{\mathcal A},{\mathcal B}}}(\mathrm{Id}\times \jmath_m)(\langle \mathbf{a} \mid \mathbf{t}\rangle,\langle \mathbf{b} \mid \mathbf{s}\rangle)=\langle \delta_{\theta}(\mathbf{a},\mathbf{b}),(a_{n},b_{m+1}) \mid \psi_{\theta}(\mathbf{t},\mathbf{s}),0\rangle$$
which gives the result.
\end{proof}
As a consequence the maps $\psi^ {{\mathcal A},{\mathcal B}}_{n+1,m}$ and $\psi^ {{\mathcal A},{\mathcal B}}_{n,m+1}$ glue together and give a map
$$\xymatrix{
\Psi_{n,m}^{{\mathcal A},{\mathcal B}}: J_X^{n+1}({\mathcal A})\times J_Y^{m}({\mathcal B})
\cup J_X^{n}({\mathcal A})\times J_Y^{m+1}({\mathcal B}) \ar[r]& J^{n+m+1}_{X\times Y}({\mathcal A}\times {\mathcal B})
}$$
which is over $X\times Y$. As before this map induces at the level of fibres a map
$$\xymatrix{
\Psi_{n,m}^{{A},{B}}:J^{n+1}({ A})\times J^{m}({ B})
\cup J^{n}({A})\times J^{m+1}({B}) \ar[r]& J^{n+m+1}({ A}\times { B}).
}$$
\subsection{The maps $\Phi_{n,m}^{A,B}$}
\label{homotopycharacterization}
For any $k$, the inclusion $J^k(A) \to J^{k+1}(A)$ factors as
$$\xymatrix{
J^k(A)\ar[rr]\ar[rd] && J^{k+1}(A)\\
& CJ^{k}(A)\ar[ru]
}$$
where the map $CJ^{k}(A)\to J^{k+1}(A)$ sends $[\langle \mathbf{a} \mid \mathbf{t}\rangle,u]$ to $\langle \mathbf{a},\ast \mid (1-u)\mathbf{t},u\rangle$ (see Remark~\ref{fibredjoins}). With these maps the following diagram is commutative
$$\xymatrix{
CJ^n(A) \times J^{m}(B) \ar[d]& J^n(A) \times J^{m}(B)\ar[l]\ar[r]\ar@{=}[d] & J^n(A) \times CJ^{m}(B)\ar[d]\\
J^{n+1}(A) \times J^{m}(B) & J^n(A) \times J^{m}(B)\ar[l]\ar[r] &J^n(A) \times J^{m+1}(B)
}$$
and induces a map $\xi: J^ n(A)\circledast J^ m(B)\to J^{n+1}(A)\times J^{m}(B)
\cup J^{n}(A)\times J^{m+1}(B)$. The composition $\Psi_{n,m}^{A,B}\circ \xi$ is a map
$$
\Phi_{n,m}^{A,B}: J^n(A)\circledast J^m(B)\to J^{n+m+1}(A\times B)
$$
which we call {the} \emph{topological {$(n,m)$} shuffle map {for $(A,B)$}}. Note that by construction this map is natural in $A$ and $B$, {and renders the commutative diagram~(\ref{shufflewjoins}) which is the key ingredient in the proof of Theorem~\ref{Hopfproducts} for identifying Hopf invariants of products of fibrations.}
Items (a) and (b) of Remarks~\ref{rmkpinchmap} and~(\ref{lasidentificacionesobvias}) imply that $\Phi_{n,m}^{A,B}$ is a map between spaces each of which has the homotopy type of a $(n+m+1)$-fold suspension. In such terms, $\Phi_{n,m}^{A,B}$ turns out to be a sum of suspended maps which we describe next.
{Start by noticing that} $\Phi_{n,m}^{A,B}$ is given on $J^n(A)\times C J^m(B)$ by
$$
(\langle\mathbf{a}\!\mid\mathbf{t}\rangle,[\langle \mathbf{b}\!\mid\mathbf{s}\rangle,u]) \mapsto
\langle \delta_{\theta}(\mathbf{a},(\mathbf{b},\ast))\mid \psi_{\theta}(\mathbf{t},(1-u)\mathbf{s},u)\rangle
$$
when $(\mathbf{t},(1-u)\mathbf{s},u)\in \Delta_{\theta}, \theta \in {\mathcal S}_{n,m+1}$, and its expression on $CJ^n(A)\times J^m(B)$ is
$$
([\langle\mathbf{a}\!\mid\mathbf{t}\rangle,u],\langle \mathbf{b}\!\mid\mathbf{s}\rangle) \mapsto
\langle \delta_{\theta}((\mathbf{a},\ast),\mathbf{b})\mid \psi_{\theta}((1-u)\mathbf{t},u,\mathbf{s})\rangle
$$
when $((1-u)\mathbf{t},u,\mathbf{s})\in \Delta_{\theta}, \theta \in {\mathcal S}_{n+1,m}$. {In particular, for the identification map}
$$
r:J^{n+m+1}(A\times B)\to \widetilde{\Sigma}^{n+m+1}(A\times B)^{n+m+2}=\left((A\times B)^{n+m+2}\right)\wedge \disfrac{\Delta^{n+m+1}}{\partial(\Delta^{n+m+1})}
$$
in~(\ref{lasidentificacionesobvias}), the composition $r\circ\Phi^{A,B}_{n,m}$ is (based-)constant on $J^n(A)\times J^m(B)=J^n(A)\times CJ^m(B) \hspace{1mm}\cap\hspace{1mm} CJ^n(A)\times J^m(B)$ as well as on all points $(\ast,[\ast,u])\in J^n(A)\times CJ^m(B)$ and $([\ast,u],\ast)\in CJ^n(A)\times J^m(B)$ for $u\in I$, where base points are as in~(\ref{puntobasedeljoin}). Consequently $r\circ\Phi^{A,B}_{n,m}$ factors through the difference pinch map:
$$\xymatrix{
J^n(A)\circledast J^m(B) \ar[r]^{\Phi^{A,B}_{n,m}\;\,}\ar[d]_-{\nu} &
J^{n+m+1}(A\times B) \ar[d]^-r \\
\widetilde{\Sigma}\left(J^n(A)\times J^m(B)\right) \vee \widetilde{\Sigma}\left(J^n(A)\times J^m(B)\right) \ar[r]^-{(\Phi',\Phi'')\;} & \left((A\times B)^{n+m+2}\right)\wedge \disfrac{\Delta^{n+m+1}}{\partial(\Delta^{n+m+1})}.
}$$
The (co)components $\Phi'$ and $\Phi''$ are given by
$$
\Phi'\left(\left[(\langle\mathbf{a}\mid\mathbf{t}\rangle,\langle\mathbf{b}\mid\mathbf{s}\rangle),u\rule{0mm}{3mm}\right]\rule{0mm}{3.5mm}\right)=\delta_{\theta}(\mathbf{a},(\mathbf{b},\ast)) \wedge \psi_{\theta}(\mathbf{t},(1-u)\mathbf{s},u)
$$
when $(\mathbf{t},(1-u)\mathbf{s},u)\in \Delta_{\theta}, \theta \in {\mathcal S}_{n,m+1}$, and by
$$
\Phi''\left(\left[(\langle\mathbf{a}\mid\mathbf{t}\rangle,\langle\mathbf{b}\mid\mathbf{s}\rangle),u\rule{0mm}{3mm}\right]\rule{0mm}{3.5mm}\right)=\delta_{\theta}((\mathbf{a},\ast),\mathbf{b}) \wedge \psi_{\theta}((1-u)\mathbf{t},u,\mathbf{s})
$$
when $((1-u)\mathbf{t},u,\mathbf{s})\in \Delta_{\theta}, \theta \in {\mathcal S}_{n+1,m}$.
\begin{prop}\label{clarificado} Both $\Phi'$ and $\Phi''$ factor through the identification map
\begin{equation*}
\begin{array}{rcl}
\widetilde{\Sigma}(J^n(A) \times J^{m}(B)) &\stackrel{\mathrm{pr}}
\longrightarrow& (A^{n+1}\times B^{m+1})\wedge \disfrac{\Delta^n\times \Delta^m \times I}{\partial(\Delta^n\times \Delta^m \times I)} \\ \left[ \left(\langle \mathbf{a}\mid \mathbf{t}\rangle, \langle \mathbf{b}\mid \mathbf{s}\rangle\right), u \rule{0mm}{3mm}\right] & \longmapsto & (\mathbf{a},\mathbf{b})\wedge [\mathbf{t},\mathbf{s},u]
\end{array}
\end{equation*}
and the resulting maps $$\widehat{\Phi}',\widehat{\Phi}''\colon (A^{n+1}\times B^{m+1})\wedge \disfrac{\Delta^n\times \Delta^m \times I}{\partial(\Delta^n\times \Delta^m \times I)} \to \left((A\times B)^{n+m+2}\right)\wedge \disfrac{\Delta^{n+m+1}}{\partial(\Delta^{n+m+1})}$$ are described up to homotopy by
\begin{eqnarray}
\widehat{\Phi}'&\simeq&\sum\limits_{\theta\in {\mathcal S}_{n,m+1}}(-1)^{|\theta|}\widetilde{\Sigma}^{n+m+1}\delta_{\theta}\circ(\mathrm{Id}_{A^{n+1}}\times \mathrm{Id}_{B^m}\times i_1)\label{lafiprima}\\
\widehat{\Phi}''&\simeq&(-1)^{m}
\sum\limits_{\theta\in {\mathcal S}_{n+1,m}}(-1)^{|\theta|}\widetilde{\Sigma}^{n+m+1}\delta_{\theta}\circ(\mathrm{Id}_{A^{n}}\times i_1\times \mathrm{Id}_{B^{m+1}})\label{lafidobleprima}
\end{eqnarray}
where $i_1:Z\hookrightarrow Z\times Z$ is the inclusion $i_1(z)=(z,*)$.
\end{prop}
\begin{proof}
The asserted factorization of $\Phi'$ and $\Phi''$ follow from an easy checking. Furthermore, {$\widehat{\Phi}'$ factors as}
$$
\resizebox{1\textwidth}{!}{
\xymatrix{
(A^{n+1}\times B^{m+1})\wedge \disfrac{\Delta^n\times \Delta^m \times I}{\partial(\Delta^n\times \Delta^m \times I)}\ar[d]^{\mathrm{Id}\wedge \bar h_{n,m+1}}\\
(A^{n+1}\times B^{m+1})\wedge \disfrac{\Delta^n\times \Delta^{m+1}}{\partial(\Delta^n\times \Delta^{m+1})}\ar[r]^{\mathrm{Id}\wedge\nu_{n,m+1}}&
\bigvee_{\theta\in{\mathcal S}_{n,m+1} }(A^{n+1}\times B^{m+1})\wedge \disfrac{\Delta_{\theta}}{\partial(\Delta_{\theta})}\ar[d]^{(\mathrm{Id}_{A^{n+1}}\times \mathrm{Id}_{B^{m}}\times i_1)\wedge{\mathrm{Id}}}\\
&\bigvee_{\theta\in{\mathcal S}_{n,m+1} }(A^{n+1}\times B^{m+2})\wedge \disfrac{\Delta_{\theta}}{\partial(\Delta_{\theta})}\ar[d]^{(\delta_{\theta}\wedge \bar\psi_{\theta})}\\
&\widetilde{\Sigma}^{n+m+1}(A\times B)^{n+m+2}=((A\times B)^{n+m+2})\wedge \disfrac{\Delta^{n+m+1}}{\partial(\Delta^{n+m+1})}
}}
$$
{where the maps $\bar\psi_{\theta}$ are as in~(\ref{diagramPinchmap}), and} $\bar h_{n,m+1}$ is induced by the (surjective) map $h_{n,m+1}:\Delta^n\times \Delta^m\times I \to \Delta^n\times \Delta^{m+1}$ given by $h_{n,m+1}(\mathbf{t},\mathbf{s},u)=(\mathbf{t},(1-u)\mathbf{s},u)$. Since the maps $\bar h_{n,m+1}$ and $\bar\psi_{\theta}$ are of degree~$1$, {(\ref{inducidoabajo}) implies that the above composition is homotopic to the right-hand term in~(\ref{lafiprima}). The equivalence in~(\ref{lafidobleprima}) is obtained in the same manner, now using the (surjective) map $h_{n+1,m}:\Delta^n\times \Delta^m\times I \to \Delta^{n+1}\times \Delta^{m}$ given by $h_{n+1,m}((\mathbf{t},\mathbf{s},u))=((1-u)\mathbf{t},u,\mathbf{s})$}
and its induced map $\bar h_{n+1,m}$ which is of degree $(-1)^m$.
\end{proof}
Proposition~\ref{clarificado}, the triangle in~(\ref{difzetadif}), and the functoriality of standard difference pinch maps yields:
\begin{cor} There is a commutative diagram
$$\xymatrix{
J^n(A)\circledast J^m(B) \ar[d]_-{\mathrm{pr}\,\cdot\,\zeta} \ar[rr]^-{\Phi^{A,B}_{n,m}}& &J^{n+m+1}(A\times B)\ar[d]^-r\\
(A^{n+1}\times B^{m+1})\wedge \disfrac{\Delta^n\times \Delta^m \times I}{\partial(\Delta^n\times \Delta^m \times I)} \ar[rr]& &\widetilde{\Sigma}^{n+m+1}(A\times B)^{n+m+2}
}$$
where the bottom horizontal map is obtained by subtracting~(\ref{lafiprima}) from~(\ref{lafidobleprima}).
\end{cor}
\subsection{Homology computations}
\label{homologyComp}
{In this section we let $R$ stand for a} commutative ring, and assume that $H_*(A;R)$ and $H_*(B;R)$ are free graded $R$-modules of finite type. {We use the short hand $H(Z):=H_*(Z;R)$ and $\widetilde{H}(Z):=\widetilde{H}_*(Z;R)$,} and apply the standard K\" unneth formulae to identify the homology {(respectively reduced homology)} of {Cartesian} products {(respectively smash products)} with the tensor product of {the homologies (respectively reduced homologies) of the factors. In particular, the map induced in reduced homology by the identification map $X\times Y\to X\wedge Y$ corresponds under the isomorphism $\widetilde{H}(X\times Y)\cong\widetilde{H}(X)\otimes R\oplus R\otimes\widetilde{H}(Y)\oplus\widetilde{H}(X)\otimes\widetilde{H}(Y)$ with projection onto the third summand.}
\begin{thm}\label{morhipsmafterdesusp} Let $\rho:(A\times B)^{n+m+2} \to (A\times B)^{\wedge n+m+2}$ be the identification map. {Up to an automorphism of $\widetilde{H}(A \times B)^{\otimes n+m+2}$,} the morphism induced in {reduced} homology
$$
s^{-(n+m+1)}(\Phi_{n,m}^{A,B})_*:\widetilde{H}(A)^{\otimes n+1} \otimes \widetilde{H}(B)^{\otimes m+1}\to \widetilde{H}(A\times B)^{\otimes n+m+2}
$$
is given by {the restriction of}
\begin{multline*}
(-1)^{m}
\sum\limits_{\theta\in {\mathcal S}_{n+1,m}}(-1)^{|\theta|}(\rho\delta_{\theta})_*\circ(\mathrm{Id}_{A^{n}}\times i_1\times \mathrm{Id}_{B^{m+1}})_*\\
-\sum\limits_{\theta\in {\mathcal S}_{n,m+1}}(-1)^{|\theta|}(\rho\delta_{\theta})_*\circ(\mathrm{Id}_{A^{n+1}}\times \mathrm{Id}_{B^m}\times i_1)_*.
\end{multline*}
\end{thm}
\begin{proof} Let $\tilde\rho$ denote $\widetilde{\Sigma}^{n+m+1}\rho$. Since the sequence of identification maps
$$
\xymatrix{
J^{n+m+1}(A\times B)\ar[r]^-r &\widetilde{\Sigma}^{n+m+1}(A\times B)^{n+m+2} \ar[r]^-{\tilde\rho} & \widetilde{\Sigma}^{n+m+1}(A\times B)^{\wedge n+m+2}
}
$$
is a homotopy equivalence, the morphism induced by $\Phi_{n,m}^{A,B}$ in {reduced} homology coincides, {up to an automorphism,} with the morphism induced by {the composition $\mathrm{pr}\circ \zeta$ followed by the morphism induced by} the difference
\begin{multline*}
(-1)^{m}
\sum\limits_{\theta\in {\mathcal S}_{n+1,m}}(-1)^{|\theta|}\widetilde{\Sigma}^{n+m+1}\rho\delta_{\theta}\circ(\mathrm{Id}_{A^{n}}\times i_1\times \mathrm{Id}_{B^{m+1}})\\
-\sum\limits_{\theta\in {\mathcal S}_{n,m+1}}(-1)^{|\theta|}\widetilde{\Sigma}^{n+m+1}\rho \delta_{\theta}\circ(\mathrm{Id}_{A^{n+1}}\times \mathrm{Id}_{B^m}\times i_1).
\end{multline*}
{The result follows since the composition $\mathrm{pr}\circ\zeta$ induces the ($n+m+1$)-suspension of the inclusion $\widetilde{H}(A)^{\otimes n+1}\otimes \widetilde{H}(B)^{\otimes m+1}\hookrightarrow\widetilde{H}(A^{n+1}\times B^{m+1})$, and because} all the suspensions involved are $(n+m+1)${-fold} suspensions, so we can globally desuspend without altering signs.
\end{proof}
We now compute {the morphism induced by $\Phi^{A,B}_{n,m}$} in some particular cases.
Recall that, if $z\in H(Z)$ is a primitive class, then, for any $k\geq 1$,
$$(\Delta_k^Z)_*(z)=\sum\limits_{i=1}^k 1\otimes \cdots 1 \otimes z_i\otimes 1\otimes \cdots \otimes 1$$
where $z_i=z$. For $k=1$, $\Delta_1^Z=\mathrm{Id}_Z$.
\begin{exs}
\begin{enumerate}[(a)]
\item Let $n=m=0$. In this case ${\mathcal S}_{1,0}=\{(1||\hspace{1.3mm})\}$ and $\hspace{.5mm}{\mathcal S}_{0,1}=\{(\hspace{1.3mm}||1)\}$ are both reduced to a single element. Thus, after a single desuspension, the morphism induced by $\Phi_{0,0}^{A,B}$ is given by\footnote{From this point on we omit any reference of the fact that we really want the obvious restriction of the morphism, and that each of the given descriptions is up to an automorphism.}
$$
\rho_*(T_2)_*(i_1\times\Delta_2^B)_*-\rho_*(T_2)_*(\Delta_2^A\times i_1)_*.
$$
For $x\in \widetilde{H}(A) $ and $y \in \widetilde{H}(B)$ primitive homology classes, we have
\begin{align*}
(T_2)_*(i_1\times\Delta^B_2)_*(x\otimes y)& {}=(T_2)_*((x\otimes 1)\otimes(1\otimes y+y\otimes 1))\\
&{}=(x\otimes 1)\otimes (1\otimes y)+(x\otimes y)\otimes (1\otimes 1).
\end{align*}
Therefore applying $\rho_*$, only the first term $(x\otimes 1)\otimes (1\otimes y)$ remains.
In the same manner,
\begin{align*}
(T_2)_*(\Delta_2^A\times i_1)_*(x\otimes y) & {}=(T_2)_*((1\otimes x+x\otimes 1)\otimes(y\otimes 1))\\
& {}=(-1)^{|x||y|}(1\otimes y)\otimes (x\otimes 1)+(x\otimes y)\otimes (1\otimes 1).
\end{align*}
Applying $\rho_*$, only the term $(-1)^{|x||y|}(1\otimes y)\otimes (x\otimes 1)$ remains and we finally get
$$s^{-1}(\Phi_{0,0}^{A,B})_*(x\otimes y)=(x\otimes 1)\otimes (1\otimes y)-(-1)^{|x||y|}(1\otimes y)\otimes (x\otimes 1).$$
\item Let $n=1$ and $m=0$. In this case, we get the following expression for $x_1,x_2\in \widetilde{H}(A) $ and $y \in \widetilde{H}(B)$ primitive homology classes:
\begin{align*}
s^{-2}(\Phi_{1,0}^{A,B})_*(x_1\otimes x_2\otimes y)
& = (x_1\otimes 1)\otimes (x_2\otimes 1)\otimes (1\otimes y)\\
&\quad -(-1)^{|x_2||y|}(x_1\otimes 1)\otimes (1\otimes y)\otimes (x_2\otimes 1)\\
&\quad +(-1)^{(|x_1|+|x_2|)|y|}(1\otimes y)\otimes (x_1\otimes 1)\otimes (x_2\otimes 1).
\end{align*}
\end{enumerate}
\end{exs}
In order to generalize these computations (in Proposition~\ref{shuffleinhomology} below), we need to fix some preliminary notation. For a permutation $\sigma\in\mathcal{S}_k$ and non-negative integers $d_1,\ldots,d_k$, let $(-1)^{s(\sigma,d_1,\ldots,d_k)}$ be the sign introduced by the morphism induced in reduced homology by maps $$\widetilde{\sigma}: Z^{\wedge k}\to Z^{\wedge k}, \hspace{2mm}(z_1,\ldots,z_k)\mapsto(z_{\sigma^{-1}(1)},\ldots,z_{\sigma^{-1}(k)}),$$ upon evaluation at a tensor $v_1\otimes \cdots \otimes v_k\in \widetilde{H}(Z)^{\otimes k}$ with $|v_i|=d_i$. Explicitly,
$$
\widetilde{\sigma}_*(v_1\otimes \cdots \otimes v_k)=(-1)^{s(\sigma, d_1,\ldots, d_k)}\hspace{.5mm} v_{\sigma^{-1}(1)}\otimes \cdots \otimes v_{\sigma^{-1}(k)}
$$
where $s(\sigma, d_1,\ldots, d_k)=\sum_{(i,j)\in I_\sigma}d_id_j$ and $I_\sigma=\{(i,j)\colon i<j \mbox{ and }\sigma(j)<\sigma(i)\}$ is the set of $\sigma$-inversions. Since we only care about the mod 2 value of $s(\sigma, d_1,\ldots, d_k)$, we can think of $s(\sigma, d_1,\ldots, d_k)$ as the number of shiftings that contribute with a $-1$ sign in permuting $v_1\otimes \cdots \otimes v_k$ to $v_{\sigma^{-1}(1)}\otimes \cdots \otimes v_{\sigma^{-1}(k)}$; that is, the number of $\sigma$-inversions $(i,j)$ for which $d_id_j$ odd.
\begin{prop}\label{shuffleinhomology}
Let $x_i\in \widetilde{H}(A) $ and $y_j \in \widetilde{H}(B)$ be primitive homology classes. The morphism
$${s^{-(n+m+1)}(\Phi_{n,m}^{A,B})_*\colon{}} \widetilde{H}(A)^{\otimes n+1} \otimes \widetilde{H}(B)^{\otimes m+1}\to
\widetilde{H}(A\times B)^{\otimes n+m+2}$$
takes $x_1\otimes \cdots \otimes x_{n+1}\otimes y_{n+2}\otimes \cdots \otimes y_{n+m+2}$ to
$$(-1)^{m}\sum\limits_{\sigma \in {\mathcal S}_{n+1,m+1}}(-1)^{\varepsilon(\sigma)}z_{\sigma^{-1}(1)}\otimes \cdots \otimes z_{\sigma^{-1}(n+m+2)}$$
where ${\mathcal S}_{n+1,m+1}$ denotes the set of $(n+1,m+1)$ shuffles,
$$
z_{i}=\begin{cases}
x_{i}\otimes 1, & \mbox{ if } 1\leq i \leq n+1; \\
1\otimes y_{i}, & \mbox{ if } n+2\leq i \leq n+m+2,
\end{cases}
$$
and ${\varepsilon}(\sigma)=|\sigma|+s(\sigma,|x_1|,\ldots,|x_{n+1}|,|y_{n+2}|,\ldots,|y_{n+m+2}|)$.
\end{prop}
In other words, the sum of the statement can be written as
$$(-1)^{m}\hspace{-2.5mm}\sum\limits_{\sigma \in {\mathcal S}_{n+1,m+1}}\hspace{-2.3mm} (-1)^{|\sigma|}\hspace{.6mm}{\widetilde{\sigma}_*} ((x_1\otimes1)\otimes \cdots \otimes (x_{n+1}\otimes1)\otimes (1\otimes y_{n+2})\otimes\cdots\otimes (1\otimes y_{n+m+2})).$$
\begin{proof} We first observe that a $(n+1,m+1)$ shuffle $\sigma$ satisfies either $\sigma(n+m+2)=n+m+2$ or $\sigma(n+1)=n+m+2$. Then the sum of the statement splits in two parts:
\begin{multline}\label{sumstatementshuffle}
(-1)^{m}\sum\limits_{\sigma(n+m+2)=n+m+2} (-1)^{{\varepsilon}(\sigma)}z_{\sigma^{-1}(1)}\otimes \cdots \otimes z_{\sigma^{-1}(n+m+2)}\\
+ (-1)^{m}\sum\limits_{\sigma(n+1)=n+m+2} (-1)^{{\varepsilon}(\sigma)}z_{\sigma^{-1}(1)}\otimes \cdots \otimes z_{\sigma^{-1}(n+m+2)}.
\end{multline}
These two parts can be identified with the two parts of the morphism described in Theorem~\ref{morhipsmafterdesusp}. Indeed, if $\sigma(n+m+2)=n+m+2$ then $\sigma$ is completely determined by the $(n+1,m)$ shuffle $\theta$ given by $\theta(i)=\sigma(i)$ and we have $|\sigma|=|\theta|$ and
\begin{multline*}
s(\sigma,|x_1|,\ldots,|x_{n+1}|,|y_{n+2}|,\ldots,|y_{n+m+2}|)\\
=s(\theta,|x_1|,\ldots,|x_{n+1}|,|y_{n+2}|,\ldots,|y_{n+m+1}|).
\end{multline*}
Then we have
\begin{multline*}
(-1)^{m}\sum\limits_{\sigma(n+m+2)=n+m+2} (-1)^{{\varepsilon}(\sigma)}z_{\sigma^{-1}(1)}\otimes \cdots \otimes z_{\sigma^{-1}(n+m+2)}\\
= (-1)^{m}\sum\limits_{\theta \in {\mathcal S}_{n+1,m}} (-1)^{|\theta|}\,\widetilde{\theta}_*(z_1\otimes \cdots \otimes z_{n+m+1})\otimes (1\otimes y_{n+m+2}).
\end{multline*}
On the other hand, the first part of the morphism described in Theorem~\ref{morhipsmafterdesusp} applied to $x_1\otimes \cdots \otimes x_{n+1}\otimes y_{n+2}\otimes \cdots \otimes y_{n+m+2}$ gives
$$(-1)^{m}
\sum\limits_{\theta\in {\mathcal S}_{n+1,m}}(-1)^{|\theta|}(\rho\delta_{\theta})_*(x_1\otimes \cdots \otimes x_{n+1}\otimes 1\otimes y_{n+2}\otimes \cdots \otimes y_{n+m+2})$$
and a straightforward calculation gives
\begin{multline*}
(\rho\delta_{\theta})_*(x_1\otimes \cdots \otimes x_{n+1}\otimes 1\otimes y_{n+2}\otimes \cdots \otimes y_{n+m+2})\\
= \widetilde{\theta}_*(z_1\otimes \cdots \otimes z_{n+m+1})\otimes (1\otimes y_{n+m+2})
\end{multline*}
which identifies the first part of $(\ref{sumstatementshuffle})$ with the first part of the morphism of
Theorem~\ref{morhipsmafterdesusp}.
We now suppose that $\sigma(n+1)=n+m+2$. Then $\sigma$ is completely determined by the $(n,m+1)$ shuffle $\theta$ given by $\theta(i)=\sigma(i)$ for $1\leq i\leq n$ and $\theta(i)=\sigma(i+1)$ for $n+1\leq i\leq n+m+1$. We have $|\sigma|=|\theta|+m+1$ and
\begin{multline*}
s(\sigma,|x_1|,\ldots,|x_{n+1}|,|y_{n+2}|,\ldots,|y_{n+m+2}|)\\
=s(\theta,|x_1|,\ldots,|x_{n}|,|y_{n+2}|,\ldots,|y_{n+m+2}|)+|x_{n+1}|\left(|y_{n+2}|+\cdots+|y_{n+m+2}|\right).
\end{multline*}
Writing $e=|x_{n+1}|(|y_{n+2}|+\cdots+|y_{n+m+2}|)$, the second part of $(\ref{sumstatementshuffle})$ becomes
\begin{align*}
& (-1)^{m} \sum\limits_{\sigma(n+1)=n+m+2} (-1)^{{\varepsilon}(\sigma)}z_{\sigma^{-1}(1)}\otimes \cdots \otimes z_{\sigma^{-1}(n+m+2)}\\
& = -\sum\limits_{\theta \in {\mathcal S}_{n,m+1}}(-1)^{\varepsilon(\theta)+ e}
(z_{\sigma^{-1}(1)}\otimes \cdots \otimes z_{\sigma^{-1}(n+m+1)})\otimes(x_{n+1}\otimes 1)\\
& = -\sum\limits_{\theta \in {\mathcal S}_{n,m+1}} (-1)^{|\theta|+e}\;{\widetilde{\theta}_*}
(z_1\otimes \cdots \otimes z_{n+m+1})\otimes(x_{n+1}\otimes 1)\\
& = -\sum\limits_{\theta \in {\mathcal S}_{n,m+1}} (-1)^{|\theta|}(\rho\delta_{\theta})_*(x_1\otimes \cdots \otimes x_{n+1}\otimes y_{n+2}\otimes \cdots \otimes y_{n+m+2}\otimes 1)
\end{align*}
which completes the proof.
\end{proof}
\begin{prop}
Consider the composite
$$ \xymatrix{
F_n(S^p)\circledast F_m(S^p) \ar[r]^{\;\;\,\Phi_{n,m}^{\Omega S^p,\Omega S^p}} & F_{n+m+1}(S^p \times S^p) \ar[r]^{\hspace{3.4mm}\bar{\chi}_{n+m+1}} & F_{n+m+1}(S^p).
}$$
If $p\ge2$, the degree $d$ of the map $S^{(n+m+2)p-1}\to S^{(n+m+2)p-1}$ induced by restriction of $\bar{\chi}_{n+m+1}\circ{\Phi_{n,m}^{\Omega S^p,\Omega S^p}}$ to the bottom sphere is given by
$$\pm d=\#(S^+_{n+1,m+1})+(-1)^p\#(S^-_{n+1,m+1})$$
where $S^+_{n+1,m+1}$ (resp.~$S^-_{n+1,m+1}$) stands for the set of $(n+1,m+1)$ shuffles of positive (resp.~negative) signature.
\end{prop}
\begin{proof}
In terms of the functorial homotopy equivalence $$F_{n+m+1}(Z)=J^{n+m+1}(\Omega Z)\simeq\widetilde{\Sigma}^{n+m+1}(\Omega Z)^{\wedge n+m+2},$$ the map
$\bar{\chi}_{n+m+1}: F_{n+m+1}(S^p\times S^p) \to F_{n+m+1}(S^p)$ takes the form
$$\widetilde{\Sigma}^{n+m+1} \bar{\chi}^{\wedge n+m+2}\colon (\Omega S^p\times \Omega S^p)^{\wedge n+m+2}\wedge\disfrac{\Delta^{n+m+1}}{\partial \Delta^{n+m+1}} \to (\Omega S^p)^{\wedge n+m+2}\wedge \disfrac{\Delta ^{n+m+1}}{\partial \Delta^{n+m+1}}.$$
Since this is a $(n+m+1)$-fold suspension, we have
$$s^{-({n+m+1})}(\bar{\chi}_{n+m+1}\circ \Phi_{n,m}^{\Omega S^p,\Omega S^p})_*=(\bar{\chi}_*)^{\otimes n+m+2} \circ s^{-({n+m+1})}(\Phi_{n,m}^{\Omega S^p,\Omega S^p})_{*}$$
where $\bar{\chi}:\Omega S^p\times \Omega S^p\to \Omega S^p$ is given by $(\alpha,\beta)\mapsto \alpha^{-1}\beta$. We take $R={\mathbb{Z}}$. {Recall that} $H_*(\Omega S^p)={\mathbb{Z}}[x]$ where $x$ is primitive with degree $|x|=p-1$, {and that} the morphism
$$\bar{\chi}_*:H_*(\Omega S^p \times\Omega S^p) \to H_*(\Omega S^p)$$
is given by $\bar{\chi}_*(x\times 1)=-x$ and $\bar{\chi}_*(1\times x)=x$.
By Proposition \ref{shuffleinhomology}, we know that $s^{-({n+m+1})}(\Phi_{n,m}^{\Omega S^p,\Omega S^p})_{*}$ takes the generator $x^{\otimes n+m+2}$ of $H_*(\Omega S^p)^{\otimes n+m+2}$ to
$$(-1)^{m}\sum\limits_{\sigma \in {\mathcal S}_{n+1,m+1}} (-1)^{{\varepsilon}(\sigma)}z_{\sigma^{-1}(1)}\otimes \cdots \otimes z_{\sigma^{-1}(n+m+2)}$$
where $z_i=x\otimes 1$ for $1\leq i\leq n+1$ and $z_i=1\otimes x$ for $n+2\leq i\leq n+m+2$.
Since in each term there are exactly $n+1$ components $x\times 1$ we obtain
$$(\bar{\chi}_*)^{\otimes n+m+2} \circ s^{-({n+m+1})}(\Phi_{n,m}^{\Omega S^p,\Omega S^p})_{*}(x^{\otimes n+m+2})=\pm\hspace{-3mm}\sum\limits_{\sigma \in {\mathcal S}_{n+1,m+1}} (-1)^{{\varepsilon}(\sigma)}x^{\otimes n+m+2}.$$
Now, if $\sigma$ is a permutation of positive signature, we have, for some integer $k$ {(depending on $\sigma$),}
$${\varepsilon}(\sigma)=|\sigma|+s(\sigma,\underbrace{|x|,\ldots,|x|}_{n+m+2})=|\sigma|+2k|x||x|\equiv 0\quad mod \,\,2$$
while, if $\sigma$ is a permutation of negative signature, we have, for some integer $k$ {(depending on $\sigma$),}
$${\varepsilon}(\sigma)=|\sigma|+s(\sigma,\underbrace{|x|,\ldots,|x|}_{n+m+2})=|\sigma|+(2k+1)|x||x|\equiv 1+|x|=p\quad mod \,\, 2.$$
This gives the result.
\end{proof}
\begin{exs}\label{degree-2cells}
\begin{enumerate}[(a)]
\item
The map $S^{4p-1}\to S^{4p-1}$ induced by restriction to the bottom cell of the composite
$$\xymatrix{
F_1(S^p)\circledast F_1(S^p) \ar[rr]^-{\Phi_{1,1}^{\Omega S^p,\Omega S^p}} && F_3(S^p \times S^p) \ar[r]^-{\bar{\chi_3}} & F_3(S^p)
}$$
has degree $\pm(4+2(-1)^p)$.
\item The composite
$$\xymatrix{
F_1(S^p)\circledast F_0(S^p) \ar[rr]^-{\Phi_{1,0}^{\Omega S^p, \Omega S^p}} && F_2(S^p\times S^p) \ar[r]^-{\bar{\chi}_2} & F_2(S^p)
}$$
induces a map $S^{3p-1}\to S^{3p-1}$ by restriction to the bottom cell. The degree of this map is $\pm(2+(-1)^p)$.
\item The composite
$$\xymatrix{
F_0(S^p)\circledast F_0(S^p) \ar[rr]^-{\Phi_{0,0}^{\Omega S^p, \Omega S^p}} && F_1(S^p\times S^p) \ar[r]^-{\bar{\chi}_1} & F_1(S^p)
}$$
induces a map $S^{2p-1}\to S^{2p-1}$ by restriction to the bottom cell. The degree of this map is $\pm(1+(-1)^p)$.
\end{enumerate}
\end{exs}
|
\section{Introduction}
Magnetism at the sub-micrometer and nano scale attracts a great deal of interest
for both fundamental reasons and for their prospective use in logic and memory
applications. As not only the static properties, such as the magnetic magnetic domain
configuration, but also the dynamics properties on the sub-nanosecond timescale
(e.g. the resonances and magnetization switching) are strongly determined by the reduced
dimensionality, appropriate characterization techniques are required.
The conventional method for high frequency characterisation of magnetic
systems is the cavity based ferromagnetic resonace technique. However, nanostuctures
can not be studied in remanence as the fixed frequency operation requires
the bias field to be swept. To address this problem, different techniques
have been developed. Dependent on how the magnetic response is detected they
can be divided in two categories. Vector Network Analyzer Ferromagnetic Resonance
(VNA-FMR \cite{neudecker06,kalarickal06}) and Pulsed Inductive Microwave Magnetometry
(PIMM \cite{kos02,silva99}) detect the resonance electrically and in Time Resolved
Magneto-Optical Kerr Effect (TR-MOKE \cite{acremann01}) and Time Resolved Scanning
Transmission X-Ray Microscopy experiments (STXM \cite{bolte08,vansteenkiste09,kammerer11}) optical
detection is used. The optical methods have a high sensisitivity to
detect the signal of a single magnetic microstructure, but have a much more complicated set-up than the electrical methods,
and require a femtosecond laser or pulsed X-ray source. On the other hand, detection in the
frequency domain, like conventional FMR and VNA-FMR can achieve much higher signal-to-noise ratios \cite{neudecker06}.
Here we present an approach which combines the frequency domain method with optical detection:
the Magneto-Optical Spectrum Analyzer (MOSA).
This method is a hybrid method between VNA-FMR and TR-MOKE in the context of measurement abilities and construction.
TR-MOKE makes use of a pulsed laser to probe the magnetization ($\vec{M}$)
through the magneto-optical Kerr effect\cite{argyres55,zak90,polisetty08,zvezdin97,qiu00} at regular intervals, while the sample
is excited using e.g. another laser pulse or a microwave signal generator. By shifting the arrival time of
the probe pulses with respect to the excitation, time domain information can be gathered.
This method, on the other hand
probes the magnetization using a continuous wave laser and measures (again through the Kerr effect) the fast change
of magnetization with an ultrafast photodiode in the frequency domain.
\section{Description of the set-up}
Shown in Fig.\ref{setup} is the optical layout of our set-up. Light (660nm) from a laser diode (LD) is linearly
polarized using a polarizer (Pol) and passes through a non-polarizing beamsplitter (BS). An objective lens (L1)
focusses it to a diffraction limited spot ($\approx 500$nm) on the sample, where microwave interconnects provide RF current for the excitation
and an electromagnet can provide a bias field of up to 50mT.
The reflected light is collected by the same objective lens and redirected by the beamsplitter onto an analyzer (An).
The tranmission axis of this analyzer is set at an angle of 45$^\circ$\ with respect to the transmission axis of the first
polarizer. The beam is then focussed using an aspheric lens (L2) onto a multimode optical fiber (MM Fiber),
transporting it over a large distance (50 m in our case) to an ultrafast photodiode (PD) with a bandwidth of approximately 12GHz. At this point intensity variations are
measured.
The objective lens is mounted on a piezo stage to allow scanning of the probe beam over the sample surface. By recording
the DC reflected light intensity, we can image the sample and correctly focus and position the probe beam on the sample.
For this purpose the light is deflected with a mirror towards a conventional photodiode (both not shown for clarity).
\begin{figure}
\includegraphics[width=8.5cm]{setup.jpg}
\caption{A basic sketch of the optical part of the set-up, used for polar Kerr detection. LD is a laser diode, Pol a polarizer, BS a non-polarizing beamsplitter,
L1 an objective lens, An an analyzer, L2 and L3 aspheric lenses and PD an ultrafast photodiode. The angle between the transmission axis of Pol and An is 45$^\circ$.}
\label{setup}
\end{figure}
The out-of-plane magnetization dynamics are measured via the polar Kerr effect. When reflecting off the sample, linearly polarized lights
gains both an ellipticity ($\epsilon_\text{K}$) and a rotation
of the major polarization axis, known as the Kerr angle ($\theta_\text{K}$). For out-of-plane saturated Permalloy, the polar Kerr angle is
typically 1 mrad\cite{veis12}. Both the sign and magnitude of this angle depend on the sign and magnitude of
the out-of-plane magnetization of the probed area.
The reflected light is analyzed with the polarizer at an angle of 45$^\circ$, thus yielding an intensity of $I = I_0(1/2+\theta_\text{K})$, where
$I_0$ is the intensity before the analyzer. Placing the analyzer at 45$^\circ$\ maximizes the signal.
The electrical part of the set-up is shown in Fig. \ref{setup2}.
One signal generator produces the RF current at frequency $f$ used for
exciting the sample. The RF power through the sample is kept level by using a diode detector measuring the power coming out
of the sample. The bias tee provides
the necessary bias voltage and shunts the DC photocurrent. The DC photocurrent flowing out of the bias tee is measured for monitoring
purposes.
Assuming a linear dependence of the light intensity on the magnetization \cite{argyres55,zvezdin97}, we can estimate the AC current induced in the photodiode
due to magnetization dynamics as $\delta I \approx I_\text{DC} \theta_\text{K,max} \delta m_z$, where $\theta_\text{K,max}$
is the Kerr angle at saturation ($M_z = M_\text{S}$),
$\delta m$ the reduced out-of-plane magnetization ($M_z/M_\text{S}$) and $I_\text{DC}$ the DC photocurrent.
The AC photocurrent is passed on to the
low noise preamplifier, which increases the signal level by 30dB.
After this preamplifier, a mixer downconverts this high frequency signal to a frequency in the audio range. To this end
a second signal generator, phase locked to the first one, produces a high frequency signal
at a frequency offseted by several kilohertz ($\Delta f$) with respect to the excitation frequency $f$. Thus, the signal that the mixer produces is
at the frequency $\Delta f$. It is further amplified by a second amplifier and finally sampled using a high end ADC.
A computer records the incoming data from the ADC and performs an FFT to compute the signal strength at $\Delta f$.
\begin{figure}
\includegraphics[width=8.5cm]{setup2.jpg}
\caption{A simplified schematic of the electrical part of the set-up. A diode detector connected to the sample helps in keeping the power transmitted through the sample level when the frequency is varied. Also shown are the photodiode (PD), termination resistor ($R_\text{Term}$), bias tee, RF preamplifier, frequency mixer, low frequency amplifier and ADC. Both signal generators produce a tone in the microwave range ($f$), but are slightly offset by frequency $\Delta f$ in the kHz range.
\label{setup2}}
\end{figure}
\section{Comparison with other methods}
To estimate the detection limit of the set-up, we compare the typical signal levels to the fundamental noise present. There are
three main contributions to this noise\cite{horowitz89}:
\begin{itemize}
\item a contribution from the DC photocurrent, known as shot noise, with a current noise density given by $i_\text{noise} = \sqrt{2 q I_\text{DC}}$;
\item a contribution from the noise intrinsic to the diode, quantified by the Noise Equivalent Power (NEP);
\item and a contribution from the terminating resistor, $R_\text{Term}$ (50$\Omega$), known as Johnson noise, with a voltage noise density given by $v_\text{noise} = \sqrt{4 kTR_\text{Term}}$.
\end{itemize} We assume that these are sources of white noise and that all other possible sources generate much less
noise in the frequency range of interest.
If we compare this noise with the signal strength, we arrive at the following expression for the Signal-to-Noise Ratio (SNR) at the input of the preamp:
\begin{equation}
\text{SNR} = 10 \text{log}_{10} \frac{ I_\text{DC}^2 \theta_\text{K}^2 \langle m_z^2 \rangle }{B (2q I_\text{DC} + 4kT/R_\text{Term} + (\mathcal{R}\cdot \text{NEP})^2)},
\end{equation} where $B$ is the measurement bandwidth and $\mathcal{R}$ the responsivity of the photodiode (A/W (at a specific wavelength)).
Typical values for our set-up would involve $I_\text{DC} = 100.0\mu\text{A}$, $B = 1\text{Hz}$ and $\langle m_z^2 \rangle = 10^{-4}$.
The Johnson noise is the largest contribution ($1.8 \cdot 10^{-11}$ A/$\sqrt{\text{Hz}}$), followed by the shot noise ($5.7 \cdot 10^{-12}$ A/$\sqrt{\text{Hz}}$).
In comparision, the NEP for the photodiode we have used is negligible (only $4.5 \cdot 10^{-17}$ A/$\sqrt{\text{Hz}}$). Because the contribution from shot noise
is still smaller than the Johnson noise, the SNR can be significantly improved by increasing the light intensity incident on the photodiode. Only when the photocurrent
reaches 1 mA (equivalent to 10 mW optical power on the diode) does the shot noise exceed the Johnson noise. However increasing light intensity
also entails heating up to sample, even to above the Curie temperature.
A further increase in SNR
could be obtained by cooling the detector, lowering the Johnson noise.
Adding these terms we find a SNR
of approximatly 34dB. In addition, the first amplifier adds a noise figure of 2dB, thus the final SNR is about 32dB.
But as the signal itself is on the order of -130dBm or $5\cdot 10^{-14}$ mW, absolute care in handling the microwave signals is still required.
To illustrate this, we have excited a sample with enough RF power (+10dBm) so that the it would precess with
$\delta m_z \approx 0.01$. The same amount of RF power but at a slightly offseted frequency
was sent through a microstrip next to the receiver, to allow for a comparison between the magnetic signal and the direct coupling
from the excitation to the receiver. The result in Fig. \ref{leakage} shows that direct coupling is much stronger than the magnetic signal and that
our estimate of the SNR is quite accurate. To this end, the detection
has been separated by a large distance from the excitation.
Low frequency $1/f$ noise is not showing due to a low frequency cut-off characteristic of the low frequency amplifier,
but the noise floor increases below 5kHz.
\begin{figure}
\includegraphics[width=8.5cm]{leakage.jpg}
\caption{Shown is a FFT when both a magnetic signal and a direct coupling are present. In this particular case
a Permalloy disc with a diameter of 20$\mu$m\ was excited at 4GHz with +10dBm of power and biased with a field of 20mT. The leakage was caused by
a microstrip, which was properly terminated and carried the same amount of power as used for the excitation, but at a slightly different frequency.\label{leakage}}
\end{figure}
Our method is complementary to both VNA-FMR and TR-MOKE. To the former, we add the advantage of spatial selectivity. This means
that several magnetic structures can be fabricated in each others vicinity and we are still able to probe each one separately, or that the
spatial variation of the magnetization dynamics can be analyzed, in contrast to VNA-FMR.
Where TR-MOKE is a time domain method, we are measuring directly in frequency domain, thus it is easier to measure resonance curves
directly and not have to rely on Fourier transforms of time domain data.
In comparison to pulsed methods, we have independent control of frequency and amplitude, resulting in non-ambiguous spectra.
Further, our method allows us to probe the signal at any arbitrary frequency,
something that is difficult to do with stroboscopic time domain measurement methods.
One last advantage of our method is that it does not require a femtosecond pulsed laser, high end oscilloscope or vector network analyzer, thus eliminating a large cost.
\section{Experimental details}
As an illustration we have measured the uniform precession of magnetization in Permalloy discs with a 20$\mu$m\ diameter.
The samples were fabricated on a silicon substrate. Structure definitions were made with electron beam lithography and the lift-off technique.
After the last metal deposition an ALD coating of Al$_2$O$_3$ was deposited. The samples were then wire bonded to a high frequency substrate.
An example of such a sample is shown in Fig. \ref{figSample}.
\begin{figure}
\includegraphics[width=6cm]{fmr_sample.jpg}
\caption{A SEM micrograph of a sample used for measurements. The dots are 75$\mu$m\ and 20$\mu$m\ in diameter. \label{figSample}}
\end{figure}
An example of a resonance curve
where the frequency is swept at a fixed field of 45mT, is shown in Fig. \ref{resonance}. The peak was fitted
to a Lorentz curve and the resulting resonance frequency was determined to be 6120$\pm$2 MHz, with a linewidth of 188$\pm$4 MHz.
This illustrates that the linewidth can be accuratly measured on single microscopic elements.
\begin{figure}
\includegraphics[width=8.5cm]{fmr_20um.jpg}
\caption{The resonance curve for a 20$\mu$m\ diameter, 50nm thickness Permalloy disc in a field of 45mT. The solid
line is the Lorentzian fit.\label{resonance}}
\end{figure}
The detector itself has a frequency dependence, making the recorded spectrum a product of the magnetic response and the
detector response. To investigate this effect, a frequency sweep was performed for a different number of bias fields and the resonance frequency was determined for each.
In Fig. \ref{kittel} the resulting datapoints are compared with the Kittel equation, which for an infinitly thin film is given by
$f_\text{Res} = 28.0 \sqrt{ \mu_0 H_\text{Bias} (\mu_0 H_\text{Bias} + \mu_0 M_s) }$ GHz/T.
Literature values have been used
for calculating the Kittel equation ($\mu_0 M_s = 1.04$T)\cite{coey10}. The datapoints are in good agreement with the Kittel equation,
ruling out strong frequency variations in the detection which might interfere with measurements.
\begin{figure}
\includegraphics[width=8.5cm]{kittel.jpg}
\caption{The evolution of the resonance peak of the uniform precession when the field is increased.
The Kittel equation for a thin film is shown as a black line. \label{kittel}}
\end{figure}
Finally, the magnetic signal can also be used for imaging purposes; when the laser beam is scanned over the sample using the piezo stage
and the magnetic signal at each point is recorded. This can be compared with the reflectivity,
which is acquired simultaneously. When the sample is excited at resonance, the contrast is highest. An example
is shown in Fig. \ref{sample2}. Here we compare the reflectivity (clearly showing the Au CPW, Si substrate and Permalloy discs) with
the magnetic signal showing only the Permalloy disc.
\begin{figure}
\includegraphics[width=8.5cm]{sample2.jpg}
\caption{At the top an image generated using the magnetic response of a uniform resonance in a 3$\mu$m\ dot at 6GHz. Shown
at the bottom is an image generated using reflectivity data that was collected simultaneously.
\label{sample2}}
\end{figure}
\section{Conclusion}
We have developed a method of probing magnetization dynamics at the multiple-GHz range using
a frequency domain method that offers spatial sensitivity. The set-up is relatively simple, yet allows
for high quality measurements, thus enabling a fast exploration of excitation parameters. We have illustrated
that our method can yield quantitative results using uniform excitation on single microscopic elements.
\begin{acknowledgments}
Mathias Helsen, Arne Vansteenkiste and Bartel Van Waeyenberge acknowledge funding by FWO-Vlaanderen and BOF-UGent.
\end{acknowledgments}
\bibliographystyle{ieeetr}
|
\section{Introduction}
The discovery of superconductivity in the ternary iron selenides A$_{y}$Fe$_{2-x}$Se$_2$ (A = K, Rb, Cs, ...)~\cite{Guo2010,Wang2011a,Krzton2011,Fang2011} has triggered great interest not only because of its relatively high value for $T_c$ (over 30~K), but also for its unique properties absent in other iron-based superconductors. First, in this system the superconductivity is found to be induced from an antiferromagnetically insulating phase~\cite{Fang2011}, different from other iron-based superconductors where the superconductivity is evolved from a spin-density-wave type metal~\cite{Dong2008,Cruz2008,Singh2009}. Further, the hole pockets at the center of the Brillouin zone are absent, leaving just the electron pockets at the corner of the zone~\cite{Nekrasov2011,Yan2010,Zhang2011,Qian2011}, which suggests that the inter-pocket scattering between the hole and electron pockets is not an essential ingredient for superconductivity in this system. Moreover, Fe vacancies are found to be present and form the $\sqrt{5} \times \sqrt{5} \times 1$ $I4/m$ ordered pattern in the lattice~\cite{Ye2011,Zhang2012,Bao2011,Wang2011}. In addition, evidence for phase separation and coexistence of magnetism and superconductivity has also been observed~\cite{Bao2011,Wang2011,Torchetti2011,Zhao2012,Texier2012,Friemel2012,Ricci2011,Charnukha2012a,Yuan2012,Homes2012}.
The superconductivity in iron-based superconductors is believed to occur beyond the conventional electron-phonon coupling mechanism~\cite{Boeri2008,Noffsinger2009}, and the pairing mechanism is proposed to be mediated by exchange of the antiferromagnetic spin fluctuations~\cite{Mazin2008,Kuroki2008}. In this newly discovered iron selenides system, a number of Raman-active phonon modes~\cite{Zhang2012,Ignatov2012} as well as phonon anomalies associated with a rather specific type of electron-phonon coupling~\cite{Zhang2012} have been observed in Raman-scattering studies. Moreover, the measurements of optical conductivity have also identified a series of infrared-active phonon modes in the far infrared region as well as a number of high-frequency bound excitations~\cite{Yuan2012,Homes2012,Chen2011,Charnukha2012}. Additionally, nuclear magnetic resonance (NMR) experiments suggest that spin fluctuations are weak~\cite{Torchetti2011,Yu2011,Kotegawa2011}. All the above results give evidence for the possible role played by electron-phonon coupling to the appearance of superconductivity in this system. In this work, we performed optical spectroscopy measurements on the superconducting single crystal of Rb$_{0.8}$Fe$_{1.68}$Se$_2$\ with $T_{c}$ $\simeq$ 31~K. Our results show that the Rb$_{0.8}$Fe$_{1.68}$Se$_2$\ presents similar optical conductivity features as the previous reports in K$_{y}$Fe$_{2-x}$Se$_2$~\cite{Yuan2012,Homes2012,Chen2011}. Moreover, we found that the phonon mode around 200\icm\ displays remarkable asymmetry effect at low temperatures. A detailed Fano line shape analysis suggests a strong coupling between lattice and electronic background.
\section{Experiments}
High quality Rb$_{0.8}$Fe$_{1.68}$Se$_2$\ ($T_c$ $\simeq$ 31~K) single crystals were grown by the Bridgeman method and the composition of the samples were determined by the ICP analysis method~\cite{Li2011}. The \emph{ab}-plane reflectivity $R(\omega)$ was measured at a near-normal angle of incidence on BOMEN DA8 FT-IR spectrometers. An \emph{in situ} gold overfilling technique~\cite{Homes1993} was used to obtain the absolute reflectivity of the sample. Data from 40 to $12\,000\icm$ were collected at different temperatures from 5~K to 300~K on freshly cleaved surfaces on an ARS-Helitran crysostat. We extended the reflectivity to $50\,000\icm$ at room temperature with an AvaSpec-2048 $\times$ 14 optical fiber spectrometer. The real part of the optical conductivity $\sigma_1(\omega)$ was determined from the reflectivity through Kramers-Kronig analysis~\cite{Dressel2002}.
\section{Results and discussion}
Figure~\ref{fig.1}(a) shows the temperature dependence of the reflectivity $R(\omega)$ for Rb$_{0.8}$Fe$_{1.68}$Se$_2$\ over broad frequencies up to 10\,000\icm. Over all, the reflectivity has a rather low value around 0.4 reflects the characteristic of a poor metal. As a comparison, we plot the reflectivity for Rb$_{0.8}$Fe$_{1.68}$Se$_2$\ and Ba$_{0.6}$K$_{0.4}$Fe$_{2}$As$_{2}$\ at 300~K in the inset of Fig.~\ref{fig.1}(a). The reflectivity for Ba$_{0.6}$K$_{0.4}$Fe$_{2}$As$_{2}$\ is about twice higher than that for Rb$_{0.8}$Fe$_{1.68}$Se$_2$. Moreover, the reflectance spectrum for Rb$_{0.8}$Fe$_{1.68}$Se$_2$\ is dominated by a series of sharp features associated with the infrared-active vibrations in the low-frequency region as well as other high-energy electronic state features.
Figure~\ref{fig.1}(b) displays the temperature dependence of the conductivity $\sigma_1(\omega)$ for Rb$_{0.8}$Fe$_{1.68}$Se$_2$\ in the low-frequency region. At room temperature $\sigma_1(\omega)$ is dominated by a series of infrared-active phonon modes (indicated by the red arrows in Fig.~1(b)) superposed on a flat and incoherent electronic background. The large number of the low-frequency infrared-active phonon modes is beyond the observation in BaFe$_2$As$_2$ ($I4/mmm$), where only two in-plane infrared-active phonon modes are observed~\cite{Akrap2009}. As the temperature is reduced, the phonon modes narrow and increase slightly in frequency, meanwhile, the electronic background increases slightly and a small Drude-like response emerges. In addition to the low-frequency vibrational features, there are several broad bound excitation features observed at high frequency, as shown in the inset of Fig.~\ref{fig.1}(b). The three prominent features are located at 700, 4300, and 5700\icm, which have also been observed in the K- and Rb-doped iron selenides in previous optical studies~\cite{Yuan2012,Homes2012,Chen2011,Charnukha2012}, and are labeled as $\gamma$, $\alpha$, and $\beta$, respectively. The high-frequency bound excitation features were thought to be highly related to the Fe vacancies and their orderings in this system~\cite{Chen2011}, where the Fe vacancies form a $\sqrt{5} \times \sqrt{5} \times 1$ superlattice pattern~\cite{Ye2011,Zhang2012,Bao2011,Wang2011}.
The optical conductivity can be described by the Drude-Lorentz model:
\begin{equation}
\label{DrudeLorentz}
\sigma_{1}(\omega)=\frac{2\pi}{Z_{0}}\biggl[\sum_{k}\frac{\Omega^{2}_{p,j}}{\omega^{2}\tau_j + \frac{1}{\tau_j}} + \sum_{k}\frac{\gamma_{k}\omega^{2}S^{2}_{k}}{(\omega^{2}_{0,k} - \omega^{2})^{2} + \gamma^{2}_{k}\omega^{2}}\biggr],
\end{equation}
where $Z_{0}$ is the vacuum impedance. The first term describes a sum of free-carrier Drude responses, each characterized by a plasma frequency $\Omega_{p,j}$ and a scattering rate $1/\tau_j$. The square of plasma frequency $\Omega^2_{p}$ is proportional to $n/m$, where $n$ and $m$ refer to the density and effective mass of conduction electron. The second term is a sum of Lorentz oscillators, each having a resonance frequency $\omega_{0,k}$, a line width $\gamma_k$ and an oscillator strength $S_k$.
Figure~\ref{fig.2} shows the results of the Drude-Lorentz fit to the conductivity at 300~K [Fig.~\ref{fig.2}(a)] and 40~K [Fig.~\ref{fig.2}(b)]. We use five Lorentz oscillators to reproduce the five major phonon modes and one Lorentz components to describe the interband transition near 700\icm. In addition to these Lorentz terms, we have to use a narrow and a broad Drude terms to reproduce the low-frequency free carrier contributions. As can be seen, the model reasonably reproduces the $\sigma_{1}(\omega)$ spectrum. The plasma frequency of the Drude terms is $\Omega_{p,1} \simeq$ 178\icm\ and $\Omega_{p,2} \simeq$ 368\icm\ at 300~K, $\Omega_{p,1} \simeq$ 230\icm\ and $\Omega_{p,2} \simeq$ 742\icm\ at 40~K. The total plasma frequency is given by $\Omega_{pD} = \sqrt{\Omega^2_{p,1} + \Omega^2_{p,2}}$. Thus the value of $\Omega_{pD}$ is $\sim$ 408\icm\ at 300~K and $\sim$ 776\icm\ at 40~K, which is about one order smaller than those of other iron-based superconductors~\cite{Hu2008,Tu2010,Dai2013}. The small value of $\Omega_{pD}$ is consistent with the low carrier density in this material.
The vibrational parameters for the three most prominent phonon modes in Rb$_{0.8}$Fe$_{1.68}$Se$_2$\ observed around 230\icm, 200\icm, and 150\icm\ at 300~K and 5~K are summarized in Table~\ref{tab.1}. The detailed temperature dependence of the phonon frequency for the three phonon modes are shown in Fig.~2(c-e). As we can see, the phonon frequency increases continuously with decreasing temperature. Additionally, as shown in Fig.~\ref{fig.2}(b), we notice that there exists a strong asymmetric effect for the phonon mode around 200\icm\ at low temperatures, while the Lorentzian oscillator was insufficient to accurately describe this line shape. The asymmetric line shape is general, and coupling of a lattice mode to a continuum of charge or spin excitations often produces such an asymmetric line shape~\cite{Fano1961,Dowty1987}.
The Fano line shape analysis can help us to understand the underlying coupling. Following the Fano theory~\cite{Fano1961}, the optical conductivity $\sigma_{1}(\omega)$ of such coupled system can described by the formula~\cite{Kuzmenko2009}:
\begin{equation}
\label{Fano}
\sigma_{1}(\omega)=\frac{2\pi}{Z_0} \frac{S^2}{\gamma}
\frac{q^2 +\frac{4q (\omega - \omega_0)}{\gamma} -1}{q^2 (1 + \frac{4(\omega - \omega_0)^2}{\gamma^2})}.
\end{equation}
The above Fano line shape description is a modified form of the standard Lorentz oscillator. The asymmetry is described by a dimensionless parameter $q$, where the symmetric Lorentz line shape is completely recovered in the limit $1/q \rightarrow 0$. The parameter $q$ is inversely related to the strength of the interaction. The sign of $q$ determines the energy range that the phonon is coupled. For example, in the case of $q < 0$, a line shape that dips on the high frequency side of the phonon frequency $\omega_0$, indicates the phonon is interacting with a continuum at lower energies.
As shown in Fig.~\ref{fig.3}(a), the sharp phonon mode around 200\icm\ can be well reproduced by replacing the Lorentz oscillator with the Fano oscillator. The Fano fit parameters at 300~K and 5~K are shown in Table~\ref{tab.1}. We can see that this phonon mode shows a strong asymmetry with $q \simeq -5.3$ at 5~K, and presents a weak asymmetry at room temperature ($q \simeq -27.3$) that tends to a Lorentz line shape. The negative sign of $q$ suggests that this phonon mode interacts with a continuum at lower energies. The temperature dependence of the Fano parameters for this phonon mode is shown in Figs.~\ref{fig.3}(b)--\ref{fig.3}(c). From Fig.~\ref{fig.3}(b), one can see that $\omega_0$ increases continuously and the line width $\gamma$ narrows almost linearly with decreasing temperature. Note that due to the Fano line shape, $\omega_0$ is larger than the position of the Lorentz oscillator fit, the difference being dependent on the asymmetry parameter $q$, $\Delta \omega_0 \simeq $ 0.3\icm\ at 300~K and $\Delta \omega_0 \simeq $ 1.1\icm\ at 5~K. The black triangle in Fig.~\ref{fig.3}(c) shows the temperature dependence of the asymmetry parameter $q$. As we can see, the asymmetry effect increases with decreasing temperature and becomes saturated at low temperatures. Meanwhile, as shown by the blue open triangle in Fig.~\ref{fig.3}(c), we can also see a corresponding increase in the lattice mode intensity.
The enhanced asymmetry effect at low temperatures implies an increased spin-lattice or electron-lattice coupling. Nevertheless, the NMR experiments have evidenced that the spin lattice relaxation rate $1/T_1$ in this system is decreased at low temperatures~\cite{Torchetti2011,Yu2011,Kotegawa2011}. Thus, this brings up the possibility of changes to the electronic states in this system. Indeed, our results as well as the previous optical results~\cite{Yuan2012,Homes2012,Charnukha2012} all suggest that the optical conductivity in this system is almost incoherent at room temperature and becomes more coherent at low temperatures. Therefore, the increase in asymmetry can be understood by considering that the coherent electronic states are continuously increased as the temperature is reduced. Such increase of the coherent quasiparticles results in the enhanced coupling of the lattice mode to the electronic states.
\section{Conclusion}
In summary, we performed optical measurements on the superconducting single crystal of Rb$_{0.8}$Fe$_{1.68}$Se$_2$. We found that the optical conductivity is dominated by a series of sharp features associated with the infrared-active phonon modes in the low-frequency region as well as several high-frequency bound excitations. As the temperature is reduced, the phonon modes increase continuously in frequency and a Drude-like low-frequency response emerges, but the fitted value $\Omega_{pD}$ is quite small, indicating low carriers density in this material. Moreover, the phonon mode around 200\icm\ shows a strong asymmetry effect and a corresponding increase in the phonon mode intensity at low temperatures due to the increased coherent electronic states, suggesting a strong electron-phonon coupling in this system. This electron-phonon coupling may play a role in the the appearance of superconductivity in this iron selenides system.
\section*{Acknowledgments}
We thank C. H. Li, and H. H. Wen for providing Rb$_{0.8}$Fe$_{1.68}$Se$_2$\ crystals. This work was supported by MOST (973 Projects No. 2012CB821403, 2011CBA00107, 2012CB921302), and NSFC (Grants No. 11374345, 11104335 and 91121004).
\section*{References}
|
\section{Introduction}
REDUCE\cite{hearn88,maccallum90}, like many modern Computer Algebra Systems (CAS) (MACSYMA, MATHEMATICA, MAPLE, SCRATCHPAD to name a few), embodies a large amount of mathematical knowledge which is spread out through thousands of procedures in the source code of the system. Most of this knowledge is, however, in an {\em implicit} form almost inaccessible to the user, who would have to decipher the source files to recover it. Also, to profit from it one needs to know in addition the capabilities of these computing systems and how to operate them.
The process of learning how to use a CAS may be done by reading throughout the manual available for the system, a sort of user's guide, a book or in the practical {\em hands on} approach. However, soon the begginers get stuck and the most confortable and easiest way to solve the problem is consulting an {\em expert} on the system. They, in general, do not want to waste their time reading either a book or manual looking for the terminological structure of the system. Particularly, to its semantics, syntax and/or examples just to be able to start using the system to solve a simple problem: an integral, for instance. That is, perhaps, the
reason why many the potential users with no previous experience with computers keep avoiding to use them.
These kinds of informations could be easily provided by an
on-line help system (and some progress in this direction have been attained\cite{rand90}) or even better, by an Intelligent Tutorial System\cite{sleeman82}. The idea behind is that existing a stored knowledge base, consisting of a large number of information elements, structured in levels of specialization, an {\em expert help system} could generate by inference (not just by pure recovering) an information, which might not be explicity stored and that might be an adequate (in terms of abstraction, difficulty, detail, etc.) directive to the user. An intelligent help facility could reason about the
user's query and then give a reasonable reply or else, suggest what has to be consulted or even what to do next.
It is our intention here only to discuss the nature, complexity and tools concerning the design of {\em Smart Help}, an expert help system with these features. Presently, the {\em Smart Help} domain knowledge base concerns REDUCE, however it may well be used to implement different CAS bases. Since the hybrid knowledge
representation system (KRS) MANTRA\cite{bittencourt90a} can be seen as a knowledge representation shell, one could make use of its (formal) reasoning facilities to build the {\em Smart Help} system. The ideas forwarded herein are in a prototypal basis, but we hope that they will raise the interest of the CAS community and evolve to reach at the end the full system implementation.
In section \ref{taxonomy} we comment briefly on the computer algebra system REDUCE and propose a taxonomy for the knowledge
embodied by this system. In section \ref{mantra} we describe the knowledge representation system MANTRA. Section \ref{shelp} and 5 are concerned with the {\it Smart Help} system design. Finally, in section 6, we give some conclusions and suggest further upgrades which would turn {\em Smart Help} into a truly {\em Intelligent Tutorial System}\cite{sleeman82}. An example of the interaction is also given. Further details can be found in \cite{dossantos90}.
\section{The CAS REDUCE}
REDUCE\cite{hearn88} is a fairly powerful system for carrying out a variety of mathematical calculations such as to manipulate polynomials in a variety of forms, simplify expressions, to differentiate and integrate algebraic expressions, do some modern differential geometry calculations, study the Lie-symmetries of systems of partial differential equations, and many others.
This system is made out of different knowledge domains (mathematical, terminological, computational, etc.). In order to represent in {\em Smart Help} the specific knowledge domain involved in operating REDUCE we propose to classify it in the following categories:
\label{taxonomy}
\begin{description}
\item[syntax:] Informations about the number and type of the arguments, and requisites and prerequisites of a REDUCE command, declaration or operator (that is the only type of information given by many help systems).
\item[terminology:] Definitions of the various terms like {\em identifier}, {\em operator}, {\em kernel}, etc., employed in documents related to REDUCE and, also important, in error messages from REDUCE (this matter can be quite confusing to the initial user).
\item[concepts:] Informations like the fact that {\em integration}
in REDUCE is represented by an operator called {\bf INT}, or a reference for the algorithm which it implements.
\item[procedures:] General sequences of commands which should be
given to perform a certain task like defining a new infix operator: one has to declare the operator infix through the INFIX declaration and then give him a precedence by means of the PRECEDENCE declaration.
\item[heuristics:] General rules and tips that simplify and improve approaches to problem-solving (this kind of information comes with
experience and is one of the most frequent in consultations from beginners). For example, ``If you want to compute a definite integration, you can try evaluating the indefinite integral, saving the resulting expression in a variable and then substituting locally the limits of integration in it and finally subtracting the results''.
\end{description}
\section{The KRS MANTRA\protect\footnotemark} \footnotetext{This description of the MANTRA System is based on \cite{bittencourt90a}} \label{mantra}
MANTRA\cite{bittencourt90a} is a hybrid
knowledge representation system which integrates three different knowledge representation methods:
\begin{description}
\item[Four-valued first-order logic language,] \qquad used to express {\em assertional knowledge}, which is decidable but presents a weaker entailment mechanism which excludes the chaining of independent facts, thus ruling out {\em modus ponens} but allowing quantifiers.
\item[Terminological language] (a kind of {\em frame} method), extended to allow the definition of concepts and n-place relations over themselves.
\item[Semantic network,] which is a representation to define hierarchies with exceptions.
\end{description}
One could think that classical logic would be able to generate
all knowledge entailed (i.e., implicity represented) by a given concept. Unfortunately, however, the systems based on the complete first-order logic, suffer from the inherent problem of {\em Combinatorial Explosion}. Also, the entailment problem is not decidable in first-order logic. In addition, the knowledge to be represented is often incomplete and incoherent.
Due to these facts, the knowledge representation system MANTRA has been designed associating the three knowledge representation
methods above, based on a common four-valued semantic, which allows the modeling of the aspects of {\em ignorance} and {\em inconsistency}, useful for representing incomplete and/or incoherent knowledge.
\section{The Conception of the {\em Smart Help} Expert System}\label{shelp}
The conception of {\em Smart Help} follows the tradition of help systems being passive. This means that the user learns how to use REDUCE playing freely with it without being interrupted by the {\em Smart Help}. To the user it looks like a normal REDUCE session but MANTRA is running behind and is accessible as an operator, by means of the interface MANTRA-REDUCE. When the user gets a confusing answer or a meaningless error message (and in fact REDUCE users know how often this happens), or even when he does not know what to do next to get his calculations done, he invokes the {\em Smart Help} to clarify the point.
The problem might have been caused by earlier mistakes (a forgotten variable definition long before, for example). An explanation facility\cite{marti84} to trace the session history and find the exact place at which the misconception first had its effect was not included in {\em Smart Help}. It is then left to the user to do the right question to get the right answer. That is, progress can be made only if, from the correct definition, usage, prerequisites, etc., of the queried topic, as returned by the {\em Smart Help}, the user can pinpoint the misconception which caused the problem.
\section{The Architecture of {\em Smart Help}}
Technically, {\em Smart Help} is a Production System on the top of a particular implementation of MANTRA which has REDUCE integrated as an additional knowledge representation module. Since the heuristic level of MANTRA has not yet been implemented, being presently represented by the Lisp language itself, {\em Smart Help} is coded in Lisp and resides in the same Lisp session of MANTRA.
Considering the taxonomy of the knowledge embodied by REDUCE, presented in section \ref{taxonomy} above, and the knowledge representation methods available in MANTRA, as described previously in
section \ref{mantra}, we had to find the best fit of both to guarantee efficiency in recovering knowledge from the base and in reasoning with it, according to the inherent structure of each knowledge category and to its adaptiveness to the specific representation method. The five categories of knowledge were implemented as {\em aspects} of the knowledge base {\em object}, making use of Corbit\cite{desmedt87}, an object-oriented extension of Common Lisp. Details of this can be found in \cite{dossantos90}.
A number of rules were implemented in the production system of {\em Smart Help}. We present them in the following. Presently, they are inserted in the code itself. We consider now to get them defined in a production rule base, what would give flexibility and clearness to our system.
When asked by the user about a topic, {\em Smart Help}
\begin{enumerate}
\item Assumes that the present level of familiarity of the user with REDUCE (student model) can be inferred from the level of specification of the queried topic, which is characterized by a numeric heuristical parameter, ranging from 1 to 3, associated to it. For example if someone queries about ``integration'' (very general -- parameter = 1) it seems probable to be a very novice user but if one queries about ``infix-operators'' (more specialized -- parameter = 2) it should be considered as an user with some familiarity.
\item Queries the domain knowledge base, to recover all informations
related to the given topic and store them in the ``answer'' structure. This process consists in querying the {\em conceptual}, {\em terminological}, {\em syntactical}, {\em procedural} and {\em
heuristical} aspects of the base.
\item Evaluates the level of generality of the terms recovered to see if they are in the same level of specialization (same value of the parameter). If a concept is too specific (much greater value of the parameter), {\em Smart Help} tries to redefine it in terms of more
general concepts. If a concept is too general, however, it is simply deleted.
\item Formats the recovered knowledge in the form of a readable answer, defining the queried concept and its usage in terms of the recovered knowledge. Presently this process is quite crude as it is a peripherical point of the implementation. It will be improved latter on.
\item Prints the answer returning control to REDUCE.
\end{enumerate}
For better understanding, suppose that an user with no experience with REDUCE wants to do some calculations, for instance, integrate an expression in terms of a certain variable. Having access to an initialized session of REDUCE (in which the heading shows how to invoke {\em Smart Help}), the interaction would consist in typing {\tt shelp integration} , and {\em Smart Help} would promptly reply: \\
\begin{verse}
{\tt "SHelp - Version 2.0: 23 Aug 1990"} \\
{\tt INDEFINITE-INTEGRATION is the default for INTEGRATION.} \\
{\tt INTEGRATION is represented in Reduce by INT.} \\
{\tt Its syntax is:} \\
{\tt INT(scalar-expression,variable)} \\
{\tt Ex.: INT(LOG(X),X); } \\
{\tt INT is implemented through the } \\
{\tt SIMPINT-PROCEDURE-IN-INT-SOURCE-FILE.} \\
{\tt For DEFINITE INTEGRATION, one may try} \\
{\tt to DO INDEFINITE-INTEGRATION,} \\
{\tt to SAVE-RESULT-IN VARIABLE,} \\
{\tt to LOCALLY-SUBSTITUTE VALUES.} \\
{\tt References: REDUCE MANUAL SECTION 7.4.} \\
{\tt See also: DERIVATION}
\end{verse}
\section{Conclusions}
We have presented and proposed in this paper a fairly general design of an expert help facility for aiding users of Computer Algebra Systems. Although the expert help system presented here has been particularly oriented to REDUCE (as a consequence of our former experience with this system), we point out that the concept of {\em Smart Help} can be extended to other Computer Algebra Systems.
The reasons for introducing {\em Smart Help} facility include:
\begin{itemize}
\item It will provide an on-line help for the system, aiding the users to find specific informations about the system terminology, structure, syntax, etc.
\item It will allow the potential user to access the whole capabilities of the system.
\item It can contribute to the development of Intelligent Computer Algebra Systems\cite{calmet87b}, which more than simply being able to do calculations, could interact with the user and free him of many details concerning the specification of his problem.
\end{itemize}
The {\em Smart Help} has no intention to {\em teach} the user how to program efficiently in REDUCE or behave like a tutoring system.
However, it can be used in the learning process as a complimentary teaching tool.
Following the ideas addressed in this paper, we intend afterwards to develop an {\em Intelligent Tutorial System}\cite{sleeman82} (ITS) for REDUCE. The ITS should inherit all the compatible facilities already available in the {\em Smart Help}. In addition, many other facilities would become available, such as, a deeper understanding of REDUCE's semantics, keeping track of history, explanations\cite{marti84}, a dynamical reasonning on the student's model\cite{sleeman82}, etc.
A prototype of {\em Smart Help} is now running on a SUN work-station on an experimental basis. The full implementation of {\em Smart Help} as a final product was not our main concern here. This task will certainly need few more people working in a close colaboration to build up a satisfactory knowledge base to reach at the end the principal objective that is helping a CAS user.
\section{Acknowledgement}
We would like to thank Prof. J. Calmet for enlightening discussions and for the warm hospitality provided by him and by all
members of his group and the Conselho Nacional de Desenvolvimento Cient\'\i fico e Tecnol\'ogico -- CNPq for the financial support.
|
\section{Introduction}
\label{sec:intro}
Binary black hole coalescences are amongst the most promising sources of gravitational-wave transients for ground based gravitational-wave observatories such as LIGO and Virgo~\cite{300years,ligo-ref,virgo-ref}. While stellar mass black holes with mass between 2.5~M${}_{\odot}${} and a few tens of M${}_{\odot}${} are formed by stellar collapse~\cite{polytrope,0004-637X-730-2-140,2041-8205-715-2-L138,0004-637X-759-1-52}, intermediate mass black holes between a few tens of M${}_{\odot}${} and $10^{5}$ M${}_{\odot}${} may result from the merger of stellar mass black holes or runaway collision of massive stars in dense globular clusters~\cite{imbh-globular-1,imbh-globular-2,0004-637X-734-2-111}.
No evidence of binary black hole coalescence has been detected so far in data from initial LIGO and Virgo~\cite{inspiral,cbclowmass_s6,cbc-highmass,cbc-highmass-s6,S5_S6_ringdown,imbh,imbh-s6}. However, according to current rate predictions, advanced LIGO~\cite{adv-LIGO} and advanced Virgo~\cite{adv-Virgo} are expected to detect several gravitational-wave signals from binary black hole coalescences~\cite{rate}.
Following the seminal work of~\cite{flanhughes}, LIGO and Virgo data have been searched for separate phases of the binary black hole coalescence: inspiral, merger and ringdown~\cite{inspiral,cbclowmass_s6,imbh,ringdown}.
Search methods have evolved to account for full inspiral-merger-ringdown waveform templates~\cite{cbc-highmass}, use signal based vetoes~\cite{chi-square-bruce}, employ multi-resolution time-frequency information~\cite{multi-resolution}, and utilize coincident and coherent methods~\cite{robinson2008,ringdown,cwb}.
This paper introduces a framework to compare different searches in real detector noise, and applies it to algorithms representative of those used in recent searches for binary black holes in LIGO and Virgo data~\cite{cbc-highmass,cbc-highmass-s6,cwb,ringdown,S5_S6_ringdown,imbh-s6}.
Section~\ref{sec:noise}, describes the data used in this study. Section~\ref{sec:algorithm} introduces the search algorithms that are compared in this study. Section~\ref{sec:method} outlines the method used in the a comparison, and Section~\ref{sec:search-performance} presents the results, using three complementary figures of merit.
\section{Detector Data}
\label{sec:noise}
We conducted this study on two months of data from the $5^{\mathrm{th}}$ science run LIGO (14 Aug to 30 Sept 2007), when initial LIGO was at design sensitivity~\cite{s5-sensitivity}. We considered the three-detectors network of the 4-km and 2-km Hanford (H1, H2)~\cite{ligo-wa}, and 4-km Livingston (L1)~\cite{ligo-la} observatories.
%
Fig.~\ref{fig:psd} shows the detectors' sensitivity, expressed as strain amplitude spectral density (ASD). The color shaded region indicates the $5^{\mathrm{th}}$ to $95^{\mathrm{th}}$ ASD percentiles in the analyzed period.
\begin{figure}[b]
\centering
\subfloat{\includegraphics[width=0.5\textwidth]{figures/quantile_all_spectra}}
\caption{ Sensitivity of the detectors during the period of data in this study.
The color shaded region indicates the $5^{\mathrm{th}}$ to $95^{\mathrm{th}}$ amplitude spectral density percentiles.}
\label{fig:psd}
\end{figure}
Data below 40 Hz is not included in this analysis, since it is limited by the seismic noise and, therefore, is not calibrated~\cite{calibration}.
We applied data selections and vetoes to account for environmental artifacts and instrumental glitches that affect the quality and reliability of the data; this, combined with the instruments' duty cycle, resulted in a 3-detector lifetime of 29 days. See~\cite{s5-glitch,first_joint_ligo_geo_virgo_burst_s5} for details of this procedure.
\section{Search algorithms}
\label{sec:algorithm}
Transient gravitational-wave searches can be broadly classified into matched filtering with templates and unmodeled searches which do not assume a specific signal model.
Matched filtering can be performed with templates for the full binary black hole coalescence or only a portion of it. Unmodeled searches look for statistically significant excess signal energy in gravitational wave data.
In this paper we consider three algorithms that have been adopted in searches of LIGO-Virgo data: the IMR-templates search~\cite{cbc-highmass,cbc-highmass-s6}, the ringdown search~\cite{ringdown,S5_S6_ringdown}, and coherent WaveBurst~\cite{imbh,imbh-s6}.
\paragraph{IMR-templates ---}
\label{sec:highmass}
Matched filtering to a bank of non-spinning EOBNR templates~\cite{eobnr} has been used to search for binary black hole mergers in 25--100~M${}_{\odot}$~\cite{cbc-highmass,cbc-highmass-s6}. Matched filtering is optimal for weak signals in Gaussian noise. However, due to the non-stationary, non-Gaussian nature of LIGO-Virgo data, this search is augmented with a sophisticated inter-site coincidence test~\cite{robinson2008}, a time-frequency, $\chi^{2}$ signal consistency test~\cite{chi-square-bruce} and a combined false alarm rate event ranking, all of which is described in Section 3 of~\cite{cbc-highmass}.
\paragraph{Ringdown search ---}
\label{sec:ringdown}
The post-merger signal is accurately described as a superposition of quasi-normal oscillation modes, the {\it ringdown} waveforms~\cite{Teukolsky-1973ha}. Searches for ringdowns utilize a matched filtering algorithm with damped sinusoid templates characterized by quality factor $Q$ and central frequency $f_0$, parameters that describe the $\ell=m=2$ oscillation mode of the final merged black hole~\cite{creighton-ring-template,ringdown,S5_S6_ringdown}.
Details on the template bank and the algorithm are given in~\cite{S5_S6_ringdown}. The central frequency and quality factor can be empirically related to the final mass and spin of the merged black hole~\cite{berti_2006}; the search space corresponds roughly to black holes with masses in the range 10 M${}_{\odot}${} to 600 M${}_{\odot}${} and spins in the range 0 to 0.99.
The ringdown matched filter search is also augmented with a sophisticated multi-detector coincidence test~\cite{robinson2008,nakano2003}, an amplitude consistency test~\cite{S5_S6_ringdown}, and an event ranking based on the {\it chopped-L statistic} described in~\cite{talukder2003}. This event ranking statistic differs from the multivariate statistic used in~\cite{S5_S6_ringdown}. Candidate events are ranked
with the combined false alarm rate detection statistic described in~\cite{cbc-highmass,S5_S6_ringdown}.
\paragraph{Coherent WaveBurst ---}
\label{sec:cwb}
The coherent WaveBurst (CWB) algorithm is designed to identify coherent excess power transients without prior knowledge of the waveform, and it is used in searches for gravitational-wave bursts in LIGO and Virgo data~\cite{first_joint_ligo_geo_virgo_burst_s5,second_joint_ligo_virgo_burst_s6}.
By imposing weak model constraints, such as the requirement of elliptical polarization, CWB has been optimized
to search for black hole binaries of total masses between 100 -- 450 M${}_{\odot}$~\cite{imbh,imbh-s6}.
The algorithm uses a wavelet basis to decompose the data into a time-frequency map with discrete signal energy pixels.
The algorithm then executes a constrained maximum likelihood analysis of the decomposed network data stream. Reconstruction of detector responses occurs for a trial set of sky locations and corresponding arrival delays. The residual data, after subtraction of the estimated detector responses from the original data, represents the reconstructed network noise.
The elliptical polarization constraint
is expected to have minimal impact on the recovery of gravitational-wave signals from compact binary coalescences, while enhancing the rejection of noisy events~\cite{elliptical-polarization}.
The coherent network amplitude $\eta$ is the detection statistics used by CWB~\cite{imbh,imbh-s6}. It is proportional to the average signal to noise ratio (SNR) per detector and is used to rank selected events and establish their significance against a sample of background events.
\section{Comparison Method}
\label{sec:method}
Since the algorithms described in section~\ref{sec:algorithm} have different targets, it is natural to expect they respond differently to weak signals in the data and to noise transients.
In this paper we compare the sensitivity of the three searches via their false alarm rate, a search independent ranking of an event's rate of occurrence, as determined from a background sample.
We ran the analyses with configurations that are representative of their applications in published results~\cite{cbc-highmass,cbc-highmass-s6,ringdown,S5_S6_ringdown,imbh,imbh-s6}.
For this analysis, we injected in the detectors' noise simulated gravitational-wave signals from the coalescence of binary black holes.
The waveforms were produced with the IMRPhenomB~\cite{spin-phenom} model: a phenomenological, non-precessing, spinning binary black hole template family, which tracks the coalescence from late inspiral to ringdown.
In this configuration, the spin vectors ($\chi_{1}$ and $\chi_{2}$) are aligned/anti-aligned with the angular momentum of the binary system. The waveforms are parametrized by three physical parameters: the component black hole masses $m_{1}$, $m_{2}$, and the mass weighted spin parameter $\chi_{s}$,
\begin{equation}
\chi_{s}=\frac{m_{1} \chi_{1} + m_{2} \chi_{2}}{m_{1}+m_{2}}.
\label{chi_s}
\end{equation}
The waveforms do not include the effects of non-aligned spin-orbit coupling, but do account for aligned / anti-aligned spin-orbit interaction, such as the orbital hang-up effect~\cite{orbital-hangup}.
To determine the false alarm rate, we time shifted the data from one or more detectors well beyond the light travel time of 10 ms between H1 and L1. We imposed a minimum shift of 5 seconds to remove inter-site correlations which could be due to a real gravitational-wave signal in the data. We did not introduce time shifts between data from the co-located H1 and H2 detectors, since the background should account for site-specific correlations~\cite{cbc-highmass}.
We applied 100 equally spaced time shift in the ringdown and the IMR-templates searches, and 600 in the CWB search.
We declared an injection {\em detected} if a coincident event was identified within 100 ms of the nominal injection time. This interval is long enough to account for the uncertainty in identifying the {\em arrival time} of a signal, where the arrival time is the maximum amplitude of the waveform.
Each pipeline ranked all the events and assigned a false alarm rate by comparison with its native background ranking statistics.
We evaluated detection efficiency and sensitive distance (as defined in section~\ref{sec:search-performance}) for a range of measurable false alarm rate thresholds. We quote the results for a false alarm rate threshold of 3 events per year which is in the middle of this range. We made sure that the searches use consistent data after the application of data quality vetoes, with small differences due to technical details in the veto implementation~\cite{s5-glitch}.
\section{Results}
\label{sec:search-performance}
\subsection{Target parameter space}
In this study, we partitioned the parameter space according to the total mass of the binary system.
\textit{Set A} includes systems with total mass between 25 and 100 M${}_{\odot}$, as searched by IMR-templates~\cite{cbc-highmass,cbc-highmass-s6}. \textit{Set B} consists of total mass between 100 and 350 M${}_{\odot}$, which overlaps the parameter space searched by the CWB algorithm~\cite{imbh,imbh-s6}.
We restricted the total mass to below 350 M${}_{\odot}${} as the peak detectable frequencies from the ringdown for some of the spin configuration is below 40 Hz for mass above 350 M${}_{\odot}$~\cite{spin-phenom,final-spin}, and thus is subject to unacceptable or ill-defined uncertainties arising from calibration~\cite{calibration}.
Simulated signals are uniformly distributed in total mass ($m$), mass ratio ($q$), and dimensionless spin parameter $\chi_{s}$, in the intervals listed in Table~\ref{tbl:inj_param}.
This distribution is not meant to reproduce the expected astrophysical distribution of binary black hole sources, but rather to probe a wide physical parameter space and evaluate the efficacy of each pipeline in detection. The injections are logarithmically distributed in distance. No correction to the waveform due to redshift at cosmological distances is included, as this effect is expected to be small ($z$ $<=$ 0.1) at the reach of initial detectors.
Injections are also uniformly distributed in sky location, polarization and inclination of the binary relative to Earth.
We analyzed $\sim$25000
injections; due to the limitations of the search (i.e. reduced efficacy of $\chi^{2}$ above 100 M${}_{\odot}${} as discussed in section~\ref{sec:algorithm}), we used
the IMR-templates search only for the lower mass set of injections. We performed ringdown matched filter and CWB analyses on both injection sets.
\begin{table}[htbp]
\begin{center}
\caption{Simulated waveform parameters.}
\begin{tabular}{lr}
\hline
Total Mass (M${}_{\odot}$), $m$: Set A & 25 -- 100 \\
Total Mass (M${}_{\odot}$), $m$: Set B & 100 -- 350 \\
Mass Ratio (both sets), $q$: & 0.1 -- 1 \\
$\chi_{s}$ (both sets) & -0.85 -- 0.85 \\
Distance (Mpc) (both sets) & 0 -- 2000 \\
\hline
\end{tabular}
\label{tbl:inj_param}
\end{center}
\end{table}
Fig.~\ref{fig:m_q-average-range} and~\ref{fig:m_chi-average-range} show the \emph{expected range} as a function of total mass, mass ratio and spin parameter. The expected range is calculated by averaging the distances over extrinsic parameters such as sky position and inclination of the binary black holes for which the network signal-to-noise ratio (SNR) is 12, following the prescription used in~\cite{obs-scenario}.
The SNR is estimated from the median value of the amplitude spectral density of the instrumental noise in Fig.~\ref{fig:psd}.
Fig.~\ref{fig:m_q-average-range} illustrates that
the expected range is higher for symmetric mass binary black holes compared to asymmetric mass system for the same total mass.
This is consistent with the fact that the SNR of the signal is proportional to its amplitude divided by the square-root of its duration in time. For a binary black hole, the gravitational-wave amplitude is proportional to $\dfrac{q}{(1+q)^2}$, while the time duration is proportional to $\dfrac{(1+q)^{2}}{q}$, hence SNR is proportional to $\dfrac{\sqrt{q}}{1+q}$~\cite{0264-9381-24-12-S04}.
Fig.~\ref{fig:m_chi-average-range} illustrates that the expected range is higher for aligned than anti-aligned spin configurations, since systems with aligned spins stay longer in orbit until merger, hence get more relativistic, leading to higher gravitational-wave luminosity, than anti-aligned systems~\cite{orbital-hangup}.
\begin{figure}[h]
\centering
\subfloat{\includegraphics[width=0.4\textwidth]{figures/m_q_expected_range_color_setA} \label{fig:m_q-average-range-IMR-a}}\\
\subfloat{\includegraphics[width=0.4\textwidth]{figures/m_q_expected_range_color_setB} \label{fig:m_q-average-range-IMR-b}}\\
\caption{Expected range (Mpc) as a function of $m$ vs $q$. This quantity is estimated from the network SNR threshold of 12, and is marginalized over the spin parameter, sky position and orientation of the binary system. The color scale is saturated at 200 Mpc for comparison with Figs.~\ref{fig:sensitive-distance-IMR-setA-m-q},~\ref{fig:sensitive-distance-IMR-setA-m-chi},~\ref{fig:sensitive-distance-IMR-setB-m-q},~\ref{fig:sensitive-distance-IMR-setB-m-chi}.}
\label{fig:m_q-average-range}
\end{figure}
\begin{figure}[h]
\centering
{\includegraphics[width=0.4\textwidth]{figures/m_chi_expected_range_color_setA} \label{fig:m_chi-average-range-IMR-a}}\\
{\includegraphics[width=0.4\textwidth]{figures/m_chi_expected_range_color_setB} \label{fig:m_chi-average-range-IMR-b}}\\
\caption{Expected range (Mpc) as a function of $m$ $\chi$. This quantity is estimated at a network SNR threshold of 12, and is marginalized over mass ratio, sky position and orientation of the black hole binary system.}
\label{fig:m_chi-average-range}
\end{figure}
\subsection{Search performance}
We now define the quantities that will be used for the comparison.
The detection efficiency $\varepsilon$ of a search is a function of the false alarm rate threshold
$\zeta$, the radial distance to the source $r$, the total mass $m$, the mass ratio $q$, and the spin parameter $\chi_{s}$.
In this work, we average over sky location, polarization and orientation, and define the average efficiency as:
\begin{equation}
\bar{\varepsilon}(\zeta,r,m,q,\chi_{s}) = \frac{N_{f}}{N_{i}},
\label{eff-eqn}
\end{equation}
where $N_{f}$ is the number of found injections and $N_{i}$ is the number of total injections averaged over all sky position and inclination.
The sensitive volume, or the volume of the sky surveyed is defined as:
\begin{equation}
V(\zeta,m,q,\chi_{s}) = \int 4 \pi r^{2} \bar{\varepsilon}\,dr.
\label{sensvol-eqn}
\end{equation}
Finally, the sensitive radius $\mathcal{R}$ is the radius of the sphere with volume of $V$:
\begin{equation}
\mathcal{R}(\zeta,m,q,\chi_{s}) = \left[ \frac{3}{4\pi}V
\right] ^{1/3}.
\label{sensdist-eqn}
\end{equation}
\paragraph*{Efficiency curves ---}
\label{sec:sig}
The detection efficiency at fixed false alarm rate and distance is estimated from Eqn.~\ref{eff-eqn}; we plot it as a function of distance in Fig.~\ref{fig:sigmoid-IMR-algos}, where $N_i$ and $N_f$ are marginalized over all other source parameters. The horizontal and the vertical error bars in Fig.~\ref{fig:sigmoid-IMR-algos} are set by the bin boundaries and binomial statistics on the number of injected signals in each amplitude bin, respectively.
For a quantitative comparison we fit a cubic spline to the efficiency curve. We then compare two characteristic parameters, the 50\% and 90\% efficiency distances, $D_{\mathrm{eff}}^{50\%}$ and $D_{\mathrm{eff}}^{90\%}$, which are the distances at which 50\% and 90\% of the signals can be found, respectively in Table~\ref{tab:compare_searches_eff_at_far}.
In the 25--100 M${}_{\odot}${} range (set A) the IMR-templates search yields 12\% and 25\% higher $D_{\mathrm{eff}}^{50\%}$, and 7\% and 30\% higher $D_{\mathrm{eff}}^{90\%}$ compared to the CWB and the ringdown searches, respectively. In the 100--350 M${}_{\odot}${} range (set B), CWB search 60\% higher $D_{\mathrm{eff}}^{50\%}$ and 170\% higher $D_{\mathrm{eff}}^{90\%}$ compared to the ringdown search.
\begin{figure}
\centering
\subfloat[Part 1][Total mass 25--100 M${}_{\odot}$.]{\includegraphics[width=0.4\textwidth]{figures/Dist_sigmoid_setA_nointerp} \label{fig:sigmoid-IMR-algos-a}}\\
\subfloat[Part 3][Total mass 100--350 M${}_{\odot}$.]{\includegraphics[width=0.4\textwidth]{figures/Dist_sigmoid_setB_nointerp} \label{fig:sigmoid-IMR-algos-c}}
\caption{Efficiency curves at a false alarm rate of 3 events per year, averaged over total mass, mass ratio, spin parameter, sky-position and inclination of the source.}
\label{fig:sigmoid-IMR-algos}
\end{figure}
\begin{table}[hb]
\begin{center}
\caption{Efficiency distances at a false alarm rate of 3 events per year. }
\label{tab:compare_searches_eff_at_far}
\begin{tabular}{ccc}
\colrule
Algorithm (on Set A) & $D^{50\%}_{\mathrm{eff}}$ (Mpc) & $D^{90\%}_{\mathrm{eff}}$ (Mpc) \\
\colrule
IMR-templates search & 94 & 30 \\
CWB & 84 & 28 \\
Ringdown & 75 & 23 \\
\end{tabular}
\begin{tabular}{ccc}
\colrule
Algorithm (on Set B) & $D^{50\%}_{\mathrm{eff}}$ (Mpc) & $D^{90\%}_{\mathrm{eff}}$ (Mpc) \\
\colrule
CWB & 97 & 24 \\
Ringdown & 60 & 9 \\
\colrule
\end{tabular}
\end{center}
\end{table}
\paragraph*{Mean sensitive distance ---}
\label{sec:roc}
Eqn.~\ref{sensdist-eqn} can be marginalized over all parameters, to compute a mean sensitive distance as a function of the false alarm rate. This quantity is show in Fig.~\ref{fig:mean-sensitive-distance-IMR}.
We notice that within a range of false alarm rates of 0.3 events per year to 30 events per year, the three searches give consistent results. We do not see any abrupt change of sensitivity for a search over this range of false alarm rates. The false alarm rate of 3 events/year we chose to plot efficiency curves and quote sensitive distances is thus representative of the relative performance of the algorithms.
Published searches have typically chosen lower FAR thresholds estimated on the loudest events~\cite{loudest1,loudest2} seen by the searches in open-box data: 0.2 events/year and 0.41 events/year for IMR-templates searches on 2005-2007~\cite{cbc-highmass} and 2009-2010~\cite{cbc-highmass-s6} LIGO-Virgo data, 0.45 events/year for the ringdown search~\cite{S5_S6_ringdown}, and 0.76 events/year for the CWB search on 2005-2007 data~\cite{imbh}.
\begin{figure}[h]
\centering
\subfloat[Part 1][Total mass 25--100 M${}_{\odot}$.]{\includegraphics[width=0.4\textwidth]{figures/mean_sensitive_distance_setA_simpler_unit} \label{fig:mean-sensitivity-distance-IMR-a}}\\
\subfloat[Part 2][Total mass 100--350 M${}_{\odot}$.]{\includegraphics[width=0.4\textwidth]{figures/mean_sensitive_distance_setB_simpler_unit} \label{fig:mean-sensitivity-distance-IMR-b}}
\caption{Mean sensitive distance as a function of false alarm rate.}
\label{fig:mean-sensitive-distance-IMR}
\end{figure}
\paragraph*{Sensitive distance ---}
\label{sec:sensitive}
To probe how different algorithms respond to different regions of the parameter space, we plot the sensitive distance, defined in Eqn.~\ref{sensdist-eqn}, as a function of the mass and spin parameters at a false alarm rate of 3 events per year. Additional sensitive distance plots in between 0.3 to 30 events per year are available at~\cite{science-summary}.
\begin{figure}
\centering
\subfloat[Part 1][Matched filter to IMR-templates.]{\includegraphics[width=0.4\textwidth]{figures/m_q_range_color_cbc_setA} \label{fig:m_q-sensitivity-distance-IMR-a}}\\
\subfloat[Part 2][Coherent WaveBurst template-less search.]{\includegraphics[width=0.4\textwidth]{figures/m_q_range_color_cwb_setA} \label{fig:m_q-sensitivity-distance-IMR-b}}\\
\subfloat[Part 3][Matched filter to ringdowns.]{\includegraphics[width=0.4\textwidth]{figures/m_q_range_color_ringdown_setA} \label{fig:m_q-sensitivity-distance-IMR-c}}\\
\caption{Sensitive distance for systems with total mass 25--100 M${}_{\odot}${} as a function of total mass $m$ and mass ratio $q$ at a false alarm rate of 3 events per year.}
\label{fig:sensitive-distance-IMR-setA-m-q}
\end{figure}
\begin{figure}
\centering
\subfloat[Part 7][Matched filter to IMR-templates.]{\includegraphics[width=0.4\textwidth]{figures/m_chi_range_color_cbc_setA} \label{fig:m_q-sensitivity-distance-IMR-f}}\\
\subfloat[Part 8][Coherent WaveBurst template-less search.]{\includegraphics[width=0.4\textwidth]{figures/m_chi_range_color_cwb_setA} \label{fig:m_q-sensitivity-distance-IMR-e}}\\
\subfloat[Part 9][Matched filter to ringdowns.]{\includegraphics[width=0.4\textwidth]{figures/m_chi_range_color_ringdown_setA} \label{fig:m_q-sensitivity-distance-IMR-g}}\\
\caption{Sensitive distance for systems with total mass 25--100 M${}_{\odot}${} as a function of total mass $m$ and mass ratio $\chi_{s}$ at a false alarm rate of 3 events per year.}
\label{fig:sensitive-distance-IMR-setA-m-chi}
\end{figure}
\begin{figure}
\centering
\subfloat[Part 4][IMR-templates and Coherent WaveBurst.]{\includegraphics[width=0.4\textwidth]{figures/m_q_volume_CBC-Highmass_CWB_ratio_setA} \label{fig:m_q-sensitivity-distance-IMR-d}}\\
\subfloat[Part 5][IMR-templates and Ringdown.]{\includegraphics[width=0.4\textwidth]{figures/m_q_volume_CBC-Highmass_Ringdown_ratio_setA} \label{fig:m_q-sensitivity-distance-IMR-b}}\\
\caption{Sensitive volume ratio for systems with total mass 25--100 M${}_{\odot}${} as a function of $m$ and $q$ at a false alarm rate of 3 events per year.}
\label{fig:sensitive-distance-difference-IMR-setA-m-q}
\end{figure}
\begin{figure}
\centering
\subfloat[Part 10][IMR-templates and Coherent WaveBurst.]{\includegraphics[width=0.4\textwidth]{figures/m_chi_volume_CBC-Highmass_CWB_ratio_setA} \label{fig:m_q-sensitivity-distance-IMR-h}}\\
\subfloat[Part 11][IMR-templates and Ringdown.]{\includegraphics[width=0.4\textwidth]{figures/m_chi_volume_CBC-Highmass_Ringdown_ratio_setA} \label{fig:m_q-sensitivity-distance-IMR-i}}
\caption{Sensitive volume ratio for systems with total mass 25--100 M${}_{\odot}${} as a function of $m$ vs $\chi_{s}$ at a false alarm rate of 3 events per year.}
\label{fig:sensitive-distance-difference-IMR-setA-m-chi}
\end{figure}
\begin{figure}[h]
\centering
\subfloat[Part 1][Coherent WaveBurst template-less search.]{\includegraphics[width=0.4\textwidth]{figures/m_q_range_color_cwb_setB} \label{fig:m_q-sensitivity-distance-IMR-k}}\\
\subfloat[Part 2][Matched filter to ringdowns.]{\includegraphics[width=0.4\textwidth]{figures/m_q_range_color_ringdown_setB} \label{fig:m_q-sensitivity-distance-IMR-l}}\\
\caption{Sensitive distance for systems with total mass 100--350 M${}_{\odot}${} as a function of $m$ and $q$ at a false alarm rate of 3 events per year.}
\label{fig:sensitive-distance-IMR-setB-m-q}
\end{figure}
\begin{figure}[h]
\centering
\subfloat[Part 1][Coherent WaveBurst template-less search.]{\includegraphics[width=0.4\textwidth]{figures/m_chi_range_color_cwb_setB} \label{fig:m_q-sensitivity-distance-IMR-n}}\\
\subfloat[Part 2][Matched filter to ringdowns.]{\includegraphics[width=0.4\textwidth]{figures/m_chi_range_color_ringdown_setB} \label{fig:m_q-sensitivity-distance-IMR-o}}\\
\caption{Sensitive distance for systems with total mass 100--350 M${}_{\odot}${} as a function of $m$ and $\chi_{s}$ at a false alarm rate of 3 events per year.}
\label{fig:sensitive-distance-IMR-setB-m-chi}
\end{figure}
Fig.~\ref{fig:sensitive-distance-IMR-setA-m-q} and~\ref{fig:sensitive-distance-IMR-setB-m-q} show the sensitive distance at FAR threshold of 3 events per year as a function of mass parameters, $m$ and $q$, and are marginalized over the spin parameter $\chi_{s}$. Across the mass ranges,
the three search algorithms are more sensitive to symmetric than asymmetric binary systems.
For set A all the three search algorithms have the highest sensitive distance in the total mass bin of 80 to 100 M${}_{\odot}$. For set B the CWB search and the ringdown search register higher sensitive distance in the total mass bin of 100 to 150 M${}_{\odot}$.
Fig.~\ref{fig:sensitive-distance-IMR-setA-m-chi} and~\ref{fig:sensitive-distance-IMR-setB-m-chi} show the sensitive distance as a function of total mass and the spin parameter, and are marginalized over mass ratio, $q$.
Across the mass ranges, the three search algorithms have higher sensitivity for detecting aligned (with respected to the orbital angular momentum) spin binary black holes compared to the anti-aligned spin binary black holes~\cite{cbc-highmass-s6,imbh-s6,ninja-mdc}. Qualitatively, this observation is in consistent with the expected range in Fig.~\ref{fig:m_chi-average-range}. Quantitatively, sensitive distance differ from the expected range, which motivates the choice of false alarm rate as the criteria for detectability, rather than the signal SNR.
The errors quoted in Fig.~\ref{fig:sensitive-distance-IMR-setA-m-q},~\ref{fig:sensitive-distance-IMR-setA-m-chi},~\ref{fig:sensitive-distance-IMR-setB-m-q} and ~\ref{fig:sensitive-distance-IMR-setB-m-chi} are derived from the binomial statistics of events in each bin.
\begin{figure}
\centering
\subfloat[Part 3][Coherent WaveBurst and Ringdown.]{\includegraphics[width=0.4\textwidth]{figures/m_q_volume_CWB_Ringdown_ratio_setB} \label{fig:m_q-sensitivity-distance-IMR-m}}\\
\subfloat[Part 6][Coherent WaveBurst and Ringdown.]{\includegraphics[width=0.4\textwidth]{figures/m_chi_volume_CWB_Ringdown_ratio_setB} \label{fig:m_q-sensitivity-distance-IMR-p}}
\caption{Sensitive volume ratio for systems with total mass 100--350 M${}_{\odot}${} as a function of $m$ and $q$, and $m$ and $\chi_{s}$ at a false alarm rate of 3 events per year.}
\label{fig:sensitive-distance-difference-IMR-setB}
\end{figure}
Fig.~\ref{fig:sensitive-distance-difference-IMR-setA-m-q} shows the ratio in the sensitive volume, defined in Eq~\ref{sensvol-eqn}, for Set A injections in total mass and mass ratio, and Fig.~\ref{fig:sensitive-distance-difference-IMR-setA-m-chi} shows that ratio for total mass and spin parameter space. In these plots the significance, \textit{i.e} $\sigma$ deviation, of the sensitive volume ratio with respect to the associated errors, is shown in each bins. For this set the IMR-templates search shows higher sensitivity compared to the ringdown search and the CWB search across the parameter space. The higher sensitivity of the IMR-templates search is significantly more with respect to the ringdown search in comparison with the CWB search. The significance is higher for binary black holes with spin aligned to the angular momentum.
CWB search has more sensitivity compared to Ringdown search for this mass range, albeit within 2 $\sigma$ deviation only.
Fig.~\ref{fig:sensitive-distance-difference-IMR-setB} show the ratio of the sensitive volume for set B injections in total mass, mass ratio and total mass, spin parameter space. Across the parameter space CWB is more sensitive than the ringdown search albeit with varying degree of significance.
\section{Conclusion}
This paper introduces a framework for comparing searches for binary black hole coalescences in ground-based gravitational wave detectors, and applies it to algorithms used in recently published searches~\cite{cbc-highmass,cbc-highmass-s6,imbh,imbh-s6,S5_S6_ringdown} on a 2 months segment of initial LIGO data.
The codes developed for this analysis are available in the LIGO-Scientific Collaboration Algorithm Library software packages~\cite{lal}.
This is the first one-on-one comparison of searches for binary black hole coalescences. The false alarm rate of 3 events per year is a prefatory choice for the demonstration of the method. This analysis provides groundwork for future work which will include comparison of algorithms at detection-level false alarm rate.
We provide a quantitative measure of how the algorithms tested in this study were more sensitive to symmetric than asymmetric systems, and to aligned rather than anti-aligned binary black holes. Additionally, we probe the different sensitivity of the algorithms to different region of the black hole binary parameter space.
We find that matched filtering search algorithm using the full Inspiral-Merger-Ringdown templates is the most sensitive search algorithm for total mass between 25 and 100 M${}_{\odot}$: 6\% more than a morphology-independent excess power search and 17\% more than matched filtering to ringdown templates, by the measurement of sensitive distance averaged over source parameters at a false alarm rate of 3 events/year. A fully coherent gravitational-wave burst search algorithm is 21\% more sensitive compared to the matched filtering search algorithm with ringdown templates, by the measurement of sensitive distance averaged over source parameters at a false alarm rate of 3 events/year.
This was a non-blind analysis, as the characteristics of simulated signals were known \emph{a priori}. The ability to detect binary black hole coalescence signals through different search algorithms in a blind analysis has been shown in a different study~\cite{ninja-mdc}.
Also, this work does not attempt to combine information from searches. A likelihood based method to combine multiple searches has been prescribed in~\cite{combine-searches}.
In this study have relied on non-precessing phenomenological waveform model. As new template families with wider coverage of the parameter space are now becoming available~\cite{ninja2catalog,eobnrHM,phenomP}, the method prescribed in this paper can readily be extended to them.
We also note the analysis presented in this paper relies on the initial LIGO sensitivity. We expect differences will ensue with the different possible noise spectra expected in advanced LIGO~\cite{aLIGO-noise,harry-aligo,aligo-imbh} and advanced Virgo~\cite{aVirgo-noise}. Additionally, we do not expect the population of non-Gaussian noise transients to be the same in advanced LIGO/Virgo as in initial LIGO/Virgo data.
New algorithms and pipeline improvements are currently under development; for instance, the inclusion of spinning templates in the bank for IMR-templates~\cite{sbank_highmass}, different clustering in the CWB search~\cite{cwb-doc}, and a more sophisticated post-processing for the ringdown search~\cite{S5_S6_ringdown}; the relative performance of the searches may need to be re-assessed once such developments have stabilized.
\section{Acknowledgment}
\input{acknowledgement}
\\
\\
This document has been assigned LIGO laboratory document number P1100198.
\\
\\
\bibliographystyle{unsrt}
|
\section{Introduction, outline of methods and main results.}
\label{sec.int}
Jaeger's directed cycle double cover conjecture~\cite{Jaeger19851} asserts that
for every 2-connected graph $G$ there exists a family of cycles $\mathcal{C}$ of $G$
such that it is possible to prescribe an orientation to each cycle of $\mathcal{C}$
in such a way that each edge $e$ of $G$ belongs to exactly two cycles of $\mathcal{C}$ and
these cycles induce opposite orientations on~$e$.
Jaeger's conjecture trivially
holds in the class of cubic bridgeless planar graphs.
And it has certainly been positively settled for some more classes of graphs,
as for example, graphs that admit a nowhere-zero
4-flow~\cite{MR1395462}
and 2-connected projective-planar graphs~\cite{Ellingham:2011:OEO:1953656.1954213}.
We kindly invite the reader interested in more details about the development of the directed
cycle double cover conjecture and related problems to consult~\cite{Jaeger19851,zhang1997integer,zhang2012circuit}.
A \emph{reduction network} is a communication network consisting of interconnected parts where each part has to perform some task.
The elements of each part need to communicate in order to successfully perform the task.
The parts are linearly ordered, and after a part performs its task, the part is no more directly functional and it is \emph{reduced} --- that is, the communication network is updated in such a way that the reduced part is removed and only its residue remains.
The goal is to make {\em reductions} whose residue help communication in yet functional parts.
The reduction network is modeled by an undirected graph $G$,
and its linearly ordered parts by the ears of an ear-decomposition of $G$ --- the order of the parts is reversed order of the ears.
At the initial step, the first part is \emph{reduced} and the updated communication network becomes
a \emph{mixed graph}; that is, the union of a subgraph of $G$ and the residues of the the initial reduction.
At each step, the currently active part
is \emph{reduced} if its \emph{correct reduction} exists.
There are multiple obstructions to the existence of \emph{correct reductions},
one of them are the \emph{cut-obstacles}.
In this work, we investigate reduction networks that are modeled by {\em robust trigraphs}.
\smallskip
In what follows, we formalize this discussion.
In order to make this paper as self-contained as possible, in the next paragraphs
we repeat some definitions of~\cite{ourwork}.
\subsubsection*{Robust trigraphs}
In this work, an \emph{ear} is a path on at least 3 vertices or a star on 4 vertices; in particular, an edge is not an ear.
A \emph{trigraph} is a cubic graph that can be obtained
from a cycle by sequentially adding \emph{short ears}, that is, ears on
at most~5 vertices.
Whenever $H$ is a trigraph, an ear decomposition of $H$ means a short-ear decomposition of $H$
and the expression $(H_0, H_i, L_i)^n$ stands for such an ear decomposition;
that is, where $H_0$ is the initial cycle of the ear decomposition, $L_i$ is the $i$-th short ear,
$H_i$ is the intermediate graph obtained from $H_0$ by adding the first $i$ short ears
and $H_n =H$.
For each ear $L$, we denote by $I(L)$ the set of its vertices of degree at least 2
and if $L$ is a path we say that $L$ is a $k$-ear if the cardinality of $I(L)$ is $k$.
Let $(H_0, H_i, L_i)^n$ be an ear decomposition of $H$.
The {\em descendant} of $I(L_i)$, where $L_i$ is a 3-ear,
is the maximal subgraph $D$ of $H-V(H_i)$ such that $H[I(L_i) \cup V(D)] $ is connected;
the notation $H[X]$ represents the induced subgraph of $H$ on vertex set $X \subset V(H)$.
In other words, $D$ is the descendant of $I(L_i)$
if and only if $D$ is the maximal subgraph of $H-V(H_i)$
such that for each component $C$ of $D$ there exists an edge
connecting $C$ and $I(L_i)$.
\begin{definition}[Robust]
\label{def.wr}
Let $(H_0, H_i, L_i)^n$ be an ear decomposition of a trigraph $H$.
We say that $(H_0, H_i, L_i)^n$ is robust if for each 3-ear, say $L_i$,
the descendant of $I(L_i)$ is composed of at most 2 connected components.
Moreover, if the descendant is composed of two connected components, then one of them is an isolated vertex
adjacent to two vertices in $V(H_0)$.
\end{definition}
\subsubsection*{Mixed graphs}
A {\em mixed graph} is a 4-tuple $(V, E, A, R)$, where $V$ is a vertex set,
$E$ is an edge set, $A$ is a set of directed edges (arcs), and $R$ is a subset of $A{\times} A$,
that is, a set of pairs of arcs. It is require that in the graph $(V,E)$, that is, the graph on vertex set $V$ and edge set $E$,
each vertex has degree at least one
and at most three, and that, in $(V, E, A, R)$, each vertex of degree one (resp. two) in $(V,E)$ is the tail
of exactly two arcs (resp. one arc) and the head of exactly two arcs (resp. one arc); note that
directed loops are allowed.
Throughout this paper, $\{u,v\}$ denotes the (non-directed) edge with end vertices $u$, $v$ and $(u,v)$ denotes
the arc directed from $u$ to~$v$; if no end vertices are specified $\vec{e}$ denotes an arc.
\subsubsection*{Correct reductions}
Let $(V, E, A, R)$ be a mixed graph and $U \subseteq V$.
A {\em correct reduction} of $U$ on $(V, E, A, R)$ is a procedure that outputs
a new mixed graph $(V', E', A', R')$ and a list $\mathcal{S}$ of directed paths and cycles (here,
a directed loop is considered
a directed cycle)
so that the following items hold:
\begin{itemize}
\item[i)] $V' = V-U$.
\item[ii)] $E'=\{ \{u,v\} \in E: \{u, v\} \cap U =\emptyset\}$.
\item[iii)] Let $\tilde{A}$ denote the set of arcs obtained by replacing each edge in $E$
incident to a vertex of $U$ by two arcs oppositely directed and
$A(U)$ be the subset of $A$ that contains all directed
edges incident to a vertex of $U$. The list $\mathcal{S}$ is the result of
partitioning $A(U){\cup} \tilde{A}$ into correct directed paths
and correct directed cycles. A cycle or a path is {\em correct}
if no pair of its arcs is an element of $R$, and if it is not a 2-cycle composed of
only arcs from $\tilde{A}$;
in particular, a directed loop is a correct cycle.
In addition, a correct path must have both end vertices in $V{-}U$.
\item[iv)] $A' = \{(x,y) \in A: \{x, y\} \cap U =\emptyset\} \cup A''$.
The set $A''$ is the set of arcs obtained by replacing each correct directed path $P \in \mathcal{S}$ by a
new directed edge, say $\vec{e}_P$, with both end vertices in $V{-} U$ and such that
$\vec{e}_P$ has the orientation of $P$.
\item[v)] $R' = \{ \{(x,y), (x',y')\} \in R : \{x, y, x', y'\} \cap U =\emptyset\} \cup R'' \cup \tilde{R}$.
The set $R''$ corresponds to all pairs $\{\vec{e}_P, \vec{e}_{P'}\}$ such that
$P$ and $P'$ have a vertex in common, or there are arcs $\vec{e} \in P$ and $\vec{g} \in P'$ such that
$\{\vec{e},\vec{g}\} \in R$.
The set $\tilde{R}$ corresponds to all pairs $\{\vec{e}_P, \vec{g}\}$ such that
$P$ contains an arc $\vec{e}$ and $\{\vec{e},\vec{g}\} \in R$.
\end{itemize}
\noindent
An example of a correct reduction is provided in Figure~\ref{fig:correct-red}.
In consequence with the definition of correct reductions, we refer to the elements in $\mathcal{S}$ as \emph{correct paths}
and \emph{correct cycles},
and to the elements in $R$ as \emph{forbidden pairs}.
\begin{figure}[h]
\centering
\subfigure[]
{
\figinput{correct_reduction}
\label{fig.srmg00}
}
\subfigure[]
{
\figinput{correct_reduction_1}
\label{fig.srmg01}
}
\subfigure[]
{
\figinput{correct_reduction_2}
\label{fig.srmg02}
}
\subfigure[]
{
\figinput{correct_reduction_3}
\label{fig.bf3}
}
\caption{Correct reduction of $U \subset V$:
(a) elements from $A$ are re\-presented by dotted lines
and the pair $(\vec{e}, \vec{g})$ of arcs is not in $R$,
(b) replace edges of $E$ with an end vertex in $U$ by two arcs oppositely directed,
(c) partition of $A(U) \cup \tilde{A}$ into correct paths and cycles and
(d) resulting structure.}
\label{fig:correct-red}
\end{figure}
Let $G$ be a cubic graph and $V_0, V_1,\ldots, V_k$ be a partition~of $V(G)$ such that for all $i\in[k]$, the graph $G-(V_k \cup \cdots \cup V_i)$ is connected; by abuse of notation, we refer to such a partition as \emph{connected}.
Let $j \in \{0,1,\ldots,k\}$. A \emph{consecutive correct reduction} (\textit{\textbf{ccr}}, in short) of $V_k, V_{k-1},\ldots, V_j$ is a sequence of $k-j+1$ correct reductions such that
for each $i \in \{k, \ldots, j\}$, the correct reduction of $V_{i} \subset W_{i}$ on $(W_{i} ,E_{i}, A_{i}, R_{i})$
outputs $(W_{i-1} ,E_{i-1}, A_{i-1}, R_{i-1})$;
where $(W_{k} ,E_{k}, A_{k}, R_{k}) = (V(G) ,E(G), \emptyset, \emptyset)$
and $(W_{-1} ,E_{-1}, A_{-1}, R_{-1})= (\emptyset, \emptyset, \emptyset, \emptyset)$.
The following proposition is proved in Section~2 of~\cite{ourwork}.
It relates the notion of consecutive correct reductions to the one of a directed cycle double cover ({\dcdc}, in short).
\begin{proposition}\label{prop:consecorr}
Let $V_0, V_1,\ldots, V_k$ be a connected partition~of the vertex set of a cubic graph $G$.
Then each {\ccr} of $V_k, V_{k-1},\ldots, V_0$ constructs a {\dcdc} of~$G$.
\end{proposition}
\subsubsection*{Cut-obstacles}
There are many potential obstructions to the existence of correct reductions. One of them is a cut-obstacle.
\begin{definition}[Cut-obstacle]
\label{def.cuto}
Let $(V, E, A, R)$ be a mixed graph and $U\subseteq V$.
We denote by $C_U$ the subset of $E \cup A$ that contains all edges and arcs with exactly one end vertex in $U$.
We say that $U$ is a {\em cut-obstacle} in $(V, E, A, R)$ if the number of edges in~$C_U$ is strictly less than twice the number of arcs in $C_U$.
\end{definition}
For an illustration of a cut-obstacle see Figure~\ref{fig.cut3ear}.
Cut-obstacles are potential obstacles for the existence of correct reductions in the following sense: if we
assume that all pairs of arcs in $C_U$ belong to $R$, then there is not correct reduction of $U$ on~$(V, E, A, R)$.
The next observation helps understanding the connection between
robust ear decompositions and cut-obstacles.
\begin{observation}\label{obs:cuto1}
If $(H_0, H_i, L_i)^n$ is an ear decomposition of a trigraph $H$ and there exists
a 3-ear $L \in \{L_1, \ldots, L_n\}$ such that its descendant is composed of 3 connected
components, then any {\ccr} of $I(L_n), I(L_{n-1}),\ldots, I(L_1)$ creates a cut-obstacle at $I(L)$.
\end{observation}
In other words, Observation~\ref{obs:cuto1} says that, for a 3-ear,
having at most 2 connected components in its descendant is fundamental in order to not creating
trivial cut-obstacles. Therefore, in a robust ear decomposition (Definition~\ref{def.wr}), there are no trivial cut-obstacles
at any 3-ear. The moreover part of Definition~\ref{def.wr} is imposed by a similar technical reasons:
roughly, this condition avoids trivial cut-obstacles at 3-ears
that belongs to the ear decompositions of the trigraphs that model cubic bridgeless graphs.
\bigskip
In the following, we define ears decompositions that admit
{\ccr} that does not create cut-obstacles. In the subsequent observation we illustrate it for the planar trigraphs.
\begin{definition}[Superb]\label{def.superb}
Let $H$ be a trigraph. We say that an ear decomposition $(H_0, H_i, L_i)^n$ of $H$
is \emph{superb} if there exists a {\ccr} of $I(L_n), I(L_{n-1}), \ldots, I(L_{j})$
for some $j \in [n]$ such that the following properties hold:
\begin{itemize}
\item[(i)] $j=1$ or $I(L_{j-1})$ cannot be correctly reduced, and
\item[(ii)] a cut-obstacle at the internal vertices of a 3-ear in $\{L_{j-1}, L_{j}, \ldots, L_{n}\}$ is never created.
\end{itemize}
\end{definition}
\begin{observation}
\label{o.av}
Every robust ear decomposition of a planar trigraph $H$ is superb.
\end{observation}
\begin{proof}
Let us consider an embedding of $H$ in the plane and let $\mathcal{C}$ be the set of facial cycles defined
by this embedding. The set $\mathcal{C}$ is a {\dcdc} of $H$: we can prescribe clockwise orientation
to the facial cycles of interior faces and anticlockwise orientation to the facial cycle of the external face.
Let $(H_0, H_i, L_i)^n$ be a robust ear decomposition of $H$.
Observe that $\mathcal{C}$ encodes a {\ccr} of $I(L_n), \ldots, I(L_{1}), V(H_0)$.
We claim that such {\ccr} does not create cut-obstacles at any 3-ear.
The rest of the proof is devoted to prove this.
Let $L_i$ be a 3-ear, $i \in [n]$ and $V(L_i)=\alpha v^1 v^2 v^3 \beta$.
We assume that $I(L_{i+1}), \ldots, I(L_{n})$ has been already correctly reduced according to $\mathcal{C}$
and let $H'$ denote the obtained mixed graph.
Suppose first that there are adjacent internal vertices of $L_i$, without loss of generality $v^1$ and $v^2$,
such that if $x$ (resp. $y$)
is the neighbour of $v^1$ (resp. of $v^2$) that is not in $V(L_i)$, then
$x$ and $y$ are in the same component, say $D$, of the descendant of $I(L_i)$.
By planarity and since $H$ is cubic, it holds that the path $xv^1v^2y$ is a subgraph of a facial cycle $C$ of $\mathcal{C}$.
Again by planarity and using the fact that $V(D)$ is connected, we have that all vertices in $V(C)\setminus\{v^1, v^2\}$ belong to $V(D)$.
If so, all vertices in $V(C)\setminus\{v^1, v^2\}$ are internal vertices
of ears that were already reduced and thus, in $H'$, there is an arc with end vertices $v^1, v^2$.
We now suppose that there are no such internal vertices. Since $(H_0, H_i, L_i)^n$ is weakly robust,
it implies that $x, z$, where $z$ is the neighbour of $v^3$ that is not in $V(L_i)$ are in the same
component, say $D'$, of the descendant of $I(L_i)$. Moreover, the component of the descendant of $I(L_i)$ to which
$y$ belongs contains only $y$ (thus, $y$ in an internal vertex of a star in $L_{i+1}, \ldots, L_{n}$)
and in the planar embedding, $y$ is not in the same side (with respect to $L_i$) where $x$ and $z$ are
since $(H_0, H_i, L_i)^n$ is weakly robust and thus $H$ is 2-edge-connected.
As before, it follows that the path $xv^1v^2v^3z$ belongs to a path that is a subgraph of a facial cycle $C'$ of $\mathcal{C}$
and all vertices in $V(C')\setminus\{v^1, v^2, v^3\}$ are in $V(D)$.
We conclude that in $H'$ there is an arc with end vertices $v^1, v^3$.
\end{proof}
\begin{remark}
In Conjecture~\ref{conj:weakavoidance} we propose that a weakening of Observation~\ref{o.av} holds for every
robust trigraph. The weakening is twofold: we allow changing the ear decomposition and also the trigraph itself.
This is explained next.
\end{remark}
\begin{definition}[Admits $H_0, S$]
\label{def.S}
Let $H$ be a trigraph, $H_0$ be an induced cycle of $H$ and let $(H_0, H_i, L_i)^n$
be an ear decomposition of $H$.
\begin{itemize}
\item
We refer to a 2,3-ear as a \emph {base} if its leaves are in $H_0$, as an \emph {up} if one leaf belongs to $H_0$ and the other one to a base, as an \emph{antenna} if exactly one leaf belongs to an up.
\item
Let $S$ be a set of disjoint paths of $H$ on 3 vertices. We say that $(H_0, H_i, L_i)^n$
\emph{admits $H_0, S$} if the following three conditions hold.
(1) If $L_i$ is a 3-ear, then $L_i$ is either a base, or an up or, an antenna containing an element of $S$.
(2)~Each element of $S$ is a subset of an antenna.
(3)~No 1-ear has both leaves in $H_0$ or one leaf in a base and second one in an up.
\end{itemize}
\end{definition}
The following notation helps with the next definition. Given an ear decomposition $(H_0, H_i, L_i)^n$
of a trigraph, we say that a sequence $(L_{i_j})_{j \in [l]}$ of ears is a \emph{heel}
if $L_{i_1}$ is a 2-ear, for every $j \in \{2,\ldots,l\}$
the leaves of $L_{i_j}$ are exactly the internal vertices of $L_{i_{j-1}}$
and $(L_{i_j})_{j \in [l]}$ is maximal.
Note that $L_{i_{j}}$ is either a 2-ear, or a 1-ear.
\begin{definition}[Local exchange. Closure]
\label{def.closure}
Let $H$ be a trigraph, $H_0$ be an induced cycle of $H$ and let $S$ be a set of disjoint paths of $H$ on 3 vertices.
Let $\mathcal{H}=(H_0, H_i, L_i)^n$ be an ear decomposition of $H$ that admits $H_0, S$.
\begin{itemize}
\item A {\em local exchange} on the pair $H, \mathcal{H}$
is an operation that produces a pair $H', (H_0, H'_i, L'_i)^n$ of a trigraph with its ear decomposition (see Figure~\ref{fig:localexchange}) as follows:
Let $L_{i_0}\in \{L_1,\ldots, L_n\}$ be a 3-ear antenna with vertices $a, w_1, w_2, w_3, b$.
Let $(L_{i_j})_{j \in [l]}$ be a heel such that the leaves of $L_{i_1}$ are $w_2, w_3$,
and let $u$ be the internal vertex of $L_{i_l}$ that has the shortest distance (in the heel) to $w_3$.
Further let $L_k, L_m$ be ears such that $w_1$ is a leaf of $L_k$ and $u$ is a leaf of $L_m$
and assume $\{w_1,w'_1\}$, $\{u,u'\}$ are edges of $L_k$, $L_m$, respectively.
Then, $H', L'_k, L'_m$ are obtained by deleting the edges $\{w_1,w'_1\}$, $\{u,u'\}$
and adding the new edges $\{w_1,u'\}$, $\{u,w'_1\}$. The other ears do not change, namely $L'_i=L_i$ for all $i \notin \{k,m\}$.
Note that $(H_0, H'_i, L'_i)^n$ admits $H_0, S$.
\item
The \emph{closure} of the pair $H$, $\mathcal{H}$ is the set of all pairs $H'$,
$\mathcal{H}'$, where $H'$ is a trigraph and $\mathcal{H}'$ is an ear decomposition of $H'$ admitting $H_0, S$,
such that $H'$, $\mathcal{H}'$ are obtained from $H$, $\mathcal{H}$ by a sequence of the following two operations:
(1) a modification of the current ear decomposition to another one admitting $H_0, S$, and (2) a local exchange on the current trigraph.
\end{itemize}
\end{definition}
\begin{figure}[h]
\centering
\figinput{localexchange}
\caption{Illustration of a local exchange: edges $\{w_1,w'_1\}$, $\{u,u'\}$ are deleted, and new edges $\{w_1,u'\}$, $\{u,w'_1\}$ (the ones in blue)
are added.
The case depicted corresponds to the one that the last ear of the heel, namely $L_{i_l}$, is a 2-ear.}
\label{fig:localexchange}
\end{figure}
Note that in definition of closure (Definition~\ref{def.closure}), a modification changes only ear decomposition, while
in a local exchange both, the ear decomposition and the trigraph change.
\subsection{Main Contribution}
\label{sub.mainresults}
Aside of cut-obstacles, there are many other obstacles for the existence of a correct reduction. The main contribution of this paper is a proposition which claims that \emph {avoidance of the cut-obstacles in a very restricted setting} is sufficient for proving the {\dcdc} conjecture.
\begin{conjecture}[Cut avoidance conjecture]
\label{conj:weakavoidance}
Let $H$ be a trigraph, $H_0$ be an induced cycle of $H$ and let $S$ be a set of disjoint paths on 3 vertices.
Let $\mathcal{H}$ be a robust ear decomposition of $H$ that admits $H_0, S$.
Then, there is a pair $H'$, $\mathcal{H}'$ in the closure of $H$, $\mathcal{H}$,
such that $\mathcal{H}'$ is superb
\end{conjecture}
The main result of this paper is the following.
\begin{theorem}\label{th:mainmain}
If Conjecture~\ref{conj:weakavoidance} holds, then
the {\dcdc} conjecture holds in general graphs.
\end{theorem}
The rest of the paper is organized as follows.
In Section~\ref{sec:main} we describe the essential step in the proof
of Theorem~\ref{th:mainmain}.
Firstly, for each cubic graph~$G$ and its ear decomposition,
we construct a trigraph $H(G)$ with specified induced cycle $H_0= H_0(G)$ and set $S= S(G)$.
Simultaneously, we construct ear decomposition $(H_0, H_i, L_i)^n$ of $H(G)$ that admits $H_0, S$.
The ear decomposition $(H_0, H_i, L_i)^n$ will be robust if the ear decomposition of $G$ used in the construction
of $H(G)$ is {\em super robust}.
In Section~\ref{sec:3conn} of this paper we define super robust ear decompositions and
prove that every 3-edge-connected cubic graph admits a super robust ear decomposition.
We recall that the {\dcdc} conjecture is as hard for 3-edge-connected cubic graphs as for general graphs.
In Section~\ref{sec:main1}, we introduce the concept of \emph{relevant} ear decomposition
and show that superb relevant ear decompositions encode
directed cycle double covers of~$G$.
Then (in Lemma~\ref{l.last}) we extend this claim to each ear decomposition in the closure of $(H_0, H_i, L_i)^n$,
thus proving Theorem~\ref{th:mainmain}.
\begin{remark} In this remark we explain the reason behind the definition of the {local exchange} operation.
Assume that 3-edge-connected cubic graph $G$ has a {\dcdc}.
Does then the constructed trigraph $H(G)$ and its constructed ear-decomposition $\mathcal{H}$ satisfy Conjecture~\ref{conj:weakavoidance}?
A {\dcdc} of $G$ defines an embedding of $G$ in an orientable surface with no dual loop.
Such embedding gives rise to a special ear-decomposition, as in the toroidal example of Section~\ref{sec.surfaces}.
Let $H', \mathcal{H}'$ be the trigraph and its ear decomposition constructed from this special ear-decomposition.
Possibly $H$ is not isomorphic to $H'$, but we believe that $H', \mathcal{H}'$ is in the closure of $H, \mathcal{H}$.
We further conjecture that $\mathcal{H}'$ is superb, which is illustrated by the toroidal example of Section~\ref{sec.surfaces}.
\end{remark}
\section{Trigraph $H(G)$}
\label{sec:main}
Let $G$ be a cubic graph and $(G_0, G_i, P_i)^{k}$ be an
ear decomposition of $G$; we recall that, an ear is a path on at least three vertices or a star on four vertices.
The aim of this section is to construct the trigraph $H(G)$ along with an ear decomposition.
Let $v_0$ be a fixed vertex of $G_0$. Let $v_1$ and $v_2$ denote the neighbours of $v_0$ in $G_0$.
We obtain a cubic graph $G'$ from $G$ by subdividing edges $\{v_0,v_1\}$ and $\{v_0,v_2\}$ into
$\{v_0,x_0\}$, $\{x_0,v_1\}$ and $\{v_0,y_0\}$, $\{y_0,v_2\}$, respectively, and adding the new edge $\{x_0,y_0\}$; this operation is
known as a Y-$\Delta$ operation.
The cubic graph~$G'$ admits an ear decomposition starting at the triangle on vertex set
$\{x_0,y_0,v_0\}$, and with building ears $P_0, P_1, \ldots, P_k$, where
$P_0$ is the path obtained from $G_0$ by deleting $v_0$ and
adding the edges $\{x_0,v_1\}$, $\{y_0,v_2\}$.
Clearly, $G$ has a {\dcdc} if and only if $G'$ does so.
In the rest of this section, for each $i \in \{0,1,\ldots,k\}$,
the notation $a_i$, $c_i$ stand for the end vertices of $P_i$, whenever $P_i$ is a path; in particular
$\{a_i, c_i\}=\{x_0, y_0\}$ is the set of end vertices of $P_0$.
Let $H(G)$ be a cycle on $n(G)$ vertices, with $n(G)$ as described in Remark~\ref{o.cl1}.
Set $H_0$ as the starting cycle of the ear decomposition of~$H(G)$ and
choose 3 distinct vertices from $V(H_0)$; we denote the set of these vertices by $V_0$.
The following building block comes in handy to describe the construction~of~$H(G)$.
\begin{definition}[Basic gadget]\label{def:basicgadget}
A \emph{basic gadget} $\mathcal{B}=\mathcal{B}(x,u)$ is a sequence $E_1, E_2, E'_3, E_4, D_1, D_2, D_3$ of short
ears which, considering the vertices named according to Figure~\ref{fig:bagadget}, are defined as follows:
\begin{itemize}
\item[(i)] $I(E_1)=\{a',w',b'\}$, $I(E_2) =\{a,w,b\}$, $V(E'_3)=\{b,z,y,x\}$, $V(E_4)=\{z, u, v, y\}$,
and the end vertices of $E_1$ and $E_2$ belong to $H_0$,
\item[(iii)] $D_1, D_2, D_3$ are stars, $\{a', w, v\}$ is the set of leaves of $D_1$, and each star
$D_2$, $D_3$ has two leaves in $H_0$.
Moreover, vertex $w'$ is a leaf of $D_2$ and vertex $a$ is a leaf of $D_3$.
\end{itemize}
Depending on the context, a basic gadget may also refer to the
graph obtained by the union of the ears $E_1, E_2, E'_3, E_4, D_1, D_2, D_3$.
In addition, we say that the path on vertex set $\{b,z,y\}$ is the {\emph{fixed}} path,
$u$ is the \emph{replica} and $x$ is the \emph{joint} of the basic gadget;
whenever only the replica vertex $u$ is specified, we denote by $x_u$ the corresponding joint vertex.
\end{definition}
\begin{figure}[h]
\centering
\subfigure[Basic gadget $\mathcal B(x,u)$]
{
\figinput{gadget}
\label{fig:bagadget}
}\qquad
\subfigure[Gadget after reduction of its stars.]
{
\figinput{gadgetgadget}
\label{fig:gadget}}
\caption{}
\label{fig:basicandno}
\end{figure}
We now describe the recursive construction of $H(G)$,
along with a function~$\Gamma$ that maps the set of the vertices and edges of $G'$ into a subset of the vertices and edges of $H(G)$.
In addition, we define a set $S$ of disjoint paths of $H(G)$ on 3 vertices,
and an ear decomposition of $H(G)$ that we call {\em canonical}.
Recall that the ear decomposition of $H(G)$ starts at cycle $H_0$.
Set $V_0 = \{\Gamma(x_0), \Gamma(y_0), \Gamma(v_0)\}$ and $S_0 = \emptyset$.
According to the following rules, for each $i \in \{0,1,\ldots,k\}$,
we obtain $H_{t_i}$ from $H_{t_{i-1}}$,
$S_{t_i}$ from $S_{t_{i-1}}$, and the list $\mathcal{L}_{t_{i}}$
of short ears of the canonical ear decomposition of $H(G)$
that generates $H_{t_i}$ from $H_{t_{i-1}}$;
under this notation, $H_{t_{-1}}=H_0$ and $S=S_{t_k}$.
\begin{itemize}
\item Case that $P_i$ is a star with three leaves. Let $s$ denote the center of $P_i$ and let $x, y, z$ denote its leaves.
Graph $H_{t_i}$ is obtained from $H_{t_{i-1}}$ by adding the new vertex $\Gamma(s)= u$,
and the new edges $\{u, \Gamma(x)\}$, $\{u, \Gamma(y)\}$ and $\{u, \Gamma(y)\}$.
For $w\in \{x,y,z\}$ let $\Gamma(\{s,w\})= \{\Gamma(s), \Gamma(w)\}$.
Further, $S_{t_i}=S_{t_{i-1}}$, and
$\mathcal{L}_{t_{i}}$ contains only one element:
the star on vertex set $\{u, \Gamma(x), \Gamma(y), \Gamma(y)\}$.
\item Case that $P_i$ is a path. Let $a_i, b_{i_1}, \ldots, b_{i_l}, c_i$, with $l\geq 1$, be the sequence of vertices of path~$P_i$.
In order to obtain $H_{t_i}$ from $H_{t_{i-1}}$, we first
add {\em substantial path} $Q_i$, which is defined by the sequence of vertices $\Gamma(a_i), x_1, \ldots, x_l, \Gamma(c_i)$, where
$x_1, \ldots, x_l$ are $l$ new vertices.
Then to each $x_i$, we connect a basic gadget graph $\mathcal B_i=\mathcal B(x_i,u_i)$;
consequently, each of them is referred to as a \emph{basic gadget of $H(G)$}.
Set $\Gamma(b_{i_j})= u_j$, for each $j\in \{1, \ldots, l\}$.
Let~$\phi$ denote the natural isomorphism between $P_i$ and $Q_i$
that maps $a_i$ to $\Gamma(a_i)$ and $c_i$ to~$\Gamma(c_i)$.
For each edge $\{u,v\} \in E(P_i)$, let $\Gamma(\{u,v\})= \{\phi(u),\phi(v) \}$.
Note that, under this setting, $\Gamma(\{u,v\})\neq \{\Gamma(u), \Gamma(v)\}$.
In addition, $S_{t_i}$ is the union of the paths in $S_{t_{i-1}}$
and the $l$ fixed paths of the basic gadgets $\mathcal B_i, 1\leq i\leq l$ (see Definition~\ref{def:basicgadget}).
Finally, $\mathcal{L}_{t_{i}} = (R_1, \ldots, R_l)$,
where for each $1\leq j< l$, $R_j$
is the list of ears
\begin{equation} \label{eq:gadget}
E_1(\mathcal B_j),E_2(\mathcal B_j),E_3,E_4(\mathcal B_j),D_1(\mathcal B_j),D_2(\mathcal B_j),D_3(\mathcal B_j),
\end{equation}
$\mathcal{B}(x_j,u_j)\cup\{x_j,x_{j-1}\}$,
where for $j=1$ we let $x_0= \Gamma(a_i)$, and $E_3$ is the path obtained by the union of $E'_3(\mathcal B_j)$ and $\{x_j,x_{j-1}\}$.
The last list $R_l$ consists of the ears \begin{equation} \label{eq:gadget*} E_1(\mathcal B _l),E_2(\mathcal B _l),F,E'_3(\mathcal B _l),E_4(\mathcal B _l),D_1(\mathcal B_l),D_2(\mathcal B_l),D_3(\mathcal B_l),\end{equation}
where $F$ is the path on vertex set $\{x_{l-1},x_l, \Gamma(c_i)\}$.
\end{itemize}
Note that $S$ is exactly the set of all fixed paths of the basic gadgets of $H(G)$.
It is a routine to check that the canonical ear decomposition of $H(G)$ admits $H_0,S$.
Further, if $\mathcal{B}$ is a basic gadget of $H$, then
$E_1(\mathcal B)$ is a base, $E_2(\mathcal B)$ is an up
and, $E_3$ in (\ref{eq:gadget}), $E_3'$ in (\ref{eq:gadget*}) are antennas (see Definition~\ref{def.S}).
The following remark sets $n(G)$; number of vertices of $H_0$.
\begin{remark}\label{o.cl1}
Each path $P_i$ in $\{P_1,\ldots,P_k\}$ gives rise to $|I(P_i)|$ distinct basic gadgets in $H(G)$.
Moreover, in order to construct each basic gadget in $H(G)$, 7 vertices from $H_0$ are needed.
We set,
$$n(G) = 7 \left(\sum_{i \in [k] \,:\, P_i \, \text{path}} |I(P_i)|\right) +3.$$
\end{remark}
Because of Lemmas~\ref{lemma:supimplrob} and~\ref{th:robust} (see Section~\ref{sec:3conn}),
the canonical ear decomposition is robust provided that graph $G$ is 3-edge-connected
and the chosen ear decomposition $(G_0, G_i, P_i)^{k}$ is super robust.
Next observation follows directly from the construction of $H(G)$
and the definition of closure.
Recall that, the cubic graph $G'$ is obtained from $G$ by a $Y-\Delta$ operation, as described
in the beginning of Section~\ref{sec:main}.
\begin{observation}
\label{o.cl}
Let $\mathcal H$ be the canonical ear decomposition of $H=H(G)$
and let $H', \mathcal{H}'$ be a pair in the closure of $H$, $\mathcal H$.
By definition, $\mathcal{H}'$ admits $H_0,S$.
Therefore, each basic gadget of $H$ is a subgraph of $H'$,
and $G'$ is obtained from $H'$ by contracting the set of all vertices of each basic gadget of $H$ to a single vertex.
\end{observation}
Observation~\ref{o.cl} implies that the definition of function $\Gamma$ can be extended
from a canonical ear decomposition to every ear decomposition that belongs to its closure.
This is formalized in the next definition. Note that Definition~\ref{def.ggg} is consistent with the definition of function $\Gamma$ for canonical ear decompositions.
\begin{definition}
\label{def.ggg}
Let $\mathcal H$ be the canonical ear decomposition of $H=H(G)$
and let $H', \mathcal{H}'$ a pair in the closure of $H$, $\mathcal H$.
We define function $\Gamma$ from the set of vertices and edges of $G'$ to the set of vertices and edges of $H'$ as follows.
If $e\in E(G')$, then $\Gamma(e)$ is the edge of $H'$ which
corresponds to $e$ after the vertex-contraction of all the basic gadgets of $H$ in $H'$.
Let $v\in V(G')$ and let $W$ be the subset of vertices of $H'$ such that contraction of $W$ to a single vertex corresponds to $v$.
If $W$ is a single vertex, say $W= \{w\}$, then $\Gamma(v)= w$.
Otherwise, $W$ is the vertex set of a basic gadget, and $\Gamma(v)= u$,
where $u$ is the replica vertex of the basic gadget.
\end{definition}
\section{Gadget analysis}\label{sec:main1}
In order to analyze the canonical ear decomposition, we extend definition of basic gadget to the one of {\em gadget}.
A gadget is basically the list of ears defined in~(\ref{eq:gadget}), which is a subsequence of the canonical ear decomposition,
and it is always associated to a basic gadget. Recall that basic gadgets are defined in~Definition~\ref{def:basicgadget}.
In the rest of this work, a \emph{block} of a sequence $a_1, \ldots, a_n $ is a subsequence $a_j, \ldots, a_m $
for $1 \leq j\leq m \leq n$. Moreover, the notation $(L_{i_1}, \ldots, L_{i_t})^{-1}$ stands for $I(L_{i_t}), \ldots, I(L_{i_1})$.
\begin{definition}[Gadget]\label{def:gadget}
Let $(H_0, H_i, L_i)^n$ be an ear decomposition of a trigraph
and $\mathcal{B}=\mathcal{B}(x,u)$ be a basic gadget.
A \emph{gadget} $\mathcal{G}$ (see Figure~\ref{fig:gadget}) of $(H_0, H_i, L_i)^n$ associated to $\mathcal{B}$
is the block of ears $$\mathcal{G}=E_1(\mathcal{B}), E_2(\mathcal{B}), E_3, E_4(\mathcal{B}), D_1(\mathcal{B}), D_2(\mathcal{B}), D_3(\mathcal{B})$$
of $L_1, \ldots, L_n$, where $E_3$ is the 3-ear on edge set $E'_3(\mathcal{B}) \cup \{x',x\}$ with
$x'$ a neighbour of $x$ not in~$V(E'_3(\mathcal{B}))$.
\end{definition}
We may also say that $\mathcal{B}(x,u)$ is the basic gadget of $\mathcal{G}$.
Naturally, we use the terminology defined for basic gadgets on the gadgets as well.
\begin{definition}[Reduction process]\label{def:redprocess}
A {\em reduction process} of gadget $\mathcal{G}$ is a {\ccr}
of a subsequence $\mathcal{G}'$ of $\mathcal{G}^{-1}$, so that
\begin{itemize}
\item[(i)] either $\mathcal{G}' = \mathcal{G}^{-1}$,
\item[(ii)] or $\mathcal{G}'= I(D_3), I(D_2), I(D_1), I(E_4), \ldots, I(E_{j})$ for some $2 \leq j \leq 4$
such that $I(E_{j-1})$ does not have correct reduction or $I(E_{j-1})$ is a cut-obstacle.
\end{itemize}
If (i) holds, we say that the reduction process is \emph{complete}.
Otherwise, we refer to it as \emph{$j$-incomplete}.
\end{definition}
\subsubsection*{Correct reductions in 3-ears}
We now mention a result from~\cite{ourwork} which describes
the behavior of 3-ears with respect to consecutive correct reductions.
This result is used in the proof of Theorem~\ref{theo:important}.
\begin{definition}[Inner obstacle]\label{def.inner}
An inner obstacle at a 3-ear with internal vertices $\{v^1, v^2, v^3\}$ in a mixed graph $(V,E,A,R)$ is the configuration depicted in Figure~\ref{fig.inner} such that $\{\vec{e},\vec{g}\} \in R$.
\begin{figure}[h]
\centering
\subfigure[Cut-obstacle at $\{v^1, v^2, v^3\}$]
{
\figinput{cut3ear}
\label{fig.cut3ear}
}\qquad
\subfigure[Inner Obstacle]
{
\figinput{obs3ears}
\label{fig.inner}
}
\caption{Potential obstacles to performing correct reductions in a 3-ear.
In the case of the inner obstacle: there exist a correct reduction if and only if $\vec{e}$ and $\vec{g}$ are not forbidden
(that is, $\{\vec{e},\vec{g}\} \notin R$).}
\label{fig:obs3ears}
\end{figure}
\end{definition}
The proof of the following statement can be found in~\cite{ourwork} (Theorem 5).
\begin{theorem}
\label{thm.main2}
Let $H$ be a trigraph and $(H_0, H_i, L_i)^n$ be an ear decomposition of $H$.
Let $i \in [n]$ and $L_i$ be a 3-ear.
Let $H'_i$ denote the mixed graph obtained by a {\ccr} of $I(L_n), \ldots, I(L_{i+1})$.
Assume that $I(L_i)$ is not a cut-obstacle in $H'_i$.
Then, $I(L_i)$ does not admit a correct reduction if and only if $I(L_i)$ is an inner obstacle.
\end{theorem}
From now on let $(H_0, H_i, L_i)^n$ be the ear decomposition of a trigraph $H$
such that the gadget~$\mathcal{G}$, defined by the sequence $E_1, E_2, E_3, E_4, D_1, D_2, D_3$
(see Definition~\ref{def:gadget}), is a block of $L_1, \ldots, L_n$;
without loss of generality, assume that $D_3=L_m$.
As in the definition of gadget, we consider the vertices of
$\mathcal{G}$ named according to Figure~\ref{fig:gadget}.
Moreover, we make the following assumption.
\begin{assumption}\label{assumption1}
There exists a {\ccr} of $I(L_n), \ldots, I(L_{m+1})$ on $H$.
\end{assumption}
\noindent
Let $H'$ denote the mixed graph obtained by a {\ccr} of $I(L_n), \ldots, I(L_{m+1})$. Let us recall that $V(H_0), I(L_1), \ldots, I(L_{m})$ is a partition
of the vertex set of $H'$.
By the definition of gadget, we have that the vertices $x$ and $u$ have degree 2 in $H'$,
and thus, each of them is the tail of one arc and the head of one arc.
Let us denote such arcs, according to Figure~\ref{fig:gadget}, by $\alpha_1=(x,x_1)$, $\alpha_2=(x_2, x)$
and $\beta_1=(u_1, u), \beta_2=(u,u_2)$.
All the remaining vertices of $\mathcal{G}$ in $H'$ have degree~3;
recall that we are considering vertices named according to Figure~\ref{fig:gadget}.
Our aim is to prove the following statement.
\begin{theorem}\label{theo:important}
Each reduction process of $\mathcal{G}$ on $H'$ satisfies exactly one of the following statements.
\begin{itemize}
\item[(i)] It is complete and either $\alpha_1=\beta_1$, or $\alpha_2=\beta_2$.
\item[(ii)] It is $j$-incomplete and $I(E_{j-1})$ is a cut-obstacle for some $2\leq j \leq 4$.
\end{itemize}
In addition, if Statement~(i) holds and $\alpha_1 = \beta_1$ (resp. $\alpha_2=\beta_2$), then
$\alpha_2$ and $\beta_2$ (resp. $\alpha_1$ and $\beta_1$) belong to the
two distinct correct paths (defined by the reduction process) that contain $(x,x')$ and $(x',x)$.
\end{theorem}
\begin{proof}
Let us first suppose that a complete reduction process of $\mathcal{G}$ has been performed on $H'$.
In order to prove the statement of the theorem we need to show that either $\alpha_1=\beta_1$, or $\alpha_2=\beta_2$.
For that, let us examine the single correct reductions involved in the {\ccr}
on $\mathcal{G}$ that witness the existence of the considered complete reduction process. By definition
of complete reduction process, no cut-obstacles at 3-ears are created.
Without loss of generality, by symmetry of the stars, we can assume that the local
configuration depicted in Figure~\ref{fig:gadget} is the one generated by the correct reduction of $I(D_3), I(D_2)$ and $I(D_1)$.
Let $\tilde{H}$ denote the obtained mixed graph. Moreover, let $e=(v,w)$ and $e'=(a',v)$ in $\tilde{H}$;
recall again that we consider vertices named according to Figure~\ref{fig:gadget}.
We claim that the following holds.
\begin{observation}\label{obs:useful}
In each {\ccr} of $I(E_4), I(E_3), I(E_2), I(E_1)$ on $\tilde{H}$
that creates no cut-obstacles at $I(E_3)$, $I(E_2)$ and $I(E_1)$,
the arc $e$ belongs to the correct path that contains $(b,z)$
and $e'$ belongs to the correct path that contains $(z,b)$.
\end{observation}
\begin{proof}[Proof of Observation~\ref{obs:useful}]
The hypothesis that no cut-obstacles at $I(E_3)$, $I(E_2)$ and $I(E_1)$ are created, implies that the
{\ccr} of
$I(E_4), I(E_3)$ generates at least one arc with both end vertices in $I(E_2)=\{a,w,b\}$,
otherwise we have that $I(E_2)$ is a cut-obstacle. In order to get such an arc,
the {\ccr} of $I(E_4), I(E_3)$ is so that $e$ belongs to the correct path that contains $(b,z)$.
Now, for the sake of contradiction, we assume that in a {\ccr} of
$I(E_4), I(E_3)$ which does not create cut-obstacles, the directed edge $e'$ does not belong to the correct path
that contains $(z,b)$. Therefore, without loss of generality,
we can assume that the configuration locally depicted in Figure~\ref{fig.case1ii} is obtained
by the {\ccr} of $I(E_4), I(E_3)$;
up to some different location of $\{u_1,x_1,x'\}$.
On the obtained mixed graph there exists a correct reduction of $I(E_2)$ that does not create a cut-obstacle at $I(E_1)$.
If so, the correct reduction of $I(E_2)$ is so that
the arc $(w,a')$ belongs to the correct path that contains $(b',b)$.
Because of the fact that $(b,w)$, $(w,a')$ is a forbidden pair of arcs, the correct path $(b',b, w, a')$ exists
and does not contain the arc $(b,w)$.
Moreover, since the pair $(b,w)$, $(u_1,b)$ is also forbidden, there exists the correct path $(u_1,b,b')$.
But then, it holds that the arc $(b,w)$ and the remaining
arc $(w,b)$ (arising from the edge $\{w,b\}$) belong to the same correct path or cycle.
However, it does not correspond to a correct reduction since $(a,w)$ and $(w,a)$ are forced to be in the same cycle.
\end{proof}
It is a routine to check that the following list of reductions (encoded by their correct paths and cycles)
correspond to all possible 4 distinct correct reductions of~$I(E_4)$.
\vspace*{0.2cm}
\begin{tabular}{ll}
\vspace*{0.2cm}
Reduction 1: & $(y,v,w)$, $(a',v,u,z)$, $(z,u,u_2)$ and $(u_1,u,v,y)$. See Figure~\ref{fig.case1}.\\ \vspace*{0.2cm}
Reduction 2: & $(z,u,v,w)$, $(a',v,y)$, $(y,v,u,u_2)$ and $(u_1, u, z)$. See Figure~\ref{fig.case2}.\\\vspace*{0.2cm}
Reduction 3: & $(z,u,v,y)$, $(y,v,w)$, $(u_1,u,z)$ and $(a',v,u,u_2)$. See Figure~\ref{fig.case3}.\\\vspace*{0.2cm}
Reduction 4: & $ (y,v,u,z), (a',v,y), (z,u,u_2)$ and $(u_1,u,v,w)$. See Figure~\ref{fig.case4}\vspace*{0.2cm}
\end{tabular}
\begin{figure}[h]
\centering
\subfigure[Case 1]
{
\figinput{case1}
\label{fig.case1}
}\qquad \qquad
\subfigure[Case 2]
{
\figinput{case2}
\label{fig.case2}}
\vspace*{0.2cm}
\subfigure[Case 3]
{
\figinput{case3}
\label{fig.case3}}\qquad \qquad
\subfigure[Case 4]
{
\figinput{case4}
\label{fig.case4}}
\caption{Resulting configurations of all possible correct reductions of $I(E_4)$.
The arc incident to $w$ contains $e$ and the arc incident to $a'$ contains $e'$.}
\label{fig.reductionE4}
\end{figure}
Recall that we aim to prove that either $\alpha_1=\beta_1$, or $\alpha_2=\beta_2$.
We first study reduction 1 (the analysis of reduction 2 follows in an analogous way).
Since we are examining the steps of a complete reduction process of $\mathcal{G}$ on $H'$,
we have that $I(E_3)$ is not a cut-obstacle; thus, there exists $i\in\{1,2\}$ such that $\alpha_{i} = \beta_i$.
Hence the study of reduction 1 involves the following three cases: (i) $\alpha_{1} = \beta_1$ and
$\alpha_{2} \neq \beta_2$, (ii) $\alpha_{1} \neq \beta_1$ and $\alpha_{2} = \beta_2$,
and (iii) $\alpha_{1} = \beta_2$ and $\alpha_{1} = \beta_2$.
By Observation~\ref{obs:useful}, if case~(i) holds, then a correct reduction of $I(E_3)$
takes the correct path $(b,z,y,w)$ and makes that $e'$ belongs to the correct path that contains $(z,b)$.
Therefore, in a correct reduction of $I(E_3)$ which does not create cut-obstacles the following
holds: the arc $(z,u_2)$ ($\alpha_2$, respectively) is in the correct path that contains $(x',x)$ ($(x,x')$, respectively);
thus the statement of Theorem~\ref{theo:important} follows.
\begin{figure}[h]
\centering
\subfigure[A correct reduction of $V(E_3)$ in Case 1(i).]
{
\figinput{case1i}
\label{fig.case1i}}\qquad
\subfigure[A correct reduction of $V(E_3)$ in Case 1(ii).]
{
\figinput{case1ii}
\label{fig.case1ii}}
\caption{}
\end{figure}
Now we study case~(ii). Recall that reduction~1 is depicted in Figure~\ref{fig.case1}; to obtain case (ii) we need to set $(z,x)=(z,u_2)=(x_2,x)$.
Let us suppose that there exists a correct reduction of $I(E_3)$ that does not create cut-obstacles.
By Observation~\ref{obs:useful}, this implies that $e$ and $(b,z)$ belong to the same correct path.
Since the pair $e$, $(z,x)$ is not a forbidden pair of arcs, then there are two
potential correct paths such that $e$ and $(b,z)$ could belong to (in case that the required correct reduction exists);
namely $e$ belongs to either
$(b,z,y,w)$, or to $(b,z,x,y,w)$.
If the correct reduction takes the correct path $(b,z,y,w)$, then by Observation~\ref{obs:useful}
this correct reduction also takes the correct path $(a',z,b)$.
We complete the proof using the same argument as for case (i); meaning,
if a correct reduction of $I(E_3)$ exists then $(u_1, y)$ ($\alpha_1$, respectively)
belongs to the correct path that contains $(x,x')$ ($(x',x)$, respectively).
We now suppose that the correct reduction takes the correct path $(b,z,x,y,w)$.
This forces the correct reduction to consider the following correct paths:
$$(u_1,y,z,b), (a',z,y,x,x'), (x',x,x_1).$$
However, this contradicts Observation~\ref{obs:useful}, since $e'$ and $(z,b)$ must belong to the same correct path.
We now move to study case (iii). We can obtain a local sketch of this case if
we set $(x,y)=\alpha_1=(u_1,y)$ and $(z,x)=\alpha_2 =(z,u_2)$,
in Figure~\ref{fig.case1}.
First of all, by Observation~\ref{obs:useful}, the correct path $(a',z,b)$ is created and
therefore, the correct path $(b,z,y,w)$ exists as well. Hence,
the correct cycle $(y,x,z,y)$ exists. It implies that the arcs $(x,x')$, $(x',x)$ (arising from the edge $\{x,x'\}$)
are in the same cycle; a contradiction to the definition of correct reduction.
Let us study reductions~3 and~4. By Observation~\ref{obs:useful},
on the one hand, in the case of reduction 3 we have $(a',u_2)= (x_2,x)$ in Figure~\ref{fig.case3}.
On the other hand, in the case of reduction~4 we must have $(u_1,w)=(x,x_1)$ in Figure~\ref{fig.case4}.
Therefore, reductions 3 and 4 are analogous and it suffices to study one of them.
We study reduction 3. In any correct reduction of $I(E_3)$ that does not create cut-obstacles, there exists only one
possible correct path, namely $(b,z,y,w)$, such that Observation~\ref{obs:useful} is satisfied;
because if there were a different correct path, then this correct path
would contain $(u_1,z)$ and would
required that $(u_1,z)=(x,x_1)$, but such a path cannot contain $(b,z)$ since
the direction of $(u_1,z)$ is opposite to the one of $(b,z)$ in any potential
correct path.
Moreover, since the pair $(u_1,z)$, $(z,y)$ is forbidden, then the correct path $(u_1,z,b)$ exists.
Hence, $(a',x)$ is not in the correct path containing $(z,b)$, a contradiction to the statement of Observation~\ref{obs:useful}.
This conclude the first part of the proof of Theorem~\ref{theo:important}.
For the second part of the proof, we consider an $j$-incomplete reduction process of $\mathcal{G}$
for some $2\leq j\leq 4$ and have to show that it implies the existence of a cut-obstacle at $I(E_{j-1})$.
By definition of $j$-incomplete reduction process and Theorem~\ref{thm.main2}, the desired result follows
from the fact that no {\ccr} of $I(D_3), I(D_2), I(D_1), I(E_4), \ldots, I(E_j)$
on $H'$ creates an inner obstacle at $I(E_{j-1})$ (see Definition~\ref{def.inner}).
A quick examination of Figure~\ref{fig:gadget} show us that if $j\in\{2,3\}$, then
there is no inner obstacle at $I(E_{j-1})$; in the case that $j=2$ (resp. $j=3$),
note that all arcs incident to $w'$ (resp. $a$)
are not incident to any other vertex of $I(E_1)$ (resp. $I(E_2)$).
In the case that $j=4$, if an inner obstacle at $I(E_3)$ were created,
then it would be required that $\alpha_1= \beta_1$ and $\alpha_2= \beta_2$, because
in an inner obstacle there are exactly 2 arcs that do not have all its end vertices in
$I(E_3)$. Moreover, it would be needed that the arc that connects $x$ and $z$ forms a forbidden pair
with the arc that has exactly one-end vertex in $I(E_3)$ and this end vertex is $y$; however, this does never occur.
\end{proof}
The following theorem states that the condition either $\alpha_1=\beta_1$, or $\alpha_2=\beta_2$
is also sufficient for the existence of a complete reduction process of $\mathcal{G}$.
\begin{theorem}\label{theo:existence}
There exists a complete reduction process of $\mathcal{G}$
if and only if either $\alpha_{1} = \beta_1$ and $\alpha_{2} \neq \beta_2$,
or $\alpha_{2} = \beta_2$ and $\alpha_{1} \neq \beta_1$.
\end{theorem}
\begin{proof}
The necessity of the condition is stated in Theorem \ref{theo:important}.
It remains to show that the condition is sufficient.
Without loss of generality, $\alpha_{1} = \beta_1$ and
$\alpha_{2} \neq \beta_2$ can be assumed.
We want to prove that a {\ccr} of $\mathcal{G}^{-1}$
that creates no cut-obstacles at $I(E_3)$, $I(E_2)$ and $I(E_1)$ exists.
By the proof of Theorem~\ref{theo:important}, it is possible to correctly reduce $I(D_3), I(D_2), I(D_1), I(E_4)$ in
such a way that we end up in Case 1 (depicted in Figure~\ref{fig.case1}).
We consider the notation from Figure~\ref{fig.case1};
recall that we have $(x,x_1)=(u_1,y)$.
We reduce $I(E_3)$ in such a way that the following list of correct paths and cycles
determines the reduction:
$$(b,z,y,w), \quad (a',z,b), \quad (x',x,y,z,u_2), \quad (x_2,x,x'), \quad (x,y,x).$$
This reduction is correct and we obtain the configuration depicted in Figure~\ref{fig.case1i}.
\begin{figure}[h]
\centering
\subfigure[A correct reduction of $I(E_3)$ in Case 1(i).]
{\figinput{case1i}
\label{fig.case1i}}\qquad
\subfigure[A correct reduction of $I(E_3)$ in Case 1(ii).]
{
\figinput{case1ii}
\label{fig.case1ii}}
\caption{}
\end{figure}
After this reduction of $I(E_3)$, we complete the {\ccr} by reducing $I(E_2)$ and $I(E_1)$, respectively
by the reductions determined by the lists of correct paths and cycles:
$$ (b,w,b), (a',b,b'), (b',b, w, a, a_1), (a'', a, w, a'), (a_2, a, a'') \,\,\,\, \text{and}$$
$$ (a'',a',\tilde{a}), (\tilde{a}, a', w', w'_1), (w'_2, w', b', \tilde{b}), (\tilde{b}, b', a_1), (a',b',w',a'), \,\,\, \text{respectively.}$$
\end{proof}
\subsection{Gadgets*, gadgets** and double gadget}
\label{sub.gg}
Note that, in the construction of the canonical ear decomposition of $H(G)$,
there are two blocks of ears involved; both of these blocks are extensions of basic gadgets.
One of them is block~(\ref{eq:gadget}), and it corresponds
to the gadget, which is defined in Definition~\ref{def:gadget} and whose
analysis is worked out earlier in Section~\ref{sec:main1}.
In addition to the gadget, a slightly different block of ears
arises (namely, block~(\ref{eq:gadget*}));
we shall refer to this block as a \emph{1-gadget*}. It turns out that,
gadgets and 1-gadgets* behave exactly in the same way with respect to
consecutive correct reductions without cut-obstacles.
In addition, we introduce other useful extensions of basic gadgets, namely a \emph{2-gadget*,
1-gadget**, 2-gadget**}.
\begin{figure}[h]
\subfigure[Gadget* after reduction of its stars.]
{
\figinput{gadget1}
\label{fig:gadget*}}\qquad
\subfigure[Gadget** after reduction of its stars.]
{
\figinput{gadget2}
\label{fig:gadget**}}\caption{A 1-gadget* (resp. 1-gadget**) contains the 1-ear on vertex set $\{x',x,s\}$,
and a 2-gadget* (resp. 2-gadget**) contains the 2-ear on vertex set $\{x',x,s,x^*\}$.
In the case of a 2-gadget* or a 2-gadget**, we say that vertex $s$ is the \emph{pending vertex} of the gadget.}\label{fig:*and**}
\end{figure}
Let $(H_0, H_i, L_i)^n$ be an ear decomposition of a trigraph $H$
and $\mathcal{B}=\mathcal{B}(x,u)$ be a basic gadget of $H$.
A \emph{1-gadget*} $\mathcal{G}$ of $(H_0, H_i, L_i)^n$ associated
to $\mathcal{B}$ is the block of ears
\begin{equation}
\label{eq:1gadget*} \mathcal{G}= F, \mathcal{B}\end{equation}
of $L_1, \ldots, L_n$, where $F$ is the 1-ear with internal vertex $x$ and end vertices not in $V(\mathcal{B})$;
in the case that $F$ is a 2-ear with internal vertices $x, s$ and $s$ is neither a joint vertex, nor a replica vertex
of a basic gadget of $H$, we say that $\mathcal{G}$ is a \emph{2-gadget*}. We refer to 1-gadgets* and to 2-gadgets*
as \emph{gadgets*}.
In addition to gadgets*, we need to introduce a natural extension of them; namely, \emph{gadgets**}.
Informally, a gadget** is obtained from a gadget* by removing the 2-ear $E_4$ and adding
two new 1-ears instead.
A gadget** $\mathcal{G}$ of $(H_0, H_i, L_i)^n$ associated to $\mathcal{B}$ is the block of ears
\begin{equation}
\label{eq:1gadget**}
\mathcal{G}= F, E_1(\mathcal{B}), E_2(\mathcal{B}), E_3(\mathcal{B}), F_1, F_2, D_1(\mathcal{B}), D_2(\mathcal{B}), D_3(\mathcal{B}),
\end{equation}
where $F$ is as described for gadgets* (accordingly we have 1-gadgets** and 2-gadgets** --- see Figure~\ref{fig:*and**}),
$F_2$ is the 1-ear with vertex set $\{y,v,u\}$ and $F_1$ is the 1-ear with vertex set $\{z,u,u'\}$
with $u'$ the neighbour of $u$ that is not in $V(\mathcal{B})$ (as described in Figure~\ref{fig:gadget**}).
Again, the terminology defined for basic gadgets and gadgets is naturally transfered to gadgets* and gadgets**.
Further, we suppose that $D_3(\mathcal{B})=L_m$ and that Assumption~\ref{assumption1} holds.
If $\mathcal{G}$ is a gadget*, suppose that $\beta_1$ and $\beta_2$ are the arcs incident
to the replica vertex of $\mathcal{G}$ with the orientations
according to Figure~\ref{fig:gadget*}.
If $\mathcal{G}$ is a gadget**, let $\beta_1$ denote the arc $(u',u)$
and $\beta_2$ the arc $(u,u')$.
The following statement for the gadgets* and gadgets** follows from the proof of Theorems~\ref{theo:important} and~\ref{theo:existence}.
\begin{theorem}\label{theo:bstargad}
Let $\mathcal{G}$ be a gadget* or a gadget**.
If there exists a {\ccr} of $\mathcal{G}^{-1}$
that creates no cut-obstacles at a 3-ear from $\mathcal{G}$,
then $\beta_1$ and $\beta_2$ belong to the two distinct correct paths that contain $(y,x)$ and $(x,y)$,
respectively. Moreover, such a {\ccr} always exists.
\end{theorem}
Because of technical reasons, we present some extra block of ears, which are concatenations of basic gadgets
in gadget* and/or gadget** fashion.
We refer to them as \emph{double gadgets} and illustrate them in Figure~\ref{fig:double}.
A block of ears $\mathcal{D}$ of $L_{1}, \ldots, L_{n}$ is called a \emph{double gadget} if
\begin{equation*}
\mathcal{D} = F', \mathcal{G}_{\mathcal{B}}-F(\mathcal{G}_{\mathcal{B}}), \mathcal{G}_{\mathcal{B}'}-F(\mathcal{G}_{\mathcal{B'}})
\end{equation*}
where $\mathcal{B}=\mathcal{B}(x,u)$, $\mathcal{B}'=\mathcal{B}'(x',u')$ are basic gadgets,
$\mathcal{G}_{\mathcal{B}}-F(\mathcal{G}_{\mathcal{B}})$ and $\mathcal{G}_{\mathcal{B}'}-F(\mathcal{G}_{\mathcal{B'}})$
are obtained from gadgets* or gadgets** $\mathcal{G}_{\mathcal{B}}$ and $\mathcal{G}_{\mathcal{B}'}$
associated to $\mathcal{B}$ and $\mathcal{B}'$ respectively (see~(\ref{eq:1gadget*}) and~(\ref{eq:1gadget**}))
by removing the ear $F(\mathcal{G}_{\mathcal{B}})$ and $F(\mathcal{G}_{\mathcal{B'}})$, respectively, and
$F'$ is a 2-path with internal vertices $x$, $x'$ and end vertices
not in $V(\mathcal{G}_{\mathcal{B}} \cup \mathcal{G}_{\mathcal{B}'})$.
We denote a double gadget by $\mathcal{D}(\mathcal{B},\mathcal{B}')$.
\begin{figure}[h]
\centering
\figinput{double}
\caption{Example of a double gadget where $\mathcal{G}_{\mathcal{B}}-F(\mathcal{G}_{\mathcal{B}}), \mathcal{G}_{\mathcal{B}'}-F(\mathcal{G}_{\mathcal{B'}})$ are obtained from gadgets*.}
\label{fig:double}
\end{figure}
As expected double gadgets have the same behavior, with respect to
the reduction process, as gadgets* and gadgets** do. Therefore, its analysis follows from the discussion regarding gadgets.
As before, we suppose that the last ear of the double gadget $\mathcal{D}$ is $L_m$ and that Assumption~\ref{assumption1} holds.
Note that in $\mathcal{D}(\mathcal{B},\mathcal{B}')$,
when $\mathcal{G}_{\mathcal{B}}$ and $\mathcal{G}_{\mathcal{B}'}$ are gadgets*,
the replica vertices of $\mathcal{B}$ and $\mathcal{B}'$ have degree 2 in the mixed graph obtained by the {\ccr}
of $I(L_n),\ldots I(L_{m+1})$ and therefore each of them is incident to two arcs,
say $\beta_1$, $\beta_2$ in the case of the replica vertex of $\mathcal{B}$,
and $\beta'_1$, $\beta'_2$ in the case of the replica vertex of $\mathcal{B}'$ (their directions
are as in Figure~\ref{fig:double}). In the case that $\mathcal{G}_{\mathcal{B}}$ is
a gadget**, we have $u$ has degree 3. If $u''$ denotes the neighbour of $u$ that does not belong to $V(\mathcal{B})$,
we denote $\beta_1=(u'',u)$ and $\beta_2=(u,u'')$; analogously for
$\mathcal{G}_{\mathcal{B}'}$. We point out that possibly $\beta_1= \beta_2'$ or $\beta_2= \beta_1'$.
The following statement for double gadgets follows.
\begin{theorem}\label{theo:double}
Let $\mathcal{D}=\mathcal{D}(\mathcal{B},\mathcal{B}')$ be a double gadget
such that $\{b,z,y\}$ (resp. $\{b',z',y'\}$) is the {\emph{fixed}} path,
$u$ (resp. $u'$) is the \emph{replica} vertex and $x$ (resp. $x'$) is the \emph{joint} vertex of $\mathcal{B}$ (resp. $\mathcal{B}'$).
If there exists a {\ccr} of $\mathcal{D}^{-1}$ that creates no cut-obstacles at a 3-ear from $\mathcal{D}$,
then the arcs $\beta_1, \beta_2$ (resp. $\beta'_1, \beta'_2$)
belong to the correct paths that contains $(y,x)$ and $(x,y)$, respectively (resp. $(y',x')$ and $(x',y')$).
Moreover, such a {\ccr} always exists.
\end{theorem}
\subsection{A directed cycle double cover for $G$}
In this section we complete the proof of Theorem~\ref{th:mainmain};
we want to prove that Conjecture~\ref{conj:weakavoidance}
implies the {\dcdc} conjecture in general graphs.
Let $G'$, $H(G)$, $S$ and $\Gamma$ as defined in Section~\ref{sec:main}.
We recall that the canonical ear decomposition of $H(G)$
admits $H_0,S$ and it is robust provided that graph $G$ is cyclically 3-edge-connected
and the ear decomposition of $G$ used in the construction of $H(G)$ is super robust.
In order to understand the aim of the next definition, recall that
Observation~\ref{o.cl} claims that if $H', \mathcal{H}'$ is an element in the closure
of $H(G)$ and its canonical ear decomposition, then every
basic gadget of $H(G)$ is contained in $H'$.
\begin{definition}[Relevant ear decompositions
\label{def.rell}
Let $\mathcal H$ be the canonical ear decomposition of $H=H(G)$ and
$H', \mathcal H'$, be a pair in the closure of $H$, $\mathcal H$.
Set $\mathcal H'=(H_0, H'_i, L'_i)^n$.
We say that $\mathcal H'$ is \emph {relevant} if
the sequence $L'_1, \ldots, L'_n$ (possibly after
some reordering) can be partitioned into blocks of ears, with each block
being either a star (on 3 or 4 vertices),
or a gadget, or a gadget*, or a gadget**, or a double gadget.
\end{definition}
Note that, by definition, the canonical ear decomposition is relevant, since its sequence of building
ears can be partitioned into blocks of ears with each block
consisting of a star on 4 vertices, or of a gadget, or of a 1-gadget*.
The following observation follows directly from the construction of trigraph $H(G)$, and
it helps understanding relevant ear decompositions
in the closure and the role of 2-gadgets* and 2-gadgets**.
Recall that, by definition, the pending vertex of 2-gadgets* and 2-gadgets**
is neither a joint vertex, nor a replica vertex.
\begin{observation}
\label{obs.2gad*}
Let $\mathcal H=(H_0, H_i, L_i)^n$ be a canonical ear decomposition of $H=H(G)$ and
$H', \mathcal H'=(H_0, H'_i, L'_i)^n$, be a pair in the closure of $H$, $\mathcal H$
such that $\mathcal H'$ is relevant.
If $\mathcal{G}$ is a 2-gadget* of $\mathcal H'$ with $x$ and $s$ the internal vertices
of $F(\mathcal{G})$ and $x$ the joint vertex of $\mathcal{G}$ and $s$ the pending
vertex of $\mathcal{G}$ (see definition in Figure~\ref{fig:gadget**}), then $s$ is the internal vertex of a star
on~4 vertices of $\{L_1,\ldots,L_n\}$.
\end{observation}
We now extend Definition~\ref{def:redprocess}.
\begin{definition}[Reduction process --- extension]\label{def:redprocess2}
Let $(H_0, H_i, L_i)^n$ be an ear decomposition of~$H(G)$.
A {\em reduction process} of $L_1, \ldots, L_n$ is a {\ccr}
of $I(L_n), \ldots, I(L_j)$ so that
\begin{itemize}
\item[(i)] either $j= 1$, or
\item[(ii)] $j>1$ and $I(L_{j-1})$ does not have correct reduction or $I(L_{j-1})$ is a cut-obstacle.
\end{itemize}
If (i) holds, we say that the reduction process is \emph{complete}.
Otherwise, we refer to it as \emph{$j$-incomplete}.
\end{definition}
In what follows, $\mathcal{H}=(H_0, H_i, L_i)^n$ denotes the canonical ear decomposition of $H=H(G)$.
Lemma~\ref{lemma:m11} is implied by Theorem~\ref{theo:important}, Theorem~\ref{theo:bstargad} and Theorem~\ref{theo:double}.
\begin{lemma}
\label{lemma:m11}
Let $(H_0, H'_i, L'_i)^n$ be a relevant ear decomposition in the closure of $H$, $\mathcal{H}$.
Each reduction process of $L'_1, \ldots, L'_n$ satisfies exactly one of the following two statements.
\begin{itemize}
\item[\textendash]
The reduction process is complete.
\item[\textendash]
The reduction process is $j$-incomplete and $I(L'_{j-1})$ is a cut-obstacle.
\end{itemize}
\end{lemma}
In Lemma~\ref{lemma:main} we state that superb relevant ear decompositions
encode directed cycles double covers.
\begin{lemma}\label{lemma:main}
Let $H', \mathcal{H}'=(H_0, H'_i, L'_i)^n$ be in the closure of $H$, $\mathcal{H}$
such that $\mathcal{H}'$ is a relevant ear decomposition.
If $\mathcal{H}'$ is superb, then it encodes a {\dcdc} of $G$.
\end{lemma}
\begin{proof}
By Lemma~\ref{lemma:m11} and definition of superb,
it follows that a {\ccr} of $I(L'_n), \ldots, I(L'_1)$ that does not create
cut-obstacle at any 3-ear exists. We show that such a {\ccr} encodes a {\dcdc} of~$G$.
We recall that, by Observation~\ref{o.cl}, the cubic graph $G'$ can be obtained from the trigraph $H'$ by contracting each basic
gadget of $H$ into a single vertex.
Since $\mathcal H'$ is relevant, its sequence of ears
$L'_1, \ldots, L'_n$ can be partitioned into blocks of ears
$\mathcal E_1,\ldots, \mathcal E_t$ where each
$\mathcal E_i$ is either a star on 3 or 4 vertices, or a gadget, or a gadget*, or a gadget**, or a double gadget.
This partition gives rise to a connected partition $W_1, \ldots W_t$
of the union of the internal vertices of the ears of $G'$,
where each $W_i$ contains either
a single vertex, or two vertices in the following way:
\begin{itemize}
\item if $\mathcal E_i$ is a star with center $\Gamma(s)$, then $W_i=\{s\}$,
\item if $\mathcal E_i$ is a gadget or a 1-gadget*, or a 1-gadget** with replica vertex $\Gamma(b)$, then $W_i=\{b\}$,
\item if $\mathcal E_i$ is a 2-gadget*, or a 2-gadget** with replica vertex $\Gamma(b)$ and pending vertex $\Gamma(s)$,
then $W_i=\{b,s\}$,
\item if $\mathcal E_i$ is a double gadget with replica vertices $\Gamma(b), \Gamma(b')$, then $W_i=\{b, b'\}$.
\end{itemize}
Recall that function $\Gamma$ (see Definition \ref{def.ggg}) embeds the set of vertices and edges of $G'$ into a subset of the vertices and edges of $H'$.
Assume now that it is possible to perform a {\ccr} of $W_t, \ldots W_1$.
After performing this {\ccr} of $W_t, \ldots W_1$ on $G'$, we obtain a mixed graph
on the vertex set $V_0$, where the underlying graph is a triangle.
Moreover, by Observation~3 of~\cite{ourwork}, we have that each mixed triangle has a correct reduction;
thus, by Proposition~\ref{prop:consecorr}, $G'$ has a {\dcdc}, and also $G$ does.
Thus, to conclude the proof of the lemma, it suffices to prove the following claim.
\begin{claim}\label{cl:cons}
A {\ccr} of $\mathcal E_t^{-1},\ldots, \mathcal E_1^{-1}$
induces a {\ccr} of~$W_t, \ldots W_1$.
\end{claim}
The rest of the proof is devoted to prove Claim~\ref{cl:cons}.
Let $i \in \{1, \ldots, t\}$. We assume that a {\ccr} of $\mathcal E_t^{-1},\ldots, \mathcal E_{i}^{-1}$
and the corresponding {\ccr} of $W_t, \ldots W_{i}$
is performed on $H'$ and $G'$, respectively. Let $M(i)$ and $M(G',i)$
denote the generated mixed graphs, with $M(t)=H'$, $M(G',t) =G'$.
Without loss of generality, $\mathcal E_t^{-1}$ is a star, and hence
the correct reduction of $W_t$ is natural since $W_t$ is the center of a star.
Thus, it suffices to prove that the correct reduction of $\mathcal E_{i-1}^{-1}$ on $M(i)$
induces a correct reduction of $W_{i-1}$ on $M(G',i)$.
The following statements, namely Properties~\ref{prop:keep1} and~\ref{prop:keep2},
are satisfied for every $u, v \in V(G')$ and for $j = t$.
We assume that they also hold for all $j\geq i$ and we argue that they are satisfied for~$i-1$.
\begin{property}\label{prop:keep1}
\emph{The arc $(u,v)$ is in $M(G',j)$ if and only if there exists an arc $\vec{a}$
in $M(j)$ with its head in $\{x_{\Gamma(v)}, \Gamma(v)\}$ and its
tail in $\{x_{\Gamma(u)}, \Gamma(u)\}$; in other words, after
vertex-contraction of all basic gadgets of $M(i)$ into their replica vertices,
the arc $(\Gamma(u), \Gamma(v))$ exists. Recall that $x_{\Gamma(u)}$ denotes
the joint vertex of the basic gadget with replica $\Gamma(u)$; in case $\Gamma(u)$
is not part of a basic gadget, set $x_{\Gamma(u)}=\Gamma(u)$.
By abuse of notation, we denote $\vec{a}$ by $\Gamma((u,v))$.
Furthermore, note that if a pair of arcs is forbidden in $M(G',i)$,
then their images under $\Gamma$ form a pair of arcs forbidden in $M(i)$.}
\end{property}
\begin{property}\label{prop:keep2}
\emph{Suppose that $(u',v')$ is an arc that belong to the correct path of
the correct reduction of $W_i$ on $M(G',i+1)$ which is replaced by arc $(u,v)$ in order to obtain $M(G',i)$.
Then, either $(u',v')$ is an arc of $M(G',i+1)$, or $\{u',v'\}$ is an edge of $M(G',i+1)$.
If $(u',v')$ is an arc of $M(G',i+1)$, then the arc $\Gamma((u',v'))$ of $M(H(G),i+1)$ (defined in Property~\ref{prop:keep1})
is in the correct path replaced by the arc $\Gamma((u,v))$.
If $(u',v')$ is an arc obtained by the orientation of edge $\{u',v'\}$ of $M(G',i)$, then the corresponding orientation of
$\Gamma(\{u',v'\})$ is in the correct path replaced by the arc $(\Gamma((u,v))$.}
\end{property}
Assuming validity of both properties for $i$, it is a routine to check that it is also valid for $i-1$ by
using Theorems~\ref{theo:important},~\ref{theo:existence},~\ref{theo:bstargad} and~\ref{theo:double}.
This finishes the proof of the statement of Claim~\ref{cl:cons}, and also of Lemma~\ref{lemma:main}.
\end{proof}
Finally, Theorem~\ref{th:mainmain} is a straightforward consequence of Lemma~\ref{lemma:main}
and Lemma~\ref{l.last}.
We first make a crucial observation, and then formulate Lemma~\ref{l.last}.
As before, $\mathcal H= (H_0, H_i, L_i)^n$ denotes the canonical ear decomposition of $H(G)$.
\begin{observation}\label{obs:suprob}
Let $H', \mathcal H'= (H_0, H'_i, L'_i)^n$ be in the closure
of $H(G)$, $\mathcal H$. If $\mathcal H'$ is superb, then
$\mathcal H'$ is a robust ear decomposition.
\end{observation}
\begin{proof}
Assume $\mathcal H'$ is not robust.
First of all note that the descendant of $E_1(\mathcal{B}), E_2(\mathcal{B})$, for all basic gadgets $\mathcal{B}$ of
$H'$ consist of exactly 2 connected components and one of them is an isolated vertex
adjacent to two vertices in $V(H_0)$, since it is exactly the center of the stars $D_2(\mathcal{B}), D_3(\mathcal{B})$, respectively.
On the other hand, note that the descendant of every 3-ear $E_3(\mathcal{B})$ has at most 2 connected components, and
there is no component corresponding to an isolated vertex adjacent to vertices in $V(H_0)$.
Thus, by the assumption, for some gadget $\mathcal{G}=\mathcal{G}(\mathcal{B})$ of $H'$, there exists a 3-ear
$E_3(\mathcal{B})=L'_{m} \in \{L'_1, \ldots, L'_n\}$ such that its descendant has exactly 2 connected components
and hence, there is no path in the descendant of $E_3(\mathcal{B})$ connecting $u$ to $x$; we consider
name of the vertices as in Figure~\ref{fig:gadget}. Note that for every {\ccr}
of $I(L'_n), \ldots, I(L'_{m+1})$ we have $\alpha_1 \neq \beta_1$ and $\alpha_2 \neq \beta_2$.
Therefore, by Theorem~\ref{theo:important}, each reduction process of $\mathcal{G}$ is
$j$-incomplete and $I(E_{j-1}(\mathcal{G}))$ is a cut-obstacle for some $2\leq j \leq 4$.
This implies that $\mathcal H'$ is not superb, a contraction.
\end{proof}
\begin{lemma} \label{l.last}
Let $H', \mathcal H'= (H_0, H'_i, L'_i)^n$ be in the closure
of $H(G)$, $\mathcal H$. Assume there exists
a {\ccr}, say $\mathcal R$, of $I(L'_n), \ldots, I(L'_1)$ that makes $\mathcal H'$ superb.
Then there is a pair $H'', \mathcal H''= (H_0, H''_i, L''_i)^n$ in the closure of $H(G)$, $\mathcal H$ such that
$\mathcal H''$ is a relevant ear decomposition and $\mathcal R$
induces a {\ccr} $\mathcal R'$ of $I(L''_n), \ldots, I(L''_1)$ that makes $\mathcal H''$ superb.
\end{lemma}
\begin{proof}
Note that by Observation~\ref{obs:suprob}, we have $\mathcal H'$ is a robust ear decomposition.
In the following, we show how to obtain $H'', \mathcal H''= (H_0, H''_i, L''_i)^n$
of Lemma~\ref{l.last} from $H', \mathcal H'= (H_0, H'_i, L'_i)^n$
by modifying $\mathcal H'$ only; meaning $H'=H''$.
Since $H', \mathcal H'$ is in the closure of $H(G)$, $\mathcal H$, we get that
$\mathcal H'$ admits $H_0, S$ (see Definition~\ref{def.S}).
By Observation~\ref{o.cl},
for each basic gadget $\mathcal{B}$ of $H(G)$, the ears
\begin{equation*}\label{eq:core} E_1(\mathcal{B}),E_2(\mathcal{B}),
E_3^*(\mathcal{B}), D_1(\mathcal{B}),D_2(\mathcal{B}), D_3(\mathcal{B})\end{equation*} belong to $\{L'_1,\ldots,L'_n\}$,
where $E_3^*(\mathcal{B})$ is a 2-ear or a 3-ear that contains the fixed path of $\mathcal{B}$,
which is an element of $S$ by construction; recall that
(as it is mentioned after the construction of $H(G)$), the ear $E_1(\mathcal{B})$
is a base, $E_2(\mathcal{B})$ is an up, $E_3^*(\mathcal{B})$ is an antenna,
and by condition (3) in the second part of Definition~\ref{def.S}, the
stars $D_1(\mathcal{B}), D_2(\mathcal{B}), D_3(\mathcal{B})$ cannot be modified.
In this proof, for a given $\mathcal B$, we call the sequence of ears $E_1(\mathcal{B}),E_2(\mathcal{B}),
E_3^*(\mathcal{B})$
the \emph{core} of $\mathcal B$. Without loss of generality, from now on, we assume that
the list of ears $L'_1, \ldots, L'_n$ is ordered so that
each core forms a block. Therefore, we can consider
a natural ordering, say $\prec$, of the basic gadgets of $H(G)$:
$\mathcal{B} \prec \mathcal{B}'$ if $E_3(\mathcal{B})=L_{k}$, $E_3(\mathcal{B}')=L_{m}$ and $k<m$.
We denote by $\mathcal L$, the set of ears consisting of the cores of all basic gadgets of $H$.
We have every 3-ear of $\mathcal H'$ is in $\mathcal L$, because: $\mathcal H'$ admits $H_0, S$, which
implies that if $L$ is a 3-ear of $\{L'_1, \ldots, L'_n\}$, then $L$ is either a base, or an up, or an antenna containing
an element of $S$. Hence, by Remark~\ref{o.cl1},~$L \in \mathcal L$.
From now on, we refer to a gadget, to a gadget*, to a gadget** and to a double gadget
as a \emph{block-gadget}.
If every basic gadget of $H(G)$ is contained is some block-gadget of $\mathcal H'$,
then we have that $\mathcal H'$ is relevant and we can put $\mathcal H''= \mathcal H'$.
Therefore, we assume that the set of basic gadgets of $H(G)$ that are not contained in some block-gadget
of $\mathcal H'$ is not empty and we denote the set of such basic gadgets of $H(G)$ by $\Sigma(\mathcal H')$.
The following claim therefore proves Lemma~\ref{l.last}.
\begin{claim}\label{cl.last}
There exists an ear decomposition $\mathcal H^*$ of $H'$
such that $|\Sigma(\mathcal H^*)| < |\Sigma(\mathcal H')|$ and $R$ induces a {\ccr}, say $\mathcal R^*$, of $\mathcal H^*$. Moreover, each 3-ear in the set of building ears of $\mathcal H^*$ also belongs to $\{L'_1, \ldots, L'_n\}$.
Further, $\mathcal H^*$ is superb and
in the closure of $H(G)$, $\mathcal H$.
\end{claim}
The rest of this proof is devoted to prove Claim~\ref{cl.last}.
We obtain $\mathcal H^*$ from $\mathcal H'$ by modifying the list $L'_1, \ldots, L'_n$.
Let $\mathcal{B}$ be the first basic gadget, according to the order $\prec$, such that
$\mathcal{B} \in \Sigma(\mathcal H')$ (namely, is not contained in a block-gadget of $\mathcal H'$).
There are two possible scenarios: (1) either the ear $E_4(\mathcal{B})$ is in $\{L'_1, \ldots, L'_n\}$, (2) or not.
\begin{itemize}
\item[Case (1)] $E_4(\mathcal{B})$ is in $\{L'_1, \ldots, L'_n\}$.
\end{itemize}
We assume the first ear of $\mathcal{B}$ is $L'_r$.
If $E_4(\mathcal{B})$ is in $\{L'_1, \ldots, L'_n\}$, then $E_3^*$ is a 2-ear, otherwise,
$\mathcal{B}$ would belong to a gadget. Let $x$ be the joint vertex of $\mathcal{B}$.
Clearly, $x$ is a leaf of $E_3^*$.
This implies that there exists an ear $L \in \{L'_1, \ldots, L'_{r-1}\}$
that contains $x$ as an internal vertex, with $L$ a 1-ear or a 2-ear.
If $L$ was a 1-ear, then $\mathcal{B}$ would be contained
in a 1-gadget*.
Thus, $L$ is a 2-ear. Since $L$ is not contained in a 2-gadget*, we have that
for $I(L)=\{x,x'\}$, the vertex $x'$ is the joint or the replica vertex of a basic gadget $\mathcal{B}'\neq \mathcal{B}$.
If $x'$ was the replica vertex of $\mathcal{B}'$, then $\mathcal{B}'$ would not be contained in a block gadget (because
$E_4(\mathcal{B}')$ would not be in $\{L'_1, \ldots, L'_n\}$), which contradicts the choice of $\mathcal{B}$.
Therefore, $x'$ is a joint vertex of $\mathcal{B}'$. Since $\mathcal{B}$ is not contained in a double gadget,
then neither $E_4(\mathcal{B}')$, nor both $F_1(\mathcal{G}_{\mathcal{B}'})$, $F_2(\mathcal{G}_{\mathcal{B}'})$
belong to $\{L'_1, \ldots, L'_n\}$, where $F_1(\mathcal{G}_{\mathcal{B}'})$, $F_2(\mathcal{G}_{\mathcal{B}'})$,
denote the 1-ears of the gadget** $\mathcal{G}_{\mathcal{B}'}$ (see~(\ref{eq:1gadget**})).
Therefore, $F_2(\mathcal{G}_{\mathcal{B}'}), F'_1 \in \{L'_1, \ldots, L'_n\}$, where
$F'_1$ is a 2-ear that contains the edges of $F_1(\mathcal{G}_{\mathcal{B}'})$
and an extra edge $e$ not in $E(\mathcal{G}_{\mathcal{B}'})$. Clearly, $F'_1$
is not contained in a block-gadget.
In this case we let the list of building ears of $\mathcal H^*$ be obtained by
removing $F'_1, F_2(\mathcal{G}_{\mathcal{B}'})$ from $\{L'_1, \ldots, L'_n\}$
and adding two new ears $E_4(\mathcal{B}')$
and the 1-ear contained in $F'_1$ that contains $e$ and has end vertex
the replica vertex of $\mathcal{B}'$. Hence, $\mathcal{B}$ and $\mathcal{B}'$
belong to a double gadget of $\mathcal H^*$.
As $F'_1$ is not contained
in a block-gadget of $\mathcal H'$, we have that $\Sigma(\mathcal H^*)+2=\Sigma(\mathcal H')$. The result follows.
\begin{itemize}
\item[Case (2)] $E_4(\mathcal{B})$ is not in $\{L'_1, \ldots, L'_n\}$.
\end{itemize}
Let $\{z,u,v,y\}$ be the vertex set of $E_4(\mathcal{B})$, where $u$ is the replica vertex
of $\mathcal{B}$ (as in Figure~\ref{fig:bagadget}). As $E_4(\mathcal{B})$ is not in $\{L'_1, \ldots, L'_n\}$,
the 1-ear, say $F_2$, on vertex set $\{u,v,y\}$ is in $\{L'_1, \ldots, L'_n\}$
and a 1-ear or a 2-ear $F_1$ containing $\{z,u\}$ is in $\{L'_1, \ldots, L'_n\}$.
First let $F_1$ be a 1-ear. If $E^*_3(\mathcal{B})$ is a 2-ear, then the proof
is the same as the proof of Case~(1): just use $F_1(\mathcal{B}), F_2(\mathcal{B})$ instead of $E_4(\mathcal{B})$.
If $E^*_3(\mathcal{B})$ is a 3-ear, then $\mathcal H'$ is not robust: to see this, note that
the descendant of $E^*_3(\mathcal{B})$ has exactly
2 connected components --- since $v$ is a leaf of the star $D_1(\mathcal{B})$
and $u$ has degree 3 in $F_1 \cup F_2$, there is no path in the descendant
of $E^*_3(\mathcal{B})$ that connects the joint of $\mathcal{B}$ to a vertex from $\{u,v\}$ (see Figure~\ref{fig:gadget}) --- and none of them is an isolated vertex connected to two vertices in $V(H_0)$ (see Definition~\ref{def.wr}).
Because of Observation~\ref{obs:suprob}, this contradicts the assumption that $\mathcal H'$ is superb.
If $F_1$ is a 2-ear, then we obtain a list of building ears $\mathcal L$ by
replacing $F_2,F_1$ in $\{L'_1, \ldots, L'_n\}$, by $E_4(\mathcal{B})$ and the 1-ear $F_1 -\{z,u\}$.
Therefore, either $\mathcal{B}$ is in a block-gadget of $\mathcal L$ (which implies $\Sigma(\mathcal H^*)+1=\Sigma(\mathcal H')$), or not. If not, the result follows by Case~(1).
This finishes the proof of Claim~\ref{cl.last}.
\end{proof}
\section{Ear decompositions of 3-edge-connected cubic graphs}\label{sec:3conn}
Let $G$ be a cubic bridgeless graph.
Let $G_0, G_1, \ldots, G_l$ and $P_1, \ldots, P_l$ be two sequences of subgraphs of~$G$
such that $G_0$ is a cycle of~$G$, $G_l = G$, for each $i \in [l]$,
the subgraph $P_i$ is an ear, $V(P_i) \cap V(G_{i-1})$
is the set of leaves of $P_i$, $E(P_i) \cap E(G_{i-1})=\emptyset$, and
$G_{i}$ is the union of $G_{i-1}$ and $P_i$, that is, $V(G_{i})= V(G_{i-1}) \cup V(P_i)$ and $E(G_{i})=E(G_{i-1}) \cup E(P_i)$.
We say that $(G_0, G_1, \ldots, G_l, P_1, \ldots, P_l)$, in short $(G_0, G_i, P_i)^l$,
is an \emph{ear decomposition} of $G$. Moreover, in the case that $t \leq l$, we say that
$(G_0, G_i, P_i)^t$ is a partial ear decomposition of $G$
and if $\{P_1, \ldots, P_l\}$ also contains edges, then we refer to $(G_0, G_i, P_i)^l$
as an \emph{edge+ear decomposition} of $G$.
We now need to generalize the concept of descendant.
Let $i \in [l]$ such that $|I(P_i)|\geq 3$ and $V(P_i)= \{\alpha_i, v^1_{i},\ldots, v^{k}_{i},\beta_i\}$, where $V(P_i) \cap V(G_{i-1})=\{\alpha_i, \beta_i\}$.
We say that $S \subset \{v^1_{i},\ldots, v^{k}_{i}\}$ is a {\em segment} of $P_i$ if
$|S|\geq 3$ and $S$ induces a connected subgraph of~$P_i$.
Let $G'_i$ denote the graph obtained from $G$ by deleting $V(G_{i})$.
For each segment $S$ of $P_i$, the maximal subgraph $G^S_i$ of $G'_i$
such that $G[V(G^S_i) \cup S]$ is connected is called the \emph{descendant of~$S$}.
\begin{definition}[Super robust ear decomposition]\label{def:superrobust}
An ear decomposition, say $(G_0, G_i, P_i)^l$, of $G$ is \emph{super robust}
if the following conditions hold:
\begin{itemize}
\item[(a)] $G-V(G_0)$ is a connected graph, and
\item[(b)] for each $i \in [l]$ such that $|I(P_i)|\geq 3$ and for each segment $S$ of $P_i$,
the descendant of $S$ is connected.
\end{itemize}
\end{definition}
By the construction of the canonical ear decomposition $\mathcal{H}$
of $H(G)$ and the definition of super robust ear decompositions of cubic graphs,
it is immediate that $\mathcal{H}$ is robust if the ear decomposition of $G$ used in the construction of $\mathcal{H}$ is super robust.
Lemma~\ref{lemma:supimplrob} follows.
\begin{lemma} \label{lemma:supimplrob}
The canonical ear decomposition $H(G)$ is robust if the ear decomposition of $G$ chosen for the construction of $H(G)$ is super robust.
\end{lemma}
It is a well-know fact that the {\dcdc} conjecture holds if and only if
it holds in the class of 3-edge-connected cubic graphs.
The next lemma (whose proof is in~\cite{Cheriyan}) implies that each 3-edge-connected cubic graph admits a super robust ear decomposition.
\begin{lemma}\label{th:robust}
Let $G$ be a 3-edge-connected cubic graph.
There exist $e \in E(G)$ and an ear decomposition $(G_0, G_i, P_i)^l$ of $G-e$ such that
$P_l$ is a path of lenght 2, and for each $i \in \{1, \ldots, l-1\}$,
denoting $V_i = V(G_0) \cup V(P_1) \cup \cdots \cup V(P_i)$, the graph $G[V-V_i]$ is connected.
\end{lemma}
\section{A toroidal example}
\label{sec.surfaces}
Let $G$ be a bridgeless graph embedded on the torus obtained from the toroidal square grid $T$
by deleting a perfect matching composed of horizontal edges only in such a way that
the obtained embedding
has neither loops in the dual, nor multiple edges.
In addition, we fix a super robust ear decomposition of $G$ starting at cycle $C$
such that $C$ contains only vertical edges of $T$;
the facial cycles of the embedding are seminal for constructing such ear decomposition.
Let $G'$ be the cubic graph obtained from $G$ as described in Section~\ref{sec:main}.
We take the canonical ear decomposition $\mathcal{H}$ of $H(G')$ and discuss the question: Does $H(G'), \mathcal{H}$ satisfy Conjecture~\ref{conj:weakavoidance}?
There is a natural directed cycle double cover of $G$, which consists of all the facial cycles
defined by the embedding of $G$ on the torus. This induces a reduction process, say $\mathcal{R}$, on
each ear decomposition that belongs to the closure of $H(G'), \mathcal{H}$. In particular, a reduction process on $\mathcal{H}$.
We specify neither the details of the chosen super robust ear decomposition of $G$, nor the canonical ear decomposition of $H(G')$.
However, a critical part of any reduction process occurs at the end; namely, when
sequence of ears of $\mathcal{H}$ corresponding to the vertices of cycle $C$ are reduced.
In order to study this case, we need to assume that the ears of $\mathcal{H}$ of the part of $H(G')$
corresponding to the vertices in $G'\setminus C$ are correctly reduced (according to $\mathcal{R}$) without
creating cut-obstacles. Let $M$ denote the resulting mixed graph
and $H_M$ denote its underlying graph.
Recall that according to the construction of $H(G')$, graph $H_M$ is obtained following the next steps.
\begin{itemize}
\item First let $H_0$ be the starting cycle of $H_M$ and $V_0 = \{u, v, w\}$.
\item Secondly, let $e= \{u', v'\}$ be an edge of $C$ and let $C_0$ the path with set of edges $E(C)-\{e\}$.
We attach $C_0$ to $H_0$ by identification of $u$ with $u'$ and $v$ with $v'$.
\item Finally, for each vertex $z$ of $V(C_0)\setminus \{u',v'\}$ we introduce a basic gadget $B_z := B(x_z,u_z)$ and identify $z$ with $x_z$.
\end{itemize}
We now describe the set of arcs of $M$.
Since the previous correct reductions were performed according to $\mathcal{R}$,
we have $M$ is obtained from $H_M$ by
adding two disjoint directed cycles so that the union of their vertex sets is $\{w\} \cup \{u_z: z~\text{ vertex of } V(C_0)\setminus \{u',v'\}\}$. This finishes the construction of $M$.
The canonical ear decomposition of $\mathcal{H}$ induces an ear decomposition $\mathcal{M}$ of $H_M$.
Let $C_0 = u' z_1 \ldots z_t v'$. The ear decomposition $\mathcal{M}$ consist of gadgets $B_{z_1}, \ldots ,B_{z_{t-1}}$
and exactly one 1-gadget* $B_{z_t}$.
We claim that $\mathcal{M}$ can be modified to $\mathcal{M}'$ so that
$\mathcal{R}$ induces a complete reduction
on $\mathcal{M}'$ without cut obstacles. Indeed, $\mathcal{M}'$ is obtained as follows:
Consider two adjacent vertices $z, z'$ in $\{z_1, \ldots, z_t\}$
such that the directed cycle incident to $u_{z}$ is distinct from the directed cycle incident to $u_{z'}$;
note that these vertices always exist. Then, to obtained the ear decomposition $\mathcal{M}'$,
we replace $\mathcal{B}_{z}$, $\mathcal{B}_{z'}$
by the 2-gadget* $\mathcal{D}(\mathcal{B}_{z},\mathcal{B}_{z'})$.
|
\section{Introduction}
\label{section1_introduction}
Primordial non--Gaussianity (NG) in the cosmic microwave background
(CMB) radiation has emerged as one of the key tests for the physics of
the early Universe, as different models of e.g.\ inflation predict
slightly different deviations from Gaussian primordial fluctuations
\citep[see
e.g.][]{bartolo2004,bartolo2010,yadav2010,liguori2010,martinez2012}. The
latest constraints by the {\it Planck\footnote{{\it Planck}
(http://www.esa.int/Planck) is a project of the European Space
Agency --ESA-- with instruments provided by two scientific
consortia funded by ESA member states with contributions from
NASA.}} satellite put strong constraints on the amount of
primordial NG that is present in the data
\citep{planck_xxiv_2013}. But the precision of the {\it Planck} data
requires great care concerning the subtraction of astrophysical
contributions to the observed CMB anisotropies (so--called
foregrounds). It is important to check all possible contributions for
their expected level of contamination of the primordial NG estimate
both on sky maps and on foreground-cleaned maps. At least the
non-negligible foreground contributions should then be estimated
jointly with the primordial ones, which requires the construction of
an estimator also for the foregrounds.
Conventionally the CMB anisotropies $\Delta T(\hat{\mathbf n})$, being a real-valued random field
on the sky sphere, are expanded in spherical harmonics,
\begin{equation}
\Delta T(\hat{\mathbf n}) = \sum_{\ell m} a_{\ell m} Y_{\ell m}(\hat{\mathbf n})
\end{equation}
and schematically we can write the coefficients $a_{\ell m}$ as a superposition of different
contributions,
\begin{equation}
a_{\ell m} = \tilde{a}^{({\mathtt{CMB}})}_{\ell m} + a^{({\mathtt{fg}})}_{\ell m} + n_{\ell m} \, . \label{eq:contributions}
\end{equation}
Here the first term on the right hand side $\tilde{a}_{\ell m}$ is the
primordial contribution, lensed by the intervening large--scale
structure. The second term is the contribution due to foregrounds --
in general there are both Galactic and extragalactic contributions,
but in this paper we will neglect the former and use the term
`foreground' to denote the extragalactic contribution only. For the
extragalactic foreground radiation we expect that radio sources
dominate at low frequencies, and dusty star--forming galaxies creating
the cosmic infrared background (CIB) at high frequencies. The final
contribution is instrumental noise, obviously uncorrelated with the
CMB and the foreground contributions, that we assume to be Gaussian.
The main tool to study primordial NG is the angular bispectrum, the
three--point function of the $a_{\ell m}$,
\begin{equation}
\langle a_{\ell_1 m_1} a_{\ell_2 m_2} a_{\ell_3 m_3} \rangle = G_{\ell_1 \ell_2 \ell_3}^{m_1 m_2 m_3} b_{\ell_1 \ell_2 \ell_3} \, ,
\label{bispectrum_three_alm}
\end{equation}
where the Gaunt integral \citep{komatsu2001}
\begin{equation}
G_{\ell_1 \ell_2 \ell_3}^{m_1 m_2 m_3} = \int d^2\hat{\mathbf n} Y_{\ell_1 m_1}(\hat{\mathbf n}) Y_{\ell_2 m_2}(\hat{\mathbf n}) Y_{\ell_3 m_3}(\hat{\mathbf n})
\label{eq:gaunt_integral}
\end{equation}
takes care of rotational symmetry, and where the non-trivial
contribution to the three-point function is encoded in the reduced
bispectrum $b_{\ell_1 \ell_2 \ell_3}$.
The product of three $a_{\ell m}$ as written in Eq.\
(\ref{eq:contributions}) will not only contain a primordial
contribution. In addition there are additional elements that involve
non-primordial terms, some of them already studied in previous works,
such as the `foreground' contribution given schematically by $\langle
a^{{({\mathtt{fg}})} 3}_{\ell m} \rangle$ \citep[see
e.g.][]{lacasa2013,penin2013}, a contribution from the correlation
between the lensing of the CMB and the integrated Sachs--Wolfe (ISW)
effect contained in the $\langle \tilde{a}^{({\mathtt{CMB}}) 3}_{\ell m}
\rangle$ term \citep[see e.g.][]{Mangilli:2013sxa} and finally a
contribution from the correlation between the lensing of the CMB and
extragalactic foregrounds. This article is focusing on the last
correlation, already detected with a significance of 42$\sigma$ by
\cite{planck_xviii_2013} considering statistical errors only
(19$\sigma$ when systematics are included). The CIB--lensing
correlation arises as the large--scale structure (LSS) both lenses the
CMB and emits the `foreground' (radio or CIB) radiation. The main
contribution here is expected due to the CIB--lensing correlation, as
in radio galaxies the clustering signal is highly diluted by the
broadness of their luminosity function and of their redshift
distribution \citep[e.g.,][]{tof05}.
In this paper we focus on the CIB--lensing bispectrum, for two
reasons. Firstly, in order to ensure that the CMB constraints on
primordial NG are accurate, we need to check that the additional
contributions are under control. In \cite{planck_xxiv_2013} the
ISW--lensing contribution was fit simultaneously with the primordial
contribution, and was shown to be small. \cite{lacasa2012a} and
\cite{curto2013} studied the bispectrum of unresolved point sources
and concluded that it is small enough to neglect. However, the
situation of the CIB--lensing contribution was so far not investigated
in detail, and this paper aims to close this gap. Secondly, probing
the non--Gaussianity due to the large--scale structure is not only
important for assessing the contamination of the primordial NG, but is
also interesting in its own right. The LSS contains important
information on the late--time evolution and content of the universe,
as well as on the formation and evolution of galaxies. Studying the
CIB--lensing correlation is thus not only important to assess the
reliability of the constraints on primordial $f_{\rm nl}$, but also
potentially useful for cosmology and astrophysics.
The outline of the paper is as follows: in Section \ref{section2} we
discuss the CIB--lensing contribution as well as the CIB model that we
will use. In Section \ref{section3_delta_fnl} we estimate the bias on
$f_{\rm nl~}$ for local, equilateral and orthogonal configurations. We perform
the calculation both for raw frequency maps and for linear
combinations of maps that remove most of the astrophysical
foregrounds, following the SEVEM component separation method used by
the {\it Planck}~ collaboration \citep[see, e.g.][]{planck_xii_2013}. We
then construct in Sections \ref{section:detectability} and
\ref{section5} an optimal estimator for measuring the CIB--lensing
bispectrum and assess the level at which we expect to be able to
detect it with {\it Planck} data before presenting our conclusions in
Section \ref{section6}. In the four attached appendices we provide
more details on our model of CIB anisotropies power spectra
(including a new parameter fit of the {\it Planck} measurement of
CIB power spectra) and on the calculation of power spectra and
bispectra.
\section{Modelling the CIB and CIB--lensing power spectra}
\label{section2}
\begin{figure*}
\centering
\includegraphics[width=80mm]{f1.eps}
\caption{Predicted power spectra of the lensing correlation with ISW (red
curve), CIB (black curves) and radio sources (green curves) at the
{\it Planck}~ frequencies relevant for cosmological analysis.}
\label{f1}
\end{figure*}
\begin{figure}
\includegraphics[width=90mm]{f2a.eps}
\includegraphics[width=90mm]{f2b.eps}
\includegraphics[width=90mm]{f2c.eps}
\includegraphics[width=90mm]{f2d.eps}
\caption{Predicted bispectra for the local ({\it upper panels}) and
equilateral ($\ell_1=\ell_2=\ell_3$; {\it lower panels})
configurations. Contributions are as in Figure\,\ref{f1}, plus
bispectra from radio sources (cyan curves) and IR galaxies (blue
curves). Frequencies are indicated inside the plots. In the {\it
upper left panel}, ``local'' bispectra are plotted at
different frequencies (from 70 to 217\,GHz)
and for $\ell_1=\ell_2$ and $\ell_{3}=\ell_{\rm min}=2$; in
the {\it upper right panel}, they are plotted at 217\,GHz
for $\ell_3=\ell_{\rm min}=2,~50,~200$.}
\label{f2}
\end{figure}
The interplay between the CMB gravitational lensing and CIB intensity
fluctuations are studied in this work in terms of the cross-bispectrum
(the so called CIB--lensing bispectrum), its detectability levels and
the bias on the primordial non-Gaussianity through the $f_{\rm nl~}$
parameter. The observed temperature fluctuations, neglecting other
foreground sources, can be expanded at first order as \citep[see
e.g.][]{Goldberg:1999xm,hu2000}:
\begin{equation}
\Delta T(\hat{\mathbf n}) = \Delta T_{{\mathtt{CMB}}}(\hat{\mathbf n} + \nabla \phi (\hat{\mathbf n})) +
\Delta T_{{{\mathtt{CIB}}}} (\hat{\mathbf n}) \simeq \Delta T_{{\mathtt{CMB}}} (\hat{\mathbf n}) + \nabla
\big(\Delta T_{{\mathtt{CMB}}} (\hat{\mathbf n})\big) \nabla \phi (\hat{\mathbf n}) +
\Delta T_{{\mathtt{CIB}}} (\hat{\mathbf n})\,,
\end{equation}
where $\phi (\hat{\mathbf n})$ is the lensing potential, $\Delta T_{{\mathtt{CMB}}}(\hat{\mathbf n})$
are the primordial CMB anisotropies and $\Delta T_{{\mathtt{CIB}}}(\hat{\mathbf n})$ are the
anisotropies due to the CIB. Going into the spherical harmonic space,
the observed anisotropies are:
\begin{equation}
a_{\ell m} = \tilde{a}^{({\mathtt{CMB}})}_{\ell m} + a^{({\mathtt{CIB}})}_{\ell m} =
a^{({\mathtt{CMB}})}_{\ell m} +\sum_{\ell' m' \ell'' m''}(-1)^m
G_{\ell \ell' \ell''}^{m m' m''}
\Big[\frac{\ell'(\ell'+1)-\ell(\ell+1)+\ell''(\ell''+1)}{2}
a^{({\mathtt{CMB}})}_{\ell' m'}\phi_{\ell'' m''} \Big] + a^{({\mathtt{CIB}})}_{\ell m},
\label{almcmb}
\end{equation}
where $\tilde{a}^{({\mathtt{CMB}})}_{\ell m}$, $a^{({\mathtt{CMB}})}_{\ell m}$,
$a^{({\mathtt{CIB}})}_{\ell m}$ and $\phi_{\ell m}$ are the spherical harmonic
coefficients of the observed/primordial CMB, CIB and gravitational
potential anisotropies, respectively, and $G_{\ell \ell' \ell''}^{m m' m''}$
is the Gaunt coefficient (see Eq.\ \ref{eq:gaunt_integral}).
The angular power spectra of CIB fluctuations and of their
cross--correlation with the CMB lensing are typically written in the
Limber approximation as \citep[e.g.,][see also Appendix
\ref{appendix_cib_lensing} for a full derivation]{son03}:
\begin{eqnarray}
C_{\ell}^{({\mathtt{CIB}})}(\nu,\nu') =\langle a^{({\mathtt{CIB}})*}_{\ell
m}a^{({\mathtt{CIB}})}_{\ell m}\rangle & = & \int_0^{\chi_*}{d\chi \over
\chi^2}\, W_{\nu}^{({\mathtt{CIB}})}(\chi)W_{\nu'}^{({\mathtt{CIB}})}(\chi)\,
P_{\rm gg}(k=\ell/\chi,\chi)\,; \nonumber \\
C_{\ell}^{({\mathtt{CIB-Lens}})}(\nu) =\langle\phi^*_{\ell
m}a^{({\mathtt{CIB}})}_{\ell m}\rangle & = & \int_0^{\chi_*} {d\chi \over
\chi^2}\,W_{\nu}^{({\mathtt{CIB}})}(\chi) W^{({\mathtt{Lens}})}(k,\chi)\,P_{\delta
{\rm g}}(k=\ell/\chi,\chi)\,.
\label{s2e1}
\end{eqnarray}
The integral is over the comoving distance $\chi$ along the line of
sight, and extends up to the comoving distance of the last scattering
surface $\chi=\chi_*$ (in practice the integral is computed up to
redshift 7 because of the negligible contribution of CIB
fluctuations at higher redshifts). The $W^{({\mathtt{CIB}})}(\chi)$ and
$W^{({\mathtt{Lens}})}(k,\chi)$ functions are the redshift weights for CIB
fluctuations and for the lensing potential $\phi$, respectively,
\begin{equation}
W_{\nu}^{({\mathtt{CIB}})}(\chi)=a(\chi)\,\bar{j}_{\nu}(\chi)~~~~~~~~~~~~
W^{({\mathtt{Lens}})}(k,\chi)=3{\Omega_m \over a(\chi)}\bigg({H_0 \over ck}\bigg)^2
{\chi_*-\chi \over \chi_*\chi}\,,
\label{s2e2}
\end{equation}
where $a(\chi)$ is the scale factor and $\bar{j}_{\nu}(\chi)$ is the
mean CIB emissivity at frequency $\nu$ as a function of $\chi$:
\begin{equation} \bar{j}_{\nu}(\chi)=(1+z)\bigg({d\chi \over dz}\bigg)^{-1}
\int_0^{S_c}\,S{d^2N \over dSdz}dS\,.
\label{ss2e3}
\end{equation}
Here $d^2N/dSdz$ denotes galaxies number counts per interval of
flux density and redshift, and $S_c$ is the flux limit above which
sources are subtracted or masked\footnote{Hereafter, we use as
flux limit for the {\it Planck}~ mission the 90\% completeness level of
the {\it Planck}~ Catalogue of Compact Sources, given in
\citet{planck_xxviii_2013}.}. We compute the redshift evolution of
the CIB emissivity from the model of galaxy evolution of
\citet{bet11}\footnote{We use number counts of the so--called {\it
mean model}, see the http://www.ias.u-psud.fr/irgalaxies/ web
page.}. This is a backward evolution model based on parametric
luminosity functions for two populations of galaxies: normal and
starburst galaxies.
It uses spectral energy distribution templates for the two galaxy
populations taken from the \citet{lag04} library. The model is
described by 13 free parameters and the best--fit values are computed
using observational number counts and luminosity functions from
mid--infrared to millimetre wavelengths \citep{bet11}. This CIB model
was previously used by \citet{pen12} before and then applied to
{\it Planck}~ results \citep{planck_xviii_2011,planck_xxx_2013} in order to
compute CIB and CIB--lensing power spectra.
In Eq.\ (\ref{s2e1}), $P_{\rm gg}(k,\chi)$ and $P_{\delta {\rm
g}}(k,\chi)$ are respectively the 3D power spectrum of galaxies
and of the cross--correlation between galaxies and the dark matter
(DM) density field. In the context of the halo model
\citep{sch91,sel00,sco01,coo02}, the power spectra are the sum of the
contribution of the clustering in one single halo (1--halo term) and
in two different halos (2--halo term):
\begin{eqnarray}
P_{\rm gg}(k)=P_{\rm gg}^{1h}(k)+P_{\rm gg}^{2h}(k) & &
P_{\delta {\rm g}} (k)=P_{\delta {\rm g}}^{1h}(k)+
P_{\delta {\rm g}}^{2h}(k) \nonumber \\
P_{\rm gg}^{1h}(k)=\int\,dM\,n(M){\langle N_{\rm gal}(N_{\rm gal}-1)\rangle
\over \bar{n}^2_{\rm gal}}u^2(k,M) & &
P_{\delta {\rm g}}^{1h}(k)=\int\,dM\,n(M){M \over \bar{\rho}}
{\langle N_{\rm gal}\rangle \over \bar{n}_{\rm gal}}u^2(k,M) \nonumber \\
P_{\rm gg}^{2h}(k)=P_{\rm lin} (k)\bigg[\int\,dM\,n(M)b(M){\langle
N_{\rm gal}\rangle \over \bar{n}_{\rm gal}}u(k,M)\bigg]^2 & &
P_{\delta {\rm g}}^{2h}(k)=P_{\rm lin}(k)\,\int\,dM_1\,n(M_1)b(M_1)
{M_1 \over \bar{\rho}}u(k,M)\,\times \nonumber \\
& & ~~~~~~~~~~~~~~~~~~ \times \int\,dM_2\,n(M_2)b(M_2){\langle
N_{\rm gal}\rangle \over \bar{n}_{\rm gal}}u(k,M_2)
\label{b0}
\end{eqnarray}
The main inputs required for the calculations of $ P_{\rm gg}(k)$ and
$ P_{\delta {\rm g}}(k)$ are: (i) the mass function $ n(M)$ of DM
halos --we use the mass function fit of \citet{tin08} with its
associated prescription for the halo bias, $ b(M)$
\citep[see][]{tin10}--; (ii) the distribution of DM within halos, $
u(k,M)$, --we use the NFW \citep{nav97} profile--; and (iii) the Halo
Occupation Distribution (HOD), that is a statistical description of
how DM halos are populated with galaxies. We model the HOD using a
central-satellite formalism \citep[e.g.,][]{kra04,zhe05}: it
introduces a distinction between central galaxies, which lies at the
centre of the halo, and satellite galaxies that populate the rest of
the halo and are distributed in proportion to the halo mass
profile. The mean number of galaxies in a halo of mass $ M$ is thus
written as $ \langle N_{{\rm gal}} \rangle=\langle N_{\rm cen}
\rangle+\langle N_{\rm sat} \rangle$. Following \citet{tin10b}, the
mean occupation functions of central and satellite galaxies are:
\begin{equation}
\langle N_{\rm cen} \rangle={1 \over 2}\Bigg[1+{\rm erf}\Bigg({\log
M-\log M_{\rm min}
\over \sigma_{\log M}}\Bigg)\Bigg]\,,
\label{b1}
\end{equation}
and
\begin{equation}
\langle N_{\rm sat} \rangle={1 \over 2}\Bigg[1+{\rm erf}\Bigg({\log
M-\log 2M_{\rm min}
\over \sigma_{\log M}}\Bigg)\Bigg]\Bigg({M \over M_{\rm sat}}
\Bigg)^{\alpha_{\rm sat}}\,,
\label{b2}
\end{equation}
where $ M_{\rm min}$, $ \alpha_{\rm sat}$, $ M_{\rm sat}$ and
$ \sigma_{\log M}$ are free parameters. Within this parametrisation,
most of the halos with $ M\ga M_{\rm min}$ contain a central
galaxy. For satellite galaxies the mass threshold is chosen to be
twice $ M_{\rm min}$, so that halos with a low probability of having a
central galaxy are unlikely to contain a satellite galaxy. The number
of satellite galaxies grows with a slope $ \alpha_{\rm sat}$ for
high--mass halos. Moreover, assuming a Poisson distribution for
$N_{\rm sat}$, we can write $ \langle N_{\rm gal}(N_{\rm gal}-1)\rangle=2\langle
N_{\rm sat}\rangle+\langle N_{\rm sat}\rangle^2$ \citep{zhe05}. Finally, in
Eq.\,(\ref{b0}), $ \bar{\rho}$ is the background density, $ \bar{n}_{\rm gal}$
is the mean number of galaxies given by $ \int dMn(M)\langle
N_{\rm gal}\rangle$, and $ P_{\rm lin}$ is the linear DM power spectrum.
Following the analysis in \citet{planck_xviii_2011}, we restrict the
free HOD parameters to only $ M_{\rm min}$ and $ \alpha_{\rm sat}$ by
imposing $ M_{\rm sat}=3.3M_{\rm min}$ and $ \sigma_{\log M}=0.65$.
Moreover, because of the uncertainty of the evolution model of
galaxies at high redshift, the effective mean emissivity $ j_{\rm eff}$
at $ z>3.5$ is also constrained from data as an extra free parameter
\citep[see Appendix \ref{appendix_hod_parameters},
and][]{planck_xviii_2011}. We find the best--fit values of the model
parameters (i.e., $ M_{\rm min}$, $ \alpha_{\rm sat}$ and $ j_{\rm eff}$)
using the recent {\it Planck}~ measurements of the CIB power spectra
\citep{planck_xxx_2013}: the results are in good agreement with values
of \citet{planck_xviii_2011}. Details and results of the analysis are
provided in Appendix \ref{appendix_hod_parameters}. As shown in
Figures\,\ref{fb1}--\ref{fb2} of Appendix
\ref{appendix_hod_parameters}, the model is able to reproduce in a
quite remarkable way {\it Planck}~ measurements both for the auto-- and
cross-- CIB spectra and for the CIB--lensing spectra.
\subsection{Computing CIB and CIB--lensing spectra at very large scales}
At the very large scales, i.e. at $\ell\la10$, the Limber
approximation used in Eq.\ (\ref{s2e1}) is not further valid. Here we
provide the general expression for CIB and CIB--lensing power spectra,
that we use for the angular scales ranging from $\ell=2$ to 40:
\begin{eqnarray}
C^{({\mathtt{CIB}})}_{\ell}(\nu,\nu') & = & {2 \over \pi}\int\,dk\,k^2\,
\int_0^{\chi_*}d\chi\,W_{\nu}^{({\mathtt{CIB}})}(\chi)j_{\ell}(k\chi)
P^{1/2}_{\rm gg}(k,\chi)
\int_0^{\chi_*}d\chi'\,W_{\nu'}^{({\mathtt{CIB}})}(\chi')j_{\ell}(k\chi')
P^{1/2}_{\rm gg}(k,\chi')
\nonumber \\
C^{({\mathtt{CIB-Lens}})}_{\ell}(\nu) & = & {2 \over \pi}\int\,dk\,k^2\,
\int_0^{\chi_*}d\chi\,W_{\nu}^{({\mathtt{CIB}})}(\chi)j_{\ell}(k\chi)
P^{1/2}_{\rm gg}(k,\chi)\,
\int_0^{\chi_*}d\chi'\,W^{({\mathtt{Lens}})}(k,\chi')j_{\ell}(k\chi')
P^{1/2}_{\delta\delta}(k,\chi')\,,
\label{ss2e1}
\end{eqnarray}
where $j_{\ell}(x)$ are the spherical Bessel functions and
$P_{\delta\delta}(k,\chi)$ is the power spectrum of the DM density
field respectively. The expression for the CIB--lensing correlation
power spectrum has been derived in Appendix
\ref{appendix_cib_lensing}, following the procedure developed in
\citet{lewis2006} for the ISW--Lensing power spectrum (the CIB
spectrum can be derived in a similar way). We have verified that in
the Limber approximation power spectra are typically overestimated by
a factor $1.5$ at $\ell=2$.
\subsection{Correlation of radio sources with the CMB lensing}
Intensity fluctuations produced by extragalactic radio sources at
cm/mm wavelengths are dominated by the shot--noise term due to bright
objects. The contribution from the radio sources clustering is
expected to be significant for faint objects, i.e. for flux densities
$S\la10$\,mJy \citep{gon05,tof05}. This is actually observed in
low--frequency surveys like the NVSS survey \citep{con98}, which have
been found to be fair tracers of the underlying density field at
redshifts $z\la2$ \citep[see,
e.g.,][]{bou05,vie06,planck_xix_2013}. Therefore, although small, we
expect some level of correlation between the signal from radio
sources and the CMB lensing potential, which is primarily induced by
dark matter halos at $1\la z\la3$.
In order to estimate this contribution we use the same formalism
as for the CIB. The radio--lensing power spectrum is given
therefore by
\begin{equation}
C_{\ell}^{({\mathtt{Radio-Lens}})}(\nu)=\int_0^{\chi_*} {d\chi \over
\chi^2}\,W_{\nu}^{({\mathtt{Radio}})}(\chi) W^{({\mathtt{Lens}})}(k,\chi)\,P_{\delta
{\rm g}}(k=\ell/\chi,\chi)~~~~~~~{\rm with}~~~
W_{\nu}^{({\mathtt{Radio}})}(\chi)=a(\chi)\bar{j}_{\nu}(\chi)\,.
\label{ss2e2}
\end{equation}
The mean emissivity of radio sources is computed from
Eq. (\ref{ss2e3}) with number counts $d^2N/dSdz$ provided by the model
described in \citet{tuc11}. We estimate the integral starting
from $S=10^{-5}\,$Jy, which nearly corresponds to the limit of
validity of the model. At lower flux densities, we expect number
counts to have a break at $\sim \mu$Jy, and that the contribution from
fainter sources, although maybe not completely negligible, should not
affect the conclusions of our analysis.
The power spectrum $P_{\delta {\rm g}}(k,\chi)$ is computed as the sum
of the 1--halo and 2--halo terms, see Eq.\ (\ref{b0}). Unlike the CIB,
we assume that the mean number of galaxies per halo, $\langle N_{\rm
gal} \rangle$, is equal to 1 if the halo mass is larger than some
threshold $M_{\rm min}$, and otherwise is zero. This choice is
motivated by the fact that we are taking into account only the most
powerful radio objects that are typically associated to the centre of
dark matter halos. This assumption also agrees with results from
\citet{mar13} for the NVSS survey: they found that the average number
of galaxies within a halo of mass $M$ can be described by a step
function with the mass threshold in the range $ 12.3\la\log(M_{\rm
min}/M_{\odot})<12.4$. We take $ \log(M_{\rm
min}/M_{\odot})=12.34$. We have also verified that our results are
only weakly dependent on the value of $M_{\rm min}$. The
radio-lensing power spectra for this parametrization are shown in
Fig.\ \ref{f1}. \citet{cha12} studied the HOD of AGNs using
cosmological hydrodynamic simulations: they found that the mean
occupation function can be modelled as a softened step function for
central AGNs (same as our Eq.\,\ref{b1}) and as a power law for
satellite AGNs. Using their occupation functions for redshift $z=1$
and for the brightest sources ($L_{\rm bol}\ge10^{42}$\,erg\,s$^{-1}$;
see their Table\,2), we find radio--lensing power spectra in very good
agreement with the ones shown in Fig.\,\ref{f1}. On the other hand,
their occupation functions for fainter AGNs give significantly lower
power spectra.
We want to stress however that our estimates for the radio--lensing
power spectra should be taken only as indicative of the level of the
signal, due to the large uncertainties in modelling radio sources. It
is outside of the aim of this work to provide more accurate predictions
for this component.
\subsection{Forecasts for non--primordial bispectra}
Due to the strong clustering of Infrared (IR) galaxies, CIB
fluctuations produce a non--constant bispectrum that we compute with
the following prescription \citep[see
e.g.][]{arg03,lacasa2012a,curto2013}:
\begin{equation}
b^{({\mathtt{CIB}})}_{\ell_1\ell_2\ell_3}=b^{({\mathtt{CIB}})}_{{\rm sn}}\,
\sqrt{{C^{({\mathtt{CIB}})}_{\ell_1}C^{({\mathtt{CIB}})}_{\ell_2}C^{({\mathtt{CIB}})}_{\ell_3} \over
(C^{({\mathtt{CIB}})}_{{\rm sn}})^3}}\,,
\label{s2e4}
\end{equation}
where $C^{({\mathtt{CIB}})}_{{\rm sn}}$ and $b^{({\mathtt{CIB}})}_{{\rm sn}}$ are the shot--noise
contributions to CIB power spectra and bispectra (see
Appendix\,\ref{appendix_sn}).
Additionally the coupling of the weak lensing of the CMB with CIB
anisotropies leads also to a bispectrum that is, in the reduced form,
given by \citep{Goldberg:1999xm,coo00,lewis2011}
\begin{equation}
b^{({\mathtt{CIB-Lens}})}_{\ell_1\ell_2\ell_3}={\ell_1(\ell_1+1)-\ell_2(\ell_2+1)
+\ell_3(\ell_3+1) \over
2}\,\tilde{C}^{({\mathtt{CMB}})}_{\ell_1}\,C^{({\mathtt{CIB-Lens}})}_{\ell_3} +(5~\rm{perm})\,,
\label{s2e3}
\end{equation}
where $\tilde{C}^{({\mathtt{CMB}})}_{\ell}$ is the lensed CMB power
spectrum\footnote{We have computed the lensed and unlensed power
spectra and the CMB-lensing cross-spectrum using the cosmological
parameters that best fit the combined WMAP and {\it Planck}~ 2013 data
\citep[referred as 'Planck+WP+highL' in][]{planck_xvi_2013} using the
latest version of CAMB \citep{lewis2000}.}. Eq.\ (\ref{s2e3}) was
derived for the first time by \citet{Goldberg:1999xm} for a generic
tracer of the matter distribution expanding lensed CMB temperature
fluctuations $\Delta T(\hat{\mathbf n}) = \Delta T(\hat{\mathbf n}+\nabla\phi (\hat{\mathbf n}))$ to
the first order in $\phi$. In the original form they used the unlensed
$C^{({\mathtt{CMB}})}_{\ell}$ in the right-hand part of the
equation. \citet{lewis2011} showed instead that, when higher--order
terms are taken into account, the correct equation requires to use the
lensed CMB power spectrum.
Bispectra induced by the correlation of the CMB lensing potential with
a tracer of the matter distribution differ only for the shape of the
cross--power spectrum of the lensing and the matter tracer. In
Fig.\ \ref{f1} we compare the cross--power spectra for the case of
ISW and extragalactic sources. We see that the ISW--lensing
power spectrum is the most relevant one on scales larger than few
degrees (it is 2--3 order of magnitude larger than other spectra at
$\ell=2$), but rapidly decreases at $\ell\ga100$. On these scales the
CIB--lensing power spectrum dominates even at frequencies as low as
100--143\,GHz. On the contrary, as expected, the radio--lensing
correlation produces just a sub-dominant contribution at all the
angular scales and therefore it will not be taken into account in the
following analysis.
The ISW--lensing correlation is found to be a significant contaminant
for {\it Planck}~ mainly on local primordial NG: \citet{planck_xxiv_2013}
estimated a bias on $f_{\rm nl~}$ of 7.1, 0.4 and $-$22 (1.22, 0.01, $-$0.56 in
$\sigma$ units) for the local, equilateral and orthogonal shape,
respectively. In Fig. \,\ref{f2} we compare the non--primordial
bispectra for the local and the equilateral configurations at
frequencies between 70 and 217\,GHz (the orthogonal and equilateral
configurations are very similar). We can see that for the local shape
the ISW--lensing bispectrum is about two orders of magnitude larger
than the CIB--lensing contribution for $ \ell_{\rm
min}=2$. However, whereas the latter changes very moderately with
$ \ell_{\rm min}$, the ISW--lensing bispectrum is strongly reduced
increasing $ \ell_{\rm min}$, e.g., by a factor $ \sim10^2$ and
$ 10^3$ for $ \ell_{\rm min}=50$ and 200 respectively. For the
other shapes, the ISW--lensing bispectrum decreases rapidly with the
angular scale $ \ell$, and the dominant contributions come from the
CIB--lensing correlation at $ \ell\ga100$ and from extragalactic
sources at very small scales ($ \ell\ga1000$). Fig. \,\ref{f2}
shows that the CIB and its correlation with the CMB lensing could be a
non-negligible contaminant in NG studies, and motivates the following
deeper analysis.
Finally, in Fig.\ \ref{f2} we also consider the different
contributions for the equilateral shape at the highest {\it Planck}~
frequencies. This is interesting in terms of a possible detection of
the CIB-lensing bispectrum. Due to its strong signal at these
frequencies, it should be detectable with high significance for
{\it Planck}~ at $\nu\ge217$\,GHz. However, IR galaxies give rise themselves
to a strong contribution at high frequencies and they can be therefore
a strong contaminant for the detection of the CIB--lensing
bispectrum. We discuss later how to tackle this problem.
\section{The CIB-lensing bispectrum bias on the primordial non-Gaussianity}
\label{section3_delta_fnl}
\begin{figure*}
\includegraphics[width=90mm]{f3.eps}
\caption{The SEVEM component separation weights $w^{ (\nu_i)}_\ell$
for the combination of 143 and 217 GHz maps. Please notice that the
weights include the deconvolution/convolution process to reach the
final 5 arcmin resolution.}
\label{f3}
\end{figure*}
To continue with our study of the bispectra presented above, we
consider three scenarios for the estimation of the $f_{\rm nl~}$ bias due to
the CIB-lensing correlation: (i) raw per-frequency maps, which in
particular contain CMB lensed signal plus CIB and radio point source
contributions plus instrumental noise, (ii) foreground-reduced (clean)
maps per frequency and (iii) a combination of clean maps. Galactic
foregrounds are not taken into account. We use {\it Planck}~ ``ideal''
instrumental characteristics -- i.e. isotropic noise, spherically
symmetric beams, full sky coverage -- that are summarized in
Table\,\ref{t1}. As a representative component separation technique
already used by the {\it Planck}~ collaboration, and for reason of
simplicity, we select SEVEM
\citep{leach2008,fernandezcobos2011,planck_xii_2013}. This cleaning
technique is based on a template fitting approach. The templates used
by SEVEM are constructed using only {\it Planck}~ data and there are no
assumptions on the foregrounds or noise levels. The templates are
constructed by taking the difference of two close {\it Planck}~ frequency
maps previously smoothed to a common resolution\footnote{E.g. the
44-70 template would be constructed by subtracting the 44 and 70 GHz
maps previously smoothed to a common beam defined as the product of
the beams of the two maps in spherical harmonics space,
$b^{\nu=44}_{\ell}$ and $b^{\nu=70}_{\ell}$.}. This template is
therefore clean of CMB signal. The SEVEM foreground-reduced map at a
given frequency is computed by subtracting from the raw map at that
frequency a linear combination of selected templates. The linear
coefficients are computed by minimising the variance of the final
map. SEVEM is linear and therefore the $a_{\ell m}$ coefficients of a
cleaned map can be written as a linear combination of the coefficients
of the raw maps involved in the cleaning process. The SEVEM cleaned
maps used in this paper are the 143 and 217 GHz maps, also considered
in the non-Gaussianity analyses performed in
\citet{planck_xxiii_2013,planck_xxiv_2013}. These two maps are
computed with 4 templates -- two corresponding to the LFI channels,
namely the 30-44 and 44-70 templates, and two corresponding to the HFI
channels, namely the 545-353 and the 857-545 templates. These
templates take into account different Galactic and extra-Galactic
foregrounds at low and high frequencies and their residual amplitude
present in the data is minimised.
The spherical harmonic coefficients of maps for the three
cases mentioned above are given by:
\begin{itemize}
\item (i) {\it Planck}~ raw maps per frequency
\begin{equation}
a^{({\mathtt{raw}},{\nu})}_{\ell m}=\big[\tilde{a}^{({\mathtt{prim}})}_{\ell m}+a^{({\mathtt{CIB}},\nu)}_{\ell m}+a^{({\mathtt{Radio}},\nu)}_{\ell m}\big]b^{(\nu)}_\ell + a^{({\mathtt{noise}})}_{\ell m},
\label{alm_raw_maps}
\end{equation}
\item (ii) {\it Planck}~ clean maps per frequency
\begin{equation}
a^{({\mathtt{clean}},\nu)}_{\ell m}=\sum_{i=1}^{9}f^{(\nu)}_{\nu_i} a^{({\mathtt{raw}},\nu_i)}_{\ell m},
\label{sevem_alm_per_freq}
\end{equation}
\item (iii) {\it Planck}~ combined clean map
\begin{equation}
a^{({\mathtt{comb}})}_{\ell m}=\sum_{i=5}^{6}{w^{(\nu_i)}_\ell a^{({\mathtt{clean}},\nu_i)}_{\ell m}}.
\label{sevem_alm_comb}
\end{equation}
\end{itemize}
where $b^{(\nu)}_\ell$ is the beam for each frequency channel $\nu$,
$f^{(\nu)}_{\nu_i}$ are the SEVEM component separation weights per
frequency $\nu$, and $w^{(\nu_i)}_\ell$ are the SEVEM component
separation weights for the combined map. We use the raw maps in the
frequency range between 100 and 353 GHz. Frequencies lower than 100
GHz have a negligible IR contribution. At frequencies higher than 353
GHz, the CIB signal is clearly dominant over the CMB and the
corresponding cosmic variance completely masks the CIB--lensing
signal. The low and high frequency maps are nonetheless useful as
templates to clean the central frequency maps in Eq.\
(\ref{sevem_alm_per_freq}) where the sum runs over all nine {\it
Planck} frequencies from frequency 1 = 30 GHz to frequency 9 = 857
GHz. These maps are only produced for 143 GHz (frequency 5) and 217
GHz (frequency 6). The weights for the SEVEM templates needed to
construct the foreground-reduced maps are given in Table \ref{t1b} and
are based on \citet[][]{planck_xii_2013}. Finally the SEVEM combined
map is computed using the weights given in Fig. \ref{f3} following
Eq. (\ref{sevem_alm_comb}), reaching a resolution of 5 arcmin.
\subsection{Power spectrum}
The power spectrum for the three considered cases is:
\begin{itemize}
\item (i) {\it Planck}~ raw maps per frequency
\begin{equation}
C^{(\nu)}_{\ell} = \big[b^{(\nu)}_\ell\big]^2\big[\tilde{C}^{({\mathtt{CMB}})}_{\ell}+C^{({\mathtt{CIB}},\nu)}_{\ell}+C^{({\mathtt{Radio}},\nu)}_{\ell}\big]+C^{({\mathtt{noise}},\nu)}_{\ell},
\label{power_spectrum_i}
\end{equation}
\item (ii) {\it Planck}~ cleaned maps per frequency
\begin{equation}
C^{({\mathtt{clean}},\nu)}_{\ell} = \sum_{\{i,j\}=1}^{9}f^{(\nu)}_{\nu_i}f^{(\nu)}_{\nu_j}b^{(\nu_i)}_\ell b^{(\nu_j)}_\ell\big[\tilde{C}^{({\mathtt{CMB}})}_{\ell}+C^{({\mathtt{CIB}},\nu_i,\nu_j)}_{\ell}+C^{({\mathtt{Radio}},\nu_i,\nu_j)}_{\ell}\big]+\sum_{i=1}^{9}\big[f^{(\nu)}_{\nu_i}\big]^2C^{({\mathtt{noise}},\nu_i)}_{\ell},
\label{power_spectrum_ii}
\end{equation}
\item (iii) {\it Planck}~ combined cleaned maps
\begin{equation}
C^{({\mathtt{comb}})}_{\ell} =\sum_{\{i,j\}=1}^{9}g^{(\nu_i)}_{\ell}g^{(\nu_j)}_{\ell}b^{(\nu_i)}_\ell b^{(\nu_j)}_\ell\big[\tilde{C}^{({\mathtt{CMB}})}_{\ell}+C^{({\mathtt{CIB}},\nu_i,\nu_j)}_{\ell}+C^{({\mathtt{Radio}},\nu_i,\nu_j)}_{\ell}\big]+\sum_{i=1}^{9}\big[g^{(\nu_i)}_{\ell}\big]^2C^{({\mathtt{noise}},\nu_i)}_{\ell},
\label{power_spectrum_iii}
\end{equation}
\end{itemize}
where
\begin{equation}
g^{(\nu)}_{\ell}\equiv\sum_{i=5}^{6}w_{\ell}^{(\nu_i)}f_{\nu}^{(\nu_i)},
\end{equation}
and $C^{({\mathtt{noise}},\nu)}_{\ell}$ is the {\it Planck}~ instrumental noise power
spectrum at frequency $\nu$.
\begin{table*}
\begin{center}
\caption{{\it Planck}~ instrumental characteristics based on published information \citep[see Tables 2 and 6 in][]{planck_i_2013}.\label{t1}}
\begin{tabular}{c|ccccccccc}
\hline
\hline
Channel index & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 \\
Frequency (GHz) & 30 & 44 & 70 & 100 & 143 & 217 & 353 & 545 & 857 \\
Beam FWHM (arcmin) & 33 & 28 & 13 & 10 & 7 & 5 & 5 & 5 & 5 \\
$\sigma_{noise}$ per pixel ($\mu K$) & 9.2 & 12.5 & 23.2 & 11.0 & 6.0 & 12.0 & 43.0 & 897.5 & 37178.6 \\
$N_{side}$ & 1024 & 1024 & 1024 & 2048 & 2048 & 2048 & 2048 & 2048 & 2048 \\
\hline
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table}
\begin{center}
\caption{Linear coefficients and templates used to clean
individual frequency maps with SEVEM \citep[see Table C1
in][]{planck_xii_2013}.\label{t1b}}
\begin{tabular}{c|cccc}
\hline
\hline
Template & 30-44 & 44-70 & 545-353 & 857-545 \\
$f^{(143)}_{\nu_i}$ & -2.14$\times 10^{-2}$ & -1.23$\times 10^{-1}$ & -7.52$\times 10^{-3}$ & 6.67$\times 10^{-5}$ \\
\hline
Template & 30-44 & 44-70 & 545-353 & 857-545 \\
$f^{(217)}_{\nu_i}$ & 1.03$\times 10^{-1}$ & -1.76$\times 10^{-1}$ & -1.84$\times 10^{-2}$ & 1.21$\times 10^{-4}$ \\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table*}
\center
\caption{The expected bias $\Delta f_{\rm nl}$ produced by the
CIB-lensing bispectrum for $\ell_{\rm max}~=~$ 2000 for the raw {\it Planck}~
frequency maps between 100 and 353 GHz, the cleaned maps at 143 and
217 GHz and the combination of the two previous cleaned maps. The
expected uncertainties on $f_{\rm nl~}$ are also reported.
\label{t2}}
\begin{tabular}{c|ccccccc}
\hline
\hline
Frequency (GHz) & 100 & 143 & 217 & 353 & SEVEM 143 & SEVEM 217 & SEVEM combined\\
\hline
Local $\Delta f_{\rm nl}$ & 0.34 & 0.73 & 3.06 & 13.89 & -0.48 & -0.28 & -0.39 \\
Local $\sigma(f_{\rm nl})$ & 6.65 & 5.25 & 5.55 & 14.06 & 5.73 & 5.87 & 5.70 \\
Local $\Delta f_{\rm nl}/\sigma(f_{\rm nl})$ & 0.05 & 0.14 & 0.55 & 0.99 & -0.08 & -0.05 & -0.07 \\
\hline
Equilateral $\Delta f_{\rm nl}$ & 0.37 & -4.56 & -16.47 & 150.23 & 1.17 & 0.09 & 1.99 \\
Equilateral $\sigma(f_{\rm nl})$ & 76.39 & 68.28 & 71.00 & 134.60 & 71.08 & 72.22 & 70.97 \\
Equilateral $\Delta f_{\rm nl}/\sigma(f_{\rm nl})$ & 0.00 & -0.07 & -0.23 & 1.12 & 0.02 & 0.00 & 0.03 \\
\hline
Orthogonal $\Delta (f_{\rm nl})$ & -8.45 & -21.31 & -87.92 & -233.57 & 12.55 & 5.78 & 10.54 \\
Orthogonal $\sigma(f_{\rm nl})$ & 39.29 & 33.19 & 34.73 & 76.33 & 35.17 & 35.82 & 35.04 \\
Orthogonal $\Delta f_{\rm nl}/\sigma(f_{\rm nl})$ & -0.22 & -0.64 & -2.53 & -3.06 & 0.36 & 0.16 & 0.30 \\
\hline
\hline
\end{tabular}
\end{table*}
\begin{figure*}
\includegraphics[width=58mm,height=58mm]{f4a.eps}
\includegraphics[width=58mm,height=58mm]{f4b.eps}
\includegraphics[width=58mm,height=58mm]{f4c.eps}
\caption{The bias $\Delta f_{\rm nl}$ produced by the CIB-lensing
bispectrum as a function of $\ell_{\rm max}$ for {\it Planck}~
cosmological frequencies (143 and 217 GHz). Solid lines correspond
to the cleaned combined maps. Dash lines correspond to the raw map
at 143 GHz and long dash lines correspond to the cleaned map at 143
GHz. Dash-and-dot lines correspond to the raw map at 217 GHz and
dash-and-3-dots lines correspond to the cleaned map at 217 GHz. From
left to right, we plot results for the local, equilateral and
orthogonal $f_{\rm nl~}$ shapes. }
\label{f4}
\end{figure*}
\subsection{Bispectrum}
We derive the CIB--lensing, CMB and point sources bispectra for
the three considered types of maps.
\begin{itemize}
\item (i) {\it Planck}~ raw maps per frequency
\begin{equation}
b^{({\mathtt{CIB-Lens}}, \nu)}_{\ell_1 \ell_2 \ell_3}=\frac{\ell_1(\ell_1+1)-\ell_2(\ell_2+1)+\ell_3(\ell_3+1)}{2}C^{({\mathtt{CIB-Lens}},\nu)}_{\ell_3}\tilde{C}^{({\mathtt{CMB}})}_{\ell_1}b^{(\nu)}_{\ell_1} b_{\ell_2}^{(\nu)}b_{\ell_3}^{(\nu)}+(5~perm),
\label{bispectrum_i}
\end{equation}
\begin{equation}
b^{({\mathtt{CMB}},\nu)}_{\ell_1 \ell_2 \ell_3}=b_{\ell_1}^{(\nu)} b_{\ell_2}^{(\nu)}b_{\ell_3}^{(\nu)}b^{({\mathtt{CMB}})}_{\ell_1 \ell_2 \ell_3},~~~~~ b^{({{\rm ps}},\nu)}_{\ell_1 \ell_2 \ell_3}= b_{\ell_1}^{(\nu)}b_{\ell_2}^{(\nu)}b_{\ell_3}^{(\nu)}b^{({{\rm ps}},\nu)}. \label{prim_radio_bispectrum_i}
\end{equation}
\item (ii) {\it Planck}~ cleaned maps per frequency
\begin{equation}
b^{({\mathtt{CIB-Lens}},{\mathtt{clean}},\nu)}_{\ell_1 \ell_2 \ell_3}=\sum_{ijk=1}^{9}f^{(\nu)}_{\nu_i}f^{(\nu)}_{\nu_j}f^{(\nu)}_{\nu_k}b^{(\nu_i)}_{\ell_1} b_{\ell_2}^{(\nu_j)}b_{\ell_3}^{(\nu_k)}b^{({\mathtt{CIB-Lens}},\nu_i\nu_j\nu_k)}_{\ell_1 \ell_2 \ell_3},
\label{bispectrum_ii}
\end{equation}
\begin{equation}
b^{({\mathtt{CMB}},{\mathtt{clean}},\nu)}_{\ell_1 \ell_2 \ell_3}=\sum_{ijk=1}^{9}f^{(\nu)}_{\nu_i}f^{(\nu)}_{\nu_j}f^{(\nu)}_{\nu_k}b_{\ell_1}^{(\nu_i)} b_{\ell_2}^{(\nu_j)}b_{\ell_3}^{(\nu_k)}b^{({\mathtt{CMB}})}_{\ell_1 \ell_2 \ell_3},~~~~~b^{({{\rm ps}},{\mathtt{clean}},\nu)}_{\ell_1 \ell_2 \ell_3}=\sum_{ijk=1}^{9}f^{(\nu)}_{\nu_i}f^{(\nu)}_{\nu_j}f^{(\nu)}_{\nu_k}b_{\ell_1}^{(\nu_i)} b_{\ell_2}^{(\nu_j)}b_{\ell_3}^{(\nu_k)}b^{({{\rm ps}},\nu_i,\nu_j,\nu_k)}.
\label{prim_radio_bispectrum_ii}
\end{equation}
\item (iii) {\it Planck}~ combined cleaned maps
\begin{equation}
b^{({\mathtt{CIB-Lens}},{\mathtt{comb}})}_{\ell_1 \ell_2 \ell_3}=\sum_{ijk=1}^{9}g^{(\nu_i)}_{\ell_1}g^{(\nu_j)}_{\ell_2}g^{(\nu_k)}_{\ell_3}b^{(\nu_i)}_{\ell_1} b_{\ell_2}^{(\nu_j)}b_{\ell_3}^{(\nu_k)}b^{({\mathtt{CIB-Lens}},\nu_i\nu_j\nu_k)}_{\ell_1 \ell_2 \ell_3},
\label{bispectrum_iii}
\end{equation}
\begin{equation}
b^{({\mathtt{CMB}},{\mathtt{comb}})}_{\ell_1 \ell_2 \ell_3}=\sum_{ijk=1}^{9}g^{(\nu_i)}_{\ell_1}g^{(\nu_j)}_{\ell_2}g^{(\nu_k)}_{\ell_3}b_{\ell_1}^{(\nu_i)} b_{\ell_2}^{(\nu_j)}b_{\ell_3}^{(\nu_k)}b^{({\mathtt{CMB}})}_{\ell_1 \ell_2 \ell_3},~~~~~ b^{({{\rm ps}},{\mathtt{comb}})}_{\ell_1 \ell_2 \ell_3}=\sum_{ijk=1}^{9}g^{(\nu_i)}_{\ell_1}g^{(\nu_j)}_{\ell_2}g^{(\nu_k)}_{\ell_3}b_{\ell_1}^{(\nu_i)} b_{\ell_2}^{(\nu_j)}b_{\ell_3}^{(\nu_k)}b^{({{\rm ps}},\nu_i,\nu_j,\nu_k)}.
\label{prim_radio_bispectrum_iii}
\end{equation}
\end{itemize}
$b^{({\mathtt{CMB}})}_{\ell_1 \ell_2 \ell_3}$ is the primordial bispectrum
\citep[see e.g.][for the equations for the local, equilateral and
orthogonal shapes]{curto2013}. The term
$b^{({\mathtt{CIB-Lens}},\nu_i\nu_j\nu_k)}_{\ell_1 \ell_2 \ell_3}$ can be
straightforwardly computed as a generalisation of the CIB-lensing
bispectrum in Eq. (\ref{s2e3}) for three different frequencies
\begin{eqnarray}
\nonumber
& b^{({\mathtt{CIB-Lens}},\nu_i\nu_j\nu_k)}_{\ell_1 \ell_2 \ell_3} \equiv \frac{\ell_1(\ell_1+1)-\ell_2(\ell_2+1)+\ell_3(\ell_3+1)}{2}C^{({\mathtt{CIB-Lens}},\nu_k)}_{\ell_3}\tilde{C}^{({\mathtt{CMB}})}_{\ell_1} + \frac{\ell_1(\ell_1+1)-\ell_3(\ell_3+1)+\ell_2(\ell_2+1)}{2}C^{({\mathtt{CIB-Lens}},\nu_j)}_{\ell_2}\tilde{C}^{({\mathtt{CMB}})}_{\ell_1}&\\
\nonumber
& +\frac{\ell_2(\ell_2+1)-\ell_1(\ell_1+1)+\ell_3(\ell_3+1)}{2}C^{({\mathtt{CIB-Lens}},\nu_k)}_{\ell_3}\tilde{C}^{({\mathtt{CMB}})}_{\ell_2}+\frac{\ell_3(\ell_3+1)-\ell_1(\ell_1+1)+\ell_2(\ell_2+1)}{2}C^{({\mathtt{CIB-Lens}},\nu_j)}_{\ell_2}\tilde{C}^{({\mathtt{CMB}})}_{\ell_3}&\\
& +\frac{\ell_2(\ell_2+1)-\ell_3(\ell_3+1)+\ell_1(\ell_1+1)}{2}C^{({\mathtt{CIB-Lens}},\nu_i)}_{\ell_1}\tilde{C}^{({\mathtt{CMB}})}_{\ell_2}+\frac{\ell_3(\ell_3+1)-\ell_2(\ell_2+1)+\ell_1(\ell_1+1)}{2}C^{({\mathtt{CIB-Lens}},\nu_i)}_{\ell_1}\tilde{C}^{({\mathtt{CMB}})}_{\ell_3}.&
\end{eqnarray}
Finally, the total point sources bispectrum
$b^{({{\rm ps}},\nu_i,\nu_j,\nu_k)}_{\ell_1 \ell_2 \ell_3}$ is computed from
the standard prescription in terms of radio and CIB shot noise
bispectra (see Appendix\,\ref{appendix_sn} for details on how we
estimate shot noise bispectra of point sources when multiple
frequencies are considered)
\begin{align}
b^{({{\rm ps}},\nu_i,\nu_j,\nu_k)}_{\ell_1 \ell_2 \ell_3} =
b^{({\mathtt{Radio}},\nu_i,\nu_j,\nu_k)}_{{\rm sn}}
+ b^{({\mathtt{CIB}},\nu_i,\nu_j,\nu_k)}_{{\rm sn}} \sqrt{\frac{C^{({\mathtt{CIB}},\nu_i)}_{\ell_1}C^{({\mathtt{CIB}},\nu_j)}_{\ell_2}C^{({\mathtt{CIB}},\nu_k)}_{\ell_3}}{C^{({\mathtt{CIB}},\nu_i)}_{{\rm sn}}C^{({\mathtt{CIB}},\nu_j)}_{{\rm sn}}C^{({\mathtt{CIB}},\nu_k)}_{{\rm sn}}}}.
\label{radio_bispectrum_raw}
\end{align}
Considering weak levels of non-Gaussianity, the bias induced in the
primordial non-Gaussianity $f_{\rm nl~}$ due to a given target bispectrum
$B^{({\mathtt{targ}})}_{\ell_1 \ell_2 \ell_3}$ is given by \citep[see
e.g.][]{lewis2011,lacasa2012a}:
\begin{equation}
\Delta f_{\rm nl} = \sigma^2\big(f_{\rm nl}\big) \times \sum_{\ell_1 \le \ell_2 \le \ell_3}^{\ell_{\rm max}}
\frac{b^{({\mathtt{targ}})}_{\ell_1 \ell_2 \ell_3} b^{({\mathtt{prim}})}_{\ell_1 \ell_2 \ell_3} }{\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_1 \ell_2 \ell_3}\big)},
\label{bias_fnl_lens_cib}
\end{equation}
where $\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_1 \ell_2 \ell_3}\big)$ is the
variance of the total observed bispectrum \citep{komatsu2001}
\begin{align}
\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_1 \ell_2 \ell_3}\big) & \equiv \langle b^{({\mathtt{obs}})}_{\ell_1 \ell_2 \ell_3}b^{({\mathtt{obs}})}_{\ell_1 \ell_2 \ell_3} \rangle - \langle b^{({\mathtt{obs}})}_{\ell_1 \ell_2 \ell_3}\rangle \langle b^{({\mathtt{obs}})}_{\ell_1 \ell_2 \ell_3} \rangle \simeq \frac{1}{I_{\ell_{1} \ell_{2} \ell_{3}}^2}\Delta_{\ell_1 \ell_2 \ell_3}C_{\ell_1}C_{\ell_2}C_{\ell_3},\\
& \Delta_{\ell_1 \ell_2 \ell_3} = 1 + 2\delta_{\ell_1
\ell_2}\delta_{\ell_2 \ell_3} + \delta_{\ell_1 \ell_2} +
\delta_{\ell_2 \ell_3} + \delta_{\ell_1 \ell_3},
\label{variance_bispectrum_i}
\end{align}
$C_{\ell}$ is the total power spectrum of the map including CMB,
CIB, radio sources and instrumental noise spectra and
\begin{eqnarray}
I_{\ell_{1} \ell_{2} \ell_{3}}\equiv \sqrt{\frac{\left(2\ell_1+1\right)\left(2\ell_2+1\right)\left(2\ell_3+1\right)}{4 \pi}}\left(\begin{array}{ccc}
\ell_1 & \ell_2 & \ell_3 \\ 0 & 0 & 0
\end{array}\right).
\label{angle_averaged_bisp2}
\end{eqnarray}
Finally $\sigma^{2}\big(f_{\rm nl}\big)$ is the expected variance of
the $f_{\rm nl~}$ parameter, given in terms of its Fisher matrix
\begin{equation}
\sigma^{-2}\big(f_{\rm nl}\big) = \sum_{\ell_1 \le
\ell_2 \le \ell_3}^{\ell_{\rm max}} \frac{\big(b^{({\mathtt{prim}})}_{\ell_1 \ell_2
\ell_3}I_{\ell_{1} \ell_{2} \ell_{3}}\big)^2} {\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_1 \ell_2 \ell_3}\big)}.
\label{eq:fnl_error_point_sources}
\end{equation}
The values of $\sigma^{2}\big(f_{\rm nl}\big)$ for the different maps
are given in Table\,\ref{t2}. For the SEVEM combined map they agree
well with the values published by
\citet{planck_xxiv_2013}\footnote{During the process of
publication of this article the Planck Collaboration published new
scientific results, in particular the first results on $f_{nl}$ with
temperature and polarisation maps \citep{planck_xix_2015}. No
significant changes have been reported in the new Planck article
regarding the results with temperature only.}. Note that at 143 and
217 GHz, the component separation technique slightly increases the
error on $f_{\rm nl}$. This is due to the extra noise added by the
template subtraction and it might be seen as the price to pay to have
CMB cleaned maps.
\subsection{$f_{\rm nl~}$ bias due to CIB--lensing and extragalactic
sources bispectra}
The bias $\Delta f_{\rm nl}$ induced by the CIB-lensing correlation
and unresolved extragalactic sources is estimated for the three types
of maps previously described and for the local, equilateral and
orthogonal $f_{\rm nl~}$ shapes in the frequency range between 100 and
353\,GHz. Results are presented in Table\,\ref{t2} and Figure \ref{f4}
for the CIB-lensing, and Table\,\ref{t3} and Figure \ref{f4b} for the
unresolved extragalactic sources.
For the frequency range considered here, the CIB-lensing bispectrum
causes negligible bias in the primordial local and equilateral shapes
with respect to the uncertainty on $f_{\rm nl}$ estimated by Eq.\
(\ref{eq:fnl_error_point_sources}). The bias is also negligible for
the orthogonal shape in the 100 and 143 GHz raw maps but reaches 2 and
3 $\sigma$ detection levels for the 217 and 353 GHz raw maps
respectively. The orthogonal bias is again negligible for the
foreground-reduced maps at 143 and 217 GHz and the combined map (see
Table \ref{t2} and Fig. \ref{f4}). These results are explained as
follows. Regarding the local shape, the CMB primordial signal peaks in
squeezed configurations, such as $ (\ell_1, \ell_2, \ell_3)=(1000,
1000, 2)$. However in this regime, the CIB-lensing bispectrum loses
most of its amplitude (see top-left panel of Fig. \ref{f2}). Regarding
the equilateral shape, the CMB primordial signal is spread in
configurations such that $ \ell_1 = \ell_2 = \ell_3=\ell$ and becomes
strongest at high resolution. The CIB-lensing bispectrum has some
peaks in equilateral configurations (see bottom-left panel of
Fig. \ref{f2}) but they are located at low $ \ell$ and therefore they
do not significantly couple with the CMB primordial equilateral
signal. Finally regarding the orthogonal shape, the CMB primordial
signal is peaked in configurations such as $ \ell_2 = \ell_3 = 2
\ell_1 $ and $ \ell_2 = \ell_3 = \ell_1$ \citep{martinez2012}. They
couple with the CIB-lensing signal producing an increasing bias for
multipoles $\ell > 500$ (see Fig. \ref{f4}). This explains the bias
predicted for the raw channels at high frequency. The process of
cleaning through the component separation subtracts part of the CIB
signal and the bias for this shape is reduced in cleaned maps to about
30 per cent of the {\it Planck} uncertainty.
The unresolved extragalactic sources present levels of detection
greater than $2\sigma$ at 353 GHz for the local, equilateral and
orthogonal shapes due to the higher amplitude of IR sources at this
frequency. There is a $2\sigma$ detection at 100 GHz for the
equilateral shape which can be explained as a trace of the radio
sources. The bias is again negligible for the foreground-reduced maps
as the component separation technique is able to significantly reduce
their contamination (see Table \ref{t3} and Fig. \ref{f4b}). These
results are well in agreement with previous analyses from
\citet{lacasa2012a} and \citet{curto2013}. These results are
explained as follows. Regarding the local shape, the bispectrum of
extragalactic sources does not have a significant signal in squeezed
configurations at low frequencies (see top-left panel of
Fig. \ref{f2}) and therefore there we do not expected significant
correlations with the CMB primordial local bispectrum. However the
high frequency channels contain a significant contribution from IR
sources in squeezed configurations that couple with the local CMB
bispectrum. Regarding the equilateral shape, the extragalactic radio
sources have significant signal in equilateral configurations ($
\ell_1 = \ell_2 = \ell_3 $) at high multipoles (see bottom-left panel
of Fig. \ref{f2}) explaining the deviation seen in Table
{\ref{t3}}. The bispectrum of IR sources is dominant in equilateral
configurations and at high multipoles explaining the large bias
predicted at 353 GHz. Finally regarding the orthogonal shape, only the
high frequency IR source bispectrum has a strong signal in
configurations that couple with the CMB primordial orthogonal
bispectrum, especially for high multipoles, explaining the large bias
predicted at 353 GHz.
\begin{figure*}
\includegraphics[width=58mm,height=58mm]{f5a.eps}
\includegraphics[width=58mm,height=58mm]{f5b.eps}
\includegraphics[width=58mm,height=58mm]{f5c.eps}
\caption
The bias $\Delta f_{\rm nl}$ produced by the point sources
bispectrum as a function of $\ell_{\rm max}$ for {\it Planck}~
cosmological frequencies (143 and 217 GHz). Solid lines correspond
to the cleaned combined maps. Dash lines correspond to the raw map
at 143 GHz and long dash lines correspond to the cleaned map at 143
GHz. Dash-and-dot lines correspond to the raw map at 217 GHz and
dash-and-3-dots lines correspond to the cleaned map at 217 GHz.
From left to right, we plot results for the local, equilateral and
orthogonal $f_{\rm nl~}$ shapes. }
\label{f4b}
\end{figure*}
\begin{table*}
\center
\caption{{\it Planck}~ expected $\Delta f_{\rm nl}$ bias due to unresolved
point sources for the local, equilateral and orthogonal $f_{\rm nl~}$
shapes for
$\ell_{\rm max} =$2000.}
\label{t3}
\begin{tabular}{c|ccccccc}
\hline
\hline
Frequency (GHz) & 100 & 143 & 217 & 353 & SEVEM 143 & SEVEM 217 & SEVEM combined\\
\hline
Local $\Delta f_{\rm nl}$ & 2.95 & 0.92 & 0.98 & 29.30 & 0.50 & 0.54 & 0.35 \\
Local $\Delta f_{\rm nl}/\sigma(f_{\rm nl})$ & 0.44 & 0.18 & 0.18 & 2.08 & 0.09 & 0.09 & 0.07 \\
\hline
Equilateral $\Delta f_{\rm nl}$ & 160.07 & 54.20 & 60.32 & 1648.70 & 34.19 & 32.36 & 30.38 \\
Equilateral $\Delta f_{\rm nl}/\sigma(f_{\rm nl})$ & 2.10 & 0.79 & 0.85 & 12.25 & 0.48 & 0.45 & 0.45 \\
\hline
Orthogonal $\Delta f_{\rm nl}$ & 10.53 & 2.45 & 7.00 & 553.01 & 1.87 & 3.56 & 2.94 \\
Orthogonal $\Delta f_{\rm nl}/\sigma(f_{\rm nl})$ & 0.27 & 0.07 & 0.20 & 7.24 & 0.05 & 0.10 & 0.09 \\
%
\hline
\hline
\end{tabular}
\end{table*}
\section{Detectability of the CIB--Lensing bispectrum}
\label{section:detectability}
In this section we develop statistical tools to detect the
CIB--lensing bispectrum using an alternative approach to the widely
known technique based on the cross-correlation of lensing potential
reconstruction and temperature maps used for example in
\citet{planck_xviii_2013} and \citet{planck_xix_2013}. The
cross-correlation approach followed in these publications used a
lensing reconstruction based on quadratic combination of {\it Planck}~ CMB
maps \citep{planck_xvii_2013}. The estimators that we propose in this
article are directly defined in terms of cubic combinations of {\it Planck}~
maps where the CMB signal is dominant (100 to 217 GHz) and {\it Planck}~
maps where the CIB signal is dominant (217 to 857 GHz). Both
approaches are linearly dependent and should result in similar levels
of efficiency to detect the targeted CIB--lensing signal. The
advantage of the new approach defined here is the application for the
first time of a battery of well-known, efficient and optimal
estimators widely used in the primordial bispectrum estimation to
detect the CIB--lensing bispectrum. This approach has already been
applied to the ISW--lensing estimation by \citet{Mangilli:2013sxa}.
\subsection{Single frequency bispectrum estimator}
The optimal estimator for the amplitude of the CIB--lensing
bispectrum, assuming small departures of non-Gaussianity, for the
ideal, full--sky and isotropic instrumental noise is
\begin{equation}
\hat{A}^{({\mathtt{CIB-Lens}})}=\big(F^{-1}\big)\hat{S}^{({\mathtt{CIB-Lens}})}
\label{cib_lens_estimator}
\end{equation}
where
\begin{equation}
\hat{S}^{({\mathtt{CIB-Lens}})}=\sum_{2\le \ell_{1} \le \ell_{2} \le \ell_{3}}^{\ell_{\rm max}}\frac{b^{({\mathtt{obs}})}_{\ell_{1} \ell_{2} \ell_{3}}b^{({\mathtt{CIB-Lens}})}_{\ell_{1} \ell_{2} \ell_{3}}}{\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_{1} \ell_{2} \ell_{3}}\big)},
\label{s_estimator_cib_lensing}
\end{equation}
\begin{equation}
F=\sum_{2\le \ell_{1} \le \ell_{2} \le \ell_{3}}^{\ell_{\rm max}}\frac{b^{({\mathtt{CIB-Lens}})}_{\ell_{1} \ell_{2} \ell_{3}}b^{({\mathtt{CIB-Lens}})}_{\ell_{1} \ell_{2} \ell_{3}}}{\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_{1} \ell_{2} \ell_{3}}\big)},
\label{fisher_cib_l}
\end{equation}
and the observed bispectrum $b^{({\mathtt{obs}})}_{\ell_{1} \ell_{2} \ell_{3}}$
is based on cubic combinations of {\it Planck}~ data in the $a_{\ell m}$
decomposition on the sphere. Another interesting quantity is the
expected bispectrum signal-to-noise ratio as a function of the largest
scale mode $\ell_{\rm min}$ \citep{lewis2011}:
\begin{equation}
F_{\ell_{\rm min}}=\sum_{\ell_{\rm min}\le \ell_{1} \le \ell_{2} \le \ell_{3}}^{\ell_{\rm max}}\frac{b^{({\mathtt{CIB-Lens}})}_{\ell_{1} \ell_{2} \ell_{3}}b^{({\mathtt{CIB-Lens}})}_{\ell_{1} \ell_{2} \ell_{3}}}{\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_{1} \ell_{2} \ell_{3}}\big)},
\label{fisher_cib_lmin}
\end{equation}
This quantity helps to find the multipole configurations where the
bispectrum signal peaks.
The signal-to-noise ratio of the CIB--lensing signal of this estimator
$\sqrt{F}$ is summarised in Table \ref{t4} for the frequency range
between 100 and 857 GHz and $\ell_{\rm max}=$2000 and different sky
fractions available: 100\% (full sky), 30.4\%\footnote{This is the
percentage of available sky used in the main results of
\citet{planck_xviii_2013}.} and 10\%. We compute this ratio
considering the CIB--lensing signal alone (Eq. \ref{fisher_cib_l}) and
the joint Fisher matrix for the four non-primordial bispectra used in
this article, i.e, CIB--lensing, ISW--lensing, CIB, and extragalactic
point sources, by using the generalised Fisher matrix between the
bispectra $i$ and $j$:
\begin{equation}
F_{ij}=\sum_{2\le \ell_{1} \le \ell_{2} \le \ell_{3}}^{\ell_{\rm max}}\frac{b^{(i)}_{\ell_{1} \ell_{2} \ell_{3}}b^{(j)}_{\ell_{1} \ell_{2} \ell_{3}}}{\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_{1} \ell_{2} \ell_{3}}\big)}.
\label{fisher_cib_l_ij}
\end{equation}
The first case is simply $F_{indep}=F_{ii}$ whereas the second is
$F_{joint}=1/\left(F^{-1}\right)_{ii}$ with $i$ being the
CIB--lensing case. The significance level for the detection of the
CIB--lensing signal with this estimator increases from
approx. 0.5$\sigma$ at 100 GHz to 9.8$\sigma$ at 353 GHz considering
the full sky available. In a more realistic scenario with a 10\% sky
available for the CIB maps, the CIB--lensing signal would be detected
with a maximum precision of 1$\sigma$ at 353 GHz. At higher
frequencies, the CIB spectrum dominates in the denominator in
Eq. (\ref{fisher_cib_l}) leading to low significance levels of
detection for the CIB--lensing bispectrum.
\begin{table*}
\begin{center}
\caption
Signal-to-noise ratio $\sqrt{F}$ of the amplitude of the
CIB--lensing bispectrum for the {\it Planck}~ raw maps at frequencies
between 100 and 857 GHz using ideal conditions (isotropic
instrumental noise) from the Fisher matrix in
Eq. (\ref{fisher_cib_l}) and several sky fractions available.}
\label{t4}
\begin{tabular}{c|cccccc}
\hline
\hline
Frequency (GHz) & 100 & 143 & 217 & 353 & 545 & 857 \\
\hline
$F_{indep}^{1/2} \left(f_{sky}=1\right)$ & 0.55 & 1.80 & 7.57 & 10.19 & 0.25 & 0.00 \\
$F_{indep}^{1/2} \left(f_{sky}=0.304\right)$ & 0.17 & 0.55 & 2.30 & 3.10 & 0.08 & 0.00 \\
$F_{indep}^{1/2} \left(f_{sky}=0.100\right)$ & 0.06 & 0.18 & 0.76 & 1.02 & 0.03 & 0.00 \\
\hline
$F_{joint}^{1/2} \left(f_{sky}=1\right)$ & 0.52 & 1.66 & 6.87 & 9.80 & 0.25 & 0.00 \\
$F_{joint}^{1/2} \left(f_{sky}=0.304\right)$ & 0.16 & 0.51 & 2.09 & 2.98 & 0.07 & 0.00 \\
$F_{joint}^{1/2} \left(f_{sky}=0.100\right)$ & 0.05 & 0.17 & 0.69 & 0.98 & 0.02 & 0.00 \\
\hline
\hline
\end{tabular}
\end{center}
\end{table*}
\subsection{Asymmetric estimator for CMB--CIB correlated maps}
\label{s4ss3}
At high {\it Planck}~ frequencies the CIB bispectrum could strongly limit
our capability to detect the CIB--lensing signal, as well as the
ISW--lensing contribution could be a relevant ``noise'' at
100--217\,GHz. Therefore, a more feasible procedure to detect the
CIB--lensing bispectrum signal should correlate CMB
signal-dominated maps and CIB signal-dominated maps. We define an
estimator for the CIB-lensing signal by considering the asymmetric
configuration $\tilde{a}^{({\mathtt{raw}}, \nu_{\mathtt{CMB}})}_{\ell_1
m_1}\tilde{a}^{({\mathtt{raw}}, \nu_{\mathtt{CMB}})}_{\ell_2 m_2}a^{({\mathtt{raw}},
\nu_{\mathtt{CIB}})}_{\ell_3 m_3}$
where $\nu_{\mathtt{CMB}}$ and $\nu_{\mathtt{CIB}}$ are frequency channels where the CMB
and CIB signal are significant respectively. We consider $\nu_{\mathtt{CMB}} =
$ 100, 143 and 217 GHz and $\nu_{\mathtt{CIB}} = $ 217, 353, 545 and 857
GHz. The four non-primordial averaged bispectra considered in this
paper, namely the radio, CIB, CIB-lensing and ISW-lensing bispectra,
are written in this asymmetric configuration by:
\begin{equation}
b^{({\mathtt{Radio}})}_{\ell_{1} \ell_{2} \ell_{3}}= b^{({\mathtt{Radio}},\nu_{\mathtt{CMB}},\nu_{\mathtt{CMB}},\nu_{\mathtt{CIB}})}_{{\rm sn}}
\label{asym_radio}
\end{equation}
\begin{equation}
b^{({\mathtt{CIB}})}_{\ell_{1} \ell_{2} \ell_{3}}= b^{({\mathtt{CIB}},\nu_{\mathtt{CMB}},\nu_{\mathtt{CMB}},\nu_{\mathtt{CIB}})}_{{\rm sn}} \sqrt{\frac{C^{({\mathtt{CIB}},\nu_{\mathtt{CMB}})}_{\ell_1}C^{({\mathtt{CIB}},\nu_{\mathtt{CMB}})}_{\ell_2}C^{({\mathtt{CIB}},\nu_{\mathtt{CIB}})}_{\ell_3}}{C^{({\mathtt{CIB}},\nu_{\mathtt{CMB}})}_{{\rm sn}}C^{({\mathtt{CIB}},\nu_{\mathtt{CMB}})}_{{\rm sn}}C^{({\mathtt{CIB}},\nu_{\mathtt{CIB}})}_{{\rm sn}}}}
\label{asym_cib}
\end{equation}
\begin{align}
\nonumber
& b^{({\mathtt{CIB-Lens}})}_{\ell_{1} \ell_{2} \ell_{3}} = \\
& \bigg[{\ell_1(\ell_1+1)-\ell_2(\ell_2+1)
+\ell_3(\ell_3+1) \over 2}\,\tilde{C}^{({\mathtt{CMB}})}_{\ell_1}\,C^{({\mathtt{CIB-Lens}},\nu_{\mathtt{CIB}})}_{\ell_3}
+{\ell_2(\ell_2+1)-\ell_1(\ell_1+1)
+\ell_3(\ell_3+1) \over 2}\,\tilde{C}^{({\mathtt{CMB}})}_{\ell_2}\,C^{({\mathtt{CIB-Lens}},\nu_{\mathtt{CIB}})}_{\ell_3}\bigg]\,,
\label{asym_cib_lens}
\end{align}
and
\begin{align}
b^{({\mathtt{ISW-Lens}})}_{\ell_{1} \ell_{2} \ell_{3}} =
\bigg[{\ell_1(\ell_1+1)-\ell_2(\ell_2+1)
+\ell_3(\ell_3+1) \over 2}\,\tilde{C}^{({\mathtt{CMB}})}_{\ell_1}\,C^{({\mathtt{ISW-Lens}})}_{\ell_3}
+(5~perm)\bigg] .
\label{asym_isw_lens}
\end{align}
The covariance matrix is nearly diagonal and can be approximated by
the following expression (see Appendix\,\ref{appendix_cov_bisp}) for
the CMB x CMB x CIB configurations:
\begin{eqnarray}
\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_{1} \ell_{2} \ell_{3}}\big)= \frac{1}{I_{\ell_{1} \ell_{2} \ell_{3}}^2}\tilde{C}^{({\mathtt{CMB}})}_{\ell_1}\tilde{C}^{({\mathtt{CMB}})}_{\ell_2}C^{({\mathtt{CIB}})}_{\ell_3}\left(1+\delta_{\ell_1 \ell_2}\right).
\label{cmbcmbcib_variance_asymm}
\end{eqnarray}
The CIB-lensing estimator for this configuration is now:
\begin{equation}
\hat{S}^{({\mathtt{CIB-Lens}})}=\sum_{2\le \ell_{1} \le \ell_{2}, \ell_{3}}^{\ell_{\rm max}}\frac{b^{({\mathtt{obs}})}_{\ell_{1} \ell_{2} \ell_{3}}b^{({\mathtt{CIB-Lens}})}_{\ell_{1} \ell_{2} \ell_{3}}}{\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_{1} \ell_{2} \ell_{3}}\big)}.
\label{asymmetric_cmbcmbcib_s}
\end{equation}
The estimator is different with respect to the single map estimators
since it admits more configurations because we just have symmetry
under permutations of $\ell_1$ and $\ell_2$ whereas $\ell_3$ is
free. The Fisher matrix for the four bispectra considered in this
work is defined as \citep{komatsu2001}:
\begin{equation}
F_{ij}=\sum_{2\le \ell_{1} \le \ell_{2}, \ell_{3}}^{\ell_{\rm max}}\frac{b^{(i)}_{\ell_{1} \ell_{2} \ell_{3}}b^{(j)}_{\ell_{1} \ell_{2} \ell_{3}}}{\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_{1} \ell_{2} \ell_{3}}\big)}
\label{fisher_cib_l_assym}
\end{equation}
where the indices $i$ and $j$ cover the following bispectra: (1) radio, (2)
CIB, (3) CIB-lensing and (4) ISW-lensing. We also define the Fisher
matrix in terms of the minimum multipole $\ell_{\rm min}$:
\begin{equation}
\left(F_{ij}\right)_{\ell_{\rm min}}=\sum_{\ell_{\rm min}\le \ell_{1} \le \ell_{2}, \ell_{3}}^{\ell_{\rm max}}\frac{b^{(i)}_{\ell_{1} \ell_{2} \ell_{3}}b^{(j)}_{\ell_{1} \ell_{2} \ell_{3}}}{\sigma^2\big(b^{({\mathtt{obs}})}_{\ell_{1} \ell_{2} \ell_{3}}\big)} \, .
\label{fisher_cib_l_assym2}
\end{equation}
The Cram\'er-Rao inequality states that the inverse of the Fisher
information matrix is a lower bound on the variance of any unbiased
estimator in the best optimal conditions. Therefore the variance of
the amplitude of each bispectra can be obtained by inverting the
Fisher matrix
\begin{equation}
\sigma_{\rm joint}^2(A_i) = \left( F^{-1} \right)_{ii}.
\label{sigma2_assym}
\end{equation}
This approach performs a joint analysis including the correlations
among the four types of bispectra, in comparison to the independent
constraint that would have a lower variance:
\begin{equation}
\sigma_{\rm indep}^2(A_i) = \left( F_{ii} \right)^{-1}.
\label{sigma2_assym_indep}
\end{equation}
{The signal-to-noise ratio is given by:}
\begin{equation}
F_{\rm indep}^{1/2}(A_i) = F^{1/2}_{ii} ~~~~~~~~~~~~~~~~~~ F_{\rm joint}^{1/2}(A_i) = 1/\sqrt{\left(F^{-1}\right)_{ii}}.
\label{sigma2_assym_indep}
\end{equation}
The expected uncertainties for the CIB-lensing estimator, computed
both using the independent and the joint approach, are plotted in Fig.
\ref{f5} and is summarised in Table \ref{t5b} for $\ell_{\rm
max}=$2000. In an ideal scenario where the CMB and the CIB maps are
completely separated in two full sky maps, we would have a
detectability level of approximately $63\sigma$ in the best
configurations given in Table \ref{t5b} for $\ell_{\rm
max}=$2000. However available CIB maps cover only about 10\% of the
sky \citep{planck_xxx_2013} in the best case scenarios. For the
incomplete sky case, the detectability level of the bispectrum is
rescaled by the fraction of the available sky $f_{{\rm{sky}}}$ such that
$\sigma(A) \longrightarrow \sigma(A)/ \sqrt{f_{{\rm{sky}}}}$. As the
CIB--lensing bispectrum is not squeezed (see Fig. \ref{f5}), this
case is not affected by the loss of low multipoles, which are
unobservable for small sky fractions, so this approximation is safe up
to the mentioned 10\% of the sky. This would provide detectability
levels between 12$\sigma$ to 20$\sigma$ respectively for a mask with
10\% of the sky available using the joint estimator (see Table
\ref{t5b}).
\begin{figure*}
\includegraphics[width=58mm,height=58mm]{f6a.eps}
\includegraphics[width=58mm,height=58mm]{f6b.eps}
\includegraphics[width=58mm,height=58mm]{f6c.eps}
\includegraphics[width=58mm,height=58mm]{f6d.eps}
\includegraphics[width=58mm,height=58mm]{f6e.eps}
\includegraphics[width=58mm,height=58mm]{f6f.eps}
\caption{{\it Top:} the uncertainty for the CIB--lensing bispectrum
amplitude $\sigma(A)$ as a function of $\ell_{\rm max}$ for cleaned
CMB maps (at 100, 143 and 217\,GHz) combined with CIB dominated maps
at 353, 545 and 857\,GHz using the independent estimates (grey
lines) and using the joint estimates (black lines). {\it Bottom: }
the Fischer matrix $F_{\ell_{\rm min}}$ (multiplied by $\ell_{\rm
min}$) as a function of $\ell_{\rm min}$ for the same cases as
above. }
\label{f5}
\end{figure*}
\begin{table*}
\begin{center}
\caption{The signal-to-noise ratio $\sqrt{F}$ of the amplitude of
the CIB--lensing bispectrum for clean full--sky maps using
the joint (Eq. \ref{sigma2_assym}) and independent
(Eq. \ref{sigma2_assym_indep}) Fisher matrices for the
combinations the clean CMB at frequencies between 100 and 217
GHz and the clean CIB at frequencies between 353 and 857 GHz
using 100\%, (30.4\%), [10\%] of the sky..
\label{t5b}}
\begin{tabular}{c|c|cccc}
\hline
\hline
Case & Frequency & $\nu_{{\mathtt{CIB}}}$ = 353 GHz & $\nu_{{\mathtt{CIB}}}$ = 545 GHz & $\nu_{{\mathtt{CIB}}}$ = 857 GHz \\
\hline
Joint & $\nu_{{\mathtt{CMB}}}$ = 100 GHz & 42.70 ( 23.54) [ 13.50] & 41.03 ( 22.62) [ 12.98] & 37.60 ( 20.73) [ 11.89] \\
Joint & $\nu_{{\mathtt{CMB}}}$ = 143 GHz & 62.66 ( 34.55) [ 19.82] & 60.15 ( 33.17) [ 19.02] & 55.40 ( 30.54) [ 17.52] \\
Joint & $\nu_{{\mathtt{CMB}}}$ = 217 GHz & 62.37 ( 34.39) [ 19.72] & 59.90 ( 33.03) [ 18.94] & 55.19 ( 30.43) [ 17.45] \\
\hline
Independent & $\nu_{{\mathtt{CMB}}}$ = 100 GHz & 44.65 ( 24.62) [ 14.12] & 43.21 ( 23.83) [ 13.66] & 39.71 ( 21.90) [ 12.56] \\
Independent & $\nu_{{\mathtt{CMB}}}$ = 143 GHz &68.70 ( 37.88) [ 21.73] & 66.61 ( 36.73) [ 21.06] & 61.42 ( 33.86) [ 19.42] \\
Independent & $\nu_{{\mathtt{CMB}}}$ = 217 GHz &68.51 ( 37.77) [ 21.66] & 66.42 ( 36.62) [ 21.00] & 61.26 ( 33.78) [ 19.37] \\
\hline
\hline
\end{tabular}
\end{center}
\end{table*}
We have compared our results with the estimates of the CIB--lensing
correlation published in \citet{planck_xviii_2013}. The statistical
estimator used in that work is based on a cross-correlation between
the lensing potential in harmonic space, $\phi_{\ell m}$, and a
temperature map. The amplitude of the detection is obtained with the
quadrature sum of the significance of the different multipole
bins. This amplitude takes into account effects such as correlations
among different bins but no systematic errors or point source
corrections. The final estimates presented in that paper are computed
using the lensing reconstruction at 143 GHz and the {\it Planck}~ HFI
foreground reduced maps with a mask of 30.4\% of available sky. The
results obtained in this way and our predictions for the same
configuration are given in Table \ref{t7}. \citet{planck_xviii_2013}
provides two types of detection significances: one that only includes
the statistical errors only and another one that includes statistical
and systematic errors. Compared to the first one -- as we do not
consider systematic errors here -- the detection significance by
\citet{planck_xviii_2013} and our model are nearly equivalent for the
353 to 857 GHz range. Note that the values measured by
\citet{planck_xviii_2013} are quite sensitive to the systematic
effects (see Table \ref{t7}). We have additionally repeated our
analysis without adding instrumental noise, i.e.\ for ideal
conditions, and have found $\sqrt{F}=$ 39, 38, 35 for the 353, 545 and
857 GHz bands. Both cases show a decreasing trend in the
signal-to-noise level of the CIB-lensing correlation as we increase
the frequency, due to the higher contamination of the CIB. We do not
see a peak at 545 GHz, and we think that the peak observed in
\citet{planck_xviii_2013} at 545 GHz might be explained by an unknown
systematic artefact present in the data.
\begin{table*}
\begin{center}
\caption{{ Significance of the amplitude of the CIB--lensing
correlation using the lensing reconstruction at 143 GHz
and the {\it Planck}~ HFI foreground reduced maps with a mask of
30.4\% of available sky. {\it Top line:} measurements by
\citet{planck_xviii_2013}. {\it Bottom line:} our predictions
for an optimal bispectrum estimator for the same
configuration.
\label{t7}}}
\begin{tabular}{c|ccc}
\hline
\hline
Case & $\nu_{{\mathtt{CIB}}}$ = 353 GHz & $\nu_{{\mathtt{CIB}}}$ = 545 GHz & $\nu_{{\mathtt{CIB}}}$ = 857 GHz \\
\hline
Number of standard deviations$^a$ \citep{planck_xviii_2013} & 31 (24) & 42 (19) & 32 (16) \\
\hline
$\sqrt{F}$ & 35 & 33 & 31 \\
\hline
\hline
\end{tabular}
\begin{tablenotes}
\small
\item $^a$ Statistical error and statistical plus systematic errors in parenthesis.
\end{tablenotes}
\end{center}
\end{table*}
The Wick expansions, used to compute the variance of the observed
bispectra (see Appendix\,\ref{appendix_cov_bisp}), are good
approximations when the departures from non-Gaussianity are
limited. Therefore, we could expect some contributions to the
covariance matrix of the bispectrum in
Eq.\,(\ref{cmbcmbcib_variance_asymm}) due to higher order moments
\citep[see Section 4.3 in][]{planck_xxx_2013}. We have computed higher
order contributions to the variance in Appendix
\ref{appendix_cov_bisp}. The result is that higher order moments do
not add a significant contribution to the covariance and that the
Wick expansions used in this paper hold.
\section{KSW-based estimators for the CIB-lensing bispectrum}
\label{section5}
We present the formalism of an optimal estimator for the amplitude of
the bispectrum induced by the CIB--lensing correlation, based on the
estimator developed by \citet[][KSW]{Komatsu:2003iq} for the
primordial non-Gaussianity and extended to the ISW--lensing bispectrum
by \citet{Mangilli:2013sxa}. Here we consider the ideal case without
noise and beam function. The case of a realistic experiment can be
straightforward extended \citep[see,
e.g.,][]{Lacasa:2012zx,Mangilli:2013sxa}.
\subsection{Single frequency bispectrum estimator}
The optimal estimator $\hat{S}$ of the CIB--lensing bispectrum for a
single frequency map
is given by
Eq. (\ref{s_estimator_cib_lensing}). Using the identity
\begin{equation}
\sum_{\ell_1 \le \ell_2 \le \ell_3}^{\ell_{\rm max}}F_{\ell_1 \ell_2 \ell_3} = \frac{1}{6}\sum_{\ell_1 \ell_2 \ell_3}^{\ell_{\rm max}}F_{\ell_1 \ell_2 \ell_3}\Delta_{\ell_1 \ell_2 \ell_3}
\end{equation}
for any given $F_{\ell_1\ell_2\ell_3}$ symmetric in $\ell_1$, $\ell_2$,
$\ell_3$, we can write the estimator $\hat{S}$ as:
\begin{eqnarray}
\hat{S}^{{\mathtt{CIB-Lens}}} & = & {1 \over 6}\,\sum_{\ell_1\ell_2\ell_3}\,\sum_{m_1m_2m_3}
\left(\begin{array}{ccc}
\ell_1 & \ell_2 & \ell_3\\ m_1 & m_2 & m_3
\end{array}\right)
{a_{\ell_1m_1}a_{\ell_2m_2}a_{\ell_3m_3} \over
C_{\ell_1}C_{\ell_2}C_{\ell_3}}\,
\sqrt{(2\ell_1+1) (2\ell_2+1) (2\ell_3+1) \over 4\pi}
\left(\begin{array}{ccc}
\ell_1 & \ell_2 & \ell_3 \\ 0 & 0 & 0
\end{array}\right)
b^{({\mathtt{CIB-Lens}})}_{\ell_1\ell_2\ell_3}\nonumber \\
& = & {1 \over 6}\,\int\,d^2{\hat{\mathbf n}}\,\sum_{\ell_1\ell_2\ell_3}
\,\sum_{m_1m_2m_3}
{a_{\ell_1m_1}a_{\ell_2m_2}a_{\ell_3m_3} \over
C_{\ell_1}C_{\ell_2}C_{\ell_3}}
Y_{\ell_1 m_1}({\hat{\mathbf n}}) Y_{\ell_2 m_2}({\hat{\mathbf n}})
Y_{\ell_3 m_3}({\hat{\mathbf n}})\,b^{({\mathtt{CIB-Lens}})}_{\ell_1\ell_2\ell_3}\,
\label{s5e4}
\end{eqnarray}
where $C_{\ell}$ is the total power spectrum in the single frequency
map. By including the bispectrum formula (Eq.\,\ref{s2e3}) into
Eq.\,(\ref{s5e4}) and factorizing the $\ell$ dependence, the integral
becomes
\begin{eqnarray}
\hat{S}^{{\mathtt{CIB-Lens}}} & = & {1 \over 12}\int\,d^2{\hat{\mathbf n}}\,\sum_{\ell_1\ell_2\ell_3}
\sum_{m_1m_2m_3}\Bigg\{
\Bigg[\ell_1(\ell_1+1)a_{\ell_1m_1}Y_{\ell_1 m_1}({\hat{\mathbf n}})
{\tilde{C}^{{\mathtt{CMB}}}_{\ell_1} \over C_{\ell_1}}\Bigg]
\Bigg[{a_{\ell_2m_2}Y_{\ell_2 m_2}({\hat{\mathbf n}})
\over C_{\ell_2}}\Bigg]
\Bigg[a_{\ell_3m_3}Y_{\ell_3 m_3}({\hat{\mathbf n}})
{C^{({\mathtt{CIB-Lens}})}_{\ell_3} \over C_{\ell_3}}\Bigg] \nonumber \\
& & -\Bigg[a_{\ell_1m_1}Y_{\ell_1 m_1}({\hat{\mathbf n}})
{\tilde{C}^{{\mathtt{CMB}}}_{\ell_1} \over C_{\ell_1}}\Bigg]
\Bigg[\ell_2(\ell_2+1){a_{\ell_2m_2}Y_{\ell_2 m_2}({\hat{\mathbf n}})
\over C_{\ell_2}}\Bigg]
\Bigg[a_{\ell_3m_3}Y_{\ell_3 m_3}({\hat{\mathbf n}})
{C^{({\mathtt{CIB-Lens}})}_{\ell_3} \over C_{\ell_3}}\Bigg] \nonumber \\
& & +\Bigg[a_{\ell_1m_1}Y_{\ell_1 m_1}({\hat{\mathbf n}})
{\tilde{C}^{{\mathtt{CMB}}}_{\ell_1} \over C_{\ell_1}}\Bigg]
\Bigg[{a_{\ell_2m_2}Y_{\ell_2 m_2}({\hat{\mathbf n}})
\over C_{\ell_2}}\Bigg]
\Bigg[\ell_3(\ell_3+1)a_{\ell_3m_3}Y_{\ell_3 m_3}({\hat{\mathbf n}})
{C^{({\mathtt{CIB-Lens}})}_{\ell_3} \over C_{\ell_3}}\Bigg]+5perm.
\Bigg\}.
\label{s5e5}
\end{eqnarray}
Now, if we define the following filtered maps
\begin{equation}
P({\hat{\mathbf n}}) \equiv \sum_{\ell m}\,a_{\ell m}Y_{\ell m}({\hat{\mathbf n}})
{\tilde{C}^{({\mathtt{CMB}})}_{\ell} \over C_{\ell}} ~~~~~~~~~~
Q({\hat{\mathbf n}}) \equiv \sum_{\ell m}\,a_{\ell m}Y_{\ell m}({\hat{\mathbf n}})
{C^{({\mathtt{CIB-Lens}})}_{\ell} \over C_{\ell}} ~~~~~~~~~~
E({\hat{\mathbf n}}) \equiv \sum_{\ell m}\,{a_{\ell m}Y_{\ell m}({\hat{\mathbf n}})
\over C_{\ell}} \,,
\end{equation}
with the corresponding $\delta^2X$ maps (with $X=P,Q,E$) obtained by
substituting $a_{\ell m}$ with $\ell(\ell+1) a_{\ell m}$,
Eq.\,(\ref{s5e5}) becomes
\begin{equation}
\hat{S}^{{\mathtt{CIB-Lens}}}={1 \over 2}\int\,d^2{\hat{\mathbf n}}\,\bigg[
\delta^2P({\hat{\mathbf n}})E({\hat{\mathbf n}})Q({\hat{\mathbf n}})-P({\hat{\mathbf n}})\delta^2
E({\hat{\mathbf n}})Q({\hat{\mathbf n}})+P({\hat{\mathbf n}})E({\hat{\mathbf n}})\delta^2Q({\hat{\mathbf n}})
\bigg]\,.
\label{s5e6}
\end{equation}
The expressions in this section assume that the full inverse
covariance matrix can be replaced by a diagonal covariance term,
$(C^{-1}a)_{\ell m}\rightarrow a_{\ell m}/C_{\ell}$. In a real
experiment, this approximation might not be valid. In fact, we have to
take into account that clean CIB maps can be obtained only in small
areas of the sky due to the Galactic dust contamination. Here we are
using this approximation just for simplicity. Moreover, when
rotational invariance is broken by, e.g., a Galactic mask or an
anisotropic noise, a linear term should be subtracted from the
estimator in Eq.\,(\ref{s5e4}) \citep[see, e.g.][]{Mangilli:2013sxa}.
This linear term correction for a generic bispectrum shape
$b_{\ell_1 \ell_2 \ell_3}$ is given by:
\begin{equation}
\hat{S}_{lin}= -\frac{1}{2} \int d^2 \hat{n} \sum_{\ell m} b_{\ell_1 \ell_2 \ell_3}
\frac{a_{\ell_3m_3}}{C_{\ell_1}C_{\ell_2}C_{\ell_3}} Y_{\ell_1 m_1}({{\hat{n}}}) Y_{\ell_2 m_2}({{\hat{n}}}) Y_{\ell_3 m_3}({{\hat{n}}}).
\label{generic_linear_term}
\end{equation}
Using the explicit form of the CIB--lensing bispectrum, the linear
term correction for the single-frequency CIB--lensing estimator defined
in Eq. (\ref{s5e6}) is given by:
\begin{eqnarray}
S_{lin}^{{\mathtt{CIB-Lens}}} &=& \nonumber \\
&-& \frac{1}{2} \int d^2 \hat{n} \Big\{ Q( \hat{n}) \Big [ \langle P( \hat{n}) \delta^2E( \hat{n}) \rangle - \langle E( \hat{n}) \delta^2P( \hat{n}) \rangle \Big ] - \delta^2Q( \hat{n}) \langle P( \hat{n}) E( \hat{n}) \rangle - E( \hat{n}) \Big [ \langle Q( \hat{n}) \delta^2P( \hat{n}) \rangle - \langle P( \hat{n}) \delta^2Q( \hat{n}) \rangle \Big ] \nonumber \\
&+& \delta^2E( \hat{n}) \langle P( \hat{n}) Q( \hat{n}) \rangle
- \delta^2P( \hat{n}) \langle E( \hat{n}) Q( \hat{n}) \rangle + P( \hat{n}) \Big [ \langle Q( \hat{n}) \delta^2E( \hat{n}) \rangle - \langle E( \hat{n}) \delta^2Q( \hat{n}) \rangle \Big ] \Big\}.
\label{linear_term_single_freq}
\end{eqnarray}
The averages in Eq. (\ref{linear_term_single_freq}) correspond to
realistic Monte Carlo simulations which contain the instrumental
properties (beams, noise, masks) and also the type of
non-Gaussianity we are testing (in this case the CIB--lensing).
In the above treatment we have also supposed that the CIB--lensing
term is the only relevant non--Gaussian contribution. This is of
course not the case when a single frequency map is used. In fact,
the bispectrum from CIB sources and from ISW--lensing correlation
can give stronger contributions as shown in Fig.\,\ref{f2}. A joint
estimation of their amplitude can be applied
\citep[see][]{Lacasa:2012zx} using estimators developed considering
only one source of non--Gaussianity.
\subsection{Asymmetric estimator for CMB--CIB correlated maps}
The estimator for CMB--CIB correlated maps is defined in Eq.
(\ref{asymmetric_cmbcmbcib_s}). If we indicate the lensed CMB map and
the CIB map at frequency $\nu$ as
\begin{equation}
\bigg({\Delta T \over T}\bigg)^{({\mathtt{CMB}})}(\hat{\mathbf n})=
\sum_{\ell m}\,\tilde{a}^{({\mathtt{CMB}})}_{\ell m}Y_{\ell m}({\hat{\mathbf n}})~~~~~~~~
T_{\nu}^{({\mathtt{CIB}})}({\hat{\mathbf n}})=\sum_{\ell m}\,a^{({\mathtt{CIB}})}_{\ell m}Y_{\ell m}({\hat{\mathbf n}})\,,
\label{s5e7}
\end{equation}
the estimator in Eq. (\ref{asymmetric_cmbcmbcib_s}) becomes
now:
\begin{equation}
\hat{S}^{{\mathtt{CIB-Lens}} *}={1 \over 2}\,\sum_{\ell_1\ell_2\ell_3}\,\sum_{m_1m_2m_3}
G^{\ell_1\ell_2\ell_3}_{m_1m_2m_3}
{\tilde{a}^{({\mathtt{CMB}})}_{\ell_1m_1}\tilde{a}^{({\mathtt{CMB}})}_{\ell_2m_2}
a^{({\mathtt{CIB}})}_{\ell_3m_3} \over
\tilde{C}^{({\mathtt{CMB}})}_{\ell_1}\tilde{C}^{({\mathtt{CMB}})}_{\ell_2}C^{({\mathtt{CIB}})}_{\ell_3}}\,
b^{({\mathtt{CIB-Lens}})*}_{\ell_1\ell_2\ell_3}\,
\label{s5e8}
\end{equation}
where here the CIB--lensing bispectrum includes only two permutations
\begin{equation}
b^{({\mathtt{CIB-Lens}})*}_{\ell_1\ell_2\ell_3}={\ell_1(\ell_1+1)-\ell_2(\ell_2+1)
+\ell_3(\ell_3+1) \over 2}\,\tilde{C}^{({\mathtt{CMB}})}_{\ell_1}\,C^{({\mathtt{CIB-Lens}})}_{\ell_3}
+{\ell_2(\ell_2+1)-\ell_1(\ell_1+1)
+\ell_3(\ell_3+1) \over 2}\,\tilde{C}^{({\mathtt{CMB}})}_{\ell_2}\,C^{({\mathtt{CIB-Lens}})}_{\ell_3}\,.
\label{s5e8a}
\end{equation}
The factor $1/2$ in Eq.\,(\ref{s5e8}) is due to the fact the bispectrum
is now symmetric only in $\ell_1$ and $\ell_2$ and
$\sum_{\ell_1 \le \ell_2,\ell_3} \longrightarrow
\frac{1}{2}\sum_{\ell_1 \ell_2 \ell_3}(1+\delta_{\ell_1 \ell_2})\,$.
By defining new filtered maps as:
\begin{equation}
P_{{\mathtt{CMB}}}({\hat{\mathbf n}})\equiv \bigg({\Delta T \over T}\bigg)^{({\mathtt{CMB}})}(\hat{\mathbf n})=
\sum_{\ell m}\,\tilde{a}^{({\mathtt{CMB}})}_{\ell m} Y_{\ell m}({\hat{\mathbf n}}) ~~~~~
Q_{{\mathtt{CIB}}}({\hat{\mathbf n}})\equiv \sum_{\ell m}\,a^{({\mathtt{CIB}})}_{\ell m}Y_{\ell m}({\hat{\mathbf n}})
{C^{({\mathtt{CIB-Lens}})}_{\ell} \over C^{({\mathtt{CIB}})}_{\ell}}~~~~~
E_{{\mathtt{CMB}}}({\hat{\mathbf n}})\equiv\sum_{\ell m}\,{\tilde{a}^{({\mathtt{CMB}})}_{\ell m}Y_{\ell
m}({\hat{\mathbf n}}) \over \tilde{C}^{({\mathtt{CMB}})}_{\ell}}\,.
\label{s5e9}
\end{equation}
the estimator can be written in terms of filtered maps as:
\begin{equation}
\hat{S}^{{\mathtt{CIB-Lens}} *}=-{1 \over 2}\int\,d^2{\hat{\mathbf n}}\,\bigg[
\delta^2P_{{\mathtt{CMB}}}({\hat{\mathbf n}})E_{{\mathtt{CMB}}}({\hat{\mathbf n}})Q_{{\mathtt{CIB}}}({\hat{\mathbf n}})
-P_{{\mathtt{CMB}}}({\hat{\mathbf n}})\delta^2E_{{\mathtt{CMB}}}({\hat{\mathbf n}})Q_{{\mathtt{CIB}}}({\hat{\mathbf n}})+
P_{{\mathtt{CMB}}}({\hat{\mathbf n}})E_{{\mathtt{CMB}}}({\hat{\mathbf n}})\delta^2Q_{{\mathtt{CIB}}}({\hat{\mathbf n}})
\bigg]\,.
\label{s5e10}
\end{equation}
The linear term correction of this estimator can be straightforwardly
computed following the same steps already developed for the
single-frequency estimator:
\begin{eqnarray}
S_{lin}^{{\mathtt{CIB-Lens}} *} &=& \nonumber \\
&-& \frac{1}{2} \int d^2 \hat{n} \Big\{ Q_{{\mathtt{CIB}}}( \hat{n}) \Big [ \langle P_{{\mathtt{CMB}}}( \hat{n}) \delta^2E_{{\mathtt{CMB}}}( \hat{n}) \rangle - \langle E_{{\mathtt{CMB}}}( \hat{n}) \delta^2P_{{\mathtt{CMB}}}( \hat{n}) \rangle \Big ] - \delta^2Q_{{\mathtt{CIB}}}( \hat{n}) \langle P_{{\mathtt{CMB}}}( \hat{n}) E_{{\mathtt{CMB}}}( \hat{n}) \rangle \nonumber \\
&-& E_{{\mathtt{CMB}}}( \hat{n}) \Big [ \langle Q_{{\mathtt{CIB}}}( \hat{n}) \delta^2P_{{\mathtt{CMB}}}( \hat{n}) \rangle - \langle P_{{\mathtt{CMB}}}( \hat{n}) \delta^2Q_{{\mathtt{CIB}}}( \hat{n}) \rangle \Big ] \nonumber \\
&+& \delta^2E_{{\mathtt{CMB}}}( \hat{n}) \langle P_{{\mathtt{CMB}}}( \hat{n}) Q_{{\mathtt{CIB}}}( \hat{n}) \rangle
- \delta^2P_{{\mathtt{CMB}}}( \hat{n}) \langle E_{{\mathtt{CMB}}}( \hat{n}) Q_{{\mathtt{CIB}}}( \hat{n}) \rangle \nonumber \\
&+& P_{{\mathtt{CMB}}}( \hat{n}) \Big [ \langle Q_{{\mathtt{CIB}}}( \hat{n}) \delta^2E_{{\mathtt{CMB}}}( \hat{n}) \rangle - \langle E_{{\mathtt{CMB}}}( \hat{n}) \delta^2Q_{{\mathtt{CIB}}}( \hat{n}) \rangle \Big ] \Big\}.
\label{linear_term_double_freq}
\end{eqnarray}
\section{Summary and Conclusions}
\label{section6}
In this paper we have investigated the NG signal arising from the CIB
and its correlation with the lensing signal imprinted in the CMB
temperature anisotropies, and we have estimated the bias they can
induce on the local, equilateral and orthogonal $f_{\rm nl~}$ parameter using
{\it Planck}~ data. The bias is computed for `raw' single--frequency
temperature maps, i.e. maps on which no component separation is
applied, and for maps cleaned by the SEVEM component separation
technique. For these maps, we have used the {\it Planck}~ instrumental
characteristics and we have assumed they are free from Galactic
foregrounds. We have then studied the possibility to detect the
CIB--lensing bispectrum in the {\it Planck}~ data.
CIB intensity fluctuations have been modelled following
\citet{planck_xviii_2011}. The parameters of the model have been
updated in order to have a better agreement with the recent {\it Planck}~
measurements of the CIB power spectra \citep{planck_xxx_2013}. We have
also considered the contribution from extragalactic radio sources and
from their correlation with CMB lensing. As expected,
radio--lensing power spectra and bispectra are found to be small and
negligible at the frequencies used for the cosmological analysis in
{\it Planck}.
Below we summarise and discuss our results.
\begin{itemize}
\item The bias $\Delta$$f_{\rm nl~}$ induced by the CIB--lensing correlation is
small but not negligible for orthogonal shapes in the ``raw'' 143
and 217\,GHz {\it Planck} maps, approximately $ -21$ and $ -88$
respectively. However, when maps are cleaned with a component
separation technique, the bias is strongly reduced and becomes
almost negligible for {\it Planck} results (the largest bias appears
for the orthogonal shape and amounts to $0.3\, \sigma$).
\item We have estimated the bias produced by the intrinsic bispectrum
of extragalactic sources. In agreement with the discussion in
\citet{planck_xxiv_2013}, point sources turn out to be not a severe
contaminant for $f_{\rm nl~}$ studies with {\it Planck} foreground--reduced
data, even though not completely negligible (the largest bias is for
the equilateral shape amounting to $0.45\, \sigma$). In ``raw''
maps they produce a significant bias only for equilateral shapes,
that is 160, 54 and 60 at 100, 143 and 217\,GHz respectively.
\item Our results confirm the capability and stress the importance of
component separation techniques in removing extragalactic
foregrounds as well. On the other hand, our results also
predict that future experiments, with better sensitivity to the
$f_{\rm nl~}$ parameter, might have to consider extragalactic sources and the
CIB--lensing correlation as further serious contaminants in some
particular shapes.
\item The detection of the CIB--lensing bispectrum signal directly
from {\it Planck} maps is not straightforward and it is feasible
only for the 217\,GHz channel (with a significance of $
\sim7.5\,\sigma$, assuming full sky coverage). Nevertheless, we have
shown in this paper that a more efficient way to detect the
bispectrum cross--correlates CMB maps with CIB maps at different
frequencies. In this case the CIB--lensing bispectrum signal could
be detected with very high significance ($ \la63\,\sigma$) if
accurate CIB maps can be extracted at 353, 545 and 857\,GHz over a
large area of the sky. \citet{planck_xxx_2013} were able to produce
clean CIB maps just over $ \approx$ 10\% of the sky; in this case we
still expect a high level of significance, of approximately {
20}\,$\sigma$ or more. We have to note however that the cross
correlation between CMB and CIB maps is not a simple procedure as
possible residuals of the CIB (CMB) in ``clean'' CMB (CIB) maps
could produce spurious signals that can be easily misinterpreted.
\item We have compared our predictions for the detectability levels of
the CIB--lensing bispectrum on {\it Planck}~ data with the results already
obtained by the {\it Planck}~ Collaboration using the cross-spectrum
estimator \citep{planck_xviii_2013}. Our results are nearly
equivalent to the ones presented in that work when we compare the
case with $f_{sky}=30.4\%$, the lensing reconstruction at 143
GHz and the CIB dominated bands of 353, 545 and 857 GHz (see Table
\ref{t7}). This shows that both approaches are equivalent in terms
of efficiency and both of them can be used to provide more robust
results as different estimators might have different sensitivity to
systematics.
\item Finally, we have developed an optimal estimator for the
CIB--lensing bispectrum, based on the KSW formalism. Two different
estimators have been constructed, one for use with single--frequency
maps and a second one for separate CMB and CIB maps.
\end{itemize}
\section*{acknowledgements}
The authors acknowledge financial support from the Spanish Ministerio
de Econom\'ia y Competitividad project AYA-2012-39475-C02-01 and the
Consolider Ingenio-2010 Programme project CSD2010-00064, as well as
from the Swiss National Science Foundation. AC acknowledges the
Spanish Consejo Superior de Investigaciones Cient\'ificas (CSIC) and
the Spanish Ministerio de Educaci\'on, Cultura y Deporte for a
postdoctoral fellowship at the Cavendish Laboratory of the University
of Cambridge (UK). AC is thankful to Airam Marcos Caballero and Marina
Migliaccio for their useful comments that have helped in the
production of this paper. MK thanks Peter Wittwer for pointing out
helpful mathematical identities. The authors acknowledge useful
discussions and the feedback provided regarding the lensing to Anthony
Challinor and Antony Lewis. The authors acknowledge useful discussions
regarding the SEVEM component separation technique to Bel\'en Barreiro
and Patricio Vielva. The authors are grateful to the anonymous
referee whose revision helped improving this article. The authors
acknowledge the computer resources, technical expertise and assistance
provided by the Spanish Supercomputing Network nodes at Universidad de
Cantabria and Universidad Polit\'ecnica de Madrid. Some of the
calculations used the Andromeda cluster of the University of
Geneva. We have also used the software packages HEALPix
\citep{healpix} and CAMB \citep{lewis2000}.
|
\section{Introduction}
On April 20, 2010, methane gas leakage on the Deepwater Horizon oil rig operated by Transocean, a subcontractor of British Petroleum (BP), caused a big explosion \cite{bps_13b}. This leakage not only killed 11 workers instantly but destroyed and sank the rig, and caused millions of gallons of oil to pour into the Gulf of Mexico. The gushing well, about a mile under the sea, was finally brought under control after more than three months of frenetic attempts. The spill, which is considered to be the largest accidental marine oil spill in the history of the petroleum industry, caused extensive damage to marine and wildlife habitats as well as the Gulf's fishing and tourism industries and its impact still continues. Just like the BP pipeline, there are tens of thousands of miles long oil and gas pipelines around the world. All of these pipelines are aging and are becoming more and more susceptible to failures, which may lead to disasters like the BP one. Hence, it is very important to do rigorous reliability analysis of oil and gas pipelines to detect and rectify potential problems.
The reliability analysis of a pipeline system involves a three-step process: (i) partitioning the given pipeline into segments and constructing its equivalent reliability block diagram (RBD), (ii) assessing the reliability of the individual segments and (iii) evaluating the reliability of the complete pipeline system based on the RBD and the reliability of its individual segments. The reliability of an individual segment is usually expressed in terms of its failure rate $\lambda$ and a random variable, like exponential \cite{Zhang_08} or Weibull random variable \cite{Kolowrocki_09}, which models the failure time. A single oil or gas pipeline can be simply modeled as a series RBD \cite{Zhang_08}. However, in many cases, these pipeline systems have either reserved components or subsystems and such pipeline systems exhibit a combination of series and parallel RBDs \cite{Soszynska_10}.
The reliability analysis of oil and gas pipelines has predominantly been accomplished by first gathering data
from in-line inspection tools to detect cracks, corrosion or damage \cite{pipe_integ_sol_13,pipe_check_13}. This information is then manipulated using the
paper-and-pencil based analytical analysis and computer simulations to deliver diagnostics and
insightful pipeline integrity reports (e.g. \cite{panday_98,Zhang_08,Soszynska_10}). However, due to the complex nature of large pipeline system analysis, paper-and-pencil proof methods are error prone and the exhaustive testing of all possible system behaviors using simulation is almost impossible. Thus, these traditional analysis techniques cannot guarantee accurate results, which is a severe limitation in the case of oil and gas pipelines as an uncaught system bug may endanger human and animal life or lead to a significant financial loss.
The inaccuracy limitations of traditional analysis techniques can be overcome by using formal methods \cite{boca_09}, which use computerized mathematical reasoning to precisely model the system's intended behavior and to provide irrefutable proof that a system satisfies its requirements. Both model checking and theorem proving have
been successfully used for the precise probabilistic analysis of a
broad range of systems (e.g. \cite{hasan_ispass_08,KNP_10a,elleuch_11,hasan_jal_11,fruth_11}). However, to the best of our knowledge, no formal analysis approach has been used for the reliability analysis of oil and gas pipelines so far. The foremost requirement for conducting the formal reliability analysis
of underground oil and gas pipelines is the ability to formalize RBDs recursively and continuous random variables.
Model checking is a state-based formal method technique. The inherent limitations of model checking is the state-space explosion problem and the inability to model complex datatypes such as trees, lists and recursive definitions \cite{kaufman2004some}.
On the other hand, higher-order logic \cite{cebrown_07} is a system of deduction with a
precise semantics and can be used to formally model any system that
can be described mathematically including recursive definitions, random variables, RBDs, and continuous components. Similarly, interactive theorem provers are computer based formal
reasoning tools that allow us to verify higher-order-logic properties under user
guidance. Higher-order-logic theorem provers can be used to reason about recursive definitions using induction methods \cite{kapur1996lemma}. Thus, higher-order-logic theorem proving can be used to conduct the formal analysis of oil and gas pipelines.
A number of higher-order-logic formalizations of probability theory are available in higher-order logic
(e.g. \cite{hurd_02,mhamdi_11,holzl_11}). Hurd's formalization of probability theory \cite{hurd_02} has been utilized to
verify sampling algorithms of a number of commonly used discrete
\cite{hurd_02} and continuous random variables \cite{hasan_cade_07}
based on their probabilistic and statistical properties
\cite{hasan_icnaam_07,hasan_fm_09}. Moreover, this formalization has been used to conduct the reliability analysis of a number of applications, such as memory arrays \cite{hasan_tc_10}, soft errors \cite{abbasi_13b} and electronic components \cite{abbasi_13}. However, Hurd's formalization of probability theory only supports having the whole universe as the probability space. This feature limits its scope and thus this probability theory cannot be used to formalize more than a single continuous random variable. Whereas, in the case of reliability analysis of pipelines, multiple continuous random variables are required. The recent formalizations of probability theory by Mhamdi
\cite{mhamdi_11} and H\"{o}lzl~\cite{holzl_11} are based on extended
real numbers (including $\pm\infty$) and provide the formalization of
Lebesgue integral for reasoning about advanced statistical
properties. These theories also allow using any arbitrary probability space that is a subset of the universe and thus are more flexible than Hurd's formalization. However, to the best of our knowledge, these foundational theories have not been used to formalize neither reliability and RBDs nor continuous random variables so far.
In this paper, we use Mhamdi's formalization of probability theory \cite{mhamdi_11}, which is available in the HOL theorem prover \cite{norris_hol}, to formalize reliability and the commonly used series RBD, where its individual segments are modeled as random variables. Our formalization includes various formally verified properties of reliability and series RBD that facilitate formal reasoning about the reliability of some simple pipelines using a theorem prover. To analyze more realistic models of pipelines, it is required to formalize other RBDs, such as parallel, series-parallel and parallel-series \cite{Bilinton_1992}. In order to illustrate the utilization and effectiveness of the proposed idea, we utilize the above mentioned formalization to analyze a simple pipeline that can be modeled as a series RBD with an exponential failure time for individual segments.
\section{Preliminaries}
\label{sec_2}
In this section, we give a brief introduction to theorem proving in general and the HOL theorem
prover in particular. The intent is to introduce the main
ideas behind this technique to facilitate the understanding of the paper for the reliability analysis community. We also summarize Mhamdi's formalization of probability theory \cite{mhamdi_11} in this section.
\subsection{Theorem Proving}
Theorem proving \cite{gordon_89} is a widely used formal
verification technique. The system that needs to be analysed is
mathematically modelled in an appropriate logic and the properties of
interest are verified using computer based formal tools. The use of
formal logics as a modelling medium makes theorem proving a very
flexible verification technique as it is possible to formally verify
any system that can be described mathematically. The core of theorem
provers usually consists of some well-known axioms and primitive
inference rules. Soundness is assured as every new theorem must be
created from these basic or already proved axioms and primitive inference rules.
The verification effort of a theorem in a theorem prover varies from
trivial to complex depending on the underlying logic
\cite{harrison_96a}. For instance, first-order logic
\cite{fitting_96} utilizes the propositional calculus and terms
(constants, function names and free variables) and is
semi-decidable. A number of sound and complete first-order logic
automated reasoners are available that enable completely automated
proofs. More expressive logics, such as higher-order logic
\cite{cebrown_07}, can be used to model a wider range of problems
than first-order logic, but theorem proving for these logics cannot
be fully automated and thus involves user interaction to guide the
proof tools. For reliability analysis of pipelines, we need to formalize
(mathematically model) random variables as functions and their distribution properties are verified by quantifying over random variable functions. Henceforth, first-order logic does not support such formalization and we need to use higher-order logic to formalize the foundations of reliability analysis of pipelines.
\subsection{HOL Theorem Prover}
HOL is an interactive theorem prover developed at the
University of Cambridge, UK, for conducting proofs in higher-order logic.
It utilizes the simple type theory of Church \cite{church_40} along
with Hindley-Milner polymorphism \cite{milner_77} to implement
higher-order logic. HOL has been successfully used as a verification
framework for both software and hardware as well as a platform for
the formalization of pure mathematics.
The HOL core consists of only 5 basic
axioms and 8 primitive inference rules, which are implemented as ML
functions. Soundness is assured as every new theorem must be verified by applying these basic axioms and primitive inference rules or any other previously verified theorems/inference rules.
We utilized the HOL theories of
Booleans, lists, sets, positive integers, \emph{real} numbers,
measure and probability in our work. In fact, one of the primary
motivations of selecting the HOL theorem prover for our work was to
benefit from these built-in mathematical theories.
Table \ref{hol_basics} provides the mathematical interpretations of
some frequently used HOL symbols and functions, which are inherited
from existing HOL theories, in this paper.
\begin{table}[!htb]
\begin{center}
\begin{tabular}{|c|c|c|} \hline
{\bfseries HOL Symbol} & {\bfseries Standard Symbol} & {\bfseries Meaning} \\
\hline \hline
$\mathtt{\wedge}$& $and$ & Logical $and$ \\ \hline
$\mathtt{\vee}$ & $or$ & Logical $or$ \\ \hline
$\mathtt{\neg}$ & $not$ & Logical $negation$ \\ \hline
$\mathtt{::}$ & $cons$ & Adds a new element to a list \\ \hline
$\mathtt{++}$ & $append$ & Joins two lists together \\ \hline
$\mathtt{HD\ L}$ & $head$ & Head element of list $L$ \\ \hline
$\mathtt{TL\ L}$ & $tail$ & Tail of list $L$\\ \hline
$\mathtt{EL\ n\ L}$ & $element$ & $n^{th}$ element of list L \\ \hline
$\mathtt{MEM\ a\ L}$ & $member$ & True if $a$ is a member of list $L$\\ \hline
$\mathtt{\lambda x.t}$& $\lambda x.t$ & Function that maps $x$ to $t(x)$ \\ \hline
$\mathtt{SUC\ n}$& $n + 1$ & Successor of a $num$ \\ \hline
$\mathtt{lim(\lambda n.f(n))}$ & $\mathop {\lim }\limits_{n \to \infty } f(n)$ & Limit of a $real$ sequence $f$ \\
\hline
\end{tabular}
\caption{HOL Symbols and Functions}
\label{hol_basics}
\end{center}
\end{table}
\subsection{Probability Theory and Random Variables in HOL}
Mathematically, a measure space is defined as a triple ($\Omega,\Sigma, \mu$), where
$\Omega$ is a set, called the sample space, $\Sigma$ represents a $\sigma$-algebra of subsets of
$\Omega$, where the subsets are usually referred to as measurable sets, and $\mu$ is a measure with domain
$\Sigma$. A probability space is a measure space ($\Omega,\Sigma, Pr$), such that the measure,
referred to as the probability and denoted by $Pr$, of the sample space is 1. In Mhamdi's formalization of probability theory \cite{mhamdi_11}, given a probability space $p$, the functions \texttt{space}
and \texttt{subsets} return the corresponding
$\Omega$ and $\Sigma$, respectively. This formalization also includes the formal verification of some of the most widely used probability axioms, which play a pivotal role in formal reasoning about reliability properties.
Mathematically,
a random variable is a measurable function between a probability space and a
measurable space. A measurable space refers to a pair
($S,\mathcal{A}$), where $S$ denotes a set and $\mathcal{A}$ represents a nonempty collection of sub-sets of $S$. Now, if $S$ is a set with finite elements, then the corresponding random variable is termed as a discrete random variable and else it is called a continuous one. The
probability that a random variable $X$ is less than or equal to some value
$x$, $Pr(X \le x)$ is called the cumulative distribution function (CDF) and it characterizes the distribution of both discrete and continuous random variables. Mhamdi's formalization of probability theory \cite{mhamdi_11} also includes the formalization of random variables and the formal verification of some of their classical properties using the HOL theorem prover.
\section{Reliability}
\label{sec_3}
In reliability theory \cite{Bilinton_1992}, reliability $R(t)$ of a system or
component is defined as the probability that it performs its
intended function until some time $t$.
\begin{equation}
\label{reliability_eq} R(t) = Pr (X > t) = 1 - Pr (X \le t) = 1 - F_X(t)
\end{equation}
where $F_X(t)$ is the CDF. The random variable $X$, in the above definition, models
the time to failure of the system. Usually, this time to failure is
modeled by the exponential random variable with parameter $\lambda$
that represents the failure rate of the system. Now, the CDF can be modeled in HOL as follows:
\begin{flushleft}
\texttt{\bf{Definition 1: }} \label{CDF_def}
\emph{Cumulative Distributive Function} \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p X x. CDF p X x = distribution p X \{y | y $\leq$ Normal x\}
}
\end{flushleft}
\noindent where $p$ represents the probability space, $X$ is the random variable and $x$ represents a $real$ number. The function \texttt{Normal} converts a $real$ number to its corresponding value in the $extended-real$ data-type, i.e, the $real$ data-type including the positive and negative infinity. The function \texttt{distribution} accepts a probability space $p$, a random variable $X$ and a set and returns the probability of $X$ acquiring all the values of the given set in the probability space $p$. Now, Definition 1 can be used to formalize the reliability definition, given in Equation \ref{reliability_eq}, as follows:
\begin{flushleft}
\texttt{\bf{Definition 2: }} \label{Reliability_def}
\emph{Reliability} \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p X x. Reliability p X x = 1 - CDF p X x
}
\end{flushleft}
We used the above mentioned formal definition of reliability to formal verify some of the classical properties of reliability in HOL. The first property in this regard relates to the fact that the reliability of a good component is 1, i.e., maximum, prior to its operation, i.e., at time 0. This property has been verified in HOL as the following theorem.
\begin{flushleft}
\texttt{\bf{Theorem 1: }} \label{Reliability_AT_ZERO}
\emph{Maximum Reliability} \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p X. prob\_space p $\wedge$ (events p = POW (p\_space p)) $\wedge$ \\
\ \ \ \ \ \ \ ($\forall$ y. X y $\neq$ NegInf $\wedge$ X y $\neq$ PosInf) $\wedge$ \\
\ \ \ \ \ \ \ ($\forall$ z. 0 $\le$ z $\Rightarrow$ ($\lambda$x. CDF p X x) contl z) $\wedge$\\
\ \ \ \ \ \ \ ($\forall$ x. Normal 0 $\le$ X x) $\Rightarrow$\\
\ \ \ \ \ \ \ (Reliability p X 0 = 1)
}
\end{flushleft}
\noindent The first two assumptions of the above theorem ensure that the variable \emph{p} represents a valid probability space based on the formalization of Mhamdi's probability theory \cite{mhamdi_11}. The third assumption constraints the random variable to be well-defined, i.e., it cannot acquire negative or positive infinity values. The fourth assumption states that the CDF of the random variable $X$ is a continuous function, which means that $X$ is a continuous random variable. This assumption utilizes the HOL function \texttt{contl}, which accepts a lambda abstraction function and a real value and ensures that the function is continuous at the given value. The last assumption ensures that the random variable $X$ can acquire positive values only since in the case of reliability this random variable always models time, which cannot be negative. The conclusion of the theorem represents our desired property that reliability at \emph{time=0} is \emph{1}.
The proof of the Theorem 1 exploits some basic probability theory axioms and the following property according to which
the probability of a continous random variable at a point is zero.
The second main characteristic of the reliability function is its decreasing monotonicity, which is verified as the following theorem in HOL:
\begin{flushleft}
\texttt{\bf{Theorem 2: }} \label{Reliability_MONOTONE}
\emph{Reliability is a Monotone Function} \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p X a b.
prob\_space p $\wedge$ (events p = POW (p\_space p)) $\wedge$ \\
\ \ \ \ \ \ \ ($\forall$ y. X y $\neq$ NegInf $\wedge$ X y $\neq$ PosInf) $\wedge$ \\
\ \ \ \ \ \ \ ($\forall$ x. Normal 0 $\le$ X x) $\wedge$ a $\le$ b $\Rightarrow$\\
\ \ \ \ \ \ \ (Reliability p X (b)) $\leq$ (Reliability p X (a))
}
\end{flushleft}
\noindent The assumptions of this theorem are the same as the ones used for Theorem 1 except the last assumption, which describes the relationship between variables $a$ and $b$. The above property clearly indicates that the reliability cannot increase with the passage of time.
The formal reasoning about the proof of Theorem 2 involves some basic axioms of probability theory and a property that the CDF is a monotonically increasing function.
Finally, we verified that the reliability tends to 0 as the time approaches infinity. This property is verified under the same assumptions that are used for Theorem 1.
\begin{flushleft}
\texttt{\bf{Theorem 3: }} \label{Reliability_TENDS}
\emph{Reliability Tends to Zero As Time Approaches Infinity} \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p X. prob\_space p $\wedge$ (events p = POW (p\_space p)) $\wedge$ \\
\ \ \ ($\forall$ y. X y $\neq$ NegInf $\wedge$ X y $\neq$ PosInf) $\wedge$ ($\forall$ x. Normal 0 $\le$ X x) $\Rightarrow$\\
\ \ \ \ \ \ \ (lim ($\lambda$n. Reliability p X (\&n)) = 0)
}
\end{flushleft}
\noindent The HOL function \texttt{lim} models the limit of a real sequence. The proof of Theorem 3 primarily uses the fact that the CDF approches to 1 as its argument approaches infinity.
These three theorems completely characterize the behavior of the reliability function on the positive real axis as the argument of the reliability is time and thus cannot be negative. The formal verification of these properties based on our definition ensure its correctness. Moreover, these formally verified properties also facilitate formal reasoning about reliability of systems, as will be demonstrated in Section \ref{sec_5} of this paper. The proof details about these properties can be obtained from our proof script \cite{waqar_ftscs_13}.
\section{Formalization of Series Reliability Block Diagram}
\label{sec_4}
In a serially connected system \cite{Bilinton_1992}, depicted in Figure 1, the reliability of the complete system mainly depends upon the failure of a single component that has the minimum reliability among all the components of the system. In other words, the system stops functioning if any one of its component fails. Thus, the operation of such a system is termed as reliable at any time $t$, if all of its components are functioning reliably at this time $t$. If the event $A_{i}(t)$ represents the reliable functioning of the $i^{th}$ component of a serially connected system with $N$ components at time $t$ then the overall reliability of the system can be mathematically expressed as \cite{Bilinton_1992}:
\begin{equation}\label{eq2}
R_{series}(t) = Pr (A_{1}(t) \cap A_{2}(t) \cap A_{3}(t) \cdots \cap A_{N}(t))
\end{equation}
\begin{figure}\label{series_fig}
\centering
\includegraphics[width = 0.5\textwidth]{latest_series.eps}
\begin{center}
\caption{System with a Series Connection of Components}
\end{center}\label{series_fig_caption}
\end{figure}
\noindent Using the assumption of mutual independence of individual reliability events of a series system \cite{Bilinton_1992}, the above equation can be simplified as:
\begin{equation}\label{eq3}
R_{series}(t) = \prod_{i=1}^{N}R_{i}(t)
\end{equation}
Moreover, an intrinsic property of a series system is that its overall reliability is always less than or equal to the reliability of the sub-component with the least reliability.
\begin{equation}\label{eq4}
R_{series}(t) \leq min(R_{i}(t))
\end{equation}
We proceed with the formalization of the series RBD by first formalizing the notion of mutual independence of more than two random variables, which is one of the most essential prerequisites for reasoning about the simplified expressions for RBD. Two events $A$ and $B$ are termed as mutually independent iff $Pr(A\cap B) = Pr(A)Pr(B)$. All the events involved in reliability modeling are generally assumed to be mutually independent. Since we often tackle the reliability assessment of systems with more than two components, we formalize the mutual independence of a list of random variables in this paper as follows:
\begin{flushleft}
\texttt{\bf{Definition 3: }} \label{mutual_indep_def}
\emph{Mutual Independence of Events } \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p L. mutual\_indep p L = \\
\ \ $\forall$ L1 \vspace{1pt} n. PERM L L1 $\wedge$ 2 $\leq$ n $\wedge$ n $\leq$ LENGTH L $\Rightarrow$ \\
\ \ \ \ \ \ prob p (inter\_set p (TAKE n L1)) = \\ \ \ \ \ \ \ list\_prod (list\_prob p (TAKE n L1))
}
\end{flushleft}
\noindent The function \texttt{mutual\_indep} takes a list of events or sets $L$ along with the probability
space $p$ as input and returns True if the given list of events are mutually independent in $p$.
The formal definitions for the HOL functions used in the above definition are given in Table 1. The predicate \texttt{PERM} ensures that its two list arguments form a permutation of one another, the function \texttt{LENGTH} returns the length of a list, the function \texttt{TAKE} returns a list that contains the first $n$ elements of its argument list, the function \texttt{inter\_set} performs the intersection of all the sets in a list of sets and returns the probability space in case of an empty list argument, the function \texttt{list\_prob} returns a list of probabilities associated with the given list of events in the given probability space and the function \texttt{list\_prod} recursively multiplies all the elements of its argument list of real numbers. Thus, using these functions the function \texttt{mutual\_indep} ensures that for any 2 or more elements $n$, taken in any order, of the given list of events $L$, the property $Pr(\bigcap_{i=0}^nL_i) = \prod_{i=0}^nPr(L_i)$ holds.
\begin{table}
\begin{tabular}[c]{|l |l|}
\hline
\texttt{\textbf{Function Name}} \ \ & \texttt{\textbf{HOL Definition}} \\ \hline
\hline
\texttt{PERM} & \texttt{$\vdash$ $\forall$ L1 L2. PERM L1 L2 =} \\& \ \ \ \texttt{$\forall$ x.
FILTER (\$= x) L1 = FILTER (\$= x)L2 } \\ \hline
\texttt{LENGTH } & \texttt{$\vdash$ (LENGTH [] = 0 )} $\wedge$ \\
& \ \ \ \texttt{ $\forall$ h t. LENGTH (h::t) = SUC (LENGTH t)} \\
\hline
\texttt{TAKE } & \texttt{$\vdash$ ($\forall$ n. TAKE n [] = [])} $\wedge$ \\
& \ \ \ \texttt{$\forall$ n x xs. TAKE n (x::xs) = if n = 0 then [] else } \\
& \ \ \ \ \ \texttt{x::TAKE (n - 1) xs}\\
\hline
\texttt{inter\_set } & \texttt{$\vdash$ ($\forall$ p. inter\_set p [] = p\_space p )} $\wedge$ \\
& \ \ \ \texttt{$\forall$ p h t. inter\_set p (h::t) = h $\cap$ inter\_set p t} \\
\hline
\texttt{list\_prod } & \texttt{$\vdash$ ($\forall$ list\_prod [] = 1)} $\wedge$ \\
& \ \ \ \texttt{$\forall$ h t. list\_prod (h::t) = h * list\_prod t} \\
\hline
\texttt{list\_prob } & \texttt{$\vdash$ ($\forall$ p. list\_prob p [] = [])} $\wedge$ \\
& \ \ \ \texttt{$\forall$ p h t. list\_prob p (h::t) =} \\
& \ \ \ \ \ \ \ \texttt{prob p (h $\cap$ p\_space p) * list\_prob p t} \\
\hline
\texttt{min} & \texttt{$\vdash$ $\forall$ x y. min x y = if x $\leq$ y then x else y }\\
\hline
\texttt{min\_rel } & \texttt{$\vdash$ ($\forall$ f. min\_rel f [] = 1)} $\wedge$ \\
& \ \ \ \texttt{$\forall$ f h t. min\_rel f (h::t) = min (f h) (min\_rel f t)} \\
\hline
\end{tabular}\caption{HOL Functions used in Definition 3}
\end{table}
Next, we propose to formalize the RBDs in this paper by using a list of events, where each event models the proper functioning of a single component at a given time based on the corresponding random variable. This list of events can be modeled as follows:
\begin{flushleft}
\texttt{\bf{Definition 4: }} \label{list_RV_def}
\emph{Reliability Event List} \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p x. rel\_event\_list p [] x = [] $\wedge$ \\
\ \ $\forall$ p x h t. rel\_event\_list p (h::t) x = \\
PREIMAGE h \{y | Normal x < y\} $\cap$ p\_space p :: rel\_event\_list p t x
}
\end{flushleft}
\noindent The function \texttt{rel\_event\_list} accepts a list of random variables, representing the time
to failure of individual components of the system, and a $real$ number $x$, which represents the time index
where the reliability is desired, and returns a list of sets corresponding to the events that the individual
components are functioning properly at the given time $x$. This list of events can be manipulated, based on the
structure of the complete system, to formalize various RBDs.
Similarly, the individual reliabilities of a list of random variables can be modeled as the following recursive
function:
\begin{flushleft}
\texttt{\bf{Definition 5: }} \label{list_reliability_function_def}
\emph{Reliability of a List of Random Variables } \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p x . rel\_list p [] x = [] $\wedge$\\
\ \ $\forall$ p h t x. rel\_list p (h::t) x = \\ \ \ Reliability p h x :: rel\_list p t x
}
\end{flushleft}
\noindent The function \texttt{rel\_list} takes a list of random variables and a $real$ number $x$, which
represents the time index where the reliability is desired, and returns a list of the corresponding
reliabilities at the given time $x$. It is important to note that all the above mentioned definitions are
generic enough to represent the behavior of any RBD, like series, parallel, series-parallel
and parallel-series.
Now, using Equation (\ref{eq2}), the reliability of a serially connected structure can be defined as:
\begin{flushleft}
\texttt{\bf{Definition 6: }} \label{series_def}
\emph{System with a Series Connection of Components } \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p L. rel\_series p L = prob p (inter\_set p L)
}
\end{flushleft}
\noindent The function \texttt{rel\_series} takes a list of random variables $L$, representing the failure times of the individual components of the system, and a probability space $p$ as input and returns
the intersection of all the events corresponding to the reliable functioning of these components using the
function \texttt{inter\_set}, given in Table 2. Based on this definition, we formally verified the result of Equation (\ref{eq2}) as follows:
\begin{flushleft}
\texttt{\bf{Theorem 4: }} \label{series_connected_system_def}
\emph{Reliability of a System with Series Connections} \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p L x. prob\_space p $\wedge$ (events p = POW (p\_space p)) $\wedge$\\
\ \ \ \ \ 0 $\leq$ x $\wedge$
2 $\leq$ LENGTH (rel\_event\_list p L x) $\wedge$ \\
\ \ \ mutual\_indep p (rel\_event\_list p L x) $\Rightarrow$ \\
\ (rel\_series p (rel\_event\_list p L x) = list\_prod (rel\_list p L x))
}
\end{flushleft}
\noindent The first two assumptions ensure that $p$ is a valid probability space based on Mhamdi's probability
theory formalization \cite{mhamdi_11}. The next one ensures that the variable $x$, which models time,
is always greater than or equal to 0. The next two assumptions of the above theorem guarantee that we have a
list of at least two mutually exclusive random variables (or a system with two or more components).
The conclusion of the theorem represents Equation (\ref{eq2}) using Definitions 4 and 6. The proof of Theorem 4
involves various probability theory axioms, the mutual independence of events and the fact that the probability
of any event that is in the returned list from the function \texttt{rel\_event\_list} is equivalent to its
reliability. More proof details can be obtained from our proof script \cite{waqar_ftscs_13}.
Similarly, we verified Equation (4) as the following theorem in HOL:
\begin{flushleft}
\texttt{\bf{Theorem 5: }} \label{series_connected_system_min_reliability_def}
\emph{Reliability of a System depends upon the minimum reliability of the connected components} \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p L x. prob\_space p $\wedge$ (events p = POW (p\_space p)) $\wedge$ \\
\ \ \ \ \ 0 $\leq$ x $\wedge$
2 $\leq$ LENGTH (rel\_event\_list p L x) $\wedge$ \\
\ \ \ \ \ mutual\_indep p (rel\_event\_list p L x) $\Rightarrow$ \\
\ \ \ \ \ \ (rel\_series p (rel\_event\_list p L x) $\leq$ \\
\ \ \ \ \ \ \ \ min\_rel ($\lambda$ L. Reliability p L x) L)
}
\end{flushleft}
The proof of the Theorem 5 uses several probability theory axioms and the fact that any subset of a mutually independent set is also mutually independent.
The definitions, presented in this section, can be used to model parallel RBD \cite{Bilinton_1992} and formally verify the corresponding simplified reliability relationships as well. The major difference would be the replacement of the function \texttt{inter\_set} in Definition 6 by a function that returns the union of a given list of events.
\section{Reliability Analysis of a Pipeline System}
\label{sec_5}
A typical oil and gas pipeline can be partitioned into a series connection of $N$ segments,
where these segments may be classified based on their individual failure times. For example, a 60
segment pipeline is analyzed in \cite{Zhang_08} under the assumption that the segments, which exhibit
exponentially distributed failure rates, can be sub-divided into 3 categories according to their failure
rates ($\lambda$), i.e., 30 segments with $\lambda= 0.0025$, 20 segments with $\lambda= 0.0023$ and
10 segments with $\lambda= 0.015$. The proposed approach for reliability analysis of pipelines allows us to
formally verify generic expressions involving any number of segments and arbitrary failure rates.
In this section, we formally verify the reliability of a simple pipeline, depicted in Figure 2, with $N$ segments having arbitrary exponentially distributed failure times.
\begin{figure}\label{series_pipeline_fig}
\centering
\includegraphics[width = 0.47\textwidth]{latest_pipeline.eps}\\
\caption{A Simple Pipeline}\label{series_pipeline_caption_fig}
\end{figure}
We proceed with the formal reliability analysis of the pipeline, shown in Figure 2, by formalizing the exponential random variable in HOL.
\begin{flushleft}
\texttt{\bf{Definition 7: }} \label{Exponential_distribution_def}
\emph{Exponential Distribution Function } \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p X l. exp\_dist p X l = \\
\ \ \ $\forall$ x. (CDF p X x = if 0 $\leq$ x then 1 - exp (-l * x) else 0)
}
\end{flushleft}
\noindent The predicate \texttt{exp\_dist} ensures that the random variable $X$ exhibits the CDF of an
exponential random variable in probability space $p$ with failure rate $l$. We classify a list of exponentially distributed random variables based on this definition as follows:
\begin{flushleft}
\texttt{\bf{Definition 8: }} \label{list_of exponential_distribution_function_def}
\emph{List of Exponential Distribution Functions} \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p L. list\_exp p [] L = T $\wedge$\\
\ \ $\forall$ p h t L. list\_exp p (h::t) L = \\ \ \ exp\_dist p (HD L) h $\wedge$ list\_exp p t (TL L)
}
\end{flushleft}
\noindent The \texttt{list\_exp} function accepts a list of failure rates, a list of random variables $L$ and
a probability space $p$. It guarantees that all elements of the list $L$ are exponentially distributed with corresponding failure rates given in the other list within the probability space $p$. For this purpose, it utilizes the list functions \texttt{HD} and \texttt{TL}, which return the \emph{head} and \emph{tail} of a list, respectively.
Next, we model the pipeline, shown in Figure 2, as a series RBD as follows:
\begin{flushleft}
\texttt{\bf{Definition 9: }} \label{Reliab_series_def}
\emph{Reliability of Series Pipeline System } \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p L . pipeline p L = rel\_series p L
}
\end{flushleft}
\noindent Now, we can use Definition 8 to guarantee that the random variable list argument of the function \texttt{pipeline} contains exponential random variables only and thus verify the following simplified expression for the pipeline reliability.
\begin{flushleft}
\texttt{\bf{Theorem 6: }} \label{pipeline system}
\emph{Series Pipeline System } \\
\vspace{1pt} \texttt{$\vdash$ $\forall$ p L x C. prob\_space p $\wedge$ (events p = POW (p\_space p)) $\wedge$ \\
\ \ \ \ \ 0 $\leq$ x $\wedge$ 2 $\leq$ LENGTH (rel\_event\_list p L x) $\wedge$ \\
\ \ \ \ \ mutual\_indep p (rel\_event\_list p L x) $\wedge$ \\
\ \ \ \ \ \ list\_exp p C L $\wedge$ (LENGTH C = LENGTH L) $\Rightarrow$ \\
\ \ \ \ \ \ (pipeline p (rel\_event\_list p L x) = exp (-list\_sum C * x))
}
\end{flushleft}
\noindent The first five assumptions are the same as the ones used in Theorem 5. The sixth assumption \texttt{list\_exp p C L} ensures that the list of random variable $L$ contains all exponential random variables with corresponding failure rates given in list $C$. The next assumptions guarantees that the lengths of the two lists $L$ and $C$ are the same. While the conclusion of Theorem 6 represents desired reliability relationship for the given pipeline model. Here the function \texttt{list\_sum} recursively adds the elements of its list argument and is used to add the failure rates of all exponentially distributed random variables, which are in turn used to model the individual segments of the series RBD of the pipeline. The proof of Theorem 6 is based on Theorem 4 and some properties of the exponential function \texttt{exp}. The reasoning was very straightforward (about 100 lines of HOL code)
compared to the reasoning for the verification of Theorem 4 \cite{waqar_ftscs_13}, which involved probability-theoretic
guidance. This fact illustrates the usefulness of our core formalization for conducting the reliability
analysis of pipelines.
The distinguishing features of this formally verified result include its generic nature, i.e., all the variables are universally quantified and thus can be specialized to obtain the reliability of the given pipeline for any given parameters, and its guaranteed correctness due to the involvement of a sound theorem prover in its verification, which ensures that all the required assumptions for the validity of the result are accompanying the theorem. Another point worth mentioning is that the individual failure rates of the pipeline segments can be easily provided to the above theorem in the form of a list, i.e., $C$. The above mentioned benefits are not shared by any other computer based reliability analysis approach for oil and gas pipelines and thus clearly indicate the usefulness of the proposed approach.
\section{Conclusions}
\label{sec_6}
Probabilistic analysis techniques have been widely utilized during the last two decades to assess the reliability of oil and gas pipelines. However, all of these probability theoretic approaches have been utilized using informal system analysis methods, like simulation or paper-and-pencil based analytical methods, and thus do not ensure accurate results. The precision of results is very important in the area of oil and gas pipeline condition assessment since even minor flaws in the analysis could result in the loss of human lives or heavy damages to the environment. In order to achieve this goal and overcome the inaccuracy limitation of the traditional probabilistic analysis techniques, we propose to build upon our proposed formalization of RBDs to formally reason about the reliability of oil and gas pipelines using higher-order-logic theorem proving.
Building upon the results presented in this paper, the formalization of other commonly used RBDs, including parallel, series-parallel and parallel-series, and the Weibull random variable is underway. These advanced concepts are widely used in the reliability analysis of pipelines. However, their formalization requires some advanced properties of probability theory. For example, for formalizing the reliability block diagrams of the series-parallel and parallel-series structures, we need to first formally verify the principle of inclusion exclusion \cite{Trivedi_02}. We also plan to formalize the underlying theories to reason about more realistic series pipeline systems, such as multi-state variable piping systems, where each subcomponent of the pipeline system consists of many irreversible states from good to worst. We also plan to investigate artificial neural networks in conjunction with theorem proving to develop a hybrid semi-automatic pipeline reliability analysis framework. Besides the pipeline reliability analysis, the formalized reliability theory foundation presented in this paper, may be used for the reliability analysis of a number of other applications, including hardware and software systems.
\section*{Acknowledgments}
This publication was made possible by NPRP grant \# [5 - 813 - 1 134] from the Qatar National Research Fund (a member of Qatar Foundation). The statements made herein are solely the responsibility of the author[s].
\bibliographystyle{splncs}
|
\section{Introduction}
In this article we give a new formula for the solution
of the Dirichlet problem for the homogeneous real and complex Monge--Amp\`ere equation
(HRMA/HCMA) on the product of either a convex domain and Euclidean space in the real
case, or a tube domain and a K\"ahler manifold in the complex case.
This is partly inspired by Kiselman's minimum
principle \cite{Kiselman} and recent work of Ross--Witt-Nystr\"om \cite{RWN}.
Our formula involves the convex- or plurisubharmonic-envelope of a family of functions on the Euclidean
space or the manifold, and
the Legendre transform on the convex domain. Consequently, one could hope to develop
the existence and regularity theory
for both weak and strong solutions using such a formula. In this article and in its sequels we develop this approach.
The regularity properties of the Legendre transform
are classical. Thus, one is naturally led to study the regularity
properties of the convex- or plurisubharmonic-envelope of a family of functions.
In the case of single function with bounded second derivatives,
the regularity of such envelopes was studied by
Benoist--Hiriart-Urruty, Griewank--Rabier, and Kirchheim--Kristensen,
\cite{BenoistHiriartUrruty,GriewankRabier,KirchheimKristensen}
(see also \cite[\S X.1.5]{HL2}) in the convex case, and by
Berman and Berman--Demailly \cite{Be,BD}
in the plurisubharmonic (psh) setting.
The convex- or psh-envelope of a family of functions is, by definition, the corresponding envelope of
the (pointwise) infimum of that family. However,
already when the family consists of two functions, their minimum
is only Lipschitz. Thus, our second goal here is to extend the aforementioned regularity results to such a setting.
The approach we take to achieve this goal is to study,
more generally, the analogous {\it subharmonic envelope}.
The subharmonic
envelope of a `rooftop obstacle' of the form $\min\{b_0,\ldots,b_k\}$
is, of course, just the solution of the free-boundary
problem for the Laplace equation associated to this obstacle.
Our first regularity result concerning envelopes is that
the solution to the free-boundary problem for the Laplace equation associated to a
such a rooftop obstacle, for functions $b_i$ with finite $C^2$ norm, also has finite
$C^2$ norm, along with an a priori estimate.
Aside from basic regularity tools from the
theory of free-boundary problems associated to the Laplacian,
this involves a new a priori estimate on the size of a ball
that lies between the rooftop and the envelope.
This result stands in contrast to the results of
Petrosyan--To \cite{PetrosyanTo} that show that
the subharmonic-envelope is $C^{1,\frac12}$ and no better
for more general rootop obstacles.
Since the subharmonic-envelope always lies above both the convex- and the
psh-envelope this allows us to establish the regularity of
the latter envelopes as well.
An important application
that makes an essential use of our results is the determination of the
Mabuchi metric completion of the space of K\"ahler potentials,
that is treated in a sequel \cite{D2014}.
\section{Main results}
Our first result concerns a new formula for the solution of the HRMA/HCMA on certain
product spaces. While the real result resembles the complex result, it is
not implied by it directly. Thus, we split the exposition into two
(\S\ref{HCMRSubSec}--\S\ref{HRMRSubSec}).
Our second result concerns the regularity of subharmonic-, convex-, and psh-envelopes
of a `rooftop' obstacle. The regularity of the latter two (\S\ref{CVXPSHRegSubSec})
is a consequence of that of the former (\S\ref{FBPSubSec}).
In passing, we also establish the Lipschitz regularity of the psh-envelope
associated to a general Lipschitz obstacle.
\subsection{A formula for the solution of the HCMA}
\label{HCMRSubSec}
Suppose $(M,\omega)$ is a compact, closed and connected K\"ahler manifold and
let $K \subset \Bbb R^k$
be a bounded convex open set. Denote by $K^{\Bbb C} = K \times \Bbb R^k$
(considered as a subset of $\Bbb C^k$)
the convex tube with base $K$.
Let $\pi_2: K^\mathbb{C} \times M \to M$ denote the natural projection,
and denote by
$\mathrm{PSH}(K^\mathbb{C}\times M,\pi_2^\star\omega)$ the set of
$\pi_2^\star\omega$-plurisubharmonic functions.
We seek bounded $\mathbb{R}^k-$invariant solutions $\varphi \in L^\infty \cap \mathrm{PSH}(K^\mathbb{C}\times M,\pi_2^\star\omega)$ of the problem
\begin{equation}\label{HCMAEq}
(\pi_2^\star\omega+\sqrt{-1}\del\dbar \varphi)^{n+k}=0 \h{\ in $K^\mathbb{C}\times M$},
\quad
\varphi=v \h{\ on $\partial K^\mathbb{C}\times M$},
\end{equation}
where the boundary data $v$ is bounded, $\mathbb{R}^k-$invariant and
$v_s:=v(s,\,\cdot\,)\in \mathrm{PSH}(M,{\omega}), s \in \partial K$.
Some care is needed in defining the sense in which the boundary data is attained
since the functions involved are merely bounded.
In \eqref{HCMAEq}, by ``$\varphi=v \h{\ on $\partial K^\mathbb{C}\times M$}$"
we mean
that for each $z\in M$ the convex function
$\varphi_z:= \varphi(\,\cdot\,,z)$ is continuous up to the boundary of $K$ and satisfies
$\varphi_z|_{\partial K}=v_z$.
This choice of boundary condition implies that
$v_z \in C^0(\partial K)$,
and we will assume this condition on the boundary data throughout.
The study of the Dirichlet problem for the complex Monge--Amp\`ere equation goes back to Bremermann and Bedford--Taylor \cite{Bremermann,BT}. In particular, their results show that one should look for the solution as an upper envelope:
\begin{equation}\label{uUpperEnvEq}
\varphi:=\sup\{w\in L^\infty\cap\mathrm{PSH}(K^\mathbb{C}\times M,\pi_2^\star\omega): \ w \textup{ is }\mathbb{R}^k\textup{-invariant and }
w|_{\partial K^\mathbb{C}\times M}\le v\},
\end{equation}
generalizing the Perron method for the Laplace equation, where $w|_{\partial K^\mathbb{C}\times M}\le v$ means that $\limsup_{s \to s_0} w(s,z) \leq v(s_0,z)$ for all $z \in M, \ s_0 \in \partial K$.
It is not immediate, but as we will prove in Theorem \ref{envdual}, $\varphi$ is upper semi-continuous on $K^\mathbb{C} \times M$.
Assuming this for the moment, by Bedford--Taylor's theory
$\varphi$ solves \eqref{HCMAEq}
(in general,
further conditions are needed on $v$ in order to ensure
that $\varphi|_{\partial K^\mathbb{C}\times M}=v$, as discussed below in Remark \ref{SolutionRemark}).
Our first result gives a different formula for expressing $\varphi$, regardless of whether $\varphi$ assumes $v$ on the boundary.
It involves the psh-envelope operator solely in the $M$ variables,
and the Legendre transform solely in the $K$ variables. The psh-envelope is
the complex analogue of the convexification operator (or double Legendre transform) in the real setting,
and is different than the upper envelope in that, roughly, it involves functions and not boundary
values thereof.
Given a family of upper semi-continuous bounded functions $\{f_a\}_{a\in{\mathcal{A}}}$ parametrized
by a
set ${\mathcal{A}}$, set
$$
P\{f_a\}_{a\in{\mathcal{A}}}:=\sup\{h\in\mathrm{PSH}(M,\omega)\,:\, h(z)\le\inf_{a\in{\mathcal{A}}}\{f_a(z)\}, \,\forall\, z\in M\}.
$$
As each $f_b$ is upper semi-continuous, it follows that the upper semi-continuous regularization satisfies $\textup{usc}(P\{f_a\}_{a\in{\mathcal{A}}}) \leq f_b$, hence by Choquet's lemma $\textup{usc}(P\{f_a\}_{a\in{\mathcal{A}}})$ is a competitor for the supremum, which in turn implies $P\{f_a\}_{a\in{\mathcal{A}}}= \textup{usc}(P\{f_a\}_{a\in{\mathcal{A}}})\in \mathrm{PSH}(M,\omega)$.
Given a function $f=f(s,z)$ on $K\times M$ (that we consider
as a family of functions on $K$ parametrized by $M$), we let
\begin{equation}\label{LegTrans1Eq}
f^\star(\sigma,z)=f^\star(\sigma):=\inf_{s\in K}[f(s,z)-\langle\sigma,s\rangle].
\end{equation}
This is the {\it negative} of the usual Legendre transform solely in the $K$-variables, in particular, it maps convex functions to concave functions, and vice versa. Despite this, we also refer to it sometimes as the partial Legendre transform, and we often omit the dependence
of the function on the $M$ variables in the notation. Here $\langle\,\cdot\,,\,\cdot\,\rangle$
is the pairing between $\mathbb{R}^k$ and its dual. Conversely,
if $g=g(\sigma,z)$ is a
function on $\mathbb{R}^k\times M$ taking values in $[-\infty,\infty)$,
where $\mathbb{R}^k$ is considered as the dual vector space to the copy of $\mathbb{R}^k$
containing $K$, then
\begin{equation}\label{LegTrans2Eq}
g^\star(s,z)=g^\star(s):=\sup_{\sigma\in \mathbb{R}^k}[g(\sigma,z)+\langle\sigma,s\rangle].
\end{equation}
Note that $f^{\star\star}=f$ if and only if $f$ is convex, lower semicontinuous and nowhere
equal to $-\infty$ (we do not allow the constant function $-\infty$),
and otherwise $f^{\star\star}$ is the convexification of $f$, namely,
the largest convex function majorized by $f$
\cite{Mand,Fenchel,Rockafellar}.
\begin{theorem}
\label{envdual}
Assume that $v$ is bounded, $v_s=v(s,\,\cdot\,)\in \mathrm{PSH}(M,{\omega})$ and $v_z = v(\cdot,z) \in C^0(\partial K)$ for all $s\in\partial K$, $z \in M$. Then $\varphi$ as defined in \eqref{uUpperEnvEq} is upper semi-continuous and
\begin{equation}\label{main_identity}
\varphi(h,z)=(P\{v_s-\langle s,\sigma\rangle\}_{s\in\partial K})^\star(h,z)
=\sup_{\sigma\in\mathbb{R}^k}[P\{v_s-\langle s,\sigma\rangle\}_{s\in\partial K}(z)+\langle h,\sigma\rangle], \ h\in K,z \in M.
\end{equation}
Equivalently, $\varphi^\star(\sigma,z)=
\inf_{s\in K}[\varphi(s,z)-\langle\sigma,s\rangle]=P\{v_s-\langle s,\sigma\rangle\}_{s\in\partial K}(z)$.
\end{theorem}
To avoid confusion, we emphasize that
$P\{v_s-\langle s,\sigma\rangle\}_{s\in\partial K}$ is {\it not}
the upper envelope of a family of linear function in $\sigma$ (that
would imply it is convex, which is essentially never true).
Instead, the psh-envelope of this family is a global operation
done for each $\sigma$ separately, and it is in fact concave in $\sigma$, as
the second statement in the theorem shows.
We pause to note an important corollary of this result for the special case $K=[0,1]$,
where $K^\mathbb{C}$ is now the strip $S:=[0,1]\times\mathbb{R}$, and
\eqref{HCMAEq} becomes
\begin{equation}\label{MabuchiEq}
(\pi_2^\star\omega+\sqrt{-1}\del\dbar \varphi)^{n+1}=0, \quad \varphi|_{\{i\}\times\mathbb{R}}=v_i,\; i=0,1,
\end{equation}
\begin{corollary}
\label{MabuchiCor}
Bedford--Taylor solutions of \eqref{MabuchiEq}
with bounded endpoints $v_0,v_1\in L^\infty(M)$, are given by
\begin{equation}\label{Mabuchimain_identity}
\varphi(s,z)=P(v_0,v_1-\sigma)^\star_s(z)=\sup_{\sigma\in\mathbb{R}}[P(v_0,v_1-\sigma)(z)+s\sigma], \ s \in [0,1], \ z\in M.
\end{equation}
\end{corollary}
According to Mabuchi, Semmes, and Donaldson \cite{M,S,D1}, sufficiently regular solutions of \eqref{MabuchiEq}
are geodesics in the Mabuchi metric on the space of K\"ahler potentials with respect to $\omega$.
Thus, Corollary \eqref{MabuchiCor}
implies that Mabuchi's geometry is essentially determined
by the understanding of upper envelopes of the form
$P({v_0},v_1 - \tau)$, for all $v_0,v_1\in\mathrm{PSH}(M,\omega)\cap L^\infty(M)$
and for all $\tau \in \Bbb R$. We refer to the sequel \cite{D2014} for
applications of Corollary \ref{MabuchiCor} in this direction, in particular, determining the metric completion of the Mabuchi metric.
\begin{figure}
\begin{center}
\includegraphics[trim=0cm 24cm 0cm 0cm,clip=true,width=4.6in]{Pv_0v_1}
\caption{The barriers $v_0,v_1$ and the envelope $P(v_0,v_1)$}
\label{figP(v0v1)}
\end{center}
\end{figure}
\subsection{A formula for the solution of the HRMA}
\label{HRMRSubSec}
Theorem \ref{envdual} has a convex analogue in the setting of the HRMA.
The result does not follow directly from the seemingly harder result
for the HCMA.
For concreteness, we only state the analogue of Corollary \ref{MabuchiCor}
in this setting, that arises in the setting of the Mabuchi metric on
a toric manifold $M$.
The reader is referred to \cite[\S2]{RZII} for the
relevant background concerning the HRMA and toric geometry.
For $z$ belonging to the open orbit of the complex torus $(\mathbb{C}^n)^\star$
(that is dense in $M$), set $x=\hbox{\rm Re}\log z\in \mathbb{R}^n$. On the open orbit,
$\omega=\sqrt{-1}\del\dbar\psi_\omega$ with $\psi_\omega$ $(S^1)^n$-invariant, thus consider
$\psi_\omega$ as a function on $\mathbb{R}^n$.
Then, the HCMA \eqref{HCMAEq} reduces to the HRMA,
\begin{equation}\label{HRMAEq}
\h{MA}\psi(s,x)=0, \quad \h{on } [0,1]\times\mathbb{R}^n,
\qquad
\psi_i(x)\equiv\psi(i,x)=\psi_\omega(x)+v_i(e^x), \quad i\in\{0,1\}.
\end{equation}
Here, $\h{MA}$ is the unique continuous extension of the
operator $f\mapsto d\frac{\partial f}{\partial x_1}\wedge\cdots \wedge d\frac{\partial f}{\partial x_n}$
from $C^2(\mathbb{R}^n)$ to the cone of convex functions on $\mathbb{R}^n$.
The following is a convex version of Corollary \ref{MabuchiCor}.
\begin{proposition}\label{HRMAProp}
The solution of \eqref{HRMAEq} with convex endpoints $\psi_0,\psi_1$ is
given by
\begin{equation}\label{toricmain_identity}
\psi=(\min\{\psi_0,\psi_1-\sigma\}^{\star\star})^\star
=\sup_{\sigma\in\mathbb{R}}[\min\{\psi_0,\psi_1-\sigma\}^{\star\star}+s\sigma].
\end{equation}
\end{proposition}
Here the first (innermost) two Legendre transforms are in the $x$ variables,
while the third (outermost) negative Legendre transform is in the $\sigma$
variable.
Note that, strictly speaking, this result is not a consequence
of Corollary \ref{MabuchiCor}, since it involves the {\it potentially larger}
convex envelope (the supremum is taken over convex functions that might not
come from toric potentials) and not the psh-envelope;
rather,
Proposition \ref{HRMAProp}
implies Corollary \ref{MabuchiCor} (in this symmetric setting)
since it shows
that the psh-envelope in this setting is attained at a `toric' convex function.
This formula also has an interpretation in terms of Hamilton--Jacobi equations, in the spirit of \cite{RZ}, that we discuss elsewhere.
\subsection{Regularity of rooftop subharmonic-envelopes}
\label{FBPSubSec}
The following result plays a crucial role in our
proof of the regularity of convex- and psh-envelopes of rooftop obstacles.
It is of independent interest to the study of regularity of solutions
to the free-boundary problem for rooftop obstacles for the Laplacian.
The solution of the aforementioned free-boundary value problem
is, in fact, the subharmonic-envelope of rooftop obstacles. This is a purely local
result, and is stated on the open unit ball $B_1$ in $\mathbb{R}^n$
(we let $B_R(x_0)$ denote the ball of radius $R$ centered at $x_0\in\mathbb{R}^n$;
when $x_0=0$ we write $B_R=B_R(0)$). Denote by $\mathrm{SH}(B_1)$
the set of subharmonic functions on $B_1$.
\begin{theorem} \label{P_reg_prop}
Let $b_0,b_1 \in C^{1,1}(B_1)$, and
let
\begin{equation}\label{bminEq}
b_{\env}:=\sup\{ f\in\mathrm{SH}(B_1)\,:\, f\le\min\{b_0,b_1\}\}.
\end{equation}
Then, there exists a constant $C=C(n,\| b_0\|_{C^{2}(B_1)},\| b_1\|_{C^{2}(B_1)})$ such that
$$
\| b_{\env}\|_{C^{2}(B_{1/8})} \leq C.
$$
\end{theorem}
\subsection{Regularity of the convex-envelope or psh-envelope of a family of functions}
\label{CVXPSHRegSubSec}
Given an upper semi-continuous family $\{ f_a\}_{a \in \mathcal A}$
with additional regularity properties, one would like to study how
much regularity is preserved by the envelope
$P\{ f_a\}_{a \in \mathcal A}$.
Motivated by Corollary \ref{MabuchiCor} and Proposition \ref{HRMAProp}, we are led to study
the regularity of upper envelopes of the type $P(v_0,v_1)$.
Here, we concentrate on the case when the
{\it barriers} (sometimes also called {\it obstacles})
$v_0$ and $v_1$ are rather regular.
The sequel \cite{D2014} treats the case when $v_0$ or $v_1$
are rather irregular in the psh setting. Already in the case of smooth convex functions,
the convexification is not $C^2$ in general. Thus, the following
results gives conditions that guarantee essentially optimal regularity.
A novelty of our approach, perhaps, is that both the convex- and the psh-envelopes are handled simultaneously.
To state the results, we define the Banach space
\begin{equation}\label{CooEq}
C^{1\bar1}(M):=\{f\in L^\infty(M)\,:\, \Delta_\omega f\in L^\infty(M)\},
\end{equation}
with associated Banach norm
\begin{equation}\label{CooNormEq}
||f||_{C^{1\bar1}}:=||f||_{L^\infty(M)}+||\Delta_\omega f||_{L^\infty(M)}.
\end{equation}
If $Cf\in\mathrm{PSH}(M,\omega)$ for some $C>0$, then $f\in C^{1\bar1}(M)$ if and only if $\sqrt{-1}\del\dbar f$
is a current with bounded coefficients.
We also define, as usual, $C^{1,1}(M)$ to be the
Banach space of functions on $M$ with finite $C^2(M)$ norm.
One has $C^2(M)\subset C^{1,1}(M)\subset C^{1\bar 1}(M)$.
\begin{theorem}
\label{PRegThm}
One has the following estimates:\hfill\break
(i)
$\|P(v)\|_{C^{1}}\le C(M,\omega,\| v\|_{C^{1}})$.
\hfill\break
(ii)
$\|P(v_0,v_1)\|_{C^{1\bar1}}
\leq C(M,\omega,\| v_0\|_{C^{1\bar1}}
,\| v_1\|_{C^{1\bar1}}).
$
\hfill\break
(iii)
Suppose $[\omega_0] \in H^2(M,\Bbb Z)$. Then,
$\|P(v_0,v_1)\|_{C^{2}} \leq C(M,\omega,\| v_0\|_{C^{2}},\| v_1\|_{C^{2}}).$
\end{theorem}
Our convention here and below is that the constants $C$ on the right
hand side of the estimates just stated may equal to $\infty$
only if the corresponding norms of $v$ or $v_i$ are infinite.
An analogous result can be stated for convex rooftop envelopes. For simplicity, we only state
a representative result in the toric setting of Proposition \ref{HRMAProp}.
\begin{corollary}
\label{PRealRegThm}
Let $\psi_0,\psi_1$
be as in Proposition \ref{HRMAProp}. Then,
$$
\|\min\{\psi_0,\psi_1\}^{\star\star}\|_{C^{2}} \leq C(M,\omega,\| \psi_0\|_{C^{2}},\| \psi_1\|_{C^{2}}).
$$
\end{corollary}
By repeated application of the formula $P(v_0,v_1,\ldots,v_k)=P(v_0,P(v_1,\ldots,v_k))$,
the results just stated hold also for
envelopes of the type $P(v_0,\ldots,v_k)$.
In general, the convex- or subharonic-envelope of a Lipschitz function
will be no better than Lipschitz, as shown by Kirchheim--Kristensen \cite{KirchheimKristensen},
and by Caffarelli \cite[Theorem 2]{Caf}, respectively.
Theorem \ref{PRegThm} (i) is the analogous fact for psh-envelopes.
The psh-envelope
of a family of functions, e.g., $P(v_0,v_1)$ is
of course the psh-envelope of the single function $\min\{v_0,v_1\}$
that is in general only Lipschitz. Thus, the point of Theorem
\ref{PRegThm} (ii)--(iii) is that for special Lipschitz functions of
the form $\min\{v_0,\ldots,v_k\}$ that we refer to as {\it
rooftop functions} (see Figure \ref{figP(v0v1)}) the psh-envelope
has a regularizing effect, roughly gaining a derivative.
The proof of Theorem \ref{PRegThm} uses
basic techniques from the theory of free boundary problems for the Laplacian,
together with results of Berman \cite{Be} and Berman-Demailly \cite{BD}
on upper envelopes of psh functions. Part (i) is, in fact, a simple
consequence of the Lipschitz estimate of \Blocki\ \cite{BlockiGrad}
in conjunction with the ``zero temprature" approximation procedure of Berman
\cite{BeApprox}. The bulk of the proof is thus devoted to parts (ii)--(iii).
The key step is to show that there exists a $C^{1,1}$ function $b$
(a `barrier')
{\it along with an a priori estimate} depending only on the respective
norms of the $v_i$, such that $b$ lies below $\min\{v_0,v_1\}$ but
above $P(v_0,v_1)$. The barrier we construct is actually obtained
by first constructing local {\it subharmonic-envelopes}
of $v_0$ and $v_1$ on coordinate charts. This construction
is mostly based on well-known techniques from the study of the
free boundary Laplace equation, see, e.g.,
\cite{Caf,CafSalsa,PetrosyanSU},
but with one essential new ingredient, that we now describe.
For a general rooftop obstacle (that is, not necessarily of the form
$\min\{v_0,v_1\}$) Petrosyan--To \cite{PetrosyanTo} show that
the subharmonic-envelope is $C^{1,\frac12}$
and no better.
Yet, also in the literature on subharmonic-envelopes we were not able
to find the regularization statement for rooftop obstacles
of the form $\min\{v_0,v_1\}$
although it might very well be known to experts.
Thus,
the main new technical ingredient is the estimate of Proposition
\ref{cushintheorem}
that guarantees that around each point in the set $\{v_0=v_1\}$
there exists a ball of a priori estimable size that stays
away from the {\it contact set}, i.e., the set where
the local subharmonic envelope equals the barrier $\min\{v_0,v_1\}$.
Given this estimate, the standard quadratic growth estimate
carries over to our setting, and one obtains a priori estimates on $b$.
Then, since the subharmonic-envelope necessarily majorizes the psh-envelope,
we get $P(b)=P(v_0,v_1)$, to which one may apply Berman--Demailly's results.
A regularity result of a similar nature has been recently proved by Ross--Witt-Nystrom \cite{RWN1}
in a different setting.
Namely, they study regularity of envelopes of the type $P_{[\phi]}(\psi)= \textup{usc} \big( \sup_{c >0} P(\phi+c,\psi)\big)$,
where $\psi \in C^{1\bar 1}(M)$, $\phi \in \mathrm{PSH}(M,\omega)$ is exponentially H\"older continuous and $M$ is polarized.
Also, upon completing this article, we were informed by Berman that the
technique of \cite{BD} can be extended to prove Theorem \ref{PRegThm}(iii) \cite{Berman-personal}.
Perhaps the novel point in our approach, compared to such an extension, is
that it also gives, in passing, a useful result concerning the obstacle problem
for the Laplacian, and thus proves the regularity of the subharmonic-, convex-, and psh-envelopes, all at once.
\subsection{Applications to regularity of Bremermann upper envelopes}
A combination of Theorem \ref{envdual} and Theorem \ref{PRegThm} (i) gives
fiberwise Lipschitz regularity of the Bremermann upper envelope $\varphi$ \eqref{uUpperEnvEq}
associated to fiberwise Lipschitz boundary data. This provides an instance when one can draw conclusions about the regularity of $\varphi$ by studying first the regularity of its partial Legendre transform.
\begin{corollary}\label{LipschitzHCMACor} In the setting of Theorem \ref{envdual},
the envelope $\varphi$ satisfies
$$
\|\varphi(s,\cdot) \|_{C^1} \leq C(M,\omega,\sup_{s \in \partial K}\|v(s,\cdot)\|_{C^1}),
\quad \h{for any $s \in K$}.
$$
In other words, if the boundary data is fiberwise Lipschitz, so is the envelope,
and with a uniform estimate.
\end{corollary}
The novelty of this result is that it proves regularity of the envelope $\varphi$,
whether or not it solves the HCMA. We are not aware of any such results in the literature.
At the same time, when $\varphi$ does solve the HCMA then other techniques exist,
notably \Blocki's Lipschitz estimate \cite{BlockiGrad}. However, even then our method seems to be new
in that it furnishes fiberwise Lipschitz regularity given the same on the boundary data,
while \Blocki's estimate alone gives full Lipschitz regularity starting from
full (also in the $\partial K$ directions) Lipschitz regular data. Of course, it should
be stressed that we ultimately use \Blocki's estimate in our proof, but we do so
only in the fiberwise directions.
\subsection*{Organization}
Theorem \ref{envdual} and Corollary \ref{MabuchiCor} are proved in \S\ref{TubeSec}.
The convex analogue, Proposition \ref{HRMAProp}, is proved in \S\ref{CVXSubSec}.
Theorem \ref{PRegThm} (i) concerning Lipschitz regularity of the psh-envelope is proved in \S\ref{LipRegPSHSubSec},
where we also prove Corollary \ref{LipschitzHCMACor}.
Theorem \ref{PRegThm} (ii)--(iii) and Corollary \ref{PRealRegThm}, concerning the
regularity of second derivatives of the psh- and convex-envelopes, are proved in \S\ref{RegRoofCVXPSHSubSec}.
Finally, the main regularity result concerning the subharmonic envelope, Theorem \ref{P_reg_prop},
is proved in \S\ref{proofofP_reg_prop}.
\section{The Dirichlet problem on the product of a tube domain and a manifold}
\label{TubeSec}
Suppose that $f(s,z)$ is a convex function on $\mathbb{R}^k_s\times\mathbb{R}^m_z$. Then
$\inf_sf(s,z)$ is either identically $-\infty$, or else a convex function on $\mathbb{R}^m$
\cite[Theorem 5.7; p. 144]{Rockafellar},\cite[Theorem 1.3.1]{Kiselman-notes}.
If we replace ``convex" with ``psh" and $\mathbb{R}$ by $\mathbb{C}$ this is not true in general.
A special situation in which this is true was described by Kiselman. Let us recall a local version of this result \cite{Kiselman}
(cf. \cite[Theorem I.7.5]{De}).
As in \S\ref{HCMRSubSec}, let
$K\subset\mathbb{R}^k$ be a convex set and denote by
$K^\mathbb{C}:=K+\sqrt{-1}\mathbb{R}^k\subset \mathbb{C}^{k}$ the tube domain associated to $K$.
Denote by $s$ a coordinate on $K\subset \mathbb{R}^k$ and by $\tau:=s+\sqrt{-1} t$ a
coordinate on $K^\mathbb{C}\subset \mathbb{C}^k$.
\begin{theorem}\label{KiselmanThm} Let $D \subset \Bbb C^n$ be a domain.
If $v\in\mathrm{PSH}(K^\mathbb{C}\times D)$ is such that
that $v(s+\sqrt{-1} t,z)=v(s,z)$ for all $t\in\mathbb{R}^k$ then
\begin{equation}\label{Kiselman}
v(z)= \inf_{\tau \in K^\mathbb{C}}v(\tau,z)
\end{equation}
is either identically $-\infty$, or else psh on $D$.
\end{theorem}
\def\omega{\omega}
This immediately implies the following global version.
As in \S\ref{HCMRSubSec}, we denote by $(M,\omega)$ a K\"ahler manifold
and by $\pi_2:K^\mathbb{C}\times M\rightarrow M$, $\pi_1:K^\mathbb{C}\times M\rightarrow K^\mathbb{C}$ the natural projections.
\begin{corollary}
\label{KiselmanCor}
Assume that $f\in \mathrm{PSH}(K^\mathbb{C}\times M,\pi_2^\star\omega)$ satisfies
$f(s,z)=f(s+\sqrt{-1} t,z)$ for all $t\in\mathbb{R}^k$.
Then $f^\star(\sigma,z)$, as defined in \eqref{LegTrans1Eq},
satisfies $f^\star(\sigma,\,\cdot\,)\in\mathrm{PSH}(M,\omega)$ for each
$\sigma\in \mathbb{R}^k$.
\end{corollary}
\begin{proof}[Proof of Theorem \ref{envdual}]
We argue that $\varphi$ is upper semi-continuous parallel with the proof of the formula
\begin{equation}\label{main_identity_proof}
\varphi^\star(\sigma,z)=
\inf_{s\in K}[\varphi(s,z)-\langle\sigma,s\rangle]=P\{v_s-\langle s,\sigma\rangle\}_{s\in\partial K}(z), \ \sigma \in \Bbb R^k, \ z \in M.
\end{equation}
To start, observe that both
$\varphi(\,\cdot\,,z)$ and $(\textup{usc}\,\varphi)(\,\cdot\,,z)$ are convex and bounded functions on $K$ for each $z\in M$ (note that $\sup\textup{usc}\,\varphi= \sup\varphi$). Indeed, the former is a supremum of convex functions , where as the latter is the restriction to $K \times \{z\}$ of an $\mathbb{R}^k$-invariant $\omega-$psh function by Choquet's lemma.
Thus, it suffices to prove that
\begin{equation}
\label{LegKeyEq}
\varphi^\star(\sigma,z) = (\textup{usc}\,\varphi)^\star(\sigma,z),
\end{equation}
for all $\sigma \in \mathbb{R}^k$ since then,
by applying another partial Legendre transform it follows that
$\varphi=\textup{usc}\,\varphi$. The proof of \eqref{LegKeyEq} will be implicit in the proof of \eqref{main_identity_proof} below.
Recall that by Bedford-Taylor theory \cite[Theorem 1.22]{Kol} the set
$E = \{ \varphi < \textup{usc}\,\varphi\} \subset K^\mathbb{C} \times M$ has capacity zero, in particular its Lebesgue measure is also zero (meaning that $\int_E dV_{K^\mathbb{C} \times M}=0$ for any smooth volume form $dV_{K^\mathbb{C} \times M}$ on $K^\mathbb{C} \times M$). As both $u$ and $\textup{usc}\,u$ are $\mathbb{R}^k$-invariant, $E$ is
also $\mathbb{R}^k$-invariant with base $B \subset K \times M$ (note that $B$ is \emph{not} a subset of $K$). Clearly, the Lebesgue measure
of $B$ is zero. For $z \in M$ we introduce the sets
$$
B_z = \pi_1(B \cap K \times \{ z\}) \subset K.
$$
It follows that $B_z$ has Lebesgue measure zero for all $z \in M \setminus F$, where $F \subset M$ has Lebesgue measure zero.
Suppose $z \in M \setminus F$, we claim that in fact $B_z$ is empty. This follows, as the continuous convex functions $\varphi(\,\cdot\,,z)$ and
$(\textup{usc}\,\varphi)(\,\cdot\,,z)$ agree on the dense set $K \setminus B_z$, hence they have to agree on all of $K$, hence $B_z$ is empty. This implies that
\begin{equation}\label{uscIndetity}
\varphi^\star(\sigma,z) = (\textup{usc}\,\varphi)^\star(\sigma,z),
\quad \textup{ for all } z \in M \setminus F, \sigma \in \mathbb{R}^k.
\end{equation}
Now, by Corollary \ref{KiselmanCor}, for each $\sigma\in \mathbb{R}^k$,
the function $(\textup{usc } \varphi)^\star(\sigma,\,\cdot\,)$ belongs to $\mathrm{PSH}(M,\omega)$.
Moreover, by definition of $\varphi$ we have
\begin{equation}\label{vpest}
\varphi^\star_\sigma \le \varphi_s-\langle\sigma,s\rangle \textup{ for all }s\in K.
\end{equation}
We can in fact extend this estimate to the boundary of $\partial K$:
\begin{claim}
For all $s\in\partial K$ and $\sigma\in\mathbb{R}^k$,
$\varphi^\star(\sigma,z) \le v_s(z)-\langle\sigma,s\rangle$.
\end{claim}
Indeed, as $v_z \in C(\partial K)$ for all $z \in M$, it follows that $\varphi(p,z) \leq \mathcal P[v_z](p), \ p \in K$, where $\mathcal P[v_z] \in C(\overline K)$ is the harmonic function on $K$ satisfying $\mathcal P[v_z]|_{\partial K}=v_z.$ This implies that $\limsup_{p \to s}\varphi_p(z) \leq v_s(z)$ for all $z \in M, \ s \in \partial K$, hence we can take the $\limsup$ of the right hand side of \eqref{vpest} to conclude the claim.
Thus, by \eqref{uscIndetity} we also have $(\textup{usc}\,\varphi)^\star(\sigma,z)
\le v_s(z)-\langle\sigma,s\rangle$ for $z \in M \setminus F, \ s \in \partial K$. As $F$ has Lebesgue measure zero we claim that this inequality extends to all $z \in M$. This follows from the fact that $(\textup{usc}\,\varphi)^\star_\sigma$ and $ v_s-\langle\sigma,s\rangle$ are $\omega-$psh for fixed $s \in \partial K$, hence by the sub-meanvalue property we can write:
$$(\textup{usc}\,\varphi)^\star_\sigma(z) = \lim_{r \to 0} \oint_{B(z,r)}(\textup{usc}\,\varphi)^\star_\sigma(\xi)dV(\xi) \leq \lim_{r \to 0} \oint_{B(z,r)}(v_s(\xi)-\langle\sigma,s\rangle) dV(\xi) =\varphi_s(z)-\langle\sigma,s\rangle,$$
for all $z \in M$, where $B(z,r)$ is a coordinate ball around $z$ and $dV$ is the standard Euclidean measure in local coordinates.
Thus,
$(\textup{usc}\,\varphi)^\star(\sigma,\cdot)
$ is a competitor in
the definition of
$P\{v_s-\langle s,\sigma\rangle\}_{s\in\partial K}$ concluding that
\begin{equation}\label{LegendreIneq1}
\varphi^\star(\sigma,\,\cdot\,)
\le
(\h{usc}\,\varphi)^\star(\sigma,\,\cdot\,)
\le P\{v_s-\langle\sigma,s\rangle\}_{s \in \partial K}.
\end{equation}
Conversely, let $\chi\in\mathrm{PSH}(M,\omega)$ satisfy $\chi\le
v_a-\langle a,\sigma\rangle$ for each
${a\in\partial K}$.
We claim that $\chi \le \varphi_s-\langle s,\sigma\rangle$
for every $s\in K$. Indeed, by \eqref{uUpperEnvEq},
$$
\varphi_s-\langle s,\sigma\rangle=\sup\{w_s-\langle s,\sigma\rangle\in L^\infty\cap\mathrm{PSH}(K^\mathbb{C}\times M,\pi_2^\star\omega):
(w-\langle s,\sigma\rangle)|_{\partial K^\mathbb{C}}\le v-\langle s,\sigma\rangle\},
$$
so $\chi$ is a competitor in this last supremum, proving the claim.
Now, taking the infimum over all $s\in K$ it follows that
\begin{equation}\label{LegendreIneq2}
\varphi^\star(\sigma,\cdot) \ge P\{v_s-\langle s,\sigma\rangle\}_{s\in\partial K}.
\end{equation}
Putting together \eqref{LegendreIneq1} and \eqref{LegendreIneq2} the identities \eqref{main_identity_proof} and \eqref{LegKeyEq} follow, proving that $u$ is upper semi-continuous.
\end{proof}
\begin{remark}\label{SolutionRemark}
{\rm
To guarantee that $\varphi$ defined by \eqref{uUpperEnvEq} is an actual solution of \eqref{HCMAEq},
one can, e.g., assume that there exists a subsolution, by which we mean
an $\mathbb{R}^k$-invariant $w\in L^\infty\cap\mathrm{PSH}(K^\mathbb{C}\times M,\pi_2^\star\omega)$ satisfying $w|_{\partial K^\mathbb{C}}=v$.
In fact, if such a subsolution exists, then $w_z \leq \varphi_z$, implying that
$\varphi_z|_{\partial K}$ lies above the boundary data. On the other hand,
$\varphi_z \leq \mathcal P[v_z]$,
where $\mathcal P[v_z] \in C(\overline K)$ is the harmonic function on $K$ satisfying
$\mathcal P[v_z]|_{\partial K}=v_z.$ Thus, $\varphi_z|_{\partial K}$ also lies below the boundary data.
In sum, $\varphi|_{\partial K^\mathbb{C}}=v$.
}
\end{remark}
Providing a subsolution is often possible given special properties of $K$ or the boundary data $v$. An instance of this is the situation described in Corollary \ref{MabuchiCor}:
\begin{proof}[Proof of Corollary \ref{MabuchiCor}]
By Theorem \ref{envdual},
all one needs to verify is that $\varphi$, as defined in \eqref{uUpperEnvEq}, satisfies $\varphi|_{\{i\}\times\mathbb{R}}=v_i,\; i=0,1$. Formula \eqref{Mabuchimain_identity} follows then from \eqref{main_identity}. However, by an observation of Berndtsson
\cite{Br} we have that the function $w(s,z) = \max\{ v_0(z) - As,v_1(z) + A(1-s)\} \in PSH(K^\mathbb{C} \times M,\pi_2^*\omega)$ satisfies $\psi|_{\{i\}\times\mathbb{R}}=v_i,\; i=0,1,$ where $A = \max \{ \| v_0\|_{L^\infty},\| v_1\|_{L^\infty}\}.$
Hence, $w$ is a subsolution in the sense of Remark \ref{SolutionRemark}.
\end{proof}
We remark in passing that the general argument to prove upper semicontinuity
given in Theorem \ref{envdual} can be avoided in the special setting of Corollary \ref{MabuchiCor}
(i.e., when $K=[0,1]$). Indeed, by convexity in $s$,
$\varphi(s,z) \leq s v_0(z) + (1-s)v_1(z)$ for all $(s,z)\in[0,1]\times M$,
thus also $\textup{usc}\,\varphi$ satisfies the same inequality.
This last estimate in turn implies that $\textup{usc}\,\varphi$ is a candidate in the supremum defining
$\varphi$, thus $\textup{usc}\,\varphi = \varphi$ (cf. \cite{Br}).
\subsection{A convex version for the HRMA}
\label{CVXSubSec}
In this subsection we prove the a version of Corollary \ref{MabuchiCor}
for the homogeneous real Monge--Amp\`ere (HRMA) equation.
While a proof of Proposition \ref{HRMAProp} and even its generalization to
higher dimensional $K$ can be given along very similar lines to the proof
of Theorem \ref{envdual}, we give below a somewhat different argument.
\begin{proof}[Proof of Proposition \ref{HRMAProp}]
As observed by Semmes \cite{Semmes1988,S}, the HRMA is linearized
by the partial Legendre transform {\it in the $\mathbb{R}^n$ variables}.
Thus, the solution to the HRMA is given by
\begin{equation}\label{DoubleLegEq}
\psi(s,x)=((1-s)\psi_0^\star+s\psi_1^\star)^\star(x),
\end{equation}
where
$\psi_i^\star(y)=\sup_{y\in\mathbb{R}^n}[\langle x,y\rangle-\psi_i(y)]$.
As is well-known, this is equal to the {\it infimal convolution
of $\psi_0$ and $\psi_1$} \cite[Theorem 38.2]{Rockafellar},
\begin{equation}\label{InfConvEq}
\inf_{\{x_0,x_1\in\mathbb{R}^n\,:\, (1-s)x_0+sx_1=x\}}[(1-s)\psi_0(x_0)+s\psi_1(x_1)].
\end{equation}
This also follows directly from the fact that $\psi$ solves the HRMA,
since by \cite{RZ} a solution of the HRMA solves a Hamilton--Jacobi equation,
and \eqref{InfConvEq} is just the Hopf--Lax formula in that setting.
Now, we take the negative Legendre transform of \eqref{DoubleLegEq} in $s$
to obtain,
$$
\begin{aligned}
\psi^\star_x(\sigma)
&=\min_{s\in[0,1]}
\Big[
\inf_{\{x_0,x_1\in\mathbb{R}^n\,:\, (1-s)x_0+sx_1=x\}}[(1-s)\psi_0(x_0)+s\psi_1(x_1)]
-s\sigma
\Big]
\cr
&
=
\min_{s\in[0,1]}\Big[
\inf_{\{x_0,x_1\in\mathbb{R}^n\,:\, (1-s)x_0+sx_1=x\}}[(1-s)\psi_0(x_0)+s(\psi_1(x_1)-\sigma)]
\Big].
\end{aligned}
$$
Now we will show that this last expression is equal to
\begin{equation}\label{CvxHullEq}
\sup\{ v\,:\, v \h{\ is convex on } \mathbb{R}^n \h{\ and $v\le\min\{\psi_0,\psi_1-\sigma\}$}\}=\min\{\psi_0,\psi_1-\sigma\}^{\star\star}.
\end{equation}
Fix $x\in\mathbb{R}^n$, and
let $s\in[0,1]$ and $x_0,x_1\in\mathbb{R}^n$ be
such that $(1-s)x_0+sx_1=x$. Let $v$ be a convex
function satisfying $v\le \min\{\psi_0,\psi_1-\sigma\}$.
Then,
$$
(1-s)\psi_0(x_0)+s(\psi_1(x_1)-\sigma)
\ge
(1-s)v(x_0)+sv(x_1)
\ge v(x),
$$
by
convexity of $v$.
Thus, $\psi^\star_x(\sigma)\ge \min\{\psi_0,\psi_1-\sigma\}^{\star\star}$.
Conversely, the expression \eqref{InfConvEq}
is a convex function jointly in $s$ and $x$ (since it is evidently convex
in $x$ by \eqref{DoubleLegEq} and it solves the HRMA in all variables). By
the minimum principle for convex functions
then $\psi^\star_x(\sigma)$ is convex in $x$.
By the definition of the negative Legendre transform in $s$,
$\psi^\star_x(\sigma)\le \min_{s\in\{0,1\}}[\psi_s(x)-s\sigma]
=\min\{\psi_0(x),\psi_1-\sigma\}$. Thus, $\psi^\star_x(\sigma)$
is a competitor in the left hand side of \eqref{CvxHullEq}.
Hence, $\psi^\star_x(\sigma)\le \min\{\psi_0,\psi_1-\sigma\}^{\star\star}$.
\end{proof}
\section{Regularity of upper envelopes of families
}
The bulk of this section is devoted to the proof of Theorem \ref{PRegThm} (ii)--(iii)
and Corollary \ref{PRealRegThm}
that establish the regularity of psh- and convex-envelopes envlopes
associated to obstacles of the form $\min\{b_0,b_1\}$,
that we refer to as `rooftop' envelopes (see Figure
\ref{figP(v0v1)}). However, we begin by first proving the Lipschitz regularity
of psh-envelopes (Theorem \ref{PRegThm} (i)).
\subsection{Lipschitz regularity of psh-envelopes}
\label{LipRegPSHSubSec}
Let $v \in C^\infty(M)$.
Berman developed the following approach for constructing $P(v)$,
generalizing a related construction for obtaining ``short-time" solutions
to the Ricci continuity method, introduced in \cite{R08}, in turn based on a result of Wu \cite{Wu}
(a new approach to which has been given in \cite[\S9]{JMR}, see \cite[\S6.3]{R14}
for an exposition of these matters).
For $\beta$ positive and sufficiently large one considers the equations
\begin{equation}\label{approx_equation}
(\omega + \sqrt{-1}\partial\bar\partial u_\beta)^n = e^{\beta (u_\beta - v)}\omega^n.
\end{equation}
By the classical work of Aubin and Yau, \eqref{approx_equation} admits a smooth solution
$u_\beta$.
Berman proves that, as $\beta$ tends to infinity, $u_\beta$ converges to $P(v)$ uniformly, and that, moreover,
there is an a priori Laplacian estimate in this setting \cite{BeApprox}.
In this section we observe that, as expected, also an a priori Lipschitz estimate holds,
by directly applying \Blocki's estimate. In other words, we prove Theorem
\ref{PRegThm} (i). The proof will show that the
constant in Theorem \ref{PRegThm} (i) depends on a lower bound
of the bisectional curvature of $(M,\omega)$ and on $||v||_{C^1(M)}$.
We claim no originality in the proof below.
\begin{proof}[Proof of Theorem \ref{PRegThm} (i)]
\def\beta}\def\al{\alpha{\beta}\def\al{\alpha}
It suffices, by a standard approximation procedure, to
assume that $v$ is smooth.
For simplicity of notation, we will often denote
$u_\beta$ by just $u$.
The argument follows \cite[Theorem 1]{BlockiGrad} very closely.
Let $B'$ be some sufficiently large positive constant to be fixed later.
Let $C_0:=\sup_{\beta>2}||u_\beta}\def\al{\alpha||_{C^0}+1$.
Let $\phi: M \to \Bbb R$ be the following function:
$$\phi:= \log|\partial u|_\omega^2 - \gamma (u),$$
where
$\gamma:[-C_0,C_0]\to \Bbb R$ is a
smooth non-decreasing
function to be fixed later.
Since $\| u\|_{C^0} \leq C(M,\| v\|_{C^0}),$ independently of $\beta$
\cite{BeApprox}, $\gamma$ is thus defined on some fixed finite interval.
Suppose $\phi$ attains its maximum at $p \in M$. Let $z=(z_1,\ldots,z_n)$ denote
holomorphic normal coordinates around this point.
Let $g$ denote a local potential for $\omega$ in this chart, i.e., $\sqrt{-1}\partial\bar\partial g =\omega$. Set $h := g + u$.
We can additionally suppose that $\sqrt{-1}\partial\bar\partial u(p)$ is diagonal in our coordinates.
Since all our local calculations will be carried out at the point $p$ we omit the dependence on this point
from the subsequent computations.
Let
$$
\alpha:=|\partial u|_\omega^2.
$$
Thus,
\begin{equation}\label{first_order_id}
0 = \frac{\partial}{\partial z_j}\phi = \phi_j = \frac{\alpha_j}{\alpha} - \gamma'(u) u_j, \quad j = 1,\ldots,n,
\end{equation}
and so (omitting from now and on symbols for summation that can be understood from the context),
\begin{flalign}
0
\geq
\Delta_{\omega_u} \phi
=
\frac{\phi_{k\bar k}}{h_{k\bar k}}
=
\frac{1}{h_{k\bar k}}\Big(\frac{\alpha_{k\bar k}}{\alpha} - \frac{|\alpha_k|^2}{\alpha^2}
-
\gamma' u_{k\bar k} - \gamma''\al\Big)
=
\frac{1}{h_{k\bar k}}\Big(\frac{\alpha_{k\bar k}}{\alpha} - \gamma' u_{k\bar k} - (\gamma''+\gamma'^2) \al\Big),
\end{flalign}
The next formula holds for each fixed $k=1,\ldots,n$ (no summation)
\begin{flalign}
\alpha_{k\bar k} = 2 \hbox{\rm Re} \ u_{jk\bar k} u_{\bar j} + |u_{jk}|^2 + |u_{j\bar k}|^2 - u_j g_{j\bar l k \bar k}u_{\bar l}
\geq
2 \hbox{\rm Re} \ u_{jk\bar k} u_{\bar j} + |u_{jk}|^2 - B\al,
\end{flalign}
whenever $-B$ is a lower bound for the bisectional curvature of $\omega$.
Using this, the identity $1 + u_{k\bar k}=h_{k\bar k}$, and fact that $g_{jk\bar k}=0$,
and summing over $k$ we have,
\begin{flalign*}
0 \geq &\frac{1}{h_{k\bar k}}\Big(\frac{2 \hbox{\rm Re} \ h_{jk\bar k} u_{\bar j} + |u_{jk}|^2- B\al}{\alpha} + \gamma' -\gamma'h_{k\bar k} - (\gamma''+\gamma'^2) \al \Big),
\end{flalign*}
Multiplying across with $\alpha$,
\begin{flalign}
0 \geq
\frac{1}{h_{k\bar k}}\Big(2 \hbox{\rm Re} \ h_{jk\bar k} u_{\bar j} + |u_{jk}|^2
+
\al\big[\gamma'-B -\gamma'h_{k\bar k} - (\gamma''+\gamma'^2) \al\big] \Big)
\label{int_est}
\end{flalign}
By \Blocki's trick \cite[(1.15)]{BlockiGrad}, we also have the following estimate:
\begin{equation}\label{blocki_trick}
\frac{|u_{jk}|^2}{h_{k\bar k}} \geq \al \Big( \gamma'^2 \frac{|u_k|^2}{h_{k\bar k}} -2\gamma' \Big) -2.
\end{equation}
The computations so far are general and taken from \cite{BlockiGrad}. We now bring
the equation we are interested in,
$\log\frac{\det[h_{j\bar l}]}{\det[g_{j\bar l}]} = \beta (u - v)$,
into the picture.
Differentiating this equation at $p$ yields
$\frac{h_{jk\bar k}}{h_{k\bar k}}= \beta (u_j - v_j).$
Thus,
\begin{equation}\label{3deriv_est}
2\hbox{\rm Re} \ \frac{h_{jk\bar k}}{h_{k\bar k}}u_{\bar j}= 2\hbox{\rm Re} \ \beta (u_j - v_j)u_{\bar j} = 2\beta |u_j|^2 - 2\beta \hbox{\rm Re} \ v_j u_{\bar j}
\geq 2\beta |u_j|^2 - 2\beta|v_j|^2.
\end{equation}
Putting \eqref{3deriv_est} and \eqref{blocki_trick} into \eqref{int_est} we obtain:
\begin{flalign}
0 \geq& 2\beta \al - 2\beta |v_j|^2 + \al \Big( \gamma'^2 \frac{|u_k|^2}{h_{k\bar k}} -2\gamma' \Big)
-2 + \frac{\al}{h_{k\bar k}}\Big[\gamma' -B-\gamma'h_{k\bar k} - (\gamma''+\gamma'^2) \al\Big] \nonumber \\
=& 2(\beta-\gamma') \al - 2\beta |v_j|^2 -2 -n\gamma'+ \frac{\al}{h_{k\bar k}}(\gamma' -B - \gamma'' \al)
\end{flalign}
Our wish is to get rid of the last term in the right.
For this reason, we choose $\gamma:[-C_0,C_0] \to \Bbb R$ to be $\gamma(t)=-t^2/2 + (C_0 + B)t$. Then $2C_0 + B > \gamma'>B$, $\gamma'' <0$. With this choice, in our last estimate the rightmost term becomes positive, so we can write:
\begin{flalign}
0 \geq (2\beta-2C_0 - B) \al - 2\beta |v_j|^2 -2 -n(2C_0+B).
\end{flalign}
This gives
\begin{flalign}
\al \leq \frac{ 2\beta |v_j|^2 -2 -n(2C_0+B)}{2\beta-2C_0 - B},
\end{flalign}
concluding the proof of Theorem \ref{PRegThm} (i), since the constant on the right hand side can
be majorized independently of $\beta$.
\end{proof}
We turn to prove a corollary of this estimate and the formula for
the Bremermann upper envelope
$\varphi$ introduced in \eqref{uUpperEnvEq}
(Theorem \ref{envdual}), namely, the
Lipschitz regularity of $\varphi$.
\begin{proof}[Proof of Corollary \ref{LipschitzHCMACor}] It follows from the definition of $\varphi$ that $\| \varphi\|_{C^0} \leq \| v \|_{C^0}$. To finish the proof we need to prove that
\begin{equation}\label{GradEst}
|\varphi(s,\cdot)|_{C^{0,1}} \leq C(M,\omega,\sup_{s \in \partial K}\|v(s,\cdot)\|_{C^1}), \ s \in \partial K.
\end{equation}
Fix $h \in K$. By \eqref{main_identity} we have
\begin{equation}
\varphi(h,z)=(P\{v_s-\langle s,\sigma\rangle\}_{s\in\partial K})^\star(h,z)
=\sup_{\sigma\in\mathbb{R}^k}[P\{v_s-\langle s-h,\sigma\rangle\}_{s\in\partial K}(z)], \ z \in M.
\end{equation}
Fix $\sigma \in \Bbb R^k$. As $K$ is bounded,
by Lemma \ref{LipFamilyLemma} below, $\phi_\sigma := \inf_{s \in \partial K}(v_s-\langle s-h,\sigma\rangle) \in C^{0,1}(X)$,
with $|\phi_\sigma |_{C^{0,1}}
\leq C(\sup_{s \in \partial K}|v(s,\cdot)|_{C^{0,1}})$.
By Theorem \ref{PRegThm} (i) it follows that
$$| P(\phi_\sigma)|_{C^{0,1}} \leq C(|\phi_\sigma|_{C^{0,1}}) \leq C(\sup_{s \in \partial K}|v(s,\cdot)|_{C^{0,1}}).$$
As $\varphi(h,\cdot)=\sup_{\sigma \in \Bbb R^k}P(\phi_\sigma)$, \eqref{GradEst} follows from another application of Lemma
\ref{LipFamilyLemma}.
\end{proof}
The next lemma is a consequence of the Arzel\`a-Ascoli compactness theorem.
\begin{lemma} \label{LipFamilyLemma}Suppose $\{ f_\alpha\}_{\alpha \in A} \subset C^{0,1}(M)$ with $\sup_{\alpha \in A} |f_\alpha|_{C^{0,1}} < \infty$. Then:
\noindent
(i) Either $\phi:=\inf_{\alpha \in A} f_\alpha \equiv -\infty$, or $\phi \in C^{0,1}(M)$ with $|\phi|_{C^{0,1}} \leq \sup_{\alpha \in A} |f_\alpha|_{C^{0,1}}$.
\noindent(ii) Either $\psi:=\sup_{\alpha \in A} f_\alpha \equiv \infty$, or $\psi \in C^{0,1}(M)$ with $|\psi|_{C^{0,1}} \leq \sup_{\alpha \in A} |f_\alpha|_{C^{0,1}}$.
\end{lemma}
\subsection{Regularity of rooftop convex- and psh-envelopes }
\label{RegRoofCVXPSHSubSec}
In this subsection we prove Theorem \ref{PRegThm} by using
Theorem \ref{P_reg_prop}. The proof of the latter is postponed
to \S\ref{proofofP_reg_prop}. First, we recall the second order estimates
of Berman \cite[Theorem 1.1, Remark 1.8]{Be}
and Berman--Demailly \cite[Theorem 1.4]{BD}:
\begin{theorem}
\label{BDreg} Let $b \in C^{1,1}(M)$. Then,
(i)
$\|P(b)\|_{C^{1\bar1}} \leq C(\|b\|_{C^{1\bar1}})$, and
(ii) If $[\omega_0] \in H^2(M,\Bbb Z)$, then $\|P(b)\|_{C^{2}} \leq C(\| b\|_{C^{2}}).$
\end{theorem}
\begin{proof}[Proof of Theorem \ref{PRegThm}]
For both parts (i) and (ii) we first assume
$v_0,v_1 \in C^{1,1}(M)$.
Indeed, by an approximation argument, this suffices also for treating part (i).
Take a covering of $M$ by charts, that we assume without loss of generality are unit balls of the form $\{B_1(x_j)\}_{j = 1}^k$ (possible as $M$
is compact), such that the balls $\{B_{1/8}(x_j)\}_{j = 1}^k$ still cover $M$.
Let $\{\rho_j\}_{j=1}^k$ be a partition of unity subordinate to the
latter covering.
Without loss of generality,
we also
assume that in a neighborhood of each
$B_1(x_j)$ the metric
$\omega$ has a K\"ahler potential $w_j \in C^\infty$.
Let $h_j \in \mathrm{SH}({B_1(x_j)})$ be the upper envelope
$$
h_j:=
\sup \{v \in \mathrm{SH}(B_1(x_j))\,:\, v \leq
\min\{v_0|_{B_1(x_j)}+w_j,v_1|_{B_1(x_j)}+w_j\}\}.
$$
If $\varphi$ is an $\omega$-psh function then
$w_j+\varphi\in\mathrm{PSH}(B_1(x_j))$ and therefore
$w_j+\varphi\in\mathrm{SH}(B_1(x_j))$. Thus,
$P(v_0,v_1)|_{B_1(x_j)} \leq h_j-w_j \leq \min \{ v_0,v_1\}|_{B_1(x_j)}$, and
by Theorem \ref{P_reg_prop},
\begin{equation}\label{hjEq}
\| h_j\|_{C^{2}}
\leq
C(\| w_j\|_{C^{2}},\| v_0\|_{C^{2}},\| v_1\|_{C^{2}}).
\end{equation}
Set
$b:= \sum_{j=1}^k \rho_j (h_j - w_j)$.
Then
\begin{equation}\label{bC2EstEq}
\| b\|_{C^{2}}
\leq
C(\{\| w_j\|_{C^{2}}\}_{j=1}^k,\{\| \rho_j\|_{C^{2}}\}_{j=1}^k,\| v_0\|_{C^{2}},\| v_1\|_{C^{2}})\le C(M,\omega,\| v_0\|_{C^{2}},\| v_1\|_{C^{2}}).
\end{equation}
It follows that $P(v_0,v_1) \leq b \leq \min\{ v_0,v_1\}$ as we noticed above that $P(v_0,v_1)|_{B_1(x_j)} \leq h_j-w_j \leq \min\{ v_0,v_1\}|_{B_1(x_j)}$.
Thus, $P(b)=P(v_0,v_1)$
and so part (i) of the theorem follows from \eqref{bC2EstEq}
and Theorem \ref{BDreg}.
Part (ii) follows as well if we can show that
$$\| b\|_{C^{1\b1}}
\leq
C(\{\| w_j\|_{C^{1\b1}}\}_{j=1}^k,\{\| \rho_j\|_{C^{1\b1}}\}_{j=1}^k,\| v_0\|_{C^{1\b1}},\| v_1\|_{C^{1\b1}})\le C(M,\omega,\| v_0\|_{C^{1\b1}},\| v_1\|_{C^{1\b1}}).
$$
This estimate indeed holds since on the incidence set
$\{h_j-w_j=\min\{v_0|_{B_1(x_j)},v_1|_{B_1(x_j)}\}$ the function
$\Delta h_j$ equals either $\Delta v_0|_{B_1(x_j)}$
or $\Delta v_1|_{B_1(x_j)}$ a.e. with respect to the Lebesgue measure, while $h_j$ is harmonic on the complement of the incidence set.
\end{proof}
Corollary \ref{PRealRegThm} follows from the previous theorem because
by Proposition \ref{HRMAProp} the convex rooftop envelope solves the HRMA,
and hence also the HCMA on the associated toric manifold.
\begin{remark}
{\rm
One can give a different proof of part (ii) of Theorem \ref{PRegThm}
using results on regularity of Mabuchi geodesics
together with Theorem \ref{envdual} and Proposition \ref{family_est}.
Indeed, let $[0,1] \ni t \to a_t \in \mathrm{PSH}(M,\omega) \cap L^\infty(M)$ be the weak geodesic joining $a_0 :=P(v_0)$ with $a_1 :=P(v_1)$.
By Theorem \ref{BDreg} both $P(v_0)$ and $P(v_1)$ have bounded Laplacian. By
Berman--Demailly \cite[Corollary 4.7]{BD} (see He \cite{He} for a different proof)
so does each $a_t$ for each $t \in [0,1]$.
Since $P(v_0,v_1)=P(P(v_0),P(v_1))$, by Theorem
\ref{envdual} we have
$$P(v_0,v_1)=a^*_0.$$
Finally,
$| \Delta_{\omega_0} a^*_0|$ is bounded
by Proposition \ref{family_est} below.
}
\end{remark}
The following estimate is very likely well-known,
although we were not able to find a precise reference.
\begin{proposition}\label{family_est}
Let $\{ v_a\}_{a \in A}$ be a
uniformly locally bounded
family of functions on a domain $D \subset \Bbb C^n$. Suppose that
$| \Delta v_a|\leq B$
for all $a \in A$,
and that $v_{\min}=\inf _{a \in A}v_a$ is psh on $D$. Then,
$| \Delta v_{\min}| \leq B$.
\end{proposition}
One can also assume instead of uniform local boundedness that $v_{\min}$
itself is locally bounded.
\begin{proof} By our assumption $\Delta v_{\min} \geq 0$, hence we only have to prove that $\Delta v_{\min} \leq B$. Our assumptions also imply that the functions $B|z|^2/2n - u_a$
are subharmonic on $D$ for any $a \in \mathcal A$. By the Zygmund-Calderon estimate we also have that the $C^{0,1}$ norm of the functions $B|z|^2/2n - u_a$ is uniformly bounded on any relatively compact open subset of $D$. This implies that $B|z|^2/2n - v_{\min}=\sup_{a \in \mathcal A}(B|z|^2/2n - v_a)$ is locally Lipschitz continuous, hence by Choquet's lemma also subharmonic. This in turn implies that $\Delta v_{\min} \leq B$.
\end{proof}
\subsection{Regularity of rooftop subharmonic-envelopes}
\label{proofofP_reg_prop}
We now prove Theorem \ref{P_reg_prop}. Let us fix some notation. Let
$b_0,b_1 \in
C^{1,1}({B_1})$ with $B_1\subset \Bbb C^n =\Bbb R^{2n}$.
The envelope $b_{\env}$ \eqref{bminEq} is upper semi-continuous hence it is subharmonic by Choquet's lemma.
We call $b_{\env}$ the {\it subharmonic-envelope of the rooftop obstacle
$\min \{ b_0,b_1\}$}.
Let
\begin{equation}\label{b10Eq}
b_{10}:=
b_1 - b_0\in C^{1,1}({B_1}),
\end{equation}
and denote the {\it contact set} (or {\it coincidence set}) by
\begin{equation}\label{LambdaEq}
\Lambda:= \{x\in B_1\,:\, b_{\operatorname{env}}(x) = \min\{b_0,b_1 \}(x)\}.
\end{equation}
We call the complement of $\Lambda$ in $B_1$ the {\it non-coincidence set}.
Our first result assures that whenever $x_0$ is a regular point of the level set $b_{10}^{-1}(0)$,
then $x_0$ is contained in the non-coincidence set,
along with a small open ball of radius uniformly
proportional to $|\nabla b_{10}(x_0)|$.
\begin{proposition}
\label{cushintheorem}
For $b_0,b_1 \in C^{1,1}({B_1})$,
using the notation we introduced, there exists
$C=C(n)/(1 + \|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})$
such that for any $x_0 \in b_{10}^{-1}(0)\cap B_{1/2}$
(recall (\ref{b10Eq}) and (\ref{LambdaEq})),
$$
\Lambda \cap B_{C|\nabla b_{10}(x_0)|}(x_0) = \emptyset.
$$
\end{proposition}
\begin{proof} We fix $x_0 \in b_{10}^{-1}(0)\cap B_{1/2}$.
We will prove that $b_{\env}<\min\{b_0,b_1\}$ on $B_{C|\nabla b_{10}(x_0)|}(x_0)$
by finding a linear function sandwiched between these two functions.
More precisely, the proposition follows from the estimate
\begin{equation}\label{cushin}
b_{\env}(x) < b_0(x_0)-2C |\nabla b_{10}(x_0)|^2 + \langle \nabla b_0(x_0),x-x_0 \rangle
< \min\{b_0,b_1\}(x) , \ x \in B_{C|\nabla b_{10}(x_0)|}(x_0),
\end{equation}
for $C$ as in the statement.
For the second inequality in \eqref{cushin},
observe that for any $x \in B_r(x_0), \ r \leq 1/2$,
\begin{flalign*}
\min\{b_0,b_1\}(x)-&b_0(x_0)
\geq
\min_{i\in\{0,1\}} \langle\nabla b_i(x_0),x-x_0 \rangle
-(\|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})|x-x_0|^2 &\\
\geq&
\langle \nabla b_0(x_0),x-x_0 \rangle
+ \min\{0,\langle \nabla b_{10}(x_0),x-x_0 \rangle\}
- (\|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})|x-x_0|^2 &\\
>&
\langle \nabla b_0(x_0),x-x_0 \rangle - |\nabla b_{10}(x_0)|r
- (\|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})r^2.
\end{flalign*}
Set $r=r'|\nabla b_{10}(x_0)|$. Then,
whenever $r' \leq 1/(1 + \|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})$,
$$
(\|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})r^2
=
(\|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})r'|\nabla b_{10}(x_0)|r
\le r' |\nabla b_{10}(x_0)|^2.
$$
Thus, as desired,
\begin{equation}\label{cushin1}
\min\{b_0,b_1\}(x) > b_0(x_0)-2r' |\nabla b_{10}(x_0)|^2
+ \langle \nabla b_0(x_0),x-x_0\rangle, \ x \in B_{r'|\nabla b_{10}(x_0)|}(x_0).
\end{equation}
Now we turn to the first inequality in \eqref{cushin}.
Fix $r \leq 1/2$.
As before, by Taylor's formula, for $x \in B_r(x_0)$,
\begin{equation}\label{minEst1Eq}
\min\{b_0,b_1\}(x)\leq
b_0(x_0)+ \langle \nabla b_0(x_0),x-x_0 \rangle
+\min\{0,\langle \nabla b_{10}(x_0),x-x_0 \rangle\} +
(\|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})r^2.
\end{equation}
Note that $h$ is subharmonic,
while
$B_r(x_0)\ni x
\mapsto
\langle \nabla b_0(x_0),x-x_0 \rangle
+ (\|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})r^2$ is harmonic.
Combining this with \eqref{minEst1Eq} and the fact that
$h \leq \min\{ b_0,b_1\}$, it follows that
\begin{flalign*}
b_{\env}(x) \leq& b_0(x_0)+ \langle \nabla b_0(x_0),x-x_0 \rangle
+
\int_{\partial B_r(x_0)} P_{ r}(x-x_0,\xi)\min\{0,\langle\nabla b_{10}(x_0),\xi\rangle\}d\sigma(\xi)
\\&+ (\|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})r^2,
\end{flalign*}
where
$P_{r}(x,\xi)=(r^2 - |x|^2)/(2n\omega_{2n} r|x-\xi|^{2n})$
is the Poisson kernel
of the ball $B_r(x)$ which is positive.
For any $x \in B_{r/2}(x_0)$ and $\xi\in \partial B_r(x_0)$,
there is a uniform estimate $|P_{ r}(x-x_0,\xi)|\le C(n)r^{1-2n}$.
Also, $\langle\nabla b_{10}(x_0),\xi\rangle = |\nabla b_{10}(x_0)||\xi|\cos\alpha,$
where $\alpha$ is the angle betwen $\xi$
and $\nabla b_{10}(x_0)$ in the plane they generate.
Now,
since the integrand is negative, one can estimate
it by considering only the quarter sphere
$\partial B^{++}_r(x_0)$ where the angle between $\xi$ and $x-x_0$
is in the range $(-\pi/4,\pi/4)$. Then,
$$
\int_{\partial B_r(x_0)} P_{ r}(x-x_0,\xi)\min\{0,\nabla b_{10}(x_0)\xi\}d\sigma(\xi)
<
\frac1{\sqrt2}
\int_{\partial B^{++}_r(x_0)} P_{ r}(x-x_0,\xi)
|\nabla b_{10}(x_0)|rd\sigma(\xi),
$$
which, in turn, is bounded from above by $-C|\nabla b_{10}(x_0)|r$ for $x \in B(x_0, r/2)$.
Thus,
there exists $C'=C'(n) < 1$ such that
$$
b_{\env}(x)\leq b_0(x_0)+\langle \nabla b_0(x_0),x-x_0 \rangle - C'|\nabla b_{10}(x_0)|r+(\|
b_0\|_{C^{2}}+\|b_1\|_{C^{2}})r^2,
$$
$x \in B(x_0, r/2)$. By taking any
$
\tilde r\leq \frac{C'}{2(1 + \|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})}$
one has that
$\tilde r |\nabla b_{10}(x_0)| < 1$. Thus,
$$
b_{\env}(x)
\leq
b_0(x_0)+
\langle \nabla b_0(x_0),x-x_0 \rangle
-\frac{C'}{2}|\nabla b_{10}(x_0)|^2\tilde r,
\quad
\h{for any $x\in B_{\tilde r|\nabla b_{10}(x_0)|/2}(x_0)$}.
$$
Therefore, for any choice $r'' < {C'\tilde r}/{4}$,
\begin{equation}\label{cushin2}
b_{\env}(x)
<
b_0(x_0)+\langle \nabla b_0(x_0),x-x_0 \rangle - 2|\nabla b_{10}(x_0)|^2r'',
\quad
\h{for any $x \in B_{r''|\nabla v(x_0)|}(x_0)$}.
\end{equation}
The estimate \eqref{cushin} with
$C=\min \{r',r''\}$ follows from \eqref{cushin1} and \eqref{cushin2}.
\end{proof}
Before we consider the interior regularity of $b_{\env}$,
we prove an adaptation to our setting
of the standard quadratic growth lemma (cf. \cite[Lemma 3]{Caf}).
It shows, roughly, that the envelope $b_{\env}$ approximates the obstacle
$\min\{b_0,b_1\}$ at least to second order.
This is quite intuitive in the classical case of an obstacle of class
$C^{1,1}$. In our setting where the obstacle is only Lipschitz,
the proof relies on Proposition \ref{cushintheorem}.
\begin{proposition}
\label{QuadraticProp}
Let $b_0,b_1 \in C^{1,1}({B_1}(x_0))$.
Suppose $x_0\in\Lambda\cap B_{1/4}(x_0)$,
with
$\min \{ b_0,b_1\}(x_0)= b_i(x_0)$ for $i\in\{0,1\}$.
Then, there exists $C = C(\| b_0\|_{C^{2}(B_1)},\| b_1\|_{C^{2}(B_1)})$
such that (recall
(\ref{LambdaEq}) and (\ref{bminEq})),
\begin{equation}\label{QuadGrowthEstEq}
|b_{\env}(x) - b_i(x_0) - \langle \nabla b_i(x_0), x - x_0\rangle| \leq C |x-x_0|^2,
\quad
\h{for all $x \in
B_{1/8}(x_0)$}.
\end{equation}
\end{proposition}
Of course, $b_{\env}$ equals $b_i$ up to infinite order on the interior of
$\Lambda$, so
one could phrase \eqref{QuadGrowthEstEq} as
$$
|b_{\env}(x) - b_{\env}(x_0) - \langle \nabla b_{\env}(x_0), x - x_0\rangle| \leq C |x-x_0|^2,
$$
whenever $x_0$ lies in the interior of $\Lambda$.
However, the key is, of course, that the estimate \eqref{QuadGrowthEstEq}
also holds on $\partial\Lambda$ (the {\it free boundary}),
and it precisely shows that $b_{\env}$ is therefore differentiable
at points on $\partial\Lambda$, and in fact that its $C^{1,1}$ norm there
is uniformly bounded. These are the problematic points,
since $b_{\env}$ is harmonic (and thus, well-behaved) on the complement
of $\Lambda$.
\begin{proof}
Let $x \in \Lambda\cap B_{1/4}(x_0)$,
and suppose that
$\min\{b_0,b_1\}(x_0)=b_0(x_0)$ (the case $i=1$ is treated in the same manner).
Set
\begin{equation}\label{MEq}
M := \| b_0\|_{C^{2}}.
\end{equation}
Then,
\begin{equation}\label{benvaboveEstEq}
b_{\env}(x) - b_0(x_0) -
\langle \nabla b_0(x_0), x - x_0\rangle \leq b_{\env}(x) - b_0(x)+ M|x -x_0|^2 \leq M|x-x_0|^2.
\end{equation}
Hence, it remains
to prove that
\begin{equation}\label{quadest}
-C|x-x_0|^2 \leq b_{\env}(x) - b_0(x_0) - \langle \nabla b_0(x_0), x - x_0\rangle,
\quad
\h{for all $x \in
B_{1/8}(x_0)$
}.
\end{equation}
Fix now $r \leq 1/4$. On $B_r(x_0)$,
decompose
\begin{equation}\label{sEq}
s(x):=b_{\env}(x) - b_0(x_0) - \langle \nabla b_0(x_0), x - x_0\rangle - Mr^2
\end{equation}
into the sum
$s|_{B_r(x_0)}= s_1 + s_2$, with
$s_1$ is harmonic on $B_r(x_0)$ with $s_1|_{\partial B_r(x_0)}=s|_{\partial B_r(x_0)}$.
Since $s_1$ is harmonic, $s \leq s_1 \leq0$.
Also, by the Harnack inequality for non-positive harmonic functions
it follows that
\begin{equation} \label{s_1}
-Mr^2 =s(x_0)\leq s_1(x_0) \leq C \inf_{B_{r/2}(x_0)} s_1,
\end{equation}
with $C$ independent of $r$.
\begin{claim}
\label{muClaim}
Let $\mu_{s_2}$ denote the measure associated to $\Delta s_2$.
Then either $s_2 \equiv 0$ or $\inf_{x \in B_r(x_0)} s_2$ is attained inside $B_r(x_0)$ on the support of $\mu_{s_2}$.
\end{claim}
\begin{proof}
First, since the obstacle $\min\{b_0,b_1\}$ is Lipschitz, it follows
from \cite[Lemma 3(a)]{Caf} that $b_{\env}$ is Lipschitz.
In particular, $b_{\env}$ is continuous and so $\inf_{x \in \overline{B_r(x_0)}} s_2$ is attained.
Now, suppose that the infimum is attained at a point $p$ on the complement of the support of $\mu_{s_2}$.
By definition of support, there is an open ball containing $q$ on which $s_2$ is harmonic. But, a harmonic function cannot obtain an interior minimum, which implies
that $p$ must be on the boundary of $B_r(x_0)$. But we have $s_2|_{\partial B_r(x_0)}=0$ and $s \leq s_1 \leq 0$ implies $s_2 \leq 0$. Hence, if the infimum of $s_2$ is obtained on the boundary then $s_2 \equiv 0.$
\end{proof}
If $s_2 \equiv 0$ then \eqref{quadest} follows from \eqref{s_1}. Hence we can suppose that $\inf_{x \in B_r(x_0)}s_2$ is attained at $x_1 \in B_r(x_0)$. By Claim \ref{muClaim}, $x_1\in\Lambda$ since the support of $\mu_{s_2}$
in $B_r(x_0)$ is
equal to the support $\mu_{b_{\env}}$ (the measure associated to $\Deltab_{\env}$) in $B_r(x_0)$ that is, in turn, contained in
$\Lambda\cap B_r(x_0)$.
Suppose first that $\min\{b_0,b_1\}(x_1)=b_0(x_1).$
Thus, since $x_1 \in \Lambda$, $b_{\env}(x_1)=b_0(x_1)$. Thus, using \eqref{MEq}
and \eqref{sEq},
\begin{equation} \label{s_2a}
\inf_{B_r(x_0)} s_2
=
s_2(x_1)
\ge s(x_1)= b_0(x_1) - b_0(x_0) - \langle \nabla b_0(x_0), x_1 - x_0\rangle - Mr^2 \geq -2Mr^2.
\end{equation}
Combining \eqref{benvaboveEstEq}, \eqref{s_1}, and \eqref{s_2a}
and the definition of $s$ \eqref{sEq},
proves \eqref{quadest} in this case.
\begin{figure}
\begin{center}
\includegraphics[trim=3cm 21cm 3cm 0cm, clip=true,width=3.9in]{Pb_0b_1}
\caption{The barriers $b_0,b_1$ and the envelope $P(b_0,b_1)$}
\label{figP(b0b1)}
\end{center}
\end{figure}
Suppose now that $\min\{b_0,b_1\}(x_1)=b_1(x_1)$
(see Figure \ref{figP(b0b1)}).
This case is new compared with the classical setting of Caffarelli \cite{Caf}
and will rely crucially on Proposition \ref{cushintheorem}.
Since $\min\{b_0,b_1\}(x_0)=b_0(x_0)$, it follows by continuity of $b_0$ and $b_1$
that there exists a point $\tilde x$
on the straight line segment $\{(1-t)x_0+tx_1\,:\, t\in[0,1]\}$ connecting $x_0$ and $x_1$
such that $b_1(\tilde x)=b_0(\tilde x)$, i.e. $\tilde x \in
b_{10}^{-1}(0)\cap B_{r}(x_0)$. Hence,
$$
\begin{aligned}
\label{s2_est}
\inf_{B_r(x_0)} s_2
=
s_2(x_1)
\ge s(x_1)
& = b_1(x_1) - b_0(x_0) - \langle \nabla b_0(x_0), x_1 - x_0\rangle -Mr^2
\cr
& =
(b_1(x_1) - b_1(\tilde x) -
\langle \nabla b_1(\tilde x), x_1 - \tilde x\rangle) &\nonumber
\cr
&\quad+ (\langle \nabla b_1(\tilde x), x_1 - \tilde x\rangle - \langle \nabla b_0(\tilde x), x_1 - \tilde x\rangle) & \nonumber
\cr
&\quad+ (\langle \nabla b_0(\tilde x), x_1 - \tilde x\rangle - \langle \nabla b_0(x_0), x_1 - \tilde x\rangle)
&
\cr
&\quad + (b_0(\tilde x) - b_0(x_0) - \langle \nabla b_0(x_0), \tilde x - x_0\rangle) - Mr^2\nonumber
\end{aligned}
$$
We now estimate from below the last four lines.
The first line is minorized by $-2\|b_1
\|_{C^{2}}|x_1-\tilde x|^2\ge -cr^2$, while
the third and fourth lines are
minorized by $-2\|b_0\|_{C^{2}}(|x_1-\tilde x|^2+|x_0-\tilde x|^2+r^2)\ge-cr^2$
(recall \eqref{MEq} and that $|x_1-x_0|\le r$, thus $|x_i-\tilde x|\le r$), for some $c=c(||b_0||_{C^{2}},||b_1||_{C^{2}})$.
In sum,
\begin{equation}\label{InSumEq}
\inf_{x \in B_r(x_0)} s_2
\ge
(\langle \nabla b_1(\tilde x), x_1 - \tilde x\rangle - \langle \nabla b_0(\tilde x),
x_1 - \tilde x\rangle)- Cr^2
\geq -| \nabla b_{10}(\tilde x)| |x_1 - x_0|- Cr^2.
\end{equation}
Now, by Proposition \ref{cushintheorem},
for some
$C=C(n)/(1 + \|b_0\|_{C^{2}}+\|b_1\|_{C^{2}})$,
there is a ball of radius $C|\nabla b_{10}(\tilde x)|$
around $\tilde x$ that does not intersect $\Lambda$.
But $x_0,x_1$ are both in $\Lambda$. Thus,
$$
C|\nabla b_{10}(\tilde x)| \le |x_i - \tilde x|, \quad \h{for $i=0,1$},
$$
hence,
$$
2C|\nabla b_{10}(\tilde x)| \le |x_1 - x_0|.
$$
Plugging this back into \eqref{InSumEq} yields
\begin{equation}\label{s_2b}
\inf_{B_{r/2}(x_0)} s_2 \ge
\inf_{B_r(x_0)} s_2\geq - C'r^2,
\end{equation}
for $C'=C'(||b_0||_{C^{2}},||b_1||_{C^{2}})$.
Thus, \eqref{quadest} holds also in this case.
This concludes the proof of the Proposition.
\end{proof}
Finally, we are in a position to prove the interior $C^{1,1}$ regularity of $b_{\env}$
\eqref{bminEq}.
\begin{proposition}
\label{InteriorProp}
\label{localreg} Let $b_0,b_1 \in C^{1,1}({B_1})$.
There exists $C = C(\| b_0\|_{C^{2}},\| b_1\|_{C^{2}})$ such that
$$
\| b_{\env}\|_{C^{2}(B_{1/8})}\leq C.
$$
\end{proposition}
\begin{proof}
First, $b_{\env}$ is differentiable on $ B_{1/4}$.
This is immediate on $\Lambda^c\cap B_{1/4}$ since $b_{\env}$ is harmonic there,
while
on $\Lambda\cap B_{1/4}$ this follows from Proposition \ref{QuadraticProp}.
Now, $\nabla h$ is Lipschitz continuous on $B_{1/8}$ with
Lipschitz constant $C$ if
$$
|b_{\env}(x) - b_{\env}(x_0) - \langle \nabla b_{\env}(x_0), x - x_0\rangle| \leq C |x-x_0|^2,
\quad \forall x_0,x \in B_{1/8}.
$$
This is shown in Proposition \ref{QuadraticProp} for $x_0 \in \Lambda \cap B_{1/8}$,
so suppose that
$x_0 \in \Lambda^c \cap B_{1/8}$.
Denote by $\rho$
the distance of $x_0$ to $\Lambda$.
If $\rho > 1/16$, then we are done since $b_{\env}$ is harmonic on $B_{1/8}(x_0)$
and so $||b_{\env}||_{C^{2}(B_{1/8}(x_0))}\le C||b_{\env}||_{L^\infty(B_{1/4}(x_0))}
\le C(\|b_0\|_{L^\infty(B_{1/4}(x_0))},
\|b_1\|_{L^\infty(B_{1/4}(x_0))})$
(here we used the fact that (i) $b_{\env}\le\min\{b_0,b_1\}\le\min\{\max b_0,\max b_1\}\}$,
(ii) since $b_0,b_1$ are bounded from below, the constant function $\min\{\min b_0,\min b_1\}$
is a candidate in the supremum for $b_{\env}$; thus, $b_{\env}\ge \min\{\min b_0,\min b_1\}$,
(iii) the $C^k$ norm of a harmonic function on a half-ball is estimated by its $C^0$
norm on the ball, divided by the radius of the ball to the $k$-th power---this
follows from the Poisson representation formula).
If $\rho\leq 1/16$ a different argument is needed since the radius of the ball
on which $b_{\env}$ is harmonic can be arbitrarily small. Thus, let $x_1 \in \partial\Lambda \cap B(0,1/4)$
be a point at distance exactly $\rho$ from $x_0$.
Since $b_{\env}$ is harmonic on $B_{\rho}(x_0)$
so is $b_{\env}(x) - b_{\env}(x_1) - \langle\nabla b_{\env}(x_1),x-x_1\rangle$.
Thus, one may express the latter
in terms of its boundary values and the Poisson kernel.
Since,
$$
\nabla^2b_{\env}
=
\nabla^2(b_{\env}(x) - b_{\env}(x_1) - \langle\nabla b_{\env}(x_1),x-x_1\rangle)
$$
then differentiating the aforementioned integral representation twice under the integral
sign yields that
$$
\|\nabla^2 b_{\env}(x_0)\|
\leq
C
\frac{\sup_{x \in B_{\rho}(x_0)}|b_{\env}(x) - b_{\env}(x_1) - \langle\nabla b_{\env}(x_1),x-x_1\rangle|}{\rho^2}.
$$
Finally, since $B_{\rho}(x_0)\subset B_{2\rho}(x_1)$ it follows
from Proposition \ref{QuadraticProp} that the right hand side is majorized by
$$
C\frac{\sup_{x \in B(x_1,2\rho)}|b_{\env}(x) - b_{\env}(x_1) - \langle\nabla b_{\env}(x_1),x-x_1\rangle|}
{\rho^2}
\leq C,
$$
as desired.
\end{proof}
\section*{Acknowledgments}
The first version of this article was written in Spring 2013. The final touches took place a year later while YAR was visiting Chalmers
University, and he is grateful to the Mathematics Department for an excellent working atmosphere,
and to R. Berman and B. Berndtsson for making the visit possible and for their warm hospitality.
We thank them, as well as C. Kiselman, L. Lempert, and A. Petrosyan for
their interest, encouragement, and related discussions.
This research was supported by NSF grants DMS-0802923, 1162070, 1206284,
and a Sloan Research Fellowship.
\begin{spacing}{0}
\def\bibitem{\bibitem}
|
\section[#1]{#1\label{#2}}\ignorespaces}
\ifdefined\endtheorem{}\else
\newtheorem{theorem}{Theorem}[section]
\fi
\newcommand*{\btheorem}[1]{\begin{theorem}\label{#1}\ignorespaces}
\newcommand*{\unskip\end{theorem}}{\unskip\end{theorem}}
\ifdefined\endlemma{}\else
\newtheorem{lemma}[theorem]{Lemma}
\fi
\newcommand*{\blemma}[1]{\begin{lemma}\label{#1}\ignorespaces}
\newcommand*{\unskip\end{lemma}}{\unskip\end{lemma}}
\ifdefined\endcorollary{}\else
\newtheorem{corollary}[theorem]{Corollary}
\fi
\newcommand*{\bcorollary}[1]{\begin{corollary}\label{#1}\ignorespaces}
\newcommand*{\unskip\end{corollary}}{\unskip\end{corollary}}
\newcommand*{\begin{equation}\begin{array}{lcll}}{\begin{equation}\begin{array}{lcll}}
\newcommand*{\eequations}[1]{\end{array}\label{#1}\end{equation}\ignorespaces}
\newcommand*{\begin{trivlist}\item[]\textbf{Claim.}\hspace{\labelsep}\em}{\begin{trivlist}\item[]\textbf{Claim.}\hspace{\labelsep}\em}
\newcommand*{\unskip\end{trivlist}}{\unskip\end{trivlist}}
\newcommand*{\begin{itemize}}{\begin{itemize}}
\newcommand*{\unskip\end{itemize}}{\unskip\end{itemize}}
\newcommand*{\begin{description}}{\begin{description}}
\newcommand*{\unskip\end{description}}{\unskip\end{description}}
\newcommand*{\ditem}[1]{\item[\mdseries{#1}]\ignorespaces}
\newcommand*{\[\begin{array}{lcll}}{\[\begin{array}{lcll}}
\newcommand*{\end{array}\]}{\end{array}\]}
\ifdefined\endproof{}\else
\newenvironment{proof}{\noindent\emph{Proof\/}:\hspace{\labelsep}\ignorespaces}{\bigbreak}
\fi
\newcommand*{\begin{proof}}{\begin{proof}}
\newcommand*{\end{proof}}{\end{proof}}
\hyphenation{au-tom-a-ta con-fig-u-ra-tion equiv-a-lent ex-clud-ing
ex-po-nen-tial im-ple-ment-ed pre-serv-ing push-down
straight-for-ward tran-si-tion tran-si-tions}
\title{Classical Automata on Promise Problems%
\thanks{A~preliminary version of this work was presented at the
16\textsuperscript{th}~International Workshop on Descriptional
Complexity of Formal Systems (DCFS~2014), August 5--8, 2014,
Turku, Finland [vol.~8614 of LNCS, pp.~126--137, Springer-Verlag, 2014]\@.
An~ArXiv version is at http://arxiv.org/abs/1405.6671.}}%
\author{Viliam Geffert\thanks{Partially supported by the Slovak grant contracts VEGA 1/0479/12
and APVV-0035-10\@.}
\\
Department of Computer Science,
P.\,J.\,\v{S}af\'{a}rik University, Ko\v{s}ice, Slovakia
\\
<EMAIL>
\\ \\
Abuzer Yakary{\i}lmaz\thanks{Partially supported by CAPES with grant 88881.030338/2013-01,
ERC Advanced Grant MQC, and FP7 FET project QALGO\@. Moreover,
the part of the research work was done while Yakary{\i}lmaz was
visiting Kazan Federal University\@.}%
\\
National Laboratory for Scientific Computing,
Petr\'{o}polis, RJ, Brazil
\\
<EMAIL>
}
\date{\today}%
\begin{document
\maketitle
\begin{abstract}
Promise problems were mainly studied in quantum automata theory. Here we focus on state complexity of classical automata for promise problems. First, it was known that there is a family of unary promise problems solvable by quantum automata by using a single qubit, but the number of states required by corresponding one-way deterministic automata cannot be bounded by a constant. For this family, we show that even two-way nondeterminism does not help to save a single state. By comparing this with the corresponding state complexity of alternating machines, we then get a tight exponential gap between two-way nondeterministic and one-way alternating automata solving unary promise problems. Second, despite of the existing quadratic gap between Las Vegas realtime probabilistic automata and one-way deterministic automata for language recognition, we show that, by turning to promise problems, the tight gap becomes exponential. Last, we show that the situation is different for one-way probabilistic automata with two-sided bounded-error. We present a family of unary promise problems that is very easy for these machines; solvable with only two states, but the number of states in two-way alternating or any simpler automata is not limited by a constant. Moreover, we show that one-way bounded-error probabilistic automata can solve promise problems not solvable at all by any other classical model.
\\
\textbf{Keywords:} descriptional complexity, promise problems, nondeterministic automata, probabilistic automata, alternating automata
\end{abstract}
\bsection{Introduction}{s:intro}
Promise problem is a generalization of language recognition.
Instead of testing all strings over a given alphabet for language
membership, we focus only on a given subset of strings as
potential inputs that should be tested. The language under
consideration and its complement form this subset.
Promise problems have already provided several different
perspectives in computational complexity. (See the survey by
Goldreich~\cite{Gol06A}\@.) For example, it is not known whether
the class of languages recognizable by bounded-error probabilistic
polynomial-time algorithms has a complete problem, but the class
of promise problems solvable by the same type of algorithms has
some complete problems. A~similar phenomenon has been observed for
quantum complexity classes~\cite{Wat09A}\@. The first known result
on promise problems for restricted computational models we are
aware of was given by Condon and Lipton in 1989~\cite{CL89}\@.
They showed that a promise version of emptiness problem is
undecidable for one-way probabilistic finite automata. In the
literature, some separation results regarding restricted
computational models have also been given in the form of promise
problems. (See, e.g.,~\cite{DS90,DS92}\@.) The first result for
restricted quantum models was given by Murakami
\mbox{et\,al.\,}\cite{MNYW05}\@: There exists a promise problem
that can be solved by quantum pushdown automata exactly but not by
any deterministic pushdown automaton. Recently, Ambainis and
Yakary{\i}lmaz~\cite{AY12} showed that there is an infinite family
of promise problems which can be solved exactly just by tuning
transition amplitudes of a realtime two-state quantum finite
automaton (\textsc{qfa}), whereas the size of the corresponding classical
automata grows above any fixed limit. Moreover, Rashid and
Yakary{\i}lmaz~\cite{RY14A} presented many superiority results of
\textsc{qfa}s\ over their classical counterparts. There are also some new
results on succinctness of
\textsc{qfa}s~\cite{ZQGLM13,GQZ14A,GQZ14B,ZGQ14A,BMP14B} and a new result
on quantum pushdown automata~\cite{Nak14A}\@.
\medbreak
In this paper, we turn our attention to classical models and
obtain some new results. In the next two sections, we give some
preliminaries and present some basic facts for the classical
models. Among others, we show that \iit{i}~the computational power
of deterministic, nondeterministic, alternating, and Las Vegas
probabilistic automata is the same, even after turning {}from the
classical language recognition to solving promise problems, and,
\iit{ii}~on promise problems, neither existing gaps of
succinctness for classical language recognition can be violated,
between any two of deterministic, nondeterministic, and
alternating automata models.
Then, in Section~\ref{s:expensive}, we consider a family of unary
promise problems given by Ambainis and Yakary{\i}l\-maz
in~\cite{AY12}, solvable by only two-state realtime quantum finite
automata. On the other hand, for solving this family by the use
classical automata, we show that the exact number of states for
one-way/two-way deterministic/nondeterministic automata is the
same. Thus, for this family, replacing the classical one-way
deterministic automata even by two-way nondeterministic automata
does not help to save a single state. However, for the same
family, already the one-way alternation does help, which gives us
the tight exponential gap between one-way alternating and two-way
sequential models of automata solving unary promise problems.
In Section~\ref{s:LV}, we show that the trade-off for the case of
Las Vegas probabilistic versus one-way deterministic automata is
different; by turning {}from language recognition to promise
problems the tight quadratic gap changes to a tight exponential
gap.
In Section~\ref{s:bounded-error}, we show how the situation
changes for finite state automata with two-sided bounded-error.
First, we present a probabilistically very easy family of unary
promise problems, i.e., solvable by one-way two-state automata
with bounded-error. Then, we show that this family is hard for
two-way alternating automata, and hence for any simpler classical
automata as well. Finally, we prove that, on promise problems,
bounded-error probabilistic automata are more powerful than any
other classical model.
\bsection{Preliminaries}{s:prel}
We assume the reader is familiar with the basic standard models of
finite state automata (see e.g.~\cite{HMU07})\@. For a more
detailed exposition related to probabilistic automata, the reader
is referred to~\cite{Pa71,Buk80}\@. Here we only recall some
models and introduce some elementary notation.
By~$\Sigma\str$, we denote the set of strings over an
alphabet~$\Sigma$\@. The set of strings of length~$n$ is denoted
by~$\Sigma^n$ and the unique string of length zero by~\eps\@.
By~$\nnumber$, we denote the set of all positive integers. The
cardinality of a finite set~$S$ is denoted by~$\card{S}$\@.
\medbreak
A~\emph{one-way nondeterministic finite automaton} with \eps-moves
(\owe\textsc{nfa}, for short) is a quintuple
$\ensuremath{\mathcal{A}}= (S,\Sigma,\lbr H,\lbr s\ini,S\acc)$, where $S$~is a finite
set of states, $\Sigma$~an input alphabet, $s\ini\in S$ an initial
state, and $S\acc\subseteq S$ a set of accepting (final) states.
Finally, $H\subseteq S\times(\Sigma\cup\set{\eps})\times S$ is a
set of \emph{transitions}, with the following meaning.
$s\trs{a} s'\in H$\@: if the next input symbol is~$a$, \ensuremath{\mathcal{A}}~gets
{}from the state~$s$ to the state~$s'$ by reading~$a$ {}from the
input. $s\trs{\eps} s'\in H$\@: \ensuremath{\mathcal{A}}~gets {}from~$s$ to~$s'$
without reading any symbol {}from the input.%
\footnote{Traditionally~\cite{HMU07}, the behavior is defined by a
function $\delta: S\times(\Sigma\cup\set{\eps})\rightarrow 2^S$\@.
However, a traditional transition $\delta(s,a)\ni s'$ can be
easily converted to $s\trs{a} s'$ and vice versa. In the same
way, all models introduced here are consistent with the
traditional definitions. Using transitions instead of a
$\delta$-function, we just make the notation more readable\@.}
The machine \ensuremath{\mathcal{A}}~\emph{halts} in the state~$s$, if there are no
executable transitions in~$H$ at a time. Typically, this happens
after reading the entire input, but the computation may also be
blocked prematurely, by \emph{undefined transitions}, i.e., by
absence of transitions for the current state~$s$ and the next
input symbol~$a$ (or~$\eps$)\@.
An~\emph{accepting computation} starts in the initial
state~$s\ini$ and, after reading the entire input, it halts in any
accepting state $s'\in S\acc$\@. The language \emph{recognized}
(accepted) by~\ensuremath{\mathcal{A}}, denoted by~$L(\ensuremath{\mathcal{A}})$, is the set of all input
strings having at least one accepting computation. The inputs with
no accepting computation are \emph{rejected}\@.
A~\emph{two-way nondeterministic finite automaton} (\tw\textsc{nfa}) is
defined in the same way as \owe\textsc{nfa}, but now \ensuremath{\mathcal{A}}~can move in both
directions along the input. For these reasons, the transitions
in~$H$ are upgraded as follows. $s\trs{\!\!a\rightd} s'$,
$s\trs{\!\!a\leftd} s'$, $s\trs{\!\!a\stayd} s'$\@: if the current
input symbol is~$a$, \,\ensuremath{\mathcal{A}}~gets {}from $s$ to~$s'$ and moves its
input head one position to the right, to the left, or keeps it
stationary, respectively. The input is enclosed in between two
special symbols \quo{$\vdash$} and~\quo{$\dashv$}, called the
\emph{left and right endmarkers}, respectively. By definition, the
input head cannot leave this area, i.e., there are no transitions
moving to the left of~$\vdash$ nor to the right of~$\dashv$\@.
\,\ensuremath{\mathcal{A}}~starts in~$s\ini$ with the input head at the left endmarker
and accepts by halting in $s'\in S\acc$ anywhere along the input
tape.
\medbreak
A~\emph{deterministic finite automaton} (\owe\textsc{dfa}\ or \tw\textsc{dfa}) can
be obtained {}from \owe\textsc{nfa}\ or \tw\textsc{nfa}\ by claiming that the
transition set~$H$ does not allow executing more than one possible
transition at a time. Consequently, a \owe\textsc{dfa}~\ensuremath{\mathcal{A}}\ can have at
most one transition $s\trs{a} s'$, for each $s\in S$ and each
$a\in\Sigma$\@. In theory, one can consider \eps-transitions (even
though this feature is rarely utilized)\@: a~transition
$s\trs{\eps} s'$ implies that there are no other transitions
starting in the same state~$s$\@. Similar restrictions can be
formulated for \tw\textsc{dfa}s\@.
\medbreak
A~\emph{self-verifying finite automaton} (\owe\textsc{svfa}\ or
\tw\textsc{svfa})~\cite{JP11} is a \owe\textsc{nfa}\ or \tw\textsc{nfa}~\ensuremath{\mathcal{A}}\ which is
additionally equipped with $S\rej\subseteq S$, the set of
rejecting states disjoint {}from~$S\acc$, the set of accepting
states. The remaining states form
$S\neu= S\setminus(S\acc\cup S\rej)$, the set of neutral
\quo{don't know} states. In this model, \iit{i}~for each accepted
input, \ensuremath{\mathcal{A}}~has at least one computation path halting in an
accepting state and, moreover, no path halts in a rejecting state,
\iit{ii}~for each rejected input, \ensuremath{\mathcal{A}}~has at least one
computation path halting in a rejecting state and, moreover, no
path halts in an accepting state. In both these cases, some
computation paths may return \quo{don't know} answers, by halting
in neutral states or by getting into infinite loops.
\medbreak
A~\emph{one-way probabilistic finite automaton} with \eps-moves
(\owe\textsc{pfa}) is defined in the same way as \owe\textsc{nfa}, but the
transitions in~$H$ are upgraded with probabilities, as follows.
$s\trs{\!\!a,p} s'\in H$\@: if the current input symbol is~$a$,
\ensuremath{\mathcal{A}}~gets {}from the state~$s$ to the state~$s'$ by reading this
symbol with probability $p\in [0,1]$\@. For $a=\eps$, no input
symbol is consumed. A~transition with $p=1$ may be displayed as
$s\trs{a} s'$, since the next step is deterministic; transitions
with $p=0$ can be completely erased, as not executable. By
definition,%
\footnote{Traditionally~\cite{Ra63}, the behavior of \textsc{pfa}s\ is
defined by stochastic matrices\@.}
for each $s\in S$ and each $a\in\Sigma$, the sum of probabilities
over all transitions beginning in~$s$ and labeled by
$a'\in\set{a,\eps}$ must be equal to~$1$\@. The overall accepting
probability of~\ensuremath{\mathcal{A}}\ on a given input $w\in\Sigma\str$, denoted
by~$f_{\ensuremath{\mathcal{A}}}(w)$, is calculated over all accepting paths. Hence, the
overall rejecting probability is $1\!-\!f_{\ensuremath{\mathcal{A}}}(w)$\@.
A~two-way version, \tw\textsc{pfa}, was first introduced in~\cite{Ku73}\@.
A~\emph{Las Vegas probabilistic finite automaton} (Las Vegas
\owe\textsc{pfa}\ or Las Vegas \tw\textsc{pfa})~\cite{HS01} is a \owe\textsc{pfa}\ or
\tw\textsc{pfa}\ of special kind, obtained {}from self-verifying automata
(i.e., {}from nondeterministic automata of special kind)\@. Thus,
transitions are upgraded with probabilities, and the state set is
partitioned into the sets of accepting, rejecting, and neutral
\quo{don't know} states. This gives three overall probabilities on
a given input; accepting, rejecting, and neutral.
\medbreak
A~\emph{one-way alternating finite automaton} with \eps-moves
(\owe\textsc{afa}) is obtained {}from \owe\textsc{nfa}\ by partitioning the state
set~$S$ into two disjoint subsets $S\exs$ and~$S\uns$, the sets of
\emph{existential} and \emph{universal} states, respectively. The
global rules are defined as is usual for alternating devices (see
e.g.~\cite{CKS81})\@: if, at the given moment, there are several
executable transitions, the machine~\ensuremath{\mathcal{A}}\
\,\iit{i}~nondeterministically chooses one of them, if it is in an
existential state~$s$, but \iit{ii}~follows, in parallel, all
possible branches, if the current state~$s$ is universal.
For better readability, a nondeterministic switch {}from the
existential~$s$ into one of the states $s_1,\ldots,s_k$, by reading
an input symbol~$a$, will be displayed as
$s\trs{a} s_1\vee \cdots\vee s_k$, while a parallel branching {}from
the universal~$s$ to all these states as
$s\trs{a} s_1\wedge \cdots\wedge s_k$\@. The same applies for
$a=\eps$\@.
The input is accepted, if all branches in the nondeterministically
chosen computation subtree, rooted in the initial state at the
beginning of the input and embedded in the full tree of all possible
computation paths, halt in accepting states at the end of the
input.%
\footnote{\label{ft:boolean}The alternating automaton should not
be confused with a \emph{Boolean finite automaton}, which is
(quite inappropriately) also sometimes called \quo{alternating}
in some literature. Instead of alternating existential and
universal branching, the Boolean automaton can control
acceptance by the use of arbitrarily complex Boolean functions.
As an example, $s\trs{a} s_1\wedge(s_2\vee s_3)$ specifies that,
for the current input symbol~$a$, the subtree rooted in the
state~$s$ is successful if and only if the subtree rooted
in~$s_1$ and at least one of those rooted in~$s_2,s_3$ are
successful\@.}
Also this model was extended to a two-way version, \tw\textsc{afa}\ (see
e.g.~\cite{Ge12})\@.
\medbreak
A~\emph{realtime automaton} (\rt\textsc{dfa}, \rt\textsc{svfa}, \rt\textsc{nfa}, \rt\textsc{pfa},
\rt\textsc{afa},\dots\ ) is the corresponding one-way finite automaton%
\footnote{All realtime \emph{quantum} models use a classical input
head and so they are not allowed to create a superposition of
the head positions on the input tape\@.}
that executes at most a constant number of \eps-transitions in a
row. (After that, the next input symbol is consumed or the machine
halts\@.)
A~\emph{one-way \eps-free automaton} (\ow\textsc{dfa}, \ow\textsc{svfa}, \ow\textsc{nfa},
\ow\textsc{pfa}, \ow\textsc{afa},\dots\ ) is the corresponding one-way finite
automaton with no \eps-transitions at all. (This is a special case
of a realtime machine, with the number of executed
\eps-transitions bounded by zero\@.)
It should be pointed out that, in \owe\textsc{nfa}s, the \eps-transitions
can be removed \emph{without increasing} the number of
states~\cite[Sect.~2.11]{Ku10}\@. Therefore, \ow\textsc{nfa}s, \rt\textsc{nfa}s,
and \owe\textsc{nfa}s\ agree in upper/lower bounds for states. The same
works for \ow\textsc{svfa}s, \rt\textsc{svfa}s, and \owe\textsc{svfa}s, as well as for
\ow\textsc{dfa}s, \rt\textsc{dfa}s, and~\owe\textsc{dfa}s\@. For this reason, we consider
only \ow\textsc{nfa}s, \ow\textsc{svfa}s, and \ow\textsc{dfa}s\@.
\medbreak
Occasionally, we shall also mention some restricted versions of
two-way machines, e.g., \emph{sweeping automata}, changing the
direction of the input head movement only at the
endmarkers~\cite{Sip80,KKM12}, or a very restricted version of
sweeping automaton called a \emph{restarting one-way automaton},
running one-way in an infinite loop~\cite{YS10B}\,---\,if the
computation does not halt at the right endmarker, then it jumps
back to the initial state at the beginning of the input.
\medbreak
A~\emph{promise problem} is a pair of languages
$\ensuremath{\mathtt{P}}= (\ensuremath{\mathtt{P}}\yes,\ensuremath{\mathtt{P}}\no)$, where
$\ensuremath{\mathtt{P}}\yes,\ensuremath{\mathtt{P}}\no\subseteq \Sigma\str$, for some~$\Sigma$, and
$\ensuremath{\mathtt{P}}\yes\cap\ensuremath{\mathtt{P}}\no= \emptyset$~\cite{Wat09A}\@.
The promise problem~\ensuremath{\mathtt{P}}\ is \emph{solved} by a finite
automaton~\ensuremath{\mathcal{P}}, if \ensuremath{\mathcal{P}}\ accepts each $w\in\ensuremath{\mathtt{P}}\yes$ and rejects
each $w\in\ensuremath{\mathtt{P}}\no$\@. (That is, we do not have to worry about the
outcome on inputs belonging neither to~$\ensuremath{\mathtt{P}}\yes$ nor
to~$\ensuremath{\mathtt{P}}\no$\@.) If $\ensuremath{\mathtt{P}}\yes\cup\ensuremath{\mathtt{P}}\no= \Sigma\str$ and \ensuremath{\mathtt{P}}~is
solved by~\ensuremath{\mathcal{P}}, we have a classical language recognition, and say
that $\ensuremath{\mathtt{P}}\yes$~is \emph{recognized} by~\ensuremath{\mathcal{P}}\@.
If \ensuremath{\mathcal{P}}~is a probabilistic finite automaton, we say that \ensuremath{\mathcal{P}}\
solves~\ensuremath{\mathtt{P}}\ with \emph{error bound} $\eps\in [0,\frac{1}{2})$, if
\ensuremath{\mathcal{P}}\ accepts each $w\in\ensuremath{\mathtt{P}}\yes$ with probability at least
$1\!-\!\eps$ and rejects each $w\in\ensuremath{\mathtt{P}}\no$ with probability at
least $1\!-\!\eps$\@. (The remaining probability goes to erroneous
answers\@.) \ensuremath{\mathcal{P}}~solves \ensuremath{\mathtt{P}}\ with a \emph{bounded-error}, if, for
some~\eps, it solves~\ensuremath{\mathtt{P}}\ with the error bound~\eps\@. If
$\eps=0$, the problem is solved \emph{exactly}\@.
A~special case of bounded-error is a \emph{one-sided bounded-error},
where either each $w\in\ensuremath{\mathtt{P}}\yes$ is accepted with probability~$1$ or
each $w\in\ensuremath{\mathtt{P}}\no$ is rejected with probability~$1$\@.
If \ensuremath{\mathcal{P}}\ is a Las Vegas probabilistic finite automaton, we say
that \ensuremath{\mathcal{P}}\ solves~\ensuremath{\mathtt{P}}\ with a \emph{success probability}
$p\in (0,1]$, if \iit{i}~for each $w\in\ensuremath{\mathtt{P}}\yes$,
\,\ensuremath{\mathcal{P}}~accepts~$w$ with probability at least~$p$ and never rejects
(i.e., the probability of rejection is~$0$) and \iit{ii}~for each
$w\in\ensuremath{\mathtt{P}}\no$, \,\ensuremath{\mathcal{P}}~rejects~$w$ with probability at least~$p$
and never accepts. (In both these cases, the remaining probability
goes to \quo{don't know} answers\@.)
\bsection{Basic Facts on Classical Automata}{s:basics}
We continue with some basic facts regarding classical automata. We start with a basic observation: The cardinality of the class of the promise problems defined by \ow\textsc{dfa}s~ is uncountable. We can show this fact even for the class defined by a fixed 2-state \ow\textsc{dfa}~ on unary languages. Let $ \mathtt{L} \subseteq \mathcal{N} $ be any language defined on natural numbers. Then, the following unary promise problem $ \mathtt{P(L)} $ can be solvable by a 2-state \ow\textsc{dfa}, say $ \mathcal{D} $:
\[
\begin{array}{llll}
\mathtt{P_{yes}(L)} & = & \{ a^{2n} & \mid n \in \mathtt{L} \}
\\
\mathtt{P_{no}(L)} & = & \{ a^{2n+1} & \mid n \in \mathtt{L} \}
\end{array} ,
\]
where $ \mathcal{D} $ has two states, the initial and only accepting state $s_1$ and $ s_2 $, and, the automaton alternates between $s_1$ and $s_2$ for each input symbol. There is a one-to-one correspondence between $ \{ \mathtt{L} \mid \mathtt{L} \subseteq \mathcal{N} \} $ and $ \{ \mathtt{P(L)} \mid \mathtt{L} \subseteq \mathcal{N} \} $. Therefore, the cardinality of the class of the promise problems solvable by $ \mathcal{D} $ is uncountable. Moreover, if $ \mathtt{L} $ is an uncomputable language, then both $ \mathtt{P_{yes}(L)} $ and $ \mathtt{P_{no}(L)} $ are uncomputable. The interested reader can easily find some other promise problems (solvable by very small \ow\textsc{dfa}s) whose yes- and no-instances satisfy some special properties (i.e. non-context-free but context-sensitive, NP-complete, semi-decidable, etc.). One trivial construction for a binary language $ \mathtt{L} \in \{ a,b \}^* $ is to define the promise problem as $ \mathtt{P(L)} = ( \{ aw \mid w \in \mathtt{L} \},\{ bw \mid w \in \mathtt{L} \} ) $. (A promise problem solvable by \ow\textsc{dfa}s~ whose yes- and no-instances are non-regular languages previously given in \cite{GQZ14B}.) Despite of the above facts, all trade-offs established between various models of automata stay valid also for promise problems.
First, we show that the classes of promise problems solvable by
deterministic, nondeterministic, alternating, and Las Vegas
probabilistic finite automata are identical.
\btheorem{t:afa-to-dfa}
If a promise problem $\ensuremath{\mathtt{P}}=(\ensuremath{\mathtt{P}}\yes,\ensuremath{\mathtt{P}}\no)$ can be solved by a
\tw\textsc{afa}~\ensuremath{\mathcal{A}}\ with $n$ states, it can also be solved by a
\ow\textsc{dfa}~\ensuremath{\mathcal{D}}\ with $t(n)\le 2^{n\cdot 2^n}$ states.
\unskip\end{theorem}
\begin{proof}
Let~\ensuremath{\mathtt{R}}\ denote the regular language recognized by~\ensuremath{\mathcal{A}}\@. (That
is, even though we are given a \tw\textsc{afa}~\ensuremath{\mathcal{A}}\ for solving a promise
problem~\ensuremath{\mathtt{P}}, such machine is still associated with some classical
regular language $L(\ensuremath{\mathcal{A}})=\ensuremath{\mathtt{R}}$\@.) The complement of this
language will be denoted by~\ensuremath{\overline{\mathtt{R}}}\@. Since the problem \ensuremath{\mathtt{P}}\ is
solvable by~\ensuremath{\mathcal{A}}, we have that $\ensuremath{\mathtt{P}}\yes\subseteq \ensuremath{\mathtt{R}}$ and
$\ensuremath{\mathtt{P}}\no\subseteq \ensuremath{\overline{\mathtt{R}}}$\@.
Now, by~\cite{LLS84}, each \tw\textsc{afa}~\ensuremath{\mathcal{A}}\ with $n$ states can be
transformed into an equivalent \ow\textsc{dfa}~\ensuremath{\mathcal{D}}\ with
$t(n)\le 2^{n\cdot 2^n}\!$ states. This gives that
$\ensuremath{\mathtt{P}}\yes\subseteq \ensuremath{\mathtt{R}}= L(\ensuremath{\mathcal{D}})$ and
$\ensuremath{\mathtt{P}}\no\subseteq \ensuremath{\overline{\mathtt{R}}}= \overline{L(\ensuremath{\mathcal{D}})}$, that is,
\ensuremath{\mathcal{D}}~accepts each $w\in\ensuremath{\mathtt{P}}\yes$ and rejects each $w\in\ensuremath{\mathtt{P}}\no$\@.
Therefore, \ensuremath{\mathcal{D}}~can be used for solving~\ensuremath{\mathtt{P}}\@.
\end{proof}
The proof of the above theorem can be easily updated for other
classical models of automata\,---\,using the corresponding
trade-off~$t(n)$ for the conversion known {}from the
literature\,---\,by which we obtain the following corollary.
\bcorollary{c:not-violated}
Any trade-off~$t(n)$ in the number of states for language
recognition, between any two of deterministic, nondeterministic,
or alternating automata models, is also a valid trade-off for
promise problems.
\unskip\end{corollary}
Next, we show that neither Las~Vegas \textsc{pfa}s\ gain any computational
power.
\btheorem{t:lv-to-dfa}
If a promise problem $\ensuremath{\mathtt{P}}=(\ensuremath{\mathtt{P}}\yes,\ensuremath{\mathtt{P}}\no)$ can be solved by a
Las Vegas \tw\textsc{pfa}~\ensuremath{\mathcal{P}}\ with $n$ states and any success
probability $p>0$, it can also be solved by a \ow\textsc{dfa}~\ensuremath{\mathcal{D}}\ with
$t(n)\le 2^{n^2+n}$ states.
\unskip\end{theorem}
\begin{proof}
First, for each $w\in\ensuremath{\mathtt{P}}\yes$, the machine~\ensuremath{\mathcal{P}}\ has at least one
accepting computation path and no rejecting path. Conversely, for
each $w\in\ensuremath{\mathtt{P}}\no$, the machine has at least one rejecting path
and no accepting path. In both cases, the paths that are not
successful return a \quo{don't know} answer.
Therefore, by removing the probabilities and converting neutral
\quo{don't know} states into rejecting states, we obtain an
$n$-state \tw\textsc{nfa}~\ensuremath{\mathcal{N}}\ recognizing a regular language~\ensuremath{\mathtt{R}}\ such
that $\ensuremath{\mathtt{P}}\yes\subseteq \ensuremath{\mathtt{R}}$ and
$\ensuremath{\mathtt{P}}\no\cap\ensuremath{\mathtt{R}}\ = \emptyset$\@. Therefore,
$\ensuremath{\mathtt{P}}\no\subseteq \ensuremath{\overline{\mathtt{R}}}$\@.
Now, by~\cite{Ka05m}, the \tw\textsc{nfa}~\ensuremath{\mathcal{N}}\ can be converted to an
equivalent \ow\textsc{dfa}~\ensuremath{\mathcal{D}}\ with the number of states bounded by
$t(n)\le
\sum_{i=0}^{n-1} \sum_{j=0}^{n-1} ({n\atop i})({n\atop j})(2^i\!-\!1)^j\le
2^{n^2+n}$,
which gives $\ensuremath{\mathtt{P}}\yes\subseteq \ensuremath{\mathtt{R}}= L(\ensuremath{\mathcal{D}})$ and
$\ensuremath{\mathtt{P}}\no\subseteq \ensuremath{\overline{\mathtt{R}}}= \overline{L(\ensuremath{\mathcal{D}})}$\@. That is,
\ensuremath{\mathcal{D}}~accepts each $w\in\ensuremath{\mathtt{P}}\yes$ and rejects each $w\in\ensuremath{\mathtt{P}}\no$,
and hence it can be used for solving~\ensuremath{\mathtt{P}}\@.
\end{proof}
A~similar idea can be used for Las Vegas \owe\textsc{pfa}s\@:
\bcorollary{c:lv-to-dfa}
If a promise problem $\ensuremath{\mathtt{P}}=(\ensuremath{\mathtt{P}}\yes,\ensuremath{\mathtt{P}}\no)$ can be solved by a
Las Vegas \owe\textsc{pfa}~\ensuremath{\mathcal{P}}\ with $n$ states and any success
probability $p>0$, it can also be solved by a \ow\textsc{dfa}~\ensuremath{\mathcal{D}}\ with
$t(n)\le 1\!+\!3^{(n-1)/3}\le 2^{0.529\cdot n}$ states.
\unskip\end{corollary}
\begin{proof}
By removing the probabilities, we first obtain an $n$-state
\owe\textsc{svfa}~\ensuremath{\mathcal{N}}\ recognizing a regular language~\ensuremath{\mathtt{R}}\ such that
$\ensuremath{\mathtt{P}}\yes\subseteq \ensuremath{\mathtt{R}}$ and $\ensuremath{\mathtt{P}}\no\subseteq \ensuremath{\overline{\mathtt{R}}}$\@. Then,
after elimination of \eps-transitions by the method described
in~\cite[Sect.~2.11]{Ku10} (still not increasing the number of
states), we have a \ow\textsc{svfa}\ that can be converted to an
equivalent \ow\textsc{dfa}~\ensuremath{\mathcal{D}}\ with $t(n)\le 1+3^{(n-1)/3}$
states~\cite{JP11}\@. The resulting \ow\textsc{dfa}\ is then used for
solving~\ensuremath{\mathtt{P}}\@.
\end{proof}
In Section~\ref{s:LV}, we show that the corresponding lower bound
is also exponential.
\bsection{A~Classically Expensive Unary Promise Problem}{s:expensive}
Recently, Ambainis and Yakary{\i}lmaz~\cite{AY12} presented a
family of promise problems $\mathtt{EvenOdd}(k)$, for
\mbox{$k\!\in\!\nnumber$}, where
\[\begin{array}{lcll}
\mathtt{EvenOdd}\yes(k) &=& \set{ a^{m\cdot 2^{k}} \mid\,} %"such that" in the middle of a set: \set{ ... \st ... m \mbox{ is even} } \,,\\
\mathtt{EvenOdd}\no(k) &=& \set{ a^{m\cdot 2^{k}} \mid\,} %"such that" in the middle of a set: \set{ ... \st ... m \mbox{ is odd} } \,,
\end{array}\]
such that, for each~$k$, the promise problem $\mathtt{EvenOdd}(k)$ can be
solved exactly by only a two-state realtime quantum finite
automaton. On the other hand, for \ow\textsc{dfa}s, $2^{k+1}$~states are
necessary and also sufficient. Later, it was shown that
$2^{k+1}$~states are also necessary in any bounded-error
\ow\textsc{pfa}~\cite{RY14A} and that \tw\textsc{dfa}s\ must use at least
$2^{k+1}\!-\!1$ states~\cite{BMP14B}\@.%
\footnote{Recently~\cite{AGKY14A}, it was also shown that the
width of any nondeterministic or bounded-error stable \textsc{obbd}\
cannot be smaller than~$2^{k+1}$, for a functional version of
$\mathtt{EvenOdd}(k)$\@.}
Here we show that even two-way nondeterminism does not help us to
save a single state for this promise problem; and hence
$2^{k+1}$~states is the \emph{exact} value for one-way/two-way
deterministic/nondeterministic automata. After that, we show that
alternation does help to reduce the number of states
to~$\Theta(k)$, already by the use of one-way \mbox{\eps-}free
machines. More precisely, $\mathtt{EvenOdd}(k)$ can be solved by a
\ow\textsc{afa}\ with $11k\!-\!14$ states (for $k\ge 3$); the corresponding
lower bound for \ow\textsc{afa}s\ is~$k\!+\!1$\@. For two-way
machines,~\tw\textsc{afa}s, the exact lower bound is an open problem, but
we know that it must grow in~$k$, which is an easy consequence of
Theorems \ref{t:afa-to-dfa} and~\ref{t:nfa-even-odd}\@. Similarly,
we do not know whether bounded-error \tw\textsc{pfa}s\ can work with fewer
than $2^{k+1}$ states.
\medbreak
The method used for proving the lower bound in the following
theorem is not new\,---\,inspired
by~\cite{SHL65,CHR91,Ge91}\,---\,but it needs some modifications
on promise problems.
\btheorem{t:nfa-even-odd}
For each $k\in\nnumber$, any \tw\textsc{nfa}\ for solving the promise problem
$\mathtt{EvenOdd}(k)$ needs at least $2^{k+1}$ states.
\unskip\end{theorem}
\begin{proof}
Suppose, for contradiction, that the promise problem can be solved
by a \tw\textsc{nfa}\ \ensuremath{\mathcal{N}}\ with $\card{S}<2^{k+1}$ states.
Consider now the input~$a^{2^{k+1}}\!$\@. Clearly, it must be
accepted by~\ensuremath{\mathcal{N}}, and hence there must exist at least one
accepting computation path, so we can fix one such path. (Besides,
being nondeterministic, \ensuremath{\mathcal{N}}~can have also other paths, not all
of them necessarily accepting, but we do not care for them\@.)
Along this fixed path, take the sequence of states
\[\begin{array}{lcll}
& q_0,\,p_1,q_1,\,p_2,q_2,\,p_3,q_3,\ldots,p_r,q_r,\,p_{r+1} \,,
\end{array}\]
where $q_0$~is the initial state, $p_{r+1}$~is an accepting state,
and all the other states $\set{p_i\cup q_i\mid\,} %"such that" in the middle of a set: \set{ ... \st ... 1\le i\le r}$ are at
the left/right endmarkers of the input~$2^{k+1}$, such that:
\begin{itemize}
\item If $p_i$~is at the left (resp., right) endmarker, then
$q_i$ is at the opposite endmarker, and the path {}from $p_i$
to~$q_i$ traverses the entire input {}from left to right
(resp., right to left), not visiting any of endmarkers in the
meantime.
\item The path between $q_i$ and~$p_{i+1}$ starts and ends at
the same endmarker, possibly visiting this endmarker several
times, but not visiting the opposite endmarker. Such a path is
called a \quo{U-turn}. This covers also the case of
$q_i=p_{i+1}$ with zero number of executed steps in between.
\unskip\end{itemize}
Now, the path connecting $p_i$ with~$q_i$ must visit all input
tape positions in the middle of the input~$a^{2^{k+1}}\!$, and
hence \ensuremath{\mathcal{N}}\ must enter $2^{k+1}$ states, at least one state per
each input tape position. However, since $\card{S}< 2^{k+1}$,
some state $r_i$ must be repeated. That is, between $p_i$
and~$q_i$, there must exist a loop, starting and ending in the
same state~$r_i$, traveling some $\ell_i$~positions to the right,
not visiting any of the endmarkers. For each traversal, we fix one
such loop (even though the path {}from $p_i$ to~$q_i$ may contain
many such loops)\@. Note that the argument for traveling across
the input {}from right to left is symmetrical. Since this loop is
in the middle of the traversal, we have that $\ell_i<2^{k+1}$\@.
But then $\ell_i$ can be expressed in the form
\[\begin{array}{lcll}
\ell_i &=& \gamma_i\!\cdot\!2^{\alpha_i} ,
\end{array}\]
where $\gamma_i>0$ is odd (not excluding $\gamma_i=1$) and
$\alpha_i\in \set{0,1,\ldots,k}$\@. Note that if
$\alpha_i\ge k\!+\!1$, we would have a contradiction with
$\ell_i<2^{k+1}$\@.
Now, consider the value
\[\begin{array}{lcll}
\ell &=& \lcm\set{ \ell_1,\ell_2,\ldots,\ell_r } \,,
\end{array}\]
the least common multiple of all fixed loops, for all traversals
of the input~$a^{2^{k+1}}\!$\@. Clearly, $\ell$~can also be
expressed in the form
\[\begin{array}{lcll}
\ell &=& \gamma\!\cdot\!2^{\alpha} ,
\end{array}\]
where $\gamma>0$ is odd and $\alpha\in \set{0,1,\ldots,k}$
\,(actually, $\gamma= \lcm\set{\gamma_1,\ldots,\gamma_r}$ and
$\alpha= \max\set{\alpha_1,\ldots,\alpha_r}$)\@.
We shall now show that the machine \ensuremath{\mathcal{N}}\ must also accept the
input
\[\begin{array}{lcll}
& a^{2^{k+1}+\ell\cdot 2^{k-\alpha}} .
\end{array}\]
First, if $q_i$ and~$p_{i+1}$ are connected by a U-turn path on
the input~$a^{2^{k+1}}\!$, they will be connected also on the
input $a^{2^{k+1}+\ell\cdot 2^{k-\alpha}}$ since such path does
not visit the opposite endmarker and
$2^{k+1}+\ell\!\cdot\!2^{k-\alpha}\ge 2^{k+1}$\@. Second, if $p_i$
and~$q_i$ are connected by a left-to-right traversal
along~$a^{2^{k+1}}\!$, they will stay connected also along the
input $a^{2^{k+1}+\ell\cdot 2^{k-\alpha}}\!$\@. This only requires
to repeat the loop of the length~$\ell_i$ beginning and ending in
the state~$r_i$\@. Namely, we can make
$\frac{\ell}{\ell_i}\cdot 2^{k-\alpha}$ more iterations. Note that
$\ell= \lcm\set{\ell_1,\ldots,\ell_r}$ and hence $\ell$ is
divisible by~$\ell_i$, for each $i=1,\ldots,r$, and that
$\alpha\le k$, which gives that
$\frac{\ell}{\ell_i}\cdot 2^{k-\alpha}$ is an integer. Note also
that these $\frac{\ell}{\ell_i}\cdot 2^{k-\alpha}$ additional
iterations of the loop of the length~$\ell_i$ travel exactly
$(\frac{\ell}{\ell_i}\cdot 2^{k-\alpha})\times\ell_i =
\ell\!\cdot\!2^{k-\alpha}$
additional positions to the right. Thus, if \ensuremath{\mathcal{N}}~has an
accepting path for~$a^{2^{k+1}}\!$, it must also have an accepting
path for $a^{2^{k+1}+\ell\cdot 2^{k-\alpha}}$ (just by a
straightforward induction on~$r$)\@. Therefore, \ensuremath{\mathcal{N}}~accepts
$a^{2^{k+1}+\ell\cdot 2^{k-\alpha}}\!$\@. (Actually, \ensuremath{\mathcal{N}}~can
have many other paths for this longer input, but they cannot rule
out the accepting decision of the path constructed above\@.)
However, the input $a^{2^{k+1}+\ell\cdot 2^{k-\alpha}}$ should be
rejected, since
$2^{k+1}+ \ell\!\cdot\!2^{k-\alpha}=
2^{k+1}+ \gamma\!\cdot\!2^{\alpha}\!\cdot\!2^{k-\alpha}=
(2\!+\!\gamma)\cdot 2^k$,
where $\gamma$ is odd. This is a contradiction. So, \ensuremath{\mathcal{N}}~must
have at least $2^{k+1}$ states.
\end{proof}
The above argument cannot be extended to alternating automata:
using \quo{cooperation} of several computation paths running in
parallel, the number of states can be reduced {}from $2^{k+1}$
to~$O(k)$, even by the use of one-way \eps-free machines. We first
present a conceptually simpler realtime version.
\blemma{l:afa-even-odd}
For each $k\in\nnumber$, the language $\mathtt{EvenOdd}\yes(k)$ can be
recognized\,---\,and hence the promise problem $\mathtt{EvenOdd}(k)$ can
be solved\,---\,by an \rt\textsc{afa}~\ensuremath{\mathcal{A}}\ with $7k\!+\!2$ states.
\unskip\end{lemma}
\begin{proof}
Recall that
$\mathtt{EvenOdd}\yes(k)= \set{a^{m\cdot 2^{k}} \mid\,} %"such that" in the middle of a set: \set{ ... \st ... m \mbox{ is even}}$,
and hence the unary input string~$a^n$ is in this language if and
only if $n$~is an integer multiple of~$2^{k+1}$\@. In turn, this
holds if and only if $b_{n,k}= b_{n,k-1}= \cdots= b_{n,0}= 0$,
where $b_{n,i}$~denotes the $i$\xth~bit in the binary
representation of the number~$n$\@. (The bit positions are
enumerated {}from right to left, beginning with zero for the least
significant bit\@.) That is, written in binary, $n$~should be of
the form~$\gamma 0^{k+1}$, for some $\gamma\in\set{0,1}\str$\@.
Based on this, the \rt\textsc{afa}~\ensuremath{\mathcal{A}}\ counts the length of the input
modulo~$2^{k+1}$ and verifies, for the binary representation of
this length, that the last $k\!+\!1$ bits are all equal to zero.
The important detail is that \ensuremath{\mathcal{A}}~counts \emph{down}, starting
{}from zero. That is, along the input, the value of the counter
changes as follows:
\[\begin{array}{lcll}
\ell &:=& 0,\,2^{k+1}\!-\!1,\,2^{k+1}\!-\!2,\,\ldots,
\,2,\,1,\,0,\,2^{k+1}\!-\!1,\,\ldots
\end{array}\]
Thus, the counter $\ell$ is interpreted as the number of input
symbols not processed yet (and taken modulo~$2^{k+1}$)\@. The
machine~\ensuremath{\mathcal{A}}\ accepts if and only if this value is zero when it
halts at the end of the input.
Implemented deterministically, this counting requires $2^{k+1}$
states. However, an alternating automaton can use several
processes running in parallel. (See~\cite{RY14B} for a similar
idea\@.) Each parallel process will keep only \emph{one bit} of
the current value of~$\ell$\@. That is, for
$i\in\set{k,k\!-\!1,\ldots,0}$, we shall use the states
\quo{$i_{0}$} and~\quo{$i_{1}$} to indicate that $i$\xth\ bit in
the current value of~$\ell$ is cleared to zero or set to one,
respectively. This will hold for each \quo{proper} sequence of
existential guesses along each input. The correctness of this
guessing will be verified when \ensuremath{\mathcal{A}}\ halts at the end of the
input. (Actually, the alternating tree of computation paths will
be highly redundant; the same bit will be remembered by a huge
number of activated identical copies of $i_{0}$ or~$i_{1}$,
running in parallel\@.) Clearly, for each symbol~$a$ along the
input, we need to decrement the counter~$\ell$ by one and hence to
update properly the corresponding bits. This is implemented by the
use of some additional auxiliary states.
We are now ready to present the set of states for the
\rt\textsc{afa}~\ensuremath{\mathcal{A}}, together with their brief description:
\begin{description}
\ditem{$i_{x}$,} for $i=k,\ldots,0$ and $x\in\set{0,1}$\@: the
$i$\xth\ bit of~$\ell$ is~$x$
\ ($b_{\ell,i}=x$),
\ditem{$i_{x,\allone}$,} for $i=k,\ldots,1$ and $x\in\set{0,1}$\@:
the $i$\xth\ bit of~$\ell$ is~$x$ and, moreover, all bits to
the right are set to~$1$
\ ($b_{\ell,i}=x$ and $b_{\ell,j}=1$ for each $j<i$),
\ditem{$i_{x,\exzero}$,} for $i=k,\ldots,1$ and $x\in\set{0,1}$\@:
the $i$\xth\ bit of~$\ell$ is~$x$ and, moreover, at least one
bit to the right is set to~$0$
\ ($b_{\ell,i}=x$ and $b_{\ell,j}=0$ for some $j<i$),
\ditem{$i_{\exzero}$,} for $i=k,\ldots,2$\@: the $i$\xth\ bit
of~$\ell$ may be quite arbitrary, but at least one bit to the
right is set to~$0$
\ ($b_{\ell,j}=0$ for some $j<i$),
\ditem{$s\ini$\@:} the initial state.
\unskip\end{description}
All states of type $i_{0},\,i_{1},\,i_{\exzero}$ are existential,
whereas those of type
$i_{0,\allone},\lbr \,i_{1,\allone},\lbr \,i_{0,\exzero},\lbr
\,i_{1,\exzero},\lbr \,s\ini$
are universal. The states of type $i_{0},\,s\ini$ are accepting
and none of the remaining states is accepting.
Clearly, the total number of states is
$2(k\!+\!1) +2k +2k +(k\!-\!1) +1= 7k\!+\!2$\@.
Transitions in~\ensuremath{\mathcal{A}}\ are defined as follows:
\begin{description}
\ditem{$i_{x}\trs{a} i_{x,\exzero}\vee i_{1-x,\allone}$,} for
$i=k,\ldots,1$ and $x\in\set{0,1}$\@: The $i$\xth\ bit
of~$\ell$ is set to~$x$ if and only if either \iit{i}~the
$i$\xth\ bit of $\ell'=\ell\!-\!1$ is set to the same
value~$x$, provided that at least one bit to the right
in~$\ell'$ is set to zero, that is, $b_{\ell',j}=0$ for some
$j<i$, or \iit{ii}~the $i$\xth\ bit of~$\ell'$ is flipped to
the complementary value~$1\!-\!x$, provided that all bits the
right in~$\ell'$ are set to one, that is, $b_{\ell',j}=1$ for
each $j<i$\@. Therefore, in the state~$i_{x}$, branching
existentially while reading the next input symbol~$a$,
\ensuremath{\mathcal{A}}~guesses between switching to $i_{x,\exzero}$ or
to~$i_{1-x,\allone}$\@.
\ditem{$0_{x}\trs{a} 0_{1-x}$,} for $x\in\set{0,1}$, a special
case of~$i_{x}$ for $i=0$\@: Transitions for the rightmost bit
are simplified, since there is no bit position with $j<i$ and
hence we do not need to use
$0_{0,\allone},\,0_{1,\allone},\lbr \,0_{0,\exzero},\,0_{1,\exzero}$\@.
Therefore, in~$0_{x}$, the machine reads the next input
symbol~$a$ deterministically, by switching to~$0_{1-x}$\@.
\ditem{$i_{x,\allone}\trs{\eps} i_{x}\wedge (i\!-\!1)_{1}\wedge
\cdots\wedge 0_{1}$,}
for $i=k,\ldots,1$ and $x\in\set{0,1}$\@: In the auxiliary
state~$i_{x,\allone}$, the machine~\ensuremath{\mathcal{A}}\ has to verify the bit
setting at the corresponding positions in~$\ell'$, namely,
that $b_{\ell',i}=x$ and that $b_{\ell',j}=1$ for each
$j<i$\@. Thus, branching universally with \eps-transitions for
each bit position $j\le i$, it splits into parallel processes
$i_{x},(i\!-\!1)_{1},\ldots,0_{1}$\@.
\ditem{$i_{x,\exzero}\trs{\eps} i_{x}\wedge i_{\exzero}$,} for
$i=k,\ldots,2$ and $x\in\set{0,1}$\@: The role of the
auxiliary state~$i_{x,\exzero}$ is similar to that
of~$i_{x,\allone}$, verifying the bit setting in~$\ell'$\@.
This time we verify that \iit{i}~$b_{\ell',i}=x$ and that
\iit{ii}~$b_{\ell',j}=0$ for at least one $j<i$\@. In order
not to mix universal and existential decisions, we delay the
condition~\iit{ii} until the next computation step, and hence
\ensuremath{\mathcal{A}}~branches universally in~$i_{x,\exzero}$, splitting into
two parallel processes $i_{x}$ and~$i_{\exzero}$, by the use
of \eps-transitions.
\ditem{$1_{x,\exzero}\trs{\eps} 1_{x}\wedge 0_{0}$,} for
$x\in\set{0,1}$, a special case of~$i_{x,\exzero}$ for
$i=1$\@: Transitions are simplified, since we can work
without~$1_{\exzero}$\@. Therefore, in~$1_{x,\exzero}$, the
machine splits universally into $1_{x}$ and~$0_{0}$, using
\eps-transitions.
\ditem{$i_{\exzero}\trs{\eps} (i\!-\!1)_{0}\vee \cdots\vee \,0_{0}$,}
for $i=k,\ldots,2$\@: In the auxiliary state~$i_{\exzero}$, we
verify that $b_{\ell',j}=0$ for at least one $j<i$\@.
Therefore, branching existentially with \eps-transitions,
\ensuremath{\mathcal{A}}~guesses which bit is set to zero, switching to~$j_{0}$,
for some $j<i$\@.
\ditem{$s\ini\trs{a} k_{1,\allone}$\@:} At the very beginning,
$\ell=n$\@. Therefore, for each accepted input~$a^n$, the
initial value of~$\ell$ must be an integer multiple
of~$2^{k+1}$, with the last $k\!+\!1$ bits all equal to zero.
Thus, after reading the first input symbol~$a$, we should get
$\ell'=\ell\!-\!1$ in which all these bits are set to one. For
this reason, in the initial state, the machine reads the first
input symbol~$a$ by switching deterministically
to~$k_{1,\allone}$\@.
\unskip\end{description}
It is easy to see {}from the above transitions that \ensuremath{\mathcal{A}}\ is a
realtime machine: After reading an input symbol, a computation
path can pass through at most two \eps-transitions in a row, after
which it gets to a state~$i_{x}$, for some $i=k,\ldots,0$ and some
$x\in\set{0,1}$\@. In~$i_{x}$, the machine waits for a next input
symbol.
To show that \ensuremath{\mathcal{A}}\ indeed recognizes $\mathtt{EvenOdd}\yes(k)$, we need to
prove the following claim.
\begin{trivlist}\item[]\textbf{Claim.}\hspace{\labelsep}\em
For each $\ell\ge 0$, each $i=k,\ldots,0$, and each
$x\in\set{0,1}$, the \rt\textsc{afa}~\ensuremath{\mathcal{A}}\ has an accepting alternating
subtree of computation paths rooted in the state~$i_{x}$ at the
input position~$\ell$ (measured {}from the end of the input) if and
only if the $i$\xth~bit of~$\ell$ is equal to~$x$\@.
\unskip\end{trivlist}
The claim is proved by induction on $\ell=0,1,2,\ldots$
First, let $\ell=0$, that is, \ensuremath{\mathcal{A}}~has already processed all input
symbols. Clearly, all bits of~$\ell$ are equal to zero, i.e.,
$b_{\ell,i}=0$ for each~$i$\@. Now, since there are no more input
symbols to be processed and there are no \eps-transitions starting
in the state~$i_{x}$, the machine \ensuremath{\mathcal{A}}\ must halt in~$i_{x}$\@.
Therefore, \ensuremath{\mathcal{A}}~has an accepting alternating subtree rooted
in~$i_{x}$ at the end of the input if and only if $i_{x}$~is an
accepting state. By definition of accepting states in~\ensuremath{\mathcal{A}}\ (the
states of type~$i_{0}$ are accepting but those of type~$i_{1}$ are
not), this holds if and only if $x=0$, which in turn holds if and
only if $x=0=b_{\ell,i}$, the $i$\xth~bit of~$\ell$\@.
Second, by induction, assume that the claim holds for
$\ell'=\ell\!-\!1$\@. Now, by definition of the transitions in the
state~$i_{x}$ for $i>0$, taking into account subsequent
\mbox{\eps-}transitions passing through auxiliary states, it can
be seen%
\footnote{With obvious modifications, the argument presented here
can be easily extended also for the case of $i=0$ (the rightmost
bit), which we leave to the reader\@.}
that \ensuremath{\mathcal{A}}\ has an accepting alternating subtree of computation
paths rooted in the state~$i_{x}$ at the input position~$\ell$ if
and only if either \iit{i}~\ensuremath{\mathcal{A}}~has an accepting alternating
subtree rooted in the state~$i_{x}$ at the input position
$\ell'=\ell\!-\!1$ and, moreover, the same holds for at least one
subtree rooted in a state~$j_{0}$ at the position~$\ell'$, for
some $j<i$, or \iit{ii}~\ensuremath{\mathcal{A}}~has an accepting alternating subtree
rooted in~$i_{1-x}$ at the input position~$\ell'$ and, moreover,
the same holds for all subtrees rooted in~$j_{1}$ at the
position~$\ell'$, for each $j<i$\@. However, by induction, using
$\ell'=\ell\!-\!1$, this statement holds if and only if either
\iit{i}~$b_{\ell',i}=x$ and, moreover, $b_{\ell',j}=0$, for some
$j<i$, or \iit{ii}~$b_{\ell',i}=1\!-\!x$ and, moreover,
$b_{\ell',j}=1$, for each $j<i$\@. Using $\ell=\ell'\!+\!1$, it is
easy to see that this holds if and only if $b_{\ell,i}=x$\@.
This completes the proof of the claim.
\medbreak
Now, recall that the machine~\ensuremath{\mathcal{A}}\ reads the first input
symbol~$a$ by switching deterministically {}from the initial state
to the state~$k_{1,\allone}$\@. After that, branching universally
with \eps-transitions for each bit position $j\le k$, the
computation splits into parallel processes
$k_{1},(k\!-\!1)_{1},\ldots,0_{1}$\@. Thus, using the above claim,
\ensuremath{\mathcal{A}}~accepts~$a^n$ with $n>0$ if and only if, in $\ell=n\!-\!1$,
the last $k\!+\!1$ bits are all equal to~$1$\@. Clearly, this
holds if and only if the last $k\!+\!1$ bits in~$n$ are all equal
to~$0$, that is, if and only if $n$~is and integer multiple
of~$2^{k+1}$\@. This gives that the language recognized by~\ensuremath{\mathcal{A}}\
agrees with $\mathtt{EvenOdd}\yes(k)$ on all inputs of length $n>0$\@.
For the special case of $n=0$, we have that
$a^0= a^{0\cdot 2^{k}}\in \mathtt{EvenOdd}\yes(k)$, since $m=0$ is even.
However, $a^0=\eps$ is accepted by~\ensuremath{\mathcal{A}}\ due to the fact that the
initial state~$s\ini$ is also an accepting state, which completes
the argument.
\end{proof}
By increasing slightly the number of states, we can replace the
realtime alternating machine {}from the above lemma by a one-way
\textsc{afa}\ not using any \mbox{\eps-}transitions:
\btheorem{t:afa-even-odd}
For each $k\ge 3$, the language $\mathtt{EvenOdd}\yes(k)$ can be
recognized\,---\,and hence the promise problem $\mathtt{EvenOdd}(k)$ can
be solved\,---\,by a \ow\textsc{afa}~$\ensuremath{\mathcal{A}}''$ with $11k\!-\!14$ states.
\unskip\end{theorem}
\begin{proof}
By inspection of the \rt\textsc{afa}~\ensuremath{\mathcal{A}}\ constructed in
Lemma~\ref{l:afa-even-odd}, it can be seen that \ensuremath{\mathcal{A}}\ can read an
input symbol only by transitions starting in the states of type
$i_{x},\,s\ini$\@. After reading, the computation path can pass
through at most two subsequent \eps-transitions until it gets to a
state of type~$i_x$, ready to read {}from the input again.
First, as an intermediate product, we replace~\ensuremath{\mathcal{A}}\ by an
equivalent machine~$\ensuremath{\mathcal{A}}'$ in which reading an input symbol is
always followed by \emph{exactly three} subsequent
\mbox{\eps-}transitions, along each computation path. This only
requires to delay artificially the moment when the automaton gets
to a state of type~$i_x$ by the use of some additional states,
namely, the following additional states:
\begin{description}
\ditem{$i'_{x},i''_{x}$,} for $i=k,\ldots,0$ and $x\in\set{0,1}$\@:
delay switching to the state~$i_{x}$ by one or two steps,
respectively\,---\,implemented by transitions
$i''_{x}\trs{\eps} i'_{x}\trs{\eps} i_{x}$,
\ditem{$0'''_{x}$,} for $x\in\set{0,1}$\@: delay switching to
the state~$0_{x}$ by three steps\,---\,implemented by
$0'''_{x}\trs{\eps} 0''_{x}$\@.
\unskip\end{description}
Taking also into account the states used by the original
machine~\ensuremath{\mathcal{A}}, the total number of states increases to
$(7k\!+\!2) +4(k\!+\!1) +2= 11k\!+\!8$\@.
Transitions for the \quo{original} states (cf.~construction in
Lemma~\ref{l:afa-even-odd}) are modified as follows:
\begin{itemize}
\item[] $i_{x}\trs{a} i_{x,\exzero}\vee i_{1-x,\allone}$, for
$i=k,\ldots,1$ and $x\in\set{0,1}$ \ (not modified),
\item[] $0_{x}\trs{a} 0'''_{1-x}$, for $x\in\set{0,1}$,
\item[] $i_{x,\allone}\trs{\eps} i''_{x}\wedge
(i\!-\!1)''_{1}\wedge \cdots\wedge 0''_{1}$,
for $i=k,\ldots,1$ and $x\in\set{0,1}$,
\item[] $i_{x,\exzero}\trs{\eps} i''_{x}\wedge i_{\exzero}$, for
$i=k,\ldots,2$ and $x\in\set{0,1}$,
\item[] $1_{x,\exzero}\trs{\eps} 1''_{x}\wedge 0''_{0}$, for
$x\in\set{0,1}$,
\item[] $i_{\exzero}\trs{\eps} (i\!-\!1)'_{0}\vee \cdots\vee \,0'_{0}$,
for $i=k,\ldots,2$,
\item[] $s\ini\trs{a} k_{1,\allone}$ \ (not modified)\@.
\unskip\end{itemize}
Since $\ensuremath{\mathcal{A}}'$ and~\ensuremath{\mathcal{A}}\ share the same set of accepting states, it
should be obvious that they recognize the same language.
Second, replace all \eps-transitions in~$\ensuremath{\mathcal{A}}'$ by transitions
reading the input symbol~$a$\@. That is, each transition of type
$s_1\trs{\eps} s_2$ is replaced by $s_1\trs{a} s_2$; the
transitions already reading input symbols do not change; nor does
the set of accepting states. This way we obtain a one-way
\textsc{afa}~$\ensuremath{\mathcal{A}}''$ without any \eps-transitions such that, for each
$n\ge 0$, it accepts the input~$a^{4n}$ if and only if $\ensuremath{\mathcal{A}}'$
accepts~$a^n$, i.e., if and only if $a^n\in \mathtt{EvenOdd}\yes(k)$\@.
The new machine~$\ensuremath{\mathcal{A}}''$ does not accept any input the length of
which is not an integer multiple of~$4$\@.
Therefore, with $11k\!+\!8$ states, $\ensuremath{\mathcal{A}}''$~recognizes the
language
\[\begin{array}{lcll}
& \set{ (a^4)^{m\cdot 2^{k}} \mid\,} %"such that" in the middle of a set: \set{ ... \st ... m \mbox{ is even} } =
\set{ a^{m\cdot 2^{k+2}} \mid\,} %"such that" in the middle of a set: \set{ ... \st ... m \mbox{ is even} } =
\mathtt{EvenOdd}\yes(k\!+\!2) \,.
\end{array}\]
Using the substitution $k'=k\!+\!2$, we thus get that
$\mathtt{EvenOdd}\yes(k')$ can be recognized by a one-way \eps-free
\textsc{afa}~$\ensuremath{\mathcal{A}}''$ with $11(k'\!-\!2) +8 =11k'\!-\!14$ states, for each
$k'\ge 3$\@.
\end{proof}
To show that the gap exhibited by Theorems \ref{t:nfa-even-odd}
and~\ref{t:afa-even-odd} is tight, we need an upper bound for
converting alternation into determinism on unary promise problems:
\blemma{l:unary-afa-to-dfa}
If a promise problem $\ensuremath{\mathtt{P}}=(\ensuremath{\mathtt{P}}\yes,\ensuremath{\mathtt{P}}\no)$ built over a unary
input alphabet can be solved by a \ow\textsc{afa}\ with $n$ states, it
can also be solved by a \ow\textsc{dfa}\ with $t(n)\le 2^n$ states.
\unskip\end{lemma}
\begin{proof}
As pointed out in~\cite[Sect.~5]{MePi01}, each \ow\textsc{afa}~\ensuremath{\mathcal{A}}\ with
$n$~states recognizing a unary language~\ensuremath{\mathtt{U}}\ can be replaced by
an equivalent \ow\textsc{dfa}~\ensuremath{\mathcal{D}}\ with $2^n$~states. This is a simple
consequence of the fact that \ensuremath{\mathcal{A}}~can be replaced by a
\ow\textsc{dfa}~\ensuremath{\mathcal{D}}\ with $2^n$ states recognizing~$\ensuremath{\mathtt{U}}\rvs$, the
reversal of the original language~\cite{CKS81}\@. (Actually, the
construction in~\cite{CKS81} works for Boolean automata\,---\,see
Footnote~\ref{ft:boolean}\,---\,which covers, as a special case,
the classical alternating automata as well\@.) However, for unary
languages, we have $\ensuremath{\mathtt{U}}\rvs=\ensuremath{\mathtt{U}}$\@.
Then, by the same reasoning as used in the proof of
Theorem~\ref{t:afa-to-dfa}, this trade-off easily extends to unary
promise problems.
\end{proof}
By the above lemma, the lower bound for solving $\mathtt{EvenOdd}(k)$ by
\ow\textsc{afa}s\ is $k\!+\!1$ states, since a \ow\textsc{afa}\ with less than
$k\!+\!1$ states would give us a \ow\textsc{dfa}\ with less than $2^{k+1}$
states for this promise problem, which contradicts
Theorem~\ref{t:nfa-even-odd}\@.
By combining Theorems \ref{t:nfa-even-odd}
and~\ref{t:afa-even-odd} with Lemma~\ref{l:unary-afa-to-dfa}, we
thus get:
\bcorollary{c:gap-even-odd}
The tight gap of succinctness between two-way \textsc{nfa}s\ and one-way
\eps-free \textsc{afa}s\ is asymptotically exponential on promise
problems, even for a unary input alphabet.
\unskip\end{corollary}
\bsection{Las Vegas Probabilistic Finite Automata}{s:LV}
In the case of language recognition, the gap of succinctness
between Las Vegas realtime \textsc{pfa}s\ and one-way \textsc{dfa}s\ can be at
most quadratic and this gap is tight~\cite{DHRS97,HS01}\@. Quite
surprisingly, despite of this fact and despite of the fact that
the existing gaps for language recognition are not be violated on
promise problems by classical models of automata
(cf.~Corollary~\ref{c:not-violated}), we shall show here that this
does not hold for the case of one-way Las Vegas probabilistic
versus one-way deterministic automata; by switching {}from
language recognition to promise problems, the quadratic gap
increases to a tight exponential gap.
\medbreak
For this purpose, we introduce a family of promise problems
solvable by Las Vegas \ow\textsc{pfa}s\ with a linear number of states and
success probability arbitrarily close to~$1$, but for which any
\ow\textsc{dfa}s\ require an exponential number of states.%
\footnote{Remark that, in~\cite{GY14A}, we argued the same
exponential separation by using some other families of promise
problems, but it appeared that some of these families can also
be solved by \ow\textsc{dfa}s\ with a linear number of states, and hence
they cannot be used as witness families for proving such
separation. More precisely, a single sentence statement given
just before Corollary~4 in~\cite{GY14A} is not correct\@.}
Consider a family of promise problems
$\mathtt{TRIOS}(n,r)$, for $n,r\in\nnumber$, such that
\[\begin{array}{lcll}
\mathtt{TRIOS}\yes(n,r) &=& \set{\, \#x_1x_1y_1\#x_2x_2y_2\cdots \#x_rx_ry_r \mid\,} %"such that" in the middle of a set: \set{ ... \st ...
x_1,y_1,\ldots,x_r,y_r\in\set{0,1}^n, \\
&& \phantom{\{\,}
\forall i\in\set{1,\ldots,r}\ \exists j\in\set{1,\ldots,n}: x_{i,j}<y_{i,j} \,} \,,\\[1.0ex]
\mathtt{TRIOS}\no(n,r) &=& \set{\, \#x_1y_1x_1\#x_2y_2x_2\cdots \#x_ry_rx_r \mid\,} %"such that" in the middle of a set: \set{ ... \st ...
x_1,y_1,\ldots,x_r,y_r\in\set{0,1}^n, \\
&& \phantom{\{\,}
\forall i\in\set{1,\ldots,r}\ \exists j\in\set{1,\ldots,n}: x_{i,j}>y_{i,j} \,} \,,
\end{array}\]
where $x_{i,j},y_{i,j}\in\set{0,1}$ denote the $j$\xth~bit in the
respective strings $x_i,y_i\in\set{0,1}^n$\@.
\btheorem{t:lv-trios}
For each $n,r\in\nnumber$, the promise problem $\mathtt{TRIOS}(n,r)$ can
be solved by a Las Vegas \ow\textsc{pfa}~\ensuremath{\mathcal{P}}\ with $4n\!+\!3$ states and
success probability $1-(\frac{n-1}{n})^r$\@.
\unskip\end{theorem}
\begin{proof}
The automaton~\ensuremath{\mathcal{P}}\ uses the following state set:
\[\begin{array}{lcll}
Q &=& \set{q\ini,q\acc,q\rej} \,\cup\, \set{ s_i \mid\,} %"such that" in the middle of a set: \set{ ... \st ... i\in\set{1,\ldots,n} } \,\cup\, \\
&& \set{ t_{i,0} \mid\,} %"such that" in the middle of a set: \set{ ... \st ... i\in\set{1,\ldots,2n} } \,\cup\, \set{ t_{i,1} \mid\,} %"such that" in the middle of a set: \set{ ... \st ... i\in\set{1,\ldots,n} } \,,
\end{array}\]
where $q\ini$~is the initial and also the \quo{don't know} state,
while $q\acc,q\rej$ are the accepting and rejecting states,
respectively.
Recall that we do not have to worry about the outcome of the
computation if the input is not of the form $(\#\set{0,1}^{3n})^r$\@.
Now, let~$\#x_iu_iv_i$, with $x_i,u_i,v_i\in\set{0,1}^n$, denote
the $i$\xth~input segment. Typically, for $i>1$, the processing
of the $i$\xth~segment is activated already somewhere in the
middle of the previous segment, by switching to the initial
state~$q\ini$\@. For $i=1$, the machine begins with the first
symbol~$\#$\@. For these reasons, in the state~$q\ini$, the
machine ignores all symbols along the input until it reads the
separator~$\#$, when it enters the state~$s_1$\@. This is
implemented by transitions $q\ini\trs{b} q\ini$, for each
$b\in\set{0,1}$, together with $q\ini\trs{\#} s_1$\@. (Unless
otherwise stated explicitly, all transitions throughout this proof
are \quo{deterministic}, i.e., with probability~$1$\@.)
Next, starting in the state~$s_1$, the automaton~\ensuremath{\mathcal{P}}\ moves along
the string~$x_i$ and chooses, with equal
probability~$\frac{1}{n}$, one of the symbols
$x_{i,1},\ldots,x_{i,n}\in \set{0,1}$ for subsequent verification.
This is implemented as follows. If \ensuremath{\mathcal{P}}~is in the state~$s_j$, for
$j\in\set{1,\ldots,n}$, and the next input symbol is
$b\in\set{0,1}$, then \ensuremath{\mathcal{P}}~reads this symbol by switching to the
state~$t_{1,b}$ with probability
$p_j= \frac{1}{p'_1\cdots p'_{j-1}}\cdot \frac{1}{n}$ \,and to the
state~$s_{j+1}$ with probability $p'_j= 1\!-\!p_j$\@. (As usual in
mathematical notation, we take $p'_1\cdots p'_{j-1}=1$ for
$j=1$\@.) The reader may easily verify that the sequence of
probabilities $p_1,\ldots,p_n$ ends with $p_n=1$, which gives
$p'_n=0$, and hence there is no need to introduce a
state~$s_{n+1}$\@. It is also easy to see that \ensuremath{\mathcal{P}}~gets {}from
the state~$s_1$ at the beginning of $x_i= x_{i,1}\cdots x_{i,n}$
to the state~$t_{1,x_{i,j}}$ positioned just behind the
symbol~$x_{i,j}$ with probability~$\frac{1}{n}$\@.
The next phase depends on whether \ensuremath{\mathcal{P}}~is in the state $t_{1,0}$
or in~$t_{1,1}$, that is, on whether the chosen bit $x_{i,j}$
is equal to $0$ or to~$1$\@.
For the case of $x_{i,j}=0$, the routine starting in the
state~$t_{1,0}$ is responsible to verify that $v_{i,j}=1$, which
corresponds to the case of $\#x_iu_iv_i= \#x_ix_iy_i$, with
$x_{i,j}<y_{i,j}$\@. Since \ensuremath{\mathcal{P}}~gets to~$t_{1,0}$ just after
reading the symbol~$x_{i,j}$, this only requires to examine the
bit that is placed exactly $2n$~positions to the right. This is
implemented by transitions $t_{j,0}\trs{b} t_{j+1,0}$, for each
$j\in \set{1,\ldots,2n\!-\!1}$ and each $b\in\set{0,1}$\@.
Finally, if $v_{i,j}=1$, then \ensuremath{\mathcal{P}}~accepts, using
$t_{2n,0}\trs{1} q\acc$\@. Otherwise, \ensuremath{\mathcal{P}}~switches back to the
initial state, by $t_{2n,0}\trs{0} q\ini$\@. In the accepting
state~$q\acc$, the machine just consumes the remaining part of the
input, using $q\acc\trs{a} q\acc$, for each $a\in\set{0,1,\#}$\@.
Conversely, in the initial state~$q\ini$, the machine proceeds to
examine the next input segment~$\#x_{i+1}u_{i+1}v_{i+1}$, by
transitions already described above. However, if \ensuremath{\mathcal{P}}~switches
to~$q\ini$ while processing the last input segment (for $i=r$), it
halts in~$q\ini$ at the end of the input. This is interpreted as a
\quo{don't know} answer.
The case of $x_{i,j}=1$ is very similar. This time the routine
starting in the state~$t_{1,1}$ verifies that $u_{i,j}=0$, which
corresponds to the case of $\#x_iu_iv_i= \#x_iy_ix_i$, with
$x_{i,j}>y_{i,j}$\@. This only requires to examine the bit placed
exactly $n$~positions to the right of the symbol~$x_{i,j}$\@. This
is implemented by transitions $t_{j,1}\trs{b} t_{j+1,1}$, for
each $j\in \set{1,\ldots,n\!-\!1}$ and each $b\in\set{0,1}$\@.
After that, if $u_{i,j}=0$, then \ensuremath{\mathcal{P}}~rejects, using
$t_{n,1}\trs{0} q\rej$\@. Otherwise, \ensuremath{\mathcal{P}}~switches back to the
initial state, by $t_{n,1}\trs{1} q\ini$\@. In the rejecting
state~$q\rej$, the machine consumes the rest of the input, using
$q\rej\trs{a} q\rej$, for each $a\in\set{0,1,\#}$\@. Conversely,
in the initial state~$q\ini$, the machine proceeds in the same way
as in the case of $x_{i,j}=0$; either it proceeds to the next
input segment, or else it halts in~$q\ini$ at the end of the
input, giving a \quo{don't know} answer.
\smallbreak
Summing up, if the $i$\xth~segment is of the form
$\#x_iu_iv_i= \#x_ix_iy_i$, with $x_{i,j}<y_{i,j}$ for some
$j\in\set{1,\ldots,n}$, then~\ensuremath{\mathcal{P}}, starting to process this
segment in the state~$q\ini$, accepts the entire input with
probability at least~$\frac{1}{n}$ (the exact value increases in
the number of bit pairs satisfying $x_{i,j}<y_{i,j}$) and proceeds
to the next segment in the state~$q\ini$ with probability at most
$1\!-\!\frac{1}{n}$\@. (The machine never rejects on such segment,
since $u_i=x_i$ and there are no bit pairs satisfying
$x_{i,j}>u_{i,j}$\@.)
Now, recall that each $w\in \mathtt{TRIOS}\yes(n,r)$ consists of $r$
segments having this form, and hence \ensuremath{\mathcal{P}}~gives the
\quo{don't know} answer with probability at most
$(1\!-\!\frac{1}{n})^r$ for~$w$\@. Therefore, $w$~is accepted with
probability at least $1-(1\!-\!\frac{1}{n})^r$ and never rejected.
\smallbreak
Conversely, if $i$\xth~segment is of the form
$\#x_iu_iv_i= \#x_iy_ix_i$, with $x_{i,j}>y_{i,j}$ for some
$j\in\set{1,\ldots,n}$, then, in the course of processing this
segment, \ensuremath{\mathcal{P}}~rejects the entire input with probability at
least~$\frac{1}{n}$ and proceeds to the next segment with
probability at most $1\!-\!\frac{1}{n}$\@. (Here the machine never
accepts, since $v_i=x_i$ and there are no bit pairs with
$x_{i,j}<v_{i,j}$\@.)
But then each $w\in \mathtt{TRIOS}\no(n,r)$, consisting of $r$ segments of
this form, is rejected with probability at least
$1-(1\!-\!\frac{1}{n})^r$ and never accepted, by the same
reasoning as for $\mathtt{TRIOS}\yes(n,r)$\@.
\end{proof}
It is obvious that, for arbitrarily small fixed constant
$\eps>0$, by taking $r= \ceil{\log\eps/\log(\frac{n-1}{n})}$, we
shall obtain a family of promise problems, namely,
$\mathtt{TRIOS}'_{\eps}(n)$, for $n\in\nnumber$, where
$\mathtt{TRIOS}'_{\eps}(n)= \mathtt{TRIOS}(n,\ceil{\log\eps/\log(\frac{n-1}{n})})$,
such that the $n$\xth~member of this family can be solved by a Las
Vegas \ow\textsc{pfa}\ with $4n\!+\!3$ states and success probability at
least \mbox{$1\!-\!\eps$}\@.
On the other hand, for each fixed $r\in\nnumber$, \ow\textsc{dfa}s\ need
an exponential number of states for solving $\mathtt{TRIOS}(n,r)$\@.
\btheorem{t:dfa-trios}
For each $n,r\in\nnumber$, any \ow\textsc{dfa}\ needs at least $2^n$
states for solving the promise problem $\mathtt{TRIOS}(n,r)$\@.
\unskip\end{theorem}
\begin{proof}
For contradiction, suppose that this promise problem can be solved
by a \ow\textsc{dfa}~\ensuremath{\mathcal{D}}\ using $\card{S}<2^n$ states. Let $s\ini$~be its
initial state, and let $\delta(q,u)\in S$ denote the state reached
by~\ensuremath{\mathcal{D}}\ {}from the state $q\in S$ by reading the string
$u\in\Sigma\str$\@. Clearly,
$\delta(q,uv)= \delta(\delta(q,u),v)$, for each state~$q$ and each
two strings~$u,v$\@.
Now, for each state $q\in S$, consider the enumeration of states
$\delta(q,\#u)$, running over all strings $u\in\set{0,1}^n$\@.
Since the number of states in~$S$ is smaller than the number of
strings in $\set{0,1}^n$, we have, for each $q\in S$, that there
must exist at least two different strings $u,v\in\set{0,1}^n$ with
$\delta(q,\#u)= \delta(q,\#v)$\@. Using some lexicographic
ordering for $\set{0,1}^n$, we can pick $u_q$ and~$v_q$, the first
pair of strings satisfying $u_q\ne v_q$ and
$\delta(q,\#u_q)= \delta(q,\#v_q)$\@. Moreover, without loss of
generality, we can assume that, for some bit position
$j\in \set{1,\ldots,n}$, we have $u_{q,j}< v_{q,j}$\@. (Here
$u_{q,j},v_{q,j}$ denote the $j$\xth~bit in the respective
strings~$u_q,v_q$\@.) Otherwise, by renaming, the roles of $u_q$
and~$v_q$ can be swapped\@.
Consider now the input
\[\begin{array}{lcll}
w\acc &=& \#u_{q_1}u_{q_1}v_{q_1} \#u_{q_2}u_{q_2}v_{q_2}\cdots
\#u_{q_r}u_{q_r}v_{q_r} \,, \mbox{ \ where}\\
q_1 &=& s\ini \,,\\
q_{i+1} &=& \delta(q_i,\#u_{q_i}u_{q_i}v_{q_i}) \,,
\mbox{ \ for $i= 1,\ldots,r$} .
\end{array}\]
It should be easily seen that, along the input~$w\acc$,
\,\ensuremath{\mathcal{D}}~passes the segment boundaries by the sequence of states
$q_1,q_2,\ldots,q_r,q_{r+1}$\@. More precisely, for each
$i= 1,\ldots,r$, the machine is in the state~$q_i$ at the
beginning of the $i$\xth\ segment $\#u_{q_i}u_{q_i}v_{q_i}$,
halting in the state~$q_{r+1}$ at the end of the last segment. It
is also obvious that $w\acc\in \mathtt{TRIOS}\yes(n,r)$, since, for each
segment $\#u_{q_i}u_{q_i}v_{q_i}$, we have a bit position
$j\in \set{1,\ldots,n}$ satisfying $u_{q,j}< v_{q,j}$\@.
Therefore, $q_{r+1}$~is an accepting state.
Next, consider the input
\[\begin{array}{lcll}
w\rej &=& \#v_{q_1}u_{q_1}v_{q_1} \#v_{q_2}u_{q_2}v_{q_2}\cdots
\#v_{q_r}u_{q_r}v_{q_r} \,.
\end{array}\]
Clearly, also here the computation starts in the state $q_1=s\ini$\@.
Now, using the fact that $\delta(q,\#u_q)= \delta(q,\#v_q)$ for
each state $q\in S$, we have that $q_{i+1}$ can also be expressed
as
\[\begin{array}{lcll}
q_{i+1} &=& \delta(q_i,\#u_{q_i}u_{q_i}v_{q_i}) =
\delta(\delta(q_i,\#u_{q_i}),u_{q_i}v_{q_i}) \\[1.0ex]
&=& \delta(\delta(q_i,\#v_{q_i}),u_{q_i}v_{q_i}) =
\delta(q_i,\#v_{q_i}u_{q_i}v_{q_i}) \,,
\end{array}\]
for each $i= 1,\ldots,r$\@. Therefore, along the input~$w\rej$,
\,\ensuremath{\mathcal{D}}~passes the segment boundaries by the same sequence of
states $q_1,q_2,\ldots,q_r,q_{r+1}$, halting in~$q_{r+1}$\@. But
then, since $q_{r+1}$~is an accepting state, \ensuremath{\mathcal{D}}~accepts~$w\rej$\@.
However, $w\rej\in \mathtt{TRIOS}\no(n,r)$, since, for each segment
$\#v_{q_i}u_{q_i}v_{q_i}$, we have a bit position
$j\in \set{1,\ldots,n}$ satisfying $v_{q,j}>u_{q,j}$\@.
Thus, each \ow\textsc{dfa}~\ensuremath{\mathcal{D}}\ with less than $2^n$ states accepts an
input that should be rejected, a contradiction.
\end{proof}
As a consequence of the above theorem, we get the asymptotically
optimal exponential size~$\Theta(2^n)$ for \ow\textsc{dfa}s\ solving
$\mathtt{TRIOS}(n,r)$; the construction of a \ow\textsc{dfa}\ with $O(2^n)$ states
is quite straightforward.
On the other hand, Theorem~\ref{t:lv-trios} guarantees $4n\!+\!3$
states for Las Vegas \ow\textsc{pfa}s\ solving $\mathtt{TRIOS}(n,r)$, and hence we
have an exponential lower bound for the gap between these two
models of automata. Finally, by Corollary~\ref{c:lv-to-dfa}, we
have an exponential upper bound, which gives:
\bcorollary{c:gap-lv}
The tight gap of succinctness between one-way \textsc{dfa}s\ and Las Vegas
one-way \eps-free \textsc{pfa}s\ is asymptotically exponential on promise
problems.
\unskip\end{corollary}
Due to Theorem~\ref{t:lv-trios}, we know that $\mathtt{TRIOS}(n,n^2)$ can
be solved by a Las Vegas \ow\textsc{pfa}\ with $4n\!+\!3$ states and
success probability $\sigma(n)\ge 1-(1\!-\!\frac{1}{n})^{n^2}\!$\@.
Thus, using the fact~\cite[Lm.~A.3.60+]{Hr05} that
$(1\!-\!\frac{1}{x})^x< \frac{1}{e}$ \,for each real $x>1$, the
probability of a \quo{don't know} outcome is at most
\[\begin{array}{lcll}
& 1\!-\!\sigma(n) \le \left(1\!-\!\frac{1}{n}\right)^{n\cdot n} <
\left(\frac{1}{e}\right)^n = \frac{1}{e^n} \,.
\end{array}\]
This Las Vegas \ow\textsc{pfa}\ can be converted to an equivalent
restarting one-way \textsc{pfa}, executing the procedure in an infinite
loop, by restarting the entire computation when the original
machine halts in the \quo{don't know} state at the end of the
input. (This can be easily achieved by using a single additional
state\@.) Therefore, we can obtain a restarting \ow\textsc{pfa}\ with
$4n\!+\!4$ states, solving $\mathtt{TRIOS}(n,n^2)$ exactly. The expected
number of the input tape sweeps can be bounded by
\[\begin{array}{lcll}
& \varrho(n) =
\sum_{i=1}^{\infty} i\!\cdot\! \sigma(n)\!\cdot\! (1\!-\!\sigma(n))^{i-1} \le
\sum_{i=1}^{\infty} i\!\cdot\! 1\!\cdot\! \left(\frac{1}{e^n}\right)^{i-1} =
\frac{1}{\textstyle \left(1\!-\!\frac{1}{e^n}\right)^2} =
\left(1\!+\!\frac{1}{e^n-1}\right)^2 \le
1\!+\!o(1) \,.
\end{array}\]
Thus, for the input of length~$\abs{w}$, the expected running time
is $(1\!+\!o(1))\!\cdot\!\abs{w}$\@.
\medbreak
Finally, the family of promise problems $\mathtt{TRIOS}(n,r)$, for
$n,r\in\nnumber$, does not give corresponding separations for
two-way machines, since already \tw\textsc{dfa}s\ can solve $\mathtt{TRIOS}(n,r)$
with $O(n)$ states, for each~$r$\@. This is done as follows.
Let~$\#x_iu_iv_i$, with $x_i,u_i,v_i\in\set{0,1}^n$, denote the
$i$\xth~segment along the input. Processing of this segment begins
with positioning the input head at the $n$\xth~bit of~$u_i$, by
moving $2n$ positions to the right {}from the symbol~$\#$\@. Then,
for $j=n,\ldots,1$ (in that order), the machine verifies that
$u_{i,j}=x_{i,j}$\@. (To move {}from the $j$\xth~bit of~$u_i$ to
the $j$\xth~bit of~$x_i$, just travel exactly $n$ positions to the
left; to move {}from the $j$\xth~bit of~$x_i$ to the
$(j\!-\!1)$\xst~bit of~$u_i$, just travel exactly $n\!-\!1$
positions to the right. The machine detects that it has tested all
$n$ bits {}from the fact that, after traveling exactly $n$ positions
to the left for the $j$\xth~bit of~$x_i$, it finds the symbol~$\#$
at this position, instead of zero or one\@.) In a similar way, the
machine can also search for a bit position~$j$ satisfying
$u_{i,j}<v_{i,j}$, this time iterating in ascending order, for
$j=1,\ldots,n$\@. If the $i$\xth~input segment passes the test
successfully, the machine proceeds to examine the next
segment~$\#x_{i+1}u_{i+1}v_{i+1}$ or, for $i=r$, it accepts.
\bsection{Two-Sided Bounded-Error Probabilistic Finite Automata}{s:bounded-error}
In Section~\ref{s:basics}, we have shown the limitations of Las
Vegas \textsc{pfa}s\@. One-sided error%
\footnote{We also cover the case of \emph{unbounded-error}, i.e.,
the machines with no bound on the error and so the error can be
arbitrarily close to~$1$ for some inputs\@.}
\textsc{pfa}s\ have similar limitations since they can be simulated by
nondeterministic or universal finite automata by removing the
probabilities: If a promise problem $\ensuremath{\mathtt{P}}=(\ensuremath{\mathtt{P}}\yes,\ensuremath{\mathtt{P}}\no)$ can
be solved by a \textsc{pfa}\ rejecting each $w\in\ensuremath{\mathtt{P}}\no$ with
probability~$1$ (only members of $\ensuremath{\mathtt{P}}\yes$ are accepted), then we
obtain an \textsc{nfa}\ {}from the one-sided error \textsc{pfa}\@. Conversely, if
each $w\in\ensuremath{\mathtt{P}}\yes$ is accepted with probability~$1$ (only members
of $\ensuremath{\mathtt{P}}\no$ are rejected), then we obtain a finite automaton
making only universal decisions (or an \textsc{nfa}\ for~$\ensuremath{\mathtt{P}}\no$) {}from
the one-sided error \textsc{pfa}\@. In this section, we show that we have
a different picture for \textsc{pfa}s\ with two-sided bounded-error.
\medbreak
First, we show that bounded-error \ow\textsc{pfa}s\ can be very succinct
compared to Las Vegas \tw\textsc{pfa}s, \tw\textsc{afa}s, or any other simpler
machines. In fact, we present a parallel result to that of
Ambainis and Yakary{\i}lmaz~\cite{AY12} but with bounded error
instead of exact acceptance.
To this aim, consider~$\ensuremath{\mathcal{U}}_p$, a \ow\textsc{pfa}\ with the unary input
alphabet $\Sigma=\set{a}$ and two states. Namely, let $s\ini$~be
the initial and also the only accepting state, and $s\rej$~be the
rejecting state. In the state~$s\ini$, the machine reads the
symbol~$a$ staying in~$s\ini$ with probability~$p$ and switching
to~$s\rej$ with the remaining probability $p'= 1\!-\!p$\@.
In~$s\rej$, the machine just consumes the remaining part of the
input, executing a single-state loop.
Now, let $\mathtt{UP}(p)$~be a promise problem such that
\[\begin{array}{lcll}
\mathtt{UP}\yes(p) &=& \set{ a^j \mid\,} %"such that" in the middle of a set: \set{ ... \st ... f_{\ensuremath{\mathcal{U}}_p}(a^j)\ge \frac{3}{4} } \,,\\[1.0ex]
\mathtt{UP}\no(p) &=& \set{ a^j \mid\,} %"such that" in the middle of a set: \set{ ... \st ... f_{\ensuremath{\mathcal{U}}_p}(a^j)\le \frac{1}{4} } \,.
\end{array}\]
Clearly, for each $p\in (0,1)$, the promise problem $\mathtt{UP}(p)$ can
be solved by a bounded-error one-way \eps-free \textsc{pfa}\ with only two
states, namely, by~$\ensuremath{\mathcal{U}}_p$\@.
On the other hand, it is easy to show that the number of states
required by any \tw\textsc{afa}\ for solving $\mathtt{UP}(p)$ \,(which covers all
simpler models of automata as well) increases when $p$
approaches~$1$\@: Since it is straightforward that the string
$a^j$ is accepted by~$\ensuremath{\mathcal{U}}_p$ with probability~$p^j$ for each
$j\ge 0$, we get $a^j\in \mathtt{UP}\yes(p)$ only for finitely many values
satisfying $j\le \flor{\log_p\frac{3}{4}}$, whereas $\mathtt{UP}\no(p)$
contains infinitely many strings satisfying
$j\ge \ceil{\log_p\frac{1}{4}}$\@. Let us denote these two
critical lengths as
\[\begin{array}{lcll}
A_p = \flor{\log_p\frac{3}{4}} &\mbox{ and }&
R_p = \ceil{\log_p\frac{1}{4}} \,.
\end{array}\]
Note also that $A_p\rightarrow +\infty$ as $p\rightarrow 1$\@.
\btheorem{t:up-exact}
For one-way/two-way deterministic/nondeterministic finite
automata, and each $p\in (0,1)$, the number of states that is
sufficient and necessary for solving the promise problem $\mathtt{UP}(p)$
is exactly $A_p\!+\!1$\@.
\unskip\end{theorem}
\begin{proof}
First, we can easily design a \ow\textsc{dfa}~\ensuremath{\mathcal{D}}\ (hence, any more
powerful model as well) solving $\ensuremath{\mathcal{U}}_p$ with $A_p\!+\!1$ states.
After the initial chain of $A_p\!+\!1$ accepting states,
responsible to accept~$a^j$ for $j\in \set{0,\ldots,A_p}$, the
automaton enters a rejecting state~$s\rej$, in which it executes a
single-step loop, consuming the rest of the input. By allowing
undefined transitions in~\ensuremath{\mathcal{D}}, the state~$s\rej$ can be
eliminated.
Second, before providing the matching lower bound, we need to
present the so-called \quo{$n\rightarrow n\!+\!n!$}
method~\cite{SHL65,CHR91,Ge91} (cf.~also the proof of
Theorem~\ref{t:nfa-even-odd})\@. To make the paper self-contained,
we include also a proof, for a simplified version of this method
that is sufficient for our purposes.
\begin{trivlist}\item[]\textbf{Claim.}\hspace{\labelsep}\em
Let \ensuremath{\mathcal{N}}~be a \tw\textsc{nfa}\ (or any less powerful model) with $n$
states and let $a^m$ be a unary string, with $m\ge n$\@. Then, if
\ensuremath{\mathcal{N}}~has a computation path traversing across~$a^m$ {}from left to
right, starting in some state~$q_1$ and ending in some~$q_2$, it
has also a computation path traversing
across~$a^{m+h\cdot m!}$, for each $h\ge 1$, starting and ending
in the same states $q_1$ and~$q_2$\@.
\unskip\end{trivlist}
The proof is based on the fact that \ensuremath{\mathcal{N}}\ uses only $n$~states,
but the sequence of states along the left-to-right traversal
across~$a^m$ is of length $m\!+\!1>n$ (including $q_1$ at the
beginning and~$q_2$ at the end)\@. Thus, in the course of this
traversal, some state must be repeated. Hence, \ensuremath{\mathcal{N}}~executes a
loop, by which it travels some $\ell\le m$ positions to the right.
Note that $m!/\ell$ is a positive integer. Hence, for each
$h\ge 1$, by using $h\!\cdot\!m!/\ell$ additional iterations of
this loop traveling $\ell$~positions, the machine traverses the
string~$a^{m+h\cdot m!}$, starting and ending in the same states
$q_1$ and~$q_2$\@. This completes the proof of the claim.
\medbreak
Now, let \ensuremath{\mathcal{N}}~be a \tw\textsc{nfa}\ (or any less powerful model) solving
$\mathtt{UP}(p)$ with at most $A_p$ states. Consider the inputs $a^m$
and~$a^{m+h\cdot m!}$, for $m=A_p$ and arbitrary $h\ge 1$\@.
Clearly, $a^m= a^{A_p}\in \mathtt{UP}\yes(p)$, and hence \ensuremath{\mathcal{N}}\ has at
least one accepting computation path for this input.
However, by the above claim, whenever this computation path
traverses {}from the left endmarker of~$a^m$ to the opposite
endmarker, starting in some state~$q_1$ and ending in some~$q_2$,
an updated version of this path can traverse
across~$a^{m+h\cdot m!}$ starting and ending in the same states
$q_1$ and~$q_2$\@. By symmetry, the same holds for right-to-left
traversals. Moreover, a U-turn starting and ending at the same
endmarker of~$a^m$ can obviously do the same on the
input~$a^{m+h\cdot m!}$\@. Thus, for each $h\ge 1$, \,\ensuremath{\mathcal{N}}~has a
computation path accepting~$a^{m+h\cdot m!}$\@. But then
\ensuremath{\mathcal{N}}~accepts infinitely many inputs, which contradicts the fact
that all inputs longer than~$R_p$ belong to $\mathtt{UP}\no(p)$ and should
be rejected. Thus, \ensuremath{\mathcal{N}}~must use at least $A_p\!+\!1$ states.
\end{proof}
A~lower bound for \tw\textsc{afa}s\ solving the promise problem $\mathtt{UP}(p)$
can be obtained by combining Corollary~\ref{c:not-violated},
Theorem~\ref{t:up-exact}, and the result of Geffert and
Okhotin~\cite{GO14}, stating that an arbitrary $n$-state \tw\textsc{afa}\
can be converted into an equivalent \ow\textsc{nfa}\ with
$2^{O(n\cdot\log n)}$ states. {}From this one can derive that the
number of states in any \tw\textsc{afa}\ solving $\mathtt{UP}(p)$ is at least
$\Omega(\frac{\log A_p}{\log\log A_p})$\@.
\bcorollary{c:gap-bounded-error}
There exists $\set{\mathtt{UP}(p) \mid\,} %"such that" in the middle of a set: \set{ ... \st ... p\in (0,1)}$, a family of unary
promise problems such that bounded-error one-way \eps-free \textsc{pfa}s\
with only two states are sufficient to solve all family, but the
number of states required by \tw\textsc{afa}s\ (or by any simpler
machines) to solve this family cannot be bounded by any constant.
\unskip\end{corollary}
The error bound~$\frac{1}{4}$ in the definition of~$\mathtt{UP}(p)$ can be
replaced with an arbitrarily small error; the same results will
follow by the use of the same reasoning.
\medbreak
Next, we present a separation result between bounded-error
\ow\textsc{pfa}s\ and \ow\textsc{dfa}s\ (and so \tw\textsc{afa}s, or any other model
capable of recognizing only regular languages)\@. For this
purpose, we use an idea given by Jibran and
Yakary{\i}lmaz~\cite{RY14A}\@. It is known that \tw\textsc{pfa}s\ can
recognize some nonregular languages, e.g.,
$\mathtt{EQ}= \set{a^nb^n \mid\,} %"such that" in the middle of a set: \set{ ... \st ... n\in\nnumber}$~\cite{Fr81}, with bounded
error but this requires exponential time~\cite{DS90}\@. It was
also shown that $\mathtt{EQ}$ can be recognized by a restarting \ow\textsc{pfa}\
for any error bound~\cite{YS10B}\@. If the given string~$a^mb^n$,
where $m,n\in\nnumber$, can be examined exponentially many times,
then the algorithm given in~\cite{YS10B} for~$\mathtt{EQ}$ can distinguish
between the cases of $m=n$ and $m\ne n$ with high probability.
Based on this, we introduce a new family of promise problems
$\mathtt{ExpEQ}(c)$, for integer $c\ge 3$\@:
\[\begin{array}{lcll}
\mathtt{ExpEQ}\yes(c) &=& \set{ (a^mb^n)^{3\cdot(2c^2)^{m+n}\cdot\ceil{\ln c}}
\mid\,} %"such that" in the middle of a set: \set{ ... \st ... m,n\in\nnumber, \,m=n } \,,\\
\mathtt{ExpEQ}\no(c) &=& \set{ (a^mb^n)^{3\cdot(2c^2)^{m+n}\cdot\ceil{\ln c}}
\mid\,} %"such that" in the middle of a set: \set{ ... \st ... m,n\in\nnumber, \,m\ne n } \,.
\end{array}\]
By~\cite{YS10B}, for each integer $c\ge 3$, there exists a
restarting \ow\textsc{pfa}~$\ensuremath{\mathcal{P}}_c$ recognizing~$\mathtt{EQ}$ such that, in one
pass along each given input~$a^mb^n$, the probability of giving an
erroneous decision (rejecting if the input should be accepted or
accepting if it should be rejected) is at least $c$~times smaller
than the probability of giving a correct decision. {}From this
moment on, one pass of~$\ensuremath{\mathcal{P}}_c$ along the given input~$a^mb^n$
will be called a \quo{round}\@. To be more exact, let $\mba$
and~$\mbr$ denote, respectively, the probability that
$\ensuremath{\mathcal{P}}_c$~accepts and rejects the input~$a^mb^n$ in a single round.
Then, by~\cite{YS10B}, the following holds:
\begin{equation}\begin{array}{lcll}
\mbr &\le& \frac{1}{c}\!\cdot\!\mba \,, &\mbox{ \ if $m=n$} \,,\\[1.0ex]
\mba &\le& \frac{1}{c}\!\cdot\!\mbr \,, &\mbox{ \ if $m\ne n$} \,.
\eequations{e:ar}
The remaining probability $\mbn= 1\!-\!\mba\!-\!\mbr$ represents
computations ending with a \quo{don't know yet} result, after
which $\ensuremath{\mathcal{P}}_c$ restarts another round. In addition, we know
{}from~\cite{YS10B} that
\begin{equation}\begin{array}{lcll}
\mba &=& \frac{1}{3\cdot(2c^2)^{m+n}} \,,
& \mbox{ \ independently of whether $m=n$} \,.
\eequations{e:a}
Note also that by combining (\ref{e:a}) and~(\ref{e:ar}) we get
$\mba>0$ for $m=n$ and $\mbr>0$ for $m\ne n$\@.
Finally, one of nice properties of these restarting \ow\textsc{pfa}s\ is
that they all use the same number of states, which is a constant
that does not depend on~$c$\@. (This is achieved by using
arbitrarily small transition probabilities\@.)
Now we are ready to construct a \ow\textsc{pfa}~$\ensuremath{\mathcal{P}}'_c$ for solving
$\mathtt{ExpEQ}(c)$\@. For the given input~$(a^mb^n)^t$, where
\begin{equation}\begin{array}{lcll}
t &=& 3\!\cdot\!(2c^2)^{m+n}\!\cdot\!\ceil{\ln c} \,,
\eequations{e:t}
$\ensuremath{\mathcal{P}}'_c$ simulates~$\ensuremath{\mathcal{P}}_c$ on~$a^mb^n$ but, each time
$\ensuremath{\mathcal{P}}_c$~restarts its computation {}from the very beginning
of~$a^mb^n$, the machine~$\ensuremath{\mathcal{P}}'_c$ proceeds to the next copy
of~$a^mb^n$ along its own input. More precisely, if $\ensuremath{\mathcal{P}}_c$~gives
the decision of \quo{acceptance} in the course of one round,
$\ensuremath{\mathcal{P}}'_c$~immediately accepts, by switching to an accepting
state~$s\acc$\@. Similarly, if $\ensuremath{\mathcal{P}}_c$~gives the decision of
\quo{rejection}, $\ensuremath{\mathcal{P}}'_c$~immediately rejects, switching to a
rejecting~$s\rej$\@. (In both these states, $\ensuremath{\mathcal{P}}'_c$~consumes the
rest of the input by executing single-state loops\@.) Otherwise,
$\ensuremath{\mathcal{P}}'_c$~gets to the end of the current copy of~$a^mb^n$ in a
state~$s\neu$, representing here a \quo{don't know yet} outcome.
That is, $\ensuremath{\mathcal{P}}'_c$~is ready to retry the entire procedure with the
next copy of~$a^mb^n$\@. Clearly, along the input~$(a^mb^n)^t$,
this can be repeated $t$~times, after which $\ensuremath{\mathcal{P}}'_c$~halts in the
state~$s\neu$\@.
Since the number of states in~$\ensuremath{\mathcal{P}}_c$ is a constant not depending
on~$c$, it is obvious that $\ensuremath{\mathcal{P}}'_c$~also uses a constant number
of states that does not depend on~$c$\@.
Now, let $\mba_t$ and~$\mbr_t$ denote the respective probabilities
that $\ensuremath{\mathcal{P}}'_c$~accepts and rejects the input~$(a^mb^n)^t$, by
halting in the respective state $s\acc$ or~$s\rej$\@. The
remaining probability $\mbn_t= 1\!-\!\mba_t\!-\!\mbr_t$ represents
computations not deciding about acceptance or rejection, halting
in the state~$s\neu$\@.
Let us estimate~$\mbn_t$ first. By combining $\mbr\ge 0$ with
(\ref{e:a}), (\ref{e:t}), and the fact~\cite[Lm.~A.3.60+]{Hr05}
that $(1\!-\!\frac{1}{x})^x< \frac{1}{e}$ \,for each real $x>1$
(in that order), we get, independently of whether $m=n$, that
\[\begin{array}{lcll}
\mbn_t &=& \mbn^t =
(1\!-\!\mba\!-\!\mbr)^t \le
(1\!-\!\mba)^t =
\left( 1\!-\!\frac{1}{3\cdot(2c^2)^{m+n}} \right)^t =
\left( 1\!-\!\frac{1}{3\cdot(2c^2)^{m+n}} \right)^{ 3\cdot(2c^2)^{m+n}\cdot\ceil{\ln c} } \\
&<& \left( \frac{1}{e} \right)^{\ceil{\ln c}} \le
\frac{1}{c} \,.
\end{array}\]
Therefore, using~(\ref{e:ar}) and $\mba>0$, we can derive that the
input~$(a^mb^n)^t$ satisfying $m=n$ is accepted with probability
at least
\[\begin{array}{lcll}
\mba_t &=& \sum_{i=1}^t \mba\!\cdot\!\mbn^{i-1} =
\mba\!\cdot\!\frac{1-\mbn^t}{1-\mbn} =
\frac{\mba}{\mba+\mbr}\!\cdot\!(1\!-\!\mbn^t) \ge
\frac{\mba}{\mba\,+\,\mba/c}\!\cdot\!(1\!-\!\mbn^t) =
\frac{c}{c+1}\!\cdot\!(1\!-\!\mbn^t) \\[1.0ex]
&>& \frac{c}{c+1}\!\cdot\!(1\!-\!\frac{1}{c}) =
1\!-\!\frac{2}{c+1} \,.
\end{array}\]
Similarly, using~(\ref{e:ar}) and $\mbr>0$, the input~$(a^mb^n)^t$
satisfying $m\ne n$ is rejected with probability at least
\[\begin{array}{lcll}
\mbr_t &=& \sum_{i=1}^t \mbr\!\cdot\!\mbn^{i-1} =
\mbr\!\cdot\!\frac{1-\mbn^t}{1-\mbn} =
\frac{\mbr}{\mba+\mbr}\!\cdot\!(1\!-\!\mbn^t) \ge
\frac{\mbr}{\mbr/c\,+\,\mbr}\!\cdot\!(1\!-\!\mbn^t) =
\frac{c}{c+1}\!\cdot\!(1\!-\!\mbn^t) \\[1.0ex]
&>& \frac{c}{c+1}\!\cdot\!(1\!-\!\frac{1}{c}) =
1\!-\!\frac{2}{c+1} \,.
\end{array}\]
Summing up, the success probability is always above
$1\!-\!\frac{2}{c+1}$ and consequently the error probability is
always below~$\frac{2}{c+1}$\@. (It should be pointed out that the
standard \textsc{pfa}s\ do not use \quo{don't know} states. However, we
can declare, by definition, the state~$s\neu$ to be a rejecting
state. This may potentially change all \quo{don't know} answers to
errors, rejecting inputs that should be accepted. However, this
does not change the fact that $\mba_t> 1\!-\!\frac{2}{c+1}$ for
$m=n$, nor does it decrease~$\mbr_t$ for $m\ne n$\@.)
By taking $c= \max\set{3,\ceil{\frac{2}{\eps}}\!-\!1}$ for
arbitrarily small but fixed $\eps>0$, we obtain the error
probability $\frac{2}{c+1}\le \eps$, keeping the same constant
number of states for each~\eps\@. This gives:
\btheorem{t:gap-nonregular}
For each fixed $\eps>0$, there exists a promise problem solvable
by a \ow\textsc{pfa}\ with bounded error~\eps, using a constant number of
states that does not depend on~\eps, but there is no \ow\textsc{dfa}\
(hence, no other machine capable of recognizing only regular
languages) solving the same problem.
\unskip\end{theorem}
\begin{proof}
We only need to show that $\mathtt{ExpEQ}(c)$ cannot be solved by any
\ow\textsc{dfa}, for no \mbox{$c\ge 3$}\@. For contradiction, let \ensuremath{\mathcal{D}}\ be
a \ow\textsc{dfa}\ solving $\mathtt{ExpEQ}(c)$, for some \mbox{$c\ge 3$}\@.
Without loss of generality, we assume that \ensuremath{\mathcal{D}}\ does not have
undefined transitions, and hence it always halts at the end of the
input. (Otherwise, we can define all missing transitions by
switching to a single new rejecting state, in which \ensuremath{\mathcal{D}}~scans the
rest of the input\@.) Let $n\ge 1$ denote the number of states
in~\ensuremath{\mathcal{D}}\@.
Consider now the unary string~$a^n$\@. Using the Claim presented
in the proof of Theorem~\ref{t:up-exact}, we see that if \ensuremath{\mathcal{D}}\
traverses across~$a^n$, starting in some state~$q_1$ and ending in
some~$q_2$, it will do the same also on the
string~$a^{n+h\cdot n!}$, for each $h\ge 1$\@. Clearly, the same
holds for traversals of~$b^n$ and~$b^{n+h\cdot n!}$\@.
Next, consider the input~$(a^nb^n)^t$, where
$t =3\!\cdot\!(2c^2)^{2n+2n!}\!\cdot\!\ceil{\ln c}$\@. If, on this
input, \ensuremath{\mathcal{D}}~halts in a state~$q'$, then, by the observation above,
\ensuremath{\mathcal{D}}~must halt in the same state~$q'$ also on the inputs
$(a^{n+n!}b^{n+n!})^t$ and~$(a^nb^{n+2n!})^t$\@. Therefore, \ensuremath{\mathcal{D}}\
accepts $(a^{n+n!}b^{n+n!})^t$ if and only if it
accepts~$(a^nb^{n+2n!})^t$\@. But this is a contradiction, since
the former string should be accepted by~\ensuremath{\mathcal{D}}\ while the latter
should be rejected. Consequently, there is no \ow\textsc{dfa}\ solving
$\mathtt{ExpEQ}(c)$\@.
\end{proof}
\bsection{Final Remarks}{s:final}
A~thorough study of promise problems can reveal several
interesting properties of computational models and give new
fundamental insights about them. In automata theory, promise
problems have mainly been used to show how quantum models can do
much better than the classical ones when compared to the case of
language recognition. In this paper, we initiated a systematic
work on promise problems for classical one-way finite automata
(deterministic, nondeterministic, alternating, and
probabilistic)\@. In this context, we have also shown that
randomness can do much better than other classical resources.
Promise problems can further be investigated for different
computational models and {}from different perspectives. Two-way
finite state machines and counter or pushdown automata models are
the first ones coming to the mind. Moreover, we believe that some
long-standing open problems, formulated for language recognition,
might be solved more easily in the case of promise problems.
\\
\noindent
\textbf{Acknowledgements.}
The authors thank Beatrice Palano for providing us a copy of~\cite{BMP14B}\@.
\bibliographystyle{plain
|
\section{Introduction}
\subsection{Preliminaries on integer additive set-indexers}
For all terms and definitions, not defined in this paper, we refer to \cite{FH} and for more about graph labeling, we refer to \cite{JAG}. Unless mentioned otherwise, all graphs considered here are simple, finite and have no isolated vertices.
The sum set of two sets $A$ and $B$, denoted by $A+B$, is defined as $A + B = \{a+b: a \in A, b \in B\}$. If at least one of two sets $A$ and $B$ is countably infinite, then their sum set $A+B$ will also be countably infinite. Hence, all sets mentioned in this paper are finite sets. We denote the cardinality of a set $A$ by $|A|$. Using the concepts of the sum set of two sets, the notion of an integer additive set-indexer of a given graph $G$ is defined in \cite{GA} as follows.
Let $\mathbb{N}_0$ denote the set of all non-negative integers and $\mathcal{P}(\mathbb{N}_0)$ be its power set. An {\em integer additive set-indexer} (IASI, in short) of a given graph $G$ is an injective function $f:V(G)\to \mathcal{P}(\mathbb{N}_0)$ such that the induced function $f^{+}:E(G) \to \mathcal{P}(\mathbb{N}_0)$ defined by $f^{+} (uv) = f(u)+ f(v)$ is also injective. A graph $G$ which admits an integer additive set-indexer is called an {\em integer additive set-indexed graph} (IASI-graph).
An IASI is said to be a {\em $k$-uniform IASI} if $|f^{+}(e)| = k$ for all $e\in E(G)$. That is, a connected graph $G$ is said to have a $k$-uniform IASI if all of its edges have the same set-indexing number $k$.
The cardinality of the labeling set of an element (vertex or edge) of a graph $G$ is called the {\em set-indexing number} of that element.
The vertex set $V$ of a graph $G$ is defined to be {\em $l$-uniformly set-indexed}, if all the vertices of $G$ have the same set-indexing number $l$.
Let $f$ be an IASI defined on a graph $G$ and let $u, v$ be any two adjacent vertices in $G$. Two ordered pairs $(a,b)$ and $(c,d)$ in $f(u)\times f(v)$ are said to be {\em compatible} if $a+b=c+d$. If $(a,b)$ and $(c,d)$ are compatible, then we write $(a,b)\sim (c,d)$. Clearly, $\sim$ is an equivalence relation.
A {\em compatibility class} of an ordered pair $(a,b)$ in $f(u)\times f(v)$ with respect to the integer $k=a+b$ is the subset of $f(u)\times f(v)$ defined by $\{(c,d)\in f(u)\times f(v):(a,b)\sim (c,d)\}$ and is denoted by $\mathsf{C}_k$. Since $f(u)$ and $f(v)$ are finite sets, then each compatibility class $\mathsf{C}_k$ in $f(u)\times f(v)$ contains finite number of elements.
It is to be noted that no compatibility class in $f(u)\times f(v)$ can be non-empty. If a compatibility class $\mathsf{C}_k$ contains only one element, then it is called a {\em trivial class}. A compatibility class $\mathsf{C}_k$ that contains maximum number of elements are called a {\em maximal compatibility class}.
\begin{lemma}\label{L-CardCC}
\cite{GS0} For a compatibility class $\mathsf{C}_k$ in $f(u)\times f(v)$, we have $1\le |\mathsf{C}_k| \le \min\,(|f(u)|,\,|f(v)|)$.
\end{lemma}
A compatibility class which contain the highest possible number of elements is called {\em saturated class}. That is, the cardinality of a saturated class in $f(u)\times f(v)$ is $\min(|f(u)|,|f(v)|)$.
It is to be noted that all saturated classes in $f(u)\times f(v)$ are maximal compatible classes, but a maximal compatible class need not be a saturated class of $f(u)\times f(v)$. That is, the existence of a saturated class depends on the nature of elements in the set-labels $f(u)$ and $f(v)$.
Based on the relation between the set-indexing numbers of an edge and its end vertices in $G$, the following notion is introduced in \cite{GS2}.
A {\em strong IASI} is an IASI $f$ such that $|f^{+}(uv)|=|f(u)|\,|f(v)|$ for all $u,v\in V(G)$. A graph which admits a strong IASI may be called a {\em strong IASI graph}. A strong IASI is said to be {\em strongly uniform IASI} if $|f^{+}(uv)|=k$, for all $u,v\in V(G)$ and for some positive integer $k$.
\subsection{Arithmetic Integer Additive Set-Indexers}
By the term, an arithmetically progressive set, (AP-set, in short), we mean a set whose elements are in an arithmetic progression. In this context, since the set-labels of the elements of $G$ need to be AP-sets, we take the sets having at least three elements for labeling the vertices of a given graph $G$.
The common difference of the set-label of an element of a graph $G$ is called the {\em deterministic index} of that element. The {\em deterministic ratio} of an edge $e$ of $G$ is the ratio, greater than or equal to $1$, between the deterministic indices of its end vertices.
A study about the graphs whose elements are labeled by AP-sets, has been done in \cite{GS7} and proposed the following notions and results.
Let $f:V(G)\to \mathcal{P}(\mathbb{N}_0)$ be an IASI on $G$. For any vertex $v$ of $G$, if $f(v)$ is an AP-set, then $f$ is called a {\em vertex-arithmetic IASI} of $G$. For an IASI $f$ of $G$, if $f^+(e)$ is an AP-set, for all $e\in E(G)$, then $f$ is called an {\em edge-arithmetic IASI} of $G$. A graph that admits a vertex-arithmetic IASI (or an edge-arithmetic IASI) is called a {\em vertex-arithmetic IASI graph} (or an {\em edge-arithmetic IASI graph}).
An IASI is said to be an {\em arithmetic integer additive set-indexer} if it is both vertex-arithmetic and edge-arithmetic. That is, an arithmetic IASI of a given graph $G$ is an IASI $f$, under which the set-labels of all elements of $G$ are AP-sets. A graph that admits an arithmetic IASI is called an {\em arithmetic IASI graph}.
The admissibility of an arithmetic IASI by a graph is established in the following theorem.
\begin{theorem}\label{T-AIASI-g}
\cite{GS7} A graph $G$ admits an arithmetic IASI $f$ if and only if $f$ is a vertex arithmetic IASI and the deterministic ratio any edge of $G$ is a positive integer, which is less than or equal to the set-indexing number of its end vertex having smaller deterministic index.
\end{theorem}
\noindent In other words, if $v_i$ and $v_j$ are two adjacent vertices of $G$, with deterministic indices $d_i$ and $d_j$ respectively with respect to an IASI $f$ of $G$, where $d_i\le d_j$, then $f$ is an arithmetic IASI if and only if $d_j=k\,d_i$, where $k$ is a positive integer such that $1\le k \le |f(v_i)|$.
In this paper, we study the characteristics given graphs, the set-labels of whose vertices and edges are AP-sets, with certain properties.
\section{Isoarithmetic IASI of Graphs}
If two AP-sets have the same common difference $d$, then their sum set is also an AP-set with the same common difference $d$. In view of this property, we introduce the following notion.
\begin{definition}{\rm
Let $f$ be an arithmetic IASI defined on a given graph $G$. If all the elements of $G$ have the same deterministic index under $f$, then $f$ is said to be an {\em isoarithmetic IASI} of $G$. A graph which admits an isoarithmetic IASI is called an {\em isoarithmetic IASI graph}.}
\end{definition}
Note that if an IASI $f$ of a graph $G$ is an isoarithmetic IASI, then the set-labels of all elements of $G$ are AP-sets with the same common difference and the deterministic ratio of every edge of $G$ is $1$.
\begin{definition}{\rm
Let $f$ be an isoarithmetic IASI of a given graph $G$, under which $V(G)$ is $l$-uniformly set-indexed, then $f$ is called an $l$-uniform {\em $l$-uniform isoarithmetic IASI} of $G$.}
\end{definition}
In the following discussions, we study certain characteristics of isoarithmetic IASI graphs. The following theorem verifies the existence of isoarithmetic IASIs for given graphs.
\begin{theorem}
Every graph $G$ admits an isoarithmetic integer additive set-indexer.
\end{theorem}
\begin{proof}
Let $f$ be an IASI defined on a given graph $G$ such that, for any vertex $v_i$ of $G$, $f(v_i)$ is an AP-set with the same common difference $d$ , where $d>1$ is a non-negative integer. Then, $f^+(v_iv_j)$ is also an AP-set with the same common difference $d$, for all edges $v_iv_j\in E(G)$. Therefore, $f$ is an isoarithmetic IASI of $G$.
\end{proof}
The following result establishes the hereditary nature of the existence of an isoarithmetic IASI of a graph $G$.
\begin{proposition}\label{P-APSL0}
A subgraph of an isoarithmetic IASI graph $G$ admits an (induced) isoarithmetic IASI. That is, the existence of an isoarithmetic IASI is a hereditary property.
\end{proposition}
\begin{proof}
Let $H$ be a subgraph of the graph $G$. Let $f$ be an isoarithmetic IASI of $G$. Then, the restriction $f|_H$ of $f$ to $V(H)$ is an isoarithmetic IASI of $H$. Hence $H$ is also an isoarithmetic IASI graph.
\end{proof}
An interesting question that arises here is about the set-indexing number of edges of an isoarithmetic IASI graph. To proceed in this direction, we need the following result.
\begin{lemma}\label{L-SS-5a}
Let $A$ and $B$ be finite AP-sets of integers having the same common difference $d$. Then, $|A+B|=|A|+|B|-1$.
\end{lemma}
\begin{lemma}\label{L-SS-5b}
\cite{MBN} Let $A$ and $B$ be finite sets of integers with $|A|=k\ge 2, |B|=l\ge 2 $. If $|A+B|=k+l-1$, then $A$ and $B$ are arithmetic progressions with the same common difference.
\end{lemma}
\noindent Invoking the above lemma, we propose the following theorem.
\begin{theorem}\label{T-APSL}
Let $G$ be a graph with an arithmetic IASI $f$ defined on it. Then, $f$ is an isoarithmetic IASI on $G$ if and only if the set-indexing number of every edge of $G$ is one less than the sum of the set-indexing numbers of it end vertices.
\end{theorem}
\begin{proof}
Let $v_i$ and $v_j$ be two adjacent vertices on $G$. Then, $f(v_i)$ and $f(v_j)$ are two AP-sets with cardinalities $m$ and $n$ respectively. Since $f$ is an arithmetic IASI of $G$, $f^+(v_iv_j)$ is also an AP-set.
First, assume that $f$ is an isoarithmetic IASI on $G$. Then, $f(v_i)$ and $f(v_j)$ are AP-sets with the same common difference, say $d$. Then, the set-label of the edge $v_iv_j$ is the set $f^+(v_iv_j)$, which is also an AP-set with the same common difference $d$. Therefore, by Lemma \ref{L-SS-5a}, the set-indexing number of the edge $v_iv_j$ is $m+n-1$.
Conversely, assume that he set-indexing number of every edge of $G$ is one less than the sum of the set-indexing numbers of it end vertices. That is, for any edge $v_iv_j$ in $G$, we have $|f^+(v_iv_j)|=|f(v_i)|+|f(v_j)|-1$. Since $f$ is an arithmetic IASI of $G$, $|f(v_i)|\ge 3 ~~ \forall ~v_i\in V(G)$. Therefore, by Lemma \ref{L-SS-5b}, both $f(v_i)$ and $f(v_j)$ also have the same common difference that of $f^+(v_iv_j)$. Hence, $f$ is an isoarithmetic IASI of $G$.
\end{proof}
\noindent The following theorem is an immediate consequence of Theorem \ref{T-APSL}.
\begin{theorem}\label{T-APSL3}
Let $f$ be an arithmetic IASI defined on a given graph $G$ such that $V(G)$ is $l$-uniformly set-indexed. Then, $f$ is an isoarithmetic IASI of $G$ if and only if $G$ is a $(2l-1)$-uniform IASI graph.
\end{theorem}
\begin{proof}
Let $V(G)$ is $l$-uniformly set-indexed under an arithmetic IASI $f$. Then, we have $|f(v_i)|=|f(v_j)|=l$ for any two (adjacent) vertices of $G$. Then, by Theorem \ref{T-APSL}, $f$ is an isoarithmetic IASI of $G$ if and only if $|f^+(v_iv_j)=2l-1$ for every edge $v_iv_j$ in $G$.
\end{proof}
The following result addresses the question whether an isoarithmetic IASI could be a strong IASI.
\begin{proposition}\label{P-APSL2}
No isoarithmetic IASI defined on a given graph $G$ can be a strong IASI of $G$.
\end{proposition}
\begin{proof}
Let $f$ be an isoarithmetic IASI of a graph $G$. Then, the set-labels of the vertices of $G$ under $f$ are AP-sets with the same common difference $d$. If possible, let $f$ be a strong IASI. Then, by Theorem \ref{T-APSL}, we have $m+n-1=mn$. This condition holds only when $m=1$ or $n=1$, which is a contradiction to the fact that the set-labels of the elements of $G$ contain at least $3$ elements. Hence, $f$ is not a strong IASI of the graph $G$.
\end{proof}
In view of Proposition \ref{P-APSL2}, for any two adjacent vertices $v_i, v_j\in V(G)$, it can be seen that under an isoarithmetic IASI $f$ on $G$, some compatibility classes in $f(v_i)\times f(v_j)$, contain more than one element. Then, the question about the number of elements in various compatibility classes arises much interest. The following theorem discusses the number of elements in the compatibility classes of $f(v_i)\times f(v_j)$ in $G$.
\begin{theorem}\label{T-NCC}
Let $G$ be a graph which admits an isoarithmetic IASI, say $f$. Then, the number of saturated classes in the Cartesian product of the set-labels of any two adjacent vertices in $G$ is one greater than the difference between cardinality of the set-labels of these vertices. More over, exactly two compatibility classes, other than the saturated classes, have the same cardinality in the Cartesian product of the set-labels of these vertices.
\end{theorem}
\begin{proof}
Let $v_i$ and $v_j$ be two adjacent vertices in $G$. Also, let $|f(v_i)|=m$ and $|f(v_j)|=n$. Without loss of generality, let $m\ge n$. Then, by lemma \ref{L-CardCC}, the maximum cardinality of a compatible class in $f(v_i)\times f(v_j)$ is $n$.
Let $f(v_i)=\{a, a+d,a+2d,\ldots,a+(m-1)d\}$ and $f(v_j)=\{b, b+d,b+2d,\ldots,b+(n-1)d\}$, where $a$ and $b$ are two positive integers. Consider the set-label of the edge $v_iv_j$ defined by $f^{+}(v_iv_j)=f(v_i)+f(v_j)$. Then, $f^{+}(v_iv_j)=\{a+b,a+b+d,a+b+2d,\ldots,a+b+(m+n-2)d\}$.
By Lemma \ref{L-CardCC}, a compatibility class can have at most of $n$ elements. Let $r=a+b$. Then, the compatibility classes in $f(v_i)\times f(v_j)$ are given by,
\begin{eqnarray*}
\mathsf{C}_r & = & \{(a,b)\},\\
\mathsf{C}_{r+d} & = & \{(a+d,b),(a,b+d)\},\\
\mathsf{C}_{r+2d} & = & \{(a+2d,b),(a+d,b+d),(a,b+2d)\},\\
\mathsf{C}_{r+3d} & = & \{(a+3d,b),(a+2d,b+d),(a+d,b+2d),(a,b+3d)\},\\
........&...&......................................................\\
........&...&.......................................................\\
\mathsf{C}_{r+(m+n-3)d} & = & \{(a+(n-1)d, b+(m-2)d), (a+(n-2)d, b+(m-1)d)\},\\
\mathsf{C}_{r+(m+n-2)d} & = & \{(a+(n-1)d, b+(m-1)d)\}.
\end{eqnarray*}
\noindent Hence, the cardinality of different compatibility classes are,
\begin{eqnarray*}
|\mathsf{C}_r| & = & |\mathsf{C}_{r+(m+n-2)d}| = 1.\\
|\mathsf{C}_{r+d}| & = & |\mathsf{C}_{r+(m+n-3)d}| = 2.\\
|\mathsf{C}_{r+2d}| & = & |\mathsf{C}_{r+(m+n-4)d}| = 3.\\
|\mathsf{C}_{r+3d}| & = & |\mathsf{C}_{r+(m+n-5)d}| = 4.\\
|\mathsf{C}_{r+4d}| & = & |\mathsf{C}_{r+(m+n-6)d}| = 5.\\
.........&...&...........................\\
.........&...&...........................\\
|\mathsf{C}_{r+(n-2)d}| & = & |\mathsf{C}_{r+md}| = n-1.
\end{eqnarray*}
Then, each of the remaining compatibility classes $\mathsf{C}_{r+(n-1)d}, \mathsf{C}_{r+(n)d}, \ldots, \mathsf{C}_{r+(m-1)d}$ contains $n$ elements, which is the highest number of elements possible in a compatibility class $\mathsf{C_r}$. Hence, all these classes are saturated classes. Therefore, the number of saturated classes in $f(v_i)\times f(v_j)$ = $m+n-1-2(n-1) =m-n+1$.
Also, it can be noted from the above equations that there are exactly two compatibility classes, other than the saturated classes, have the same cardinality $p$, where $1\le p\le n-1$. This completes the proof.
\end{proof}
\begin{corollary}
Let $f$ be an isoarithmetic IASI defined on a graph $G$, under which $V(G)$ is $l$-uniformly set-indexed. Then, there is exactly one saturated class in the Cartesian product of the set-labels of any two adjacent vertices in $G$.
\end{corollary}
\begin{proof}
Let $f$ be an isoarithmetic IASI defined on a graph $G$, under which $V(G)$ is $l$-uniformly set-indexed. Then, for any two adjacent vertices $v_i$ and $v_j$ in $G$, in $|f(v_i)|=|f(v_j)=l$. By Theorem \ref{T-NCC}, the number of saturated classes in $f(v_i)\times f(v_j)$ is $|f(v_i)|-|f(v_j)+1 = 1$.
\end{proof}
Can an isoarithmetic IASI $f$ defined on a given graph $G$ be a uniform IASI of $G$? If so, what are the conditions required for $f$ to be a uniform IASI? The following theorem provides a solution to these questions.
\begin{theorem}\label{T-AUIASI1}
An isoarithmetic IASI of a $G$ is a uniform IASI if and only if $V(G)$ is uniformly set-indexed or $G$ is bipartite.
\end{theorem}
\begin{proof}
Let $f$ be an isoarithmetic IASI of a graph $G$. If $V(G)$ is $l$-uniformly set-indexed, then by Theorem \ref{T-APSL3}, $G$ is $(2l-1)$-uniform IASI. Now, assume that $V(G)$ is not uniformly set-indexed. Then, for at least one edge of $G$, say $v_iv_j$, $|f(v_i)|\neq |f(v_j)|$. Let $G$ is bipartite with bipartition $(X,Y)$. Label the vertices of $X$ by distinct $m$-element AP-sets having the common difference $d$ and the vertices of $Y$ by distinct $n$-element AP-sets with the same common difference $d$. Then, by Theorem \ref{T-APSL}, every edge of $G$ has the set-indexing number $m+n-1$. That is, $f$ is $(m+n-1)$-uniform IASI.
Conversely, assume that $f$ is an $r$-uniform IASI of a connected graph $G$. If $V(G)$ is uniformly set-indexed, the proof is complete. Hence, assume that $V(G)$ is not uniformly set-indexed. Since $G$ is connected, there exist a unique pair of distinct positive integers $m$ and $n$ such that $r=m+n-1$ and every edge of $G$ has one vertex with set-indexing number $m$ and other end vertex with set-indexing number $n$. Let $X$ and $Y$ be the sets of all vertices of $G$ with set-indexing number $m$ and $n$ respectively. Let $v_i\in X$. Then, $v_iv_j\in E(G)\implies f^+(v_iv_j)=m+n-1\implies v_j\in Y$. Similarly, for $v_j\in Y, v_jv_k\in E(G)\implies f^+(v_kv_j)=m+n-1\implies v_k\in X$. Therefore, $(X,Y)$ is a bipartition of $G$.
\end{proof}
In view of Theorem \ref{T-AUIASI1}, it is natural to enquire whether an isoarithmetic IASI of a disconnected graph $G$ can be a uniform IASI and to determine the conditions, if exist, required for an isoarithmetic IASI of such a graph $G$ to be a uniform IASI? Let us establish a solution to all these questions in the following theorem.
\begin{theorem}\label{T-AUIASI2}
An isoarithmetic IASI $f$ of a graph $G$ is an $r$-uniform IASI if and only if every component $G$ is either bipartite or its vertex set is $l$-uniformly set-indexed, where $l=\frac{1}{2}(r+1)$.
\end{theorem}
\begin{proof}
Let $G$ be a graph with $q$ components, say $G_1, G_2,\ldots, G_q$ and $f$ be an isoarithmetic IASI defined on $G$. Let $f$ be an $r$-uniform IASI on $G$. Since each $G_i$ is a subgraph of $G$, by Proposition \ref{P-APSL0}, a restriction $f_i$ of $f$ to $V(G_i)$ induces an isoarithmetic IASI on $G_i$, which is also an $r$-uniform IASI on $G_i$. Since $G_i$ is a connected graph, by Theorem \ref{T-AUIASI1}, $G_i$ is a bipartite graph or $V(G_i)$ is $l$-uniformly set-indexed.
Conversely, assume that every component $G$ is either bipartite or its vertex set is $l$-uniformly set-indexed. If the vertex sets of all components of $G$ are $l$-uniformly set-indexed, then $V(G)$ will also be $l$-uniformly set-indexed. Then by Theorem \ref{T-AUIASI1}, the isoarithmetic IASI $f$ will be a uniform IASI of $G$. If for a component $G_i$ of $G$, $V(G_i)$ is not uniformly indexed, then $G_i$ is a bipartite graph with bipartition $(X_i,Y_i)$. We can label the vertices in $X_i$ by distinct AP-sets having $m_i$ elements and the common difference $d>1$ and label the vertices in $Y_i$ by distinct AP-sets having $n_i$ elements and the same common difference $d$, where $m_i, n_i\ge 3$ are the positive integers such that $m_i+n_i-1=r$. Then, the corresponding IASI, say $f_i$, is an $r$-uniform IASI of $G_i$. Label all the vertices of every component of $G$ by distinct AP-sets having the same common difference $d$, as explained above, according to whether it is bipartite or not. Then, the function $f:V(G)\to \mathcal{P}(\mathbb{N}_0)$ defined by $f(v)=f_i(v)$, if $v\in V(G_i)$ is an isoarithmetic IASI of $G$, which is a uniform IASI of $G$.
\end{proof}
Now that we have discussed the characteristics of arithmetic IASIs of certain graphs, all of whose elements have the same deterministic indices, we now proceed to consider the graphs whose different vertices have different deterministic indices.
\section{Biarithmetic IASI graphs}
By Theorem \ref{T-AIASI-g}, a graph admits an arithmetic IASI if and only if the deterministic ratios of all its edges are positive integers greater than or equal to $1$. We have considered the case when the deterministic ratio of all edges of $G$ is $1$. For studying the remaining cases, we introduce the following notion.
\begin{definition}{\rm
An arithmetic IASI $f$ of a graph $G$, under which the deterministic ratio of each edge of $G$ is a positive integer greater than $1$ and less than or equal to the set-indexing number of the end vertex of $e$ having smaller deterministic index.}
\end{definition}
In other words, a biarithmetic IASI of a graph $G$ is an arithmetic IASI $f$ of $G$, for which the deterministic indices of any two adjacent vertices $v_i$ and $v_j$ in $G$, denoted by $d_i$ and $d_j$ respectively such that $d_i < d_j$, holds the condition $d_j=kd_i$ where $k$ is a positive integer such that $1< k \le |f(v_i)|$.
In general, all edges of $G$ may not have the same deterministic ratio. Hence, we introduce the following notion.
\begin{definition}{\rm
Let $f$ be a biarithmetic IASI defined on a graph $G$. If the deterministic ratio of every edge of $G$ is the same, say $k$, then $f$ is called an {\em identical biarithmetic IASI} of $G$ and $G$ is called an {\em identical biarithmetic IASI graph}. }
\end{definition}
The existence of a biarithmetic IASI for a given graph $G$ depends upon the cardinality of set-labels of vertices of $G$ and the adjacency between the vertices. The following result establishes the admissibility of biarithmetic IASI by a given graph.
\begin{proposition}
Every graph $G$ admits a biarithmetic IASI.
\end{proposition}
The above result can be verified by taking the elements and cardinalities of the set-labels properly so that the conditions on the deterministic indices of the elements of $G$, as mentioned in \ref{T-AIASI-g}, are fulfilled.
An identical biarithmetic IASI may not exist for every graph $G$. The following theorem discusses the conditions required for a graph $G$ to admit an identical biarithmetic IASI.
\begin{theorem}\label{T-AIIBA1}
A graph $G$ admits an identical biarithmetic IASI if and only if it is bipartite.
\end{theorem}
\begin{proof}
Let $G$ be a bipartite graph having a bipartition $(X,Y)$ of $V(G)$. Now, define a function $f:V(G)\to \mathcal{P}(\mathbb{N}_0)$ in such a way that $f$ assigns distinct AP-sets having the same common difference, say $d>1$, to distinct vertices in $X$ and distinct AP-sets having the same common difference, say $k\,d$, to distinct vertices of $Y$, where$k=\min\{|f(u_i)|,u_i\in X\}$. Then, $f$ is an identical biarithmetic IASI of $G$.
Conversely, let $G$ admits an identical biarithmetic IASI. If possible, assume that $G$ is not a bipartite graph. Then, $G$ contains at least one odd cycle. For a positive integer $n=2i+1; i\ge 1$, let $C_{n}=v_1v_2v_3\ldots v_{2i+1}v_1$ be an odd cycle in $G$. Let $k$ be a positive integer such that $k\le |f(v_j)|$, where $1\le j \le n$ and $d$ be a positive integer, greater than $1$. Label the first vertex $v_1$ by an AP-set with common difference $d$. Now, label the vertices $v_2, v_3,\ldots,v_{2n}$ of $C_n$ by distinct AP-sets of non-negative integers in such a way that the edges connecting these vertices in $C_n$ have the deterministic ratio $k$. Then, the vertices of $C_n$ at the odd positions have the deterministic index $k^l\,d$, where $l$ is an even integer, positive or negative, and the vertices of $C_n$ at the even positions have the deterministic index $k^s\,d$, where $s$ is an odd integer, positive or negative.
Now, it remains to find a set-label for the vertex $v_{2i+1}$. If we choose an AP-set, which is not used for labeling the previous vertices, to label the vertex $v_{2i+1}$ in such a way that the edge $v_{2i}v_{2i+1}$ has the deterministic ratio $k$, then the deterministic index of the vertex $v_{2i+1}$ is $k^r\,d$, where $r$ is an even integer. Therefore, the deterministic ratio of the edge $v_{2i+1}v_1$ is greater than $k$. If we choose a set-label for $v_{2i+1}$ in such a way that the edge $v_{2i+1}v_1$ has the deterministic ratio $k$, then the deterministic index of the vertex $v_{2i+1}$ is $k^r\,d$, where $r$ is an odd integer. But, we know that $v_{2n}$ is $k^{r_1}\,d$, where $r_1$ is also an odd integer. Therefore, the deterministic ratio of the edge $v_{2i}v_{2i+1}$ can not be $k$. In both cases, $C_n$ do not admit an identical biarithmetic IASI. Then, by Remark \ref{R-IBIASISG}, $G$ can not be an identical biarithmetic IASI graph, which is a contradiction to the hypothesis. Hence, $G$ must be bipartite. This completes the proof.
\end{proof}
In the following discussion, we study certain characteristics of identical and non-identical biarithmetic IASI graphs.
Analogous to Proposition \ref{P-APSL0}, we propose following result on biarithmetic IASI graphs.
\begin{proposition}
Any subgraph of a biarithmetic IASI graph $G$ also admits a (induced) biarithmetic IASI. That is, existence of biarithmetic IASI is a hereditary property.
\end{proposition}
This proposition can be verified from the fact that an IASI of a graph $G$ induces an IASI to all its subgraphs. By the above proposition, it can be noted that an identical biarithmetic IASI of a graph $G$ also induces an identical biarithmetic IASI to any subgraph of $G$. This statement can also be re-stated as follows.
\begin{remark}\label{R-IBIASISG}{\rm
If a graph $G$ does not admit an identical biarithmetic IASI, then no supergraph of $G$ can be an identical biarithmetic IASI graph.}
\end{remark}
To learn about the set-indexing number of edges of a biarithmetic graph, we need the following theorem which estimates the set-indexing number of edges of a arithmetic IASI graph.
\begin{theorem}\label{T-AIASI1}
\cite{GS7} Let $G$ be a graph which admits an arithmetic IASI, say $f$ and let $v_i$ and $v_j$ be two adjacent vertices in $G$ with the deterministic indices $d_i$ and $d_j$, such that $d_i\le d_j$. Then, the set-indexing number of the edge $v_iv_j$ is $|f(v_i)|+k(|f(v_j)|-1)$, where $k\le |f(v_i)|$ is the deterministic ratio of the edge $v_iv_j$.
\end{theorem}
The set-indexing number of the edges a biarithmetic IASI graph can be written as a special case of Theorem \ref{T-AIASI1} as follows.
\begin{theorem}\label{T-AIASI1a}
Let $G$ be a graph which admits an arithmetic IASI, say $f$ and let $v_i$ and $v_j$ be two adjacent vertices in $G$ with the deterministic indices $d_i$ and $d_j$, such that $d_j=k\,d_i$, where $k$ is a positive integer such that $1<k\le |f(v_i)|$. Then, the set-indexing number of the edge $v_iv_j$ is $|f(v_i)|+k(|f(v_j)|-1)$.
\end{theorem}
Our next aim is to verify whether a biarithmetic IASI of a given graph can be a strong IASI of $G$. The following theorem explains a necessary and sufficient condition for a biarithmetic IASI of $G$ to be a strong IASI.
\begin{theorem}\label{T-AIASI2}
Let $G$ be a graph which admits a biarithmetic IASI, say $f$. Then, $f$ is a strong IASI of $G$ if and only if the deterministic ratio of every edge of $G$ is equal to the set-indexing number of its end vertex having smaller deterministic index.
\end{theorem}
\begin{proof}
Let $f$ be an arithmetic IASI of $G$. Let $v_i$ and $v_j$ are two adjacent vertices in $G$ and $d_i$ and $d_j$ be their deterministic indices under $f$. Without loss of generality, let $d_i<d_j$. Then, by Theorem \ref{T-AIASI1a}, the set-indexing number of the edge $v_iv_j$ is $|f(v_i)|+k(|f(v_j)|-1)$.
Assume that $f$ is a strong IASI. Therefore, $f^{+}(v_iv_j)=mn$. Then,
\begin{eqnarray*}
|f(v_i)|+k(|f(v_j)|-1) & = & |f(v_i)|\,|f(v_j)|\\
\implies k(|f(v_j)|-1) & = & |f(v_i)|\,(|f(v_j)|-1)\\
\implies k & = & |f(v_i)|.
\end{eqnarray*}
Conversely, assume that the deterministic indices $d_i$ and $d_j$ of two adjacent vertices $v_i$ and $v_j$ respectively in $G$, where $d_i<d_j$ such that $d_j=|f(v_i)|.d_i$. Assume that $f(v_i)=\{a_r = a+rd_i:0 \le r < |f(v_i)|\}$ and $f(v_j)=\{b_s=b+s\,k\,d_i:0\le s < |f(v_j)|\}$, where $k\le |f(v_i)|$.
Now, arrange the terms of $f^+(v_iv_j)=f(v_i)+f(v_j)$ in rows and columns as follows. For $b_s\in f(v_j), 0\le s < |f(v_j)|$, arrange the terms of $f(v_i)+b_s$ in $(s+1)$-th row in such a way that equal terms of different rows come in the same column of this arrangement. Then the common difference between consecutive elements in each row is $d_i$. Since $k=|f(v_i)|$, the difference between the final element of any row (other than the last row) and first element of its succeeding row is also $d_i$. That is, no column in this arrangement contains more than one element. Hence, all elements in this arrangement are distinct. Therefore, total number of elements in $f(v_i)+f(v_j)$ is $|f(v_i)|\,|f(v_j)|$. Hence, $f$ is a strong IASI.
\end{proof}
Invoking Theorem \ref{T-AIASI2}, the condition for an identical biarithmetic IASI to be a strong IASI is established in the following theorem.
\begin{theorem}\label{T-AIASI3}
An identical biarithmetic IASI of a graph $G$ is a strong IASI of $G$ if and only if one partition of $V(G)$ is $k$-uniformly set-indexed, where $k$ is the deterministic ratio of the edges of $G$.
\end{theorem}
\begin{proof}
Let $f$ be an identical arithmetic IASI of $G$. Then, by Theorem \ref{T-AIIBA1}, $G$ is bipartite. Let $(X,Y)$ be the bipartition of $G$, where $X=\{u_i, 1\le i \le r\}$ and $Y=\{v_j, 1\le j \le s\}$, $r+s=|V(G)|$. Also let $d_i$ be the deterministic index of the vertex $u_i \in X$ and $d'_j$ be the deterministic index of $v_j \in Y$. Without loss of generality, let $X$ is $k$-uniformly set-indexed, where $k$ is the deterministic ratio of the edges of $G$. Therefore, $|f(u_i)|=k~~ \forall u_i \in X$. Then, since $f$ is an identical biarithmetic IASI, we have $d'_j=|f(u_i)|\,d_i$ for every edge $u_iv_j \in V(G)$. Hence, by Theorem \ref{T-AIASI2}, $f$ is a strong IASI of $G$.
Conversely, assume that the identical biarithmetic IASI $f$ of $G$ is a strong IASI. Then, for every edge $e$ of $G$, the set-label of the end vertex of $e$ having smaller deterministic index must have exactly $k$ elements, where $k$ is the deterministic ratio of the edges of $G$. Let $X$ be the set of all these vertices having set-indexing number $k$. Since $f$ is an identical biarithmetic IASI, no vertices in $X$ can be adjacent to each other. Therefore, $(X, V-X)$ is a bipartition of $V(G)$, where $X$ is $k$-uniformly set-indexed. This completes the proof.
\end{proof}
In this context, it is interesting to check the existence of saturated classes or maximal compatibility classes and their cardinalities. The following theorem provides the necessary and sufficient condition for the existence of saturated classes and the number of saturated classes in the Cartesian product of the set-labels of two adjacent vertices.
\begin{theorem}\label{T-NSC-II}
Let $G$ be a graph that admits a biarithmetic IASI, say $f$. Let $v_i$ and $v_j$ be two adjacent vertices in $G$, where $v_i$ has the smaller deterministic index. Let $k$ be the deterministic ratio of the edge $v_iv_j$. Then, a compatible class in $f(v_i)\times f(v_j)$ is a saturated class if and only if $|f(v_i)|=(|f(v_j)|-1)\,k+r, ~r>0$. Also, number of saturated classes in $f(v_i)\times f(v_j)$ is $|f(v_i)|-(|f(v_j)|-1)\,k$. Moreover, for $1\le p \le n-1$, there are exactly $2k$ compatibility classes contain $p$ elements.
\end{theorem}
\begin{proof}
Let $f(v_i)=\{a, a+d_i,a+2d_i,\ldots,a+(m-1)d_i\}$ and $f(v_j)=\{b, b+kd_i,b+2kd_i,\ldots,b+(n-1)kd_i\}$, where $a$ and $b$ are positive integers. Consider the set-label of the edge $v_iv_j$ defined by $f^{+}(v_iv_j)=f(v_i)+f(v_j)$. Let $a+b=q$. Then, $f^{+}(v_iv_j)=\{q,q+d,q+2d,\ldots,q+[(m-1)+k(n-1)]d\}$.
Arrange the elements of $f(v_i)\times f(v_j)$ in rows and columns as follows. Write the elements of $f(v_i)+\{b_s\},~b_s\in f(v_j),~0\le s\le (n-1)$ in $(s+1)$-th row in such a way that equal terms in these rows come in the same column. Hence, we have $n$ rows containing $m$ elements in each row. Then, each column of this arrangement corresponds to a compatibility class and the number of elements in a column is the cardinality of the corresponding compatibility class. It is to be noted that the last $(m-k)$ elements of each row, except the last row, will be the first $m-k$ elements of the succeeding row. Hence, for $j\le n$, if $m>jk$, then last $(m-(j-1)k)$ elements of the first row will be the first $(m-(j-1)k)$ elements of the $j$-th row.
Assume that there are $r>0$ saturated classes in $f(v_i)\times f(v_j)$. Then, clearly $m>n$ and hence by Lemma \ref{L-CardCC}, a saturated class in $f(v_i)\times f(v_j)$ can have a maximum of $n$ elements. Then, $r$ columns of the arrangement contains $n$ elements. Therefore, the last $m+(n-1)k$ elements of the first row are the first $m+(n-1)k$ elements in the $n$-th (the last) row. That is, $m-(n-1)k=r$ or $m=(n-1)k+r$ where $r$ is a positive integer.
Conversely, assume that $m=(n-1)k+r, r>0$. From the above step, we note that $m-(n-1)k=r$ elements of the first row are common to all $n$ rows in the above row and column arrangement. That is, $r$ elements are common to all the $n$ rows of this arrangement. Hence, there are $r=m-(n-1)k$ saturated classes in $f(v_i)\times f(v_j)$. This completes the proof.
Since each column in the row and column arrangement, we mentioned above, corresponds to a compatibility class and the set-indexing number of an edge $v_iv_j$ is equal to the number of distinct compatibility classes in $f(v_i)\times f(v_j)$, by Theorem \ref{T-AIASI1a}, we have $|f(v_i)|+k(|f(v_j)|-1)$ columns in the arrangement. Note that the first $k$ elements of the first row and the last $k$ elements of the last row do not appear in any other rows. Therefore, the number columns having exactly one element is $2k$. That is, the number of compatibility classes with one element is $2k$.
Now remove these $2k$ columns from the arrangement. Then, in the revised arrangement, the first $k$ elements of the first two rows are the same and the last $k$ elements of the last two rows are the same and these element do not appear in any other rows. Therefore, the number of compatibility classes with $2$ elements is $2k$.
Proceeding like this, we have the number of compatibility classes having $p$ elements is $2k$, where $1\le p \le (n-1)$.
\end{proof}
It is clear that if $f(v_i)<f(v_j)$ in Theorem \ref{T-NSC-II}, then there is no saturated class in $f(v_i)\times f(v_j)$. Then, our next intention is to study the case when $|f(v_i)|=pk+q$, where $p$ and $q$ are non-negative integers such that $p<(|f(v_j)|-1)$ and $q<k$. Hence, we need to study further to determine the number of maximal compatibility classes and their cardinality. The following theorem provides the number of maximal compatibility classes in $f(v_i)\times f(v_j)$.
\begin{theorem}\label{T-NMCC-II}
Let $G$ be a graph that admits a biarithmetic IASI, say $f$. Let $v_i$ and $v_2$ be two adjacent vertices of $G$, where $v_i$ has the smaller deterministic index and $k\le|f(v_i)|$, be the deterministic ratio of the edge $v_iv_j$. If $|f(v_i)|=pk+q$, where $p,q$ are non-negative integers such that $p\le (|f(v_j)|-1)$ and $q<k$, then
\begin{enumerate}
\item[(i)] if $q=0$, then $(|f(v_j)|-p+1)k$ compatibility classes are maximal compatibility classes and contain $p$ elements.
\item[(ii)] if $q>0$, then $(|f(v_j)|-p-1)k+q$ compatibility classes are compatibility classes and contain $(p+1)$ elements.
\end{enumerate}
\end{theorem}
\begin{proof}
Let $v_i$ and $v_j$ be two adjacent vertices of $G$ with deterministic indices $d_i$ and $d_j$, such that $d_j=k.d_i$, where $k\le |f(v_i)|$ is the deterministic ratio of the edge $v_iv_j$. Also, let $|f(v_i)|=m$ and $|f(v_j)|=n$. Since $f$ is a biarithmetic IASI, we have $k\le m$. By Theorem \ref{T-APSL3}, the set-indexing number of the edge $v_iv_j$ is $m+k(n-1)$. Now, assume that $m=pk+q$, where $p,q$ are non-negative integers such that $1\le p\le (n-1)$ and $0\le q<k$. Arrange the elements of $f(v_i)\times f(v_j)$ in such a way that $f(v_i)+\{b_s\}$, where $b_s=b+sd\in A_j, 0\le s\le k(n-1)\}$, in $(s+1)$-th row and equal terms in these rows come in the same column.
\noindent {\em Case-1:} Let $q=0$. That is, $m= pk$. From the above arrangement, we observe that the last $k$ elements of the each row in the first half of the arrangement and the first $k$ elements of each row in the second half of this arrangement are common to exactly $p$ rows. Therefore, the cardinality of a maximal class in this case is $p$.
Moreover, as explained in Theorem \ref{T-NSC-II}, for $1\le j \le p-1$ there exist exactly $2k$ classes containing $j$ elements. Therefore, the total number of non-maximal compatibility classes is $2k(p-1)$. Therefore, the number of maximal compatibility classes is $m+k(n-1)-2k(p-1)=pk+(n-1)k-2pk=(n-p+1)k$.
\noindent {\em Case-2:} Let $q\ge 0$. That is, $m=pk+q$. From the above arrangement, we observe that the last $q$ elements of the each row in the first half of the arrangement and the first $k$ elements of each row in the second half of this arrangement are common to exactly $p+1$ rows. Therefore, the cardinality of a maximal class in this case is $p+1$.
Now, for $1\le j \le p$ there exist exactly $2k$ classes containing $j$ elements. Therefore, the total number of non-maximal compatibility classes is $2kp$. Therefore, the number of maximal compatibility classes is $m+k(n-1)-2kp = pk+q+(n-1)k-2pk=(n-p-1)k+q$.
That is, the number of maximal classes here is $(n-p-1)k+q$ and the number of elements in each of these maximal classes is $p+1$. This completes the proof.
\end{proof}
\section{Conclusion}
In this paper, we have discussed some characteristics of graphs which admit certain types of IASIs called isoarithmetic and biarithmetic IASIs. We have formulated some conditions for some graph classes to admit these types of arithmetic IASIs and discussed about certain properties characteristics of isoarithmetic and biarithmetic IASI graphs. Problems related to the characterisation of different biarithmetic IASI graphs are still open. The problems regarding the admissibility of certain graph operations and products which admit isoarithmetic and biarithmetic IASIs, characterisation of given graphs which admit biarithmetic IASIs, uniform and non-uniform, etc. are promising and worth studying. The IASIs which are vertex arithmetic, but not edge arithmetic can also be studied in detail.
The IASIs under which the vertices of a given graph are labeled by different standard sequences of non negative integers, are also note worthy. The problems of establishing the necessary and sufficient conditions for various graphs and graph classes to have certain IASIs still remain unsettled. All these facts highlight a wide scope for further studies in this area.
|
\section{Introduction}
\Label{s1}
It has been proved in \cite{KZ10} that if the boundary of a pseudoconvex domain of $\C^n$ has geometric ``type $F$", then there is an ``$f$-estimate" for the $\dib$-Neumann problem for $f=F^*(t^{-1})^{-1}$ where $F^*$ is the inverse function to $F$. The converse is also true (cf. \cite{KZ12}), apart from a loss of accuracy in the estimate which is in most cases negligeable. The succesful approach in establishing the equivalence between the $F$-type and the $f$-estimate consists in triangulating through a potential theoretical condition, namely, the ``$f$-property", that is, the existence of a bounded weight whose Levi-form grows with the rate of $f^2$ at the boundary. This generalizes former work by Kohn \cite{K79}, Catlin \cite{C83}, \cite{C87}, McNeal \cite{MN92} et alii.
What we prove here is that the $F$ type implies the $f$-estimate for the tangential system $\dib_b$; this is a generalization of Kohn \cite{K02}.
In greater detail, let $M\subset \C^n$ be a pseudoconvex manifold of hypersurface type
and $v$ or $u$ a form in $M$ of a certain degree $h$. We use the microlocal decomposition into wavelets $u=\sum_{k=1}^{+\infty}\Gamma_ku$ (cf. \cite{K02} proof of Theorem~6.1).
We consider
a submanifold $S\subset M$ of CR dimension $0$, and a real function $F$ satisfying $\frac F{d_S^2}\searrow0$ as the distance $d_S$ to $S$ decreases to $0$. We also use the notation Id for the identity of the complex tangent bundle $T^\C M=TM\cap iTM$. We assume that $M$ has type $F$ along $S$ in a neighborhood $U$ of point $z_o\in S$ in the sense that the Levi form $(c_{ij})$ of $M$ satisfies
$
(c_{ij})\simgeq \frac{F(d_S)}{d_S^2}\,\T{Id}.
$
Then, there is a bounded family of weights $\{\phi^k\}$ by the aid of which we get the estimate
of the $f$-norm by the Levi form $(c_{ij})$ of $M$ and $(\phi^k_{ij})$ of the $\phi^k$'s.
\bt
\Label{t1.1}
Let $M$ have type $F$ along $S$; then
\begin{equation}
\Label{1.0,1}
\begin{cases}
\begin{split}
\no{f(\Lambda)v}&\simleq \int_M(c_{ij})(\Lambda^{\frac12}v,\overline{\Lambda^{\frac12}v})\,dV+\sum_{k=1}^{+\infty}\int_M(\phi_{ij}^k)(\Gamma_kv,\overline{\Gamma_kv})\,dV
\\
&+\NO{v}_0,\quad \T{for any $v$ of degree $h\in[1,\dim_{CR}(M)]$},
\end{split}
\\
\begin{split}
\no{f(\Lambda)v}&\simleq \int_M\Big(\T{Trace}(c_{ij})\T{Id}-(c_{ij})\Big)(\Lambda^{\frac12}v,\overline{\Lambda^{\frac12}v})\,dV+\sum_{k=1}^{+\infty}\int_M\Big(\T{Trace}(\phi^k_{ij})\T{Id}-(\phi_{ij})\Big)\times
\\&
\times (\Gamma_kv,\overline{\Gamma_kv})\,dV+\NO{v}_0,\quad \T{for any $v$ of degree $h\in[0,\dim_{CR}(M)-1]$}.
\end{split}
\end{cases}
\end{equation}
\et
The proof is the content of Section~\ref{s2} below.
We denote by $u=u^++u^-+u^0$ the microlocal decomposition of $u$ (cf. \cite{K02} Section~2)
and also use the notation $Q^b $ for the energy $Q^b=\NO{\dib_b v}+\NO{\dib_b^* v}$, and $\mathcal H$ for the space of harmonic forms $\mathcal H=\ker\dib_b\cap\ker\dib_b^*$.
We apply the first of \eqref{1.0,1} for $v=u^+$, resp. the second for $v=u^-$, and plug into a basic estimate. We also use the elliptic estimate for $u^0$ and conclude
\bt
\Label{t1.1,5}
We have
\begin{equation}
\Label{f}
\NO{f(\Lambda)u}\simleq Q^b(u,\bar u)+\NO{u}_0,\quad \T{ for any $u$ of degree }h\in[0,\dim_{CR}(M)].
\end{equation}
\et
As it has been already said, \eqref{f} follows from \eqref{1.0,1} for the common range of degrees $h\in[1,\dim_{CR}(M)-1]$. As for the critical top and bottom degrees, we get the estimate for $u\in\mathcal H^\perp$ from the estimate in nearby degree from closed range of $\dib_b$ and $\dib_b^*$ (\cite{K02} proof of Theorem~7.3 p. 237).
Next, we prove a general basic weighted estimate twisted by a pseudodifferential operator $\Psi$, that is, \eqref{3.1} and \eqref{3.1bis} of Theorem~\ref{t3.1} below.
We have to mention that our formula is classical (cf. McNeal \cite{MN06}, \cite{S10}) when $\Psi$ is a function. A recent application, in which $\Psi$ is a family of cut-off, has been given in \cite{BPZ14} in the problem of the local regularity of the Green operator $G=\Box_b^{-1}$.
We choose a smooth orthonormal basis of $(1,0)$ forms $\om_1,...,\om_{n-1}$, supplement by a purely imaginary form $\gamma$ and denote the dual basis of vector fields by $\di_{\om_1},...,\di_{\om_{n-1}},T$. We define various constants $c^h_{ij}$'s as the coefficients of the commutator
$
[\di_{\om_i},\dib_{\om_j}]=c_{ij}^nT+\sum_{j=1}^{n-1}c_{ij}^h\di_{\om_h}-\sum_{j=1}^{n-1}\bar c_{ji}^h\dib_{\om_h}$; sometimes, we also write $c_{ij}$ instead of $c_{ij}^n$.
We use the notation $\Op$ for an operator of order smaller than $\Psi$ whose support is contained in a conical neighborhood of that of $\Psi$.
Combination of the $f$ estimate with the basic twisted estimate yields
\bt
\Label{t1.2}
Let $M$ have type $F$ along a CR manifold $S$ of CR dimension $0$ at $z_o$ and $U=U_t$ be suitably small.
For any form $v=u^+\in C^\infty_c(M\cap U)$ of degree $h\in[1,\dim_{CR}(M)-1]$ we have
\begin{equation}
\Label{1.0,3}
\begin{split}
||f(\Lambda)\Psi v||^2&\le \int(c_{ij})(\Psi T^{\frac12} v,\overline{\Psi T^{\frac12}v})\,dV+\sum_k\int(\phi^k)_{ij}(\Gamma_k\Psi v,\overline{\Gamma_k\Psi v})dV +t\NO{\Psi v}_0
\\
&\simleq Q_{\Psi}^b(v,\overline{v})+\NO{[\di_b,\Psi]\contrazione v}_0+\Big|\sum_h\int (c_{ij}^h)([\di_{\om_h},\Psi](v),\overline{\Psi v})\,dV\Big|
\\
&\qquad+\Big|\int [\di_b,[\dib_b,\Psi^2]](v,\overline{ v})dV\Big|+Q^b_{\Op}(v,\bar v)+\NO{\Op v}_0+\NO{\Psi v}_0.
\end{split}
\end{equation}
Here $Q^b_{\Psi}=\NO{\Psi\dib_bv}+\NO{\Psi\dib_b^*v}$.
\noindent
{\bf (ii)}\hskip0.2cm
The similar equation holds for $u^-$ in degree $[0,\dim_{CR}(M)-1]$ if we replace $(c_{ij})$, $(\phi^k_{ij})$ and $[\di_b,[\dib_b,\Psi^2]]$ by $-(c_{ij})+\sum_j c_{jj}\T{Id}$, $-(\phi^k_{ij})+\sum_j \phi_{jj}\T{Id}$ and $-[\di_b,[\dib_b,\Psi^2]]+\T{Trace}([\di_b,[\dib_b,\Psi^2]])\T{Id}$ respectively.
\noindent
{\bf (iii)} \hskip0.2cm
Taking summation of the estimate for $v=u^+, v=u^-$ together with the elliptic estimate for $v=u^0$, and using the closed range of $\dib_b$ and $\dib^*_b$ for the critical degrees we get for the full $u\in \mathcal H^\perp$ in degree $h\in [0,\dim_{CR}(M)]$
\begin{equation}
\Label{1.2}
\begin{split}
\NO{f(\Lambda)\Psi u}_0&\simleq Q^b_{\Psi}(u,\bar u)+\NO{[\di_b,\Psi]\contrazione u}_0+\Big|\int_M[\di_b,[\dib_b,\Psi^2]](u^+,\overline{u^+})\,dV\,\Big|
\\
& +\,\Big|\sum_h\int (c_{ij}^h)([\di_{\om_h},\Psi](u),\overline{\Psi u})\,dV\Big|+\Big|\int_M\Big(-[\di_b,[\dib_b,\Psi^2]](u^-,\overline{u^-})
\\
&\qquad+\T{Trace}([\di_b,[\dib_b,\Psi^2]])\T{Id}\Big)(u^-,\overline{u^-})\,dV\Big|+Q^b_{\Op}(u,\bar u)+\NO{\Op u}_0+\NO{\Psi u}_0.
\end{split}
\end{equation}
\et
The proof is just the superposition of the items (i) and (ii) of Theorem~\ref{t3.1} below. We have indeed, in Theorem~\ref{t3.1} (i) and (ii) a more general, weighted version of this estimate.
We give an application of the general twisted estimate in which $\Psi$ includes a cut-off $\eta$ and a differentiation of arbitrarily high order $s$ (such as $R^s$ of Section~\ref{s4} below).
To introduce it, we need the notion of superlogarithmic multipliers which are an obvious variant of the subelliptic multipliers (cf. \cite{K02} Definition~8.1). The crucial point in our discussion is that we consider vector multipliers $g=(g_j)$ and also require a more intense property in which energy is replaced by Levi form, that is, for any $\epsilon$, suitable $c_\epsilon$, and for an uniformly bounded family of weights $\{\phi^k\}$
\begin{equation}
\Label{1.0,4}
\NO{\log(\Lambda)g\contrazione v}\simleq\epsilon\Big(\int_M(c_{ij}(\Lambda^{\frac12}v,\overline{\Lambda^{\frac12}v})\,dV+\sum_{k=1}^{+\infty}\int_M(\phi_{ij}^k)(\Gamma_kv,\overline{\Gamma_kv})\,dV\Big)+c_\epsilon\NO{v}_0.
\end{equation}
We also require that the same estimate holds for $(c_{ij})$ and $(\phi^k_{ij})$ replaced by $-(c_{ij})+\T{Trace}(c_{ij})\T{Id}$ and $-(\phi^k_{ij})+\T{Trace}(\phi^k_{ij})\,\T{Id}$ respectively.
With this preliminary we have
\bt
\Label{t1.3}
Assume that there is a system of cut-off $\{\eta\}$ at $z_o$ such that $[\dib_b,\eta]$ and $[\di_b,[\dib_b,\eta]]$ are vector and matrix superlogarithmic multipliers respectively,
and $(c_{ij}^h)$ are subelliptic multipliers.
Then $G$ is regular at $z_o$.
\et
The proof is found in Section~\ref{s4}. We combine Theorem~\ref{t1.3} with \ref{t1.1}. This gives back the conclusion of \cite{BKZ14} (in a tangential version) which was in turn a generalization of \cite{K00}. It also provides a larger class of hypersurfaces for which $G$ is regular. Let $M$ be the ``block decomposed" hypersurface of $\C^n$ defined by $x_n=\sum_{j=1}^m h^{I^j}(z_{I^j},y_n)$ where $z=(z_{I^1},...,z_{I^m},z_n)$ is a decompostion of coordinates.
\bt
\Label{t1.4}
Assume that
\begin{equation}
\Label{1.4}
\begin{cases}
\T{ (a) $h^{I^j}$ has infraexponential type along a totally real $S^{I^j}\setminus\Gamma^{I^j}$ where $S^{I^j}$ is}
\\
\quad\T{ totally real in $\C^{I^j}\times \C_{z_n}$ and $\Gamma^{I^j}$ is a curve of $\C^{I^j}\times\C_{z_n}$ transversal to $\C^{I^j}\times\{0\}$,}
\\
\T{(b) $h^j_{z_j}$ are superlogarithmic multipliers},
\\
\T{(c) $c_{ij}^h$ are subelliptic multipliers}.
\end{cases}
\end{equation}
Then, we have local regularity of $G$ at $z_o=0$.
\et
In case of a single block $x_n=h^{I^1}$ we regain \cite{BPZ14} and \cite{K00}.
The proof is found in Section~\ref{s4} below.
\noindent
{\it Example}
Let
$$
(i)\quad
x_n=\sum_{j=1}^{n-1}e^{-\frac1{|z_j|^a}}e^{-\frac1{|x_j|^b}}\qquad\T{ for any $a\geq0$ and for $b<1$}.
$$
Then, \eqref{1.4} (a) is obtained starting from $h^j_{z_j\bar z_j}\simgeq \frac{e^{-\frac1{|x_j|^b}}}{|x_j|^2}$, that is, the condition of type $F_j:=e^{-\frac1{|x_j|^b}}$ along $S_j=\R_{y_j}\times\{0\}$. This yields the estimate of the $f$ norm for $f(t)=
\log^{\frac1b} (t)$; since $\frac1b>1$, this is superlogarithmic. \eqref{1.4} (b) follows from $|h^j_{z_j}|^2\simleq h^j_{z_j\bar z_j}$ which says that the $h^j_{z_j}$'s are not only superlogarithmic, but indeed $\frac12$-subelliptic, multipliers.
Finally, (c) follows from $c_{jj}^h\simleq c_{jj}$ (a consequence of the ``rigidity" of $M$) which shows that these constant are $\frac12$ subelliptic multipliers.
\eqref{1.4} is the ultimate step of a long sequence of criteria of regularity of $G$, not reduceable in one another, described by the hypersurface models below, in which $a>0$ and $0<b<1$,
\begin{itemize}
\item [(ii)] $x_n=\sum_{j=1}^{n-1}e^{-\frac1{|x_j|^b}}$ Kohn \cite{K02},
\item[(iii)] $x_n=e^{-\sum_{j=1}^{n-1}\frac1{|z_j|^a}}$ Kohn \cite{K00},
\item[(iv)] $x_n=e^{-\frac1{\sum_{j=1}^{n-1}|x_j|^a}}\Big(\sum_{j=1}^{n-1}e^{-\frac1{|x_j|^b}}\Big)$ Baracco-Khanh-Zampieri \cite{BKZ14},
\item[(v)] $x_n=\sum_{j=1}^{n-1}e^{-\frac1{|z_j|^a}}$ Baracco-Pinton-Zampieri \cite{BPZ13},
\item[(vi)] $x_n=\sum_{j=1}^{n-1}e^{-\frac1{|z_j|^a}}x_j^a$ Baracco-Pinton-Zampieri \cite{BPZ14}.
\end{itemize}
Thus, the degeneracy in our model (i) comes as the combination of those of (ii) with (v) (or (vi)).
\section{Estimate of the $f$-norm by the Levi form}
\Label{s2}
Let $M$ be a $C^\infty$ CR-manifold of $\C^n$ of hypersurface-type, $z_o$ a point of $M$, $U$ an open neighborhood of $z_o$.
Our setting being local, we can find a local CR-diffeomeorphism which reduces $M$ to a hypersurface of $TM+iTM$; therefore, it is not restrictive to assume that $M$ is a hypersurface of $\C^n$ from the beginning.
We choose a smooth orthonormal basis of $(1,0)$ forms $\om_1,...,\om_{n-1}$, supplement by a purely imaginary form $\gamma$ and denote the dual basis of vector fields by $\di_{\om_1},...,\di_{\om_{n-1}},T$. We also use the notation $\dib_b$ for the tangential CR-system. For a smooth real function $\phi$, we denote by
$(\phi_{ij})$ the matrix of the Levi form $\di_b\dib_b\phi$. Note that $\phi_{ij}$ differs from $\di_{\om_i}\dib_{\om_j}(\phi)$ because of the presence of the derivatives of the coefficients of the forms $\bar\di_{\om_j}$.
Let $(c_{i\bar j})_{i,j=1,...n-1}$ be the Levi-form $d\gamma|_{T^\C {M}}$ where $T^\C {M}=T{M}\cap iT{M}$.
Let $S\subset M$ be a submanifold of CR-dimension $0$, $d_S$ the Euclidean distance to $S$, and $f:\,\R^+\to\R^+$ a smooth monotonic increasing function such that $f\simleq t^{\frac12}$.
We use the notation $a_k$ for the constant $a_k:=f^{-1}(2^k)$ and $S_{a_k}$ for the strip $S_{a_k}:=\{z\in M:\,\,d_S(z)\le a_k\}$.
\bl
\Label{l2.1}
There is an uniformly bounded family of smooth weights $\{\phi^k\}$ with supp $\phi^k\subset S_{2a_k}$ whose Levi-form satisfies
\begin{equation}
\Label{2.1}
\di_b\dib_b\phi^k\simgeq \begin{cases}
f^2(2^k)&\T{on $S_{a_k}$}
\\
-f^2(2^k)&\T{on $S_{2a_k}\setminus S_{a_k}$},
\\
0&\T{ on $M\setminus S_{2a_k}$}.
\end{cases}
\end{equation}
This also readily implies the same inequalities as \eqref{2.1} with $\di_b\dib_b\phi^k$ replaced by $\Big(\T{Trace}(\di_b\dib_b\phi^k)\,\T{Id}-\di_b\dib_b\phi^k\Big)$.
\el
Note that there is no assumption about the behavior of $M$ at $S$ in this Lemma.
\bpf
Set
\begin{equation}
\Label{2.2}
\phi^k=c\chi(\frac{d_S(z)}{a_k})\log(\frac{d_S^2(z))}{a_k^2}+1),
\end{equation}
where $c$ is a constant that will be specified later and $\chi\in C^\infty(0,2)$ is a decreasing cut-off function which satisfies
\begin{equation*}
\begin{cases}
\chi\equiv1&\T{on $[0,1]$},
\\
0\le \chi\le 1 &\T{on $[1,\frac32]$},
\\
\chi\equiv0 &\T{on $[\frac32,2]$}.
\end{cases}
\end{equation*}
Remark that
\begin{equation*}
\begin{split}
\di_b\dib_b d_S^2&=2\di_b d_S\otimes \dib_b d_S+2d_S\di_b\dib_b d_S
\\
&\ge 2\di_b d_S\otimes \dib_b d_S\\
&\simgeq\T{Id},
\end{split}
\end{equation*}
where the last inequality follows from $\dim_{CR}(M)=0$ (with the agreement that Id denotes the identity of $T^\C M$).
Now, when $\di_b\dib_b$ hits $\log$, we have
\begin{equation}
\Label{2.3}
\begin{split}
\di_b\dib_b \log (\frac{d_S^2(z)}{a_k^2}+1)
&\simgeq\frac{\di_b d_S\otimes\dib_b d_S+d_S\di_b\dib_b d_S}{a_k^2}
\\
&\simgeq \frac{\T{Id}}{a_k^{2}}=f^2(2^k)\,\T{Id}.
\end{split}
\end{equation}
On the other hand, on $S_{a_k}$, the function $\chi$ is constant and therefore $\di_b\dib_b\phi^k=\di_b\dib_b \log $. Thus \eqref{2.3} yields the first of
\eqref{2.1}. When, instead, $\di_b\dib_b$ hits $\chi$, we have
\begin{equation}
\Label{2.3,5}
\begin{split}
\Big|\di_b\dib_b\chi\Big(\frac{d_S(z)}{a_k}\Big)\Big|&\leq|\ddot\chi|\frac{\di_bd_S\otimes \dib_b d_S}{a_k^2}+|\dot\chi|\frac{\di_b\dib_bd_S}{a_k}
\\
&\underset{\T{ since $\dim_{CR}(S)=0$}}\simleq \frac{\T{Id}}{a_k^2}.
\end{split}
\end{equation}
On the other hand, $\log$ stays bounded on $S_{2a_k}$ and therefore $\di_b\dib_b(\chi)\log\simgeq -a_k^{-2} =-f^{2}(2^k)$. Finally, when $\di_b$ and $\dib_b$ hit $\chi$ and $\log$ separately, we get
\begin{equation}
\Label{3.5,6}
\begin{split}
\Big|2\Re e \di_b\chi\Big(\frac{d_S}{a_k}\Big)\dib_b\log(\frac{d_S^2}{a_k^2}+1)\Big|&\simleq\Big|2\Re e\dot\chi\frac{\di_b d_S}{a_k}\otimes \frac{2a_k^2d_S\dib_bd_S}{2d_S^2a_k^2}\Big|
\\
&\underset{\T{since $d_S\sim a_k$ on supp$\,\dot\chi$}}\simleq \frac{\di_bd_S\otimes \dib_b d_S}{a_k^2}=f^2(2^k)\,\T{Id}.
\end{split}
\end{equation}
Thus, again, $2\Re e\dib_b\chi\dib_b\log\simgeq -f^2(2^k)\,\T{Id}$.
\epf
As we have seen in the proof of Lemma~\ref{l2.1}, when $\dot\chi$ and $\ddot\chi\neq0$, the Levi form of $\phi^k$ can get negative. However, this annoyance can be well behaved by the aid of the Levi form of $M$.
Let $F$ be a smooth real function such that $\frac{F(d)}{d^2}\searrow 0$ as $d\searrow0$, denote by $F^*$ the inverse to $F$ and define $f(t):=(F^*(\delta))^{-1}$, for $\delta=t^{-1}$.
Let $f(\Lambda)$ be the tangential pseudodifferential operator with symbol $f$.
This is defined by introducing a local straightening ${M}\simeq \R^{2n-1}\times\{0\}$ for a defining function $r=0$ of $M$, taking local coordinates $x\in M$, dual coordinates $\xi$ of $x$ and setting
$$
f(\Lambda)(u)=\int \left(e^{ix\xi}f(\sqrt{1+\xi^2})\int e^{-iy\xi} u(y)dy\right)d\xi.
$$
In particular $\Lambda$ is the standard elliptic pseudodifferential operator with symbol $\sqrt{1+\xi^2}$.
\bd
\label{type}
We say that $M$ has type $F$ along $S$ in a neighborhood $U$ of $z_o$, if
\begin{equation}
\Label{2.4}
(c_{ij})\simgeq \frac{F(d_S)}{d_S^2}\T{Id}\quad\T{on $U$}.
\end{equation}
\ed
Note that \eqref{2.4} implies
\begin{equation}
\Label{2.4bis}
\Big(\T{Trace}(c_{ij})\T{Id}-(c_{ij})\Big)\simgeq \frac{F(d_S)}{d_S^2}\T{Id}\quad\T{on $U$}.
\end{equation}
\bp
\Label{p2.1}
Let $M$ have type $F$ along S of CR dimension 0. Then
\begin{equation}
\Label{2.5}
\begin{cases}
\no{f(\Lambda)\Gamma_kv}^2_0\simleq \int_M(c_{ij})(\Gamma_k\Lambda^{\frac12}v,\overline{\Gamma_k\Lambda^{\frac12}v})\,dV+ \int_M (\phi^k_{ij})(\Gamma_kv,\overline{\Gamma_kv})\, dV+\no{\Gamma_kv}_0^2,\,\,h\in[1,n-1],
\\
\begin{split}
\no{f(\Lambda)\Gamma_kv}^2_0&\simleq \int_M\Big(\T{Trace}(c_{ij})\T{Id}-(c_{ij})\Big)(\Gamma_k\Lambda^{\frac12}v,\overline{\Gamma_k\Lambda^{\frac12}v})\,dV
\\
&+ \int_M \Big(\T{Trace}(\phi^k_{ij})\T{Id}-(\phi^k_{ij})\Big)(\Gamma_kv,\overline{\Gamma_kv})\, dV+\no{\Gamma_kv}_0^2,\,\,h\in[0,n-2].
\end{split}
\end{cases}
\end{equation}
\ep
\bpf
We set $a_k=f^{-1}(2^k)=F^*(2^{-k})$, $S_{a_k}=\{z:\,d_S(z)<a_k\}$ and denote by $\lambda(z)$ the minimum of the $n-1$ eigenvalues of $(c_{ij})$ at $z$.
We start from the first of \eqref{2.5}. We have
\begin{equation}
\Label{2.7}
\begin{split}
\no{\Gamma_kv}^{M\setminus S_{a_k}}_0&\simleq \underset{z\in M\setminus S_{a_k}}\max \frac{2^{-\frac k2}}{\lambda(z)^{\frac12}}\Big(\no{
\lambda^{\frac12}\Gamma_k\Lambda^{\frac12} v}^{M\setminus S_{a_k}}+\no{\Gamma_kv}_{-\frac12}\Big)
\\
&\simleq \frac{a_k2^{-\frac k2}}{F(a_k)^{\frac12}}\Big(\sqrt{\int_{M\setminus S_{a_k}}(c_{ij})(\Gamma_k\Lambda^{\frac12}v,\overline{\Gamma_k\Lambda^{\frac12}v})\,dV}+2^{-\frac k2}\no{\Gamma_kv}^{M\setminus S_{a_k}}_0\Big)
\\
&\simleq f^{-1}(2^{k})\Big(\sqrt{\int_{M\setminus S_{a_k}}(c_{ij})(\Gamma_k\Lambda^{\frac12}v,\overline{\Gamma_k\Lambda^{\frac12}v})\,dV}+\no{\Gamma_kv}_0\Big).
\end{split}
\end{equation}
Recalling that $f(\Lambda_\xi)\equiv f(2^k)$ on supp$\,\Gamma_k$, this gives
\begin{equation}
\Label{*}
\no{f(\Lambda)\Gamma_kv}^{M\setminus S_{a_k}}_0\simleq \sqrt{\int_{M\setminus S_{a_k}}(c_{ij})(\Gamma_k\Lambda^{\frac12}v,\overline{\Gamma_k\Lambda^{\frac12}v})\,dV}+\no{\Gamma_kv}_0.
\end{equation}
Now, on $S_{2a_k}\setminus S_{a_k}$, $(\phi^k)_{ij}$ can get negative. However, using the second of \eqref{2.1} and tuning the choice of $c$, independent of $k$, in the definition of $\phi^k$ so that $2^k(c_{ij})+(\phi^k)_{ij}\geq \frac{f^2(2^k)}2$ on $S_{2a_k}\setminus S_{a_k}$, we have that not only \eqref{*} but also \eqref{2.5} holds on $M\setminus S_{a_k}$.
Finally, on $S_{a_k}$, $(\phi^k)_{ij}$ satisfies the first of \eqref{2.1} and therefore
$$
\no{f(\Lambda)\Gamma_kv}^{ S_{a_k}}_0\simleq \sqrt{\int_{ S_{a_k}}(\phi^k_{ij})(\Gamma_kv,\overline{\Gamma_kv})\,dV}+\no{\Gamma_kv}_0.
$$
This shows how \eqref{2.5} follows from \eqref{2.4}. In the same way we can see that the second follows from \eqref{2.4bis}.
\epf
\noindent
{\it Proof of Theorems~\ref{t1.1} and \ref{t1.1,5}.} The proof of \eqref{1.0,1} just consists in taking summation over $k$ in \eqref{2.5}. As for \eqref{f} in degrees $h\in [1,n-2]$, it follows from the combination of the first (resp. the second) of \eqref{1.0,1} for $v=u^+$ (resp. $v=u^-$), in addition to the classical basic tangential estimates and the elliptic estimate for $u^0$. As for the critical degree $h=0$ and $h=n-1$ in \eqref{f}, it follows from writing $u=\dib_b^*w$ and $u=\dib_bw$ respectively (by closed range) and by using the estimate already established for $w$ in the non-critical degrees $1$ and $n-2$ respectively.
\section{The tangential H\"ormander-Kohn-Morrey formula twisted by a pseudodifferential operator}
\Label{s3}
Let $M$ be a CR manifold of hypersurface type of $\C^n$, $\dib_b$ the tangential Cauchy-Riemann system, $\dib_b^*$ the adjoint system. Our discussion is local and we can therefore assume that $M$ is in fact a hypersurface. For a neighborhood $U$ of a point $z_o\in M$, we identify $U\cap M$ to $\R^{2n-1}$ with coordinates $x$ and dual coordinates $\xi$, and consider a pseudodifferential operator $\Psi$ with symbol $\mathcal S(\Psi)(x,\xi)$. For notational convenience we assume that the symbol is real. We also use the notation $L^2_\phi$ for the $L^2$ space weighted by $e^{-\phi}$, $Q^b=\no{\dib_b u}^2+\no{\dib^*_b u}^2$ for the energy, and $Q^{b\,\phi}_\Psi=\no{\Psi\dib_b u}^2_\phi+\no{\Psi\dib^*_b u}^2_\phi$ for the energy weighted by $\phi$ and twisted by $\Psi$. We consider the pseudodifferential decomposition of the identity by Kohn $\T{Id}=\Phi^++\Phi^-+\Phi^0$ modulo $\T{Op}^{-\infty}$. We consider a basis of $(1,0)$ forms $\om_1,...,\om_{n-1}$ the conjugate basis $\bom_1,...,\bom_{n-1}$ and complete by a purely imaginary form $\gamma$. We denote by $\di_{\om_1},...,\di_{\om_{n-1}},\dib_{\om_1},...,\dib_{\om_{n-1}},T$
the dual basis of vector fields.
$M$ being a hypersurface defined, say, by $r=0$, we can supplement the $\om_j$'s to a full basis of $(1,0)$ forms in $\C^n$ by adding $\om_n=\di r$. Then $\gamma=\om_n-\bar\om_n$ and $T=\di_{\om_n}-\di_{\bar\om_n}$. We describe the commutators by
\begin{equation}
\Label{commutator}
\begin{split}
[\di_{\om_i},\di_{\bar\om_j}]&=\sum_{j=1}^{n}c_{ij}^h\di_{\om_h}-\sum_{j=1}^{n}\bar c_{ji}^h\dib_{\om_h}
\\
&=c^n_{ij}T+,\sum_{j=1}^{n-1}c_{ij}^h\di_{\om_h}-\sum_{j=1}^{n-1}\bar c_{ji}^h\dib_{\om_h};
\end{split}
\end{equation}
We also write $c_{ij}$ instead of $c^n_{ij}$.
For a cut-off $\eta\in C^\infty_c(U\cap M)$ we write $u^+:=\eta\Phi^+u,\,\,u^-=\eta\Phi^-u,\,\,u^0=\eta\Phi^0u,\,\,T^{\overset\pm0}=\eta T\Phi^{\overset\pm0}$. We note that $\mathcal S(T)>0$ on supp$\,\mathcal S(\Phi^+)$ (resp. $\mathcal S(T^-)>0$ on supp$\,\mathcal S(\Phi^-)$) and therefore $T^{\frac12}$ (resp. $(T^-)^{\frac12}$) makes sense when acting on $u^+$ (resp. $u^-$).
We make the relevant remark that
\begin{equation*}
\begin{cases}
\mathcal S(T)\sim \Lambda\T{ on supp$\,\mathcal S(\Phi^+)$}, \qquad \mathcal S(T^-)\sim \Lambda\T{ on supp$\,\mathcal S(\Phi^-)$},
\\
\{\mathcal S(\di_{\om_j}\}_{j=1,...,n-1}\sim \Lambda\T{ and } \mathcal S(\dib_{\om_j}\}_{j=1,...,n-1}\sim \Lambda\quad\T{on supp$\,\mathcal S(\Phi^0)$}.
\end{cases}
\end{equation*}
We denote by $\Op$ , resp. $\T{Op}^0$, an operator of order $2\T{ord}(\Psi)-\frac12$, resp. $0$, whose support is contained in supp$\,\Psi$; we also assume that $\T{Op}^0$ only depends on the $C^2$-norm of $M$ and, in particular, is independent of $\phi$ and $\Psi$.
\bt
\Label{t3.1}
(i) We have for every smooth form $v=u^+$ of degree $h\in[1,n-1]$
\begin{equation}
\Label{3.1}
\begin{split}
\int_M&e^{-\phi}(c_{ij})(T^{\frac12}\Psi v,\overline{T^{\frac12}\Psi v})dV+\int_Me^{-\phi}\Big((\phi_{ij})-\frac12(c_{ij})T(\phi)\Big)(\Psi v,\overline{\Psi v})dV+\NO{\Psi \bar\nabla v}_\phi
\\
&\simleq Q^{b\,\phi}_{\Psi}(v,\overline{v})+\NO{[\di_b,\Psi]\contrazione v}_\phi+\NO{[\di_b,\phi]\contrazione \Psi v}_\phi+\Big|\sum_{h=1}^{n-1}\int (c_{ij}^h)([\di_{\om_h},\Psi](v),\overline{\Psi v})\,dV\Big|
\\
&+\Big|\int_M e^{-\phi}[\di_b,[\dib_b,\Psi^2]](v,\overline{v})dV\Big|+Q^{b\,\phi}_{\Op}(v,\bar v)+\NO{\Op v}_\phi+\no{\Psi v}^2_\phi.
\end{split}
\end{equation}
Here we are using the notation $Q^{b\,\phi}_{\Psi}=\NO{\Psi\dib_b v}_\phi+\NO{\Psi\dib_b^* v}_\phi$.
\noindent (ii) We also have, for $v=u^-$ smooth of degree $h\in [0,n-2]$
\begin{equation}
\Label{3.1bis}
\begin{split}
\int_M&e^{-\phi}\Big(-(c_{ij})+\sum_jc_{jj}\T{Id}\Big)((T^-)^{\frac12}\Psi v,\overline{(T^-)^{\frac12}\Psi v})+\NO{\Psi \nabla v}_\phi
\\
&+\int_Me^{-\phi}\left(\left(-(\phi_{ij})+\sum_j\phi_{jj}\T{Id}\right)+\frac12\left((c_{ij})T(\phi))-(\sum_jc_{jj})T(\phi)\right)\right)(\Psi v,\overline{\Psi v})dV
\\
&\simleq Q^{b\,\phi}_\Psi(v,\overline{v})+\NO{[\di_b,\Psi]\contrazione v}_\phi+\NO{[\di_b,\phi]\contrazione \Psi v}_\phi+\Big|\sum_{h=1}^{n-1}\int\Big(-(c_{ij}^h)+\sum_jc_{jj}^h\T{Id}\Big)([\di_{\om_h},\Psi](v),\times
\\
&\times\overline{\Psi v})\,dV\Big|+\Big|\int_M e^{-\phi}\Big(-[\di_b,[\dib_b,\Psi^2]]+\T{Trace}([\di_b,[\dib_b,\Psi^2]])\T{Id}\Big)(v,\overline{v})dV\Big|
\\
&+Q^{b\,\phi}_{\Op}(v,\bar v)+\NO{\Op v}_\phi+\no{\Psi v}^2_\phi.
\end{split}
\end{equation}
\et
Clearly $u^0$ is subject to elliptic estimates. These, combined with \eqref{3.1}, \eqref{3.1bis} yield an estimate for the full $u$ in degrees $[1,n-2]$ and then also for $u\in\mathcal H^\perp$ in degree $k\in[0,n-1]$ by closed range.
\br
The formula also holds for $\Psi$ complex: in this case one replaces $\Psi^2$ by $|\Psi|^2$ and add the additional error term $[\di_b,\bar\Psi]\contrazione$ to the already existing $[\di_b,\Psi]\contrazione$.
\er
\bpf
We start from
\begin{equation}
\Label{35}
\begin{split}
\di_b\dib_b\phi&=\di_b(\sum_j\bar \di_{\om_j}(\phi)\bom_j)
\\&\sum_{ij}\left(\di_{\om_i}\dib_{\om_j}(\phi)+\sum_h\overline{c_{ji}^h} \dib_{\om_h}(\phi)\right)\om_i\wedge\bom_j.
\end{split}
\end{equation}
Similarly,
\begin{equation}
\Label{36}
\begin{split}
\dib_b\di_b\phi&=\dib_b(\sum_j \di_{\om_j}(\phi)\om_j)
\\&=\sum_{ij}\left(-\dib_{\om_j}\di_{\om_i}(\phi)-\sum_h{c_{ij}^h}\di_{\om_h}(\phi)\right)\om_i\wedge\bom_j.
\end{split}
\end{equation}
Differently from the ambient $\dib$-system on $\C^n$, we do not have $\di_b\dib_b=\dib_b\di_b$ and in fact,
combining \eqref{35} with \eqref{36}, we can describe $(\phi^b_{ij})$, the matrix of $\frac12(\di_b\dib_b-\dib_b\di_b)(\phi)$, by
\begin{equation}
\Label{3.3}
\begin{split}
\phi_{ij}^b&=\langle\frac12(\di_b\dib_b-\dib_b\di_b)(\phi),\di_{\om_i}\wedge\dib_{\om_j}\rangle
\\
&\underset{\T{by \eqref{35}, \eqref{36}}}=\frac12\Big(\Big(\di_{\om_i}\dib_{\om_j}+\dib_{\om_j}\di_{\om_i}\Big)(\phi)+\sum_{h=1}^{n-1}\bar c_{ji}^h\dib_{\om_h}(\phi)+c_{ij}^h\di_{\om_h}(\phi)\Big)
\\
&=\dib_{\om_j}\di_{\om_i}(\phi)+\frac12\Big([\di_{\om_i},\dib_{\om_j}](\phi)+\sum_h\bar c_{ji}^h\dib_{\om_h}(\phi)+\sum_hc_{ij}^h\di_{\om_h}(\phi)\Big)
\\
&\underset{\T{\eqref{commutator}}}=\dib_{\om_j}\di_{\om_i}(\phi)+\frac12c_{ij}T(\phi)+\sum_hc^h_{ij}\di_{\om_h}(\phi).
\end{split}
\end{equation}
We consider now
\begin{equation}
\Label{3.3bis}
e^{\phi}\Psi^{-2}[\dib_{\om_i},e^{-\phi}\Psi^2]=-\phi_{\bar \om_i}+2\frac{[\dib_{\om_i},\Psi]}\Psi+\frac{\Opbis}{\Psi^2},
\end{equation}
whose sense is fully clear when both sides are multiplied by $\Psi^2$.
In other terms, we have
\begin{equation}
\Label{3.3quater}
\dib^*_{e^{-\phi}\Psi^2}=\dib^*+\di\phi\contrazione-2\frac{[\di,\Psi]}\Psi\contrazione+\frac{\Opbis}{\Psi^2}+\T{Op}^0.
\end{equation}
This leads us to define the transposed operator $\delta_{\om_i}$ to $\dib_{\om_i}$ by
\begin{equation}
\Label{3.3ter}
\delta_{\om_i}:=\di_{\om_i}-\phi_{\om_i}+2\frac{[\di_{\om_i},\Psi]}\Psi+\frac{\Opbis}{\Psi^2}+\T{Op}^0.
\end{equation}
With these preliminaries we have
\begin{equation}
\Label{3.4}
\begin{split}
[\delta_{\om_i},\bar\di_{\om_j}]&=c_{ij}T+\sumh c_{ij}^h\delta_{\om_h}-\sumh\bar c_{ji}^h\dib_{\om_h}+\Big(\phi_{ij}^b-\frac12c_{ij}T(\phi)\Big)
\\&-2\sum_h c_{ij}^h\frac{[\di_{\om_h},\Psi]}{\Psi}+\frac{[\di_{\om_i},[\dib_{\om_j},\Psi]]}\Psi+\frac{[\di_{\om_i},\Psi]\otimes[\dib_{\om_j},\Psi]}{\Psi^2}+\frac{\Opbis}{\Psi^2}+\T{Op}^0.
\end{split}
\end{equation}
We remember now that there are two equally reasonable definition of the pseudodifferential action
\begin{equation}
\Label{pseudo}
\Psi(w)=\begin{cases}
(i)\hskip0.2cm\int e^{ix\xi}\SPsi(x,\xi)\tilde w(\xi)d\xi
\\
(ii)\hskip0.2cm \int e^{ix\xi}(\widetilde{\SPsi}(\cdot,\xi)*\tilde w)d\xi,
\end{cases}
\end{equation}
where $\tilde w$ denotes the Fourier transform.
Up to error terms of type $\Op$, we have
\begin{equation*}
\begin{split}
\NO{\Psi(w)}&\sim(\Psi w, \Psi w)
\\
&\underset{\T{Plancherel and \eqref{pseudo} (ii)}}\sim(\widetilde{\Psi(w)},\widetilde{\SPsi}(\cdot,\xi)*\tilde w)
\\
&\sim\int\widetilde {\Psi(w)}(\xi)\overline{\int\widetilde{\SPsi}(\xi-\eta,\xi)\tilde w(\eta)d\eta}d\xi
\\
&\underset{\T{$\overline{\widetilde{\SPsi}}(\xi-\eta,\xi)\sim\widetilde{\overline{\SPsi}}(\eta-\xi,\xi)$}}=\int\Big(\int\widetilde{\Psi(w)}(\xi)\widetilde{\overline{\SPsi}}(\eta-\xi,\xi)d\xi\Big)\bar{\tilde w}(\eta)d\eta
\\
&\underset{\T{\eqref{pseudo} (i)}}\sim\int\widetilde{\bar\Psi\Psi(w)}(\eta)\bar{\tilde w}(\eta)d\eta
\\
&\underset{\T{Plancherel}}\sim(|\Psi|^2w,w).
\end{split}
\end{equation*}
For the same reason $(\Psi^2w,w)\sim\int |\Psi|^2|w|^2\,dV$ and therefore
$$
\NO{\Psi(w)}\sim \int|\Psi|^2|w|^2\,dV.
$$
Adding the weight $\phi$ and recalling that in our discussion $\Psi$ is real,
\begin{equation}
\Label{reasonable}
\NO{\Psi\dib_b^{(*)}v}_\phi=\int e^{-\phi }\Psi^2|\dib_b^{(*)}v|^2\,dV+\NO{\Op(\dib^{(*)}v)}_\phi,
\end{equation}
where $\dib_b^{(*)}$ denotes either $\dib_b$ or $\dib_b^*$.
We are ready for the proof of \eqref{3.1}; we prove it only for $v=u^+$, the proof of \eqref{3.1bis} for $v=u^-$ being similar. We have
\begin{multline}
\int_\Om e^{-\phi}(c_{ij})(T\Psi v,\overline{\Psi v})+\int_\Om[\di_b,[\dib_b,e^{-\phi}\Psi^2]](v,\overline{v})dV
\\
-\NO{[\di_b,\phi]\contrazione \Psi v}_\phi-\NO{[\di_b,\Psi]\contrazione v}_\phi
+\NO{\Psi\bar\nabla v}_\phi
\\
\simleq \NO{\Psi\dib_b v}_\phi+\NO{\Psi(\dib_b)^*_{e^{-\phi}\Psi^2}v}_\phi+sc\NO{\Psi\bar\nabla v}_\phi+\Big|\sum_h\int_\Om e^{-\phi}(c_{ij}^h)([\di_{\om_h},\Psi]v,\overline{\Psi v})\,dV\Big|
\\
+Q^{b\,\phi}_{\Op}(v,v)+\NO{\Op v}_\phi+\NO{\Psi v}_\phi,
\end{multline}
or, according to \eqref{3.4} and after absorbing the term which comes with sc,
\begin{multline}
\Label{3.6}
\int_{M}e^{-\phi} c_{ij}(T\Psi v,\overline{\Psi v})dV+\int_{M}e^{-\phi} \phi_{ij}(\Psi v,\overline{\Psi v})dV-\NO{[\di_b,\phi]\contrazione \Psi v}_\phi
\\
+\int_M e^{-\phi}[\di_i,[\dib_j,\Psi^2]](v,\overline{v})dV
-\NO{[\di_b,\Psi]\contrazione v}_\phi+\NO{\Psi\bar\nabla v}_\phi
\\
\simleq \NO{\Psi\dib_b v}_\phi+\NO{\Psi(\dib_b)^*_{e^{-\phi}\Psi^2}v}_\phi
+\Big|\sum_h\int_\Om e^{-\phi}(c_{ij}^h)([\di_{\om_h},\Psi]v,\overline{\Psi v})\,dV\Big|
\\+Q_{\Op}(v,v)
+\NO{\Op v}_\phi+\NO{\Psi v}_\phi.
\end{multline}
To carry out our proof we need to replace $(\dib_b)^*_{e^{-\phi}\Psi^2}$ by $\dib_b^*$. We have from \eqref{3.4}
\begin{multline}
\Label{3.7}
\NO{\Psi (\dib_b)^*_{e^{-\phi}\Psi^2}v}_\phi\leq \NO{\Psi\dib_b^* v}_\phi+\NO{\Psi\di \phi\contrazione \Psi^2v}_\phi+\NO{[\di_b,\Psi]\contrazione v}_\phi+\NO{\Op v}_\phi
\\
+\underset{\T{\#}}{\underbrace{2\Big|\Re e (\Psi\dib_b^* v,\overline{\Psi\di_b\phi\contrazione v})_\phi\Big|+2\Big|\Re e (\Psi\dib_b^* v,\overline{[\di_b,\Psi]\contrazione v})_\phi+2\Big|\Re e (\Psi\di_b\phi\contrazione v,\overline{[\di_b,\Psi]\contrazione v})_\phi\Big|}}.
\end{multline}
We next estimate by Cauchy-Schwarz inequality
$$
\#\simleq \NO{\Psi\dib_b^* v}_\phi+\NO{\Psi\di_b\phi\contrazione v}_\phi+\NO{[\di_b,\Psi]\contrazione v}_\phi.
$$
We move the third, forth and fifth terms from the left to the right of \eqref{3.6}, and get \eqref{3.1} with $(T\Psi v,\Psi v)$ instead of $(T^{\frac12}\Psi v,T^{\frac12}\Psi v)$. But they only differ for
$$
\Big|\int_M e^{-\phi}\Big ([(c_{ij},T^{\frac12}](T^{\frac12}\Psi v,\Psi v)\Big)\,dV\Big|\simleq \NO{\Psi v}_0,
$$
which is negligeable.
\epf
We go back to the family of weights of Theorem~\ref{t1.1} and Proposition~\ref{p2.1}. We apply \eqref{3.1} (resp. \eqref{3.1bis}) for $\phi=\phi^k+t|z'|^2$ (resp. $\phi=\phi^k-t|z'|^2$). First, we note that they are absolutely uniformly bounded with respect to $k$; they can be made bounded in $t$ by taking $U=\{z:\,|z'|<\frac1t\}$. (In particular, by boundedness, they can be removed from the norms.) Possibly by raising to exponential, boundednes implies ``selfboundedness of the gradient" when $\phi$ is plurisubharmonic. In our case, in which to be positive is not $(\phi^k_{ij})$ itself but $2^k(c_{ij})+(\phi^k_{ij})$, we have, for $|z'|$ small
\begin{equation}
\Label{supernova}
\begin{split}
|\di_b\phi|^2&=|\di_b(\phi^k+t|z'|^2)|^2
\\
&\simleq |\di_b\phi^k|^2+t^2|z|^2
\\
&\leq 2^k(c_{ij})+(\phi^k_{ij})+t.
\end{split}
\end{equation}
So $\no{\di_b\phi\contrazione \Psi u^\pm}^2$ can be removed from the right side of both \eqref{3.1} and \eqref{3.1bis}.
Also, the term $-\frac12(c_{ij})T(\phi)(v,\bar v)$ is controlled by $(c_{ij})(T^{\frac12}v,\overline{T^{\frac12}v})$ by Sobolev interpolation.
We then combine Proposition~\ref{p2.1} with Theorem~\ref{t3.1} formula \eqref{3.1} for the weight $\phi^k+t|z'|^2$ (resp. formula \eqref{3.4} for the weight $(\phi^k-t|z'|^2$), and notice that $T^{\frac12}\sim\Lambda^{\frac12}$ on supp$\,\Psi^+$ (resp. $(T^-)^{\frac12}\sim\Lambda^{\frac12}$ on supp$\,\Psi^-$).
Also, on the right of \eqref{3.1} and \eqref{3.1bis}, one reduces $\NO{\Op v}_\phi$ to $\NO{v}_\phi$ by induction and estimates all terms $Q^\phi_{\Op}$ and $Q^\phi_{{\T{Op}}^{\T{ord}(\Psi)-j}}\,\,j\ge1$ by a common $Q^\phi_{\Psi'}$.
\noindent
{\it Proof of Theorem~\ref{t1.2}.}\hskip0.2cm We have to use \eqref{3.1} with the above choice of the weight $\phi$ and take summation over $k$; this yields \eqref{1.0,3} for $v=u^+$. The twin estimate for $v=u^-$ follows from \eqref{3.1bis} by similar procedure. Finally, \eqref{1.2} comes as the combination of \eqref{1.0,3} for $v=u^+$, the twin for $v=u^-$ and the elliptic estimate for $v=u^0$.
\hskip13cm$\Box$
\section{A criterion of hypoellipticity of the Kohn Laplacian}
\Label{s4}
Let $M$ be a pseudoconvex, hypersurface type manifold of $\C^n$, $\Box_b=\dib_b\dib^*_b+\dib^*_b\dib_b$ the Kohn Laplacian of $M$, and $G:=\Box_b^{-1}$ the Green operator.
{\it \bf Proof of Theorem~\ref{t1.3}}\hskip0.2cm
Our program is to prove that for any cut-off $\eta_o\in C^\infty_c(U)$ with $\eta_o\equiv1$ in a neighborhood of $z_o$, for suitable $\eta\succ\eta_o$, that is $\eta|_{\T{supp}\,\eta_o}\equiv1$, for any $s$ and suitable $U$, we have
\begin{multline}
\Label{4.1}
\no{\eta_o u}_s\simleq \no{\eta\dib_b u}_s+\no{\eta\dib^*_b u}_s+\no{u}_0\quad \T{for any $u\in \mathcal H^\perp\cap C^\infty(M\cap U)$}
\\
\T{ in any degree $k\in[0,n-1]$}.
\end{multline}
If we are able to prove \eqref{4.1}, we have immediately the exact local $H^s$-regularity of $\dib^*_bG$ and $\dib G$ over $\ker\dib$ and $\ker\dib^*$ respectively.
From this, we get the (non-exact) regularity of the Szeg\"o $S=\T{Id}-\dib_b^*G\dib_b$ and anti-Szeg\"o $S^*=\T{Id}-\dib_bG\dib_b^*$ projection respectively.
(At this stage we need to apply the method of the elliptic regularization to pass from $C^\infty$- to $H^s$-forms.)
From this the (non-exact) regularity of $G$ itself follows (cf. e.g. the proof of Theorem~2.1 of \cite{BKZ14}).
Along with $\eta_o\prec\eta$, we consider an additional
cut-off $\sigma$ with $\eta_o\prec\sigma\prec\eta$ and denote by $R^s$ the pseudodifferential operator with symbol $(1+|\xi|^2)^{\frac {s\sigma(a)}2}$.
According to Proposition~2.1 of \cite{BKZ14}, there is no restriction on the degree of $u$; thus $u$ can be either a form or a function.
By Section~\ref{s3} above, we can prove \eqref{4.1} separately on each term of the microlocal decomposition of $u=u^++u^-+u^0$; since $u^0$ has elliptic estimate and $u^-$ can be reduced to $u^+$ by star-Hodge correspondence, we prove the result only for $v=u^+$. We start from
\begin{equation}
\Label{4.11}
\begin{split}
\no{\Lambda^s\eta_o v}&\simleq \no{R^s\eta_o v}+\no{v}
\\
&=\no{R^s\eta_o\eta^2v}+\no{v}
\\
&\le \no{R^s\eta^2v}+\no{[R^s,\eta_o]\eta^2v}+\no{v}
\\
&\simleq\no{R^s\eta^2v}+\no{v}
\\
&\simleq \no{\eta R^s\eta v}+\no{[R^s,\eta]\eta v}+\no{v}
\\
&\simleq \no{\eta R^s\eta v}+\no{v},
\end{split}
\end{equation}
(cf. \cite{K02} Section~7).
Next, we apply Theorem~\ref{t3.1} for $\Psi=\eta R^s\eta$,
What we have to describe are the error terms in the right of \eqref{3.1}, \eqref{3.1bis}, that is, $[\di_b,\eta R^s\eta]$ and $[\di_b,[\dib_b,\eta R^s\eta]]$. Since the argument is similar for the two, we only treat the first. We have by Jacobi identity
\begin{equation}
\Label{4.6}
\begin{split}
[\di_b,\eta R^s\eta]&=[\di_b,\eta]R^s\eta+\eta[\di_b,R^s]\eta+\eta R^s[\di_b, \eta]
\\
&=[\di_b,R^s]+\T{Op}^{-\infty}.
\end{split}
\end{equation}
In fact, since supp$\,\di_b\eta\cap\,\T{supp}\,\sigma=\emptyset$, then the first and last terms in the right of the first line of \eqref{4.6} are operators of order $-\infty$ and can therefore be disregarded. As for the central term, we have
\begin{equation}
\Label{**}
[\di_b,R^s]=\di_b(\sigma)\log(\Lambda)R^s.
\end{equation}
Now, our hypothesis is that
\begin{multline}
\Label{4.7}
\no{\log(\Lambda)\di_b\sigma\contrazione \eta R^s\eta v}^2\le\epsilon\Big(\int_M(c_{ij})(\Lambda^{\frac12}\eta R^s\eta v,\overline{\Lambda^{\frac12}\eta R^s\eta v})\,dV
\\
+\sum_{k=1}^{+\infty}\int_M \Big((\phi^k_{ij})(\eta R^s\eta \Gamma_kv,\overline{\eta R^s\eta \Gamma_kv})\Big)\,dV\Big)
+c_\epsilon\NO{\eta R^s\eta v}.
\end{multline}
Altogether, we get
\begin{equation}
\Label{4.8}
\begin{split}
t\NO{\Lambda^s\eta_ov}&\underset{\T{\eqref{4.11}}}\simleq t\NO{\eta R^s\eta v}_0+\NO{v}_0
\\
&\simleq \Big(\int_M(c_{ij})(\Lambda^{\frac12}\Psi v,\overline{\Lambda^{\frac12}\Psi v})\,dV+\sum_{k=1}^{+\infty}\int_M(\phi_{ij}^k)(\Gamma_k\Psi v,\overline{\Gamma_k\Psi v})\,dV\Big)+t\NO{\eta R^s\eta v}_0+c_\epsilon\NO{v}_0
\\
&\underset{\T{by the second of \eqref{1.0,3}}}\simleq Q^b_{\eta R^s\eta}(v,\bar v)+\NO{[\di_b,\eta R^s\eta]\contrazione v}_0+\Big|\int_M[\di_b,[\dib_b,\eta R^s\eta]](v,\bar v)\,dV\Big|
\\&\hskip0.3cm+\Big|\sum_h\int (c_{ij}^h)([\di_{\om_h},\Psi](v),\overline{\Psi v})\,dV\Big|+Q_{\Op}^b(v,\bar v)+\NO{\Op v}_0
\\
&
\underset{\T{\eqref{**} and \eqref{1.4} (c)}}\simleq Q^b_{\eta R^s\eta}(v,\bar v)+\NO{\di_b(\sigma)\log(\Lambda)\eta R^s\eta v}_0+Q_{\Op}^b(v,\bar v)+\NO{\eta'v}_{s-\epsilon}
\\
&\underset{\T{\eqref{4.7}}}\simleq Q^b_{\eta R^s\eta}(v,\bar v)+\epsilon\Big(\int_M(c_{ij})(\Lambda^{\frac12}\eta R^s\eta v,\overline{\Lambda^{\frac12}\eta R^s\eta v})\,dV+\sum_k\int \Big((\phi^k_{ij})(\eta R^s\eta \Gamma_kv,\times
\\
&\times \overline{\eta R^s\eta \Gamma_kv})\Big)\,dV\Big)
+c_\epsilon\NO{\eta R^s\eta v}_0+Q_{\Op}^b(v,\bar v)+\NO{\eta'v}_{s-\epsilon}
\\
&\underset{\T{absorbtion in the second line}}\simleq Q^b_{\eta R^s\eta}(v,\bar v)+Q_{\Op}^b(v,\bar v)+\NO{\eta R^s\eta v}_0+\NO{\eta'v}_{s-\epsilon}
\\
&\underset{\T{absortion by means of $t$}}\simleq Q^b_{\eta R^s\eta}(v,\bar v)+Q_{\Op}^b(v,\bar v)+\NO{\eta'v}_{s-\epsilon}.
\end{split}
\end{equation}
Now, the $s-\epsilon$ norm is reduced to $0$ norm by induction over $j$ with $j\epsilon>s$, and $Q_\eta R^s\eta$ and the various $Q_{\T{Op}^{s-\epsilon j-1}}$ are estimated by a common $Q_{\eta'\Lambda^s}$. In conclusion, we have got \eqref{4.1} with the notational difference of $\eta'$ instead of $\eta$.
\hskip13cm $\Box$
\noindent
{\it Proof of Theorem~\ref{t1.4}.} \hskip0.2cm
We choose our cut-off starting from a cut-off $\chi$ in one real variable and setting $\eta=\Pi_j\chi(|z_{I^j}|)\chi(|y_n|)$. We have
\begin{itemize}
\item[(a)] supp$\,\di_{z_{I^j}}\chi(|z_{I^j}|)$ is contained in $z_{I^j}\neq0$
in particular, outside the ``critical" curve $\Gamma$ where superlogarithmic estimates hold by Theorem~\ref{t1.1} and Theorem~\ref{t1.1,5};
thus $\di_b(\Pi_j\chi(|z_{I^j}|)$ are superlogaritmic multipliers.
\item[(b)] $\di_b\chi(|y_n|)\sim\cdot(h^{I^j}_{z_{I^j}})_j$ and hence it is by hypothesis a superlogarithmic multiplier.
\end{itemize}
Altogether, $\di_b\eta\contrazione$ are superlogarithmic multipliers.
Remember that we are assuming that $(c_{ij}^h)$ are subelliptic multipliers.
Finally, supp$\,[\di_b,[\dib_b,\chi(z_{I^j})]]$ is contained in $z_{I^j}\neq0$ and $[\di_b,[\dib_b,\chi(y_n)]]\sim h^{I^j}_{z_{I^j},\overline{z_{I^j}}}$ are subelliptic multipliers; in conclusion, $[\di_b,[\dib_b,\eta]]$ are superlogarithmic multipliers.
We can then apply Theorem~\ref{t1.3} and this completes the proof of Theorem~\ref{t1.4}
\hskip12cm $\Box$
|
\section{Introduction}
The Physics of the non-planar two-dimensional electron gas has become an active area of research \cite{magaril}. Curvature induces novel phenomena which might be important to engineering quantum devices \cite{saxena}. On the other hand, curvature can influence well known two-dimensional phenomena. For example, in references \cite{ferrari1}, the cylindrical geometry has profound effects in a two-dimensional electron gas (2DEG) in an external magnetic field. Hall conductance in a spherical cap was investigated in \cite{cap}. Landau levels and the Hall conductivity were also investigated theoretically on a surface of constant negative curvature in \cite{bulaev}. The magnetic moment of a 2DEG in this geometry was addressed in \cite{bula}. A free particle moving on a one-sheeted hyperboloid was discussed both at the classical and quantum levels and novel equations of motion were found in \cite{k2}. Quantization on a prolate and on a oblate ellipsoid, considering applications to multiple-shell fullerenes, can be found in \cite{ivailo}. A special surface on which many theoretical investigations are realized is the cone. Its geometry is associated to topological line defects in semiconductors \cite{katanaev} known as disclinations . We can cite some examples as the investigation of the influence of this surface on the specific heat\cite{heat} and on the persistent currents \cite{sergio} of an electron gas. A particle constrained to move on a cone and bound to its tip by harmonic oscillator and Coulomb-Kepler potentials was considered in \cite{adam}. The quantum mechanics on a cone was studied by the path integral method in \cite{inomata}. In \cite{janaira}, the Landau levels on conical graphene were obtained. An interesting problem, concerned with the possibility of holonomic quantum computation based on the defect-mediated properties of graphite cones, was investigated in \cite{knut}.
A fundamental problem that has been discussed in \cite{cone1} is how the cone tip affects the quantum dynamics on a cone. From a mathematical point of view, there exists a coupling between the wave functions and the singular scalar curvature on a cone \cite{remarks}. However, real cones have smoothed out tips and therefore no singularity. The influence of such coupling in real systems can only be investigated from an experimental point of view. Considering this, we investigate the Landau levels on a cone in both cases: first without taking into account the influence of the singularity and then using the self-adjoint extension, which is the mathematically correct way to incorporate it. We then compare both cases and describe how should be the Integer Quantum Hall effect in both scenarios, hopping that these analysis would inspire experimental investigations of singular effects on $ \rm 2DEG^{'}s$. Although our findings are applied only to common semiconductors, they give a general idea on how singularities can affect a 2DEG in other materials. For instance, we expect the work addressed may lead to further theoretical and experimental investigations in graphene sheets \cite{graphene} as well as in 2D topological insulators \cite{top} with conical tips.
We will see that the profile of the Hall conductivity versus the external magnetic field changes whether or not we have a coupling with the conical tip. We also investigate the integer quantum Hall effect using two different models for electrons on a curved surface. One is based on the dimensional reduction of squared Dirac theory on surfaces employing an intrinsic approach which implies an effective scalar potential proportional to the Gaussian curvature on the surface in the non relativistic limit. The other employs an extrinsic approach which implies an effective scalar potential which contains both the Gaussian and the mean curvature on the surface. This discrepancy was addressed in \cite{remarks}. We will show that the profile of the Hall conductivity may show discrepancies as well. Depending on the opening angle of the cone, the contribution from the mean curvature potential may be irrelevant.
This work is divided as follows: in the section 2, we obtain the Landau levels on a cone and we discuss how the conical tip affect them. In section 3, we compute the Hall conductivity at zero temperature in order to show how the presence/absence of singular effects changes its profile. We also investigate how the results change depending on which theoretical model for carries on curved surfaces is considered. In the last section we have the concluding remarks.
\section {Landau levels for a particle on a cone}
Let us consider the coordinates $l$ and $\varphi$ of a particle confined to the surface of a cone as defined by
\begin{equation}
\left\lbrace
\begin{array}{lll}
$ $x=l\sin\alpha\cos\varphi\\
y=l\sin\alpha\sin\varphi\\
z=l\cos\alpha$
$ \end{array}
\right.\;,\label{cone}
\end{equation}
where $0\leq\varphi\leq 2\pi$ is the usual cylindrical $\varphi$ coordinate and $0< l <+\infty$ (see Fig. 1). Note that $l\sin\alpha$ is the usual cylindrical $\rho$ coordinate.
Consider an applied uniform magnetic field in the $z-$direction, $\vec{B}=(0,0,B_{z})$, as depicted in Fig. 1. For this field, we may choose the vector potential $\vec{A}=\frac{1}{2}\rho B_z \vec{\hat{\varphi}}$ such that $\vec{B}=\vec{\nabla}\times \vec{A}=B_z \hat{\vec{z}}$. Then,
\begin{equation}
\vec{A}=\frac{B_{z}l}{2}\sin\alpha \vec{\hat{\varphi}}\;.
\end{equation}
\begin{figure}[!htb]
\begin{center}
\includegraphics[height=7cm]{cone1.eps}
\caption{Geometrical setting of the problem: section of height $l\cos\alpha$ of the infinite straight circular cone of opening angle $2\alpha$ and uniform magnetic field parallel to the cone axis. We consider $\alpha$ between $0$ and $\pi/2$, since, for $\pi/2<\alpha <\pi$ we just have an inverted cone.}
\label{cone}
\end{center}
\end{figure}
We can now write the Schr\"odinger equation for a charged particle confined to the surface of a cone immersed in an external magnetic
field in the $z-$direction. From the relation (\ref{cone}), the map of a cone is $\vec{X}(l,\varphi)=(l\sin\alpha\cos\varphi,l\sin\alpha\sin\varphi,l\cos\alpha)$. The coefficients of the induced metric are given by \cite{docarmo}
\begin{equation}
g_{11}=X_l\cdot X_l=\frac{\partial \vec{X}}{\partial l}\cdot \frac{\partial \vec{X}}{\partial l}=1\;,
\end{equation}
\begin{equation}
g_{21}=g_{12}=X_l\cdot X_\varphi=\frac{\partial \vec{X}}{\partial l}\cdot \frac{\partial \vec{X}}{\partial \varphi}=0\;,
\end{equation}
\begin{equation}
g_{22}=X_\varphi\cdot X_\varphi=\frac{\partial \vec{X}}{\partial \varphi}\cdot \frac{\partial \vec{X}}{\partial \varphi}=l^2\sin^2\alpha\;.
\end{equation}
So, the metric tensor and its inverse are
\begin{equation}
g_{ij}=
\begin{pmatrix}
1&0\\
0&l^{2}\sin^{2}(\alpha)\\
\end{pmatrix}
\end{equation}
and
\begin{equation}
g^{ij}=
\begin{pmatrix}
1&0\\
0&\frac{1}{l^{2}\sin^{2}(\alpha)}\\
\end{pmatrix} \;,
\end{equation}
respectively.
We consider the usual minimal coupling Hamiltonian for a charged particle
\begin{equation}
\hat{H}=\frac{1}{2m}(\hat{\vec{p}}-e\hat{\vec{A}})^2\;.
\end{equation}
In order to ensure the Hermiticity of the radial momentum operator, $\hat{p}_l$, acting on the radial wave functions must have the form \cite{acsy}
\begin{equation}
\hat{p}_l\psi(l)=-i\hbar\left(\frac{\partial}{\partial l}+\frac{1}{2l}\right)\psi(l)\;\label{herm}
\end{equation}
whereas \cite{cone1}
\begin{equation}
p_{\varphi}=-\frac{i\hbar}{l\sin\alpha }\partial_{\varphi}.
\end{equation}
So, we achieve
\begin{equation}
\hat{H}=-\frac{\hbar^2}{2m}\left[\frac{1}{l}\frac{\partial}{\partial_l}\left(l\frac{\partial}{\partial_l}\right)-\frac{1}{4l^2}-\left(-\frac{i}{l\sin\alpha}\frac{\partial}{\partial_\varphi}-\frac{e\sin\alpha B_{z}l}{2\hbar}\right)^2\right]
\end{equation}
for the particle on a cone under the action of a magnetic field. It is important to notice that the Schr\"odinger equation for a particle moving on curved surfaces exhibits a scalar potential which describes its interaction with geometry. It has been shown by Ferrari and Cuoghi {\cite{ferrari} that this potential does not couple with the magnetic field. The geometric interaction is given by the potential
\begin{equation}
V_S=-\frac{\hbar^2}{2m}\left({\rm M}^2-{\rm K}\right)=-\frac{\hbar^2}{2m}\left(k_1-k_2\right)^2\;,\label{dacosta}
\end{equation}
where, $k_1$ and $k_2$ are the surface's {\it principal curvatures}, ${\rm M\equiv1/2(k_1+k_2)}$ is the {\it mean curvature} and $K\equiv k_1k_2$ is the {\it Gaussian curvature} of the surface \cite{docarmo}. This potential was first derived in \cite{dacosta}. Curiously, in Geometry, the term $W=\int_{S} \left(k_1-k_2\right)^2 dA$ is known as the Willmore energy \cite{Willmore} which, remarkably, is invariant under conformal transformations of $\mathbb{R}^3$ \cite{White}. The Willmore energy gives a measure of how much a given compact surface deviates from a sphere (which has $W=0$). It has appeared naturally in other physical contexts, like for example, the Helfrich model \cite{Helfrich} for elasticity of cell membranes.
Finally, the Schr\"odinger equation is $\left[\hat{H}+V_S\right]\Psi(l,\varphi)=E\Psi(l,\varphi)$. So, we have
\begin{equation}
-\frac{\hbar^2}{2m}\left[\frac{1}{l}\frac{\partial}{\partial l}\left(l\frac{\partial}{\partial l}\right)-\frac{1}{4l^2}-\left(\frac{-i}{l\sin\alpha}\frac{\partial}{\partial_\varphi}-\frac{e\sin\alpha B_{z}l}{2\hbar}\right)^2\right]\Psi(l,\varphi)+V_S\Psi(l,\varphi)=E\Psi(l,\varphi)\label{scho}\;.
\end{equation}
We may calculate $\rm M$ and $\rm K$ by the formula \cite{docarmo}
\begin{equation}
{\rm M}=\frac{1}{2}\frac{eG-2fE+gE}{EG-F^2}\;
\end{equation}
and
\begin{equation}
{\rm K}=\frac{1}{2}\frac{eg-f^2}{EG-F^2}\;,
\end{equation}
respectively. For a cone, $E=g_{11}=1$, $F=g_{12}=0$, $G=g_{22}=l^2\sin^2\alpha$, $e=\vec{n}\cdot \frac{\partial^2\vec{X}}{\partial\varphi^2}=0$, $f=\vec{n}\cdot \frac{\partial^2\vec{X}}{\partial\varphi\partial l}=0$ and $g=\vec{n}\cdot\frac{\partial^2\vec{X}}{\partial l^2}=l\sin\alpha\cos\alpha$. This way, ${\rm M}=\frac{\cos\alpha}{2\sin\alpha l}$ and $K=0$ except at the tip of the cone where it is a delta-function singularity. We consider first the case of real cones, which do not have singularities, we neglect the influence the effect of the singular Gaussian curvature on the wave function. By considering the wavefunction as $\Psi(l,\varphi)=\Psi(l)e^{ij\varphi}$, with $j=0,\pm1,\pm2,...$, we arrive at
\begin{equation}
-\frac{\hbar^2}{2m}\left[\frac{1}{l}\frac{d}{d l}\left(l\frac{d}{d l}\right)-\frac{\frac{1}{4}+\frac{\mu^2}{\sin^2\alpha}}{l^2}\right]\Psi(l)+\frac{mw^2l^2}{2}\Psi(l)=\Sigma\Psi(l)\label{harm}
\end{equation}
where $\mu^2=j^2-\frac{1-\sin^2\alpha}{4}$, $\omega=\left(\omega_c/2 \right)\sin\alpha$, with $\omega_c=\frac{eB_z}{m}$(cyclotron frequency), and
\begin{equation}
\Sigma=E+j\hbar\omega\;.\label{cond}
\end{equation}
Notice that, in general, $\frac{1}{4}+\frac{\mu^2}{\sin ^2 \alpha}>0$ but for the specific case of $j=0$ and $0<\alpha<\pi/4$ it is negative. This means that we have an attractive $1/l^2$ potential near the cone tip. This is a pathological potential whose bound state spectrum is unbound from below \cite{case} and should be treated with proper regularization. As it will be seen below, in order for the wave function to be square integrable we must have $j=0$ and then our study, for simplicity, will be restrict to the cones with $\alpha$ in the interval $[\pi/4,\pi/2]$.
The differential equation (\ref{harm}) describes a two-dimensional quantum oscillator on a conical background. This problem was addressed in \cite{kowaslki} for a circular double cone. In the case of 2DEG on a single cone, the range of the coordinate $l$ is $0< l <+\infty$ while that for a circular double cone it is $-\infty< l <+\infty$ . Like in \cite{kowaslki}, the wavefunctions are given by
\begin{equation}
\Psi(l)={\rm C} \left|l\right|^s \exp(-\frac{m\omega^2 l^2}{2})U\left(\frac{s}{2}+\frac{1}{4}-\frac{E}{2\omega},s+\frac{1}{2},m\omega l^2\right)\label{wf}
\end{equation}
with
\begin{equation}
s=\frac{1}{2}\left(1\pm\sqrt{1+\frac{4\mu^2}{\sin^2\alpha}}\right)\label{ss}
\end{equation}
and {\rm C} being the normalization constant. $U(a,b,c)$ is the Kummer function \cite{abra}. In order to have proper normalization of the wavefunction,we must have $$\lim_{l\rightarrow\infty}\Psi(l)\rightarrow0\;.$$ In order to get this condition, the series $U\left(\frac{s}{2}+\frac{1}{4}-\frac{E}{2\omega},s+\frac{1}{2},m\omega l^2\right)$ in (\ref{wf}) must be a polynomial of degree n. This is achieved when \cite{abra}
\begin{equation}
\frac{s}{2}+\frac{1}{4}-\frac{E}{2\omega}=-n\;.
\end{equation}
No extra condition must be imposed on $\Psi$. On the other hand, we have to investigate the behavior of the density of probability as $l\rightarrow0$, that is,
\begin{equation}
\lim_{l\rightarrow0}\left|\Psi(l)\right|^2 ldl\rightarrow0\;.
\end{equation}
Considering \cite{abra} $$\lim_{x\rightarrow0}U(a,b,x)\rightarrow \frac{\pi}{\sin(b\pi)}\left[\frac{1}{\Gamma(1+a-b)\Gamma(b)}-\frac{x^{1-b}}{\Gamma(a)\Gamma(2-b)}\right]\;,$$
we get
\begin{equation}
\lim_{l\rightarrow0}\left|\Psi(l)\right|^2 ldl=C l^{2s+1}\;,\label{si}
\end{equation}
where $C$ is a constant. Before doing any further analysis, we must mention that, as it was argued in \cite{cone1,kowaslki}, the tip of the cone correspond to a singularity and the the quantum dynamics requests {\it self-adjoint extensions} \cite{fr,selfadj}, in contrast with the double cone case, where such singular effects do not show up. Notice that the case of quantum particles around a double cone is similar to the one where we have quantum particles on a single cone without the coupling between the wavefunctions and a singular curvature. This is a situation without $\delta$-function potentials, which correspond the the so-called {\it Friedrichs extension} \cite{fr}. It is characterized by the following boundary condition,
\begin{equation}
\lim_{l\rightarrow0}l\partial_l \psi(l)=0\;.
\end{equation}
This condition means that, although we have a conical tip, the wavefunctions are {\it regular} at $l=0$. When we consider the $\delta-$function potential, the quantum dynamics requests self-adjoint extensions since the wavefunctions now are {\it irregular} as \cite{eu} $l\rightarrow0$.
In what follows, we evoke the results of \cite{kowaslki} for the regular case and after that we discuss what is going to change when we take into account the wavefunctions being irregular at the conical tip.
For electrons on a single cone without the coupling with the singular Gaussian curvature, the parameter $s$ above must be just that with a plus sign. This statement comes from equation (\ref{si}). Notice that, for positive $s$, $\left|\Psi(l)\right|^2 ldl$ goes to zero as $l\rightarrow0$. Since we do not have singularity, this means that the wavefunctions are regular at the conical tip. So, the eigenvalues are given by
\begin{equation}
\Sigma_{j,n}=2\hbar\omega\left(n+\frac{1}{2}+\frac{1}{4}\sqrt{1+\frac{4\mu^2}{\sin^2\alpha}}\right).
\end{equation}
Combining this equation with (\ref{cond}), we arrive at the {\it Landau levels } for electrons on a cone, that is,
\begin{equation}
E_{j,n}=\hbar\omega_c\sin\alpha\left(n+\frac{1}{2}+\frac{1}{4}\sqrt{1+\frac{4\mu^2}{\sin^2\alpha}}-\frac{j}{2}\right).\label{ll}
\end{equation}
The problem addressed here has an important difference from that investigated in \cite{kowaslki}. First of all, (\ref{ll}) has an extra piece proportional to $-j/2$. Also, the term inside the square root in (\ref{ll}) shows the parameter $\mu^2$( instead of just $j^2$) which can never be null for integer values of $j$. If we take $j=0$ in equation (\ref{harm}) we get the same result as taking $j\rightarrow0$ in (\ref{ll}), in contrast with \cite{kowaslki}, where the results are different. This solution corresponds to the case of an harmonic oscillator on a meridian of the cone. This difference comes from the geometric potential $V_S$ which leads to an inverse square distance dependent quantum potential.
We now turn our attention to the case where there is a coupling between the wavefunctions and the singular Gaussian curvature. The problem of the quantum harmonic oscillator with singularities was addressed in \cite{eu}. Everything we have done so far remains the same but the parameter $s$ (\ref{ss}) now admit both signs, plus and minus. The case with minus sign correspond to wave solutions which are irregular at the tip of the cone. In order to have $\int\left|\Psi(l)\right|^2 ldl=C l^{2s+2}\rightarrow0$ as $l\rightarrow0$, we must impose the following constrain
\begin{equation}
-1<s<1\;.
\label{cond1}
\end{equation}
This constraint guarantees that the wavefunctions are square integrable. Whenever we have singularity, this condition must happens \cite{hagen}. In order to fulfill (\ref{cond1}), the only possible value for the angular momentum is $j=0$. Then, the Landau levels split as
\begin{equation}
E_{0,n}^+=\hbar\omega_c\sin\alpha\left(n+\frac{1}{2}+\frac{1}{4}\sqrt{2-\frac{1}{\sin^2\alpha}}\right)\label{ll1}
\end{equation}
and
\begin{equation}
E_{0,n}^-=\hbar\omega_c\sin\alpha\left(n+\frac{1}{2}-\frac{1}{4}\sqrt{2-\frac{1}{\sin^2\alpha}}\right)\;.\label{ll2}
\end{equation}
In summary, without singular effects, the Landau levels are those obtained as (\ref{ll}), with $j=0,\pm1,\pm2,...$. If the singular Gaussian curvature plays a role, then the only possibility will be $j=0$. In the former case, the degeneracy is infinitely broken in comparison to the flat case($j=0,\pm1,\pm2,...$ ) and in the latter one, each Landau level is split twice. We sketch the Landau levels for both problems in figure \ref{eg}. In both cases, each energy labeled with index $n$ splits in two energy values but they show different energy spacing. This will have some impact in the Hall conductivity profile.
We note that the energy of the Landau levels as expressed by (\ref{ll1}) and (\ref{ll2}) becomes complex when the cone angle $\alpha$ is not in the interval $[\pi/4,\pi/2]$. As pointed out above, for cones with such values for $\alpha$, a pathological contribution to the potential appears and the problem requires special regularization at $l=0$. For these cones the result given by expressions (\ref{ll1}) and (\ref{ll2}) is therefore not valid. Since we are only considering the cones with $\alpha$ in the interval $[\pi/4,\pi/2]$, the energy levels given by (\ref{ll1}) and (\ref{ll2}) are always real.
\begin{figure}[!htb]
\begin{center}
a\includegraphics[height=4cm]{imageI.eps}\label{f1}
\hspace{0.2cm}
b\includegraphics[height=4cm]{ll.eps }\label{f2}
\caption{{\bf a} Some Landau levels for a 2DEG on a single cone without singular effects(we considered just three filled angular momentum states, $j=-1,0,1$) and {\bf b} with singular effects. In both cases, each energy labeled with index $n$ splits in two energy values but they show different energy spacing. We considered $\alpha=\pi/3$ rad.}\label{eg}
\end{center}
\end{figure}
With respect to the Landau levels given by (\ref{ll1}) and (\ref{ll2}) we remark that, for $\alpha=\pi/4$, the levels become degenerate. This is probably related to some specific symmetry of the $\alpha=\pi/4$ cone. A possible explanation might come from Supersymmetric Quantum Mechanics (SUSYQM) \cite{cooper} which relates pairs of Hamiltonians which share a particular mathematical relationship (the original SUSY concept related bosons and fermions). A hint for the specific puzzle of this degeneracy might be in the hidden SUSY found in the delta potential problem \cite{correa1} and in the Aharonov-Bohm system \cite{correa2}. This point deserves further investigation as well as the inclusion of the spin degree of freedom which, in the field of a magnetic vortex, shows additional SUSY properties\cite{correa3}.
\section{The Hall conductivity}
In this section, we compute the Hall conductivity of a 2DEG on a cone. The system which we are interested in is essentially the one depicted in Fig. 1 with the electrical contacts missing. It consists in a conical surface in the presence of a uniform magnetic field parallel to its axis. This, in principle, can be achieved with a carbon nanocone or with semiconductors heterostructures grown as a cone. We consider the integer quantum Hall effect, which consists of a quantization of the conductivity and appearance of plateaus at particular values of the external magnetic field. It manifests at low temperatures. We take into account just the ground state, $T=0$. Each Landau level contributes one quantum of conductivity to the electronic transport. If $n_0$ Landau levels are completely filled, the Hall conductivity, $\sigma_H$, will be given by\cite{hall}
\begin{equation}
\sigma_H=-n_0\sigma_o\;,
\end{equation}
where $\sigma_o\equiv e^2/h$ is the quantum of conductivity, $e$ is the electron charge and $h$ is the Planck$^{,}$s constant.
We start the calculations determining $n_0$ for a specific value of magnetic field. This is achieved by counting how many states are occupied below the {\it Fermi energy}, $E_F$. We compare $E_F$ with the Landau levels showed in figures \ref{eg}, making it equal to the highest Landau level filled. Next, we determine the quantity of magnetic field required to move the Hall conductivity from $-n_0\sigma_o$ to $-(n_o-1)\sigma_o$. In doing that, we are, in fact, determining the size of the plateaus at zero temperature for a specific $n_o$. Continuing this procedure, we can plot the Hall conductivity versus the plateaus. In our calculations, we start at $B=10T$.
In figure \ref{alpha} we plot the Hall conductivity for three different cases: a flat sample without singularity and the cone with and without singularity. In the last case we considered the filled states $j=-1,0,1$ only. If we take more values of $j$, which is not allowed for the Landau levels in the singularity case, we will obtain other intermediate plateaus. We see that the plateaus of conductivity on the cone are shifted to higher magnetic fields and there is an enhancement of each Hall conductivity in both cases. We can see clearly that the singular boundary conditions change significantly the Hall conductivity profile. We also probe the influence of the opening angle $\alpha$. In the interval of magnetic field showed, the Hall conductivity profile crosses the one for a flat sample for some values of $\alpha$ and some plateaus are shifted to lower magnetic fields. We can also observe that, for $\alpha=\frac{\pi}{2}$, we do not reach what we are calling {\it flat case}. For it, we considered just the Landau levels $E_n=\hbar\omega_c(n+1/2)$. We can not recover it here because of equation (\ref{herm}). This makes sense since we can not make any transformation with the conical tip which is a singular point. We believe that the case $\alpha=\pi/2$ correspond to a flat sample with an extracted point. The self-adjoint extensions in this case was addressed in \cite{extra} and we think it could be probed experimentally based on the Hall effect as discussed here. With these plots, we can more easily see that we have a reduction of the size of the plateaus in comparison to the flat case. In analyzing figure \ref{alpha}, we observe that considering or not the singular effects on the wavefunctions is an important issue on a theoretical basis. We hope this fact may be experimentally probed.
\begin{figure}[!htb]
\begin{center}
(a)\includegraphics[height=7cm]{alphaSCc.eps}
\hspace{0.5cm}
(b)\includegraphics[height=7cm]{alpha_flat.eps}
\hspace{0.5cm}
\caption{Hall conductivity on the cone for different opening angles. {\bf (a)} irregular and {\bf (b)} regular case. There is an enhancement on the Hall conductivity as $\alpha$ is reduced. Notice that considering the singularity in the wavefunctions has important consequences.}
\label{alpha}
\end{center}
\end{figure}
We now move to see the influence of the mean curvature, that is, we probe the two models we are dealing with for electrons on a curved surface. We consider both cases, with and without singular effects. In the figure \ref{M}, we can see that, depending on the opening angle $\alpha$, the two models may yield or not different quantitative results.
Indeed, for the singular case and for $\alpha=\frac{\pi}{3}$, the plateaus are shifted to lower magnetic fields. This happens because of the size of the plateaus, which are smaller when we consider the mean curvature potential. For $\alpha=\frac{2\pi}{5}$, taking or not into account the mean curvature potential does not make any difference.
Without singular effects, the same thing happens. The difference is only that pointed out in figures \ref{alpha}.
\begin{figure}[!htb]
\begin{center}
(a)\includegraphics[height=6cm]{SC_M_piover3.eps}
\hspace{0.5cm}
(b)\includegraphics[height=6cm]{SC_M_2piover5.eps}
\hspace{1.0cm}
(c)\includegraphics[height=6cm]{DC_M_piover3.eps}
\hspace{0.5cm}
(d)\includegraphics[height=6cm]{DC_M_2piover5.eps}
\caption{Hall conductivity on the cone with and without the mean curvature potential ${\rm M}$: without singularity, we have {\bf (a)} $\alpha=\frac{\pi}{3}$ and {\bf (b)} $\alpha=\frac{2\pi}{5}$, while with singularity we have {\bf (c)} $\alpha=\frac{\pi}{3}$ and {\bf (d)} $\alpha=\frac{2\pi}{5}$. The difference between both approaches may be noted depending on the opening angle $\alpha$ considered(see figure \ref{fonction}).}
\label{M}
\end{center}
\end{figure}
Plotting the mean curvature potential $M$ in function of the opening angle $\alpha$, we obtain figure \ref{fonction}. That explains clearly what happens in figure \ref{M}. Indeed, when we have a big opening angle, we can see that the mean curvature potential goes to zero. This is why we have seen in our plots that when we take a big opening angle, taking or not into account $M$ do not make a big difference, contrary to a small opening angle. This said, we conclude that depending on the opening of the cone, the difference between the two approaches discussed here for electrons on curved surfaces will not show up. This fact may happens for other geometries. So, comparing the mean curvature with the Gaussian curvature first can tell us when it is important to take it into account in our calculations.
Obviously, for {\it minimal surfaces} ($M\equiv0$), the two approaches yield identical Hall profiles.
\begin{figure}[!htb]
\begin{center}
\includegraphics[height=7cm]{M.eps}
\caption{Profile of the mean curvature $M$ depending on the opening angle $\alpha$. We can see that, for a smaller values of $\alpha$, $M$ grows while for higher values of $\alpha$ approaching $\pi/2$, $M\longrightarrow 0$.}
\label{fonction}
\end{center}
\end{figure}
\newpage
\section{Conclusion}
In this work we studied the integer quantum Hall effect on a cone. We have showed that considering the coupling between the wavefunctions and the singular scalar curvature (the conical tip), the profile of the Hall conductivity versus the external magnetic field modifies in a significant way. We have also considered two different models for electrons on a curved surface. One contains an effective scalar potential proportional to the Gaussian curvature on the surface while the other one shows an effective scalar potential which contains both the Gaussian and mean curvature on the surface. So, we have seen that the profile of the Hall conductivity show discrepancies when we compare the two theories, depending on the opening angle of the cone. We restricted our study to cones with opening angle $\alpha$ in the interval $[\pi/4,\pi/2]$ which is enough to show the effects of the conical geometry in the Hall conductivity. As an open problem we leave the case of more acute cone angles, $0<\alpha<\pi/4$, where the boundary condition at $l=0$ has to include some kind of regularization in order to fix the pathology that gives bound states with no lower bound.
The theory based on the da Costa approach(with the presence of mean curvature) is a result of particle confinement and it is the same for electrons and holes. A similar confining procedure
for relativistic carriers in graphene was first addressed in \cite{tun}. From the results we addressed here, we may note that the model one takes to investigate carries on curved graphene is going to be relevant. Without the confinement procedure, which in our case led
to the mean curvature contribution to the quantum Hall effect, the theory may not fit well the experiments on curved 2DEG. In the case of graphene cones \cite{grapcone}, the conical geometry, which appears due to the presence of topological defects, must be investigated carefully since the singular Gaussian curvature may also affect the electronic transport in this material. As pointed out in \cite{hagen}, an inadequate choice of the boundary condition for the electronic wavefunction can lead to contradictory physical results. The authors have showed that the irregular solutions at the conical tip must be considered. So, although we do not consider graphene here, our results may corroborate with this statement but we think that it should be probed experimentally in the context of quantum Hall effect.
In summary, the study carried out here can be realized in common semiconductors. However, it shows important features if one has intention to deal with non planar 2DEG on other classes of materials.
|
\section{Introduction and key results}
One of the most basic notions of condensed matter physcis is the quantum mechanical problem of a particle in a periodic potential. Yet, there are still quite fundamental questions relating to the physics of Bloch bands that have not been conclusively answered: How can optimally localized real space representations of band insulators in terms of Wannier functions (WFs) be found systematically and computationally efficiently? Under which circumstances can even compactly supported WFs exist for a given lattice Hamiltonian,
or at least for some representative of its topological equivalence class? These questions are of key importance not only for electronic band structure calculations within the single particle approximation, e.g., in the framework of density functional theory \cite{KohnSham}, but also for the dissipative preparation of topological band structures \cite{DiehlTopDiss} and their variational representation as a starting point for tensor network methods. In this work, we report substantial progress towards a comprehensive answer to these questions, \jcb{building on a {\it compressed sensing} (CS) based approach to the problem of finding maximally localized WFs recently proposed by Ozolins et al. \ \cite{OszolinsCompressedModes,OszolinsTranslation}}.
\subsection{\je{Localized Wannier functions}}
The crucial optimization problem of finding maximally localized WFs $\lvert w_R^\alpha\rangle$ associated with a family of $n$ occupied Bloch vectors $\lvert \psi_k^\alpha\rangle, \alpha=1,\ldots,n$\jens{,}
\je{and} $k\in \text{BZ}$ defined in the first Brillouin zone (BZ) has been subject of active research for many years \cite{VanderbiltReview}. The main difficulty is a local $U(n)$ gauge degree of freedom
in reciprocal space acting on the Bloch functions as
\begin{align}
\lvert \psi_k^\alpha\rangle \je{\mapsto} \sum_{\beta=1}^n U_{\alpha\je{,}\beta}(k)\lvert \psi_k^\beta\rangle.
\label{eqn:gauge}
\end{align}
This redundancy in the definition of the Bloch functions renders the Wannier representation highly non-unique: A different gauge choice on the Bloch functions can modify the localization properties of the associated WFs which are defined as
\begin{align}
\lvert w_R^\alpha\rangle=\frac{V}{(2\pi)^{\je{d}}}\int_{\text{BZ}}\text{d}^{\je{d}}k\, \text{e}^{-ikR}\lvert \psi_k^\alpha\rangle,
\label{eqn:WFDef}
\end{align}
where $V$ is the volume of the primitive cell in real space and \je{$d$} is the spatial dimension of the crystal.
Interestingly, the search for maximally localized WFs is substantially influenced by topological aspects of the underlying band structure. The recent study of band structure topology has led to fundamental discoveries like topological insulators and superconductors \cite{HasanKaneXLReview} that have given a new twist to the basic physics of Bloch bands: Roughly speaking, the topology of insulating band structures measures the winding of the submanifold of occupied bands, represented by their projection $\mathcal P_k=\sum_{\alpha=1}^n \lvert \psi_k^\alpha\rangle\langle \psi_k^\alpha \rvert$, in the total space of all bands as a function of the lattice momentum $k$. The archetype of a topological invariant for band structures is the first Chern number
\begin{align}
\mathcal C=\frac{i}{2\pi}\int_{\text{BZ}} \text{d}^2 k\, \text{Tr}\left( \mathcal P_k \left[(\partial_{k_x} \mathcal P_k),(\partial_{k_y} \mathcal P_k)\right]\right),
\label{eqn:Chern}
\end{align}
an integer quantized monopole charge associated with the gauge structure of the Bloch functions that distinguishes topologically inequivalent insulators in two spatial dimensions \cite{TKNN1982}. A non-vanishing monopole charge can be viewed as a fundamental obstruction to finding a global smooth gauge for the family of Bloch functions \cite{TKNN1982, Kohmoto1985}. However, it is precisely this analytical structure of the Bloch functions which determines the asymptotic decay of the associated WFs obtained by Fourier transform (cf.\ Eq.\ (\ref{eqn:WFDef})). This makes it intuitively plausible why a non-trivial band topology can have notable implications on the localization of WFs. Most prominently in this context, it is known that exponentially localized Wannier functions exist if and only if the first Chern number
is zero \cite{chernalgebraic, WannierChern, Matt}. In contrast, in one spatial dimension, Kohn could prove \cite{Kohn1959} that exponentially localized WFs always exist.
For so called symmetry protected topological states \cite{classification}, the situation is less simple. The topological nature of these band structures is rooted in the presence of a discrete symmetry, i.e., they are topologically distinct from an atomic insulator only if these underlying symmetries are maintained. Due to their vanishing Chern numbers, the existence of exponentially localized WFs is guaranteed for symmetry protected topological band structures. However, the possibility of representatives with even compactly supported WFs is unknown for many symmetry protected states. A conclusive understanding of this issue is of particular interest since compactly supported WFs imply the existence of exact flat-band models with finite range hopping \cite{chernflat} and the possibility of dissipative analogs by virtue of local dissipation \cite{BardynTopDiss}.
A remarkably widely adopted and practically very useful approach to maximally localized WFs has been reported in Ref.\ \cite{MarzariVanderbilt1997}, see Ref.\ \cite{VanderbiltReview} for a recent review article. The guiding idea in Ref.\ \cite{MarzariVanderbilt1997} is to localize the WFs in real space by optimizing the gauge of the associated Bloch functions in reciprocal space based on a gradient search algorithm. Generically, this class of algorithms finds a local optimum that depends on the initial choice of gauge.
Very recently, a different paradigm for the construction of localized functions that approximately block-diagonalize a Hamilton operator has been formulated \cite{OszolinsCompressedModes}. This approach is rooted in the theory of CS \cite{Candes}, a contemporary branch of research at the interface between signal processing and fundamental information theory \cite{CSI}, which has also found applications in quantum theory \cite{CS}.
In CS, the expected sparsity of a signal in some basis is employed for its exact reconstruction from significantly under-sampled measurement data, {without having to make use of the exact
sparsity pattern}. To this end, the sparsity of the signal is optimized under the constraint that it be compatible with the incomplete measurement data at hand. Translated to the spectral problem of a Hamiltonian, the analog of the incomplete measurement data is the ambiguity in the choice of basis functions that span a subspace associated with a certain energy range. Under the constraint of not involving basis states outside of this energy range, the sparsity of the basis functions in real space, i.e., their localization, is then optimized. First progress applying this program to the calculation of Wannier functions has been reported in Ref.\ \cite{OszolinsTranslation}.
\subsection{\je{Key results}}
In this work, we \jcb{extend} a CS based approach to the search for maximally localized WFs \jens{\cite{OszolinsCompressedModes,OszolinsTranslation} to study topological equivalence classes of band structures}. The physical motivation of our study is twofold: a comprehensive understanding of the interplay between band structure topology and localization properties of WFs at a fundamental level, and its impact on applications ranging from electronic band structure calculations over variational tensor network methods to dissipative state preparation.
To this end, \jcb{elaborating on the concepts introduced in Refs.\ \cite{OszolinsCompressedModes,OszolinsTranslation}}, we propose a numerically feasible and practical class of algorithms that are capable of manifestly maintaining the underlying physical symmetries of the band structure under investigation. Most interestingly, this allows us to search for maximally localized representatives of a {\it topological equivalence class of band structures} via adiabatic continuity -- an unprecedented approach. The method exploring this possibility does not only search for a gauge of maximally localized WFs for a given Hamiltonian. Instead, the model Hamiltonian flows continuously within the symmetry class of band structures under consideration towards a topologically equivalent sweet-spot with compactly supported WFs. The starting point is in this case a set of Wannier functions of a generic representative of the topological state of interest.
The asymptotic scaling of each step of
\jcb{the present} iterative
method is $O(N\log(N))$, where $N$ is the number of lattice sites in the system. We argue that for each step this is up to constants the optimal effort:
any algorithm on such a translation invariant problem will at some point involve a fast Fourier transform which has the same scaling. \jcb{This speedup compared to Ref.\ \cite{OszolinsTranslation} is rooted in the use of a local orthogonality constraint imposed on the Bloch functions in reciprocal space that is equivalent to a non-local shift-orthogonality constraint on the WFs.} \jcb{Furthermore}, the \jcb{extended} algorithms proposed in this work are capable of exactly preserving the fundamental physical symmetries of the system under investigation. \jcb{From a practical perspective,} this can be of key importance to obtain physically meaningful results when constructing approximate Wannier functions for a given model. For example, if one is concerned with mean field superconductors in the {\it Bogoliubov de Gennes} (BdG) formulation, the fermionic algebra necessarily entails a formal {\it particle hole symmetry} (PHS) constraint;
its violation would lead to inherently unphysical results. \jcb{From a more fundamental perspective, the capability of manifestly maintaining the underlying symmetries at every iterative step opens us the way to study equivalence classes of topological bands structures instead of individual Hamiltonian representatives.}
We present benchmark results for a one-dimensional (1D) topological superconductor (TSC) \cite{Kitaev2001} demonstrating the efficiency of our method: Starting from a generic representative Hamiltonian of a 1D TSC, the algorithm converges towards a set of WFs \resub{corresponding to a projection $\mathcal P_k$ onto an occupied band that obeys} the BdG PHS to high numerical accuracy. In the adiabatic continuity mode described above, our algorithm finds the maximally localized representative of the 1D TSC equivalence class, a state with compactly supported Wannier functions delocalized over two lattice sites. While for this particular state of matter, this ``sweet-spot'' point has been constructed analytically in Ref.\ \cite{Kitaev2001}, our search algorithm is capable of discovering it numerically starting from a generic model Hamiltonian represented by non-compact Wannier functions, as illustrated in Fig. \ref{fig:movies1D}.
For a topologically trivial starting point, the algorithm converges towards a set of atomic orbitals localized at a single lattice site -- the most localized representative of the trivial equivalence class. Finally, we give numerical evidence for the absence of compactly supported Wannier functions for {\it time reversal symmetry} (TRS) protected 2D topological insulators \cite{KaneMele2005a,KaneMele2005b,BHZ2006,koenig2007}: While our adiabatic search algorithm again converges to the WFs of an atomic insulator from a generic topologically trivial starting point, it does not find a compactly supported representative as soon as the initial Hamiltonian has gone through the phase transition to the topological insulator equivalence class. This indicates that there are no two-dimensional topological insulators with compact WFs.
\begin{figure}[htp]
\centering
\includegraphics[width=0.76\columnwidth]{movies1Dv2.pdf}
\caption{(Color online) Evolution of the extent of Wannier functions under the adiabatic continuity algorithm for a trivial 1D superconductor (upper panel) and a non-trivial 1D topological superconductor (lower panel). In both cases, the most localized, compactly supported representatives of the respective phases are found \resub{, i.e., a WF localized on a single site (upper panel) and on two sites (lower panel), respectively}. Plotted is the real space probability density $\rho_x$ (cf.\ Eq.\ \eqref{eq:rho}) on the horizontal $x$-axis with logarithmic color code from $10^{0}$ (yellow) to $10^{-30}$ (blue), and runtime increases on the vertical $t$-axis in units of ten iterative steps. Initial Wannier functions obtain from the gauge constructed in Eq.\ (\ref{eqn:blochKitaev}) in Sec. \ref{sec:restop1D}. Parameters are $\mu =1.5, 2t=2\Delta=1$ and $\mu =0.3, 2t=2\Delta=1$ for upper and lower panel, respectively. The home cell of both WFs is $x=101$, with total length $L=200$ for both plots.}
\label{fig:movies1D}
\end{figure}
{\emph{Outline. }}The remainder of this article is organized as follows. We define in Section \ref{sec:defopt} the search for maximally localized WFs associated with a given model Hamiltonian as an optimization problem subject to orthogonality and symmetry constraints. In Section \ref{sec:hamalg}, we \jcb{present} an efficient algorithm based on CS to numerically tackle this optimization problem. Numerical results for the 1D TSC are presented in Section \ref{sec:resham}. An algorithm which is not limited to a fixed model Hamiltonian but is designed for finding the most localized representative of a topological equivalence class of Hamiltonians is introduced in Section \ref{sec:algtop}. Benchmark results demonstrating the power of this tool are presented in Section \ref{sec:restop1D} and Section \ref{sec:restop2D}.
Finally, we sum up our findings and give an outlook to possible applications in Section \ref{sec:conclusion}.
\section{Compact Wannier functions from sparsity optimization}
\label{sec:csham}
\subsection{Formulation of the optimization problem}
\label{sec:defopt}
The problem of calculating the electronic (fermionic) band structure of a crystal within the independent particle approximation can be viewed as the quantum mechanical problem of a single particle in a lattice-periodic potential. The spectrum of its solution consists of energy bands parameterized by a lattice momentum. Both eigenvalues and eigenfunctions are periodic with the reciprocal lattice and can hence be constrained to a representative unit cell of the reciprocal lattice called the first Brillouin zone (BZ). The eigenfunctions are so called Bloch states. For a given set of energy bands, WFs, i.e., localized functions in real space that are orthogonal to their own lattice translations (shift orthogonality) can be obtained by Fourier transform of the Bloch states (cf.\ Eq.\ (\ref{eqn:WFDef})). In 1D, this problem has been addressed with methods from complex analysis by Kohn \cite{Kohn1959} showing that exponentially localized Wannier functions always exist. In higher spatial dimensions, topological obstructions can preclude the existence of exponentially localized WFs \cite{WannierChern}, e.g., due to a non-vanishing Chern number in 2D (cf.\ Eq.\ (\ref{eqn:Chern})).
The work by Kohn \cite{Kohn1959}, as well as the majority of applications for band structure calculations \cite{VanderbiltReview}\je{,} focus on periodic problems in the spatial continuum. In practice, the continuous problem is often times not approximated by a straightforward discretization in real space but by deriving a so called tight binding model. The relevant degrees of freedom of such a model are then a finite number of orbitals per site of a discrete lattice with the periodicity of the crystalline potential. Our work is concerned with such lattice models within the independent particle approximation from the outset.
{\emph{Definitions. }}We consider a system with Hamiltonian $H$ on a hypercubic lattice with unit lattice constant and $N=L^d$~sites with periodic boundary conditions. Each lattice site hosts $m$~internal degrees of freedom (orbitals), $n$ bands are occupied. Our single particle wave functions are hence normalized vectors in $\cc^{m N}$, a set of Wannier functions is represented by a {matrix}
{$\psi\in \cc^{mN\times n}$}
with shift-orthonormal {columns}, i.e. $\psi^\dag T_j \psi=\ii \delta_{j,0} $ {for all} $ j\in \mathbb Z_L^d$, where $T_j$ performs a translation by the lattice vector $j\in \mathbb Z_L^d$. We denote the matrix elements by $\psi_{\nu, j; \alpha}$, where $\nu \in \{1,\ldots,m\}$,
$ j\in \mathbb Z_L^d$, and $\alpha \in \left\{1,\ldots,n\right\}$. Among any set of shift orthogonal functions, a set of WFs associated with the $n$ occupied bands is
distinguished by minimizing the quadratic energy functional
\begin{align}
\mathcal E[\psi] = \text{Tr}(\psi^\dag H \psi).
\label{eqn:evar}
\end{align}
While the Slater determinant forming the many body ground state characterized by its minimal energy expectation value of the insulating band structure is unique up to a global phase, the set of possible single particle WFs $\psi$ representing this ground state, i.e., minimizing $\mathcal E$ is highly non-unique.
This is due to the local $U(n)$~gauge degree of freedom on the Bloch functions (cf.\ Eq.\ (\ref{eqn:gauge})).
Within this set, we would like to identify the representative where the probability density {$\rho_j^\alpha = \sum_\nu |\psi_{\nu , j; \alpha}|^2$} is most localized in real space.
In the language of compressed sensing, localization is referred to as sparsity.
\jcb{As \jens{suggested} in Ref.\ \cite{OszolinsCompressedModes}, a $l_1$-norm regularization of the energy functional (\ref{eqn:evar}) is a convenient way to enforce the localization of the WFs. Concretely, as a measure for sparsity, we use} the vector
$l_1$-norm $\| \sqrt{\rho} \|{_{l_1}}=\sum_{j,\alpha}\lvert \sqrt{\rho_j^\alpha}\rvert$ of the square root of the probability density{, as a convex relaxation with more benign properties regarding continuity \resub{than} discrete measures like the rank.}
For the WFs themselves, we define the $\rho$-norm as the $l_1$-norm of the square root of the associated probability density, i.e.,
\begin{align}\label{eq:rho}
\|\psi\|_\rho=\| \sqrt{\rho}\|_{l_1}.
\end{align}
A minimization with respect to the $\rho$-norm localizes the WFs only in real space and not in the internal degrees of freedom, as desired. The localization can be enforced by
adding a term $ \| \psi\|_\rho/\xi$~ \jcb{to the energy functional $\mathcal E$ \cite{OszolinsCompressedModes}}. The real parameter $\xi>0$~tunes the priority of the localisation respectively sparsity
condition compared to the energy minimization condition. The optimization problem considered
is hence the minimization \jcb{of the $l_1$-regularized energy expectation \cite{OszolinsCompressedModes}}
\begin{equation}
\mathcal E(\psi) + \frac{1}{\xi}\|\psi\|_\rho.
\label{eqn:l1reg}
\end{equation}
such that $\psi^\dagger T_j\psi=\ii \delta_{j,0}$. {The latter is a non-convex orthogonality constraint \cite{Yin}.}
Even if \jcb{the minimization of (\ref{eqn:l1reg})} will for finite $\xi$ in general produce approximations to the WFs of the model characterized by $H$, we would like to make sure that the \resub{band structure defined by the resulting WFs preserves the} underlying physical symmetries of the problem exactly. It is key to our algorithm that these symmetries can be exactly maintained.
Constraints that we will explicitly consider in this work are TRS $\mathcal T$, and PHS $\mathcal C$ (see Eq. (\ref{eqn:symconstraints}) for the corresponding constraints on the projection $\mathcal P_k$). Generically, we denote the set of local symmetry constraints by $\mathcal S$. With these definitions, the problem of maximally localized WFs can {for each $\xi>0$} be concisely stated as the {$l_1$ regularized minimization problem}
\begin{align}
&\psi=\text{arg min}_\phi \left(\mathcal E(\phi)+\frac{1}{\xi} \| \phi \|_\rho\right),\nonumber\\
&\text{subject to}\quad (\phi^\dag T_j \phi = {\ii }\delta_{j,0})~\text{and}~\mathcal S,
\label{eqn:minfunctional}
\end{align}
where arg gives the argument that minimizes the functional. The objective function is convex, while the symmetries \resub{and orthogonality constraints} give rise to
quadratic equality
constraints.
\subsection{Compressed sensing based algorithm}
\label{sec:hamalg}
Convex $l_1$ regularized problems can be practically and efficiently solved using a number of methods. Here, we
\jcb{use} a {\it split Bregman method} \cite{Bregman,Yin}, \jcb{which has been proposed to calculate maximally localized WFs in Refs.\ \cite{OszolinsCompressedModes, OszolinsTranslation}},
in a way that conveniently allows to include symmetries. The split Bregman method is related to the method of multipliers \cite{59}, which again can be connected to the alternating direction method of multipliers \cite{29}. Each step can then be implemented with as little as $O(N\log N)$ effort in the system size $N$.
The idea of a split Bregman iteration is to decompose the full optimization problem defined in Eq.\ (\ref{eqn:minfunctional}) into a set of coupled {subproblems} that can be solved exactly at every iterative step. We start from the simplest case without additional symmetries $\mathcal S$. In this case, our algorithm can be viewed as a numerically more efficient modification of the algorithms introduced in Refs.\ \cite{OszolinsCompressedModes, OszolinsTranslation}, adopted for {and generalized to}
a lattice Hamiltonian with internal degrees of freedom. We define the auxiliary variables $Q,R$ and associated noise terms $q,r$ that have the same dimension as the set of WFs $\psi\in
\cc^{mN\times n}$. During every step of the iteration, $\psi$ will optimise the energy functional $\mathcal E$ augmented by bilinear coupling terms (see step (i) below), $Q$ will be subject to a {\it soft thresholding procedure} stemming from the $\rho$-norm optimisation (see step (ii)), and $R$ will be subject to the shift-orthogonality constraint defining a proper set of WF (see step (iii)).
The noise terms $q$ and $r$ are incremented by the difference between $\psi$ and the auxiliary variables $Q$ and $R$, respectively (see steps (iv)-(v)). The algorithm in the absence of symmetries $\mathcal S$ then reads
{as pseudocode}
\begin{align}
&\text{Initialize } \psi=Q=R,~q=r=0. {\text{ While not converged do}}\nonumber \\
&\text{(i) } \psi\mapsto\text{arg}\min_{\psi} \left( \mathcal E[\psi]+\frac{\lambda}{2}\lVert \psi-Q+q\rVert^2_F+\frac{\kappa}{2}\lVert \psi-R+r\rVert^2_F\right),\nonumber\\
&\text{(ii) } Q\mapsto \text{arg}\min_{Q}\left( \resub{\frac{1}{\xi}}\| Q\|_\rho+ \frac{\lambda}{2}\lVert \psi-Q+q\rVert^2_F\right),\nonumber\\
&\text{(iii) } R\mapsto\text{arg}\min_R\frac{\kappa}{2}\lVert \psi-R+r\rVert^2_F,~\text{s.t.}~\tilde R_k^\dag \tilde R_k=\frac{\ii}{L^{d/2}}~\forall k,\nonumber\\
&\text{(iv) } q\mapsto q+\psi-Q,\nonumber\\
&\text{(v) } r\mapsto r+\psi-R,
\end{align}
where {$\lVert \ejb{M}\lVert_F\ejb{=({\sum_{i,j}|M_{i,j}|^2}})^{1/2}$ denotes the Frobenius matrix norm of a matrix $M$,
and $ \tilde R_k$~}the Fourier transform of $R$~at momentum $k$. $\lambda,\kappa,\xi>0$ are coupling constants.
The way this problem is split in parts, the
{subproblems} (i)-(iii) afford an explicit exact solution {not requiring any optimisation, given by}
\begin{align}
&\text{(i)}~ \psi=(2H+\lambda+\kappa)^{-1}(\kappa(R-r)+\lambda (Q-q)),\nonumber\\
&\text{(ii)}~ Q=\text{Shrink}\left(A,\frac{1}{\lambda \xi}\right),\nonumber\\
&\text{(iii)}~ \tilde R_k=\tilde B_k U\Lambda^{-\frac{1}{2}}U^\dag.
\label{eqn:exactmin}
\end{align}
Here $A=\psi+q,~ B=\psi+r$,
\begin{equation}
\text{Shrink}(\je{b},\epsilon)= \frac{\je{b}}{\lvert \je{b}\rvert} \max(0,\lvert \je{b}\rvert-\epsilon)
\end{equation}
is applied independently to each of the \resub{$m$-spinors} $B_j^\alpha$ associated with {the} Wannier function $\alpha$~evaluated at site $j$.
{Also,}
\begin{equation}
\tilde B_k^\dag \tilde B_k=U\Lambda U^\dag
\end{equation}
{with $U$ unitary and $\Lambda$ diagonal,} is {an eigenvalue}
decomposition of the positive Hermitian matrix $\tilde B_k^\dag \tilde B_k$. The orthogonality constraint $\tilde R_k^\dag \tilde R_k=\ejb{\ii}/{L^{d/2}}~\forall k$ on the Bloch functions occurring in step (iii) is equivalent
{with} the shift orthogonality {constraints} $R^\dag T_j R={\ii} \delta_{j,0}~\forall j$
on the Wannier functions. However, due to the local nature of the further, step (iii) can readily be solved exactly as explicitly done above, whereas the numerically less
efficient method of Lagrange multipliers has been proposed in Ref.\ \cite{OszolinsTranslation} to enforce the latter non-local constraint in real space.
{This is true even though it arises from a convex problem with a quadratic orthogonality constraint.}
More explicitly, the Fourier transform involved in the implementation used in the present work scales as $O(N\log N)$ if a fast Fourier algorithm is used.
Each step of the procedure is hence efficient. Rigorous convergence proofs for
split Bregman methods are known for $l_1$-regularized convex problems \cite{Convergence}. Here, including the equality constraints, there is still
evidence that the entire method is efficient and convergent as well, {in line with the findings of Ref.\ \cite{Yin}.}
Step (iii) of the above algorithm solves the following problem: Given a set of wave functions $B$, it finds the closest possible (in {Frobenius} norm)
set of proper shift orthogonal Wannier functions. Imposing additional local symmetry constraints $\mathcal S$~further complicates step (iii) of the above algorithm. From our numerical data presented below, it becomes clear that imposing constraints like PHS can be of key importance to obtain physically meaningful results. The simplest way to implement such symmetries is by considering the projection
\begin{align}
\mathcal P_k=\sum_{\alpha=1}^n \tilde \psi_k^\alpha \tilde \psi_k^{\alpha \dag}
\label{eqn:defpofk}
\end{align}
onto the occupied Bloch states at momentum $k$. Local symmetries will basically impose local constraints on this quantity, the only significant complication being the complex conjugation $K$ involved in anti-unitary constraints like TRS and PHS which connects $k$ and $-k$. Explicitly, for TRS $\mathcal T$ and PHS $\mathcal C$, we get the constraints
\begin{equation}
\mathcal T \mathcal P_k \mathcal T^{-1}=\mathcal P_{-k},\,\,
\mathcal C \mathcal P_k \mathcal C^{-1}=1-\mathcal P_{-k},
\label{eqn:symconstraints}
\end{equation}
respectively.
With these definitions, we are ready to formulate a symmetry purification procedure augmenting step (iii). To this end, we follow (iii) to obtain the closest Bloch functions for half of the BZ and calculate $\mathcal P_k$. For the other half of the BZ, $\mathcal P_k$ is then obtained by symmetry conjugation \resub{by virtue of Eq. (\ref{eqn:symconstraints})}. The Bloch functions spanning $\mathcal P_k$ for this second half are obtained by projecting \resub{the Bloch functions from the previous iteration} $\tilde B_k$ onto $\mathcal P_k$ and again performing an orthogonalization based on a eigenvalue decomposition. By this continuous gauge prescription, we make sure that an input function $\tilde B_k$ that already obeys the given symmetry is unchanged by the purification procedure. This ensures that the algorithm can become stationary for the desired solution. The choice how the BZ is divided into two halves is {to some extent}
arbitrary. However, the fact that the Bloch basis in which we perform this purification and the real space basis in which the thresholding (ii) is performed are maximally incoherent bases
{\cite{Candes}}
prevents systematic effects of this choice. For a unitary local symmetry, no constraint between $k$ and $-k$ is introduced and the symmetry purification can be done locally at every point in momentum space.
In summary, the core of our method consists of iteratively shrinking the spatial extent of the WFs by a soft thresholding prescription while reestablishing symmetry and orthogonality constraints on the associated \resub{projection $\mathcal P_k$} at every step. The localization and compact support of the WFs is enforced directly in real space by virtue of $l_1$-norm optimization. Split orthogonality and symmetry constraints enforce the defining properties of the desired WFs. For a search limited to WFs of a fixed lattice model Hamiltonian, the subspace corresponding to a certain subset of bands and with that to a certain energy range is selected by minimizing a quadratic energy functional as proposed in Ref.\ \cite{OszolinsCompressedModes}. Hence, the CS approach does not require the knowledge of an initial set of WFs as a starting point. The converged trial functions are compactly supported well-defined Wannier functions by construction. Their degree of localization can be tuned arbitrarily by a sparsity parameter $\xi$, with a tradeoff in controlling their quality in representing the given model Hamiltonian.
\subsection{Results for the 1D TSC state}
\label{sec:resham}
As an interesting benchmark example, we consider the 1D TSC proposed by Kitaev in 2001 \cite{Kitaev2001} which is distinguished from a trivial superconductor by a topological $\mathbb Z_2$-invariant. The simplest representative of this state is a 1D lattice of spinless fermions modelled by the Hamiltonian
\begin{align}
H_p =\sum_j \left(-t c_j^\dag c_{j+1} +\frac{\mu}{2} c_j^\dag c_j-\Delta c_j c_{j+1}+\text{h.c.}\right),
\end{align}
where $t$ is a real nearest neighbor hopping constant, $\mu$~models a chemical potential, $\Delta$~is a proximity induced $p$-wave pairing.
{The collection of} $c_j~(c_j^\dag)$ ~are the fermionic annihilation (creation) operators. Introducing the {collection of} Nambu spinors $\Psi_j=(c_j,c_j^\dag)^T$ and their Fourier transforms $\tilde \Psi_k=(\tilde c_k,\tilde c_{-k}^\dag)^T$, $H_p$ can be written in the BdG picture as
\begin{align}
H_p=\frac{1}{2}\int_{0}^{2\pi}\tilde \Psi_k^\dag d^i(k)\tau_i\tilde \Psi_k,
\label{eqn:KitaevBdG}
\end{align}
where $\tau_i$ are Pauli matrices in Nambu space \je{and}
\begin{align}
d^1(k)&=0,\\
d^2(k)&=-2 \Delta \sin(k),\\
d^3(k)&=\mu -2t\cos(k).
\end{align}
For simplicity, we {consider the specific case} $2\Delta=2t=1$. As a function of $\mu$, $H_p$ is then in the topologically non-trivial phase for $\lvert \mu \rvert<1$, critical for $\lvert\mu\rvert=1$, and trivial otherwise. The description in terms of Nambu spinors implies a formal doubling of the degrees of freedom while keeping the number of physical degrees of freedom fixed.
This redundancy \resub{is reflected} in an algebraic constraint on the BdG Hamiltonian that can be viewed as a formal PHS $\mathcal C=\tau_1 K$, where $K$ denotes complex conjugation.
The BdG Hamiltonian (\ref{eqn:KitaevBdG}) is formally equivalent to an insulating band structure with one occupied and one empty band. The projection $\mathcal P_k$ onto the occupied band can be expressed as
\begin{align}
\mathcal P_k=\frac{1}{2}(1-\hat d^i(k) \tau_i),
\label{eqn:ptwoband}
\end{align}
where {$\hat d(k)={d}(k)/{\lvert d(k)\rvert}$}.
\begin{figure}[htp]
\centering
\includegraphics[width=0.72\columnwidth]{CompressedPlotsJens.pdf}
\caption{(Color online) Logarithmic plots of the probability density $\rho$ of WFs with home cell $x=101$ for a non-trivial 1D TSC with $\mu =0.3, 2t=2\Delta=1$ (see Section \ref{sec:resham} for definitions). \eb{Upper} panel: Result of algorithm without additional symmetries and coupling constants $\xi=10,r=50,\lambda=20$. Central panel: Result of algorithm with $\mathcal S=\left\{\text{PHS}\right\}$ and coupling constants $\xi=10,r=50,\lambda=20$. \eb{Lower} panel: WF from the gauge constructed in Eq.\ (\ref{eqn:blochKitaev}). $L=200$
{has been} chosen
for all plots.}
\label{fig:hamcomparison}
\end{figure}
We now apply the algorithm introduced in Section \ref{sec:hamalg} to the toy model (\ref{eqn:KitaevBdG}). We first ignore the PHS constraint and apply the algorithm without further symmetries $\mathcal S$. For $\xi=10,r=50,\lambda=20,\mu=0.3,L=200$, it converges towards a set of compact WFs \resub{(see Fig.\ \ref{fig:hamcomparison} upper panel)} functions that minimize $\mathcal E$ to a relative accuracy of $1.8\je{\times}10^{-3}$ but that break PHS by as much as $2.0$ percent. The violation of PHS is measured by the deviation of $\lVert \mathcal P_k-(1-\mathcal C \mathcal P_{-k}\mathcal C^{-1})\rVert_F$ from zero \resub{(cf. Eq. (\ref{eqn:symconstraints}))} . Note that a set of WFs \resub{for which the associated projection $\mathcal P_k$} does not preserve the Nambu PHS $\mathcal C=\tau_x K$ cannot describe any physical superconductor. This demonstrates how important it is to manifestly maintain PHS here. In a next step, we apply the algorithm with $\mathcal S=\left\{\text{PHS}\right\}$ for the same parameters. It converges towards compactly supported WFs \resub{(see Fig.\ \ref{fig:hamcomparison} central panel)} which minimise $\mathcal E$ to a relative accuracy of $1.7\je{\times}10^{-3}$ and \resub{for which $\mathcal P_k$ preserves} PHS to $2.0\je{\times}10^{-8}$ accuracy within our numerical precision, i.e., six orders of magnitude better than without explicitly maintaining PHS. We show a logarithmic plot of the probability density $\rho$ of the converged WFs in Fig.\ \ref{fig:hamcomparison}. From these plots, it becomes clear why PHS is rather strongly broken if not explicitly maintained: The PHS breaking WFs \resub{(upper panel)} minimize the energy functional $\mathcal E$ to roughly the same accuracy but have a somewhat smaller $l_1$-norm at the expense of violating PHS. We compare the results of our algorithm to an analytically obtained WF \resub{(lower panel)} which has been computed from a smooth family of Bloch functions (see Eq.\ (\ref{eqn:blochKitaev}) below for its explicit construction), which clearly is much less localized (note the logarithmic scale).
\section{Maximally localized representatives of topological equivalence classes}
\label{sec:cstop}
\subsection{Adiabatic continuity algorithm}
\label{sec:algtop}
In Section \ref{sec:hamalg}, we introduced an algorithm that is designed to find compactly supported WFs for a fixed model Hamiltonian. In this Section, we present a tool which searches for the most localized compactly supported WFs not only for a given Hamiltonian but {within} an entire topological equivalence class. Topological equivalence classes are the connected components of the set of all free Hamiltonians obeying certain local symmetries $\mathcal S$. In other words, starting from any Hamiltonian that preserves $\mathcal S$, its topological equivalence class is defined by all Hamiltonians that can be reached adiabatically, i.e., continuously without closing the band gap and without breaking $\mathcal S$. We confine our attention to topological states relying on at least one symmetry constraint, i.e., states with zero Chern number. For states with non-zero Chern number, it is known that no representative with exponentially localized let alone compactly supported WFs can exist {\cite{Brouder}}.
The key idea of our adiabatic continuity algorithm is the following:
Start from a set of WFs associated with a generic representative of a given topological equivalence class. Perform the split Bregman iteration introduced in Section \ref{sec:hamalg} with the crucial difference that the energy functional $\mathcal E$~is set to zero. That way the bias towards a particular model Hamiltonian is completely removed. However, the symmetries $\mathcal S$ are again restored at every step of the iteration and the $\rho$-norm optimization is continuous on a coarse grained scale controlled by the thresholding parameter ${1}/({\lambda \xi})$. Hence, the model Hamiltonian that the instantaneous WFs represent will flow continuously in the topological equivalence class of the Hamiltonian associated with the initial set of WFs. The only bias of this flow is the $\rho$-norm optimization, i.e., the localization of the WFs in real space
by minimization of the ${{l_1}}$-norm of the square root of their probability density. Thus, the adiabatic continuity algorithm searches for the most localized representative of a given topological state of matter.
For the converged set of WFs, the corresponding Bloch functions are readily obtained by Fourier transform. From these Bloch functions, the projection onto the occupied bands $\mathcal P_k$ is straightforward to compute (see Eq. (\ref{eqn:PfromWanniers})). The generic flat band Hamiltonian $Q(k)=1-2\mathcal P_k$ then defines an explicit model Hamiltonian for the most localized representative of the topological equivalence class under investigation.
\subsection{Maximally localized representatives in symmetry class D in one dimension}
\label{sec:restop1D}
To benchmark the adiabatic continuity algorithm, we would like to apply it to the 1D TSC model (\ref{eqn:KitaevBdG}) introduced in Section \ref{sec:resham}.
{In the language of Ref.\ \cite{Altland}, this belongs to the symmetry class D.}
For this model, the result of a perfect performance is clear: From a topologically trivial starting point, we would expect our algorithm to converge towards an ``atomic'' Wannier function which has support only on a single site. From Ref.\ \cite{Kitaev2001}, we also know the simplest representatives of the topologically nontrivial class, which are of the form $\lvert t\rvert =\lvert \Delta\rvert >0=\mu$. Such exactly dispersionless models are characterized by WFs \resub{corresponding to operators} of the form $w_j=(c_j+c_j^\dag-c_{j+1}+c_{j+1}^\dag)/2$ with compact support on only two sites around $j\in \left\{1,\ldots,L\right\}$. It is clear that no topologically non-trivial state can be represented by WFs with support on a single site, as this would preclude any momentum dependence of $\mathcal P_k$. We hence expect a set of WFs \resub{annihilated by operators} similar to $w_j$ as a result of our adiabatic search in the topologically non-trivial sector.
As a starting point we calculate a set of WFs from a family of Bloch functions representing the occupied band of $H_p$ for generic $\mu$. A global gauge defining a family of Bloch functions
${k\mapsto \lvert u_-(k)\rangle}$ for the occupied BdG band can be constructed as
\begin{align}
\lvert u_-(k)\rangle = \frac{\mathcal P_k \lvert +x\rangle}{\lvert \mathcal P_k \lvert +x\rangle\rvert},
\label{eqn:blochKitaev}
\end{align}
where $\lvert +x\rangle=\tau_1 \lvert +x\rangle$ is a $\tau_1$ eigenvector. From Eq.\ (\ref{eqn:ptwoband}), it is easy to see that this gauge is regular for all $k$ since $d^1(k)=0$.
The initial WFs $\psi_0$~are then simply obtained by Fourier transform of the Bloch functions. Since $k\mapsto \lvert u_-(k)\rangle$ as resulting from Eq.\ (\ref{eqn:blochKitaev}) are $C^\infty$ functions, the corresponding Wannier functions are asymptotically bound to decay faster than every power law and exhibit in fact only exponential tails as verified in Fig.\ \ref{fig:inWFLog}. Our gauge choice turns out to be more efficient for the non-trivial WF which decays much more rapidly.
Using these functions as an input, the algorithm described in Section \ref{sec:algtop} indeed converges to the correct benchmark results in less than one minute on a regular desktop computer for a lattice of size $L=200$. In other words, our search algorithm numerically detects the ``sweet spot'' point with compactly supported WFs from Ref.\ \cite{Kitaev2001}, starting from a generic set of WFs representing some Hamiltonian with dispersive bands in the same topological equivalence class. Conversely, as soon as we tune $\mu$ over the topological quantum phase transition to a trivial 1D superconducting state, our search algorithm correctly finds an atomic WF representing the simplest possible trivial Hamiltonian.
In Fig.\ \ref{fig:movies1D}, we visualize the performance of our algorithm with a logarithmic color plot of the probability density $\rho_x$ at lattice site $x$ as a function of the computation time $t$. The final WFs concur with the anticipated perfect benchmark results to impressive numerical precision.
\begin{figure}
\includegraphics[width=0.7\columnwidth]{initialWFs.pdf}
\caption{(Color online) Logarithmic plot of the probability density $\rho$ of sets of Wannier functions from the gauge constructed in Eq.\ (\ref{eqn:blochKitaev}) for a trivial 1D superconductor with $\mu =1.5, 2t=2\Delta=1$ (lower panel) and a non-trivial 1D TSC with $\mu =0.3, 2t=2\Delta=1$ (upper panel). The home cell of both WFs is $x=101$. The linear tails demonstrate the asymptotic exponential decay. $L=200$
{is chosen}
for both plots.}
\label{fig:inWFLog}
\end{figure}
\subsection{Absence of compactly supported topological insulator WFs in symmetry class AII in 2D}
\label{sec:restop2D}
We would now like to turn our attention to time reversal symmetric 2D insulators, in symmetry class AII \cite{Altland}.
For states in symmetry class A with non-vanishing first Chern number, so called Chern insulators \cite{QAH}, only algebraically decaying WFs can be found. As a consequence, Chern insulators with exponentially localized or even compactly supported WFs cannot exist.
However, the situation is less obvious for TRS protected topological insulators, a.k.a.\ quantum spin Hall (QSH) insulators \cite{KaneMele2005a,KaneMele2005b,BHZ2006,koenig2007}. The conceptually simplest representative of this topological equivalence class consists of two TRS conjugated copies of a Chern insulator with opposite odd Chern number, one copy for each spin block
{(cf.\ Ref.\ \cite{Matt})}. While the individual spin blocks have non-zero Chern number, the total set of occupied bands has zero Chern number as required by TRS.
Hence, a smooth gauge mixing the two TRS conjugated blocks can be found for the Bloch functions \cite{Soluyanov2011}.
{Here we} would like to consider a minimal model for a QSH insulator analogous to the one introduced in Ref.\ \cite{BHZ2006} which has $m=4$ degrees of freedom per site and $n=2$ occupied bands. The four degrees of freedom are labeled by the basis vectors $\vert e,\uparrow \rangle,\vert h,\uparrow \rangle,\vert e,\downarrow \rangle,\vert h,\downarrow \rangle$. We denote the $e-h$ pseudo spin by $\sigma$ and the real spin by $s$. The Bloch Hamiltonian of the spin up block reads as
\begin{align}
& h_{\uparrow}(k)=d_{\uparrow}^i(k)\sigma_i,\quad d_{\uparrow}^1(k)=\sin(k_x),\nonumber\\
&d_{\uparrow}^2(k)=\sin(k_y),\quad d_{\uparrow}^3(k)=M-\cos(k_x)-\cos(k_y).
\label{eqn:BHZ}
\end{align}
The Hamiltonian of the TRS conjugated block is then {defined} by $h_\downarrow(k)=h^*_\uparrow(-k)$. This model is topologically nontrivial for $0<\lvert M\rvert < 2$ and trivial for $\lvert M\rvert >2$. The projection onto the occupied bands $\mathcal P_k$ can {for each $k$} be written as a sum of
\begin{equation}
\mathcal P_k^\uparrow =\frac{1}{2}\left(1-\hat d_\uparrow^i(k)\sigma_i\right)\otimes\lvert \uparrow\rangle\langle \uparrow \rvert
\end{equation}
and
\begin{equation}
\mathcal P_k^\downarrow =\frac{1}{2}\left(1-\hat d_\downarrow^i(k)\sigma_i\right)\otimes\lvert \downarrow\rangle\langle \downarrow \rvert.
\end{equation}
A smooth gauge of Bloch functions ${k\mapsto \lvert u_i(k)\rangle}$, $i=1,2$, can be found in a generic way \cite{VanderbiltReview}. One first chooses a set of trial orbitals $\lvert \tau_i\rangle,~i=1,2${,}
which are random linear combinations of the four basis orbitals. Projecting onto the occupied bands yields $\lvert \gamma_i(k)\rangle=\mathcal P_k \lvert \tau_i\rangle$. If the
{family of} Gram {matrices}
{with entries}
\begin{equation}
S_{ij}(k)=\langle \gamma_i(k)\vert \gamma_j(k)\rangle
\end{equation}
is regular for all $k$, smooth Bloch functions {defined as}
\begin{equation}
{k\mapsto \lvert u_i(k) \rangle=S^{-{1}/{2}}_{j,i}(k)
\lvert \gamma_j\rangle }
\end{equation}
can be obtained. In practice, by trying a few random choices, a gauge for which ${\det(S(k) )\ge 10^{-2}}~\forall k$ can be readily found.
The associated WFs are then obtained by Fourier transform. \resub{Note that these WFs, while still spanning the same many-body state of occupied bands, individually break all symmetries present in Eq. (\ref{eqn:BHZ}) due to the random choice of $\tau_i$}.
We employ the above prescription to find exponentially decaying WFs both in the topologically trivial and nontrivial regime on a lattice of $N=101\times 101$ sites. These WFs are then used as starting points for the adiabatic continuity algorithm introduced in Section \ref{sec:algtop}. For WFs associated with topologically trivial insulators, i.e., $\lvert M\rvert>2$, our algorithm finds a set of atomic WFs representing the most localized topologically trivial insulator to impressive numerical accuracy (see Fig. \ref{fig:movie2D}).
\begin{figure*}
\centering
\includegraphics[width=.8\linewidth]{Movie2DTrivWide2.pdf}
\caption{(Color online) Logarithmic plot of the probability density $\rho$ of an initial Wannier function for the model Hamiltonian (\ref{eqn:BHZ}) for the topologically trivial mass parameter $M=2.5$ (leftmost panel). The home cell of the WFs is $(x,y)=(51,51)$, the size of the lattice is $101\times101$. Adiabatically deformed WF after 100, 500, and 727 (rightmost panel) iterations iterations respectively with $\xi=\kappa=\lambda=50$.}
\label{fig:movie2D}
\end{figure*}
However, as soon as the initial set of WFs corresponds to a non-trivial QSH state, the algorithm does not find a compactly supported set of WFs. This result gives numerical evidence that a simple flat band model Hamiltonian with finite range hopping does not exist for the QSH state in contrast to the 1D TSC. The relation between flat band models with finite range hopping and compact WFs becomes clear from the following representation of the projection $\mathcal P_k$ onto the occupied bands at momentum $k$ in terms of Wannier functions $w^\alpha_0, \alpha=1,\ldots \je{,} n$ centered around the origin,
\begin{align}
\mathcal P_k = \sum_{\alpha=1}^n \sum_{r,r'}\text{e}^{ik(r-r')}w_0^\alpha(r) w_0^{\alpha \dag}(r').
\label{eqn:PfromWanniers}
\end{align}
An exact flat band Hamiltonian where all occupied states have energy $-\epsilon$ and all empty states have energy $+\epsilon$ is then immediately obtained as
\begin{align}
Q(k) = (+\epsilon) (1-\mathcal P_k) + (-\epsilon) \mathcal P_k=\epsilon\left(1-2\mathcal P_k\right).
\end{align}
To see if our findings are sensitive to the number of bands \resub{or to the spin rotation symmetry of Eq. (\ref{eqn:BHZ})}, we also applied the adiabatic continuity algorithm to a QSH model with 8 bands \resub{and spin mixing terms} which did not yield qualitatively different results.
\subsection{{Dissipative state preparation}}
The idea of dissipative state preparation \cite{DiehlStatePrep} in the context of topological states of matter \cite{DiehlTopDiss} relies, for pure and translation invariant target states, on the existence of a complete set of fermionic creation and annihilation operators $w_{i,\alpha},w_{i,\alpha}^\dag$ forming a Dirac algebra (the indices referring to sites and bands, respectively). In this case, the stationary state of a dissipative evolution described by a Lindblad master equation
\begin{eqnarray}
{\frac{\partial}{\partial t}}\rho = \kappa \sum_{i,\alpha} \left(w_{i,\alpha} \rho w^\dag_{i,\alpha} - \tfrac{1}{2} \{w^\dag_{i,\alpha} w_{i,\alpha} ,\rho\}\right)
\end{eqnarray}
with damping rate {$\kappa>0$}, will be identical to the ground state of the dimensionless Hamiltonian
\begin{equation}
H = \sum_{i,\alpha}h_{i,\alpha},~ h_{i,\alpha} = w^\dag_{i,\alpha} w_{i,\alpha},
\end{equation}
with mutually commuting $h_{i,\alpha}$. In typical implementations of such a dissipative dynamics in the context of cold atomic systems, the Lindblad operators $w_{i,\alpha}$ generating the dissipative dynamics are quasi-local, i.e. have a compact support on the underlying lattice \cite{BardynTopDiss}. Our algorithm is precisely constructed to find such compactly supported operators $w_{i,\alpha}$, with the mutual commutativity of the associated $h_{i,\alpha}$ being granted by the shift orthogonality of the Wannier functions \resub{corresponding to the Lindblad operators $w_{i,\alpha}$}. Unlike the one dimensional case of the topologically nontrivial ground state of Kitaev's quantum wire, where a representative with compactly supported Wannier functions exists and is indeed found by our algorithm, our results in two dimensions imply the absence of an analogous situation in two dimensions.
\section{Conclusion and outlook}
\label{sec:conclusion}
In this work, we have presented a method to search for localized Wannier functions of free quantum lattice models which explicitly takes into account the symmetry of the problem. Most interestingly, we could extend the domain of this search algorithm from
individual model Hamiltonians to entire topological equivalence classes. This allows for a numerical detection of the most localized representative of a given topological state. We did so by elaborating on a compressed sensing approach built upon Bregman split techniques, where the spatial locality takes the role of the sparsity of the problem (see Refs.\ \cite{OszolinsCompressedModes,OszolinsTranslation} ).
We close our presentation by providing some perspectives opened up by our present analysis, including a few particularly intriguing implications and applications of our new algorithm
\je{beyond the most widely known applications \cite{VanderbiltReview} of having localized Wannier functions available.}
\subsection{{Diagnostic tool of topological states}}
The possibility to identify localized Wannier functions not only for given model Hamiltonians,
but also -- if the energy functional is set to zero along with $\xi\rightarrow 0$ -- maximally localized Wannier functions within entire
topological equivalence classes opens up another interesting application of our work: That of a {\it diagnostic tool}: Whenever it converges
\je{to a compactly supported Wannier function},
it identifies a "sweet spot" characterizing the topological class of the initial Hamiltonian itself rather than minimizing the energy of a certain model. The flow towards the atomic insulator and the topological flat band (Kitaev) superconductor, starting from generic states within the same topological phase provide striking examples of this. But the parameter $\xi>0$ can be freely chosen, reflecting the $l_1$-regularization in terms of compressed sensing. \je{In condensed matter terms, this parameter allows for a precise trade-off between locality and energy.}
This freedom gives rise to a useful ``knob'' to tune, and for applications in the context of e.g., ab initio band structure calculations, a finite $\xi$ is more appropriate.
\subsection{{Applications in devising tensor network methods}}
Thinking further about our algorithm as a flow in the renormalization group sense is likely to be fruitful also in the context of interacting as well as disordered systems. In fact our protocol bears some (non-accidental) resemblance with tensor network algorithms (quantum state renormalization methods) such as DMRG and TEBD in one dimension and PEPS and MERA more generally
\cite{R1,R2,R3}.
More specifically, it seems that in order to simulate weakly interacting (and/or disordered) fermionic lattice models,
the efficiently localized Wannier functions which are still orthogonal appear to be a very suitable starting point for devising
variational sets of states, as real space operators remain short {ranged} (and close to diagonal) when projected to the pertinent electronic band. Most saliently, tensor network approaches augmented with an initial preferred basis selection based on our algorithm appear particularly promising in two-dimensional
approaches, where having a low bond dimension in PEPS approaches is critical for the highly costly (approximate) tensor network contraction.
More specifically, two approaches seem interesting: In a first, one takes a weakly interacting model and re-expresses the non-interacting part in the Wannier basis
found by the algorithm. If the Wannier functions are exactly localized, then the new Hamiltonian will still be local. This \je{can then serve as}
an ansatz for a tensor network approach including
interactions. In a second, one starts from a generalized mean field approach for the interacting model,
generates Wannier functions and then applies a variational
tensor network method.
\subsection{Symmetry breaking by truncation of exponential tails}
Finally, a fundamental question arises due to the apparent lack of \resub{compactly supported} Wannier functions for the quantum spin Hall phase, namely that of the importance of exponentially decaying tails. We have found that any truncation of the tail of the Wannier functions inevitably leads to the breaking of time-reversal symmetry at a corresponding rate. In fact, cutting exponential tails seems continuous, but the QSH phase can be left continuously by breaking TRS. Despite being a conceptual problem it may not be a practical one. In any solid-state realization of a finite size QSH insulator, there will be weak TRS breaking terms, yet the physical response can -- at least in principle -- be experimentally indistinguishable from that of a truly TRS invariant system. In this sense, even though the Wannier functions with compact support and formally do not represent a QSH phase, they may still be used for practical purposes.
Our algorithm provides a tool to systematically assess these questions. Yet these are merely a few of many intriguing directions, and we anticipate that our findings will inspire future research in diverse branches of physics, as well as in applied mathematics.\\
\jcb{\emph{Note added. A key result of the present paper is the use of {\emph{local}} orthogonality constraints on the Bloch functions. In this context, we note the recent arXiv submissions by Barekat \emph{et al.} \cite{Barekat1, Barekat2}. In Ref. \cite{Barekat1}, Barekat {\emph{et al.}} independently derive a similar algorithm with the same asymptotic scaling. In Ref. \cite{Barekat2}, the same authors use orthogonality constraints in terms of Bloch functions in the context of certain (topologically trivial) band structures. These papers do not address the maximally localized representatives of topological equivalence classes of band structures
which is the main focus of our present work.}}
\section{Acknowledgements}
We would like to thank C.\ Krumnow and H.\ Wilming for discussions. \eb{We also thank V. Ozolins for helpful correspondence on Refs. \cite{OszolinsCompressedModes,OszolinsTranslation} and for making us aware of Ref. \cite{Barekat1}.}
{Support from the ERC \je{Synergy Grant} UQUAM and \je{Consolidator Grant TAQ}, the EU (RAQUEL, SIQS, \je{COST}),
the BMBF (QuOReP), the START Grant No. Y 581-N16, the SFB FoQuS (FWF Project No. F4006- N16) and DFG's Emmy Noether program (BE 5233/1-1) is gratefully acknowledged.}
\bibliographystyle{apsrev}
|
\section{Introduction}
\label{sec:int}
Dark matter (DM) provides a strong connection between the two
phenomenologically rich arenas: particle astrophysics and beyond
Standard Model (SM) physics. While the existence of DM is part of the
standard model of cosmology, its particle physics origins are largely
unknown. The WIMP (weakly interacting massive particle) miracle
however provides a tantalizing hint that DM is associated with
new physics (NP) at the weak scale, and such candidates should be
accessible to various ongoing experiments. Signals at these
experiments depend strongly on the nature of interactions of the DM
with SM fields, and are less sensitive to other details of the model.
This motivates the study of simplified models, which minimally extend
the SM to include couplings of DM particles with the SM. Each
simplified model can then capture the dark matter phenomenology of a
wide range of models.
Once we consider different classes of simplified models, one new
category of models arises in analogy with SM flavor:
flavored
DM~\cite{Kile:2011mn,Kamenik:2011nb,Batell:2011tc,Agrawal:2011ze,
Batell:2013zwa,Kile:2013ola,Lopez-Honorez:2013wla,Kumar:2013hfa,
Zhang:2012da}.
In this setup DM particles come in multiple copies, and
have a non-trivial flavor structure in their couplings with quarks and
leptons.\footnote{An alternative scenario in which dark matter arises from a discrete $A_4$ symmetry in the neutrino sector has been considered in \cite{Hirsch:2010ru,Boucenna:2011tj}.}
This framework does show up in a very
specific way in supersymmetric models as sneutrino DM models
\cite{Ibanez:1983kw, Ellis:1983ew, Hagelin:1984wv, Goodman:1984dc,
Freese:1985qw, Falk:1994es,MarchRussell:2009aq}, but clearly there are
more general possibilities.
This class of models is constrained, like other DM models, by both
indirect and direct detection DM experiments as well as collider
searches. The relevant schematic interaction responsible for these
signatures is shown in the left panel of figure \ref{fig:DMloop}.
Additionally precision flavor experiments have to be taken into
account due to the flavor violation introduced by the dark sector.
Schematically this contribution is displayed in the right panel of
figure \ref{fig:DMloop}, adding a new class of diagrams to the
well-studied DM-SM interaction.
\begin{figure}[h!]
\centering
\includegraphics[width=0.35\textwidth]{figures/figdmdmsmsm.pdf}\qquad\qquad
\includegraphics[width=0.35\textwidth]{figures/figdmloop.pdf}
\caption{Schematic diagrams contributing to experimental constraints on flavored DM.}
\label{fig:DMloop}
\end{figure}
Flavored dark matter models can have significantly distinct
phenomenology. For indirect detection experiments, the spectrum of
photons and leptons arising from DM annihilation depends on the
relative annihilation into various final states. For example, it was
shown that a DM candidate annihilating exclusively to $b$-quarks
provides a good fit to the spectrum of excess photons observed in a
recent analysis of Fermi-LAT data from the galactic
center~\cite{Daylan:2014rsa,
Berlin:2014tja,Agrawal:2014una,
Izaguirre:2014vva,
Boehm:2014hva,
Ipek:2014gua,
Kong:2014haa,
Ko:2014gha,
Boehm:2014bia,
Abdullah:2014lla,
Ghosh:2014pwa,
Martin:2014sxa,
Berlin:2014pya,
Basak:2014sza,
Modak:2013jya} . Direct detection predictions for scattering vary
widely depending upon whether the ambient DM particles couple to the
first generation quarks directly or not. The absence of direct
detection signals so far then point to the possibility of suppression
of such a coupling, which can be achieved through either a loop
suppression or a small mixing angle.
Collider searches for DM are also sensitive to the DM couplings to
various quark flavors, both in terms of the DM production cross section,
as well as the flavor pattern of visible final states which
can be produced in association with the DM.
While some of these effects have been explored, the study of flavor
phenomenology has largely been restricted to elaborate models such as
the MSSM. Previous analyses often assume for simplicity universality
or minimal flavor violation (MFV)
\cite{Chivukula:1987py,Hall:1990ac,Buras:2000dm,D'Ambrosio:2002ex,Buras:2003jf},
so that flavor changing neutral current (FCNC) effects are
automatically suppressed. On one hand this is welcome due to the
good agreement of the flavor data with the SM prediction, but on the
other hand interesting effects in the flavor sector are eliminated.
In this paper we abandon the MFV principle and consider instead a
general flavor violating coupling of DM particles with quarks. DM is
introduced as a triplet under a new global flavor symmetry
$U(3)_\chi$. While in our analysis the coupling matrix (denoted by
$\lambda$) is taken to be completely general, we make one simplifying
assumption that turns out to be helpful in various respects. We impose
that $\lambda$ is the only new source of flavor breaking, in addition
to the SM Yukawa couplings. As this assumption generalizes
the MFV principle to the DM sector, we call it {\it Dark Minimal
Flavor Violation} (DMFV). We will point out the following features of
DMFV:
\begin{itemize}
\item The DMFV framework, while bearing some conceptual similarity to MFV,
goes well beyond the latter framework, as it allows for large FCNC
effects. The structure of $\lambda$ needs to be determined from the
available constraints.
\item
The DMFV ansatz naturally preserves a residual $\mathbb{Z}_3$
symmetry, which guarantees the stability of the DM particle.
\item
DMFV significantly reduces the number of new parameters in the Lagrangian, as the DM mass term $m_\chi$ must be flavor conserving up to corrections of the form $\lambda^\dagger \lambda$.
\item
DMFV guarantees ``flavor-safety'' of the UV complete theory. It is therefore sufficient to identify flavor-safe scenarios for the structure of $\lambda$ within the simplified model framework.
\end{itemize}
In the phenomenological part of our paper we will restrict ourselves
to the study of the simplest version of DMFV, which we refer to as the
minimal DMFV (mDMFV) model in order to distinguish it from the more
general framework. The DM is taken to be a Dirac fermion $\chi$, interacting
with the right-handed down-type quarks via the coupling
\begin{align}
\lambda \bar d_R \chi \phi
\end{align}
with a scalar mediator $\phi$. While leaving the question of a
possible UV completion unanswered, this study captures the most
important phenomenological effects accessible to current experiments.
Our studies extend the existing literature on the phenomenology of
flavored DM in the following ways:
\begin{itemize}
\item
{We go beyond the simple MFV hypothesis that automatically suppresses
all flavor effects to an acceptable level. Instead we study the
implications of a completely general coupling matrix $\lambda$,
embedded in the DMFV ansatz, and derive its structure from the
experimental constraints. }
\item
We consider a large number of relevant precision observables which can
potentially be affected by the mDMFV model. These are in
particular the constraints from meson-antimeson mixing, radiative and
rare $B$ and $K$ decays, electroweak precision observables and
electric dipole moments.
\item
From the analysis of meson-antimeson mixing observables we identify a
number of ``flavor-safe'' scenarios for the structure of $\lambda$. These
scenarios will be useful for future studies of flavored DM
models beyond MFV, as they can be imposed simply and render detailed
re-analyses unnecessary.
\item
Subsequently we perform a simultaneous analysis of flavor and DM
constraints, such as the relic abundance from thermal
freeze-out, and direct detection data from LUX~\cite{Akerib:2013tjd}.
While restricting ourselves to the phenomenologically interesting case
of $b$-flavored DM, we consider several mass hierarchies in the
dark sector, i.\,e.\ large and small splittings between the DM
particle and the heavier flavors.
\item
We reveal a non-trivial interplay of the complementary flavor and DM
constraints, such that the combined constraint on the parameter
space of $\lambda$ turns out to be interesting intersections of the
individual ones. This result underlines the importance of taking into
account the various constraints simultaneously.
\item
We point out a cancellation between various mDMFV one-loop contributions
(photon penguin and box diagram) to the WIMP-nucleon scattering,
occurring for a certain range of coupling parameters. As the photon
penguin is only present for scattering off protons, while the box
diagram contributes to proton and neutron scattering cross-sections,
this cancellation provides a possible realization of Xenophobic
DM~\cite{Feng:2011vu,Feng:2013vod}.
\item
We review the constraints from collider searches on the mDMFV model
with $b$-flavored DM. The most stringent bounds are placed by
searches for bottom squark pair production, constraining the parameter
space of the model up to a mediator mass $m_\phi\sim 800-900\, {\rm GeV}$. Monojet
searches can be important for very compressed spectra, or for very
heavy $\phi$ such that its direct production is suppressed.
\end{itemize}
Our paper is organized as follows. In section \ref{sec:model} we
introduce the concept of Dark Minimal Flavor Violation (DMFV), and
describe the minimal model realizing this hypothesis, the mDMFV model.
Section \ref{sec:DMFV} deals with the implications of the DMFV
hypothesis that are valid beyond the minimal model. In section
\ref{sec:DF2constraints} we provide the formalism for a detailed study
of the constraints from meson anti-meson mixing on the mDMFV model. We
also consider potential new contributions to radiative and rare $B$
and $K$ decays, electroweak precision observables and electric dipole
moments and find all of these observables to be SM-like.
Section \ref{sec:DF2numerics} is devoted to a detailed numerical
analysis of the constraints on the coupling matrix $\lambda$ arising
from meson-antimeson mixing. We identify a number of ``flavor-safe'' scenarios for the coupling matrix $\lambda$. In section \ref{sec:flavor-consequences}
we provide a comprehensive summary of the results of the numerical
flavor analysis and the different scenarios emerging for the analysis
of DM constraints. In section \ref{sec:DMpheno} we study the
DM phenomenology of the mDMFV model, considering both the
relic abundance constraint from thermal freeze-out and the emerging
WIMP-nucleon cross section observed in direct detection experiments. A
combined numerical analysis of flavor and DM constraints is
performed in section \ref{sec:numerics}, studying the various possible
mass hierarchies in turn. In section \ref{sec:collider} we estimate
the constraints on the mDMFV model from the LHC, stemming in
particular from monojet searches and searches for supersymmetric
bottom squarks. We also mention some new signatures for these models.
In section \ref{sec:conclusions} we summarize our results. Some
technical details are relegated to the appendices.
\section{Flavored dark matter beyond MFV -- a minimal model}\label{sec:model}
We consider a setup where DM $\chi$ transforms in the
fundamental representation of a new flavor symmetry $U(3)_\chi$, in
analogy with the SM flavor symmetry. While we posit this symmetry as
an ansatz, it will be an interesting future direction to study
possible UV completions (e.g. extended Grand Unified Theories) which
incorporate this structure.
We assume that the global
\begin{equation}\label{eq:flavor-group}
U(3)_q \times U(3)_u \times U(3)_d \times U(3)_\chi
\end{equation}
flavor symmetry is broken only by the SM Yukawa couplings $Y_u$,
$Y_d$ and the DM-quark coupling $\lambda$. This ansatz generalizes the
MFV hypothesis \cite{D'Ambrosio:2002ex,Buras:2000dm,Buras:2003jf,Chivukula:1987py,Hall:1990ac} to include an extra
$U(3)_\chi$ symmetry under which the DM field transforms, and an
additional Yukawa coupling $\lambda$. We refer to this
assumption as {\emph{Dark Minimal Flavor Violation (DMFV)}.
Depending on the type of quark to which the DM couples, different
classes of DMFV can be defined, see appendix
\ref{app:DMFV-classification} for details.
In what follows we restrict
ourselves to the coupling of $\chi$ to right-handed down-type quarks
via a scalar mediator
$\phi$. While the DM particle $\chi$ is a gauge singlet, the mediator
$\phi$ has to carry color and hypercharge.
This helps to keep the model simple, since no further electroweak
structure is required when assuming the new particles to be singlets
under $SU(2)_L$. Further the choice of down-type quarks ensures to
have an effect in relevant flavor observables such as {$K$ and
$B_{d,s}$ meson mixing} and well-measured rare decays.
The most general renormalizable Lagrangian including the minimal field
content is then given by
\begin{eqnarray}
{\cal L}&=& \mathcal{L}_\text{SM}+
i \bar \chi \slashed{\partial} \chi
- m_{\chi} \bar \chi \chi - (\lambda_{ij} \bar {d_{R}}_i \chi_j \phi + {\rm h.c.}) \nonumber \\
&& \qquad\,
+
(D_{\mu} \phi)^{\dagger} (D^{\mu} \phi) - m_{\phi}^2 \phi^{\dagger} \phi
+\lambda_{H \phi}\, \phi^{\dagger} \phi\, H^{\dagger} H
+\lambda_{\phi\phi}\, \phi^{\dagger} \phi\, \phi^{\dagger} \phi \,,\label{eq:Lagrangian}
\end{eqnarray}
with the symmetry transformation properties summarized in
table~\ref{tab:representations}.
Note that the $U(3)_{\chi}$ flavor symmetry in the DM sector
guarantees that at the Lagrangian level all three DM flavors have the
same mass $m_{\chi}$, although they acquire a small splitting from
higher order DMFV corrections. In what follows we refer to this model
as the {\emph{minimal DMFV (mDMFV) model}.
\begin{table}
\begin{center}
\begin{tabular}{|c||ccc|cccc|}
\hline
& $SU(3)_c$ & $SU(2)_L$ & $U(1)_Y$ & $U(3)_q$ & $U(3)_u$ & $U(3)_d$ & $U(3)_\chi$ \\ \hline\hline
$q_L$ & {\bf 3} & {\bf 2} & 1/6 & {\bf 3} & {\bf 1} & {\bf 1} & {\bf 1} \\
$u_R$ & {\bf 3} & {\bf 1} & 2/3 & {\bf 1} & {\bf 3} & {\bf 1} & {\bf 1} \\
$d_R$ & {\bf 3} & {\bf 1} & -1/3 & {\bf 1} & {\bf 1} & {\bf 3} & {\bf 1} \\ \hline
$\ell_L$ & {\bf 1} & {\bf 2} & -1/2 & {\bf 1} & {\bf 1} & {\bf 1} & {\bf 1} \\
$e_R$ & {\bf 1} & {\bf 1} & -1 & {\bf 1} & {\bf 1} & {\bf 1} & {\bf 1} \\ \hline
$H$ & {\bf 1} & {\bf 2} & 1/2 & {\bf 1} & {\bf 1} & {\bf 1} & {\bf 1} \\ \hline
$\phi$ & {\bf 3} & {\bf 1} & -1/3 & {\bf 1} & {\bf 1} & {\bf 1} & {\bf 1} \\
$\chi_L$ & {\bf 1} & {\bf 1} & 0 & {\bf 1} & {\bf 1} & {\bf 1} & {\bf 3} \\
$\chi_R$ & {\bf 1} & {\bf 1} & 0 & {\bf 1} & {\bf 1} & {\bf 1} & {\bf 3} \\ \hline\hline
$Y_u$ & {\bf 1} & {\bf 1} & 0 & {\bf 3} & \boldmath{${\bar 3}$} & {\bf 1} & {\bf 1} \\
$Y_d$ & {\bf 1} & {\bf 1} & 0 & {\bf 3} & {\bf 1} & \boldmath{${\bar 3}$} & {\bf 1} \\
$\lambda$ & {\bf 1} & {\bf 1} & 0 & {\bf 1} & {\bf 1} & {\bf 3} & \boldmath{${\bar 3}$} \\ \hline
\end{tabular}
\end{center}
\caption{Symmetry transformation properties of the minimal DMFV matter content and the Yukawa spurions.\label{tab:representations}}
\end{table}
The mDMFV model has some similarities to simplified models of supersymmetry and should be understood in an analogous manner. In contrast to the SUSY case however in mDMFV the flavor charge is carried by the DM fermions and not by the scalar mediator. Further we assume $\chi$ to be a Dirac fermion (a Majorana mass term would violate the $U(3)_\chi$ symmetry), while in the minimal SUSY models the gauginos are Majorana.
We stress that the DMFV ansatz, in contrast to the MFV ansatz,
potentially allows for large flavor violating effects. A careful
analysis of FCNC constraints is therefore necessary.
\section{Implications of the Dark Minimal Flavor Violation hypothesis}\label{sec:DMFV}
In the present section we consider the consequences of the DMFV
ansatz. We stress that these implications go beyond the simple mDMFV
model introduced in section~\ref{sec:model} and hold in any scenario
with the same DMFV flavor symmetry breaking pattern.
\subsection[New flavor violating parameters and a convenient parametrization for $\lambda$]{\boldmath New flavor violating parameters and a convenient parametrization for $\lambda$}
In the DMFV setup the flavor symmetry in the quark sector is broken only by the SM Yukawa couplings $Y_u$, $Y_d$ and the DM-quark coupling~$\lambda$. In
a first step the SM flavor symmetry can be used to remove unphysical parameters from the SM Yukawas. They can be parametrized as usual in terms of the six quark masses and the CKM matrix, signaling the misalignment between $Y_u$ and $Y_d$.
In the second step we remove unphysical parameters from the coupling matrix $\lambda$.
Being an arbitrary complex matrix, it contains at first 9 real parameters and 9 complex phases. Some of them can be removed by making use of the DM flavor symmetry $U(3)_\chi$.
We start by parametrizing $\lambda$ in terms of a singular value decomposition
\begin{equation}\label{eq:sing-val}
\lambda= U_{\lambda} D_{\lambda} V_\lambda\,,
\end{equation}
where $D_\lambda$ is a diagonal matrix with real and positive entries,
and $U_{\lambda}$ and $V_\lambda$ are unitary matrices. Note that
$U_{\lambda}$ and $V_\lambda$ are not uniquely defined, as $\lambda$
is invariant under the diagonal rephasing
\begin{equation}
U'_{\lambda} = U_{\lambda}\diag(e^{i\theta_1},e^{i\theta_2},e^{i\theta_3})\,,\qquad
V'_{\lambda} = \diag(e^{-i\theta_1},e^{-i\theta_2},e^{-i\theta_3})V_{\lambda}\,.
\end{equation}
We use this freedom to reduce the number of phases in $U_{\lambda}$ to
three. Then $\lambda$ has 9 real parameters and 9 phases in the
parametrization \eqref{eq:sing-val}.
We can now use the $U(3)_\chi$ invariance to fully remove the unitary matrix $V_\lambda$. Consequently we are left with the matrix
\begin{equation}\label{eq:sing-val-red}
\lambda= U_{\lambda} D_{\lambda} \,.
\end{equation}
It contains nine parameters: three non-negative elements of
$D_{\lambda}$, and three mixing angles and three CP violating phases
in $U_{\lambda}$. Note that the mixing angles are restricted to the
range $0\le\theta_{ij}^\lambda\le\pi/4$ in order to avoid a
double-counting of parameter space.
This choice ensures that each DM flavor couples dominantly to the
quark of the same generation. For instance we can refer to $\chi_3$ as
$b$-flavored DM.
A convenient parametrization for $U_{\lambda}$ has been derived in \cite{Blanke:2006xr} in the context of the Littlest Higgs model with T-parity. It can be written as
\begin{eqnarray}
\addtolength{\arraycolsep}{3pt}
U_\lambda &=& U_{23}^\lambda U_{13}^\lambda U_{12}^\lambda \nonumber\\
&=&
\begin{pmatrix}
1 & 0 & 0\\
0 & c_{23}^\lambda & s_{23}^\lambda e^{- i\delta^\lambda_{23}}\\
0 & -s_{23}^\lambda e^{i\delta^\lambda_{23}} & c_{23}^\lambda\\
\end{pmatrix}
\begin{pmatrix}
c_{13}^\lambda & 0 & s_{13}^\lambda e^{- i\delta^\lambda_{13}}\\
0 & 1 & 0\\
-s_{13}^\lambda e^{ i\delta^\lambda_{13}} & 0 & c_{13}^\lambda\\
\end{pmatrix}
\begin{pmatrix}
c_{12}^\lambda & s_{12}^\lambda e^{- i\delta^\lambda_{12}} & 0\\
-s_{12}^\lambda e^{i\delta^\lambda_{12}} & c_{12}^\lambda & 0\\
0 & 0 & 1\\
\end{pmatrix}\,,\qquad
\end{eqnarray}
where $c_{ij}^\lambda = \cos\theta_{ij}^\lambda$ and $s_{ij}^\lambda = \sin\theta_{ij}^\lambda$.
Performing the product one obtains the expression
\begin{equation}
\addtolength{\arraycolsep}{3pt}
U_\lambda= \begin{pmatrix}
c_{12}^\lambda c_{13}^\lambda & s_{12}^\lambda c_{13}^\lambda e^{-i\delta^\lambda_{12}}& s_{13}^\lambda e^{-i\delta^\lambda_{13}}\\
-s_{12}^\lambda c_{23}^\lambda e^{i\delta^\lambda_{12}}-c_{12}^\lambda s_{23}^\lambda s_{13}^\lambda e^{i(\delta^\lambda_{13}-\delta^\lambda_{23})} &
c_{12}^\lambda c_{23}^\lambda-s_{12}^\lambda s_{23}^\lambda s_{13}^\lambda e^{i(\delta^\lambda_{13}-\delta^\lambda_{12}-\delta^\lambda_{23})} &
s_{23}^\lambda c_{13}^\lambda e^{-i\delta^\lambda_{23}}\\
s_{12}^\lambda s_{23}^\lambda e^{i(\delta^\lambda_{12}+\delta^\lambda_{23})}-c_{12}^\lambda c_{23}^\lambda s_{13}^\lambda e^{i\delta^\lambda_{13}} &
-c_{12}^\lambda s_{23}^\lambda e^{i\delta^\lambda_{23}}-s_{12}^\lambda c_{23}^\lambda s_{13}^\lambda e^{i(\delta^\lambda_{13}-\delta^\lambda_{12})} &
c_{23}^\lambda c_{13}^\lambda\\
\end{pmatrix}.\
\end{equation}
Finally it turns out to be convenient to parametrize the diagonal matrix $D_\lambda$ as
\begin{equation}\label{eq:Dlambda}
D_\lambda \equiv \diag(D_{\lambda,11},D_{\lambda,22},D_{\lambda,33})=
\lambda_0\cdot \mathbbm{1}+\diag(\lambda_1,\lambda_2,-(\lambda_1+\lambda_2))\,.
\end{equation}
The first parametrization is useful for the analysis of DM and collider constraints. The second parametrization instead is better suited for the flavor analysis, since it quantifies the deviations from a flavor universal coupling.
\subsection{Non-DMFV contributions and dark matter stability}
The flavor structure of the SM is accidental --- there exist no other
gauge-invariant operators beyond the Yukawa terms at the
renormalizable level. It is then worth asking if the DMFV ansatz
can also arise naturally in an analogous way.
In this section we study how generic the DMFV ansatz is from a UV
point-of-view. We stress however that the goal of this work is merely
to study the novel phenomenology arising from this ansatz, and a
complete UV model is beyond the current scope. We merely study the
corrections to the ansatz to the extent that they can affect low
energy phenomenology, particularly DM decay.
Interestingly, in the exact DMFV limit, all operators inducing decay
of the $U(3)_\chi$ triplet $\chi$ are forbidden, even at the
non-renormalizable level. In analogy to the stability of DM
in the MFV case \cite{Batell:2011tc} it can straightforwardly be
shown -- see appendix \ref{app:Z3} for details -- that the flavor
symmetry \eqref{eq:flavor-group} broken only by the Yukawa couplings
$Y_u$, $Y_d$ and $\lambda$, together with $SU(3)_\text{QCD}$ imply an
unbroken $\mathbbm{Z}_3$ symmetry. It is then natural to impose this
$\mathbb{Z}_3$ symmetry as exact, under which only the new particles
$\chi_i$ and $\phi$ are charged, and transform as
\begin{align}
\chi_i \to e^{2\pi i/3} \chi_i, \qquad
\phi \to e^{-2\pi i/3} \phi
\ .
\end{align}
This symmetry prevents the decay of any of
these states into SM particles only, and therefore renders the
lightest state stable.
We now estimate the size of non-DMFV effects that can arise.
We imagine a UV scale, $\Lambda$, above which the DM flavor
symmetry $U(3)_\chi$ is unbroken. While this scale could in principle
be associated with the SM flavor scale as well, for simplicity we assume that
the SM flavor structure is generated at a higher scale.
Generically, we expect all operators allowed by symmetries to be
generated at the scale $\Lambda$. The most important
contributions at low energy arise from relevant and marginal
operators.
The relevant operator
\begin{align}
\mathcal{O}_m
&=
\delta m_{ij} \, \bar{\chi}_i \chi_j
\end{align}
is the leading operator that is generated. It maximally violates the
DMFV ansatz, while preserving the $\mathbb{Z}_3$. This operator can be
prevented from being generated at the scale $\Lambda$ if the mass of
the DM fermions is generated at a
lower scale, through a flavor-blind sector. Note the analogy to
flavor-blind SUSY breaking, which yields MFV. It is an open question
whether such a scenario can be achieved simply in this framework. We
assume henceforth that this operator is negligible.
At the marginal level, we generate the following two operators,
\begin{align}
\mathcal{O}_L
&= \bar{\ell}_L \chi H + \mathrm{h.c.}
\\
\mathcal{O}_B
&= \bar{q}_L (i\sigma^2){q}_L^\dagger \phi^\dagger
+\mathrm{h.c.},
\end{align}
which are lepton and baryon number violating respectively. These are
prohibited by the discrete $\mathbb{Z}_3$ symmetry, however.
We see that additional discrete symmetries can be imposed in order to
prevent the DM from decaying, and preventing non-DMFV
contributions at the renormalizable level. Higher dimensional non-DMFV
operators can then only modify other aspects of phenomenology, which
are not as severely constrained. Therefore, for a reasonably high
scale $\Lambda$, these operators are not expected
to alter the phenomenology appreciably.
\subsection{Mass splitting in the dark sector}
As noted in section \ref{sec:model} the DMFV hypothesis ensures that to leading order in the coupling $\lambda$, the
masses for different DM particles are equal. There are three
potential sources for splittings.
Firstly, there can be contributions to the mass matrix $m_\chi$ directly
violating DMFV. We assume that such contributions are absent. Secondly,
higher dimensional DMFV-violating contributions can still induce splittings, but
these are expected to be suppressed by the heavy scale where DMFV is broken.
An unavoidable contribution is through the renormalization group running, where a universal
mass at the high scale is renormalized by the presence of the DM coupling $\lambda$ at low scales.
Generically, it is also possible that there is a DMFV preserving
contribution $\propto \lambda^\dagger\lambda$ at tree level. If present, this would be the largest
contribution to the DM splittings. Of course, the pattern of
splittings generated by the running and by such threshold effects is
identical, since both cases are consistent with DMFV.
The splittings are given by
\begin{align}\label{eq:DMFVmass}
m_{ij}
&=
m_{\chi} (\mathbbm{1}
+\eta\, \lambda^\dagger\lambda+ \cdots
)_{ij}
=
m_{\chi} (1 + \eta (D_{\lambda, ii} )^2 + \cdots
)\delta_{ij}\,,
\end{align}
where summation is not implied in the last term. Here $\eta$ is a real coefficient whose value depends on the details of the model.
If the contribution to the mass matrix arises at tree level, then $\eta$ is expected to be an $\mathcal{O}(1)$ number. On the other hand, the contribution from running is schematically given by
\begin{align}
\eta
&\sim
\frac{1}{16\pi^2} \log \left(\frac{m_\chi^2}{\Lambda^2} \right)\,,
\end{align}
where $\Lambda$ is the dark flavor scale noted above.
The DMFV expansion above is only valid if higher order corrections are
parametrically suppressed. In order to ensure convergence, in what
follows we will assume $|\eta (D_{\lambda,ii})^2| <0.3$.
\section{Constraints from flavor and precision observables}\label{sec:DF2constraints}
In this section we study all relevant constraints from flavor
observables on the mDMFV model. We start the analysis of the
well-measured and strongly constraining observables from meson
anti-meson mixing, followed by relevant rare decays. Finally we take a
brief look at electroweak precision tests and electric dipole moments.
While we will find that $\Delta F =2$ processes significantly shape
the structure of a phenomenologically viable coupling matrix
$\lambda$, effects of other flavor observables are negligible.
We restrict ourselves to providing the formulae directly relevant for our study, a more detailed description of relevant techniques and necessary formulae for the study of $\Delta F =2$ processes reaching from effective Hamiltonian to flavor observables can be found for instance in \cite{Blanke:2011ry}. A recent comprehensive review can be found in \cite{Buras:2013ooa}.
\subsection{Constraints from meson anti-meson mixing}
\begin{figure}[h!]
\centering
\includegraphics[width=0.35\textwidth]{figures/figkkbar.pdf}
\caption{New contribution to $K^0-\bar K^0$ mixing in the mDMFV model.}
\label{fig:kkbar}
\end{figure}
In the mDMFV model new contributions to $\Delta F =2 $ processes arise first at the one loop level. The relevant box diagram is shown in figure \ref{fig:kkbar} for the case of $K^0-\bar K^0$ mixing. Evaluating this diagram we obtain the following contribution to the effective Hamiltonian:
\begin{equation}\label{eq:Heff1}
{\cal H}_\text{ eff}^{\Delta S=2,\text{new}} = \frac{1}{128\pi^2 m_\phi^2} \sum_{i,j}
\lambda_{si}\lambda_{di}^* \lambda_{sj}\lambda_{dj}^* F(x_i,x_j)
\times Q^{VRR} +\text{h.c.},
\end{equation}
with $x_i=m_{\chi_i}^2/m_{\phi}^2$, and the loop function $F(x_i,x_j)$ can be found in appendix \ref{app:functions}. As the new particles $\phi$ and $\chi_i$ couple only to right-handed down-type quarks, the only effective operator which receives new contributions is
\begin{equation}
Q^{VRR} = (\bar s_\alpha \gamma^\mu P_R d_\alpha) (\bar s_\beta \gamma_\mu P_R d_\beta)\,,
\end{equation}
i.\,e. the chirality-flipped counterpart of the SM operator.
The mass splittings among the $\chi_i$ fields constitutes a
higher order correction in the DMFV expansion, which we assume to be
small. Thus we can take the limit of equal $\chi$ masses in
\eqref{eq:Heff1}. The effective Hamiltonian then simplifies to
\begin{equation}\label{eq:Heff2}
{\cal H}_\text{ eff}^{\Delta S=2,\text{new}} = \frac{1}{128\pi^2 m_\phi^2} F(x) \, \xi_K^2 \times Q^{VRR} +\text{h.c.}\,,
\end{equation}
where $x=m_\chi^2/m_\phi^2$. The loop function $F(x)$ can be found in appendix \ref{app:functions}. We also defined
\begin{equation}
\xi_K = (\lambda\lambda^\dagger)_{sd} = \sum_{i=1}^3 \lambda_{si}\lambda_{di}^*\,.
\end{equation}
The mDMFV contribution of the DM sector to the off-diagonal element of the $K^0-\bar K^0$ mass matrix can then be obtained from
\begin{equation}
M_{12}^{K,\text{new}} = \frac{1}{2m_K}\langle \bar K^0| {\cal H}_\text{ eff}^{\Delta S=2,\text{new}} | K^0\rangle ^*\,.
\end{equation}
Using
\begin{equation}
\langle Q^{VRR}(\mu=2\,\text{GeV})\rangle =\frac{2}{3} m_K^2 F_K^2 \hat B_K
\end{equation}
we obtain
\begin{equation}
M_{12}^{K,\text{new}} = \frac{1}{384\pi^2 m_\phi^2} m_K F_K^2 \hat B_K \eta_2 F(x) (\xi_K^*)^2 \,.
\end{equation}
The parameter $\eta_2$ summarizes the corrections from the
renormalization group running from the weak scale $\mu\sim m_t$ down
to the scale $\mu=2\,$ GeV, where the lattice calculations are
performed, as well as the corrections due to the matching of the full
theory to the effective theory calculated within the SM.
By parametrizing the NLO corrections by $\eta_2$, we make two
approximations. We neglect the running from the NP scale
$\mu\sim m_\phi$ to the scale $\mu\sim m_t$, as well as the difference
in the matching conditions between the SM and the NP scenario studied
here. In order to estimate the error associated to our approach, it is
useful to compare our case with the discussion of the 331 models in
\cite{Buras:2012dp}. In the latter framework the inclusion of the
next-to-leading order (NLO) corrections amounts to a few percent
correction to the size of the NP contribution. We expect similar
conclusions to hold also in our case, in particular since in the MSSM
the NLO corrections to the Wilson coefficient $C^{VRR}$ have been
found to be small \cite{Virto:2009wm}.
In an analogous manner we find
\begin{equation}
M_{12}^{q,\text{new}} = \frac{1}{384\pi^2 m_\phi^2} m_{B_q} F_{B_q}^2 \hat B_{B_q} \eta_B F(x) (\xi_{B_q}^*)^2
\qquad (q=d,s)\,,
\end{equation}
where we define
\begin{equation}\label{eq:xiBq}
\xi_{B_q} = (\lambda\lambda^\dagger)_{bq} = \sum_{i=1}^3 \lambda_{bi}\lambda_{qi}^* \qquad (q=d,s)\,.
\end{equation}
In passing we note that the mDMFV model, coupling only to down-type quarks, does not contribute to $D$ meson observables at the one loop level.
\subsection[Radiative and rare $K$ and $B$ decays]{\boldmath Radiative and rare $K$ and $B$ decays}
We now turn our attention to radiative and rare decays, starting with the electromagnetic dipole operator generating the $b\to s\gamma$ transition.
The effective Hamiltonian describing the $B\to X_s\gamma$ decay can be written as
\begin{equation}
{\cal H}_\text{ eff} = \frac{4 G_F}{\sqrt{2}} V_{ts}^* V_{tb} \left( C_7 Q_7 + C'_7 Q'_7+\cdots \right)\,,
\end{equation}
where we omitted the tree level and chromomagnetic dipole operators that contribute to $b\to s\gamma$ via renormalization group mixing. We use the normalization
\begin{eqnarray}
Q_7 &=& \frac{e}{16\pi^2} m_b \bar s_L \sigma^{\mu\nu} b_R F_{\mu\nu}\,,\\
Q'_7 &=& \frac{e}{16\pi^2} m_b \bar s_R \sigma^{\mu\nu} b_L F_{\mu\nu}\,.
\end{eqnarray}
In the SM the Wilson coefficient $C'_7$ is strongly suppressed due to the chiral structure of weak interactions:
\begin{equation}
C'_{7,\text{SM}} = \frac{m_s}{m_b} C_{7,\text{SM}}\,.
\end{equation}
Conversely the DM in our scenario couples only to right-handed SM fermions -- therefore the only relevant new contribution arises in the chirality-flipped Wilson coefficient~$C'_7$. The relevant diagrams are analogous to the ones depicting the gluino contribution in supersymmetric models, replacing the gluino by the DM particles $\chi_i$ and the squarks by the scalar mediator $\phi$, and keeping only the coupling to right-handed SM quarks. Correcting for the different coupling and taking into account that $\chi_i$ are QCD singlets while the gluino is a color octet, we can straightforwardly obtain the result for $\delta C'_7$ from \cite{Bertolini:1990if,Cho:1996we}. Adjusting eq. (A.5) of \cite{Cho:1996we} to our model, we find
\begin{equation}
\delta C'_7 =-\frac{1}{6 g_2^2 V^*_{ts} V_{tb}}\frac{m_W^2}{m_\phi^2} \, \xi_{B_s}^* \, g(x_i)\,,
\end{equation}
where the short-hand notation for the relevant combination of elements of $\lambda$ has been defined in \eqref{eq:xiBq}, and the loop function $g(x)$ is given in appendix \ref{app:functions}.
With $g(x) \sim 0.08-0.17$ the size of the new contribution $\delta C'_7$ can be estimated as
\begin{equation}
|\delta C'_7| \lsim 4\cdot 10^{-2} |\xi_{B_s}^*| \left[\frac{500\, {\rm GeV}}{m_\phi}\right]^2 \,.
\end{equation}
Comparing this result to the constraints on the size of NP
contributions, see e.\,g.\ figure~2 in \cite{Altmannshofer:2013foa}, we see that the effect on the electromagnetic dipole operators generated in the present model is completely negligible. This is very welcome in view of the good agreement of $Br(B\to X_s\gamma)$ with the data.
Contributions to the four-fermion operators mediating transitions like $b\to s\mu^+\mu^-$ or $s\to d\nu\bar\nu$ can generally be split into tree level and one loop box and penguin diagrams. In the mDMFV model new tree level diagrams are forbidden by the residual $\mathbb{Z}_3$ symmetry (see appendix \ref{app:Z3}), while box diagrams are not generated since the new particles $\phi$ and $\chi_i$ do not couple to leptons. We are hence left with potential contributions to the $Z$ and photon penguins. An explicit calculation shows that the $Z$ penguin contribution vanishes. This can be explained by the chiral structure of our model with the new particles coupling only to right-handed quarks, and is also confirmed by adapting the SUSY results of \cite{Cho:1996we} to our scenario.
The photon penguin contribution is non-zero, however numerically small, as known from supersymmetric models \cite{Cho:1996we,Altmannshofer:2013foa}.
In summary we are left with completely SM-like rare decays like $B_{s,d}\to\mu^+\mu^-$, $B\to K^*\mu^+\mu^-$ and $B\to X_s\gamma$. Consequently the mDMFV model does not ameliorate the tension in the $B\to K^*\mu^+\mu^-$ data.
A bit more care is however required in the case of semileptonic decays
with neutrinos in the final state, such as $K\to\pi\nu\bar\nu$ or
$B\to K^{(*)}\nu\bar\nu$. Since the neutrinos escape detection, the
experimental signatures are $K\to\pi+\,{}/ \hspace{-1.5ex} E_{T}$ and $B\to
K^{(*)}+\,{}/ \hspace{-1.5ex} E_{T}$ respectively. Consequently also the
decays\footnote{We denote by $\chi_\text{DM}$ the lightest flavor
which is stable and provides the DM.} $K\to\pi\chi_\text{DM}\bar\chi_\text{DM}$ and $B\to K^{(*)}\chi_\text{DM}\bar\chi_\text{DM}$, mediated by a tree level $\phi$ exchange, will contribute to the measured branching ratio if the decay is kinematically allowed. See \cite{Kamenik:2011vy} for a detailed discussion. In order to avoid these potentially stringent constraints, in the remainder of our analysis we will assume $m_\text{DM} > 10\, {\rm GeV}$ and therefore well outside the kinematically allowed region for these decays.
\subsection{Electroweak precision tests and electric dipole moments}
Besides the flavor violating $K$ and $B$ decays discussed above,
the flavor conserving electroweak precision constraints and the bounds
on electric dipole moments also put strong constraints on many NP models.
In this section we consider these observables within the mDMFV model.
We start by considering electroweak precision observables. Due to the
residual $\mathbb{Z}_3$ symmetry corrections from the mDMFV model can
arise only at the loop level and are therefore suppressed by a loop
factor $1/(16\pi^2)$. Furthermore the mDMFV model introduces no new
$SU(2)_L$ doublets, and only $\phi$ carries hypercharge. Consequently
the contributions to electroweak precision observables receive an
additional suppression by $\sim g_Y^2/(9 m_\phi^2)$. Together with the
loop factor and the scale $m_\phi$ above the electroweak scale we
conclude that all new contributions to electroweak precision
observables are safely small.
Similarly we also find no significant new contribution to electric
dipole moments. The reasons are as follows. Due to the chiral
structure of the mDMFV model with new particles coupling only to
right-handed down-type quarks no EDM is generated at the one loop
level. At the two loop level a Barr-Zee type diagram
\cite{Barr:1990vd} with $\phi$ running in the loop exists -- however
its CP-violating phase is zero because the coupling
$\lambda_{H\phi}$ is real.
\section{Flavor pre-analysis of possible structures for the DM-quark coupling}\label{sec:DF2numerics}
We are now prepared to study the allowed regions of parameter space from flavor observables as well as correlations between different parameters of the coupling matrix $\lambda$.
\subsection{Strategy of the numerical analysis}
In order to determine the constraints from $\Delta F = 2$ observables
on the mDMFV model, we use the latest New Physics Fit results of the
model-independent NP fit presented by the UTfit collaboration
\cite{Bona:2005eu}. To this end we define
\begin{equation}
M_{12}^{B_q} = C_{B_q} e^{2i \varphi_{B_q}} M_{12}^{B_q,\text{SM}}\qquad (q=d,s)\,,
\end{equation}
where $M_{12}^{B_q}$ is the full mixing amplitude containing both SM and mDMFV contributions. Furthermore
\begin{equation}
\text{Re} M_{12}^{K} = C_{\Delta M_K} \text{Re} M_{12}^{K,\text{SM}}\,, \qquad \text{Im} M_{12}^{K} = C_{\varepsilon_K} \text{Im} M_{12}^{K,\text{SM}}\,.
\end{equation}
These six parameters are constrained by a global fit of the NP amplitude to the available tree level and $\Delta F = 2$ data \cite{Bona:2005eu,Bona:2007vi}. In order to be conservative we impose the resulting constraints at the $2\sigma$ level, see table \ref{tab:inputs} for a summary. In the case of $\Delta M_K$ we allow for a $\pm40\%$ uncertainty in order to capture the poorly known long distance effects. For consistency we set the CKM parameters to their central values obtained in the UTfit fit. All other input parameters are set to their central values listed in table 3 of \cite{Buras:2013dea}.
\begin{table}[h!]
\centering{
\begin{tabular}{|l|l|}
\hline
$|V_{us}|=0.22527$ & $C_{\Delta M_K}=1.10\pm0.44$\\
$|V_{ub}|=3.76\cdot 10^{-3}$ & $C_{\varepsilon_K}=1.05 \pm 0.32$\\
$|V_{cb}|=4.061\cdot 10^{-2}$ & $C_{B_d}=1.07 \pm 0.34$\\
$\delta= 67.8^\circ$ & $\varphi_{B_d}=-(2.0 \pm 6.4)^\circ$\\
& $C_{B_s}=1.066 \pm 0.166$\\
& $\varphi_{B_s}=(0.6 \pm 4.0)^\circ$ \\\hline
\end{tabular}}
\caption{\label{tab:inputs}Summary of CKM parameters and $\Delta F=2$ constraints used in our numerical analysis, see \cite{Bona:2005eu} for details.}
\end{table}
Altogether, we have the following new parameters relevant for flavor
violating decays:
\begin{equation}
m_\phi\,,\quad m_\chi\,,\quad \lambda_0\,,\quad \lambda_1\,,\quad
\lambda_2\,,\quad \theta_{12}^\lambda\,,\quad \theta_{13}^\lambda
\,,\quad \theta_{23}^\lambda\,,\quad \delta_{12}^\lambda\,,\quad
\delta_{13}^\lambda \,,\quad \delta_{23}^\lambda \,.
\end{equation}
Our goal is to obtain a clear picture of patterns in the coupling matrix $\lambda$ that are implied by the available $\Delta F = 2$ data. To this end we fix the flavor conserving parameters $m_\phi$, $m_\chi$ and $\lambda_0$ to the values
\begin{equation}
m_\phi=850\, {\rm GeV}\,,\qquad m_\chi= 200\, {\rm GeV}\,,\qquad
\lambda_0=1\,.
\end{equation}
The impact of varying these parameters can be estimated from the functional dependence of $M_{12}^{i,\text{new}}$ ($i=K,B_d,B_s$), which is roughly given by
\begin{equation}
M_{12}^{i,\text{new}} \propto \frac{\lambda_0^2}{m_\phi^2} F(x)\,,\qquad x= m_\chi^2/m_\phi^2\,.
\end{equation}
Note that $F(x)$ is a monotonically decreasing function varying from 1 to 1/3 over the range $0<x<1$.
\begin{figure}
\centering
\includegraphics[width=.6\textwidth]{figures/xi.png}
\caption{\label{fig:xi}Allowed ranges for the flavor violating parameters $\xi_M=\xi_K$ (yellow), $\xi_M=\xi_{B_d}$ (blue), $\xi_M=\xi_{B_s}$ (red).}
\end{figure}
The $\Delta F =2$ constraints then translate directly into constraints on the values of $\xi_K$, $\xi_{B_d}$ and $\xi_{B_s}$, as shown in figure \ref{fig:xi}.
We observe that the strongest constraints come from $K^0 -\bar K^0$ mixing, and in particular the CP-violating parameter $\varepsilon_K$, which forces the phase of $\xi_K$ to be very close to $0,\pi/2,\pi,3\pi/2$ unless $|\xi_K|\lsim 10^{-3}$. The $B$ physics constraints are less stringent and in particular do not yield a specific pattern for the phases of $\xi_{B_q}$. The weakest constraints are found in the $B_s$ system.
This pattern of allowed deviations from the SM is not specific to the mDMFV model, but can be found in all models with a generic NP flavor structure that do not induce the chirally enhanced left-right operators, like the Littlest Higgs model with T-parity analyzed in detail in \cite{Hubisz:2005bd,Blanke:2006sb,Blanke:2009am}. It is a direct consequence of the CKM hierarchies that determine the size of effects within the SM, as well as the theoretical uncertainties involved.
Note that e.\,g.\ in Randall-Sundrum (RS) models with bulk fermions \cite{Csaki:2008zd,Blanke:2008zb,Bauer:2009cf} and left-right models \cite{Zhang:2007da,Blanke:2011ry}, the strong enhancement of the left-right operators in the kaon system makes the $K^0-\bar K^0$ constraints even more severe.
\subsection[``Flavor-safe'' scenarios for the structure of $\lambda$]{\boldmath ``Flavor-safe'' scenarios for the structure of $\lambda$}
We now analyze the structure of the coupling matrix $\lambda$ that is implied by the $\Delta F =2 $ constraints. To this end we show in figure \ref{fig:scenarios}
the allowed points in the $(\lambda_1,\lambda_2,s_{12}^\lambda)$, $(\lambda_1,\lambda_2,s_{13}^\lambda)$ and $(\lambda_1,\lambda_2,s_{23}^\lambda)$ spaces, respectively.
\begin{figure}[htp]
\centering{
\includegraphics[width=.6\textwidth]{figures/scenarios-s12.png}
\includegraphics[width=.6\textwidth]{figures/scenarios-s13.png}
\includegraphics[width=.6\textwidth]{figures/scenarios-s23.png}}
\caption{\label{fig:scenarios}Scenarios for the structure of $\lambda$: Universality (black), 12-degeneracy (blue), 13-degeneracy (red), 23-degeneracy (green),
small mixing (yellow).}
\end{figure}
\afterpage{\clearpage}
We observe that the allowed points fall into five distinct scenarios for the structure of $\lambda$, which we discuss in some detail in the following. In order to analytically understand the scenarios, we recall the parametrization of $\lambda$ in terms of three two-flavor rotation matrices $U_{ij}$ and a diagonal matrix $D_\lambda$:
\begin{equation}
\lambda = U_{23} U_{13} U_{12} D_\lambda\,.
\end{equation}
\begin{enumerate}
\item {\bf universality scenario} (black): $\lambda_1 \simeq \lambda_2 \simeq 0 $
In this case $\lambda \simeq U_\lambda \cdot \lambda_0$ so that $\lambda \lambda^\dagger \simeq \lambda_0^2\cdot\mathbbm{1}$. Since flavour violating effects are governed by the off-diagonal elements of $\lambda \lambda^\dagger$, the $\Delta F=2$ constraints are trivially fulfilled for arbitrary $U_\lambda$ and there are no FCNC effects beyond the SM.
\item {\bf 12-degeneracy} (blue): $\lambda_1 \simeq \lambda_2$
If the first two generations of DM fermions are quasi-degenerate, then -- as seen from the blue points in figure \ref{fig:scenarios} -- the mixing angle $s_{12}^\lambda$ can be generic while $s^\lambda_{13,23}$ have to be small. This can be understood by taking the limit $\lambda_1=\lambda_2$, in which the mixing matrix $U_{12}$ becomes non-physical, and we are left with
\begin{equation}
\lambda= U_{23} U_{13} D_\lambda\,.
\end{equation}
It is easy to see that in order to fully suppress flavor violating effects we need $U_{13,23}\simeq \mathbbm{1}$ and therefore $s^\lambda_{13,23}\simeq 0$.
\item {\bf 13-degeneracy} (red): $\lambda_2 \simeq -2\lambda_1$
In the case $\lambda_2 \simeq -2\lambda_1$, shown by the red points in figure \ref{fig:scenarios}, the first and third DM flavor are quasi-degenerate, and consequently $s_{13}^\lambda$ is unconstrained. In order to suppress the remaining flavor violating effects both $s^\lambda_{12}$ and $s^\lambda_{23}$ have to be small.
\item {\bf 23-degeneracy} (green): $\lambda_2 \simeq -1/2\lambda_1$
Finally if $\lambda_2 \simeq -1/2\lambda_1$, the second and third DM flavor are quasi degenerate. Consequently the mixing angle $s^\lambda_{23}$ is arbitrary, while $s^\lambda_{12}$ and $s^{\lambda}_{13}$ have to be small.
This scenario is shown by the green points in figure \ref{fig:scenarios}.
\item
{\bf small mixing scenario} (yellow): arbitrary $D_\lambda$
Finally if $D_\lambda$ does not exhibit any degeneracies, then FCNC effects have to be suppressed by the smallness of all three mixing angles $s_{12}^\lambda \simeq s_{13}^\lambda \simeq s_{23}^\lambda \simeq 0$. This scenario, shown by the yellow points in figure \ref{fig:scenarios}, corresponds to a diagonal but non-degenerate coupling matrix $\lambda$.
\end{enumerate}
In order to quantify the allowed size of deviations from the
degeneracy scenarios discussed above, we show in figure \ref{fig:deg}
the mixing angles $s^{\lambda}_{ij}$ as a function of the deviation
from the corresponding degeneracy line. We observe that the constraint
on $12$ and $13$ mixings are comparable with the former being
somewhat stronger, while the constraint on the $23$ mixing angle is
significantly weaker. This is a direct consequence of the allowed
sizes of NP effects in the various meson systems, see figure
\ref{fig:xi}.
\begin{figure}[h!]
\centering
\includegraphics[width=.6\textwidth]{figures/degeneracy.png}
\caption{\label{fig:deg}Allowed ranges for the mixing angles $s^\lambda_{ij}$ as a function of the deviation from the $ij$-degeneracy line $\Delta_{ij}=|D_{\lambda,ii}-D_{\lambda,jj}|$. $ij=12$ in yellow, $ij=13$ in blue, $ij=23$ in red.}
\end{figure}
\subsection{A note on flavor safety of the UV completion}
FCNC processes are known to be sensitive to NP at very high scales. It
is therefore questionable whether a study of the simplified mDMFV
model is sufficient to capture all relevant effects.
Following the DMFV principle we can write any contribution from the UV
completion in terms of higher-dimensional operators that are
suppressed by powers of the UV scale $\Lambda_\text{UV}$ and made formally
invariant under the flavor group \eqref{eq:flavor-group} by insertion
of the appropriate combination of spurion fields $Y_{u,d}$ and
$\lambda$. The leading contribution to the $\Delta F = 2$ effective
Hamiltonian is then
\begin{equation}
{\cal H}_\text{ eff}^{\Delta F=2,\text{UV}} \sim
\frac{c^\text{UV}_{\Delta F=2}}{\Lambda_\text{UV}^2} \lambda \lambda^\dagger (\bar s \gamma^\mu P_R d) (\bar s \gamma_\mu P_R d)\,,
\end{equation}
where $c^\text{UV}_{\Delta F=2}$ is an ${\cal O}(1)$ coefficient that is common to all three meson systems. Comparing this to the new contribution generated first at the one loop level in the simplified model (see \eqref{eq:Heff1}), that can schematically be written as
\begin{equation}
{\cal H}_\text{ eff}^{\Delta F=2, \text{simpl.}} \sim
\frac{c^\text{simpl.}_{\Delta F=2}}{16\pi^2 m_\phi^2} \lambda \lambda^\dagger (\bar s \gamma^\mu P_R d) (\bar s \gamma_\mu P_R d)\,,
\end{equation}
we observe that both contributions carry the same flavor structure. Furthermore the UV contribution is suppressed with respect to the simplified model one if
\begin{equation}
\Lambda_\text{UV}\gsim 4\pi m_\phi\,.
\end{equation}
Therefore, NP close to the mass of $\phi$ does not change the
flavor phenomenology as long as it respects the DMFV hypothesis.
Generic flavor violation needs to be suppressed by a much higher scale
$\sim\mathcal{O}({100-1000})$ TeV~\cite{Isidori:2010kg}.
\subsection[Recovering the MFV limit in the structures for $\lambda$]{\boldmath Recovering the MFV limit in the structures for $\lambda$}
Earlier studies of flavored DM have been restricted to the MFV
framework in order to be safe from undesired effects in flavor
observables. The flavor-safe scenarios identified above
are more general than the MFV ansatz. It is also worthwhile to study
how MFV can be recovered in the DMFV framework.
Let us first consider the case where $U(3)_\chi$ is identified with $U(3)_d$. The MFV hypothesis then requires that $\lambda$ takes the schematic form
\begin{equation}\label{eq:lambdaMFV}
\lambda
\propto \mathbbm{1} +\alpha Y_d^\dagger Y_d + \dots\,,
\end{equation}
where $\alpha$ is an arbitrary coefficient.
The matrix $\lambda$ is diagonal in the down quark mass basis. In particular $Y_d$ is
proportional to the down-type quark masses. Considering that $Y_d$ can
be approximated by $Y_d\sim \diag(0,0,y_b)$, MFV must be close to the
12-degeneracy. Additionally MFV requires
\begin{equation}
m_\chi \propto \mathbbm{1}+ \beta Y_d^\dagger Y_d + \dots\,,
\end{equation}
where $\beta$ is an arbitrary coefficient.
The same expansion for $m_\chi$ is obtained when inserting \eqref{eq:lambdaMFV} into the DMFV expansion \eqref{eq:DMFVmass}, so that MFV in this case is consistent with the DMFV hypothesis. As $\lambda$ and $m_\chi$ are diagonal in the same basis, all three flavor mixing angles are zero. Thus the MFV limit can be recovered as a very specific subset of parameter space, (determined by the specific choice of $Y_d$) close to the 12-degeneracy line with all mixing angles zero.
If instead $U(3)_\chi$ is identified with $U(3)_u$, then
\begin{equation}
\lambda
\propto Y_d^\dagger Y_u + \dots\,,
\end{equation}
while
\begin{equation}
m_\chi \propto \mathbbm{1}+ \alpha Y_u^\dagger Y_u + \dots\, ,
\end{equation}
where as before $\alpha$ is an arbitrary coefficient.
We can see immediately that the mass splittings are not
directly
correlated with the $\lambda$ matrix, as was the case for
DMFV \eqref{eq:DMFVmass} with the separate U(3)$_\chi$ symmetry.
Finally identifying $U(3)_\chi$ with $U(3)_q$, we have
\begin{equation}
\lambda
\propto Y_d + \dots
\end{equation}
and
\begin{equation}
m_\chi \propto \mathbbm{1}+ \alpha Y_u Y_u^\dagger + \beta Y_d Y_d^\dagger + \dots\,,
\end{equation}
where $\alpha$ and $\beta$ are arbitrary coefficients.
Again we observe that the pattern of splittings in $m_\chi$
are not directly correlated with the $\lambda$ matrix.
In summary we find that if $\chi$ is assumed to transform under
$U(3)_d$ then the MFV limit can be recovered as a small subset of the
scenarios for $\lambda$, with an approximate 12-degeneracy and all
mixing angles identically zero. On the other hand, the
MFV limits for
$\chi$ transforming under $U(3)_u$ or $U(3)_q$ have mass splitting
patterns which are not correlated with the coupling matrix $\lambda$,
and therefore to capture these cases one needs to consider additional
contributions to the mass splitting in \eqref{eq:DMFVmass}.
\section{From the flavor pre-analysis to dark matter scenarios}\label{sec:flavor-consequences}
Our flavor pre-analysis shows that a generic coupling matrix $\lambda$
leads to unacceptably large corrections to $\Delta F = 2$ observables.
We have identified a number of non-trivial scenarios for the structure
of $\lambda$ for which flavor violating effects are efficiently
suppressed:
\begin{enumerate}
\item
Universality scenario: all elements of the diagonal matrix $D_\lambda$ equal and arbitrary flavor mixing angles.
\item
$ij$-degeneracy scenarios ($ij=12,13,23$): $D_{\lambda,ii}=D_{\lambda,jj}$, arbitrary $s^\lambda_{ij}$ and the other mixing angles small.
\item
Small mixing scenario: small mixing angles and arbitrary $D_\lambda$.
\end{enumerate}
While these scenarios have been identified in a scan with fixed flavor
conserving parameters $m_\phi$, $m_\chi$ and $\lambda_0$, we stress
that these structures for $\lambda$ also remain valid for different
choices of parameters. Furthermore, even though our analysis has been
performed within the simplified framework of the mDMFV model, the
identified scenarios for $\lambda$ remain flavor-safe in
non-minimal versions of DMFV also. Thus they provide a useful framework for
future study of the phenomenology of DMFV models -- employing any of
these scenarios for the structure of $\lambda$ efficiently evades all
FCNC constraints, without the need for an involved study of the
latter.
\begin{table}[h!]
\begin{center}
\begin{tabular}{|l|l|l|}
\hline
scenario & specification & lightest DM particle \\ \hline\hline
universal scenario ($ m_{\chi_d} \simeq m_{\chi_s} \simeq m_{\chi_b}$) & - & all hierarchies possible \\ \hline
12 degeneracy ($m_{\chi_d} \simeq m_{\chi_s}$) & $\eta \lambda_1 > 0$ & $ {\chi_b}$ \\
& $\eta \lambda_1 < 0$ & ${\chi_d}$ or ${\chi_s} $ \\ \hline
13 degeneracy ($m_{\chi_d} \simeq m_{\chi_b}$) & $\eta \lambda_1 > 0$ & $ {\chi_s}$ \\
& $\eta \lambda_1 < 0$ & ${\chi_d}$ or ${\chi_b} $ \\ \hline
23 degeneracy ($m_{\chi_s} \simeq m_{\chi_b}$) & $\eta \lambda_1 > 0$ & ${\chi_s}$ or ${\chi_b}$ \\
& $\eta \lambda_1 < 0$ & ${\chi_d}$ \\
\hline
small mixing scenario & - & all hierarchies possible \\ \hline
\end{tabular}
\end{center}
\caption{Overview of flavor-safe scenarios and their implications
for the mass hierarchy in the DM sector. \label{tab:DMscenarios}}
\end{table}
In table \ref{tab:DMscenarios} we summarize the flavor-safe scenarios
for $\lambda$ and their implications for the mass pattern in the DM
sector. It is clear that flavor constraints do not impose a
specific mass hierarchy on the dark sector, i.\,e.\ from the point of
FCNC constraints any dark flavor can be the lightest. Note that an
exact degeneracy of two flavors is unnatural, since in the case of
universal $\lambda$ it is violated by the presence of $Y_{d,u}$ at
higher orders in the DMFV expansion. We therefore assume that the
observed DM is composed of a single $\chi$ flavor, while the
decay of the heavier states is fast enough to have happened in the
early universe. We refer the reader to appendix \ref{sec:decayheavy} for an
estimate of the life-time of the heavier states.
However not all DM flavors are equally motivated from the
point of DM and collider phenomenology. DM that
couples dominantly to first generation quarks, like $d$-flavored DM,
is strongly constrained by the direct detection experiments.
If the DM relic density is assumed to arise from thermal
freeze-out in the early universe, the relic abundance condition is in
severe tension with the experimental constraints. We will therefore
not consider the case of $d$-flavored DM further.
As far as direct detection constraints are concerned, $s$- and
$b$-flavored DM are on equal footing. Interestingly the same
holds, at least qualitatively, also for the flavor phenomenology -- as we have seen in figure
\ref{fig:deg} the amount of flavor violation allowed by $\Delta F=2$
constraints is almost symmetric under the exchange of the second and
third generation, $2\leftrightarrow 3$.
The case is however different for collider phenomenology. While pair
production of the mediator and its subsequent decay will dominantly
produce light jets and missing energy in the $s$-flavored case, in the
case of $b$-flavored DM the large coupling to the $b$ quark
will give rise to $b$-jet signatures in a significant fraction of the
events. Since events with $b$-jets are much more easily distinguished
from the QCD background, the collider phenomenology of $b$-flavored
DM is at the same time more constraining (in particular
concerning the bound on the mediator mass) and also more promising, as
quite distinctive
signatures arise.
A further motivation for $b$-flavored DM comes from indirect
detection. Recently it has been shown that a 35 GeV $\chi_b$ provides
a good fit to the excess $\gamma$-rays observed at
the galactic center \cite{Agrawal:2014una}.
Therefore, in the rest of our analysis we restrict ourselves to the
case of $b$-flavored DM, i.\,e.\ $m_{\chi_b}<m_{\chi_{d,s}}$.
We also assume that the DM relic abundance is set by the
thermal freeze-out condition, so that $D_{\lambda,33}$ has to be
large. Due to the strong constraints on the first generation coupling
from direct detection and collider data, we deduce that
$D_{\lambda,11}<D_{\lambda,33}$. Consequently in order to ensure the
correct mass hierarchy, we have $\eta < 0$.
We are then left with the following scenarios for DM freeze-out:
\begin{enumerate}
\item single flavor freeze-out: The $s$- and $d$-flavored states are split from the $b$-flavored DM by at least 10\%.
\item two flavor freeze-out:
\begin{enumerate}
\item 13-degeneracy -- $\chi_b$ and $\chi_d$ are quasi-degenerate, while $\chi_s$ is split
\item 23-degeneracy -- $\chi_b$ and $\chi_s$ are quasi-degenerate, while $\chi_d$ is split
\end{enumerate}
\item three flavor freeze-out:
All three states are quasi-degenerate. Such a scenario can either be achieved by a quasi universal coupling matrix $D_\lambda$, or if the DMFV expansion parameter $\eta$ is loop-suppressed, $|\eta|\sim 10^{-2}$.
\end{enumerate}
In our numerical analysis we will study all of these scenarios in turn.
\section{\boldmath Phenomenology of $b$-flavored dark matter}
\label{sec:DMpheno}
In this section, we study the constraints arising from requiring the
DM to be a thermal relic and from direct
detection experiments. We note that the relic abundance constraints
may be potentially relaxed in the presence of other particles in the
dark sector.
The presence of multiple flavors can affect the DM freeze-out
significantly. This occurs when the mass splitting between different
flavors of DM is much smaller than the freeze-out
temperature, ($T_f \sim m_\chi / 20$). For mass splittings much bigger
than this scale, the freeze-out follows the standard WIMP paradigm.
We show that in the range of parameter space we consider, the heavier
DM flavors decay before big bang nucleosynthesis (BBN) (see appendix
\ref{sec:decayheavy}).
The direct detection constraints then depend sensitively on the
couplings of the lightest flavor of DM. In particular, when
the lightest dark flavor couples appreciably to the first generation quarks,
it gives rise to a very large direct detection signal. If this
contribution is suppressed, the dominant contribution then arises at
1-loop level, which is seen to be within the reach of present and
future direct detection experiments.
\subsection{Relic abundance}
\begin{figure}[h!]
\centering
\includegraphics[width=0.3\textwidth]{figures/figddbb.pdf}
\caption{Feynman diagram for dark matter annihilation in the early universe.}
\label{fig:relic}
\end{figure}
We will consider two different qualitative regimes. When their masses
are nearly degenerate, then all DM flavors are present during
freeze-out and can be treated together. Otherwise only the lightest
flavor of DM remains in the thermal
bath.
We start with a single flavor freeze-out.
The dominant annihilation during freeze-out occurs in the lowest
partial wave. In
this limit
\begin{align}
\langle\sigma v\rangle_{bb}
&=
\sum_{i,j}
\frac{3\lambda_{ib} \lambda^*_{ib}
\lambda_{jb}\lambda^*_{jb} m_{\chi_b}^2}{32\pi (m_{\chi_b}^2+m_\phi^2)^2}
=
\frac{3(D_{\lambda}^{\dagger} U^\dagger U D_\lambda)^2_{33}
m_{\chi_b}^2}{32\pi (m_{\chi_b}^2+m_\phi^2)^2}
=
\frac{3D_{33}^4
m_{\chi_b}^2}{32\pi (m_{\chi_b}^2+m_\phi^2)^2},
\label{annihilationcrosssection}
\end{align}
where we have ignored the masses of the final state quarks.
The relic abundance is determined by solving the Boltzmann equation for
the DM number density $n$ at late times.
For a Dirac fermion, it is useful to convert the annihilation cross
section into an effective cross
section~\cite{Griest:1990kh,Servant:2002aq}.
\begin{align}
\langle\sigma v\rangle_\text{eff} = \frac12 \langle
\sigma v\rangle \ .
\end{align}
which is approximately required to be~\cite{Steigman:2012nb}
\begin{align}
\langle \sigma v \rangle_\text{eff}
&=
2.2\times10^{-26 }\mathrm{cm}^3/\mathrm{s}
\label{eq:effsigma}
\end{align}
in order to produce the correct relic abundance of DM.
Next we consider the case, where the mass
splitting between the DM flavors is much smaller than the
temperature at freeze-out. Consequently, we have to take into account
the co-annihilation between different flavors. We assume that flavor
changing (but DM number preserving) interactions $\chi_i q
\to \chi_j q$ are fast during the epoch of DM freeze-out. The
rate for these processes is enhanced over the DM
annihilations---which are approximately in thermal equilibrium---by a
large Boltzmann factor ($\mathcal{O}(10^9)$). Thus, this approximation
is valid as long as any individual cross sections are not suppressed
enough to overwhelm this factor.
Then the Boltzmann equation for freeze-out
has a very similar form to the single DM
case, and can be solved in exactly the same way. The
relic abundance is in fact relatively insensitive to the change in the
number of DM species, changing by only about 5\% when other
parameters are kept fixed. In the limit of small splitting,
the effective cross section
is well approximated by \cite{Griest:1990kh},
\begin{align}
\langle \sigma v \rangle_\text{eff}
&=
\frac{1}{18}
\sum_{i,j=d,s,b}
\langle \sigma v \rangle_{ij}
.
\end{align}
The co-annihilation cross section can be derived by modifying equation
(\ref{annihilationcrosssection}), e.g.
\begin{align}
\langle\sigma v\rangle_{bs}
&=
\sum_{i,j}
\frac{3 \lambda_{is} \lambda^*_{is}
\lambda_{jb}\lambda^*_{jb} m_{\chi b}^2}{32\pi (m_{\chi b}^2+m_\phi^2)^2}
=
\frac{3 D_{22}^2 D_{33}^2
m_{\chi b}^2}{32\pi (m_{\chi b}^2+m_\phi^2)^2}\,.
\label{annihilationcrosssectiontwofl}
\end{align}
Note that the splitting between DM masses is in this case
negligible ($m_{\chi d} \simeq m_{\chi s} \simeq m_{\chi b}$).
If only two states are nearly degenerate, then a two flavor freeze-out occurs. The corresponding formulae can be straightforwardly obtained from the above results.
\subsection{Direct detection}
For $b$-flavored DM, the
direct detection scattering arises either through mixing, or at one
loop.
We focus on the spin-independent contribution to the WIMP-nucleus
scattering. The reported experimental bounds are translated to the
WIMP-nucleon cross section, which can be written as
{\begin{align}
\sigma^{\rm SI}_n
&=
\frac{\mu^2_n}{\pi A^2}
\left| Z f_p + (A-Z)f_n\right|^2\,,
\end{align}}
where $\mu_n$ is the reduced mass of the WIMP-nucleon system,
$A$ and $Z$ are the mass and atomic numbers of the nucleus
respectively, and $f_n$ and $f_p$ parametrize DM coupling to
neutrons and protons. The relevant processes are shown in figure \ref{fig:dd}.
\begin{figure}[h!]
\includegraphics[width=0.3\textwidth]{figures/figtree.pdf}\hfill
\includegraphics[width=0.3\textwidth]{figures/figbox.pdf}\hfill
\includegraphics[width=0.3\textwidth]{figures/figphoton.pdf}
\caption{Diagrams contributing to WIMP-nucleon scattering in the mDMFV model.}
\label{fig:dd}
\end{figure}
For direct detection, we can safely work in the effective theory with
the $\phi$ integrated out. The Lorentz structure of the
four-fermion operator generated after
performing the Fierz transformation is given by
\begin{align}
\bar{\chi}_b \gamma^\mu (1-\gamma^5)\chi_b\,
\bar{d} \gamma^\mu (1+\gamma^5)d\,.
\label{eq:effoper}
\end{align}
There are three contributions to $f_{n,p}$:
\begin{align}
f_{n,p} &=
f^{\rm tree}_{n,p} +
f^{\rm box}_{n,p} +
f^{\rm photon}_{n,p} \,.
\end{align}
These contributions are individually given as follows:
\begin{enumerate}
\item $s$-channel $\phi$ at tree-level:\\
In presence of significant mixing in the $\lambda$ matrix,
this is the dominant contribution to direct detection:
\begin{align}
2f^{\rm tree}_p
=f^{\rm tree}_n = \frac{ |\lambda_{db}|^2}{4 m_\phi^2}\,.
\end{align}
The spin-independent part in equation (\ref{eq:effoper}) arises
from the matrix element of the quark vector current bilinear
in the nucleons. Thus,
only the valence $d$-quark contributes.
\item One-loop photon exchange:\\
The interaction of DM with nucleons via photon exchange
is conveniently parametrized as the electromagnetic form factors
of the DM coupling with those of the nucleus.
In particular, the scattering cross sections
arise from charge-charge, dipole-charge and dipole-dipole
interactions~\cite{Agrawal:2011ze}. In the region of interest, the
charge-charge
interactions dominate, leading to
\begin{align}
f^\text{photon}_p
&=
\sum_i
\frac{\left| \lambda_{ib} \right|^2 e^2}
{96 \pi^2 m_\phi^2}
\log \left[\frac{m_{q_i}^2 }{ m_\phi^2}\right]
\end{align}
in the leading-log approximation.
\item Box diagram with $\phi$ exchange in the t-channel:\\
{This new contribution depends upon} the coupling of DM with the first
generations quarks \cite{Kumar:2013hfa} and is given by
\begin{align}
4 f_p^{\rm box} &=
2 f_n^{\rm box} =
\sum_{i,j} \frac{\left| \lambda_{dj} \right|^2 \left| \lambda_{ib}\right|^2 } {16\pi^2 m_\phi^2}
F \left( \frac{m_{q_i}^2}{m_\phi^2},
\frac{m_{\chi_j}^2}{m_\phi^2}\right)\,,
\end{align}
with the loop function $F$ given in appendix \ref{app:functions}.
\end{enumerate}
The tree-level contribution, {being flavor violating}, constrains the mixing $s^\lambda_{13}$ to be
small. The LUX experiment~\cite{Akerib:2013tjd} is sensitive to even the loop level
scattering cross sections for WIMP DM. {These contributions are present even in the absence of flavor
violation.} The box and the photon
loop diagrams are seen to destructively interfere.
\section{Combined numerical analysis of flavor and dark matter constraints}\label{sec:numerics}
Having all relevant formulae for the DM phenomenology in
hand, we are now ready to perform a combined numerical analysis of
both DM and flavor constraints. We restrict ourselves to the
phenomenologically most interesting case of $b$-flavored DM,
and study in turn the scenarios identified in section
\ref{sec:flavor-consequences}.
The DM mass $m_{\chi_b}$ is allowed to vary in the
phenomenologically interesting region $10\, {\rm GeV}<m_{\chi_b}<250\, {\rm GeV}$. We
also assume $\eta<0$ in order to suppress the $\chi$ coupling to the
first generation, in order to cope with the strong direct detection
and collider constraints. Convergence of the DMFV expansion is ensured
by requiring $|\eta D_{\lambda,33}^2|<0.3$. Finally, since corrections
to $m_\chi$ are unavoidably generated at the one loop level, we take
$|\eta|>10^{-2}$. In summary, \begin{equation}
-\frac{0.3}{D_{\lambda,33}^2}<\eta<-0.01\,. \end{equation} We fix
$m_\phi=850\, {\rm GeV}$ in agreement with the collider constraints, see
section \ref{sec:collider}. The parameters of the coupling matrix
$\lambda$ are scattered, imposing the flavor and DM constraints from
section \ref{sec:DF2constraints} and \ref{sec:DMpheno}.
\subsection{Single flavor freeze-out}
If the masses of the heavier flavors $\chi_{d,s}$ are sufficiently
split from the DM mass $m_{\chi_b}$ (by $\gsim 10\%$), then at the freeze-out temperature only the lightest state $\chi_b$ is left while the heavier ones have decayed. Therefore the single flavor freeze-out condition for the relic abundance, eq.\ \eqref{eq:effsigma}, applies and fixes $D_{\lambda,33}$ as a function of $m_{\chi_b}$.
The couplings $D_{\lambda,11}$ and $D_{\lambda,22}$ on the other hand are free to vary. They both need to be split from $D_{\lambda,33}$ in order to achieve the mass splitting.
\begin{figure}[h!]
\includegraphics[width=.49\textwidth]{figures/mchi-s13.png}\hfill
\includegraphics[width=.49\textwidth]{figures/mchi-D11.png}
\caption{\label{fig:1f-s13-D11}Flavor mixing angle $s^\lambda_{13}$
and first generation coupling $D_{\lambda,11}$ as functions of the DM mass $m_{\chi_b}$ in the single flavor freeze-out scenario. The mass hierarchy $m_{\chi_b}<m_{\chi_{d,s}}$ and the relic abundance constraint are imposed. The red points satisfy the bound from LUX, while the blue points satisfy the $\Delta F=2$ constraints. For the yellow points both LUX and $\Delta F =2$ constraints are imposed.}
\end{figure}
Figure \ref{fig:1f-s13-D11} shows the result of a parameter scan over $m_{\chi_b}$ and $\lambda$, with the relic abundance constraint for the single flavor freeze-out scenario imposed. The red points satisfy the bound from LUX, while the blue points satisfy the $\Delta F=2$ constraints. The yellow points fulfill both LUX and $\Delta F =2$ constraints.
From the left panel, showing $s^\lambda_{13}$ as a function of $m_{\chi_b}$, we can see that both the LUX and the $\Delta F =2$ constraints require $s^\lambda_{13}$ to be small. In case of LUX this constraint arises from the necessary suppression of the DM-nucleon scattering at tree level. The $\Delta F =2$ constraints on the other hand require $s^\lambda_{13}$ to be small, as the mass splitting between $\chi_{d}$ and $\chi_b$ requires a deviation from the 13-degeneracy scenario $D_{\lambda,11}=D_{\lambda,33}$. This bound becomes stronger for small $m_{\chi_b}$ for the following reason: Small $m_{\chi_b}$ requires a large $D_{\lambda,33}$ from the relic abundance constraint, so that $|\eta|_\text{max}$ decreases with decreasing $m_{\chi_b}$. In turn a larger splitting between $D_{\lambda,11}$ and $D_{\lambda,33}$ is required in order to generate the $\gsim 10\%$ mass splitting, implying smaller flavor mixing $s^\lambda_{13}$.
We further observe that once both the LUX and the $\Delta F=2$
constraints are taken into account, the upper bound on
$s^\lambda_{13}$ gets stronger than the individual ones. This proves
a non-trivial interplay of the flavor and DM constraints and underlines the importance of a combined study.
In the right panel we show the allowed size of $D_{\lambda,11}$ as a function of $m_{\chi_b}$. The upper bound on $D_{\lambda,11}$ as a function of $m_{\chi_b}$ arises from the relic abundance constraint on $D_{\lambda,33}$ together with the hierarchy requirement $D_{\lambda,11}<D_{\lambda,33}$. For $m_{\chi_b}\gsim 100\, {\rm GeV}$ the whole range of $D_{\lambda,11}$ up to the relic abundance bound is allowed. However for smaller $m_{\chi_b}$ the LUX constraint disfavors large values for $D_{\lambda,11}$, so that in combination with the $\Delta F=2$ constraints a sharp cutoff arises. This cutoff decreases with decreasing mass $m_{\chi_b}$.
Below $m_{\chi_b} \sim 60\, {\rm GeV}$ also a lower bound on $D_{\lambda,11}$
arises for the parameter points that satisfy both the LUX and flavor
constraints -- interestingly this bound does not emerge if the
constraints are taken into account separately. This is another clear
sign of a non-trivial interplay of the flavor and DM constraints.
Analytically the bounds on $D_{\lambda,11}$ can be understood by having a closer look at the structure of the WIMP-nucleon cross-section in the mDMFV model. As already mentioned above, the $s$-channel $\phi$ exchange at tree level is proportional to $(D_{\lambda,33}s^\lambda_{13})^2$ and therefore places a strong constraint on $s^\lambda_{13}$. We are then left with the one-loop box and photon penguin contributions.
The box amplitude is positive and proportional to
\begin{equation}\label{eq:box-dependence}
D_{\lambda,33}^2\cdot\left[ (D_{\lambda,11}c^\lambda_{12})^2+(D_{\lambda,22}s^\lambda_{12})^2\right]\,.
\end{equation}
It can therefore be suppressed by choosing both terms in the bracket
small. This explains why the LUX constraint, which is strongest for
low DM masses, gives rise to an upper bound on $D_{\lambda,11}$.
The size of the photon penguin is determined by $D_{\lambda,33}^2$ and therefore fixed by the relic abundance (there is a small dependence on $s^\lambda_{23}$, due to the difference in quark masses). Interestingly due to the log factor the photon penguin amplitude carries an overall minus sign. This relative sign between the penguin and box amplitudes leads to a cancellation of the two contributions, provided the expression in the bracket of \eqref{eq:box-dependence} is of the right size. In the absence of flavor constraints $D_{\lambda,11}$, $D_{\lambda,22}$ and $s^\lambda_{12}$ are independent parameters. Consequently no conclusion can be drawn on the value of only one of them, while letting the other two vary. Taking into account the flavor constraints the picture changes however. In that case $s^\lambda_{12}$ is allowed to be sizable only if $D_{\lambda,22}\simeq D_{\lambda,11}$. Eq.\ \eqref{eq:box-dependence} then reduces to $D_{\lambda,33}^2\cdot D_{\lambda,11}^2$, and we can
directly read off the requirement that $D_{\lambda,11}$ has to lie in
a specific range such that the cancellation between the penguin and box contributions can work.
This mechanism provides a realization of xenophobic DM
\cite{Feng:2011vu,Feng:2013vod}. The box diagram contributes to DM
scattering off protons and neutrons, while the photon penguin is only
present for DM-proton scattering. Therefore the penguin-box
cancellation works only for a specific number of protons and neutrons
in the nucleus.
We note that a pattern of constraints analogous to the one observed in
figure \ref{fig:1f-s13-D11} arises for all the possible mass
hierarchies in the dark sector, i.\,e.\ it does not depend
qualitatively on whether a single flavor or multiple flavors are
present at freeze-out.
\begin{figure}[h!]
\centering
\begin{minipage}{7.5cm}
\includegraphics[width=\textwidth]{figures/lambda12-1flavor-mass.png}
\end{minipage}\hspace{1mm}
\begin{minipage}{2.2cm}
\includegraphics[width=\textwidth]{figures/mchi-color.png}
\vspace*{.01cm}
\end{minipage}
\caption{\label{fig:1f-lambda1-lambda2}Allowed region in the
$\lambda_1$-$\lambda_2$-plane for the single flavor freeze-out
scenario, after imposing the relic abundance, LUX and flavor
constraints. The DM mass $m_{\chi_b}$ is indicated by the color, and the $ij$-degeneracy lines are sketched.}
\end{figure}
Next let us recover which of the flavor scenarios identified in section \ref{sec:DF2numerics} are realized in the single flavor freeze-out case. To this end we parameterize the matrix $D_\lambda$ in terms of the parameters $\lambda_{0,1,2}$, used in the flavor analysis, and defined in \eqref{eq:Dlambda}.
The distribution of parameter points consistent with the LUX and flavor constraints in the $\lambda_1$-$\lambda_2$-plane is shown in figure \ref{fig:1f-lambda1-lambda2}. We observe that the requirement $m_{\chi_b}<m_{\chi_{d,s}}$ and therefore $D_{\lambda,11},D_{\lambda,22}<D_{\lambda,33}$ restricts the allowed parameter space to the region $\lambda_2< -1/2\lambda_1 \wedge \lambda_2< -2\lambda_1$. Consequently only the 12 degeneracy scenario and the small mixing scenario are realized in this case.
The DM mass $m_{\chi_b}$ is indicated by the color, with the dark
points corresponding to light DM. We see that with smaller DM mass $m_{\chi_b}$ points move further away from the universality $\lambda_1 = \lambda_2=0$. This feature is a direct consequence of the upper bound on $D_{\lambda,11}$ for $m_{\chi_b}\lsim 100\, {\rm GeV}$, requiring a sizable splitting in $D_\lambda$.
\subsection{13-degeneracy}
Next we consider the scenario of quasi-degenerate first and third dark generations, i.\,e.\ $m_{\chi_b}\lsim m_{\chi_d}< m_{\chi_s}$. In this case we have $D_{\lambda,33}\gsim D_{\lambda,11} > D_{\lambda,22}$. Comparing this structure to the scenarios identified in the flavor pre-analysis we see that we are confined to the 13-degeneracy scenario, in which $s^\lambda_{12}$ and $s^\lambda_{23}$ are small while $s^\lambda_{13}$ is in principle allowed to be large. However recalling the results from the previous section we anticipate that once the LUX bound is imposed also the latter mixing angle is constrained to be small, in order to suppress the tree level contribution to the WIMP-nucleon scattering. This is indeed confirmed by our numerical analysis.
\begin{figure}[h!]
\centering
\includegraphics[width=.49\textwidth]{figures/mchi-D11-13deg.png}
\caption{\label{fig:2f-D11-13deg}First generation coupling
$D_{\lambda,11}$ as function of the DM mass $m_{\chi_b}$ in the 13-degeneracy scenario.
The mass hierarchy $m_{\chi_b}\lsim m_{\chi_{d}}<m_{\chi_{s}}$ and the relic abundance constraint are imposed. The red points satisfy the bound from LUX, while the blue points satisfy the $\Delta F=2$ constraints. For the yellow points both LUX and $\Delta F =2$ constraints are imposed.}
\end{figure}
In figure \ref{fig:2f-D11-13deg} we show the allowed values of the
first generation coupling $D_{\lambda,11}$ as a function of the DM mass $m_{\chi_b}$. We observe that due to the quasi-degeneracy of the first and third generation for a given $m_{\chi_b}$ only a small range of parameter space is allowed by the relic abundance constraint. Consequently the upper bound on $D_{\lambda,11}$ that arises again from the need to cut off the box contribution to the WIMP-nucleon scattering translates into a lower bound on $m_{\chi_b}$. If we neglect the flavor constraints and
take into account only the LUX bound, we find a lower bound $m_{\chi_b}\gsim 70\, {\rm GeV}$. Taking into account both the LUX and flavor constraints simultaneously, the bound becomes considerably stronger, $m_{\chi_b}\gsim 100\, {\rm GeV}$.
Again we stress the non-trivial interplay of flavor and DM constraints.
\begin{figure}[h!]
\centering
\begin{minipage}{7.5cm}
\includegraphics[width=\textwidth]{figures/Dii-s12-13deg.png}
\end{minipage}
\begin{minipage}{1.4cm}
\includegraphics[width=\textwidth]{figures/s12-color.png}
\vspace*{.01cm}
\end{minipage}
\caption{\label{fig:2f-13deg-D11-D22-LUX}Allowed region for $D_{\lambda,11}$ and $D_{\lambda,22}$ in the 13-degeneracy scenario, after imposing both the relic abundance and LUX constraints, while the flavor constraints are not taken into account. The flavor mixing angle $s^\lambda_{12}$ is indicated by the color, and the 12-degeneracy line $D_{\lambda,11}=D_{\lambda,22}$ is sketched.}
\end{figure}
The interplay of flavor and DM constraints can be understood from
figure \ref{fig:2f-13deg-D11-D22-LUX}. We show the allowed parameter
range in the $D_{\lambda,11}$-$D_{\lambda,22}$-plane, applying only
the LUX bound, but not the flavor constraints. We observe that the
largest values for $D_{\lambda,11}$, corresponding to the smallest DM masses, are reached away from the 12-degeneracy line and require near-maximal mixing $s^\lambda_{12}$.
This part of parameter space, while perfectly consistent with the DM constraints, is however ruled out by the stringent $\Delta F=2$ constraints that allow a large $s^\lambda_{12}$ only very close to the 12-degeneracy line.
\begin{figure}[h!]
\centering
\begin{minipage}{7.5cm}
\includegraphics[width=\textwidth]{figures/lambda12-2flavor-13deg-mass.png}
\end{minipage}\hspace{1mm}
\begin{minipage}{2.2cm}
\includegraphics[width=\textwidth]{figures/mchi-color.png}
\vspace*{.01cm}
\end{minipage}
\caption{\label{fig:2f-13deg-lambda1-lambda2}
Allowed region in the $\lambda_1$-$\lambda_2$-plane for the
13-degeneracy scenario, after imposing the relic abundance, LUX and
flavor constraints. The DM mass $m_{\chi_b}$ is indicated by the color, and the $ij$-degeneracy lines are sketched.}
\end{figure}
In figure \ref{fig:2f-13deg-lambda1-lambda2} we show the allowed region of parameter space in the $\lambda_1$-$\lambda_2$-plane.
As expected all parameter points lie very close to the exact
13-degeneracy line $\lambda_2=-2\lambda_1$. With decreasing DM mass $m_{\chi_b}$ they move further away from the universality point $\lambda_1 = \lambda_2=0$. This is a consequence of the required splitting of the second generation $m_{\chi_s}$.
\subsection{23-degeneracy}
\begin{figure}[h!]
\centering
\begin{minipage}{7.5cm}
\includegraphics[width=\textwidth]{figures/lambda12-2flavor-mass.png}
\end{minipage}\hspace{1mm}
\begin{minipage}{2.2cm}
\includegraphics[width=\textwidth]{figures/mchi-color.png}
\vspace*{.01cm}
\end{minipage}
\caption{\label{fig:2f-23deg-lambda1-lambda2}
Allowed region in the $\lambda_1$-$\lambda_2$-plane for the
23-degeneracy scenario, after imposing the relic abundance, LUX and
flavor constraints. The DM mass $m_{\chi_b}$ is indicated by the color, and the $ij$-degeneracy lines are sketched.}
\end{figure}
The next scenario to study is the approximate 23-degeneracy, with
$m_{\chi_b}\lsim m_{\chi_s}< m_{\chi_d}$. In this case
$s^\lambda_{23}$ is arbitrary, with $s^\lambda_{12}$ is required to be
small by the $\Delta F =2 $ constraints, and both $\Delta F =2 $ and
LUX constraints demand $s^\lambda_{13}$ to be small. In figure
\ref{fig:2f-23deg-lambda1-lambda2} we show the allowed parameter space
for $\lambda_1$ and $\lambda_2$, which as expected is confined to be
very close to the 23-degeneracy line $\lambda_2=-1/2\lambda_1$. Again
small DM mass $m_{\chi_b}$
implies a sizable coupling non-universality $\lambda_{1,2}\ne 0$. In contrast to the 13-degeneracy in this case no lower bound on $m_{\chi_b}$ emerges.
\subsection{Quasi-degenerate dark sector}
We finally study the case of a quasi-degenerate dark sector with mass
splittings $<1\%$. Such a scenario can be achieved either by having a
quasi-universal coupling matrix $D_\lambda$ or by suppressing the DMFV
correction to the mass matrix by the smallness of the parameter
$\eta$. Note however that the correction is unavoidably generated at
one loop level, so $|\eta|\lsim 10^{-2}$ would involve fine-tuning. As
in the previous cases we scan over the allowed parameter space for
$\lambda$, assuming $b$-flavored DM and requiring the relic
abundance to be set by thermal freeze-out. Since the three $\chi$
flavors are quasi-degenerate, all of them are present at
freeze-out.
\begin{figure}[ht]
\centering
\begin{minipage}{7.5cm}
\includegraphics[width=\textwidth]{figures/lambda12-3flavor-mass.png}
\end{minipage}\hspace{1mm}
\begin{minipage}{2.2cm}
\includegraphics[width=\textwidth]{figures/mchi-color.png}
\vspace*{.01cm}
\end{minipage}
\caption{\label{fig:3f-lambda1-lambda2}Allowed region in the
$\lambda_1$-$\lambda_2$-plane for the quasi-degeneracy scenario, after
imposing the relic abundance, LUX and flavor constraints. The DM mass $m_{\chi_b}$ is indicated by the color, and the $ij$-degeneracy lines are sketched.}
\end{figure}
In figure \ref{fig:3f-lambda1-lambda2} we show the allowed range of
parameters in the $\lambda_1$-$\lambda_2$-plane. The DM mass
$m_{\chi_b}$ is varied and depicted by the color. We observe that for
large DM masses $m_{\chi_b}\gsim 100\, {\rm GeV}$ the allowed range of
parameters is indeed close to the universality scenario
$\lambda_{1,2}=0$. Some deviations are however allowed, as the
non-universal corrections can be suppressed by a small expansion
parameter $\eta$. For small DM masses however a significant deviation $\lambda_1\lsim -1$ from the universality scenario is necessary and therefore $\eta$ has to be loop suppressed. A sizable negative $\lambda_1$, together with a near 23-degeneracy, suppresses the coupling $D_{\lambda,11}$ with respect to the other $D_\lambda$ components. Consequently the relic abundance constraint is brought in accordance with the bound from LUX.
\section{Collider phenomenology}\label{sec:collider}
In this section we present the collider constraints {on the mDMFV model}. We expect
significant constraints from the LHC since our model contains new
colored states, as well as large coupling to quarks.
Our framework predicts a number of new particles within the reach of the LHC:
\begin{itemize}
\item
The scalar mediator $\phi$ which carries QCD charge and therefore has
a large production cross-section in proton-proton collisions.
\item
The DM particle $\chi_b$ and the heavier flavors $\chi_{d,s}$, which
are Dirac fermions and singlets under the SM gauge group.
\end{itemize}
This spectrum has some similarities to simplified models of squarks
and neutralinos in the MSSM. The scalar mediator $\phi$
is similar to a right-handed down-type squark, except that it does
not carry flavor. It couples to all three quark flavors $d,s,b$ with
strengths given by the elements of the coupling matrix $\lambda$.
The flavor is carried by $\chi$, leading to three
somewhat degenerate states, in contrast to a neutralino LSP (lightest
supersymmetric particle). Also note
that in the simplest versions of supersymmetry neutralinos are
Majorana fermions, whereas $\chi$ is a Dirac fermion.
A full treatment of collider constraints is beyond the scope of this work.
We use existing analyses and translate their bounds into limits on our
model. There are some qualitatively new features beyond the
simplified MSSM model. First is the presence of large DM
trilinear couplings with quarks. In the {MSSM, supersymmetry} constrains this
coupling to be of the electroweak strength, making the production of
squarks insensitive to the coupling with DM. In contrast, the
large couplings required for relic abundance in the {mDMFV model}
contribute non-negligibly to the production. This also makes the
production sensitive to the flavor structure of the DM-quark
couplings.
While large flavor angles in the mDMFV model are an interesting
possibility, they lead to a large number of free parameters, and
we postpone a detailed analysis of this case for future
work. Here, we focus on the cases where the mixing angles are
small. This has implications for both production and decay of $\phi$
and $\chi$.
The main constraints on the mDMFV model come from searches for
monojets and dijets in conjunction with missing transverse energy. We
now estimate what constraints these searches put on our model. We also
point out some unexplored novel signatures that arise in this framework.
\subsection{Multijet with missing energy searches}
Let us begin with the constraints from multijet + MET searches.
The relevant analyses that we
consider for this case are the 19.5 fb$^{-1}$ CMS analysis for direct
sbottom production~\cite{CMS-PAS-SUS-13-018} and multijet + MET
analysis for squark-pair
production~\cite{Chatrchyan:2014lfa}. The colored particle $\phi$ is analogous to a
squark in the SUSY simplified models considered in these analyses,
but with some important differences which we highlight below.
A dominant mode of $\phi$
pair production is through QCD interactions, in analogy with pair
production of a single squark {flavor}, in the limit of decoupled gluino mass.
There are two additional contributions that we need to consider.
There is a sizeable contribution from the $t$-channel process mediated
by $\chi_d$ due to the valence down quark pdf. Depending upon the
value of the coupling $D_{\lambda, 11}$, this can be an
$\mathcal{O}$(1) fraction of the total production rate. The
corresponding contribution from the $D_{\lambda,22}$ is suppressed by
the sea-quark $s$-quark pdf and does not contribute significantly.
There are also contributions from an off-shell quark $(gg\to d d^* \to
d \chi_d \phi)$ which yield the same dijet + MET final state. The total
production cross section for this process can be much higher than the
QCD/$t$-channel production. However, this contribution always involves
one jet arising from a gluon splitting, which typically carries a very
small $p_T$. Thus, most events in this channel do not pass the
selection cuts for dijet analysis.
As is familiar from squark pair-production, NLO QCD effects are
significant. Since there is no NLO calculation available for our
model, we present limits for two extreme cases: one is assuming the
leading order production cross section, which provides a weak bound.
The other limit is to take the squark production
K-factor~\cite{Kramer:2012bx} , and apply
it to the LO cross section obtained above. Note that for the case
$D_{\lambda, 11} =0$, the production cross section is given directly by
the squark pair production cross section.
Another important difference from the sbottom case is the decay of the
mediator $\phi$. In presence of small mixing, squarks decay into a
flavor-specific jet and the LSP, whereas $\phi$ can decay to each of
the quark flavors $d$, $s$ and $b$. This has the effect of weakening
the limits relative to the sbottom search for sizeable branching
fraction to the first and second generation. {A similar effect has
been observed for the case of stop scharm mixing in
\cite{Blanke:2013uia,Agrawal:2013kha}, where the bounds on the stop
mass get weakened due to the light jets in the final state.}
In the CMS sbottom search~\cite{CMS-PAS-SUS-13-018}, separate search regions are defined
for one and two $b$-tags. For the SUSY simplified model, the two $b$-tag search
region is the most sensitive for almost all of the parameter space. To
translate limits from this analysis to the mDMFV case, we of course
need to account for the branching fractions of $\phi$. The scaling is
different for the one vs. two $b$-tag region, and it is possible that for
some part of parameter space the one $b$-tag region is more sensitive. For
the current study we scale the cross section assuming the dominance of
the two $b$-tag signal region.
There is no $b$-tag veto employed in the CMS squark search.
Consequently, we can directly use the cross section limits obtained in
that case. These limits are seen to be stronger than the sbottom
limits for lower $m_\phi$ values, or for higher values of
$D_{\lambda,11}$. The latter is easily explained by the fact that
in sbottom searches the increased
production cross section for larger $D_{\lambda,11}$ is offset by a
smaller branching ratio into b-quarks.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=0.45\textwidth]{figures/sbottom1.pdf}
\quad
\includegraphics[width=0.45\textwidth]{figures/sbottom15.pdf}
\end{center}
\caption{Limits on the mDMFV model from the CMS 19.5
fb$^{-1}$ sbottom~\cite{CMS-PAS-SUS-13-018} and
squark~\cite{Chatrchyan:2014lfa} search. We show for comparison the
CMS $\tilde{b}\tilde{b}$ limit (grey curve), which also directly
applies to the case
$D_{\lambda,11}=D_{\lambda,22}=0$. Shaded regions are excluded for
particular choices of $D_{\lambda,11}$ and $D_{\lambda,22}$ as shown on
the plot. For the NLO limit, we estimate
the cross section by multiplying
LO prediction with the NLO K-factor from sbottom pair
production. The DM coupling to $b$-quark, $D_{\lambda,33}$
is fixed everywhere by the corresponding relic abundance
constraint, and all mixing angles are set to zero for simplicity.}
\label{fig:sbottoms1}
\end{figure}
We present our results in the $m_{\phi}$\,--\,$m_\chi$ mass plane in figure
\ref{fig:sbottoms1}. We choose three benchmark coupling values,
$D_{\lambda,11}=D_{\lambda,22}=\{0,1,1.5\}$ to
demonstrate the qualitative nature of the limits. The coupling
$D_{\lambda,33}$ is fixed to yield the correct relic abundance for each
point on the plane, assuming a single flavor freeze-out. We show the limits
from both the CMS 19.5 fb$^{-1}$ sbottom~\cite{CMS-PAS-SUS-13-018} and
squark~\cite{Chatrchyan:2014lfa} searches. The
NLO estimate is obtained from the LO cross section multiplied by
K-factors from sbottom production. Larger $m_\chi$ correspond to a
smaller $D_{\lambda, 33}$, and hence to a smaller $\phi\to b\chi_b$
branching ratio for fixed $D_{\lambda,11}$ and $D_{\lambda,22}$. This
results in a much weaker limit for higher values of $m_\chi$ in figure
\ref{fig:sbottoms1}.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=0.45\textwidth]{figures/couplingvariation.pdf}
\end{center}
\caption{Exclusion contours for $m_\phi$ {(in GeV)} for various values of
$D_{\lambda,11}$ and $D_{\lambda,22}$ from the CMS
sbottom~\cite{CMS-PAS-SUS-13-018} and
squark~\cite{Chatrchyan:2014lfa} searches, obtained for the NLO
estimate. $D_{\lambda,33}$ is fixed by the
relic abundance constraint.}
\label{fig:sbottoms2}
\end{figure}
In figure~\ref{fig:sbottoms2}, we present limits on $m_\phi$ for
different values of $D_{\lambda,11}$ and $D_{\lambda,22}$, again with
$D_{\lambda,33}$ fixed by the relic abundance constraint. We choose a
representative DM mass of $m_\chi = 100$ GeV. This plot shows
the difference in sensitivity to the two different couplings -- the
constraint is much more severe for larger values of $D_{\lambda,11}$
than for corresponding values of $D_{\lambda,22}$ due to the enhanced
production cross section in the former case.
\subsection{Monojet searches}
Another important constraint on the mDMFV parameter space is placed by
the searches for monojets and large missing $E_T$. There are two
interesting regimes for the monojet searches, the effective field
theory (EFT) regime {with a heavy mediator} and the
compressed spectrum regime.
For large mediator masses, the dominant signal arises from
pair-production of DM along with a hard initial state
radiation (ISR) jet. This can be
crudely modeled by an EFT, which has the advantage of being largely
model independent. Limits are presented as lower bounds on $\Lambda$,
the scale associated with the coefficient of the higher dimensional
operator~\cite{CMS-PAS-EXO-12-048}. However, exclusions using this
technique have to be interpreted
with care. In principle, the validity of the EFT fails for $\Lambda
\lesssim p_T$, the $p_T$ cut employed in the searches. Thus,
technically we can only exclude a band in $\Lambda$ self-consistently
where the EFT is expected to be valid. CMS and ATLAS have put limits
on various DM-quark EFT operators. The flavor specific
operators that arise in our case have not been considered. However,
we can estimate these limits by matching cross sections in our model
with those from the DM EFT used by CMS/ATLAS.
In this case it is important to note that diagrams with an initial
state $d$ quark are enhanced by the valence quark pdf. This process is
therefore directly sensitive to the coupling $D_{\lambda,11}$,
suppressing that coupling would help to effectively evade the monojet
constraints. In the DM mass regime we consider, $10\, {\rm GeV}<m_\chi<250\, {\rm GeV}$, the limits do not vary significantly. We show two
benchmarks $m_\chi=10$ GeV and $m_\chi = 210$ GeV as examples in
figure \ref{fig:monojetscoup}.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=0.45\textwidth]{figures/monocoup10.pdf}
\qquad
\includegraphics[width=0.45\textwidth]{figures/monocoup210.pdf}
\end{center}
\caption{Limits on the mDMFV model in terms of mass of $\phi$ (in
GeV) from the CMS monojet
search~\cite{CMS-PAS-EXO-12-048}. Limits are obtained by matching
production cross sections in our model with those in the DMEFT used
by CMS. }
\label{fig:monojetscoup}
\end{figure}
The second regime in which the monojets provide a constraint
complementary to the multijet searches is in the compressed region,
{$m_\phi \sim m_\chi $}. In this region, the jets from $\phi$ decays
are typically too soft to appear in multijet analyses. With a hard ISR
jet, the signal resembles the DM monojet signal. The
branching ratios of $\phi$ to different quark flavors are irrelevant
in this case, since these decay products are typically too soft to
pass the selection cuts. We show the limits obtained from recasting
the CMS search for stops decaying to charm and a neutralino
\cite{CMS-PAS-SUS-13-009} in figure
\ref{fig:monojets}.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=0.45\textwidth]{figures/monojet1.pdf}
\includegraphics[width=0.45\textwidth]{figures/monojet15.pdf}
\end{center}
\caption{Limits on the mDMFV model from the CMS 19.7
fb$^{-1}$ search~\cite{CMS-PAS-SUS-13-009} for stops decaying to a
charm and a neutralino,
using the monojet + MET final state.
The NLO cross section is estimated
from the LO prediction with the NLO K-factor from sbottom pair
production. The triangular grey regions are parts of parameter space
where $\phi$ is lighter than the dark matter particles and hence
cannot decay into them.
}
\label{fig:monojets}
\end{figure}
Since the monojet analyses typically veto additional jets, they lose
sensitivity as the mass splitting between $\phi$ and $\chi$ increases.
It has been pointed out that the razor analysis can be powerful in
this region of parameter space~\cite{Fox:2012ee,Agrawal:2013kha,
Chatrchyan:2014goa}. It is an interesting future direction to analyze
the constraints on this framework from razor.
A more quantitative analysis of the constraints on the masses of
$\phi$ and $\chi_i$ would require a dedicated recasting of the
experimental analyses which is beyond the scope of the
present paper.
\subsection{Distinctive signatures}
We see that a number of current searches are sensitive to this model.
However, the presence of multiple flavors of DM have
potentially novel signatures which could distinguish these models with
other models of DM and colored mediators.
In the decay of $\phi$ all three DM flavors $\chi_i$ will be
produced. While the lightest state $\chi_b$ is stable and simply
escapes detection, the heavier states will decay to the lightest one
through the decay channels discussed in appendix \ref{sec:decayheavy}.
In the case when the mass splitting between the $\chi_i$ are small,
the decay products of the heavier states (jets, leptons or photons)
can be too soft by themselves for the current searches to be
sensitive. However, it is an interesting possibility to look for these
relatively soft objects in conjunction with high $p_T$ jets. In the
compressed regime, the heavier flavors can also have displaced decays,
or decays in the detector.
When the mass splitting is larger, we expect a multijet signal,
including heavy-flavor jets in addition to missing transverse
momentum. Observation of these cascade decays would help in
disentangling our model from other models of DM with
colored mediators.
\section{Conclusions}\label{sec:conclusions}
In this paper we presented a simplified model of flavored DM. The fermionic DM $\chi$ carries flavor and couples to right-handed down
quarks transmitted by a colored scalar mediator $\phi$.
We introduced the Dark Minimal Flavor Violation (DMFV) framework where the extended flavor symmetry $U(3)_q \times U(3)_u \times U(3)_d \times U(3)_\chi$ is broken only by the SM Yukawa couplings $Y_u$, $Y_d$ and the DM-quark coupling~$\lambda$.
In general the flavor violating effects
generated by $\lambda$ can be large. Hence DMFV, while being conceptually similar to MFV, is
phenomenologically very different from the latter scenario.
The DMFV
assumption has the following implications that are very useful from a
model building perspective:
\begin{itemize}
\item
If DMFV is exact, an unbroken $\mathbb{Z}_3$ subgroup stabilizes the DM. It also prevents the new particles from being singly produced at the LHC.
\item
The DMFV hypothesis allows a reduction in the number of free parameters. The matrix $\lambda$ can be parametrized in terms of a unitary matrix $U_\lambda$ and a positive and diagonal matrix $D_\lambda$.
\end{itemize}
We performed a detailed phenomenological analysis of the minimal DMFV
(mDMFV) model, focusing at first on the available constraints from flavor changing neutral current processes.
Our findings can be summarized as follows:
\begin{itemize}
\item Meson antimeson mixing observables put strong constraints on the structure of the coupling matrix $\lambda$. Yet it is possible to satisfy all existing $\Delta F=2$ constraints simultaneously if $\lambda$ obeys one of the following scenarios:
\begin{enumerate}
\item Universality scenario -- The diagonal coupling matrix $D_\lambda$ is nearly universal and large mixing angles in $U_\lambda$ are allowed.
\item Small mixing scenario -- The mixing matrix $U_\lambda$ is close to the unit matrix, while $D_\lambda$ is arbitrary.
\item $ij$-degeneracy scenarios ($ij=12,13,23$) -- Two elements of $D_\lambda$ are almost equal, $D_{\lambda,ii}\simeq D_{\lambda,jj}$, and only the corresponding mixing angle $s^\lambda_{ij}$ is allowed to be large.
\end{enumerate}
Flavor violating effects are then efficiently suppressed by an almost diagonal matrix $\lambda \lambda^\dagger$. Our scenarios can straightforwardly be implemented in future studies. Note that their validity extends beyond the minimal model.
\item
The mDMFV model does not yield any new contribution to rare
(semi-)leptonic $K$ and $B$ decays.\footnote{A non-SM experimental
signature for decays with neutrinos in the final state may arise if
the DM particle is light enough that the decay into a
$\chi\bar\chi$ final state is kinematically accessible
\cite{Kamenik:2011vy}.} Also the strongly constrained $B\to X_s\gamma$
decay does not receive any significant new contribution. Similarly
new contributions to electroweak precision observables and
electric dipole moments are small in mDMFV.
\end{itemize}
In the second step the constraints from DM phenomenology are
included. We restricted our attention to the case of $b$-flavored DM
which is phenomenologically most appealing. We also assumed
that the observed DM relic density arises from thermal
freeze-out. Our results are:
\begin{itemize}
\item
The tree level contribution to DM-nucleon scattering, strongly
constrained by the recent LUX data, needs to be suppressed by a small
mixing angle $s^\lambda_{13}$.
\item
Further contributions to DM-nucleon scattering arise at the one-loop
level from a box diagram and a photon penguin. These contributions
become relevant for DM masses $m_{\chi_b}\lsim 100\, {\rm GeV}$. The
penguin and box contributions carry a different overall sign, and
hence destructively interfere if they are comparable in magnitude. The
LUX constraint requires a partial cancellation, leading to both an
upper and a lower bound on the size of $D_{\lambda,11}$.
\item
We observe a non-trivial interplay of flavor and DM constraints, such
that the combined constraint on the parameter
space of $\lambda$ is an interesting overlap of the
individual ones.
\item
We studied several mass spectra implying different freeze-out conditions.
In all cases the penguin box cancellation implies a sizable
splitting $D_{\lambda,11}<D_{\lambda,33}$ for DM masses
$m_{\chi_b}\lsim 100\, {\rm GeV}$. This rules out the 13-degeneracy scenario
below this scale.
\end{itemize}
Stringent constraints on the mDMFV model also arise from direct
searches at the LHC. We estimate the bounds arising from $b$-jets
$+\,{}/ \hspace{-1.5ex} E_{T}$ and monojet searches and point out possible new signatures.
We find:
\begin{itemize}
\item
Searches for direct sbottom quark production and for multiple jets plus missing energy
constrain the mDMFV parameter space up to $m_\phi\simeq 800-900\, {\rm GeV}$.
\item
Monojet searches are complementary to the searches for jets $+\,{}/ \hspace{-1.5ex} E_{T}$ and competitive in
two regions of parameter space, namely for large mediator masses or a compressed spectrum.
\item
The decay of the heavier $\chi$ flavors leads to additional jets or to
soft photons, which could be looked for in conjunction with the
monojet or dijet signals discussed above.
\end{itemize}
Our pioneering study of flavored DM with a generic flavor
violating coupling matrix has revealed a lot of interesting phenomenology
and showed that it is worthwhile to consider DM models beyond MFV.
\subsection*{Acknowledgements}
PA and KG acknowledge partial support by the National Science Foundation under Grant No. PHYS-1066293 and the hospitality of the Aspen Center for Physics.
KG was supported by the
Deutsche Forschungsgemeinschaft (DFG), grant number GE 2541/1-1.
Fermilab is
operated by Fermi Research Alliance, LLC under Contract No. DE-AC02-07CH11359 with the United States
Department of Energy.
\begin{appendix}
\section{Classification of DMFV symmetry structures}\label{app:DMFV-classification}
In this appendix we classify the different possible versions of DMFV.
A common feature is that the SM global flavor symmetry is
extended by a $U(3)_\chi$ in the dark sector. In addition to the SM
Yukawa couplings there is only a single interaction $\lambda$ that
breaks the flavor symmetry. In principle it is straightforward to
implement DMFV in the lepton sector, but we restrict ourselves to the
quark sector in what follows. The global flavor symmetry is then given
by
\begin{equation}
U(3)_q \times U(3)_u \times U(3)_d \times U(3)_\chi\,.
\end{equation}
Throughout this paper we assumed that the DM $\chi$ is a Dirac fermion and the mediator $\phi$ is a scalar, yet other Lorentz structures are also possible. We will not explore this option further.
In our analysis we considered the coupling of DM to the right-handed down type quarks by the structure
\begin{equation}
\mathcal{L}\ni -\lambda \phi \bar d_R \chi_L\,.
\end{equation}
In this case the coupling $\lambda$ induces the flavor symmetry breaking
\begin{equation}
U(3)_d \times U(3)_\chi \to U(1)_{d\oplus\chi}\,.
\end{equation}
We classify this symmetry breaking pattern as $\text{DMFV}_d$ in order to distinguish it from the following scenarios.
In $\text{DMFV}_u$ instead DM couples to the right-handed up quarks,
\begin{equation}
\mathcal{L} \ni -\lambda \phi \bar u_R \chi_L\,,
\end{equation}
so that the symmetry breaking pattern is
\begin{equation}
U(3)_u \times U(3)_\chi \to U(1)_{u\oplus\chi}\,.
\end{equation}
Due to the coupling to up-type quarks, the phenomenology of $\text{DMFV}_u$ is very different from the $\text{DMFV}_d$ case. We leave a detailed study for future work.
Last but not least it is also possible to couple DM to the left-handed quark doublets
\begin{equation}
\mathcal{L} \ni -\lambda \phi \bar q_L \chi_R\,.
\end{equation}
In this case either $\phi$ or $\chi$ has to transform as an $SU(2)_L$ doublet. In this so-called
{$\text{DMFV}_q$}
scenario, $\lambda$ yields the symmetry breaking
\begin{equation}
U(3)_q \times U(3)_\chi \to U(1)_{q\oplus\chi}\,.
\end{equation}
The phenomenology of this scenario is enriched by the fact that now DM couples to both up and down type quarks.
\section{\boldmath $\mathbbm{Z}_3$ as an unbroken subgroup of DMFV}\label{app:Z3}
In this appendix we present the proof for the presence of an unbroken
$\mathbbm{Z}_3$ subgroup emerging from the $SU(3)_\text{QCD} \times
U(3)_q\times U(3)_u\times U(3)_d \times U(3)_\chi$ symmetry in the
limit of exact DMFV, where $\chi$ transforms as a triplet under
$U(3)_\chi$. The argument is quite analogous to the one provided for the
MFV case in \cite{Batell:2011tc}. The unbroken $\mathbbm{Z}_3$
symmetry automatically forbids all operators inducing DM decay even at
the non-renormalizable level, {as long as they preserve DMFV}. It also prevents the new particles from
being singly produced at the LHC. Note that the validity of the proof
does not depend on the Lorentz structure as it is based purely on
internal symmetry arguments.
Let us consider the operator
\begin{equation}
\mathcal{O} \sim
\chi \dots \bar \chi\dots \phi\dots \phi^\dagger\dots q_L\dots \bar q_L\dots u_R\dots \bar u_R \dots d_R\dots \bar d_R \dots G \dots \mathcal{S}\,,
\end{equation}
where the dots represent an arbitrary number of insertions of the same fields, $G$ schematically denotes the gluon field or field strength and $S$ contains a combination of SM fields that are neutral under both QCD and the flavor symmetry. This is the most generic structure of interactions between the dark sector and SM particles. In the case $N_\chi=1, N_{\bar\chi}=N_\phi=N_{\phi^\dagger} =0$ it mediates DM decay.
This structure is invariant under QCD, if the number of $SU(3)_c$
triplet minus the number of $SU(3)_c$ anti-triplets is a multiple of three:
\begin{equation}
(N_\phi- N_{\phi^\dagger} + N_q + N_u + N_d - N_{\bar q} - N_{\bar u} - N_{\bar d}) \mod 3 = 0 \,.\label{eq:QCD}
\end{equation}
Invariance of the operator $\mathcal{O}$ under the flavor symmetry can formally be achieved by treating the Yukawa couplings $Y_u$, $Y_d$ and $\lambda$ as spurion fields and adding a combination
\begin{equation}
Y_u\dots Y_u^\dagger \dots Y_d\dots Y_d^\dagger\dots \lambda\dots \lambda^\dagger\dots
\end{equation}
to the operator $\mathcal{O}$.
Invariance under $U(3)_q$ then requires
\begin{equation}
(N_q-N_{\bar q} + N_{Y_u} - N_{Y_u^\dagger} + N_{Y_d} - N_{Y_d^\dagger} )\mod 3=0 \,.\label{eq:U3q}
\end{equation}
Similarly $U(3)_u$, $U(3)_d$ and $U(3)_\chi$ invariance give, respectively,
\begin{eqnarray}
(N_u-N_{\bar u} - N_{Y_u} + N_{Y_u^\dagger} )\mod 3&=&0 \,,\\
(N_d-N_{\bar d} - N_{Y_d} + N_{Y_d^\dagger} + N_\lambda - N_{\lambda^\dagger})\mod 3&=&0 \,,\\
(N_\chi - N_{\bar \chi}- N_\lambda + N_{\lambda^\dagger})\mod 3&=&0 \,.\label{eq:U3chi}
\end{eqnarray}
Adding \eqref{eq:U3q}--\eqref{eq:U3chi} and subtracting \eqref{eq:QCD} we find
\begin{equation}\label{eq:Z3}
(N_\chi - N_{\bar \chi} - N_\phi + N_{\phi^\dagger} )\mod 3 =0
\end{equation}
as a necessary condition for the operator $\mathcal{O}$ to be invariant under QCD and the flavour symmetries. Eq. \eqref{eq:Z3} can be interpreted as a $\mathbb{Z}_3$ symmetry under which $\chi$ and $\phi$ carry the charges $e^{+2\pi i/3}$ and $e^{-2\pi i/3}$, respectively, while the SM fields carry the charge $+1$.
Consequently the combination of QCD and the DMFV assumption, together with our choice of representations forbid both $\chi$ and $\phi$ to decay into SM fields. Therefore {in the exact DMFV limit} the lightest state out of $\phi$ and the components of $\chi$ is stable without the need of imposing an additional symmetry.
\section{Relevant loop functions}
\label{app:functions}
In this appendix we collect the one loop functions relevant for our analysis.
The NP one loop function for $\Delta F=2$ processes in mDMFV reads
\begin{equation}
F(x_i,x_j) = \frac{x_i^2 \log x_i}{(x_i-x_j)(1-x_i)^2} + \frac{x_j^2 \log x_j}{(x_j-x_i)(1-x_j)^2} + \frac{1}{(1-x_i)(1-x_j)}\,,
\end{equation}
where $x_i=m_{\chi_i}^2/m_\phi^2$. In the limit of degenerate masses $m_\chi$ it reduces to
\begin{equation}
F(x)= \frac{2x \log x}{(1-x)^3} + \frac{(1+x)}{(1-x)^2}
\end{equation}
with $x=m_\chi^2/m_\phi^2$.
The NP one loop contribution to the $b\to s\gamma$ transition is given by
\begin{equation}
g(x) = \frac{1-5x-2x^2}{6(1-x)^3}-\frac{x^2\log x}{(1-x)^4}\,,\qquad x_i = \frac{m_{\chi_i}^2}{m_\phi^2}\,.
\end{equation}
\section{Decay of heavier flavors}
\label{sec:decayheavy}
DM flavors pick up mass splitting either through threshold
corrections or through running. Consequently, the heavier $\chi$ flavors can decay into the lighter DM. The open decay modes
depend both on the size of mass splittings as well as on the structure
of the coupling matrix. As in the text, we focus on the case where
$\chi_b$ is the lightest flavor, and hence constitutes the DM
today. The relevant diagrams are shown in figure \ref{fig:decays}.
The dominant decay mode arises at tree-level, where the heavier DM
decays into a pair of quarks and a lighter DM. This
decay mode is only open if the splitting between the DM
flavors is more than the lightest allowed mesons in the final state.
Since this is a flavor-conserving decay, it is expected to dominate
whenever allowed by phase space. Further, for couplings relevant for
us, this decay mode is prompt on the timescale of BBN.
For more compressed spectra, the quark/meson final states may not be
kinematically allowed. In this case, flavor-violating decays into a
photon or a pair of leptons via an off-shell photon will be present.
These can be estimated by using the effective coupling of DM
particles with a photon~\cite{Agrawal:2011ze,Kopp:2009et},
\begin{align}
\mathcal{L}_{\rm eff} &=
\sum_{i=d,s,b}\frac{-\lambda_{ib}^* \lambda_{ih}e}{64\pi^2 m_\phi^2} \left[
\left(
\frac12
+\frac23\log\left[\tfrac{m_{q_i}^2}{m_\phi^2}\right]
\right)
\mathcal{O}_{1,bh} + \frac14 \mathcal{O}_{2,bh} \right] ,
\end{align}
where the subscript $h$ refers to the decaying heavier flavor of DM.
The operators are given by
\begin{align}
\mathcal{O}_{1,bh} &=
\left[
\bar{\chi}_b
\gamma^\mu(1-\gamma^5)\partial^\nu
\chi_h\;
+h.c.\right] F_{\mu\nu} \,,
\\
\mathcal{O}_{2,bh} &=
\left[
i\bar{\chi}_b
\gamma^\mu(1-\gamma^5)\partial^\nu
\chi_h\;
+h.c.\right]
F^{\sigma\rho} \epsilon_{\mu\nu\sigma\rho} \,,
\label{eq:ops}
\end{align}
and we have only kept the leading terms in momentum transfer,
$\mathcal{O}(k^2/m_\phi^2)$, and in $m_{q}/m_\phi$.
\begin{figure}[tp]
\centering
\includegraphics[width=0.28\textwidth]{figures/figchi3bd.pdf}
\qquad
\includegraphics[width=0.28\textwidth]{figures/figchidecay1loop.pdf}
\qquad
\includegraphics[width=0.28\textwidth]{figures/figchidecay3bd1loop.pdf}
\caption{Decays of heavier flavors of DM into $\chi_b$.}
\label{fig:decays}
\end{figure}
We see immediately that operator $\mathcal{O}_1$ does not give rise to
an on-shell photon decay, since the amplitude in that case is proportional to
$k^2$. Similarly, the operator $\mathcal{O}_2$ contribution is seen to
vanish because of a GIM-like cancellation. This is reminiscent of the
decay $b \to s\gamma$ in the SM, where an analogous GIM cancellation
causes the dominant decay to arise at a
higher order in the EFT.
We can then estimate the decay rate by the SM partonic $b\to s\gamma$
rate. We replace $G_F$ by the coefficient of the effective operator
above, and account for an additional color factor. While
the SM has spin-1 $W$ bosons running in the loop instead of
the scalar $\phi$, this estimate should suffice for our purposes.
The dominant contribution comes from the heaviest quark, i.e. the
$b$-quark in the loop. The decay rate estimate is
\begin{align}
\nonumber
\Gamma (\chi_h \to \chi_b \gamma)
&\sim
\frac{9
|\lambda_{bh} \lambda_{bb}|^2
}{2\pi}
\left[
\frac{m_b^2}{m_\phi^2}
\frac{e\ m_\chi
}{
(64 \pi^2 m_\phi^2)
}
\right]^2
\,
\delta m^3\,,
\\&\sim
|\lambda_{bh} \lambda_{bb}|^2
\left(
\frac{1}{10^5\ \mathrm{ sec}}
\right)
\left(
\frac{1000\ \mathrm{ GeV}}{m_\phi}
\right)^8
\left(
\frac{m_\chi}{10\ \mathrm{ GeV}}
\right)^2
\left(
\frac{\delta m}{0.2\ \mathrm{ GeV}}
\right)^3 ,
\end{align}
where $\delta m$ is the mass splitting between the heavy and the
$b$-flavored DM. We see that this decay mode is too slow unless the dark
matter is relatively heavy and the splitting is sizeable.
We now estimate the third contribution in figure~\ref{fig:decays}. The
operator $\mathcal{O}_1$ in equation~\eqref{eq:ops} induces a
four-fermion contact operator of the DM with SM leptons and light
quarks. We can hence calculate the three-body decay rate in analogy
with muon decay. The presence of the log prevents the GIM-like
cancellation in this case. The estimate for the decay rate to light
SM fermions is given as,
\begin{align}
\Gamma (\chi_h \to \chi_b f \bar{f})
&\sim
|\lambda_{bh} \lambda_{bb}|^2
\left[
\frac{e^2 }{96 \pi^2 m_\phi^2}
\log\left[\frac{m_b^2}{m_{q,h}^2}\right]
\right]^2
\frac{9\ \delta m^5}{192 \pi^3}
\,,
\nonumber\\&\sim
|\lambda_{bh} \lambda_{bb}|^2
\left(
\frac{1}{1\ \mathrm{ sec}}
\right)
\left(
\frac{1000\ \mathrm{GeV}}{m_\phi}
\right)^4
\left(
\frac{\delta m}{0.2\ \mathrm{GeV}}
\right)^5
,
\end{align}
where $m_{q,h}$ is the mass of the quark associated with the heavy
flavor.
We see that for relatively large values of the mass splitting,
the DM decays well before the BBN epoch. On the edges of our
parameter space, a more careful calculation is needed, especially in
the region of small mixing angles.
\end{appendix}
|
\part{\partial}
\define\emp{\emptyset}
\define\imp{\implies}
\define\ra{\rangle}
\define\n{\notin}
\define\iy{\infty}
\define\m{\mapsto}
\define\do{\dots}
\define\la{\langle}
\define\bsl{\backslash}
\define\lras{\leftrightarrows}
\define\lra{\leftrightarrow}
\define\Lra{\Leftrightarrow}
\define\hra{\hookrightarrow}
\define\sm{\smallmatrix}
\define\esm{\endsmallmatrix}
\define\sub{\subset}
\define\bxt{\boxtimes}
\define\T{\times}
\define\ti{\tilde}
\define\nl{\newline}
\redefine\i{^{-1}}
\define\fra{\frac}
\define\un{\underline}
\define\ov{\overline}
\define\ot{\otimes}
\define\bbq{\bar{\QQ}_l}
\define\bcc{\thickfracwithdelims[]\thickness0}
\define\ad{\text{\rm ad}}
\define\Ad{\text{\rm Ad}}
\define\Hom{\text{\rm Hom}}
\define\End{\text{\rm End}}
\define\Aut{\text{\rm Aut}}
\define\Ind{\text{\rm Ind}}
\define\IND{\text{\rm IND}}
\define\ind{\text{\rm ind}}
\define\Res{\text{\rm Res}}
\define\res{\text{\rm res}}
\define\Ker{\text{\rm Ker}}
\define\Gal{\text{\rm Gal}}
\redefine\Im{\text{\rm Im}}
\define\sg{\text{\rm sgn}}
\define\tr{\text{\rm tr}}
\define\dom{\text{\rm dom}}
\define\supp{\text{\rm supp}}
\define\card{\text{\rm card}}
\define\bst{\bigstar}
\define\he{\heartsuit}
\define\clu{\clubsuit}
\define\di{\diamond}
\redefine\spa{\spadesuit}
\define\a{\alpha}
\redefine\b{\beta}
\redefine\c{\chi}
\define\g{\gamma}
\redefine\d{\delta}
\define\e{\epsilon}
\define\et{\eta}
\define\io{\iota}
\redefine\o{\omega}
\define\p{\pi}
\define\ph{\phi}
\define\ps{\psi}
\define\r{\rho}
\define\s{\sigma}
\redefine\t{\tau}
\define\th{\theta}
\define\k{\kappa}
\redefine\l{\lambda}
\define\z{\zeta}
\define\x{\xi}
\define\vp{\varpi}
\define\vt{\vartheta}
\define\vr{\varrho}
\redefine\G{\Gamma}
\redefine\D{\Delta}
\define\Om{\Omega}
\define\Si{\Sigma}
\define\Th{\Theta}
\redefine\L{\Lambda}
\define\Ph{\Phi}
\define\Ps{\Psi}
\redefine\aa{\bold a}
\define\bb{\bold b}
\define\boc{\bold c}
\define\dd{\bold d}
\define\ee{\bold e}
\define\bof{\bold f}
\define\hh{\bold h}
\define\ii{\bold i}
\define\jj{\bold j}
\define\kk{\bold k}
\redefine\ll{\bold l}
\define\mm{\bold m}
\define\nn{\bold n}
\define\oo{\bold o}
\define\pp{\bold p}
\define\qq{\bold q}
\define\rr{\bold r}
\redefine\ss{\bold s}
\redefine\tt{\bold t}
\define\uu{\bold u}
\define\vv{\bold v}
\define\ww{\bold w}
\define\zz{\bold z}
\redefine\xx{\bold x}
\define\yy{\bold y}
\redefine\AA{\bold A}
\define\BB{\bold B}
\define\CC{\bold C}
\define\DD{\bold D}
\define\EE{\bold E}
\define\FF{\bold F}
\define\GG{\bold G}
\define\HH{\bold H}
\define\II{\bold I}
\define\JJ{\bold J}
\define\KK{\bold K}
\define\LL{\bold L}
\define\MM{\bold M}
\define\NN{\bold N}
\define\OO{\bold O}
\define\PP{\bold P}
\define\QQ{\bold Q}
\define\RR{\bold R}
\define\SS{\bold S}
\define\TT{\bold T}
\define\UU{\bold U}
\define\VV{\bold V}
\define\WW{\bold W}
\define\ZZ{\bold Z}
\define\XX{\bold X}
\define\YY{\bold Y}
\define\ca{\Cal A}
\define\cb{\Cal B}
\define\cc{\Cal C}
\define\cd{\Cal D}
\define\ce{\Cal E}
\define\cf{\Cal F}
\define\cg{\Cal G}
\define\ch{\Cal H}
\define\ci{\Cal I}
\define\cj{\Cal J}
\define\ck{\Cal K}
\define\cl{\Cal L}
\define\cm{\Cal M}
\define\cn{\Cal N}
\define\co{\Cal O}
\define\cp{\Cal P}
\define\cq{\Cal Q}
\define\car{\Cal R}
\define\cs{\Cal S}
\define\ct{\Cal T}
\define\cu{\Cal U}
\define\cv{\Cal V}
\define\cw{\Cal W}
\define\cz{\Cal Z}
\define\cx{\Cal X}
\define\cy{\Cal Y}
\define\fa{\frak a}
\define\fb{\frak b}
\define\fc{\frak c}
\define\fd{\frak d}
\define\fe{\frak e}
\define\ff{\frak f}
\define\fg{\frak g}
\define\fh{\frak h}
\define\fii{\frak i}
\define\fj{\frak j}
\define\fk{\frak k}
\define\fl{\frak l}
\define\fm{\frak m}
\define\fn{\frak n}
\define\fo{\frak o}
\define\fp{\frak p}
\define\fq{\frak q}
\define\fr{\frak r}
\define\fs{\frak s}
\define\ft{\frak t}
\define\fu{\frak u}
\define\fv{\frak v}
\define\fz{\frak z}
\define\fx{\frak x}
\define\fy{\frak y}
\define\fA{\frak A}
\define\fB{\frak B}
\define\fC{\frak C}
\define\fD{\frak D}
\define\fE{\frak E}
\define\fF{\frak F}
\define\fG{\frak G}
\define\fH{\frak H}
\define\fJ{\frak J}
\define\fK{\frak K}
\define\fL{\frak L}
\define\fM{\frak M}
\define\fN{\frak N}
\define\fO{\frak O}
\define\fP{\frak P}
\define\fQ{\frak Q}
\define\fR{\frak R}
\define\fS{\frak S}
\define\fT{\frak T}
\define\fU{\frak U}
\define\fV{\frak V}
\define\fZ{\frak Z}
\define\fX{\frak X}
\define\fY{\frak Y}
\define\ta{\ti a}
\define\tb{\ti b}
\define\tc{\ti c}
\define\td{\ti d}
\define\te{\ti e}
\define\tf{\ti f}
\define\tg{\ti g}
\define\tih{\ti h}
\define\tj{\ti j}
\define\tk{\ti k}
\define\tl{\ti l}
\define\tm{\ti m}
\define\tn{\ti n}
\define\tio{\ti\o}
\define\tp{\ti p}
\define\tq{\ti q}
\define\ts{\ti s}
\define\tit{\ti t}
\define\tu{\ti u}
\define\tv{\ti v}
\define\tw{\ti w}
\define\tz{\ti z}
\define\tx{\ti x}
\define\ty{\ti y}
\define\tA{\ti A}
\define\tB{\ti B}
\define\tC{\ti C}
\define\tD{\ti D}
\define\tE{\ti E}
\define\tF{\ti F}
\define\tG{\ti G}
\define\tH{\ti H}
\define\tI{\ti I}
\define\tJ{\ti J}
\define\tK{\ti K}
\define\tL{\ti L}
\define\tM{\ti M}
\define\tN{\ti N}
\define\tO{\ti O}
\define\tP{\ti P}
\define\tQ{\ti Q}
\define\tR{\ti R}
\define\tS{\ti S}
\define\tT{\ti T}
\define\tU{\ti U}
\define\tV{\ti V}
\define\tW{\ti W}
\define\tX{\ti X}
\define\tY{\ti Y}
\define\tZ{\ti Z}
\define\tcc{\ti\cc}
\define\sha{\sharp}
\define\Mod{\text{\rm Mod}}
\define\Ir{\text{\rm Irr}}
\define\sps{\supset}
\define\app{\asymp}
\define\uP{\un P}
\define\bnu{\bar\nu}
\define\bc{\bar c}
\define\bp{\bar p}
\define\br{\bar r}
\define\bg{\bar g}
\define\hC{\hat C}
\define\bE{\bar E}
\define\bS{\bar S}
\define\bP{\bar P}
\define\bce{\bar\ce}
\define\tce{\ti\ce}
\define\bul{\bullet}
\define\uZ{\un Z}
\define\che{\check}
\define\cha{\che{\a}}
\define\chg{\che g}
\define\bfU{\bar\fU}
\define\Rf{\text{\rm R}}
\define\tfK{\ti{\fK}}
\define\tfD{\ti{\fD}}
\define\tfC{\ti{\fC}}
\define\bat{\bar\t}
\define\dcl{\dot{\cl}}
\define\cir{\circ}
\define\ndsv{\not\dsv}
\define\prq{\preceq}
\define\tss{\ti{\ss}}
\define\tSS{\ti{\SS}}
\define\tcj{\ti{\cj}}
\define\tcp{\ti{\cp}}
\define\tcf{\ti{\cf}}
\define\tcb{\ti{\cb}}
\define\tcy{\ti{\cy}}
\define\y{\ti r}
\define\tgtir{\tg{\y}}
\define\sscj{\ss_\cj}
\define\tsstcj{\tss_{\tcj}}
\define\ccl{\check{\cl}}
\define\tip{\ti\p}
\define\chR{\check R}
\define\tcw{\ti{\cw}}
\define\tfc{\ti{\fc}}
\define\AL{AL}
\define\BC{BC}
\define\BL{BL}
\define\DL{DL}
\define\KL{KL}
\define\SPE{L1}
\define\OBC{L2}
\define\ORA{L3}
\define\LCE{L4}
\define\HEC{L5}
\define\BAR{L6}
\define\LV{LV}
\subhead 1.1\endsubhead
Let $W$ be a finite, irreducible Coxeter group and let S be the set of simple
reflections of $W$; let $l:W@>>>\NN$ be the length function.
Let $\Irr W$ be a set of representatives for the
isomorphism classes of irreducible representations of $W$ over $\CC$, the
complex numbers. Let $\ca=\CC[v,v\i]$, $\ca'=\CC[v^2,v^{-2}]$ ($v$ an indeterminate).
We have $\ca'\sub\ca\sub K\supset K'\supset\ca'$ where $K=\CC(v),K'=\CC(v^2)$.
Let $H$ be the Hecke algebra over $\ca$ associated to $W$; thus
$H$ has generators $T_s (s\in S)$ and generators $(T_s+1)(T_s-v^2)=0$ for
$s\in S$, $T_sT_{s'}T_s\do=T_{s'}T_sT_{s'}\do$ for $s\ne s'$ in $S$ (both
products have $m$ factors where $m$ is the order of $ss'$ in $W$).
Let $H'$ be the $\ca'$-subalgebra of $H$ generated by
$T_s (s\in S)$; note that $H=\ca\ot_{\ca'}H'$.
Let $H_K=K\ot_\ca H$, $H_{K'}=K'\ot_{\ca'}H'$ so that $H_K=K\ot_{K'}H_{K'}$.
It is known \cite{\OBC} (see also 1.2 below) that the algebra $H_K$ is canonically
isomorphic
to the group algebra $K[W]$. Hence any $E\in\Irr W$ can be viewed as a simple
$H_K$-module $E_v$. We say that $E$ is {\it ordinary} if $E_v$ is obtained by
extension of scalars from an $H_{K'}$-module; otherwise, we say that $E$ is
{\it exceptional}. Let $\Irr_0W$ (resp. $\Irr_1W$) be the set of all
$E\in\Irr W$ which are ordinary (resp. exceptional).
We define a subset $\ce W$ of $\Irr W$ as follows.
If $W$ is not of type $E_7,E_8,H_3,H_4$, we set $\ce W=\emp$.
If $W$ is of type $E_7,E_8,H_3,H_4$, then $\ce W$ consists of
$2^a$ representations of dimension $2^b$ where $2^a=2$ for $E_7,H_3$, $2^a=4$
for $E_8,H_4$ and $2^{a+b}$ is the largest power of $2$ that divides the order
of $W$; thus $2^b$ is $512,4096,4,16$ respectively.
When $W$ is crystallographic we have $\Irr W-\ce W\sub\Irr_0W$ (see \cite{\BC}) and $\ce W\sub\Irr_1W$ (a
result of Springer); hence $\Irr W-\ce W=\Irr_0W$ and $\ce W=\Irr_1W$. The same holds when $W$ is not
crystallographic. (The fact $\ce\sub\Irr_1W$ for $W$ of type $H_3$ was pointed out in \cite{\OBC}.
The fact that any $E\in\Irr W-\ce$ is ordinary for $W$ of type $H_4$ can be seen from the fact that,
according to \cite{\AL}, $E$ can be realized by a $W$-graph which is even (in the sense
that the vertices can be partitioned into two subsets so that no edge connects vertices in the same subset).
In this paper we try to understand various consequences in representation
theory of the existence of exceptional representations.
\subhead 1.2\endsubhead
Let $\{c_w;w\in W\}$ be the basis of $H$ which in \cite{\KL} was denoted by
$\{C'_w;w\in W\}$. Let $\le_{LR},\le_L$ be the preorders on $W$ defined in
\cite{\KL} and let $\si_{LR},\si_L$ be the corresponding equivalence relations
on $W$
(the equivalence classes are called the two-sided cells and left cells
respectively). For $x,y\in W$ we write $c_xc_y=\sum_{z\in W}h_{x,y,z}c_z$.
For $z\in W$ there is a unique number $a(z)\in\NN$ such that for any $x,y$ in
$W$ we have $h_{x,y,z}=\g_{x,y,z\i}v^{a(z)}\mod v^{a(z)-1}\ZZ[v\i]$
where $\g_{x,y,z\i}\in\NN$ and $\g_{x,y,z\i}>0$ for some $x,y$ in $W$.
Moreover, $z\m a(z)$ is constant on any two-sided cell. (See \cite{\HEC}.)
Let $J$ be the $\CC$-vector space with basis $\{t_w;w\in W\}$. It has an
associative $\CC$-algebra structure given by
$t_xt_y=\sum_{z\in W}\g_{x,y,z\i}t_z$; it has a unit element of the form
$\sum_{d\in\cd}t_d$ where $\cd$ is a subset of $W$ consisting of certain
involutions (that is elements with square $1$). (See \cite{\HEC}.)
Let $h\m h^\da$ be the algebra automorphism of $H$ such that $T_s^\da=-T_s\i$ for $s\in S$.
Now the $\ca$-linear map $H@>>>\ca\ot J$ given by
$c_x^\da\m\sum_{d\in\cd,z\in W,d\si_Lz}h_{x,d,z}t_z$ induces an algebra
isomorphism $H_K@>>>K\ot J$ and (by specializing $v=1$) an algebra
isomorphism $\CC[W]@>>>J$ hence an algebra isomorphism $K[W]@>>>K\ot J$. (See \cite{\HEC}.)
Now if $E\in\Irr W$ then $E_v$ in 1.1 is obtained as follows.
We first view $K\ot E$ as a $K\ot J$-module $E_\iy$ via the isomorphism
$K[W]@>>>K\ot J$ above and then view $E_\iy$ as an $H_K$-module $E_v$ via the
isomorphism $H_K@>>>K\ot J$. Note that for $x\in W$ we have
$$\tr(c_x^\da,E_v)=\sum_{d\in\cd,z\in W;d\si_Lz}h_{x,d,z}\tr(t_z,E_\iy).\tag a$$
We show:
(b) {\it If $x\in W$ satisfies $x^2=1$, or more generally, if $x\si_L x\i$ then
there exists $E\in\Irr W$ such that $\tr(t_x,E_\iy)\ne0$.}
\nl
It is enough to show that $\sum_{E\in\Irr W}\tr(t_x,E_\iy)\tr(t_{x\i},E_\iy)\ne0$. The last sum
is equal to the trace of the $K$-linear map $K\ot J@>>>K\ot J$,
$\x\m t_x\x t_{x\i}$ (we use that $t_w\m t_{w\i}$ defines an isomorphism of the algebra $K\ot J$
onto the algebra with opposed multiplication) hence it is equal to
$\sum_{u,u'\in W}\g_{x,u,u'{}\i}\g_{u',x\i,u\i}$. Thus it is enough to show that
the last sum is $\ne0$. Now each term in the last sum is in $\NN$ hence it is
enough to show that for some $u,u'$ we have
$\g_{x,u,u'{}\i}\g_{u',x\i,u\i}>0$. We take $u=x\i$ and $u'=d$ where
$d$ is the unique involution in $\cx$ such that $x\si_L d$. It is enough to show that
$\g_{x,x\i,d}\g_{d,x\i,x}>0$. But the last product is $1$ since $x\si_Ld\si_Lx\i$ so that
$\g_{x,x\i,d}=1,\g_{d,x\i,x}=\g_{x\i,x,d}=1$. (See \cite{\HEC}.) This proves (b).
\subhead 1.3\endsubhead
If $E\in\Irr W$ then there is a unique two-sided cell $c$ such that
$t_x:E_\iy@>>>E_{\iy}$ is nonzero for some $x\in c$. This gives us a (surjective) map
$E\m c$ from $\Irr W$ to the set of two-sided cells; its fibre at a two-sided cell $c$
is denoted by $\Irr^cW$. One checks that if some $E\in\Irr^cW$ is exceptional then any
$E\in\Irr^cW$ is exceptional; in this case we say that $c$ is exceptional. If some/any
$E\in\Irr^cW$ is ordinary, we say that $c$ is ordinary.
An involution $x$ in $W$ is said to be ordinary (resp. exceptional) if $l(x)=a(x)\mod2$
(resp. $l(x)=a(x)+1\mod2$). Note that any two-sided cell $c$ contains some ordinary involution
(for example, $c\cap\cd$ is a nonempty set consisting of ordinary involutions). We show:
(a) {\it If $c$ is an ordinary two-sided cell, then for any $x\in c$ such that $x\si_Lx\i$
we have $l(x)=a(x)\mod2$. In particular, any involution in $c$ is ordinary.}
\nl
By 1.2(a) we can find $E\in\Irr W$ such that $\tr(t_x,E_\iy)\ne0$. By definition we have $E\in\Irr^cW$
hence $E$ is ordinary; since $v^{l(x)}c_x^\da\in H'$, it follows that $\tr(v^{l(x)}c_x^\da,E_v)\in\CC(v^2)$.
Using this and 1.2(a) we deduce
$$v^{l(x)}\sum_{d\in\cd,z\in W;d\si_Lz}h_{x,d,z}\tr(t_z,E_\iy)\in\CC(v^2).\tag b$$
Let $a_0$ be the value of the $a$-function on $c$. For $z,d$ in the last sum such that
$\tr(t_z,E_\iy)\ne0$ we have $z\in c$ hence $a(z)=a_0$ and $h_{x,d,z}=\g_{x,d,z\i}v^{a_0}$ plus a
$\ZZ$-linear combination of strictly smaller powers of $v$; moreover we have $\g_{x,d,z\i}=\g_{z\i,x,d}$
and this is $1$ if $z=x$ and $d$ is the unique element of $\cd$ such that $d\si_Lx$ and is $0$ otherwise.
Thus (b) becomes
$$v^{l(x)}v^{a_0}\tr(t_x,E_\iy)+\text{lin.comb.of strictly smaller powers of }v\in\CC(v^2).$$
Since $\tr(t_x,E_\iy)\in\CC-\{0\}$ it follows that $l(x)+a_0\in2\ZZ$ and (a) follows.
We now show:
(c) {\it If $c$ is an exceptional two-sided cell, then $c$ contains both ordinary and exceptional
involutions. More precisely, any left cell in $c$ contains exactly one ordinary involution and
exactly one exceptional involution.}
\nl
Let $n_c$ (resp. $\tn_c$) be the number of ordinary (resp. exceptional) involutions in $c$.
Let $c'=w_0c$ where $w_0$ is the longest element of $W$. Then $c'$ is again an exceptional two-sided cell.
In type $E_7$ or $H_3$ we have $c'=c$. Since $w_0$ is central in $W$ and of odd length, for any involution
$x$ in $c$, $w_0x$ is again an involution in $c$ and $x$ is ordinary if and only if $w_0x$ is exceptional;
thus we have $n_c=\tn_c$. In type $E_8$ or $H_4$ we have $c'\ne c$; more precisely the value of the
$a$-function on $c$ has a different parity than that on $c'$. Since $w_0$ is central in $W$ and of even
length, for any involution $x$ in $c$, $w_0x$ is an involution in $c'$ and $x$ is ordinary if and only if
$w_0x$ is exceptional; thus we have $n_c=\tn_{c'}$ and $\tn_c=n_{c'}$.
Note that $\Irr^cW$ consists of two elements of dimension $m_c$ where $m_c$ is the number of left cells in
$c$. It is known that if $\G$ is any left cell in $c$ then $\G$ carries a $W$-module structure isomorphic to
the direct sum of the two representations in $\Irr^cW$. In type $E_7,E_8$, using \cite{\ORA, 12.15}, we
deduce that $\G\cap\G\i$ has exactly two elements (a similar result can be proved in type $H_3,H_4$). The
unique element of $\G\cap\cd$ is one of these two elements and is an ordinary involution. Also
any involution in $\G$ is contained in $\G\cap\G\i$. We see that $n_c\ge m_c\ge\tn_c$.
In type $E_7,H_3$ we have $n_c=\tn_c$ hence $n_c=m_c=\tn_c$; we see that (c) holds in this case.
In type $E_8,H_4$ we have $n_c\ge\tn_c$ (and similarly $n_{c'}\ge\tn_{c'}$).
Using $n_c=\tn_{c'}$ and $\tn_c=n_{c'}$ we deduce $n_c=\tn_{c'}=\tn_c=n_{c'}=m_c=n_{c'}$; we see that
(c) holds in this case.
\subhead 1.4\endsubhead
In this subsection we assume that $W$ is crystallographic.
Let $G$ be a simple algebraic group over an algebraic closure $\kk$ of a finite field $\FF_q$ with $q$
elements with a fixed split $\FF_q$-structure such that the Weyl group of $G$ is $W$ in 1.1. The variety
$\cb$ of Borel subgroups of $G$ has a natural $\FF_q$-structure with Frobenius map $F:\cb@>>>\cb$. For each
$w\in W$ let $\co_w$ be the $G$-orbit on $\cb\T\cb$ (diagonal action) indexed by $w$ and let $\bX_w$ be
the closure in $\cb$ of the variety $\{B\in\cb;(B,F(B))\in\co_w\}$ of \cite{\DL}. Now $G(\FF_q)$ acts
naturally on the
$l$-adic intersection cohomology spaces $IH^i(\bX_w)$. An irreducible representation of $G(\FF_q)$ is said to
be unipotent if it appears in the $G(\FF_q)$-module $IH^i(\bX_w)$ for some $w,i$.
Let $\cu_q$ be the a set of representatives for the isomorphism classes of unipotent representations of
$G(\FF_q)$. Let $\r\in\cu_q$. By \cite{\ORA, 3.8}, for any $\r\in\cu_q$, any $z\in W$ and any $j\in\ZZ$ we
have
$$\align&(\r:IH^j(\bX_z))_{G(\FF_q)}\\&=\text{coefficient of }v^j\text{ in }
(-1)^j\sum_{E\in\Irr W}c_{\r,E}\tr(v^{l(z)}c_z,E(v))\endalign$$
where $c_{\r,E}$ are uniquely defined rational numbers and $\tr(v^{l(z)}c_z,E(v))\in\ca$.
Moreover, by \cite{\ORA, 6.17}, given $\r$ as above, there is a unique
two-sided cell $\boc_\r$ of $W$ such that $c_{\r,E}=0$ whenever $E\n\Irr^{\boc_\r}W$.
For a two-sided cell $\boc$ we write $\cu^\boc_q=\{\r\in\cu_q;\boc_\r=\boc\}$. We see that for
any $\r\in\cu_q^\boc$, any $z\in W$ and any $j\in\ZZ$ we have
$$\align&(\r:IH^j(\bX_z))_{G(\FF_q)}\\&=\text{coefficient of }v^j\text{ in }
(-1)^j\sum_{E\in\Irr^\boc W}c_{\r,E}\tr(v^{l(z)}c_z,E(v)).\tag a\endalign$$
To any $\r\in\cu_q$ we associate a sign $\e_\r\in\{1,-1\}$ by the following requirement:
if $\r$ appears in $IH^j(\bX_z)$ with $z\in W,j\in\ZZ$ then $\e_\r=(-1)^j$; this is well defined by
\cite{\ORA, 6.6}.
We say that $\r\in\cu_q$ is ordinary (resp. exceptional) if $\e_\r=1$ (resp. $\e_\r=-1$).
We show:
(b) {\it If $\boc$ is an ordinary two-sided cell then any $\r\in\cu_q^\boc$ is ordinary. If $\boc$ is an
exceptional two-sided cell, then $\cu_q^\boc$ consists of two ordinary and two exceptional representations.}
\nl
Assume first that $\boc$ is ordinary. Since $\tr(v^{l(z)}c_z,E(v))\in\ca'$ for $E\in\Irr^\boc W$, $z\in W$,
we see from (a) that for any $\r\in\cu^\boc_q$ we have $(\r:IH^j(\bX_z))_{G(\FF_q)}=0$ if $j$ is odd. Thus
$\r$ is ordinary. Assume next that $\boc$ is exceptional. Then $\cu^\boc_q$ consists of four representations
of which two appear in $IH^0(\bX_1)$ hence are ordinary and the other two appear in $IH^7(\bX_z)$
where $z$ is an element of length $7$ in $W$.
\subhead 1.5\endsubhead
Let $S_i$ be the $i$-th symmetric power of the reflection representation of $W$ and let $S=\op_iS_i$, a
commutative algebra over $\RR$. Let $I$ be the ideal of $S$ generated by the $W$-invariant elements of $S$
of degree $>0$. Let $\bS=S/I$ and let $\bS_i$ be the image of $S_i$ in $\bS$. Note that $\bS_i$ is a
$W$-module. For any $E\in\Irr W$ we set $P_E=\sum_{i\ge0}(E:\bS_i)X^i\in\NN[X]$. We note the following
property:
(a) {\it If $E$ is ordinary then $P_E$ is palindromic. If $E$ is exceptional then $P_E$ is not palindromic.}
\nl
(A polynomial $P(X)\in\CC[X]$ is said to be palindromic if there exists
$u\in\NN$ such that $P(X\i)=X^{-u}P(X)$.) When $W$ is crystallographic this has been noted in
\cite{\BL}. When $W$ is dihedral or of type $H_3$ this is easily verified.
When $W$ is of type $H_4$ this follows from \cite{\AL}. We will now give an explanation for why (a) holds
assuming that $W$ is crystallographic.
Let $\boc$ be the two-sided cell such that $E\in\Irr^\boc W$.
It is known that $\cu_q^\boc$ (see 1.4) can be naturally indexed by a set independent
of $q$ so that when $\r\in\cu_q^\boc$, the dimension of $q$ can be regarded as a polynomial $\d_\r$ in $q$
with rational coefficients; more precisely, we have $d_\r(X)=e\i X^{a_0}(X-1)^sf(X)$ where $e\in\ZZ_{>0}$,
$a_0\in\NN$ depends only on $\boc$, not on $\r$, $s\in\NN$ is such that $\e_\r=(-1)^s$ and $f$ is a
product
of cyclotomic polynomials $\Ph_r(X)$ with $r\ge2$. Also the degree of the polynomial $X^{a_0}(X-1)^sf(q)$ is
a number $A_0$ depending only on $\boc$, not on $\r$. It follows that $d_\r(X\i)=(-1)^sX^{-a_0-A_0}d_\r(X)$.
We now assume that $\boc$ is ordinary. Then we have $\e_\r=1$ for each $\r$ as above (hence $s$ is even),
see 1.4(b). From \cite{\ORA, 4.23} it is known that $P_E(X)$ is a constant linear combination of polynomials
$d_\r(X)$ with $\r\in\cu^\boc_q$. Since $d_\r(X\i)=X^{-a_0-A_0}d_\r(X)$ for each $\r$ it follows that
$P_E(X\i)=X^{-a_0-A_0}P_E(X)$.
\subhead 1.6\endsubhead
In this subsection we assume that $W$ is crystallographic.
Let $\boc$ be a two-sided cell. Let $E_\boc$ be the special representation in $\Irr^\boc W$ (see
\cite{\ORA}). For each left cell $\G$ in $\boc$ let $[\G]$ be the
$W$-module carried by $\G$. Let $[[\boc]]$ be the $W$-module carried by the set of involutions in $\boc$
defined in \cite{\LV}. We have the following result.
(a) {\it Assume that $\boc$ is ordinary. There is a unique $E\in\Irr^\boc W$ such that $E$ appears in $[\G]$
for every $\G$ as above and $E$ appears in $[[\boc]]$, namely $E=E_\boc$.}
\nl
If $W$ is of classical type, then it is known that $[[\boc]]$ is a sum of copies of $E_\boc$ and that
$E_\boc$ appears in each $[\G]$ with multiplicity one. Hence (a) holds in this case. We now assume that $W$
is of exceptional type. If $\boc$ is not the two-sided cell containing $E$ of dimension
$4480$ (in $E_8$) or $12$ (in $F_4$) or $2$ (in $G_2$) then there is exactly one $E$ which appears in
each $[\G]$ namely $E_\boc$ and $E_\boc$ appears in $[[\boc]]$, see \cite{\LCE}; hence (a) holds in this
case.
We now assume that $\boc$ is the two-sided cell containing $E$ of dimension
$4480$ (in $E_8$) or $12$ (in $F_4$) or $2$ (in $G_2$).
Then there are exactly two $E$ which appear in each $[\G]$ namely $E_\boc$ and the $E$ of dimension
$7168$ (in $E_8$) or $16$ (in $F_4$) or $2$ (non-special) in $G_2$, see \cite{\LCE}. Now one verifies that
$E_\boc$ appears
in $[[\boc]]$ but $7168$ (in $E_8$) or $16$ (in $F_4$) or $2$ (non-special) in $G_2$ do not appear in
$[[\boc]]$. Hence (a) holds in this case.
Note that if $\boc$ is exceptional then there are exactly two $E\in\Irr^\boc W$ such that $E$ appears in
$[\G]$ for every $\G$ as above and $E$ appears in $[[\boc]]$; one of them is $E=E_\boc$.
Let $\sg$ be the sign representation of $W$. The following result has been noted in \cite{\SPE}.
(b) {\it If $\boc$ is ordinary then $E_\boc\ot\sg$ is a special representation.
If $\boc$ is exceptional then $E_\boc\ot\sg$ is not a special representation.}
\nl
Note that the first statement of (b) can be deduced from (a) applied to $w_0\boc$ (an ordinary two-sided
cell) since if $\G'$ is a left cell in $w_0\boc$ then $[\G']\cong[\G]\ot\sg$ for some left cell in $\boc$
and $[[w_0\boc]]\cong[[\boc]\ot\sg$ (this follows from the inversion formula in \cite{\BAR}).
\widestnumber\key{BC}
\Refs
\ref\key\AL\by D.Alvis and G.Lusztig\paper
The representations and generic degrees of the Hecke algebras of type $H_4$\jour J. fu"r reine und
angew.math.\vol336\yr1982\pages201-212\moreref Erratum\vol449\yr1994\pages217-218\endref
\ref\key\BC\by C.T.Benson and C.W.Curtis\paper On the degrees and rationality of certain characters of finite
Chevalley groups\jour Trans.Amer.Math.Soc.\vol165\yr1972\pages251-273\moreref \vol202\yr1975\pages405-406
\endref
\ref\key\BL\by W.M.Beynon and G.Lusztig\paper Some numerical results on the characters of exceptional Weyl
groups\jour Math.Proc.Camb.Phil.Soc.\vol84\yr1978\pages417-426\endref
\ref\key\DL\by P.Deligne and G.Lusztig\paper Representations of reductive groups over finite fields\jour
Ann.Math.\vol103\yr1976\pages103-161\endref
\ref\key\KL\by D.Kazhdan and G.Lusztig\paper Representations of Coxeter groups and Hecke algebras\jour
Inv.Math.\vol53\yr1979\pages165-184\endref
\ref\key\SPE\by G.Lusztig\paper A class of irreducible representations of a Weyl group\jour Proc.Kon.Nederl.
Akad.(A)\vol82\yr1979\pages 323-335\endref
\ref\key\OBC\by G.Lusztig\paper On a theorem of Benson and Curtis\jour J.Alg.\vol71\yr1981\pages490-498
\endref
\ref\key\ORA\by G.Lusztig\book Characters of reductive groups over a finite field\bookinfo Ann.Math.Studies
107\publ Princeton U.Press\yr1984\endref
\ref\key\LCE\by G.Lusztig\paper Sur les cellules gauches des groupes de Weyl\jour C.R.Acad.Sci.Paris(A)
\vol302\yr1986\pages5-8\endref
\ref\key\HEC\by G.Lusztig \book Hecke algebras with unequal parameters
\bookinfo CRM Monograph Ser.18\publ Amer.Math.Soc. \yr2003\endref
\ref\key\BAR\by G.Lusztig\paper A bar operator for involutions in a Coxeter group\jour
Bull.Inst.Math.Acad.Sinica (N.S.)\vol7\yr2012\pages355-404\endref
\ref\key\LV\by G.Lusztig and D.Vogan\paper Hecke algebras and involutions in Weyl groups\jour
Bull. Inst. Math. Acad. Sinica (N.S.)\vol7\yr2012\pages323-354\endref
\endRefs
\enddocument
|
\section{Introduction}
Grand Unified Theories (GUTs) offer one of the most attractive
extensions of the Standard Model of particle physics (SM), unifying
three of the four fundamental forces of nature. Towards a more fundamental
theory of nature it would, however, be desirable to also explain the pattern of fermion
masses and mixing with its plenty of parameters.
Hence, successful models of flavour in GUTs
need to address two main issues, firstly, they have to achieve a working
mechanism for GUT symmetry breaking, including sufficient suppression of proton
decay without unacceptably large fine-tuning, and secondly, they should provide
correct predictions for the flavour observables, such as the
Yukawa coupling ratios and mixing angles.
Existing models which manage to naturally suppress proton decay either predict the
unrealistic quark-lepton Yukawa relation $Y_e=Y_d^T$ (e.g.\ \cite{Berezhiani:1996nu}),
which may only be viable in the presence of extensive uncontrolled higher order
corrections (e.g.\ \cite{Altarelli:2000fu}) rendering the model non-predictive,
or the experimentally disfavoured combination of the Georgi-Jarlskog relations
$y_\mu = - 3 y_s$ and $y_e = \frac{1}{3} y_d$ \cite{GJ} (as e.g.\ in \cite{Murayama:1991ew}),
or rely on linear combinations of GUT Yukawa operators (e.g.\ \cite{Zhang:2012rc}),
which again implies the loss of predictivity. Furthermore, there exists a large
number of GUT models which focus on the flavour sector, but do not include the
Higgs potential.
So while there are several existing models focusing on one of these two concerns, we are not aware of any work to date capable of resolving the two challenges in full detail (and without invoking extra space-time dimensions) in a predictive setup, also given the present rather precise experimental data.
In this paper we employ the framework of supersymmetric (SUSY) GUTs,
as in the minimal supersymmetric extension of the SM (MSSM) the gauge couplings
unify to a surprising precision.
The scale where they unify (about $10^{16}$~GeV) is also large enough to sufficiently suppress proton decay from dimension six GUT operators \cite{Murayama:2001ur}.
However, in addition to proton decay mediated by additional heavy gauge bosons, in a GUT model the minimal embedding of the Higgs fields contains additional colour triplets which also lead to baryon number violating operators. Therefore the colour triplets have to be very heavy to suppress proton decay sufficiently or their couplings to the MSSM fields should be very strongly suppressed. Keeping the doublets of the SU(5) Higgs fields light, while generating a high enough mass for the colour triplets constitutes the so-called ``Doublet-Triplet Splitting (DTS) Problem''.
In SU(5) in four dimensions, a proposed solution to the DTS problem is the ``missing partner mechanism''
(MPM) \cite{Masiero:1982fe,Grinstein:1982um} or its improved version, the ``Double
Missing Partner Mechanism'' (DMPM) \cite{Hisano:1994fn} which we will
review briefly in the next section. In the existing models referred to above, either the MPM or the DMPM is applied.
Since GUTs not only unify the forces of the SM into a single GUT force, but also the fermions into joint GUT representations, they are indeed a promising starting point for addressing the flavour puzzle. More specifically, GUTs are capable of predicting the ratios between the Yukawa couplings of quarks and leptons at the GUT scale. After their renormalization group evolution to low energies (including supersymmetric 1-loop threshold corrections), these predictions can be compared to the experimental data for quark and lepton masses. As well known, the prediction of minimal SU(5) for the charged lepton
and down-type quark Yukawa matrix, the relation $Y_e=Y_d^T$ already mentioned above, is strongly disfavoured by the experimental results on the fermion masses.
But also the ubiquitous proposal for more realistic ratios,
the above mentioned Georgi-Jarlskog relations,
obtained from the introduction of a 45-dimensional
Higgs representation of SU(5) and certain assumptions about the Yukawa
textures \cite{GJ}, are disfavoured by the recently improved data \cite{pdg}.
To arrive at experimentally favoured predictions for GUT scale Yukawa coupling ratios, we employ an alternative approach \cite{Antusch:2009gu} involving higher-dimensional operators which contain a GUT breaking
Higgs field. Due to this, new Clebsch-Gordan (CG) factors appear as Yukawa coupling ratios, with interesting associated implications for the masses and mixing of the fermions, cf. \cite{Antusch:2011qg,Marzocca:2011dh} and \cite{Antusch:2012fb,Meroni:2012ty,Antusch:2013kna,Antusch:2013tta}. The main goal of this paper is to show how GUT flavour models featuring these promising quark-lepton mass relations can be combined with a version of the DMPM for solving the DTS problem.
As explicit examples, we will present two models with these properties that are ``UV complete'' in terms of messenger fields and
employ sets of discrete Abelian symmetries (referred to as shaping symmetries)
such that only the desired effective GUT operators are generated when the heavy degrees of freedom are integrated out. The two models predict the GUT scale quark lepton Yukawa ratios but are not yet predictive for the fermion mixing parameters (although the experimentally observed values can be fitted by both of the models), so we only view them as existence proofs which show that DTS and predictive Yukawa coupling ratios can indeed be combined. The strategies discussed here, however, provide the tools for the construction of more ambitious GUT models of flavour, which should finally also predict the quark and lepton mixings and CP phases (and include the observed neutrino masses).
The paper is organised as follows: We will start with a brief review of the MPM and DMPM and the recently proposed alternative Yukawa coupling ratios. We will also discuss the implications of replacing the 75-dimensional
representation used in the MPM and DMPM with an adjoint 24-dimensional representation
of SU(5). This choice is particularly well suited towards combining the DMPM with the novel
CG factors.
In section \ref{sec:GUT} we discuss the impact of the additional fields
on gauge coupling unification and on their implications for the
colour triplet masses. We will especially focus here on the case of superpotentials with two
adjoint Higgs representations.
We then address the Yukawa sector in section \ref{sec:yukawas}, where we describe
the above-mentioned two predictive example models.
Before we summarise and conclude in section \ref{sec:conc}, in section \ref{sec:proton}
we briefly comment on proton decay, showing it is under
control in the proposed class of models. We present additional helpful material for model building in the appendices.
\section{Strategy \label{sec:strat}}
In this section we present the general strategy that will be implemented in two example models.
We begin our presentation with the review of the MPM and DMPM and the origin of the CG coefficients
(coming from higher dimensional operators)
used in our models. Afterwards
we discuss a modification of the DMPM with a GUT Higgs field in the adjoint representation
and follow that with the actual DMPM realization implemented in our models, where a second Higgs field in the adjoint representation is added.
Throughout this section, for illustrative purposes, we consider that the bounds on proton decay rate require the effective mass of the colour triplets to be of at least \linebreak\mbox{${M^\text{dim=5}_{T}} \approx 10^{17}~{\rm GeV}$} \cite{Goto:1998qg}, while the effective mass suppressing dimension six proton decay mediated by the colour triplets is required to be ${M^\text{dim=6}_{T}} \gtrsim 10^{12}~{\rm GeV}$ \cite{Nath:2006ut}.
\subsection{The missing partner mechanism}\label{subsec:mpm}
The basic idea of the missing partner mechanism (MPM) is the introduction of two
new superfields $Z_{50}$ and $\bar Z_{50}$ in $\mathbf{50}$ and $\mathbf{\overline{50}}$
representations of SU(5).
The decomposition of a $\mathbf{50}$ of SU(5) under the SM gauge group does
not contain an SU(2) doublet, but it includes an SU(3) triplet. Thus, using the
$50$-plets to generate an effective mass term keeps the electroweak doublets massless,
while the colour triplets acquire masses of the order of the GUT scale. The superpotential
for the MPM is given by\footnote{For simplicity we omit most order one coefficients in the superpotentials, except where they are relevant to the discussion.}
\begin{equation}
\label{eq:MPM75}
W_{\text{MPM}} = \bar H_5 H_{75} Z_{50} + \bar Z_{50}H_{75}H_5 + M_{50} Z_{50} \bar Z_{50}\;,
\end{equation}
where $H_{75}$ is a superfield transforming in the $\mathbf{75}$ representation of SU(5),
which contains a SM singlet. When $H_{75}$ gets a vacuum expectation value (VEV) SU(5)
is broken to the SM gauge group\footnote{$H_{75}$ can be replaced by the effective
combination $H_{24}^2/\Lambda$, where $H_{24}$ is the usual GUT-breaking Higgs field in the
$\mathbf{24}$ representation of SU(5) \cite{Berezhiani:1996nu}, see also section \ref{subsec:dmpm24}.}. With the triplet mass contribution from $\vev{H_{75}}$ denoted by
$V$, the mass matrices of the Higgs fields $H_5$, $\bar H_5$ and $Z_{50}$, $\bar Z_{50}$
are given by
\begin{equation}
m_D = 0\;,
\quad m_T = \begin{pmatrix}
0 & V \\ V & M_{50}
\end{pmatrix}\;,
\label{eq:mpmmass}
\end{equation}
for the doublet and triplet components $D$ and $T$ of $H_5$ and $Z_{50}$, respectively.
The dangerous terms for dimension five proton decay are obtained from the
Yukawa couplings
\begin{equation}
W_{\text{Yuk}} = \mathcal{T}_i \mathcal{F}_j \bar H_5 + \mathcal{T}_i \mathcal{T}_j H_5\;,
\end{equation}
where the families of the MSSM matter superfields are embedded in the standard way
in $\mathcal{T}_i$ and $\mathcal{F}_j$, transforming as $\mathbf{10}$ and $\mathbf{\bar 5}$ of SU(5),
respectively. To calculate the effective dimension five proton decay operators all Higgs
triplets from 5- and 50-dimensional representations have to be
integrated out, but only the triplets in the 5-dimensional representations dominantly
couple to matter. We denote the triplet mass eigenvalues with $\tilde M_1$ and $\tilde M_2$,
and the corresponding mass eigenstates as $\tilde T_1$ and $\tilde T_2$, respectively. The
triplets that couple to matter are given by the combinations
\begin{equation}
T^{(5)}=\sum_i U_{1i}^*\tilde T_i\;,\quad \bar T^{(5)}=\sum_i V_{1i}\bar{\tilde T}_i\;,
\end{equation}
where $U$ and $V$ are unitary matrices defined by $m_T = U m_T^\text{diag} V^\dagger$.
Integrating out the triplet mass eigenstates $\tilde T_i$ leads to the effective dimension five
operator for proton decay, which is proportional to the inverse of the ``effective triplet mass''
\begin{equation}
({M^\text{dim=5}_{T}})^{-1}:=U_{1i}^* \left(m_T^\text{diag}\right)^{-1}_{ij}V_{1j}=U_{1i}^* \left(m_T^\text{diag}\right)^{-1}_{ij}V^T_{j1}=(m_T^{-1})_{11}\;.
\end{equation}
Integrating out the heavy colour triplet mass eigenstates also in the K\"ahler potential
\begin{equation}
K_T = T^{(5)}T^{(5)\dagger} + \bar T^{(5)}\bar T^{(5)\dagger} + T^{(50)}T^{(50)\dagger} + \bar T^{(50)}\bar T^{(50)\dagger}\;,
\end{equation}
effective dimension six K\"ahler operators emerge from inserting their equations of motion.
The Lagrangian obtained from the \mbox{D-terms} of $K_T$ contains baryon number
violating four fermion operators. These are proportional to
\begin{equation}
\left({M^\text{dim=6}_{T}}\right)^{-2}:= V_{1i}\left(m_T^\text{diag}\right)^{-1}_{ij}U^\dagger_{jk}U_{km}\left(m_T^\text{diag}\right)^{-1}
_{ml}V_{l1}^\dagger = \left(m_T^{-1}m_T^{\dagger-1}\right)_{11}
\end{equation}
from $T^{(r)}T^{(r)\dagger}$ and to $\left({M^\text{dim=6}_{\bar T}}\right)^{-2}=\left(m_T^{\dagger-1}m_T^{-1}\right)_{11}$ from $\bar T^{(r)}\bar T^{(r)\dagger}$.
With the mass matrix $m_T$ given in eq.~\eqref{eq:mpmmass}, the effective triplet mass is thus\footnote{
In the text when we quote numbers for ${M^\text{dim=5}_{T}}$, ${M^\text{dim=6}_{T}}$ and ${M^\text{dim=6}_{\bar T}}$ we will always refer to their absolute values.}
\begin{equation}
{M^\text{dim=5}_{T}} = \left(m_T^{-1}\right)_{11}^{-1}= -\frac{V^2}{M_{50}}\;,
\end{equation}
while the suppression of dimension six proton decay is given by
\begin{equation}
\left({M^\text{dim=6}_{T}}\right)^2 = \left({M^\text{dim=6}_{\bar T}}\right)^2= \left(m_T^{-1}m_T^{\dagger-1}\right)_{11}^{-1}=\frac{\abs{V}^4}{\abs{M_{50}}^2+\abs{V}^2}\;.
\end{equation}
Note that with a GUT scale value of $V \approx 10^{16}~{\rm GeV}$ and $M_{50}$ below the Planck scale, the dimension six proton decay is suppressed sufficiently with values of ${M^\text{dim=6}_{T}}$ between $10^{13}$ and $10^{16}~{\rm GeV}.$
Since the doublets obtain no mass terms, the splitting of doublet mass and effective triplet mass is achieved.
Using ${M^\text{dim=5}_{T}} \gtrsim 10^{17}~{\rm GeV}$
one obtains an upper bound for $M_{50} \lesssim 10^{15}~{\rm GeV}$. Having the large representations $\mathbf{50}$ and $\mathbf{\overline{50}}$ enter the Renormalization Group Equations (RGEs) at this low mass scale, however, leads to the break down of perturbativity just above the GUT scale. Thus, the MPM solves the DTS problem -- but trades it for SU(5) becoming non-perturbative much below the Planck scale $M_{\text{Pl}}$.
\subsection{The double missing partner mechanism}\label{subsec:dmpm}
This trade-off can be avoided in the double missing partner mechanism (DMPM),
where the number of Higgs fields in $\mathbf{5}$, $\mathbf{\bar 5}$, $\mathbf{50}$ and
$\mathbf{\overline{50}}$ representations gets doubled \cite{Hisano:1994fn}. The
fields $H_5$ and $\bar H_5$ couple to the matter fields $\mathcal{F}_i$ and $\mathcal{T}_i$,
whereas $H_5^\prime$ and $\bar H_5^\prime$ do not. The superpotential for the DMPM is given by
\begin{align}
\label{eq:DMPM75}
W_{\text{DMPM}} & = \bar H_5H_{75} Z_{50} + \bar Z_{50} H_{75}H_5^\prime + \bar H_5^\prime H_{75} Z_{50}^\prime + \bar Z_{50}^\prime H_{75}H_5 \nonumber \\
& + M_{50}Z_{50}\bar Z_{50}+ M_{50}^\prime Z^\prime_{50}\bar Z^\prime_{50}\nonumber \\
& +\mu^\prime H_5^\prime\bar H_5^\prime\;.
\end{align}
The mass matrices of the doublet and triplet components of the Higgs fields $H_5$, $H_5^\prime$,
$Z_{50}$, $Z^\prime_{50}$ and their corresponding barred fields after $H_{75}$ gets a VEV $V$
are given by
\begin{equation}
\label{eq:M_DMPM}
m_D = \begin{pmatrix}
0 & 0 \\
0 & \mu^\prime
\end{pmatrix}\;,
\quad m_T = \begin{pmatrix}
0 & 0 & 0 & V\\
0 & \mu^\prime & V & 0 \\
V & 0 & M_{50} & 0\\
0 & V & 0 & M_{50}^\prime
\end{pmatrix}\;.
\end{equation}
While the Higgs doublets coupling to matter remain massless, the second pair of Higgs doublets
contained in $H_{5}^\prime$ and $\bar H^\prime_5$ has mass $\mu^\prime$. The improvement
of the DMPM compared to the MPM can be seen from the effective triplet mass ${M^\text{dim=5}_{T}}$
\begin{equation}
{M^\text{dim=5}_{T}} = \left(m_T^{-1}\right)_{11}^{-1} = -\frac{V^4}{\mu^\prime M_{50} M^\prime_{50}}\;.
\end{equation}
The same effective triplet mass of ${M^\text{dim=5}_{T}}\approx 10^{17}~{\rm GeV}$ can now be obtained
while keeping high masses $M_{50}\approx M_{50}^\prime \approx 10^{18}~{\rm GeV}$, provided
the heavier doublet pair has a (relatively) small mass $\mu^\prime\approx 10^{11}~{\rm GeV}$. With
the large representations of SU(5) having high masses, the perturbativity of the model
can be preserved up to (almost) the Planck scale. Dimension six proton decay is suppressed by
\begin{align}
\left({M^\text{dim=6}_{T}}\right)^2 &= \left(m_T^{-1}m_T^{\dagger-1}\right)_{11}^{-1}=\frac{\abs{V}^8}{\abs{V}^6 + \abs{M_{50}}^2 (\abs{V}^4 + \abs{M_{50}^\prime\mu^\prime}^ 2 + \abs{V \mu^\prime}^2)}\approx (10^{14}~{\rm GeV})^2\;,\\
(M_{\bar T}^\text{dim=6})^2 &= (m_T^{\dagger -1}m_T^{-1})_{11}^{-1}=
\frac{\abs{V}^8}{\abs{V}^6 + \abs{M_{50}^\prime}^2 (\abs{V}^4 + \abs{M_{50}\mu^\prime}^ 2 + \abs{V \mu^\prime}^2)}\approx (10^{14}~{\rm GeV})^2\;,
\end{align}
in agreement with the bounds on proton decay.
\subsection{Planck-scale suppressed operators \label{PSO}}
The philosophy we follow in this paper is to consider all Planck-scale suppressed operators allowed by the symmetries.
The superpotentials in eq.~(\ref{eq:MPM75}) and eq.~(\ref{eq:DMPM75}) include mass terms for the 50-dimensional messengers. As such, one cannot use symmetries to forbid non-renormalizable Planck-scale suppressed operators such as $H_5 H_{75}^2 \bar{H}_5/M_{\text{Pl}}$ (for the MPM, and $H_5 H_{75}^2 \bar{H}'_5/M_{\text{Pl}}$ and $H'_5 H_{75}^2 \bar{H}_5/M_{\text{Pl}}$ for the DMPM). These Planck-scale suppressed operators do not involve the 50-dimensional messengers and therefore generate dangerously large contributions to the masses of the doublets contained in the 5-dimensional representations, effectively spoiling the mechanism.
Given our philosophy, we must forbid these operators through a shaping symmetry. The MPM and the DMPM can then be restored by adding a singlet field $S$, responsible for giving mass to the 50-dimensional superfields through couplings of the form $S Z_{50}\bar Z_{50}$, $S Z^\prime_{50}\bar Z^\prime_{50}$ and a VEV $\vev{S} \ne 0$, as seen in the diagram in figure \ref{fig:MPM_S} (note $S$ is not acting as an external field but generating the mass term). The non-trivial charge of $S$ under a shaping symmetry forbids the dangerous Planck-scale suppressed operators.
\begin{figure}
\center
\includegraphics[scale=0.75]{MPM75.pdf}
\caption{MPM diagram with an external $S$ field generating the mass term for the 50-dimensional messengers after getting a VEV.
\label{fig:MPM_S}}
\end{figure}
We will generalise this strategy of generating masses for the messenger
fields through an additional singlet field in sections \ref{subsec:dmpm24}
and \ref{sec:yukawas}, to avoid similarly Planck-scale suppressed operators
spoiling the DMPM or the predictions for the Yukawa coupling ratios.
\subsection{Yukawa coupling ratios}
\label{subsec:Clebsch}
The problem of non-viable GUT predictions for the fermion masses in the minimal SU(5)
model, such as $Y_e=Y_d^T$, can be solved through effective Yukawa couplings generated
from higher dimensional operators. When the higher dimensional operators contain a GUT
breaking Higgs field, new ratios between the Yukawa couplings of down-type quarks and
charged leptons can emerge once the GUT symmetry gets spontaneously broken \cite{Antusch:2009gu}.
Due to the introduction of extra SU(5) non-singlet fields, which participate in higher dimensional operators, there are in
general several ways to construct invariants, namely multiple ways to contract
the SU(5) indices. In the general case, such effective operators are not predictive and
introduce an arbitrary, linear combination of several CG factors.
This issue is generic in flavour models and can be resolved by constructing a specific UV completion
of the effective operators, see \cite{Varzielas:2010mp, Varzielas:2012ai} for mixing in lepton models, \cite{Antusch:2009gu, Antusch:2013rxa} for GUT relations, and
for applications in (GUT) flavour models, e.g.~\cite{Meroni:2012ty,Antusch:2013kna,Antusch:2013tta}.
By introducing pairs of heavy messengers in the SU(5) representations $\mathbf{R}$ and $\mathbf{\bar R}$, the renormalizable couplings associated with the effective operators are specified and a unique contraction of SU(5) indices of the effective operator is obtained simply by integrating out the messenger fields.
\begin{figure}
\begin{floatrow}
\ffigbox{
\includegraphics[page=1,scale=0.55]{yukawaratiooperators.pdf}
}{
\caption{Supergraphs generating effectively Yukawa couplings
upon integrating out the pair of messengers fields $R$
and $\bar{R}$.}\label{fig:messengerdiagram1}
}
\capbtabbox{
\caption{CG factors for the dimension five effective operators
$W \supset (A B)_{R} (C D)_{\bar{R}}$. See the main text,
Figure~\ref{fig:messengerdiagram1} and \cite{Antusch:2009gu,
Antusch:2013rxa} for more details.} \label{tab:effectiveFToperators11}
}{
\rule{7mm}{0pt}\newline
\begin{tabular}{cccc}
\toprule
$A \, B$ & $C \, D$ & $R$ & $(Y_e)_{ji} / (Y_d)_{ij}$ \\
\midrule
$H_{24} \, \mathcal{F}$ & $\mathcal{T} \, \bar H_{45}$ & $\overline{\mathbf{45}}$ & $-\tfrac{1}{2}$ \\[0.3ex]
$H_{24} \, \mathcal{F}$ & $\mathcal{T} \, \bar H_5$ & $\bar{\mathbf{5}}$ & $-\tfrac{3}{2}$ \\[0.3ex]
$H_{24} \, \mathcal{T}$ & $\mathcal{F} \, \bar H_5$ & $\mathbf{10}$ & $6$ \\
\bottomrule
\end{tabular}
}
\end{floatrow}
\end{figure}
In Figure~\ref{fig:messengerdiagram1} and Table~\ref{tab:effectiveFToperators11},
we briefly review the topology of the diagram and field combinations that lead
to the CG factors $(Y_e)_{ji} / (Y_d)_{ij} = -\tfrac{1}{2}$, $-\tfrac{3}{2}$ and $6$ for
SU(5) GUTs, which have been shown to be useful for flavour model building
\cite{Antusch:2009gu, Antusch:2011qg,Marzocca:2011dh}.
To generate these relations we use a GUT breaking Higgs field $H_{24}$
in the adjoint representation of SU(5), and Higgs fields $\bar H_{5}$ and
$\bar H_{45}$ in the $\mathbf{\bar{5}}$ and $\mathbf{\overline{45}}$
representation. The factor $-\tfrac{1}{2}$ is, for instance, generated by the
coupling of $H_{24}$ and $\mathcal{F}$ to a messenger field transforming as $\mathbf{45}$,
while its partner $\mathbf{\overline{45}}$ couples to $\mathcal{T}$ and $\bar H_{45}$.
In our models in section \ref{sec:yukawas} the same CG factor $-\tfrac{1}{2}$ is
obtained from a dimension six operator where $\bar{H}_{45}$ acts as heavy
messenger field coupling to $H_{24}$ and $\bar{H}_5$.
So far we have discussed CG factors between the MSSM Yukawa couplings. For an
analysis of proton decay, the CG factors to the Yukawa couplings of the colour
Higgs triplets also play an important role. In Appendix~\ref{app:Clebsch} we present
an extensive discussion of CG factors in SU(5), including those.
\subsection{The double missing partner mechanism with an adjoint}
\label{subsec:dmpm24}
As we discussed in the previous subsection, the CG factors we want to combine
with the DMPM require the GUT breaking Higgs field to be in the adjoint representation
of SU(5), $\mathbf{24}$. This motivates us to replace the Higgs field $H_{75}$
needed for the DMPM with the effective combination $H^2_{24}/\Lambda$ \cite{Berezhiani:1996nu},
which at the renormalizable level can be obtained by integrating out heavy messenger
fields in the $\mathbf{45}$ and $\mathbf{\overline{45}}$ representations of SU(5) \cite{Zhang:2012rc}. To replace the $H_{75}$ in the MPM, we have to introduce
a set of messenger fields $X_{45}$, ${\bar X}_{45}$, $Y_{45}$ and
${\bar Y}_{45}$. For the DMPM, we also need to add a second set
$X_{45}^\prime$, ${\bar X}^\prime_{45}$, $Y_{45}^\prime$ and
${\bar Y}_{45}^\prime$. In Figure~\ref{Fig:Heff} we show the supergraphs generating
the non-diagonal entries of the triplet mass matrix in the DMPM with an adjoint $H_{24}$.
\begin{figure}
\center
\includegraphics[scale=0.75,page=1]{DMPM24.pdf}
\includegraphics[scale=0.75,page=2]{DMPM24.pdf}
\caption{Supergraphs generating the non-diagonal entries of the triplet mass matrix.
\label{Fig:Heff}}
\end{figure}
In analogy with the discussion in section \ref{PSO},
we avoid direct mass terms in order to forbid dangerous Planck-scale suppressed operators that would generate universal
mass contributions for Higgs doublets and triplets. The messenger pairs $X_{45}\bar X_{45}$,
$Y_{45}\bar Y_{45}$, $Z_{50}\bar Z_{50}$ and
their corresponding primed versions obtain masses from the VEV of a singlet
field $S$, charged under an additional shaping symmetry.
We specify these symmetries for two example models in section \ref{sec:yukawas}.
One may wonder if some of these heavy 45-dimensional messengers could be the
same, so that the number of fields in the spectrum would be reduced while preserving the structure of the mechanism.
But if either $X_{45} \equiv Y_{45}$ or
$X^\prime_{45} \equiv Y^\prime_{45}$, it can be seen from Figure~\ref{Fig:Heff} that supergraphs
without the $Z_{50}\bar Z_{50}$ mass insertion would be allowed, spoiling the splitting of
doublets and triplets and generating large non-diagonal entries in $m_D$.
In turn, if either $X_{45} \equiv X^\prime_{45}$, $Y_{45} \equiv Y^\prime_{45}$ or $Z_{50} \equiv Z_{50}^\prime$, the DMPM is reduced to the MPM, reintroducing the issue of perturbativity.
Finally, an identification of $X_{45} \equiv Y^\prime_{45}$ would allow
diagrams bypassing the 50-dimensional fields, generating unwanted mass term for both doublet and triplet components of $H_5$, $\bar H_5$, and thus a too large $\mu$-term.
With this messenger superfield content, we carefully checked that no dangerous Planck-scale suppressed operators spoil the mechanism. The renormalizable superpotential is:
\begin{align}
W_{\text{DMPM24}} & = \bar H_5 H_{24}X_{45} + \bar X_{45}H_{24}Z_{50} + \bar Z_{50}H_{24}Y_{45}+\bar Y_{45}H_{24}H_5^\prime \nonumber \\
& + \bar H_5^\prime H_{24}X_{45}^\prime + \bar X_{45}^\prime H_{24}Z_{50}^\prime + \bar Z_{50}^\prime H_{24}Y_{45}^\prime+\bar Y_{45}^\prime H_{24}H_5 \nonumber \\
& + SX_{45}\bar X_{45} + SY_{45}\bar Y_{45} + SZ_{50}\bar Z_{50}
+ SX_{45}^\prime\bar X_{45}^\prime + SY_{45}^\prime\bar Y_{45}^\prime + SZ_{50}^\prime\bar Z_{50}^\prime \nonumber \\
& + \mu^\prime H_5^\prime \bar H_5^\prime\;.
\end{align}
After $H_{24}$ and $S$ obtain their VEVs and integrating out the 45-dimensional messenger fields, we find the mass matrices for the Higgs doublets
and triplets to be
\begin{equation}
m_D = \begin{pmatrix}
0 & 0 \\
0 & \mu^\prime
\end{pmatrix}\;,\quad
\quad m_T = \begin{pmatrix}
0 & 0 & 0 & -\frac{V^2}{\vev{S}}\\
0 & \mu^\prime & -\frac{V^2}{\vev{S}} & 0 \\
-\frac{V^2}{\vev{S}} & 0 & \vev{S} & 0\\
0 & -\frac{V^2}{\vev{S}} & 0 & \vev{S}
\end{pmatrix}\;,
\end{equation}
which one can easily compare to the ones of eq.~(\ref{eq:M_DMPM}) in section \ref{subsec:dmpm24}. Here $V$ is defined by $\vev{H_{24}} \equiv V \text{diag}(1,1,1,-\frac{3}{2},-\frac{3}{2})$. Integrating out the heavy 50-dimensional fields in a next step, the mass matrices become
\begin{equation}
m_D = \begin{pmatrix}
0 & 0 \\
0 & \mu^\prime
\end{pmatrix}\;,\quad
m_T = \begin{pmatrix}
0 & -\frac{V^4}{ \vev{S}^3} \\
-\frac{V^4}{\vev{S}^3} & \mu^\prime
\end{pmatrix}\;.
\end{equation}
Thus, the doublets in the pair $H_5 {\bar H}_{5}$ stay massless, while the doublet pair in $H^\prime_5{\bar H}^\prime_{5}$ is heavy. Using only $H_5$ and ${\bar H}_{5}$ for the Yukawa couplings of the SM fermions, the effective triplet component mass relevant for dimension 5 proton decay is given by
\begin{equation}\label{eq:MTeff24_definition}
{M^\text{dim=5}_{T}} = \left(m_T^{-1}\right)_{11}^{-1} = -\frac{V^8}{\vev{S}^6 \mu^\prime} \;.
\end{equation}
The effective masses suppressing dimension six proton decay mediated by colour triplets are given by
\begin{equation}\label{eq:MTeff624_definition}
\left({M^\text{dim=6}_{T}}\right)^2 = \left({M^\text{dim=6}_{\bar T}}\right)^2 \approx\frac{\abs{V}^8}{\abs{\vev{S}}^6}\;,
\end{equation}
where we used the fact that $\abs{\vev{S}}\gg \abs{V}$ and $\abs{\vev{S}^3\mu'}\ll \abs{V}^4$. Now the requirement of ${M^\text{dim=6}_{T}} \gtrsim 10^{12}~{\rm GeV}$ can only be obtained with $\vev{S}\approx 10^{18}~{\rm GeV}$ if the GUT scale value is larger than $V\gtrsim 10^{16}~{\rm GeV}$. In this case one needs $\mu'\approx 10^7~{\rm GeV}$ to obtain an effective triplet mass ${M^\text{dim=5}_{T}}\approx 10^{17}~{\rm GeV}$ for dimension five proton decay operators.
Therefore, if the GUT scale is high enough, the effective triplet mass can be large enough to
stabilize the proton, while the large SU(5) representations used in the DMPM can be heavy
enough to keep the theory perturbative up to the Planck scale.
When $H_{24}$ is uncharged under additional symmetries, having $\mu^\prime$ several
orders of magnitude smaller
than $V$ requires $\mu^\prime$ to arise from the spontaneous breakdown of a shaping symmetry,
to avoid the term $\vev{H_{24}}H^\prime_5\bar H^\prime_5$, which would give rise to a
much too small effective triplet mass. Note that the effective triplet masses entering dimension
five proton decay can be expressed
in terms of the mass eigenstates of doublet and triplet components as
${M^\text{dim=5}_{T}} = - \tfrac{\tilde M_1\tilde M_2}{\mu^\prime}$. The effective triplet mass of dimension six proton decay is then excellently approximated by ${M^\text{dim=6}_{\bar T}} = {M^\text{dim=6}_{T}} \approx \sqrt{{M^\text{dim=5}_{T}} \mu^\prime}$.
\subsection{Introducing a second adjoint field \label{subsec:adjoints}}
We have seen that the DMPM with an adjoint GUT breaking Higgs instead of a
$\mathbf{75}$ already solves the DTS problem while providing the
necessary building block for the desirable CG factors for flavour model building, if the GUT scale is high enough. To
conclude this section, we will argue why it is compelling to further introduce a second adjoint Higgs field:
\begin{itemize}
\item
In the minimal SUSY SU(5) model \cite{Dimopoulos:1981zb}, the single GUT breaking
$\mathbf{24}$ contains an SU(2) triplet component and an SU(3) octet component with
equal masses. Demanding gauge coupling unification, the mass of the Higgs colour
triplets is required to be about $10^{15}$ GeV \cite{Murayama:2001ur}, ruling out this model due to proton
decay. Non-renormalizable operators in the GUT breaking
superpotential can split the $\mathbf{24}$ component masses,
allowing a higher effective triplet mass \cite{Bajc:2002pg} (see also section
\ref{sec:GUT}). An additional $\mathbf{24}$ can be used to realize this non-renormalizable superpotential
in a renormalizable way.
\item It turns out that the introduction of an additional $\mathbf{24}$ is not just a UV-completion of the non-renormalizable superpotential of \cite{Bajc:2002pg}. When both adjoints have approximately the same mass and therefore the second $\mathbf{24}$ is not integrated out, the additional colour octet and electroweak triplet in the spectrum lead to more freedom for the GUT scale and effective triplet mass. In the following section we discuss all possible renormalizable superpotentials with two adjoints and their impact on $M_{\text{GUT}}$ and ${M^\text{dim=5}_{T}}$ from a gauge coupling unification analysis.
\item A renormalizable superpotential for one $\mathbf{24}$ requires it to be uncharged under shaping symmetries in order for it to obtain a VEV.
However, such a shaping symmetry charge is vital in the type of flavour models considered here, to avoid unwanted admixtures of additional CG factors involving less insertions of $H_{24}$.
With a second $\mathbf{24}$, the adjoint fields can acquire non-vanishing VEVs even when charged under shaping symmetries.
\end{itemize}
These features of renormalizable superpotentials for two adjoints are presented in detail in the next section.
\section{Grand unification and the effective triplet mass \label{sec:GUT}}
In GUT extensions of the SM it is quite common to have additional fields below
the GUT scale that modify the RGE running, as it is the case for the class of models
in this paper. Therefore one has to study the impact of the additional fields on
the running of the gauge couplings and especially study their unification.
The modified unification condition for the gauge couplings at one-loop reads
\begin{equation}\label{eq:unification}
\frac{1}{\alpha_u} = \frac{1}{\alpha_i} - \frac{1}{2 \pi} \left(
b_i^{\text{(SM)}} \log \frac{M_{\text{SUSY}}}{M_Z} + b_i^{\text{(MSSM)}} \log \frac{M_{\text{GUT}}}{M_{\text{SUSY}}}
+ \sum\limits_f b_i^{(f)} \log \frac{M_{\text{GUT}}}{M_f} \right)
\;,
\end{equation}
where $i=1,2,3$ labels the SM gauge interaction and $f$ labels the additional
superfields (compared to the MSSM), with masses $M_f$ and $\beta$
coefficients $b_i^{(f)}$. The one-loop $\beta$-function coefficients
for the SM are $b_i^{\text{(SM)}} = (41/10,-19/6,-7)$ and for the MSSM
$b_i^{\text{(MSSM)}} = (33/5,1,-3)$. The SUSY scale $M_{\text{SUSY}}$ is defined here as the scale
where we make the transition from the SM $\beta$ coefficients to the MSSM ones.
The $\alpha_i$ are defined at
low energies $\alpha_i \equiv \alpha_i(M_Z)$ while $\alpha_u$ is the unified
gauge coupling at the GUT scale $\alpha_u \equiv \alpha_i(M_{\text{GUT}})$.
The GUT scale $M_{\text{GUT}}$ is defined here as the scale where the last SU(5)
multiplet is completed, in other words the scale where all three one-loop $\beta$
coefficients for the SM gauge couplings become equal.
From now on, we assume that the heaviest incomplete SU(5) multiplets to enter the
RGE running are the leptoquark vector bosons, such that the GUT scale corresponds
to their mass $M_{\text{GUT}} = M_V$. While other cases can certainly arise, we focus on
this option because it is quite common in our setup
to have heavy leptoquark vector bosons,
and furthermore we verified that in this case the triplet mass can be made
very heavy as well.
In addition to the MSSM field content, the DMPM introduces one
additional pair of SU(2)-doublets, $D^{(5)}$ and ${\bar D}^{(5)}$ and two additional
pairs of SU(3)-triplets, $T^{(5)}_i$ and ${\bar T}^{(5)}_i$, $i = 1$, 2. They enter
the $\beta$-functions with the coefficients $b^{(5,D)}_i = (3/5,1,0)$ for the Dirac
pair of doublets and $b^{(5,T)}_i = (2/5,0,1)$ per Dirac pair of triplet and
anti-triplet.
Furthermore, we use two SU(5) breaking GUT Higgs fields $H_{24}$ and
$H'_{24}$ in the adjoint representation. They contain one SM-singlet component each,
one SU(2)-triplet, $T^{(24)}$, with $b^{(24,T)}_i = (0,2,0)$,
an SU(3)-octet, $O^{(24)}$ with $b^{(24,O)}_i = (0,0,3)$ and a leptoquark
superfield pair, $L^{(24)}$ with $b^{(24,L)}_i = (5,3,2)$.
Since one leptoquark superfield pair is eaten up during the breaking of SU(5),
we are left with two triplets with masses $M_{T^{(24)}_1}$ and $M_{T^{(24)}_2}$,
two octets with masses $M_{O^{(24)}_1}$ and $M_{O^{(24)}_2}$ and one
leptoquark superfield pair with mass $M_{L^{(24)}}$.
For convenience we define the geometric means of the masses $M_{T^{(5)}}^2 = M_{T^{(5)}_1} M_{T^{(5)}_2}$
for the colour triplets and analogously $M_{T^{(24)}}^2 = M_{T^{(24)}_1} M_{T^{(24)}_2}$,
$M_{O^{(24)}}^2 = M_{O^{(24)}_1} M_{O^{(24)}_2}$ for the components of $H_{24}$ and $H'_{24}$.
Having this at hand, we can solve eq.~\eqref{eq:unification} for
$M_{D^{(5)}}$, $M_{T^{(5)}}$ and $M_{\text{GUT}}$, \footnote{In the following,
``$\log$'' of a mass is to be understood as the natural logarithm of the
mass divided by one common mass scale, e.g. $\log m \equiv \log(m/{\rm GeV})$.}
\begin{align}
\log M_{D^{(5)}} &= \frac{15 \pi }{4 \alpha_1}
- \frac{17 \pi }{4 \alpha_2}
- \frac{3 \pi }{2 \alpha_3}
+ \frac{59}{3} \log M_Z \\\nonumber
&+ \frac{2 \pi}{\alpha_u}
+ \frac{3}{2} \log M_{L^{(24)}}
- \frac{17}{2} \log M_{T^{(24)}}
- \frac{9}{2} \log M_{O^{(24)}}
- \frac{43}{6} \log M_{\text{SUSY}} \;,\\
\log M_{T^{(5)}} &= \frac{35 \pi}{24 \alpha_1}
- \frac{7 \pi}{8 \alpha_2}
- \frac{19 \pi }{12 \alpha_3}
+ \frac{119}{12} \log M_Z \\\nonumber
&+ \frac{\pi}{\alpha_u}
+ \frac{3}{4} \log M_{L^{(24)}}
- \frac{7}{4} \log M_{T^{(24)}}
- \frac{19}{4} \log M_{O^{(24)}}
- \frac{19}{6} \log M_{\text{SUSY}} \;,\\
\log M_{\text{GUT}} &= \frac{5 \pi }{12 \alpha_1}
- \frac{\pi }{4 \alpha_2}
- \frac{\pi }{6 \alpha_3}
+ \frac{11}{6} \log M_Z \\\nonumber
&+ \frac{1}{2} \log M_{L^{(24)}}
- \frac{1}{2} \log M_{T^{(24)}}
- \frac{1}{2} \log M_{O^{(24)}}
- \frac{1}{3} \log M_{\text{SUSY}} \;.
\end{align}
For the study of proton decay, it is more convenient to instead
solve eq.~\eqref{eq:unification} for the GUT scale gauge
coupling $\alpha_u$ and the effective triplet mass
${M^\text{dim=5}_{T}} = M_{T^{(5)}}^2/M_{D^{(5)}}$, which gives
the suppression of the dimension 5 proton decay operators
(cf.~the discussion in section \ref{subsec:dmpm24}).
Then we get the relations
\begin{align}
\frac{\pi}{\alpha_u} &= -\frac{43 \pi}{24 \alpha_1}
+ \frac{15 \pi}{8 \alpha_2}
+ \frac{11 \pi}{12 \alpha_3}
- \frac{197}{20} \log M_Z
+ \frac{3}{5} \log M_{DT} \\\nonumber
&- \frac{3}{4} \log M_{L^{(24)}}
+ \frac{15}{4} \log M_{T^{(24)}}
+ \frac{11}{4} \log M_{O^{(24)}}
+ \frac{7}{2} \log M_{\text{SUSY}} \;,\\
\log {M^\text{dim=5}_{T}} &= -\frac{5 \pi}{6 \alpha_1}
+ \frac{5\pi}{2 \alpha_2}
- \frac{5\pi}{3 \alpha_3}
+ \frac{1}{6} \log M_Z \\\nonumber
&+ 5 \log M_{T^{(24)}}
- 5 \log M_{O^{(24)}}
+ \frac{5}{6} \log M_{\text{SUSY}} \;,\\
\log M_{\text{GUT}} &= \frac{5 \pi}{12 \alpha_1}
- \frac{\pi}{4 \alpha_2}
- \frac{\pi }{6 \alpha_3}
+ \frac{11}{6} \log M_Z \\\nonumber
&+ \frac{1}{2} \log M_{L^{(24)}}
- \frac{1}{2} \log M_{T^{(24)}}
- \frac{1}{2} \log M_{O^{(24)}}
- \frac{1}{3} \log M_{\text{SUSY}} \;,
\end{align}
where we have introduced the mass $M_{DT}^3 = M_{D^{(5)}}^2 M_{T^{(5)}}$.
As one can see only $\alpha_u$ depends on $M_{DT}$, which is due to the fact that
doublets and colour triplets together form a complete representation of SU(5).
Thus, following eq.~\eqref{eq:unification}, one can see that a simultaneous rescaling
$M_{D^{(5)}} \to q^2 M_{D^{(5)}}$ and $M_{T^{(5)}} \to q M_{T^{(5)}}$ leaves the GUT
scale invariant and only shifts $\alpha_u$, while ${M^\text{dim=5}_{T}} \propto q^0$ remains
unchanged and $M_{DT} \propto q$ parametrises this rescaling.
Further interdependencies between $\alpha_u$, ${M^\text{dim=5}_{T}}$ and $M_{\text{GUT}}$ are then
implicit via their shared dependence on the other masses.
Thus, unification implies that the effective triplet mass follows the relation
\begin{align}
{M^\text{dim=5}_{T}} &=
\exp\left(\frac{5}{6} \pi \left(\frac{3}{\alpha_2}-\frac{2}{\alpha_3}-\frac{1}{\alpha_1}\right)\right)
M_Z^{1/6} M_{\text{SUSY}}^{\frac 5 6}
\left(\frac{M_{T^{(24)}}}{M_{O^{(24)}}}\right)^5 \nonumber\\
&= \label{eq:MTeff}
2.5^{+0.6}_{-0.8} \cdot 10^{17} \; {\rm GeV}
\left(\frac{M_{\text{SUSY}}}{1 \, {\rm TeV}}\right)^{\frac 5 6}
\left(\frac{M_{T^{(24)}}}{M_{O^{(24)}}}\right)^5 \;,
\end{align}
while the GUT scale is given by
\begin{equation}\label{eq:MGUT}
M_{\text{GUT}} =
1.37^{+0.05}_{-0.05} \cdot 10^{16} \; {\rm GeV}
\left(\frac{M_{\text{SUSY}}}{1 \, {\rm TeV}}\right)^{-\frac 1 3}
\left(\frac{M_{L^{(24)}}}{10^{16}\, {\rm GeV}}\right)^{\frac 1 2}
\left(\frac{M_{T^{(24)}} M_{O^{(24)}}}{(10^{16}\, {\rm GeV})^2} \right)^{-\frac 1 2} \;.
\end{equation}
For completeness, the unified gauge coupling is given by
\begin{align}\label{eq:alphau}
\frac{1}{\alpha_u} &= 24.58 \pm 0.06
+ \frac{7}{2\pi} \ln \frac{M_{\text{SUSY}}}{1\, {\rm TeV}}
+ \frac{3}{5\pi} \ln \frac{M_{DT}}{10^{14} \, {\rm GeV}} \\\nonumber
&\quad- \frac{3}{4\pi} \ln \frac{M_{L^{(24)}}}{10^{16}\, {\rm GeV}}
+ \frac{15}{4\pi} \ln \frac{M_{T^{(24)}}}{10^{16}\, {\rm GeV}}
+ \frac{11}{4\pi} \ln \frac{M_{O^{(24)}}}{10^{16}\, {\rm GeV}} \;.
\end{align}
For these numbers, we have used the experimental values and uncertainties for
the gauge couplings found in \cite{pdg}. Note that for all three quantities
the resulting uncertainty is dominated by the experimental error on
$\alpha_s$.
In the following we will not quote any errors on the masses anymore
since the relative uncertainty changes only negligibly for the different
superpotentials and for two-loop running.
The reference scale $10^{14}$~GeV is chosen due to the fact
that $M_{DT} = 10^{14}$~GeV and $M_{T^{(5)}} = 10^{16}$~GeV
implies ${M^\text{dim=5}_{T}} = 10^{19}$~GeV.
Since the effective triplet mass ${M^\text{dim=5}_{T}}$ receives significant
two-loop contributions (cf., for instance, \cite{Murayama:2001ur}), we have also
implemented a numerical two-loop RGE analysis using the following procedure.
We start with SM values for the gauge and Yukawa couplings \cite{Antusch:2013jca}
at $M_Z$, run up to to a scale of 1~TeV with the full two-loop SM RGEs
and match the SM to
the MSSM (including $\overline{\text{MS}}$ to $\overline{\text{DR}}$ scheme conversion).
From there we run and match using full two-loop MSSM RGEs and one-loop gauge coupling
threshold corrections
\footnote{When one integrates out particles at a threshold
scale equal to their mass, these threshold corrections vanish, as can be seen in \cite{Hall:1980kf}.
}
while step-by-step including all additional multiplets at their mass scale via their
contributions to the one- and two-loop gauge coupling RGEs, see Appendix \ref{sec:2loopRGEs}.
The Yukawa couplings of the Higgs colour triplets are well approximated by using the
Yukawa couplings of the corresponding doublets.
We do not take into account any other Yukawa couplings. We verified this approximation
numerically and the results for the colour triplet masses are barely affected.
\subsection{Superpotentials with two adjoints of SU(5)}
\label{sec:W2Adjoints}
In this section, we systematically study all superpotentials with two adjoints
that can break SU(5) to the SM gauge group. We find only four possibilities with
non-vanishing VEVs and masses. Classified based on their symmetry, they are:
\begin{enumerate}
\item[(a)] $W = M_{24} \, \mathop{\rm tr} H_{24}^2 + M^\prime_{24} \mathop{\rm tr} H^{\prime 2}_{24} + \kappa^\prime \, \mathop{\rm tr} H_{24}H_{24}^{\prime 2} + \lambda \, \mathop{\rm tr} H^3_{24}$, \\
$\mathbb{Z}_2$ symmetry where $H_{24}$ is uncharged and $H_{24}^\prime$ charged.
\item[(b)] $W = \tilde{M}_{24} \, \mathop{\rm tr} H_{24} H^\prime_{24} + \lambda \, \mathop{\rm tr} H_{24}^3 + \lambda^\prime \, \mathop{\rm tr} H^{\prime 3}_{24}$, \\
$\mathbb{Z}_3$ symmetry, where $H_{24}$ has charge $2$ and $H'_{24}$ charge $1$.
\item[(c)] $W = \tilde{M}_{24} \, \mathop{\rm tr} H_{24} H^\prime_{24} + \lambda \, \mathop{\rm tr} H_{24}^3 + \kappa^\prime \, \mathop{\rm tr} H_{24} H^{\prime 2}_{24}$, \\
$\mathbb{Z}_4^R$ symmetry where $H_{24}$ has a charge of 2 (with $q_\theta = 1$) and $H'_{24}$ is uncharged.
\item[(d)] The trivial case with both fields only charged under SU(5) and all (non-linear) terms allowed.
We will not consider this case any further.
\end{enumerate}
Since we are dealing with two adjoint Higgs fields, it is convenient to define
a quantity $\tan\beta_V$ similar to $\tan \beta$ of the MSSM, so that
\begin{align}
\vev{H_{24}} &= V_1 \,\, \text{e}^{\text{i} \phi_1} \,\, \text{diag}(1,1,1,-3/2,-3/2) \;,\label{eq:vevdef1}\\
\vev{H'_{24}} &= V_2 \,\, \text{e}^{\text{i} \phi_2} \,\, \text{diag}(1,1,1,-3/2,-3/2) \;,\label{eq:vevdef2}
\end{align}
with $V_1, V_2 > 0$ and $\tan \beta_V = V_1 / V_2$.
\subsubsection{Superpotential (a)}\label{subsec:superpotentiala}
\begin{figure}
\begin{center}
\begin{subfigure}{0.45\textwidth}
\begin{center}
\subcaption*{${M^\text{dim=5}_{T}}$}
\includegraphics[width=1\textwidth]{MTeff_a_at1loop.pdf}\\
\includegraphics[width=1\textwidth]{MTeff_a_at2loop.pdf}
\end{center}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\begin{center}
\subcaption*{$M_{\text{GUT}}$}
\includegraphics[width=1\textwidth]{MGUT_a_at1loop.pdf}\\
\includegraphics[width=1\textwidth]{MGUT_a_at2loop.pdf}
\end{center}
\end{subfigure}
\end{center}
\caption{
The effective colour triplet mass ${M^\text{dim=5}_{T}}$ (left)
and GUT scale $M_{\text{GUT}}$ (right) in GeV at one-loop (upper)
and two-loop (lower) order as resulting from superpotential (a)
for $M_{\text{SUSY}} = 1$~TeV, $M = 10^{15}$~GeV,
$M_{D^{(5)}}=1000$~TeV and $\bar{\phi} = 0$.
Note the different colour coding between left and right.
For illustration, the white strips denote areas with light $M_{T^{(24)}}$
or $M_{O^{(24)}}$ ($<10^{13}$~GeV). Such relatively low values for these
components can arise either from cancellation between terms, or from a
generic suppression due to small parameters, cf.\ eqs.~\eqref{eq:24comps}.
}\label{fig:MTeff+MGUT}
\end{figure}
We will begin our discussion with superpotential (a) which turns out to
be the most complicated case since it has the most parameters.
As it contains two mass parameters, we introduce
a second angle $\beta_M$ and mean mass $M > 0$ such that
$M_{24} = M \text{e}^{\text{i} \alpha_1} \sin \beta_M$
and $M'_{24} = M \text{e}^{\text{i} \alpha_2} \cos \beta_M$. The vacuum expectation
values are given by
\begin{equation}
V_1 \text{e}^{\text{i} \phi_1}= \frac{4 M'_{24}}{\kappa^\prime} \text{\quad and\quad}
V_2^2 \text{e}^{2 \text{i} \phi_2}= 4 \, M'_{24} \frac{2 M_{24} \kappa^\prime - 3 M'_{24} \lambda}{\kappa^{\prime 3}} \;,
\end{equation}
which can as well be expressed in terms of a ratio of the coupling constants involved,
$3 \lambda / \kappa^\prime = 2 \text{e}^{\text{i} (\alpha_1 - \alpha_2)} \tan\beta_M - \text{e}^{-2 \text{i} (\phi_1 - \phi_2)} \cot^2\beta_V$.
For the geometric means of the masses of the additional fields compared to the MSSM, we find
\begin{subequations}
\label{eq:24comps}
\begin{align}
M_{T^{(24)}}^2 &= 5 \, M^2 \, \cos \beta_M \,
\sqrt{(2 \cos \beta_M - 3 \sin \beta_M \tan^2 \beta_V)^2 + \Delta} \;,\\
M_{O^{(24)}}^2 &= 5 \, M^2 \, \cos \beta_M \,
\sqrt{(3 \cos \beta_M - 2 \sin \beta_M \tan^2 \beta_V)^2 + \Delta} \;, \\
M_{L^{(24)}}^2 &= \frac{1}{4} \, M^2 \, \frac{\cos^2 \beta_M}{\sin^4 \beta_V} \;,
\end{align}
\end{subequations}
with $\Delta = 12 \, \sin(2 \beta_M) \, \cot^2(\beta_V) \, \sin^2\bar{\phi}$
and $\bar{\phi} = (\alpha_1-\alpha_2)/2 + \phi_1 - \phi_2$. Note that not only
the geometric mean masses, but also the mass eigenvalues themselves only depend
on this phase combination $\bar{\phi}$ and are invariant under
$\bar{\phi} \to \bar{\phi} + \pi$.
The effective triplet mass as of eq.~\eqref{eq:MTeff} is heaviest if the phase
$\bar{\phi}$ is $0$, $\pi$ or $2\pi$, since then the ratio
$M_{T^{(24)}} / M_{O^{(24)}}$ is not bounded from above (or below), which allows
for the maximal range for ${M^\text{dim=5}_{T}}$. Thus, in the following, we choose
$\bar{\phi}$ and thus $\Delta$ to vanish.
The resulting plots for ${M^\text{dim=5}_{T}}$ and for $M_{\text{GUT}}$ are shown
in Fig.~\ref{fig:MTeff+MGUT} for $M_{\text{SUSY}} = 1$~TeV, $M= 10^{15}$~GeV,
$M_{D^{(5)}}=1000$~TeV and $\bar{\phi} = 0$,
including a comparison between one- and two-loop results.
Note however, that gauge coupling unification depends only weakly on $M_{D^{(5)}}$.
${M^\text{dim=6}_{T}}$ and ${M^\text{dim=6}_{\bar T}}$ are again approximately given by $\sqrt{{M^\text{dim=5}_{T}} M_{D^{(5)}}}$.
\subsubsection{Superpotential (b) and (c)}
\label{sec:Superpotentialbc}
These two superpotentials have only one massive parameter $\tilde{M}_{24} = M \text{e}^{\text{i} \alpha}$
and hence the analytic results become much less cumbersome.
The vacuum solutions are given by
\begin{equation}
V_1 \text{e}^{\text{i} \phi_1}= \frac{2}{3} \frac{M \text{e}^{\text{i} \alpha}}{\sqrt[3]{\lambda^2 \lambda'}} \text{\quad and\quad}
V_2 \text{e}^{\text{i} \phi_2}= \frac{2}{3} \frac{M \text{e}^{\text{i} \alpha}}{\sqrt[3]{\lambda \lambda'^2}} \;,
\end{equation}
up to a $\mathbb{Z}_3$ symmetry transformation for superpotential (b) and
\begin{equation}
V_1 \text{e}^{\text{i} \phi_1}= \frac{1}{\sqrt{3}} \frac{M \text{e}^{\text{i} \alpha}}{\sqrt{\lambda \kappa'}} \text{\quad and\quad}
V_2 \text{e}^{\text{i} \phi_2}= \frac{M \text{e}^{\text{i} \alpha}}{\kappa'} \;,
\end{equation}
up to a minus sign in $V_1$ for superpotential (c). In other words, the couplings
fulfill the relation $\lambda' / \lambda = \text{e}^{3 \text{i} (\phi_2 - \phi_1)}\tan^3\beta_V$
for superpotential (b) and $\kappa'/ \lambda = 3 \text{e}^{2 \text{i} (\phi_1 - \phi_2)}\tan^2\beta_V$
for superpotential (c).
For the geometric means of the masses of the colour triplets,
octets and the mass of the left-over leptoquark superfield in
$H_{24}$ and $H'_{24}$, we find
\begin{equation}
M_{T^{(24)}}^2 = \frac{35}{4} \, M^2 \;,\;\;
M_{O^{(24)}}^2 = \frac{15}{4} \, M^2 \text{ and }
M_{L^{(24)}}^2 = \frac{1}{\sin^2 (2 \beta_V)} \, M^2 \;,
\end{equation}
for superpotential (b) and
\begin{equation}
M_{T^{(24)}}^2 = \frac{5}{4} \, M^2 \;,\;\;
M_{O^{(24)}}^2 = \frac{5}{4} \, M^2 \text{ and }
M_{L^{(24)}}^2 = \frac{1}{4 \sin^2 (2 \beta_V)} \, M^2 \;,
\end{equation}
for superpotential (c).
Note that in both cases also all mass eigenvalues turn out to be phase independent.
Therefore, applying eq.~\eqref{eq:MTeff} unification of
the gauge couplings on one-loop implies
\begin{equation} \label{eq:MTeffsup_bc}
{M^\text{dim=5, 1-loop}_{T}} = 2.5 \cdot 10^{17} \; {\rm GeV} \left(\frac{M_{\text{SUSY}}}{1 \, {\rm TeV}}\right)^{\frac 5 6}
\times
\begin{cases}
\left(\frac 7 3\right)^{\frac 5 2} \approx 8.3 & \text{(b)} \\
\quad1 & \text{(c)}
\end{cases} \;,
\end{equation}
and
\begin{equation}\label{eq:MGUTsup_bc}
M_{\text{GUT}}^{\text{1-loop}} = \frac{1.37 \cdot 10^{16} \; {\rm GeV}}{\sqrt{|\sin 2\beta_V|}}
\left(\frac{M_{\text{SUSY}}}{1 \, {\rm TeV}}\right)^{-\frac 1 3}
\left(\frac{M}{10^{15} \, {\rm GeV}}\right)^{-\frac 1 2}
\times
\begin{cases}
\sqrt{\frac{8}{\sqrt{21}}} \approx 1.32 & \text{(b)} \\
2 & \text{(c)}
\end{cases}
\;.
\end{equation}
Assuming the same parameters, we find an almost
ten times heavier effective triplet mass in superpotential (b) than in (c) and hence we
focus on this case in our second example model later.
At two-loop, we find the following approximate behaviour for the masses
\begin{equation} \label{eq:MTeffsup_bc_2loop}
{M^\text{dim=5, 2-loop}_{T}} = \left(\frac{M_{\text{SUSY}}}{1 \, {\rm TeV}}\right)^{0.74}
\cdot
\begin{cases}
5.2 \cdot 10^{16} \; {\rm GeV} \left(\frac{M}{10^{15} \, {\rm GeV}}\right)^{-0.15} & \text{(b)} \\
6.8 \cdot 10^{15} \; {\rm GeV} \left(\frac{M}{10^{15} \, {\rm GeV}}\right)^{-0.18} & \text{(c)}
\end{cases} \;,
\end{equation}
and
\begin{equation}\label{eq:MGUTsup_bc_2loop}
M_{\text{GUT}}^{\text{2-loop}} = |\sin 2\beta_V|^{-0.48}
\left(\frac{M_{\text{SUSY}}}{1 \, {\rm TeV}}\right)^{-0.4}
\cdot
\begin{cases}
2.89 \cdot 10^{16} \; {\rm GeV} \left(\frac{M}{10^{15} \, {\rm GeV}}\right)^{-0.61} & \text{(b)} \\
4.78 \cdot 10^{16} \; {\rm GeV} \left(\frac{M}{10^{15} \, {\rm GeV}}\right)^{-0.63} & \text{(c)}
\end{cases}
\;.
\end{equation}
The dependence on other parameters is very small. As we can see, at two-loop, having ${M^\text{dim=5}_{T}} \gtrsim 10^{17}$ GeV requires $M_{\text{SUSY}} \gtrsim 2.3$~TeV and $35$~TeV for superpotential (b) and (c) respectively. Again, $M_{D^{(5)}}=1000$~TeV has been fixed and the values of ${M^\text{dim=6, 1-loop}_{T}}$ and ${M^\text{dim=6, 2-loop}_{T}}$ can be approximated by the square root of the product of $M_{D^{(5)}}$ and ${M^\text{dim=5, 1-loop}_{T}}$ or ${M^\text{dim=5, 2-loop}_{T}}$ respectively.
There are a few more comments in order. There is a claim \cite{Fallbacher:2011xg}
that the MSSM with an additional unbroken R-symmetry can not be obtained from
the spontaneous breaking of a four-dimensional (SUSY) GUT. Note that this is not in
conflict with our superpotentials, because the R-symmetry is either absent (a, b, d)
or spontaneously broken at the GUT scale (c). Superpotential (c) is particularly
interesting for model building purposes because R-symmetries are very popular in
flavour models with non-Abelian family symmetries (and spontaneous CP violation).
We will discuss this in more detail in appendix~\ref{app:SR}.
\section{Flavour Models with DMPM}
\label{sec:yukawas}
In this section we combine the DMPM (featuring two adjoints of SU(5))
with a predictive GUT flavour model for the quark-lepton Yukawa ratios at the GUT scale.
In particular, we implement
CG factors as given in \cite{Antusch:2009gu}. Two examples with different
Yukawa matrix structures are presented: in the
first model we construct diagonal down-type quark and charged lepton
Yukawa matrices $Y_d$ and $Y_e$, with all mixing originating from the up-type
quark Yukawa matrix $Y_u$. The second model realizes the attractive feature of
the Cabibbo mixing angle $\theta_C$ originating from $Y_d$. Both models
are providing existence proofs that successful DTS and experimentally viable predictions
for the GUT scale Yukawa coupling ratios can
indeed be realised simultaneously in one model.
Let us be more specific on the predictions made by the two models:
Due to the CG factors in the down-type quark and charged
lepton sector, $\tfrac{y_\tau}{y_b}$, $\tfrac{y_\mu}{y_s}$ and
$\tfrac{y_e}{y_d}$ are predicted at the GUT scale. To confront them with the
experimental data, the RG running to low energies has to be performed,
including in particular supersymmetric 1-loop threshold corrections \cite{SUSYthresholds}
when the MSSM is matched to the SM. These threshold corrections can have
a sizeable impact on the low energy values of the Yukawa couplings (and thus the fermion masses),
depending on the sparticle spectrum and $\tan \beta$.
So the predictions here are two-fold: Firstly, the predictions for the Yukawa ratios at the GUT scale imply
constraints on the SUSY breaking parameters, which may be tested at future collider searches if
SUSY is found\footnote{To make explicit statements about the constraints on the SUSY parameters one would have to specify the model of SUSY breaking, which is beyond the scope of this paper. A discussion and an explicit example where such constraints are worked out can be found, e.g.\ in \cite{Antusch:2011sq}.}. Secondly, the ratios $\tfrac{y_d}{y_s}$ and $\tfrac{y_e}{y_\mu}$
are not affected by RG running and by the SUSY threshold corrections (as long as the first two families
of sfermions are almost degenerate in mass as commonly assumed). They can be directly used to
constrain GUT models. A particularly useful quantity in this context is indeed the
double ratio
\begin{equation}
\left| \frac{y_\mu}{y_s}\frac{y_d}{y_e} \right| = 10.7^{+1.8}_{-0.8}\;,
\end{equation}
which can be checked directly at the GUT scale \cite{Antusch:2013jca}.
In our models we will use the CG factors $\tfrac{y_\tau}{y_b} = - \tfrac{3}{2}$,
$\tfrac{y_\mu}{y_s} = 6$ and $\tfrac{y_e}{y_d} = - \tfrac{1}{2}$. This leads to
$\left| \tfrac{y_\mu}{y_s} \tfrac{y_d}{y_e} \right| =12$, which is in good agreement with the
phenomenological value. On the other hand, the ubiquitous CG factors
$y_\mu = - 3 y_s$ and $y_e = \frac{1}{3} y_d$, known as Georgi-Jarlskog relations
\cite{GJ}, would give $\left| \tfrac{y_\mu}{y_s} \tfrac{y_d}{y_e} \right|=9$ and deviate
from the central value $10.7$ by more than two sigma.
In this section we explicitly construct only the Yukawa matrices of the up-
and down-type quarks and charged leptons. Adding one of the ubiquitous mechanisms
to generate neutrino masses and lepton mixing angles would be straightforward.
However, we do not consider neutrinos in this paper, since they are not directly relevant
for the discussion of proton decay, doublet-triplet splitting and the CG factors
between $Y_d$ and $Y_e$.
\subsection[A model with diagonal $Y_d$ and $Y_e$ Yukawa matrices]
{A model with diagonal $\boldsymbol{Y_d}$ and $\boldsymbol{Y_e}$ Yukawa matrices\label{sec:modelA}}
We now turn to our first model featuring diagonal down-type quark and charged
lepton Yukawa matrices $Y_d$ and $Y_e$.\footnote{The matrices are diagonal in the
preferred basis where the different fermion generations have well defined symmetry
assignments, cf.\ Table~\ref{tab:A_MSSM}.} In this case all the mixing in the quark
sector has to come exclusively from the up-type quark Yukawa matrix $Y_u$. Explicitly,
we have the following structure for the Yukawa matrices
\begin{equation}
\label{eq:A_Y}
Y_d = \begin{pmatrix}
y_d & 0 & 0\\
0 & y_s & 0\\
0 & 0 & y_b
\end{pmatrix},\;
Y_e=\begin{pmatrix}
-\frac{1}{2}y_d & 0& 0\\
0 & 6 y_s & 0\\
0 & 0 & -\frac{3}{2}y_b
\end{pmatrix},\;
Y_u =\begin{pmatrix}
y_{11} & y_{12} & y_{13}\\
y_{12} & y_{22} & y_{23}\\
y_{13} & y_{23} & y_{33}
\end{pmatrix}\;.
\end{equation}
An approach to flavour (GUT) model building with diagonal $Y_e$ (and $Y_d$) has been
discussed recently in \cite{King:2013hoa}.
We introduce flavon fields $\theta_1$, $\theta_2$, $\theta_3$ and $\theta_4$ that obtain a VEV and generate
the hierarchical structure of the Yukawa matrices. After the flavon fields, $H_{24}$
and $H_{24}^\prime$ obtain their VEVs, the Yukawa matrices of eq.~\eqref{eq:A_Y} originate
from the following effective superpotentials
\begin{align}
W_u &= \frac{1}{\Lambda^4} H_5 {\cal T}_1 {\cal T}_1 \theta_1^2\theta_2^2
+ \frac{1}{\Lambda^3} H_5 {\cal T}_1 {\cal T}_2 \theta_1^2 \theta_2
+ \frac{1}{\Lambda^2} H_5 {\cal T}_1 {\cal T}_3 \theta_1 \theta_2
+ \frac{1}{\Lambda^2} H_5 {\cal T}_2 {\cal T}_2 \theta_1^2 \nonumber\\
&\phantom{=}+ \frac{1}{\Lambda} H_5 {\cal T}_2 {\cal T}_3 \theta_1
+ H_5 {\cal T}_3 {\cal T}_3\;,\label{eq:A_Wu}\\
W_d &= \frac{1}{\vev{S^\prime}} (H_{24}^\prime {\mathcal{F}}_3)_{\bar{5}} ({\bar H}_{5} {\mathcal{T}}_3)_{5}
+ \frac{\theta_3}{\vev{S^\prime}^2} (H_{24}^\prime {\mathcal{T}}_2)_{10} ( {\bar H}_{5} {\mathcal{F}}_2)_{\overline{10}} \nonumber\\
&\phantom{=}+ \frac{\theta_4}{\vev{S^\prime}^2 \vev{S}} (H_{24}^\prime {\mathcal{F}}_1)_{\overline{45}} ( {\mathcal{T}}_1 H_{24} {\bar H}_{5})_{45}\;,\label{eq:A_Wd}
\end{align}
where we do not show order one coefficients, and denote the different
messenger masses generating $W_u$ by a generic $\Lambda$. However, keep
in mind that this is just for the sake of simplicity and different entries
in the Yukawa matrix should be understood as independent parameters.
The ratios of flavon VEVs and messenger masses is small of about $0.01 - 0.1$.
For a list of all fields including their
charges under the additional discrete shaping symmetries, see Tables~\ref{tab:A_MSSM},
\ref{tab:A_VEV}, \ref{tab:A_DMPM}, and \ref{tab:A_messengers}.
\begin{table}
\begin{tabular}{cccccccccc}
\toprule
& SU(5) & $\mathbb{Z}_{2}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{9}$ & $\mathbb{Z}_{2}$ \\
\midrule
$H_5$ & $\mathbf{5} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ \\
$\bar H_{5}$ & $\mathbf{\bar 5} $ & $.$ & $2$ & $.$ & $.$ & $.$ & $1$ & $2$ & $.$ \\
$\mathcal{T}_1$ & $\mathbf{10} $ & $.$ & $.$ & $3$ & $.$ & $6$ & $.$ & $.$ & $1$ \\
$\mathcal{T}_2$ & $\mathbf{10} $ & $.$ & $.$ & $.$ & $.$ & $6$ & $.$ & $.$ & $1$ \\
$\mathcal{T}_3$ & $\mathbf{10} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $1$ \\
$\mathcal{F}_1$ & $\mathbf{\bar 5} $ & $1$ & $1$ & $1$ & $2$ & $1$ & $1$ & $2$ & $1$ \\
$\mathcal{F}_2$ & $\mathbf{\bar 5} $ & $1$ & $.$ & $.$ & $2$ & $1$ & $.$ & $2$ & $1$ \\
$\mathcal{F}_3$ & $\mathbf{\bar 5} $ & $1$ & $2$ & $.$ & $1$ & $.$ & $6$ & $7$ & $1$ \\
\bottomrule
\end{tabular}
\caption{
SU(5) representations and charges under discrete shaping symmetries
of the MSSM fields and colour triplets of the model presented in subsection \ref{sec:modelA}. A dot denotes charge zero.
\label{tab:A_MSSM}
}
\end{table}
\begin{table}
\centering
\begin{tabular}{cccccccccc}
\toprule
& SU(5) & $\mathbb{Z}_{2}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{9}$ & $\mathbb{Z}_{2}$ \\
\midrule
$H_{24}$ & $\mathbf{24} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ \\
$H_{24}^\prime$ & $\mathbf{24} $ & $1$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ \\
$S$ & $\mathbf{1} $ & $.$ & $3$ & $.$ & $.$ & $.$ & $2$ & $.$ & $.$ \\
$S^\prime$ & $\mathbf{1} $ & $.$ & $.$ & $.$ & $1$ & $.$ & $.$ & $.$ & $.$ \\
$\theta_1$ & $\mathbf{1} $ & $.$ & $.$ & $.$ & $.$ & $1$ & $.$ & $.$ & $.$ \\
$\theta_2$ & $\mathbf{1} $ & $.$ & $.$ & $1$ & $.$ & $.$ & $.$ & $.$ & $.$ \\
$\theta_{3}$ & $\mathbf{1} $ & $.$ & $2$ & $.$ & $.$ & $.$ & $6$ & $5$ & $.$ \\
$\theta_{4}$ & $\mathbf{1} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $5$ & $.$ \\
\bottomrule
\end{tabular}
\caption{
SU(5) representations and charges under discrete shaping symmetries
of the superfields obtaining VEVs at around the GUT scale of the model presented in subsection \ref{sec:modelA}. A dot denotes charge zero.
\label{tab:A_VEV}
}
\end{table}
\begin{table}
\begin{tabular}{cccccccccc}
\toprule
& SU(5) & $\mathbb{Z}_{2}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{9}$ & $\mathbb{Z}_{2}$ \\
\midrule
$H_{5}^\prime$ & $\mathbf{5} $ & $.$ & $3$ & $.$ & $.$ & $.$ & $5$ & $7$ & $.$ \\
$\bar H_{5}^\prime$ & $\mathbf{\bar 5} $ & $.$ & $1$ & $.$ & $.$ & $.$ & $6$ & $.$ & $.$ \\
$X_{45}$ & $\mathbf{45} $ & $.$ & $2$ & $.$ & $.$ & $.$ & $6$ & $7$ & $.$ \\
$\bar X_{45}$ & $\mathbf{\overline{45}} $ & $.$ & $3$ & $.$ & $.$ & $.$ & $6$ & $2$ & $.$ \\
$Y_{45}$ & $\mathbf{45} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $3$ & $7$ & $.$ \\
$\bar Y_{45}$ & $\mathbf{\overline{45}} $ & $.$ & $1$ & $.$ & $.$ & $.$ & $2$ & $2$ & $.$ \\
$Z_{50}$ & $\mathbf{50} $ & $.$ & $1$ & $.$ & $.$ & $.$ & $1$ & $7$ & $.$ \\
$\bar Z_{50}$ & $\mathbf{\overline{50}} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $4$ & $2$ & $.$ \\
$X_{45}^\prime$ & $\mathbf{45} $ & $.$ & $3$ & $.$ & $.$ & $.$ & $1$ & $.$ & $.$ \\
$\bar X_{45}^\prime$ & $\mathbf{\overline{45}} $ & $.$ & $2$ & $.$ & $.$ & $.$ & $4$ & $.$ & $.$ \\
$Y_{45}^\prime$ & $\mathbf{45} $ & $.$ & $1$ & $.$ & $.$ & $.$ & $5$ & $.$ & $.$ \\
$\bar Y_{45}^\prime$ & $\mathbf{\overline{45}} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ \\
$Z_{50}^\prime$ & $\mathbf{50} $ & $.$ & $2$ & $.$ & $.$ & $.$ & $3$ & $.$ & $.$ \\
$\bar Z_{50}^\prime$ & $\mathbf{\overline{50}} $ & $.$ & $3$ & $.$ & $.$ & $.$ & $2$ & $.$ & $.$ \\
\bottomrule
\end{tabular}
\caption{
SU(5) representations and charges under discrete shaping symmetries
of the fields in the DMPM sector of the model presented in subsection \ref{sec:modelA}. A dot denotes charge zero.
\label{tab:A_DMPM}
}
\end{table}
\begin{table}
\centering
\begin{tabular}{cccccccccc}
\toprule
& SU(5) & $\mathbb{Z}_{2}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{9}$ & $\mathbb{Z}_{2}$ \\
\midrule
$Z_{5,1}$ & $\mathbf{5} $ & $.$ & $2$ & $.$ & $3$ & $.$ & $1$ & $2$ & $1$ \\
$\bar{Z}_{5,1}$ & $\mathbf{\bar 5} $ & $.$ & $2$ & $.$ & $.$ & $.$ & $6$ & $7$ & $1$ \\
$Z_{10,1}$ & $\mathbf{10} $ & $1$ & $.$ & $.$ & $3$ & $6$ & $.$ & $.$ & $1$ \\
$\bar{Z}_{10,1}$ & $\mathbf{\overline{10}} $ & $1$ & $.$ & $.$ & $.$ & $1$ & $.$ & $.$ & $1$ \\
$Z_{10,2}$ & $\mathbf{10} $ & $1$ & $2$ & $.$ & $2$ & $6$ & $6$ & $5$ & $1$ \\
$\bar{Z}_{10,2}$ & $\mathbf{\overline{10}} $ & $1$ & $2$ & $.$ & $1$ & $1$ & $1$ & $4$ & $1$ \\
$Z_{45,1}$ & $\mathbf{45} $ & $.$ & $3$ & $3$ & $2$ & $6$ & $6$ & $7$ & $1$ \\
$\bar{Z}_{45,1}$ & $\mathbf{\overline{45}} $ & $.$ & $1$ & $1$ & $1$ & $1$ & $1$ & $2$ & $1$ \\
$Z_{45,2}$ & $\mathbf{45} $ & $.$ & $3$ & $3$ & $3$ & $6$ & $6$ & $2$ & $1$ \\
$\bar{Z}_{45,2}$ & $\mathbf{\overline{45}} $ & $.$ & $1$ & $1$ & $.$ & $1$ & $1$ & $7$ & $1$ \\
\midrule
$Z_{10,3}$ & $\mathbf{10} $ & $.$ & $.$ & $.$ & $.$ & $1$ & $.$ & $.$ & $1$ \\
$Z_{10,4}$ & $\mathbf{10} $ & $.$ & $.$ & $1$ & $.$ & $1$ & $.$ & $.$ & $1$ \\
$Z_{1}$ & $\mathbf{1} $ & $.$ & $.$ & $.$ & $.$ & $5$ & $.$ & $.$ & $.$ \\
$Z_{2}$ & $\mathbf{1} $ & $.$ & $.$ & $3$ & $.$ & $5$ & $.$ & $.$ & $.$ \\
\bottomrule
\end{tabular}
\caption{
SU(5) representations and charges under discrete shaping symmetries
of the flavon and flavour messenger fields of the model presented in subsection \ref{sec:modelA}.
Note that the messengers $Z_{5,1}\bar Z_{5,1}$, $Z_{10,1}\bar Z_{10,1}$,
$Z_{10,2}\bar Z_{10,2}$, $Z_{45,1}\bar Z_{45,1}$ and $Z_{45,2}\bar Z_{45,2}$
have no direct mass term, but get their masses through the VEVs of $S$ and $S'$.
The other messenger fields have direct mass terms, so their corresponding barred
field is omitted in this table. A dot denotes charge zero.
\label{tab:A_messengers}
}
\end{table}
The adjoint $H_{24}'$ required to
construct the desired CG factors must be charged under shaping symmetries.
We will therefore implement superpotential (a). Note that it leaves the second
adjoint $H_{24}$ uncharged, which could in principle lead to a problem.
A direct mass term of messenger fields $M_i Z_i\bar Z_i$ would
in this case always show a up with a term of the form $H_{24} Z_i\bar Z_i$.
Such a contribution would inevitably spoil the desired CG factors~\cite{Antusch:2013rxa}
between $Y_d$ and $Y_e^T$ as long as the mass and the
adjoint VEV are not very hierarchical. To avoid this and still generate the desired operators
the masses of the messenger fields that give rise to $W_d$ in eq.~(\ref{eq:A_Wd}) originate
from the VEVs of the fields $S$ and $S^\prime$
charged under the shaping symmetry (but with different charges than $H_{24}'$)
\footnote{The VEV of the charged singlet $S$ gives masses to
the heavy messengers of the DMPM (see section~\ref{subsec:dmpm24}).}.
\begin{figure}
\centering
\includegraphics[scale=0.5]{supergraph_optionA_yb.pdf}
\includegraphics[scale=0.5]{supergraph_optionA_ys.pdf}
\includegraphics[scale=0.5]{supergraph_optionA_yd.pdf}
\caption{Supergraphs leading to the effective superpotential $W_d$ of
eq.~\eqref{eq:A_Wd} when the heavy messenger fields get integrated out in the model presented in subsection \ref{sec:modelA}.
\label{fig:optionA_Yd}
}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.5]{supergraph_optionA_yu23.pdf}
\includegraphics[scale=0.5]{supergraph_optionA_yu13.pdf}
\includegraphics[scale=0.5]{supergraph_optionA_yu22.pdf}
\includegraphics[scale=0.5]{supergraph_optionA_yu12a.pdf}
\includegraphics[scale=0.5]{supergraph_optionA_yu12b.pdf}
\includegraphics[scale=0.5]{supergraph_optionA_yu12c.pdf}
\includegraphics[scale=0.5]{supergraph_optionA_yu11.pdf}
\caption{Supergraphs leading to the effective superpotential $W_u$ of
eq.~\eqref{eq:A_Wu} when the heavy messenger fields are integrated out in the model presented in subsection \ref{sec:modelA}. Note that there are three supergraphs contributing to the superpotential term generating $y_{12}$. A detailed discussion of the messenger sector is
presented in the appendix \ref{app:mess}.}
\label{fig:optionA_Yu}
\end{figure}
The full messenger sector can be read off from the supergraphs presented in Figures~\ref{fig:optionA_Yd} and \ref{fig:optionA_Yu}, see also Table~\ref{tab:A_messengers}.
After the heavy messenger fields get integrated out, the effective
superpotentials $W_d$ and $W_u$ are obtained.
The predictivity of the down-type quarks and charged lepton Yukawa
couplings has already been discussed above. Note that the hierarchical
structure is enforced due to the use of higher order effective
operators in $W_d$. In a small angle approximation the leading order
estimates for the eigenvalues of $Y_u$ and the mixing angles are given as
\begin{equation}
y_u \approx y_{11} - \frac{y_{12}^2}{y_{22}} \;,\quad
y_c \approx y_{22} \;,\quad
y_t \approx y_{33}\;,\quad
\theta_{C} \approx \frac{y_{12}}{y_{22}} \;,\quad
\theta_{23} \approx\frac{y_{23}}{y_{33}} \;,\quad
\theta_{13} \approx\frac{y_{13}}{y_{33}}\;.
\end{equation}
Phenomenology requires that all parameters $y_{ij}$ of $Y_u$ are independent, which
needs to be carefully considered in the construction of the messenger
sector, as discussed in Appendix~\ref{app:mess}.
In section \ref{subsec:dmpm24} we argued that for the case of an uncharged $H_{24}$,
as it appears in the selected superpotential (a), the mass term for the additional,
five-dimensional Higgs fields must come from the VEV of some singlet field.
In our model an effective $\mu'$ term is generated from a higher-dimensional operator and
with an even higher suppression, there is also a $\mu$-term for the Higgs fields coupling to matter,
\begin{equation}
W_5^{\text{eff}} = \mu H_5\bar H_5 + \mu' H_5^\prime\bar H_5^\prime \;,
\end{equation}
where $\mu' \equiv \vev{\theta_3}^4/M_{\text{Pl}}^3$ and
$\mu \equiv \vev{\theta_3} \vev{\theta_4}^4/M_{\text{Pl}}^4$.\footnote{Alternatively,
we checked that a UV-complete generation of these $\mu$ and $\mu^\prime$ operators
via messenger fields would be possible. However, the necessary messenger fields
are not included in the model and it turns out that the masses are already generated
by Planck-scale suppressed effective operators.}
The mass matrices for the doublet and triplet components are then given by
\begin{equation}
m_D = \begin{pmatrix}
\mu & 0 \\
0 & \mu'
\end{pmatrix}\;,\quad\;
\quad m_T = \begin{pmatrix}
\mu & 0 & 0 & -\frac{V_1^2}{\vev{S}}\\
0 & \mu' & -\frac{V_1^2}{\vev{S}} & 0 \\
-\frac{V_1^2}{\vev{S}} & 0 & \vev{S} & 0\\
0 & -\frac{V_1^2}{\vev{S}} & 0 & \vev{S}
\end{pmatrix} \;,
\end{equation}
where $V_1$ is defined in Eq.~\eqref{eq:vevdef1}.
After the heavy $50$-dimensional fields are integrated out and SU(5)
gets spontaneously broken, the mass matrices for the doublet and triplet
components of the Higgs fields $H_5$, $H^\prime_5$ and their corresponding
barred fields are given by
\begin{equation}
m_D = \begin{pmatrix}
\mu & 0 \\
0 & \mu'
\end{pmatrix}\;,\quad
m_T = \begin{pmatrix}
\mu & -\frac{V_1^4}{ \vev{S}^3} \\
-\frac{V_1^4}{\vev{S}^3} & \mu'
\end{pmatrix}\;.
\end{equation}
The effective triplet mass for dimension five proton decay is then given by
\begin{equation}
\label{eq:A_mteff}
{M^\text{dim=5}_{T}} = \left(m_T^{-1}\right)_{11}^{-1} = \mu - \frac{V_1^8}{\vev{S}^6 \mu'}
\approx -\frac{V_1^8 M_{\text{Pl}}^3}{\vev{S}^6 \vev{\theta_3}^4} \;.
\end{equation}
and the effective triplet masses suppressing dimension six proton decay are given by
\begin{equation}
\left({M^\text{dim=6}_{T}}\right)^2 = \left({M^\text{dim=6}_{\bar T}}\right)^2 \approx\frac{\abs{V_1}^8}{\abs{\vev{S}}^6}\;.
\end{equation}
Let us give an explicit example for the scales involved in the model: Because of
perturbativity the mass of the $50$-dimensional Higgs fields has to be almost at the
Planck scale. We therefore assume $\vev{S} \sim 10^{-1} M_{\text{Pl}}$. Using the known
values of the Yukawa couplings, we can estimate the values of the relevant masses of
our model. At the GUT scale with $\tan\beta = 30$ the Yukawa couplings are approximately
given by \cite{Antusch:2013jca}
\begin{equation}
\label{eq:A_yuk_num}
y_d \approx 1.6 \cdot 10^{-4}\;, \;\;
y_s \approx 3 \cdot 10^{-3}\;, \text{ and }
y_b \approx 0.18\;.
\end{equation}
In our example model these Yukawa couplings are (up to order one couplings)
given by the operators
\begin{equation}
\label{eq:A_yuk_eqs}
y_d \sim \frac{\vev{H_{24}}\vev{H_{24}^\prime}\vev{\theta_4}}{\vev{S^\prime}^2\vev{S}}\;,\quad
y_s\sim\frac{\vev{H_{24}^\prime}\vev{\theta_3}}{\vev{S^\prime}^2} \text{ and }
y_b\sim\frac{\vev{H_{24}^\prime}}{\vev{S^\prime}}\;.
\end{equation}
Then we find from eqs.~\eqref{eq:A_mteff}--\eqref{eq:A_yuk_eqs} for the
effective triplet masses, the $\mu$-term and the mass of
the additional heavy doublet the following values
\begin{equation}
{M^\text{dim=5}_{T}}\approx 1.4\cdot 10^{19}~{\rm GeV}\;,\quad
{M^\text{dim=6}_{T}}\approx 1.4\cdot 10^{12}~{\rm GeV}\;,\quad
\mu \approx 225~{\rm GeV} \;, \quad
\mu^\prime \approx 130~{\rm TeV} \;.
\end{equation}
Recalling from section \ref{subsec:superpotentiala} the parameters governing superpotential (a) and using here SUSY scale of $M_{\text{SUSY}}=1~{\rm TeV}$, $\lambda\sim 0.19$, $\kappa^\prime\sim0.08$, $M_{24}^\prime = M_{24} =10^{15}~{\rm GeV}$ and $V_2/V_1=1.2$ the GUT scale is given by
\begin{equation}
M_{\text{GUT}} \approx 6.4\cdot 10^{16} {\rm GeV}\;.
\end{equation}
Although these numbers are only estimates, which neglect order one couplings,
they illustrate the model's features: the DTS problem is solved
with large effective triplet masses, therefore the proton decay rate is suppressed
and the fermion mass ratios are realistic. The $\mu$-term emerges from
a Planck-scale suppressed operator.
We remark that the DMPM does not suffer from any dangerous Planck-scale suppressed operators, due to the charge assignment of the singlet field $S$.
\subsection[A model with $\theta_C$ from $Y_d$]{A model with $\boldsymbol{\theta_C}$ from $\boldsymbol{Y_d}$ \label{sec:modelB}}
We now turn to our second example model, which realises the attractive feature of
$\theta_C$ emerging dominantly from the down-type quark mixing $\theta_C \approx \theta_{12}^d$.
The Yukawa matrices are given by the following structure
\begin{equation}
\label{eq:B_Y}
Y_d=\begin{pmatrix}
0 & y_{d,12}& 0\\
y_{d,21} & y_s& 0\\
0 & 0& y_b
\end{pmatrix} ,
Y_e=\begin{pmatrix}
0 & 6 y_{d,21}& 0\\
-\frac{1}{2}y_{d,12} & 6 y_s& 0\\
0 & 0& -\frac{3}{2}y_b
\end{pmatrix} ,
Y_u=\begin{pmatrix}
y_{11} & y_{12} & 0\\
y_{12} & y_{22} & y_{23}\\
0 & y_{23} & y_{33}
\end{pmatrix} .
\end{equation}
A similar structure \cite{Antusch:2013kna,Antusch:2013tta} has been used to explain the relation $\theta_{13}^\text{PMNS}= \theta_C / \sqrt{2}$ via charged lepton corrections, and a right-handed quark unitarity triangle \cite{oai:arXiv.org:0910.5127}.
From the matrices we can see that we want to couple both $\mathcal{F}_1$ and $\mathcal{F}_2$ to $\mathcal{T}_2$
but still distinguish both fields from each other to forbid $(Y_d)_{11}$. This suggests to
use at least a $\mathbb{Z}_3$ symmetry and hence we will use superpotential (b) from
section~\ref{sec:W2Adjoints} where both adjoints are charged.
The effective superpotentials that lead to the desired Yukawa matrices after integrating out
heavy messenger fields and breaking of the GUT gauge group are
\begin{align}
W_u &= \frac{1}{\Lambda^2} H_5 {\cal T}_1 {\cal T}_1 \theta_5^2
+ \frac{1}{\Lambda^2} H_5 {\cal T}_1 {\cal T}_2 \theta_2^2
+ \frac{1}{\Lambda^2} H_5 {\cal T}_2 {\cal T}_2 \theta_1^2
+ \frac{1}{\Lambda} H_5 {\cal T}_2 {\cal T}_3 \theta_1
+ H_5 {\cal T}_3 {\cal T}_3\;,\label{eq:B_Wu} \\
W_d &= \frac{1}{\vev{S^\prime}} (H_{24}^\prime {\mathcal{F}}_3)_{\bar{5}} ({\bar H}_{5} {\mathcal{T}}_3)_{5}
+ \frac{\theta_3}{\vev{S^\prime}^2} (H_{24} {\mathcal{T}}_2)_{10} ( {\bar H}_{5} {\mathcal{F}}_1)_{\overline{10}}
+ \frac{\theta_4}{\vev{S^\prime}^2} (H_{24}^\prime {\mathcal{T}}_2)_{10} ( {\bar H}_{5} {\mathcal{F}}_2)_{\overline{10}}\nonumber\\
&+ \frac{1}{\vev{S}^2} (H_{24} {\mathcal{F}}_2)_{\overline{45}} ( {\mathcal{T}}_1 H_{24} {\bar H}_{5})_{45}\;,\label{eq:B_Wd}
\end{align}
where again VEVs of singlet fields appear in the denominators by virtue of messenger masses generated by them. Note that in comparison to $Y_u$ of the previous example model of eq.~\eqref{eq:A_Y}, the vanishing $(Y_u)_{13}$ element requires to introduce an additional flavon
field $\theta_5$. The supergraphs that generate these effective operators are shown in Figure~\ref{fig:optionB_Yd} and \ref{fig:optionB_Yu}. A complete list of all fields including their charges and representations is given in Tables~\ref{tab:B_MSSM}, \ref{tab:B_VEVS}, \ref{tab:B_DMPM} and \ref{tab:B_messengers}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.5]{supergraph_optionA_yb.pdf}
\includegraphics[scale=0.5]{supergraph_optionB_y21.pdf}
\includegraphics[scale=0.5]{supergraph_optionB_ys.pdf}
\includegraphics[scale=0.5]{supergraph_optionB_y12.pdf}
\end{center}
\caption{Supergraphs leading to the effective superpotential $W_d$ of eq.~(\ref{eq:B_Wd}) when the heavy messenger fields get integrated out in the model presented in subsection \ref{sec:modelB}.}
\label{fig:optionB_Yd}
\end{figure}
\begin{table}
\begin{tabular}{cccccccccc}
\toprule
& SU(5) & $\mathbb{Z}_{2}$ & $\mathbb{Z}_{3}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{5}$ & $\mathbb{Z}_{6}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{2}$ \\
\midrule
$H_5$ & $\mathbf{5} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ \\
$\bar H_{5}$ & $\mathbf{\bar 5} $ & $.$ & $2$ & $.$ & $1$ & $4$ & $6$ & $6$ & $.$ \\
$\mathcal{T}_1$ & $\mathbf{10} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $6$ & $5$ & $1$ \\
$\mathcal{T}_2$ & $\mathbf{10} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $1$ & $.$ & $1$ \\
$\mathcal{T}_3$ & $\mathbf{10} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $1$ \\
$\mathcal{F}_1$ & $\mathbf{\bar 5} $ & $.$ & $.$ & $.$ & $3$ & $2$ & $5$ & $6$ & $1$ \\
$\mathcal{F}_2$ & $\mathbf{\bar 5} $ & $.$ & $2$ & $.$ & $3$ & $.$ & $2$ & $3$ & $1$ \\
$\mathcal{F}_3$ & $\mathbf{\bar 5} $ & $.$ & $1$ & $3$ & $4$ & $.$ & $1$ & $1$ & $1$ \\
\bottomrule
\end{tabular}
\caption{
SU(5) representations and charges under discrete shaping symmetries
of the MSSM fields and colour triplets of the model presented in subsection \ref{sec:modelB}. A dot denotes charge zero.
\label{tab:B_MSSM}
}
\end{table}
\begin{table}
\begin{tabular}{cccccccccc}
\toprule
& SU(5) & $\mathbb{Z}_{2}$ & $\mathbb{Z}_{3}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{5}$ & $\mathbb{Z}_{6}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{2}$ \\
\midrule
$H_{24}$ & $\mathbf{24} $ & $.$ & $.$ & $.$ & $.$ & $4$ & $.$ & $.$ & $.$ \\
$H_{24}^\prime$ & $\mathbf{24} $ & $.$ & $.$ & $.$ & $.$ & $2$ & $.$ & $.$ & $.$ \\
$S$ & $\mathbf{1} $ & $1$ & $2$ & $.$ & $2$ & $.$ & $.$ & $.$ & $.$ \\
$S^\prime$ & $\mathbf{1} $ & $.$ & $.$ & $3$ & $.$ & $.$ & $.$ & $.$ & $.$ \\
$\theta_1$ & $\mathbf{1} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $6$ & $.$ & $.$ \\
$\theta_2$ & $\mathbf{1} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $1$ & $.$ \\
$\theta_{3}$ & $\mathbf{1} $ & $.$ & $1$ & $2$ & $1$ & $2$ & $2$ & $2$ & $.$ \\
$\theta_{4}$ & $\mathbf{1} $ & $.$ & $2$ & $2$ & $1$ & $.$ & $5$ & $5$ & $.$ \\
$\theta_5$ & $\mathbf{1} $ & $.$ & $.$ & $.$ & $.$ & $3$ & $1$ & $2$ & $.$ \\
\bottomrule
\end{tabular}
\caption{
SU(5) representations and charges under discrete shaping symmetries
of the superfields obtaining VEVs at around the GUT scale of the model presented in subsection \ref{sec:modelB}. A dot denotes charge zero.
\label{tab:B_VEVS}
}
\end{table}
\begin{table}
\begin{tabular}{cccccccccc}
\toprule
& SU(5) & $\mathbb{Z}_{2}$ & $\mathbb{Z}_{3}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{5}$ & $\mathbb{Z}_{6}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{2}$ \\
\midrule
$H_{5}^\prime$ & $\mathbf{5} $ & $1$ & $1$ & $.$ & $.$ & $4$ & $1$ & $1$ & $.$ \\
$\bar H_{5}^\prime$ & $\mathbf{\bar 5} $ & $1$ & $.$ & $.$ & $1$ & $2$ & $.$ & $.$ & $.$ \\
$X_{45}$ & $\mathbf{45} $ & $.$ & $1$ & $.$ & $4$ & $4$ & $1$ & $1$ & $.$ \\
$\bar X_{45}$ & $\mathbf{\overline{45}} $ & $1$ & $.$ & $.$ & $4$ & $2$ & $6$ & $6$ & $.$ \\
$Y_{45}$ & $\mathbf{45} $ & $.$ & $2$ & $.$ & $3$ & $2$ & $1$ & $1$ & $.$ \\
$\bar Y_{45}$ & $\mathbf{\overline{45}} $ & $1$ & $2$ & $.$ & $.$ & $4$ & $6$ & $6$ & $.$ \\
$Z_{50}$ & $\mathbf{50} $ & $1$ & $.$ & $.$ & $1$ & $.$ & $1$ & $1$ & $.$ \\
$\bar Z_{50}$ & $\mathbf{\overline{50}} $ & $.$ & $1$ & $.$ & $2$ & $.$ & $6$ & $6$ & $.$ \\
$X_{45}^\prime$ & $\mathbf{45} $ & $1$ & $.$ & $.$ & $4$ & $.$ & $.$ & $.$ & $.$ \\
$\bar X_{45}^\prime$ & $\mathbf{\overline{45}} $ & $.$ & $1$ & $.$ & $4$ & $.$ & $.$ & $.$ & $.$ \\
$Y_{45}^\prime$ & $\mathbf{45} $ & $1$ & $1$ & $.$ & $3$ & $4$ & $.$ & $.$ & $.$ \\
$\bar Y_{45}^\prime$ & $\mathbf{\overline{45}} $ & $.$ & $.$ & $.$ & $.$ & $2$ & $.$ & $.$ & $.$ \\
$Z_{50}^\prime$ & $\mathbf{50} $ & $.$ & $2$ & $.$ & $1$ & $2$ & $.$ & $.$ & $.$ \\
$\bar Z_{50}^\prime$ & $\mathbf{\overline{50}} $ & $1$ & $2$ & $.$ & $2$ & $4$ & $.$ & $.$ & $.$ \\
\bottomrule
\end{tabular}
\caption{
SU(5) representations and charges under discrete shaping symmetries
of the fields in the DMPM sector of the model presented in subsection \ref{sec:modelB}. A dot denotes charge zero.
\label{tab:B_DMPM}
}
\end{table}
\begin{table}
\centering
\begin{tabular}{cccccccccc}
\toprule
& SU(5) & $\mathbb{Z}_{2}$ & $\mathbb{Z}_{3}$ & $\mathbb{Z}_{4}$ & $\mathbb{Z}_{5}$ & $\mathbb{Z}_{6}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{7}$ & $\mathbb{Z}_{2}$ \\
\midrule
$Z_5$ & $\mathbf{5} $ & $.$ & $2$ & $1$ & $1$ & $4$ & $6$ & $6$ & $1$ \\
$\bar{Z}_5$ & $\mathbf{\bar 5} $ & $.$ & $1$ & $.$ & $4$ & $2$ & $1$ & $1$ & $1$ \\
$Z_{10,1}$ & $\mathbf{10} $ & $.$ & $1$ & $.$ & $1$ & $.$ & $3$ & $2$ & $1$ \\
$\bar{Z}_{10,1}$ & $\mathbf{\overline{10}} $ & $.$ & $2$ & $1$ & $4$ & $.$ & $4$ & $5$ & $1$ \\
$Z_{10,2}$ & $\mathbf{10} $ & $.$ & $.$ & $1$ & $.$ & $4$ & $1$ & $.$ & $1$ \\
$\bar{Z}_{10,2}$ & $\mathbf{\overline{10}} $ & $.$ & $.$ & $.$ & $.$ & $2$ & $6$ & $.$ & $1$ \\
$Z_{10,3}$ & $\mathbf{10} $ & $.$ & $2$ & $.$ & $1$ & $2$ & $6$ & $5$ & $1$ \\
$\bar{Z}_{10,3}$ & $\mathbf{\overline{10}} $ & $.$ & $1$ & $1$ & $4$ & $4$ & $1$ & $2$ & $1$ \\
$Z_{10,4}$ & $\mathbf{10} $ & $.$ & $.$ & $1$ & $.$ & $2$ & $1$ & $.$ & $1$ \\
$\bar{Z}_{10,4}$ & $\mathbf{\overline{10}} $ & $.$ & $.$ & $.$ & $.$ & $4$ & $6$ & $.$ & $1$ \\
$Z_{45,1}$ & $\mathbf{45} $ & $.$ & $1$ & $.$ & $2$ & $2$ & $5$ & $4$ & $1$ \\
$\bar{Z}_{45,1}$ & $\mathbf{\overline{45}} $ & $1$ & $.$ & $.$ & $1$ & $4$ & $2$ & $3$ & $1$ \\
\midrule
$Z_{10,5}$ & $\mathbf{10} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $6$ & $.$ & $1$ \\
$Z_{10,6}$ & $\mathbf{10} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $1$ & $2$ & $1$ \\
$Z_1$ & $\mathbf{1} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $2$ & $.$ & $.$ \\
$Z_{2}$ & $\mathbf{1} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $.$ & $5$ & $.$ \\
$Z_{3}$ & $\mathbf{1} $ & $.$ & $.$ & $.$ & $.$ & $.$ & $5$ & $3$ & $.$ \\
\bottomrule
\end{tabular}
\caption{
SU(5) representations and charges under discrete shaping symmetries
of the flavon and flavour messenger fields of the model presented in subsection \ref{sec:modelB}.
Note that the messengers $Z_{5,1}\bar Z_{5,1}$, $Z_{10,1}\bar Z_{10,1}$,
$Z_{10,2}\bar Z_{10,2}$, $Z_{10,3}\bar Z_{10,3}$, $Z_{10,4}\bar Z_{10,4}$
and $Z_{45,1}\bar Z_{45,1}$ have no direct mass term, but get their masses
through VEVs of $S$ and $S'$. The other messenger fields have direct mass
terms, so their corresponding barred field is not shown in the table. A dot denotes charge zero.
\label{tab:B_messengers}
}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[scale=0.5]{supergraph_optionB_yu23.pdf}
\includegraphics[scale=0.5]{supergraph_optionB_yu22.pdf}
\includegraphics[scale=0.5]{supergraph_optionB_yu12a.pdf}
\includegraphics[scale=0.5]{supergraph_optionB_yu12b.pdf}
\includegraphics[scale=0.5]{supergraph_optionB_yu11.pdf}
\end{center}
\caption{Supergraphs leading to the effective superpotential $W_u$ of eq.~(\ref{eq:B_Wu}) when the heavy messenger fields get integrated out.}
\label{fig:optionB_Yu}
\end{figure}
In a small angle approximation the mixing angles and Yukawa couplings are given by
\begin{align}
y_d\approx \frac{y_{d,12}y_{d,21}}{y_s}\;,\quad & \theta_C\approx\theta_{12}^d\approx\frac{y_{d,12}}{y_s}\;,\nonumber\\
y_u\approx y_{11}-\frac{y_{12}^2}{y_{22}}\;,\quad y_c\approx y_{22}\;,\quad y_t\approx y_{33}\;,\quad & \theta_{23}\approx\frac{y_{23}}{y_{33}}\;,\quad\theta_{13}\approx\frac{y_{12}}{y_{22}}\theta_{23}\;,
\end{align}
with $y_s$ and $y_b$ given as parameters of $Y_d$ in eq.~(\ref{eq:B_Y}). Thus the Yukawa matrices can fit the experimental values without tension.
We continue now by discussing additional details concerning the use of the DMPM in this model.
Like in the previous model, Planck-scale suppressed operators generate mass terms for the
five-dimensional Higgs representations
\begin{equation}
W_5^{\text{eff}} = \mu H_5\bar H_5 + \mu' H_5^\prime\bar H_5^\prime \;,
\end{equation}
where $\mu = \vev{\theta_3}^4 / M_{\text{Pl}}^3$ and $\mu' = \vev{\theta_4}^4 / M_{\text{Pl}}^3$. Note that although they appear at the same order, a modest hierarchy $\vev{\theta_3} < \vev{\theta_4}$, as we have in our model, will sufficiently split their masses. Furthermore, after integrating out the 45-dimensional messengers of the DMPM, the mass matrices for doublet and triplet components of the $5$- and $50$-dimensional
superfields are given by
\begin{equation}
m_D = \begin{pmatrix}
\mu & 0 \\
0 & \mu'
\end{pmatrix}\;,\quad\;
\quad m_T = \begin{pmatrix}
\mu & 0 & 0 & -\frac{V_1^2}{\vev{S}}\\
0 & \mu' & -\frac{V_1^2}{\vev{S}} & 0 \\
-\frac{V_1^2}{\vev{S}} & 0 & \vev{S} & 0\\
0 & -\frac{V_1^2}{\vev{S}} & 0 & \vev{S}
\end{pmatrix} \;,
\end{equation}
where $V_1$ is defined in Eq.~\eqref{eq:vevdef1}.
After the heavy 50-dimensional fields get integrated out the mass matrices become
\begin{equation}
m_D = \begin{pmatrix}
\mu & 0 \\
0 & \mu'
\end{pmatrix}\;,\quad
m_T = \begin{pmatrix}
\mu & - \frac{V_1^4}{ \vev{S}^3} \\
- \frac{V_1^4}{\vev{S}^3} & \mu'
\end{pmatrix}\;.
\end{equation}
This leads to an effective triplet mass for dimension five proton decay of
\begin{equation}
{M^\text{dim=5}_{T}} = \left(m_T^{-1}\right)_{11}^{-1}
= \mu - \frac{V_1^8}{\vev{S}^6 \mu'}
\approx -\frac{V_1^8 M_{\text{Pl}}^3}{\vev{S}^6 \vev{\theta_4}^4} \;,
\end{equation}
and the effective triplet masses suppressing dimension six proton decay are given by
\begin{equation}
\left({M^\text{dim=6}_{T}}\right)^2 = \left({M^\text{dim=6}_{\bar T}}\right)^2 = \left(m_T^{-1}m_T^{\dagger-1}\right)_{11}^{-1}\approx\frac{\abs{V_1}^8}{\abs{\vev{S}}^6}\;.
\end{equation}
As for the first example model we can estimate the values of the relevant masses in an explicit example from the known Yukawa couplings. In a small angle approximation the down-type Yukawa couplings are given by
\begin{align}
y_d & \sim \frac{\vev{H_{24}}^2}{\vev{S}^2}\frac{\vev{H_{24}}\vev{\theta_3}}{\vev{S^\prime}^2}\frac{1}{y_s}\approx 1.6\cdot 10^{-4} \text{\, with \,}y_s\sim\frac{\vev{H_{24}^\prime}\vev{\theta_4}}{\vev{S^\prime}^2}\approx 3\cdot 10^{-3}
\end{align}
and
\begin{align}
y_b&\sim\frac{\vev{H_{24}^\prime}}{\vev{S^\prime}} \approx 0.18\;,
\end{align}
where the numerical values for the Yukawa couplings taken from~\cite{Antusch:2013jca} are valid for
$\tan\beta=30$ and order one coefficients have been neglected. For the effective triplet masses of dimension five and six proton decay, respectively, the $\mu$-term and the mass of the additional, heavier Higgs doublets the following estimates emerge
\begin{equation}\label{eq:numbersB}
{M^\text{dim=5}_{T}}\approx 1.4 \cdot 10^{18}~{\rm GeV} \;,\quad {M^\text{dim=6}_{T}}\approx 10^{12}~{\rm GeV}\;,\quad \mu\approx 7~{\rm TeV} \;,\quad\mu^\prime\approx 800~{\rm TeV}\;.
\end{equation}
The parameters of superpotential (b) from section \ref{sec:Superpotentialbc} have been chosen to be real with values $\tan\beta_V= 0.5$, $\lambda=10^{-4}$ and $M=2.4\cdot 10^{12}~{\rm GeV}$. The mass of the 45-dimensional and 50-dimensional superfields has been set to $\vev{S}= 10^{18}~{\rm GeV}$. With the numbers of Eq.~\eqref{eq:numbersB} we find
\begin{equation}
M_{\text{SUSY}}\approx 24~{\rm TeV} \;,\quad M_{\text{GUT}}\approx 4\cdot 10^{17}~{\rm GeV}\;.
\end{equation}
Thus, also in this model, proton decay can be suppressed by large effective
triplet masses. DTS is achieved through the DMPM and the light doublet gets
its mass from an Planck-scale suppressed operator. The Yukawa sector of the model
features viable fermion masses and quark mixing angles.
\section{Proton Decay}
\label{sec:proton}
We have split our discussion of proton decay into two parts. In the first part we comment on
proton decay induced by dimension five operators in the superpotential.
These operators are usually considered to be more dangerous for the
validity of any SUSY GUT model. We will argue why we are more
predictive than ordinary models and still should be able to evade
current experimental bounds. In the second part we will comment
on dimension six proton decay operators, which appear also in non-SUSY GUTs. The dimension six proton decay operator emerging from the exchange of heavy gauge bosons are considered to be not as dangerous in
SUSY GUTs due to the usually higher unification scale~\cite{Murayama:2001ur}. The dimension six proton decay mediated by colour triplets, however, has to be suppressed by a high enough effective triplet mass, as for the case of dimension five proton decay.
\subsection{Proton Decay from Dimension Five Operators}
We will adopt for this section a notation similar to the SLHA
convention \cite{Skands:2003cj}, for more details see appendix \ref{app:Clebsch}.
The superpotential describing the couplings of the matter fields
to the heavy colour triplet reads
\begin{equation}
\begin{split}
W_T =&\; \epsilon_{\alpha\beta} \left(
- \frac{1}{2} (Y_{qq})_{ij} \epsilon_{abc} T^a Q_i^{\alpha b} Q_j^{\beta c}
+ (Y_{ql})_{ij} \bar{T}^a Q_i^{\alpha a} L_j^\beta
\right) \\
+&\;
(Y_{ue})_{ij} T^a \bar{U}_i^a \bar{E}_j
- (Y_{ud})_{ij} \epsilon_{abc} \bar{T}^a \bar{U}_i^{b} \bar{D}_j^{c} +
M_T T^a \bar{T}^a \;.
\end{split}
\end{equation}
In many SU(5)
models at least for the first two generations it was assumed that the Yukawa
couplings are generated or significantly corrected by some set of higher-dimensional
operators, see for instance, \cite{Bajc:2002pg, EmmanuelCosta:2003pu, Wiesenfeldt:2004qa}. But, usually,
there is no control over which operator is the dominant one and hence it is not
possible to calculate how strong exactly
the heavy triplets couple to the MSSM fields.
In our setup this is not the case. For every entry of the Yukawa matrix we have specified
the operator with only very small corrections, if any. Therefore we know how strong
the MSSM fields couple to the heavy triplets if we know the MSSM Yukawa couplings
at the GUT scale. To be more precise the new Yukawa couplings are related to the MSSM
ones only via CG coefficients which are fixed by the gauge structure of the
underlying operator. We find for our first example model
\begin{align}
Y_d &= \text{Diag}\left( y_d, y_s, y_b \right) \;,\quad
Y_e = \text{Diag}\left( - \frac{1}{2} y_d, 6 y_s, - \frac{3}{2} y_b \right) \;, \\
Y_{ql} &= \text{Diag}\left( y_d, y_s, -\frac{3}{2} y_b \right) \;,\quad
Y_{ud} = \text{Diag}\left( \frac{2}{3} y_d, - 4 y_s, y_b \right) \;,\quad \\
Y_{qq} &= Y_u \;,\quad Y_{ue} = Y_u \;,
\end{align}
where the structure of $Y_u$ can be read off from eq.~\eqref{eq:A_Y}.
For the second model we find
\begin{align}
Y_d &= \begin{pmatrix} 0 & y_{d,12} & 0 \\ y_{d,21} & y_s & 0 \\ 0 & 0 & y_b \end{pmatrix} \;,\quad
Y_e^T =\begin{pmatrix} 0 & -\tfrac{1}{2} y_{d,12} & 0 \\ 6 y_{d,21} & 6 y_s & 0 \\ 0 & 0 & -\tfrac{3}{2} y_b \end{pmatrix} \;, \\
Y_{ql} &= \begin{pmatrix} 0 & y_{d,12} & 0 \\ y_{d,21} & y_s & 0 \\ 0 & 0 & -\tfrac{3}{2} y_b \end{pmatrix} \;,\quad
Y_{ud} = \begin{pmatrix} 0 & \tfrac{2}{3} y_{d,12} & 0 \\ -4 y_{d,21} & -4 y_s & 0 \\ 0 & 0 & y_b \end{pmatrix} \;,\quad \\
Y_{qq} &= Y_u \;,\quad Y_{ue} = Y_u \;,
\end{align}
where the structure of $Y_u$ can be read off from eq.~\eqref{eq:B_Y}.\footnote{GUT
textures for proton decay, without fully constructed models, have been considered,
for example, in \cite{Nath:1996qs, Wiesenfeldt:2004qa}.}
The dimension five operators which violate baryon number after integrating
out the triplets read
\begin{equation}
W_{\cancel{B}} = \frac{1}{M_T^{\text{eff}}} \left[ \frac{1}{2} Y_{qq}^{ij} Y_{ql}^{mn} Q_i Q_j Q_m L_n + Y_{ue}^{ij} Y_{ud}^{mn} \bar U_i \bar E_j \bar U_m \bar D_n \right] \;,
\label{eq:Dim5pDecay}
\end{equation}
where we have suppressed $\epsilon$-tensors. The first operator is called
the $LLLL$ operator and the second one the $RRRR$ operator.
To make definite predictions for the proton decay rate one would have to
take the RGE evolution of these operators to low energies into account
and dress the operators with a closed loop including MSSM particles
to calculate the decay rate of the proton, see e.g.~\cite{Nath:2006ut} and
references therein. This goes clearly beyond the scope of this paper. Before we
give some more qualitative statements about what we expect for the proton decay rate in
comparison to other models we want to argue first that the operators
in eq.~\eqref{eq:Dim5pDecay} together with $M_T^{\text{eff}}$ from
eq.~\eqref{eq:MTeff24_definition} give the dominant dimension five contribution.
Inside the additional higher dimensional representations used in the DMPM
there are additional colour triplets which could in principle give the same
operators like in eq.~\eqref{eq:Dim5pDecay} but with a weaker suppression.
We have explicitly checked that the DMPM by itself is safe: the additional
colour triplets from 5-, 45- and 50-dimensional representations can only
mediate operators that are suppressed compared to the leading contribution by one or more powers of $H_{24} / M_{\text{Pl}}$ and $H_{24}^\prime / M_{\text{Pl}}$.
We have also checked for each of our models that no other Yukawa matrix entries
(through components in the messengers) lead to a lower ${M^\text{dim=5}_{T}}$. This is,
of course, model dependent and has to be checked for any specific model.
We now turn to a qualitative discussion
of dimension five proton decay in the considered class of models. Firstly, we want to
point out that especially in the first model one could expect
some decay modes of the proton to be suppressed because $Y_{ql}$ and $Y_{ud}$
are diagonal in flavour space. Therefore, decays which need a flavour transition
in these matrices would be suppressed. This reinforces the statement that our setup is
more predictive than conventional ones, due to the better
control over the flavour structure of all the Yukawa matrices governing proton decay.
Secondly, introducing a second adjoint in a renormalizable way gives us
enough freedom to enhance the effective triplet mass ${M^\text{dim=5}_{T}}$
to comfortable levels, cf.~section \ref{sec:GUT}. In the case of superpotential (b),
we get a mass of roughly $5 \cdot 10^{16} \, {\rm GeV}$ for a SUSY scale of 1~TeV
which increases with increasing $M_{{\text{SUSY}}}$, see section~\ref{sec:Superpotentialbc}.
For superpotential (a) the situation is
more involved and depends on the specific GUT scale, and the effective
triplet mass can be easily above $10^{18} \, {\rm GeV}$.
To conclude the discussion of the dimension five operators mediating proton decay
we also want to stress that in our setup we
need only a moderate value of $\tan \beta \approx 25$. As it was pointed out
by Lucas and Raby \cite{Lucas:1996bc} especially the contribution from the
$RRRR$ operator is enhanced by $\tan^2 \beta$ for large $\tan \beta$
which poses a challenge for many GUT scenarios which rely on $\tan \beta \approx 50$.
Our setup with $y_\tau / y_b = \frac{3}{2}$ and $\tan\beta\approx 25$ has therefore a
suppression of proton decay via the $RRRR$ operator by a factor of four compared to
these models.
In summary, together with the large effective triplet mass in the double missing partner
mechanism we hence expect the proton decay rate to be sufficiently small but
possibly in the reach of the next generation of proton decay experiments.
\subsection{Proton Decay from Dimension Six Operators}
We now turn to the discussion of dimension six proton decay. As discussed in section
\ref{sec:strat}, the dimension six proton decay mediated by colour triplets originates
from the K\"ahler potential
\begin{equation}
K_T = T^a T^{\dagger a} + \bar T^a \bar T^{\dagger a}\;.
\end{equation}
When the colour triplets are integrated out, the dimension six baryon number violating K\"ahler operators emerge as
\begin{equation}
\label{eq:Dim6pDecayTriplets}
K_{\cancel{B}} =
- \frac{1}{\left({M^\text{dim=6}_{\bar T}}\right)^2} \frac{1}{2}Y_{qq}^{ij}Y_{ue}^{* mn}Q_iQ_j\bar U^\dagger_m\bar E^\dagger_n-
\frac{1}{\left({M^\text{dim=6}_{T}}\right)^2} Y_{ql}^{ij}Y_{ud}^{* mn}Q_iL_j\bar U^\dagger_m\bar D_n^\dagger+ h.c.\;,
\end{equation}
where the Yukawa coupling matrices are defined by $W_T$ of Eq.~\eqref{eq:Dim5pDecay}.
As for the dimension five proton decay, these operators need to be subject to a RGE evolution from the GUT scale to the proton mass scale. Note however, that these operators do not need to be dressed with superparticles and therefore the proton decay rates obtained from dimension six operators are independent of the details of the SUSY spectrum.
For completeness, we will now briefly discuss as well proton decay from exchange of heavy gauge bosons.
There is no substantial difference in our models compared to other SUSY SU(5)
models since the gauge structure is exactly the same.
To make the discussion a little bit more analogous to the previous discussion
we define an effective GUT scale
\begin{equation}
M_{\text{GUT}}^\text{eff} = \frac{M_{\text{GUT}}}{\sqrt{\alpha_u}}\;,
\end{equation}
where $\alpha_u$ is the unified gauge coupling at the GUT scale and
$M_{\text{GUT}} \equiv M_V$ is the mass of the leptoquark vector bosons,
cf.~the discussion in section~\ref{sec:GUT}.
Just integrating out the heavy gauge bosons in the tree-level
diagrams governing proton decay, and from dimensional
analysis, we can estimate the lifetime of the proton to be
\begin{equation}
\Gamma_p \approx \alpha_u^2 \frac{m_p^5}{M_V^4} =
\frac{m_p^5}{(M_{\text{GUT}}^\text{eff})^4} \;,
\end{equation}
where $m_p$ is the proton mass and we have neglected RGE effects
and order one coefficients
from nuclear matrix elements and such. The most stringent bound
on the proton lifetime is $\tau(p \to \pi^0 e^+) > 8.2 \cdot 10^{33}$~years
\cite{pdg} which yields an effective GUT scale of about $M_{\text{GUT}}^\text{eff}
\gtrsim 2 \cdot 10^{16} {\rm GeV}$. In section~\ref{sec:GUT} we have seen that
the GUT scale can be easily above $10^{16}$~GeV and
$\sqrt{\alpha_u} \approx 1/5$ such that the proton decay rate from dimension
six operators is sufficiently small but possibly in the reach of the next generation
of proton decay experiments.
\section{Summary and Conclusions \label{sec:conc}}
In this work we have discussed how the double missing partner mechanism solution to the
doublet-triplet splitting problem in four-dimensional supersymmetric
SU(5) Grand Unified Theories can be combined with predictive models featuring novel predictions
for the quark-lepton Yukawa coupling ratios at the GUT scale.
We have argued that towards this goal a second SU(5) breaking Higgs
field in the adjoint representation is very useful. We systematically discussed all
possible renormalizable superpotentials with two adjoint Higgs fields,
also calculating the corresponding constraints on the GUT scale and effective triplet
mass from a two-loop gauge coupling unification analysis. We found that the effective
masses of the colour triplet, which enter dimension five and six proton decay, can easily be raised enough to avoid problems with
proton decay (more than feasible with standard non-renormalizable Higgs potentials with only one
adjoint GUT Higgs field).
We have constructed two explicit flavour models with different predictions for the GUT scale Yukawa sector.
A set of shaping symmetries and a renormalizable messenger sector for the models is presented,
which guarantees that only the desired effective GUT operators are generated when the heavy degrees of freedom are integrated out.
In addition, we also include all possible effective Planck-scale suppressed operators consistent with our symmetries,
and make sure that they do not spoil our results. The models stay perturbative until close to the Planck scale,
such that our predictions do not suffer from large uncertainties due to these Planck-scale suppressed operators.
They serve as existence proofs that predictive models for the GUT scale quark-lepton mass relations
can be combined successfully with the DMPM solution for solving the DTS problem.
We also provide several useful appendices for GUT flavour model building: For instance, one appendix contains the
Clebsch-Gordan coefficients for the couplings of the colour triplets, which are required for calculating the rates for proton
decay induced by their exchange.
We provide detailed tables with the Clebsch-Gordan coefficients for the possible dimension five and six GUT Yukawa operators.
We also discuss there how one can use GUT Higgs potentials for flavour model building, where a (discrete) R-symmetry is
broken spontaneously around the GUT scale. R-symmetries are a helpful ingredient of many flavour models, especially when they include non-Abelian family symmetries.
In summary, we have demonstrated that four-dimensional supersymmetric SU(5) GUTs
with successful doublet-triplet splitting can be combined with predictive models
featuring promising predictions for the quark-lepton Yukawa coupling ratios at
the GUT scale. We have provided the tools for the construction of even more
ambitious GUT models of flavour with additional non-Abelian family symmetries, as
well as towards the calculation of the predictions for the rates of the various
nucleon decay channels in such models by systematically providing the required
Clebsch-Gordan coefficients.
\section*{Acknowledgements}
This work is supported by the Swiss National Science Foundation. We thank Borut Bajc and David Emmanuel-Costa for useful discussions.
\section*{Appendix}
\begin{appendix}
\section{Yukawa Coupling Ratios including Colour Triplets}
\label{app:Clebsch}
The MSSM superpotential is given by
\begin{equation}
W = \epsilon_{\alpha\beta} \left(
(Y_e)^{ij} H_d^\alpha L_i^{\beta} \bar{E}_j +
(Y_d)^{ij} H_d^\alpha Q_i^{\beta a} \bar{D}_j^{a} +
(Y_u)^{ij} H_u^\beta Q_i^{\alpha a} \bar{U}_j^{a} +
\mu H_u^\alpha H_d^\beta
\right) \;,
\end{equation}
where $i$,$j$ are generation indices, $\epsilon_{\alpha \beta}$ the Levi-Civita
tensor ($\epsilon_{12} = 1$), $\alpha$,$\beta$ are SU(2) indices and
$a$, $b$ and $c$ are SU(3) indices\footnote{This definition coincides with the
definitions of \cite{Skands:2003cj}.}. Adding a pair of colour triplets $T$ and $\bar{T}$,
we get the additional terms
\begin{align}
W_T =&\; \epsilon_{\alpha\beta} \left(
- \frac{1}{2} (Y_{qq})_{ij} \epsilon_{abc} T^a Q_i^{\alpha b} Q_j^{\beta c}
+ (Y_{ql})_{ij} \bar{T}^a Q_i^{\alpha a} L_j^\beta
\right) \nonumber\\
+&\;
(Y_{ue})_{ij} T^a \bar{U}_i^a \bar{E}_j
- (Y_{ud})_{ij} \epsilon_{abc} \bar{T}^a \bar{U}_i^{b} \bar{D}_j^{c} +
M_T T^a \bar{T}^a \;,
\end{align}
where $\epsilon_{abc}$ is the three indices Levi-Civita tensor (with $\epsilon_{123} = 1$).
Extending the SM gauge group to SU(5), we embed the MSSM superfields in a $5$-plet
$H_5$, $\bar{5}$-plets ${\bar H}_{5}$ and $\mathcal{F}_i$, and $10$-plets $\mathcal{T}_i$ as in
\begin{align}
H_5 &= \begin{pmatrix} T^r & T^g & T^b & H_u^+ & H_u^0 \end{pmatrix}^T \;,\\
{\bar H}_{5} &= \begin{pmatrix} \bar{T}^r & \bar{T}^g & \bar{T}^b & H_d^- & -H_d^0 \end{pmatrix} \;,\\
\mathcal{F}_i &= \begin{pmatrix} \bar{D}_i^r & \bar{D}_i^g & \bar{D}_i^b & E_i & -\nu_i \end{pmatrix} \;,\\
\mathcal{T}_i &= \frac{1}{\sqrt{2}} \begin{pmatrix}
0 & -\bar{U}_i^b & \bar{U}_i^g & -U_i^r & -D_i^r \\
\bar{U}_i^b & 0 & -\bar{U}_i^r & -U_i^g & -D_i^g \\
-\bar{U}_i^g & \bar{U}_i^r & 0 & -U_i^b & -D_i^b \\
U_i^r & U_i^g & U_i^b & 0 & -\bar{E}_i \\
D_i^r & D_i^g & D_i^b & \bar{E}_i & 0 \\
\end{pmatrix} \;,
\end{align}
where $r,g,b$ are the SU(3) colours and $U$, $D$ and $\nu$, $E$ are the
components of SU(2)-doublets $Q$ and $L$.\footnote{Likewise
$H_d = \begin{pmatrix}H_d^0 & H_d^- \end{pmatrix}^T$ and $H_u =
\begin{pmatrix} H_u^+ & H_u^0 \end{pmatrix}^T$.}
We can write down the renormalizable superpotential terms
\begin{equation}\label{eq:YukawaSU5}
W = (Y_{TF})_{ij} \mathcal{T}_i^{ab} (\mathcal{F}_j)_a (\bar{H}_{5})_b +
\frac{1}{2} (Y_{TT})_{ij} \epsilon_{abcde} \mathcal{T}_i^{ab} \mathcal{T}_j^{cd} H_5^e +
\mu_5 H_5^a (\bar{H}_{5})_a \;,
\end{equation}
where now $a,b,c,d,e$ are SU(5)-indices and $\epsilon_{abcde}$ is the
respective Levi-Civita tensor. From the embedding of the MSSM fields,
one obtains the minimal SU(5) GUT scale relations
\begin{align}
\mu &= M_T = \mu_5 \;, \\
Y_d &= Y_e^T = Y_{ql} = Y_{ud} = \frac{1}{\sqrt{2}} Y_{TF} \;, \\
Y_u &= Y_u^T = Y_{qq} = Y_{ue} = 2 \, Y_{TT} \;.
\end{align}
The relation $Y_d = Y_e^T$ is highly disfavoured as was already
discussed in section \ref{subsec:Clebsch}.
The conventional approach is to add a 45-dimensional Higgs representation
which generates a relative factor of $-3$ between the Yukawa couplings
of the charged leptons and down-type quarks \cite{GJ}.
In this work we instead focus on an approach where the ratios between
Yukawa couplings are fixed by the CG coefficients of
higher-dimensional operators where in addition an adjoint Higgs
representation of SU(5) is added \cite{Antusch:2009gu, Antusch:2013rxa}.
This approach was briefly reviewed in section \ref{subsec:Clebsch} and for
a thorough discussion we refer the interested reader to the original papers, see also
\cite{Spinrath:2010dh}.
\begin{figure}
\begin{center}
\includegraphics[page=1,scale=0.55]{yukawaratiooperators.pdf}
\includegraphics[page=2,scale=0.55]{yukawaratiooperators.pdf}
\end{center}
\caption{Supergraphs generating Yukawa couplings upon integrating out messengers fields in representation $R$,$\bar{R}$, etc.}\label{fig:messengerdiagram}
\end{figure}
In this appendix we will extend the previous discussions to include also the
relative CG coefficents to the triplets. In \cite{Antusch:2009gu, Antusch:2013rxa}
only $Y_d$, $Y_e$ and $Y_u$ were discussed while here we will also
discuss in detail the implications
of this approach for $Y_{ql}$, $Y_{ud}$, $Y_{qq}$, $Y_{ue}$.
The list of the resulting ratios for dimension 4 and 5 operators with Higgs fields
in a 5- and 45-dimensional representation can be found in
Tabs.~\ref{tab:effectiveFToperators1} and \ref{tab:effectiveTToperators1},
where the labels for the representations is defined by figure \ref{fig:messengerdiagram}.
The corresponding results for dimension six operators
are given in Tabs.~\ref{tab:effectiveFToperators2} - \ref{tab:effectiveTToperators3}.
The ones used in the example models are marked with a ``$\rightarrow$''
in the tables.
There are a few comments in order. First, note that several topologies
involving a 45-dimensional messenger field exhibit a free parameter,
simply because the tensor product $\mathbf{45} \otimes \mathbf{24}$
contains two 45-dimensional representations. Hence, there are two
operators possibly giving two different ratios so that any ratio is possible
depending on the coefficients of the two operators. For these cases we
write $x$ in the tables.
We want to mention as well that, unlike at the renormalizable level,
the up-type quark Yukawa and related matrices do not have
to be symmetric or antisymmetric.
Consider, for example, the operator
$(H_{24} \mathcal{T}_1)_{\mathbf{10}} (H_5 \mathcal{T}_2)_{\overline{\mathbf{10}}}$.
Due to the symmetries and messenger content the operator
$(H_{24} \mathcal{T}_2)_{\mathbf{10}} (H_5 \mathcal{T}_1)_{\overline{\mathbf{10}}}$
could be forbidden. In this case we find $(Y_u)_{12} / (Y_u)_{21} = -4$.
Hence we have adopted the following notation for the ratios in the tables
for the Yukawa couplings related to $Y_u$
\begin{align}
(Y_u)_{ij} : (Y_u)_{ji} : (Y_{qq})_{ij} : (Y_{ue})_{ij} : (Y_{ue})_{ji} &=
a : b : c : d : e \;,
\end{align}
which reduces for the diagonal entries of the Yukawa matrices to
\begin{align}
(Y_u)_{ii} : (Y_{qq})_{ii} : (Y_{ue})_{ii} &= (a + b) : c : (d + e) \;.
\end{align}
The ratios related to $Y_{d}$ do not have this extra complication since none
of them could be expected to be symmetric or anti-symmetric in the first place.
\begin{table}
\begin{center}
\setlength{\extrarowheight}{3pt}
\begin{tabular}{ccccc@{ : }c@{ : }c@{ : }c}
\toprule
&$A \, B$ & $C \, D$ & $R$ & $(Y_d)_{ij}$ & $(Y_e)_{ji}$ & $(Y_{ql})_{ij}$ & $(Y_{ud})_{ij}$ \\
\midrule
& \multicolumn{2}{c}{$\mathcal{F}_j \, \mathcal{T}_i \, \bar{H}_5$} & --- & $1$ & $1$ & $1$ & $1$ \\
$\rightarrow$ & $H_{24} \, \mathcal{T}_i$ & $\mathcal{F}_j \, \bar{H}_5$ & 10 & $1$ & $6$ & $1$ & $-4$ \\
& $H_{24} \, \mathcal{T}_i$ & $\mathcal{F}_j \, \bar{H}_5$ & 15 & $1$ & $0$ & $-1$ & $0$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{F}_j \, \mathcal{T}_i$ & $\bar{5}$ & $1$ & $1$ & $-\frac{2}{3}$ & $-\frac{2}{3}$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{F}_j \, \mathcal{T}_i$ & $\overline{45}$ & $1$ & $-3$ & $-2$ & $2$ \\
$\rightarrow$ & $H_{24} \, \mathcal{F}_j$ & $\mathcal{T}_i \, \bar{H}_5$ & $\bar{5}$ & $1$ & $-\frac{3}{2}$ & $-\frac{3}{2}$ & $1$ \\
& $H_{24} \, \mathcal{F}_j$ & $\mathcal{T}_i \, \bar{H}_5$ & $\overline{45}$ & $1$ & $\frac{3}{2}$ & $-\frac{1}{2}$ & $-1$ \\
\midrule
& \multicolumn{2}{c}{$\mathcal{F}_j \, \mathcal{T}_i \, \bar{H}_{45}$} & --- & $1$ & $-3$ & $\sqrt{3}$ & $-\sqrt{3}$ \\
& $H_{24} \, \mathcal{T}_i$ & $\mathcal{F}_j \, \bar{H}_{45}$ & 10 & $1$ & $-18$ & $\sqrt{3}$ & $4 \sqrt{3}$ \\
& $H_{24} \, \mathcal{T}_i$ & $\mathcal{F}_j \, \bar{H}_{45}$ & 40 & $1$ &$ 0$ & $-\frac{\sqrt{3}}{2}$ & $-\frac{\sqrt{3}}{2}$ \\
& $H_{24} \, \mathcal{T}_i$ & $\mathcal{F}_j \, \bar{H}_{45}$ & 175 & $1$ & $\frac{36}{23}$ & $-\frac{19 \sqrt{3}}{23}$ & $-\frac{16 \sqrt{3}}{23}$ \\
& $H_{24} \, \bar{H}_{45}$ & $\mathcal{F}_j \, \mathcal{T}_i$ & $\bar{5}$ & $1$ & $1$ & $-\frac{2}{\sqrt{3}}$ & $-\frac{2}{\sqrt{3}}$ \\
& $H_{24} \, \bar{H}_{45}$ & $\mathcal{F}_j \, \mathcal{T}_i$ & $\overline{45}$ & $1$ & $-3$ & $x$ & $-x$ \\
& $H_{24} \, \mathcal{F}_j$ & $\mathcal{T}_i \, \bar{H}_{45}$ & $\bar{5}$ & $1$ & $\frac{9}{2}$ & $-\frac{3 \sqrt{3}}{2}$ & $-\sqrt{3}$ \\
& $H_{24} \, \mathcal{F}_j$ & $\mathcal{T}_i \, \bar{H}_{45}$ & $\overline{45}$ & $1$ & $-\frac{1}{2}$ & $-\frac{\sqrt{3}}{2}$ & $-\frac{1}{\sqrt{3}}$ \\
& $H_{24} \, \mathcal{F}_j$ & $\mathcal{T}_i \, \bar{H}_{45}$ & $\overline{70}$ & $1$ & $\frac{9}{4}$ & $-\frac{3 \sqrt{3}}{4}$ & $-\sqrt{3}$ \\
\bottomrule
\end{tabular}
\end{center}
\caption{$Y_{TF}$-like CG ratios for the dimension 4 operator and
effective dimension 5 operators $W \supset (A B)_R (C D)_{\bar{R}}$
(involving 5- and 45-dimensional Higgs fields) corresponding to the
the left diagram in figure \ref{fig:messengerdiagram}. Note that one
combination has a free parameter $x$ due to the ambiguity of the index
contraction. See main text for more details.}\label{tab:effectiveFToperators1}
\setlength{\extrarowheight}{0pt}
\end{table}
\begin{table}
\begin{center}
\setlength{\extrarowheight}{3pt}
\begin{tabular}{ccccc@{ : }c@{ : }c@{ : }c@{ : }c}
\toprule
&$A \, B$ & $C \, D$ & $R$ & $(Y_u)_{ij}$ & $(Y_u)_{ji}$ & $(Y_{qq})_{ij}$ & $(Y_{ue})_{ij}$ & $(Y_{ue})_{ji}$\\
\midrule
$\rightarrow$ & \multicolumn{2}{c}{$\mathcal{T}_i \mathcal{T}_j H_5$} & --- & $1$ & $1$ & $1$ & $1$ & $1$\\
&$H_{24} H_5$ & $\mathcal{T}_i \mathcal{T}_j$ & $5$ & $1$ & $1$ & $-\frac{2}{3}$ & $-\frac{2}{3}$ & $-\frac{2}{3}$\\
&$H_{24} H_5$ & $\mathcal{T}_i \mathcal{T}_j$ & $45$ & $1$ & $-1$ & $0$ & $-2$ & $2$\\
&$H_{24} \mathcal{T}_i$ & $\mathcal{T}_j H_5$ & $10$ & $1$ & $-4$ & $1$ & $-4$ & $6$\\
&$H_{24} \mathcal{T}_i$ & $\mathcal{T}_j H_5$ & $40$ & $1$ & $\frac{1}{2}$ & $-\frac{1}{2}$ & $-1$ & $0$\\
\midrule
& \multicolumn{2}{c}{$\mathcal{T}_i \mathcal{T}_j H_{45}$} & --- & $1$ & $-1$ & $0$ & $\sqrt{3}$ & $-\sqrt{3}$ \\
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j H_{45}$ & $10$ & $1$ & $4$ & $0$ & $-4 \sqrt{3}$ & $-6 \sqrt{3}$ \\
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j H_{45}$ & $15$ & $1$ & $0$ & $-\frac{\sqrt{3}}{2}$ & $0$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j H_{45}$ & $40$ & $1$ & $-\frac{7}{2}$ & $\frac{3 \sqrt{3}}{2}$ & $-\sqrt{3}$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j H_{45}$ & $175$ & $1$ & $\frac{16}{19}$ & $-\frac{21 \sqrt{3}}{38}$ & $-\frac{16 \sqrt{3}}{19}$ & $-\frac{12 \sqrt{3}}{19}$ \\
& $H_{24} H_{45}$ & $\mathcal{T}_i \mathcal{T}_j$ & $5$ & $1$ & $1$ & $-\frac{2}{\sqrt{3}}$ & $-\frac{2}{\sqrt{3}}$ & $-\frac{2}{\sqrt{3}}$ \\
& $H_{24} H_{45}$ & $\mathcal{T}_i \mathcal{T}_j$ & $45$ & $1$ & $-1$ & $0$ & $x$ & $-x$ \\
& $H_{24} H_{45}$ & $\mathcal{T}_i \mathcal{T}_j$ & $50$ & $0$ & $0$ & $1$ & $-2$ & $-2$ \\
\bottomrule
\end{tabular}
\end{center}
\caption{$Y_{TT}$-like CG ratios for the dimension 4 operator and
effective dimension 5 operators $W \supset (A B)_R (C D)_{\bar{R}}$
(involving 5- and 45-dimensional Higgs fields) corresponding to the
the left diagram in figure \ref{fig:messengerdiagram}. Note that one
combination has a free parameter $x$ due to the ambiguity of the
index contraction. See main text for more details.}\label{tab:effectiveTToperators1}
\setlength{\extrarowheight}{0pt}
\end{table}
\begin{table}
\begin{center}
\setlength{\extrarowheight}{3pt}
\begin{tabular}{cccccc@{ : }c@{ : }c@{ : }c}
\toprule
& $A \, B$ & $C$ & $D \, E$ & $R_1, R_2$ & $(Y_d)_{ij}$ & $(Y_e)_{ji}$ & $(Y_{ql})_{ij}$ & $(Y_{ud})_{ij}$ \\
\midrule
& $\mathcal{T}_i \, \bar{H}_5$ & $\mathcal{F}_j$ & $H_{24} \, H_{24}$ & 5, 1 & $1$ & $1$ & $1$ & $1$ \\
& $\mathcal{T}_i \, \bar{H}_5$ & $\mathcal{F}_j$ & $H_{24} \, H_{24}$ & 5, 24 & $1$ & $-\frac{3}{2}$ & $-\frac{3}{2}$ & $1$ \\
& $\mathcal{T}_i \, \bar{H}_5$ & $\mathcal{F}_j$ & $H_{24} \, H_{24}$ & 45, 24 & $1$ & $\frac{3}{2}$ & $-\frac{1}{2}$ & $-1$ \\
& $\mathcal{T}_i \, \bar{H}_5$ & $\mathcal{F}_j$ & $H_{24} \, H_{24}$ & 45, 75 & $1$ & $-3$ & $1$ & $-1$ \\
& $H_{24} \, \bar{H}_5$ & $H_{24}$ & $\mathcal{F}_j \, \mathcal{T}_i$ & $\bar{5}$, 5 & $1$ & $1$ & $\frac{4}{9}$ & $\frac{4}{9}$ \\
& $H_{24} \, \bar{H}_5$ & $H_{24}$ & $\mathcal{F}_j \, \mathcal{T}_i$ & $\bar{5}$, 45 & $1$ & $-3$ & $\frac{4}{3}$ & $-\frac{4}{3}$ \\
& $H_{24} \, \bar{H}_5$ & $H_{24}$ & $\mathcal{F}_j \, \mathcal{T}_i$ & $\overline{45}$, 5 & $1$ & $1$ & $\frac{4}{3}$ & $\frac{4}{3}$ \\
& $H_{24} \, \bar{H}_5$ & $H_{24}$ & $\mathcal{F}_j \, \mathcal{T}_i$ & $\overline{45}$, 45 & $1$ & $-3$ & $x$ & $-x$ \\
& $H_{24} \, \bar{H}_5$ & $H_{24}$ & $\mathcal{F}_j \, \mathcal{T}_i$ & $\overline{70}$, 5 & $1$ & $1$ & $\frac{8}{9}$ & $\frac{8}{9}$ \\
& $H_{24} \, \bar{H}_5$ & $H_{24}$ & $\mathcal{F}_j \, \mathcal{T}_i$ & $\overline{70}$, 45 & $1$ & $-3$ & $\frac{8}{3}$ & $-\frac{8}{3}$ \\
& $\mathcal{F}_j \, \mathcal{T}_i$ & $\bar{H}_5$ & $H_{24} \, H_{24}$ & 5, 1 & $1$ & $1$ & $1$ & $1$ \\
& $\mathcal{F}_j \, \mathcal{T}_i$ & $\bar{H}_5$ & $H_{24} \, H_{24}$ & 5, 24 & $1$ & $1$ & $-\frac{2}{3}$ & $-\frac{2}{3}$ \\
& $\mathcal{F}_j \, \mathcal{T}_i$ & $\bar{H}_5$ & $H_{24} \, H_{24}$ & 45, 24 & $1$ & $-3$ & $-2$ & $2$ \\
& $\mathcal{F}_j \, \mathcal{T}_i$ & $\bar{H}_5$ & $H_{24} \, H_{24}$ & 45, 75 & $1$ & $-3$ & $1$ & $-1$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{T}_i$ & $H_{24} \, \mathcal{F}_j$ & $\bar{5}$, $\bar{5}$ & $1$ & $-\frac{3}{2}$ & $1$ & $-\frac{2}{3}$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{T}_i$ & $H_{24} \, \mathcal{F}_j$ & $\bar{5}$, $\overline{45}$ & $1$ & $\frac{3}{2}$ & $\frac{1}{3}$ & $\frac{2}{3}$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{T}_i$ & $H_{24} \, \mathcal{F}_j$ & $\overline{45}$, $\bar{5}$ & $1$ & $\frac{9}{2}$ & $3$ & $2$ \\
$\rightarrow$ & $H_{24} \, \bar{H}_5$ & $\mathcal{T}_i$ & $H_{24} \, \mathcal{F}_j$ & $\overline{45}$, $\overline{45}$ & $1$ & $-\frac{1}{2}$ & $1$ & $\frac{2}{3}$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{T}_i$ & $H_{24} \, \mathcal{F}_j$ & $\overline{45}$, $\overline{70}$ & $1$ & $\frac{9}{4}$ & $\frac{3}{2}$ & $2$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{T}_i$ & $H_{24} \, \mathcal{F}_j$ & $\overline{70}$, $\overline{45}$ & $1$ & $\frac{3}{2}$ & $\frac{2}{3}$ & $\frac{4}{3}$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{T}_i$ & $H_{24} \, \mathcal{F}_j$ & $\overline{70}$, $\overline{70}$ & $1$ & $\frac{3}{4}$ & $1$ & $\frac{2}{3}$ \\
& $\mathcal{F}_j \, \bar{H}_5$ & $H_{24}$ & $H_{24} \, \mathcal{T}_i$ & $\overline{10}$, 10 & $1$ & $36$ & $1$ & $16$ \\
& $\mathcal{F}_j \, \bar{H}_5$ & $H_{24}$ & $H_{24} \, \mathcal{T}_i$ & $\overline{10}$, 15 & $1$ & $0$ & $1$ & $0$ \\
& $\mathcal{F}_j \, \bar{H}_5$ & $H_{24}$ & $H_{24} \, \mathcal{T}_i$ & $\overline{10}$, 40 & $1$ & $0$ & $1$ & $1$ \\
& $\mathcal{F}_j \, \bar{H}_5$ & $H_{24}$ & $H_{24} \, \mathcal{T}_i$ & $\overline{10}$, 175 & $1$ & $\frac{72}{61}$ & $1$ & $\frac{64}{61}$ \\
& $\mathcal{F}_j \, \bar{H}_5$ & $H_{24}$ & $H_{24} \, \mathcal{T}_i$ & $\overline{15}$, 10 & $1$ & $0$ & $-1$ & $0$ \\
& $\mathcal{F}_j \, \bar{H}_5$ & $H_{24}$ & $H_{24} \, \mathcal{T}_i$ & $\overline{15}$, 15 & $1$ & $0$ & $-1$ & $0$ \\
& $\mathcal{F}_j \, \bar{H}_5$ & $H_{24}$ & $H_{24} \, \mathcal{T}_i$ & $\overline{15}$, 175 & $1$ & $0$ & $-1$ & $0$ \\
\bottomrule
\end{tabular}
\end{center}
\caption{$Y_{TF}$-like CG ratios for the effective dimension 6 operators $W \supset (A B)_{R_1} C (D E)_{R_2}$ corresponding to the the right diagram in figure \ref{fig:messengerdiagram}. Note one combination has a free parameter $x$.
See main text for more details.
}\label{tab:effectiveFToperators2}
\setlength{\extrarowheight}{0pt}
\end{table}
\begin{table}
\begin{center}
\setlength{\extrarowheight}{3pt}
\begin{tabular}{cccccc@{ : }c@{ : }c@{ : }c}
\toprule
& $A \, B$ & $C$ & $D \, E$ & $R_1, R_2$ & $(Y_d)_{ij}$ & $(Y_e)_{ji}$ & $(Y_{ql})_{ij}$ & $(Y_{ud})_{ij}$ \\
\midrule
& $H_{24} \, \mathcal{F}_j$ & $H_{24}$ & $\mathcal{T}_i \, \bar{H}_5$ & $\bar{5}$, 5 & $1$ & $\frac{9}{4}$ & $\frac{9}{4}$ & $1$ \\
& $H_{24} \, \mathcal{F}_j$ & $H_{24}$ & $\mathcal{T}_i \, \bar{H}_5$ & $\bar{5}$, 45 & $1$ & $-\frac{9}{4}$ & $\frac{3}{4}$ & $-1$ \\
& $H_{24} \, \mathcal{F}_j$ & $H_{24}$ & $\mathcal{T}_i \, \bar{H}_5$ & $\overline{45}$, 5 & $1$ & $\frac{3}{4}$ & $\frac{3}{4}$ & $1$ \\
& $H_{24} \, \mathcal{F}_j$ & $H_{24}$ & $\mathcal{T}_i \, \bar{H}_5$ & $\overline{45}$, 45 & $1$ & $x$ & $-\frac{x}{3}$ & $-1$ \\
& $H_{24} \, \mathcal{F}_j$ & $H_{24}$ & $\mathcal{T}_i \, \bar{H}_5$ & $\overline{70}$, 5 & $1$ & $\frac{9}{8}$ & $\frac{9}{8}$ & $1$ \\
& $H_{24} \, \mathcal{F}_j$ & $H_{24}$ & $\mathcal{T}_i \, \bar{H}_5$ & $\overline{70}$, 45 & $1$ & $-\frac{9}{8}$ & $\frac{3}{8}$ & $-1$ \\
& $H_{24} \, \mathcal{F}_j$ & $\bar{H}_5$ & $H_{24} \, \mathcal{T}_i$ & $\bar{5}$, 10 & $1$ & $-9$ & $-\frac{3}{2}$ & $-4$ \\
& $H_{24} \, \mathcal{F}_j$ & $\bar{H}_5$ & $H_{24} \, \mathcal{T}_i$ & $\bar{5}$, 15 & $1$ & $0$ & $\frac{3}{2}$ & $0$ \\
& $H_{24} \, \mathcal{F}_j$ & $\bar{H}_5$ & $H_{24} \, \mathcal{T}_i$ & $\overline{45}$, 10 & $1$ & $9$ & $-\frac{1}{2}$ & $4$ \\
& $H_{24} \, \mathcal{F}_j$ & $\bar{H}_5$ & $H_{24} \, \mathcal{T}_i$ & $\overline{45}$, 40 & $1$ & $0$ & $1$ & $1$ \\
& $H_{24} \, \mathcal{F}_j$ & $\bar{H}_5$ & $H_{24} \, \mathcal{T}_i$ & $\overline{45}$, 175 & $1$ & $\frac{18}{19}$ & $\frac{23}{38}$ & $\frac{16}{19}$ \\
& $H_{24} \, \mathcal{F}_j$ & $\bar{H}_5$ & $H_{24} \, \mathcal{T}_i$ & $\overline{70}$, 15 & $1$ & $0$ & $\frac{3}{4}$ & $0$ \\
& $H_{24} \, \mathcal{F}_j$ & $\bar{H}_5$ & $H_{24} \, \mathcal{T}_i$ & $\overline{70}$, 175 & $1$ & $\frac{9}{7}$ & $\frac{33}{28}$ & $\frac{8}{7}$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{F}_j$ & $H_{24} \, \mathcal{T}_i$ & $\bar{5}$, 10 & $1$ & $6$ & $-\frac{2}{3}$ & $\frac{8}{3}$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{F}_j$ & $H_{24} \, \mathcal{T}_i$ & $\bar{5}$, 15 & $1$ & $0$ & $\frac{2}{3}$ & $0$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{F}_j$ & $H_{24} \, \mathcal{T}_i$ & $\overline{45}$, 10 & $1$ & $-18$ & $-2$ & $-8$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{F}_j$ & $H_{24} \, \mathcal{T}_i$ & $\overline{45}$, 40 & $1$ & $0$ & $1$ & $1$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{F}_j$ & $H_{24} \, \mathcal{T}_i$ & $\overline{45}$, 175 & $1$ & $\frac{36}{23}$ & $\frac{38}{23}$ & $\frac{32}{23}$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{F}_j$ & $H_{24} \, \mathcal{T}_i$ & $\overline{70}$, 15 & $1$ & $0$ & $\frac{4}{3}$ & $0$ \\
& $H_{24} \, \bar{H}_5$ & $\mathcal{F}_j$ & $H_{24} \, \mathcal{T}_i$ & $\overline{70}$, 175 & $1$ & $\frac{12}{11}$ & $\frac{28}{33}$ & $\frac{32}{33}$ \\
& $\mathcal{F}_j \, \bar{H}_5$ & $\mathcal{T}_i$ & $H_{24} \, H_{24}$ & $\overline{10}$, 1 & $1$ & $1$ & $1$ & $1$ \\
& $\mathcal{F}_j \, \bar{H}_5$ & $\mathcal{T}_i$ & $H_{24} \, H_{24}$ & $\overline{10}$, 24 & $1$ & $6$ & $1$ & $-4$ \\
& $\mathcal{F}_j \, \bar{H}_5$ & $\mathcal{T}_i$ & $H_{24} \, H_{24}$ & $\overline{10}$, 75 & $1$ & $-3$ & $1$ & $-1$ \\
& $\mathcal{F}_j \, \bar{H}_5$ & $\mathcal{T}_i$ & $H_{24} \, H_{24}$ & $\overline{15}$, 24 & $1$ & $0$ & $-1$ & $0$ \\
\bottomrule
\end{tabular}
\end{center}
\caption{Continuation of table \ref{tab:effectiveFToperators2}: $Y_{TF}$-like CG ratios for the effective dimension 6 operators $W \supset (A B)_{R_1} C (D E)_{R_2}$ corresponding to the the right diagram in figure \ref{fig:messengerdiagram}. Note another combination with a free parameter $x$. See main text for more details.}\label{tab:effectiveFToperators3}
\setlength{\extrarowheight}{0pt}
\end{table}
\begin{table}
\begin{center}
\setlength{\extrarowheight}{3pt}
\begin{tabular}{cccccc@{ : }c@{ : }c@{ : }c@{ : }c}
\toprule
& $A \, B$ & $C$ & $D \, E$ & $R_1,R_2$ & $(Y_u)_{ij}$ & $(Y_u)_{ji}$ & $(Y_{qq})_{ij}$ & $(Y_{ue})_{ij}$ & $(Y_{ue})_{ji}$\\
\midrule
& $H_{24} \mathcal{T}_i$ & $H_ 5$ & $H_{24} \mathcal{T}_j$ & $10, 10$ & $1$ & $1$ & $-\frac{1}{4}$ & $6$ & $6$ \\
& $H_{24} \mathcal{T}_i$ & $H_ 5$ & $H_{24} \mathcal{T}_j$ & $10, 40$ & $1$ & $-8$ & $-1$ & $0$ & $-12$ \\
& $H_{24} \mathcal{T}_i$ & $H_ 5$ & $H_{24} \mathcal{T}_j$ & $15, 40$ & $1$ & $0$ & $1$ & $0$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $H_ 5$ & $H_{24} \mathcal{T}_j$ & $40, 10$ & $1$ & $-\frac{1}{8}$ & $\frac{1}{8}$ & $\frac{3}{2}$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $H_ 5$ & $H_{24} \mathcal{T}_j$ & $40, 15$ & $0$ & $1$ & $1$ & $0$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $H_ 5$ & $H_{24} \mathcal{T}_j$ & $40, 175$ & $1$ & $\frac{23}{32}$ & $\frac{19}{32}$ & $\frac{3}{4}$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $H_ 5$ & $H_{24} \mathcal{T}_j$ & $175, 40$ & $1$ & $\frac{32}{23}$ & $\frac{19}{23}$ & $0$ & $\frac{24}{23}$ \\
& $H_{24} \mathcal{T}_i$ & $H_ 5$ & $H_{24} \mathcal{T}_j$ & $175, 175$ & $1$ & $1$ & $\frac{41}{40}$ & $\frac{6}{5}$ & $\frac{6}{5}$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_ 5$ & $H_{24} H_{24}$ & $\bar{5}, 1$ & $1$ & $1$ & $1$ & $1$ & $1$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_ 5$ & $H_{24} H_{24}$ & $\bar{5}, 24$ & $1$ & $1$ & $-\frac{2}{3}$ & $-\frac{2}{3}$ & $-\frac{2}{3}$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_ 5$ & $H_{24} H_{24}$ & $\overline{45}, 24$ & $1$ & $-1$ & $0$ & $-2$ & $2$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_ 5$ & $H_{24} H_{24}$ & $\overline{45}, 75$ & $1$ & $-1$ & $0$ & $1$ & $-1$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_ 5$ & $H_{24} H_{24}$ & $\overline{50}, 75$ & $0$ & $0$ & $1$ & $-2$ & $-2$ \\
& $\mathcal{T}_i H_ 5$ & $\mathcal{T}_j$ & $H_{24} H_{24}$ & $\overline{10}, 1$ & $1$ & $1$ & $1$ & $1$ & $1$ \\
& $\mathcal{T}_i H_ 5$ & $\mathcal{T}_j$ & $H_{24} H_{24}$ & $\overline{10}, 24$ & $1$ & $-\frac{1}{4}$ & $-\frac{1}{4}$ & $-\frac{3}{2}$ & $1$ \\
& $\mathcal{T}_i H_ 5$ & $\mathcal{T}_j$ & $H_{24} H_{24}$ & $\overline{10}, 75$ & $1$ & $-1$ & $-1$ & $3$ & $1$ \\
& $\mathcal{T}_i H_ 5$ & $\mathcal{T}_j$ & $H_{24} H_{24}$ & $\overline{40}, 24$ & $1$ & $2$ & $-1$ & $0$ & $-2$ \\
& $\mathcal{T}_i H_ 5$ & $\mathcal{T}_j$ & $H_{24} H_{24}$ & $\overline{40}, 75$ & $1$ & $-1$ & $\frac{1}{2}$ & $0$ & $-2$ \\
\bottomrule
\end{tabular}
\end{center}
\caption{$Y_{TT}$-like CG ratios for the effective dimension 6 operators $W \supset (A B)_{R_1} C (D E)_{R_2}$ corresponding to the the right diagram in figure \ref{fig:messengerdiagram}. }\label{tab:effectiveTToperators2}
\setlength{\extrarowheight}{0pt}
\end{table}
\begin{table}
\begin{center}
\setlength{\extrarowheight}{3pt}
\begin{tabular}{cccccc@{ : }c@{ : }c@{ : }c@{ : }c}
\toprule
& $A \, B$ & $C$ & $D \, E$ & $R_1,R_2$ & $(Y_u)_{ij}$ & $(Y_u)_{ji}$ & $(Y_{qq})_{ij}$ & $(Y_{ue})_{ij}$ & $(Y_{ue})_{ji}$\\
\midrule
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j$ & $H_{24} H_ 5$ & $10, 5$ & $1$ & $-4$ & $-\frac{2}{3}$ & $\frac{8}{3}$ & $-4$ \\
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j$ & $H_{24} H_ 5$ & $10, 45$ & $1$ & $4$ & $0$ & $8$ & $12$ \\
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j$ & $H_{24} H_ 5$ & $15, 45$ & $1$ & $0$ & $1$ & $0$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j$ & $H_{24} H_ 5$ & $40, 5$ & $1$ & $\frac{1}{2}$ & $\frac{1}{3}$ & $\frac{2}{3}$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j$ & $H_{24} H_ 5$ & $40, 45$ & $1$ & $-\frac{7}{2}$ & $-3$ & $2$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j$ & $H_{24} H_ 5$ & $40, 70$ & $1$ & $\frac{1}{2}$ & $\frac{2}{3}$ & $\frac{4}{3}$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j$ & $H_{24} H_ 5$ & $175, 45$ & $1$ & $\frac{16}{19}$ & $\frac{21}{19}$ & $\frac{32}{19}$ & $\frac{24}{19}$ \\
& $H_{24} \mathcal{T}_i$ & $\mathcal{T}_j$ & $H_{24} H_ 5$ & $175, 70$ & $1$ & $\frac{8}{7}$ & $\frac{20}{21}$ & $\frac{16}{21}$ & $\frac{8}{7}$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_{24}$ & $H_{24} H_ 5$ & $\bar{5}, 5$ & $1$ & $1$ & $\frac{4}{9}$ & $\frac{4}{9}$ & $\frac{4}{9}$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_{24}$ & $H_{24} H_ 5$ & $\bar{5}, 45$ & $1$ & $1$ & $\frac{4}{3}$ & $\frac{4}{3}$ & $\frac{4}{3}$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_{24}$ & $H_{24} H_ 5$ & $\bar{5}, 70$ & $1$ & $1$ & $\frac{8}{9}$ & $\frac{8}{9}$ & $\frac{8}{9}$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_{24}$ & $H_{24} H_ 5$ & $\overline{45}, 5$ & $1$ & $-1$ & $0$ & $\frac{4}{3}$ & $-\frac{4}{3}$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_{24}$ & $H_{24} H_ 5$ & $\overline{45}, 45$ & $1$ & $-1$ & $0$ & $x$ & $-x$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_{24}$ & $H_{24} H_ 5$ & $\overline{45}, 70$ & $1$ & $-1$ & $0$ & $\frac{8}{3}$ & $-\frac{8}{3}$ \\
& $\mathcal{T}_i \mathcal{T}_j$ & $H_{24}$ & $H_{24} H_ 5$ & $\overline{50}, 45$ & $0$ & $0$ & $1$ & $-2$ & $-2$ \\
& $H_{24} \mathcal{T}_i$ & $H_{24}$ & $\mathcal{T}_j H_ 5$ & $10, \overline{10}$ & $1$ & $16$ & $1$ & $16$ & $36$ \\
& $H_{24} \mathcal{T}_i$ & $H_{24}$ & $\mathcal{T}_j H_ 5$ & $10, \overline{40}$ & $1$ & $-2$ & $-\frac{1}{2}$ & $4$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $H_{24}$ & $\mathcal{T}_j H_ 5$ & $15, \overline{10}$ & $1$ & $0$ & $1$ & $0$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $H_{24}$ & $\mathcal{T}_j H_ 5$ & $40, \overline{10}$ & $1$ & $1$ & $1$ & $1$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $H_{24}$ & $\mathcal{T}_j H_ 5$ & $40, \overline{40}$ & $1$ & $x$ & $-\frac{1}{2}$ & $-2 x$ & $0$ \\
& $H_{24} \mathcal{T}_i$ & $H_{24}$ & $\mathcal{T}_j H_ 5$ & $175, \overline{10}$ & $1$ & $\frac{64}{61}$ & $1$ & $\frac{64}{61}$ & $\frac{72}{61}$ \\
& $H_{24} \mathcal{T}_i$ & $H_{24}$ & $\mathcal{T}_j H_ 5$ & $175, \overline{40}$ & $1$ & $4$ & $-\frac{1}{2}$ & $-8$ & $0$ \\
\bottomrule
\end{tabular}
\end{center}
\caption{Continuation of table \ref{tab:effectiveTToperators2}: $Y_{TT}$-like CG ratios for the effective dimension 6 operators $W \supset (A B)_{R_1} C (D E)_{R_2}$ corresponding to the the right diagram in figure \ref{fig:messengerdiagram}. Note two combinations have a free parameter $x$. See main text for more details.}\label{tab:effectiveTToperators3}
\setlength{\extrarowheight}{0pt}
\end{table}
If the considered model contains Higgs fields in 5- and 45-dimensional representations
\footnote{One could also imagine using 45-dimensional Higgs fields exclusively. However, this severely exacerbates the doublet-triplet splitting problem as now one has to split the doublets from even more component fields that can generate proton decay operators.},
there are two Higgs doublet pairs in the spectrum and
care should be taken that unification is still possible. One solution is mixing both and making one linear combination heavy while one stays at the electroweak scale. The simplest term generating such a mixing is
\begin{equation}\label{eq:5to45mixing}
W \supset H_{24} H_5 \bar{H}_{45}
\propto H_u H_d^{\overline{45}} - \frac{2}{\sqrt{3}} T \bar{T}^{\overline{45}} \;,
\end{equation}
where we suppressed any additional MSSM multiplets in $\bar{H}_{45}$.
Since a $\overline{\mathbf{45}}$ contains more potentially dangerous MSSM multiplets,
it is natural to have the heavy linear combination be predominantly in the $\overline{\mathbf{45}}$.
Then it is possible to treat $\bar{H}_{45}$ like a messenger field and the
renormalizable operator $\mathcal{F} \mathcal{T} \bar{H}_{45}$ turns into the non-renormalizable
operator $(\mathcal{F} \mathcal{T})_{45} (H_{24} \bar{H}_5)_{{\overline{45}}}$, cf.\ table \ref{tab:effectiveFToperators1}.
Analogous limits can be deduced trivially.
If the approximation $m_{45} \gg \vev{H_{24}}$ does not hold, one has to
take into account the full mass matrix for the Higgs doublets including the term in eq.~\eqref{eq:5to45mixing}.
\section{Two-loop RGEs in Extensions to the MSSM \label{sec:2loopRGEs}}
\def1{1.2}
The renormalization group equations for gauge couplings at two-loop are given by \cite{Jones:1974pg+1983vk}
\begin{equation}
\mu \frac{d}{d\mu} g_a= \frac{g_a^3}{16 \pi^2} b_a + \frac{g_a^3}{(16\pi^2)^2} \left(
\sum\limits_{b=1}^3 B_{ab} g_b^2 - \sum\limits_{f} C_a^f \mathop{\rm tr}(Y_f^\dagger Y_f)
\right)\;,
\end{equation}
in the $\overline{\text{DR}}$ renormalization scheme, where $\mu$ is the renormalization
scale and $f$ runs over all Yukawa coupling matrices. In the MSSM, the beta function coefficients
are given by (in GUT normalisation for $g_1$)
\begin{equation}
b_a = \begin{pmatrix} \frac{33}{5} \\ 1 \\ -3 \end{pmatrix}, \;
B_{ab} = \begin{pmatrix}
\frac{199}{25} & \frac{27}{5} & \frac{88}{5} \\
\frac{9}{5} & 25 & 24 \\
\frac{11}{5} & 9 & 14 \\
\end{pmatrix} \;,
\end{equation}
and
\begin{equation}
C_a^{u,d,e} = \begin{pmatrix}
\frac{26}{5} & \frac{14}{5} & \frac{18}{5} \\
6 & 6 & 2 \\
4 & 4 & 0 \\
\end{pmatrix} \;.
\end{equation}
where the first column stands for $u$, the second column for $d$
and the third for $e$. Additional colour triplet and weak doublet pairs,
as those contained in $\mathbf{5}$, $\bar{\mathbf{5}}$ representations,
contribute (per pair) at one-loop with
\begin{equation}
b^{(5,T)}_a = \begin{pmatrix} \frac{2}{5} \\ 0 \\ 1 \end{pmatrix}, \;
b^{(5,D)}_a = \begin{pmatrix} \frac{3}{5} \\ 1 \\ 0 \end{pmatrix}, \;
\end{equation}
and at two-loop with
\begin{equation}
B^{(5,T)}_{ab} = \begin{pmatrix}
\frac{8}{75} & 0 & \frac{32}{15} \\
0 & 0 & 0 \\
\frac{4}{15} & 0 & \frac{34}{3}
\end{pmatrix}, \;
B^{(5,D)}_{ab} = \begin{pmatrix}
\frac{9}{25} & \frac{9}{5} & 0 \\
\frac{3}{5} & 7 & 0 \\
0 & 0 & 0 \\
\end{pmatrix} \;,
\end{equation}
SU(2) triplets, SU(3) octets and leptoquark superfields\footnote{Note that leptoquark superfields can only appear in Dirac pairs due to their charges.} from an adjoint contribute
(per chiral superfield) at one-loop with
\begin{equation}
b^{(24,T)}_a = \begin{pmatrix} 0 \\ 2 \\ 0 \end{pmatrix}, \;
b^{(24,O)}_a = \begin{pmatrix} 0 \\ 0 \\ 3 \end{pmatrix}, \;
b^{(24,L)}_a = \begin{pmatrix} \frac{5}{2} \\ \frac{3}{2} \\ 1 \end{pmatrix}, \;
\end{equation}
and at two-loop with
\begin{equation}
B^{(24,T)}_{ab} = \begin{pmatrix}
0 & 0 & 0 \\
0 & 24 & 0 \\
0 & 0 & 0
\end{pmatrix} ,\;
B^{(24,O)}_{ab} = \begin{pmatrix}
0 & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & 54
\end{pmatrix} ,\;
B^{(24,L)}_{ab} = \begin{pmatrix}
\frac{25}{6} & \frac{15}{2} & \frac{40}{3} \\
\frac{5}{2} & \frac{21}{2} & 8 \\
\frac{5}{3} & 3 & \frac{34}{3}
\end{pmatrix} \;.
\end{equation}
Additional Yukawa couplings between SM fermion superfields and a pair of
colour triplet/anti-triplet contribute with
\begin{equation}
C_a^{qq,ue,ql,ud} = \begin{pmatrix}
\frac{6}{5} & \frac{28}{5} & \frac{14}{5} & \frac{24}{5} \\
6 & 0 & 6 & 0 \\
6 & 2 & 4 & 6
\end{pmatrix} \;.
\end{equation}
In our numerical analysis we have assumed that $Y_{qq} = Y_{ue} = Y_u$
and $Y_{ql} = Y_{ud} = Y_d$ (as motivated by minimal SU(5)). We have checked
that this approximation changes our results only negligibly.
\def1{1}
\section{Discussion of messenger fields}
\label{app:mess}
We discuss now the messenger fields appearing in Figure~\ref{fig:optionA_Yu} of the model presented in section~\ref{sec:modelA}. There are three supergraphs generating the superpotential term $y_{12}$. The renormalizable superpotential corresponding to Figure~\ref{fig:optionA_Yu} is
\begin{align}
W \ \supset\ & \gamma_1 \, H_5 \mathcal{T}_1 Z_{10,4} + \gamma_2 \, H_5 \mathcal{T}_2 Z_{10,3} + \gamma_3 \, \mathcal{T}_3 \theta_1 \bar Z_{10,3} + \gamma_\theta \, \theta_1^2 Z_1\nonumber \\
+\ & \lambda_1 \, \theta_2 Z_2 \bar Z_1 + \lambda_{10} \, \theta_2 Z_{10,3} \bar Z_{10,4} \nonumber \\
+\ & \eta_1 \, \mathcal{T}_1 \bar Z_{10,3} \bar Z_2 + \eta_2 \, \mathcal{T}_2 \bar Z_{10,4} \bar Z_2 + \eta'_2 \, \mathcal{T}_2 \bar Z_{10,3} \bar Z_1 \nonumber \\
+\ & M_1 \, Z_1\bar Z_1 + M_2 \, Z_2\bar Z_2 + M_{10,3} \, Z_{10,3}\bar Z_{10,3} + M_{10,4} \, Z_{10,4}\bar Z_{10,4}\;,
\end{align}
where we explicitly denote the coupling constants at the ends of a diagram with $\gamma_x$,
coupling constants in the middle of the diagrams with $\lambda_x$ if they involve $\theta_j$ and $\eta_i$ if they involve $\mathcal{T}_i$ and messenger masses with $M_x$.
After
$Z_x$ and $\bar Z_x$ are integrated out and the flavons $\theta_i$ obtain VEVs, the elements of the up-type Yukawa matrix $Y_u$ of eq.~\eqref{eq:A_Y} are given by
\begin{align}
y_{11} &= \gamma_1 \, \lambda_{10} \, \eta_1 \, \lambda_1 \, \gamma_\theta \, \frac{\vev{\theta_1}^2 \vev{\theta_2}^2}{M_1 M_2 M_{10,3} M_{10,4}} \;,\nonumber \\
y_{12} &= \gamma_2 \, \eta_1 \, \lambda_1 \, \gamma_\theta \, \frac{\vev{\theta_1}^2 \vev{\theta_2}}{M_1 M_2 M_{10,3}} + \gamma_1 \, \eta_2 \, \lambda_1 \, \gamma_\theta \, \frac{\vev{\theta_1}^2 \vev{\theta_2}}{M_1 M_2 M_{10,4}} + \gamma_1 \, \lambda_{10} \, \eta'_2 \, \gamma_\theta \, \frac{\vev{\theta_1}^2 \vev{\theta_2}}{M_1 M_{10,3} M_{10,4}} \;,\nonumber \\
y_{22} &= \gamma_2 \, \eta'_2 \, \gamma_\theta \frac{\vev{\theta_1}^2}{M_1 M_{10,3}} \;,\nonumber \\
y_{13} &= \gamma_1 \, \lambda_{10} \, \gamma_3 \, \frac{\vev{\theta_1}\vev{\theta_2}}{M_{10,3} M_{10,4}} \;, \nonumber \\
y_{23} &= \gamma_2 \, \gamma_3 \, \frac{\vev{\theta_1}}{M_{10,3}}\;,
\end{align}
and $y_{33}$ is a renormalizable Yukawa coupling coefficient.
Removing the messenger pair $Z_2$, $\bar Z_2$ from the spectrum eliminates two
supergraphs and thus the first two terms contributing to $y_{12}$.\footnote{A different pair of fields $Z_2'$, $\bar Z_2'$ would need to be introduced anyway (using different charge assignment than $Z_2$,$\bar Z_2$) in order to generate $y_{11}$. The charges can be assigned such that no extra contributions to $y_{12}$ appear.} However, without $Z_2 \bar Z_2$, evaluating $y_{12}$ yields
\begin{align}
y_{12} & = \gamma_1 \, \lambda_{10} \, \eta'_2 \, \gamma_\theta \, \frac{\vev{\theta_1}^2\vev{\theta_2}}{M_1 M_{10,3} M_{10,4}} = y_{13} \, \frac{\vev{\theta_1}}{M_1} \, \frac{\eta'_2\, \gamma_\theta}{\gamma_3} \nonumber \\
& = y_{13} \, y_{22} \, \frac{M_{10,3}}{\vev{\theta_1}}\frac{1}{\gamma_2\gamma_3} = \frac{y_{13}y_{22}}{y_{23}}\;.
\end{align}
This relation is not phenomenologically viable as it would imply $\theta_C = \theta_{13} / \theta_{23}$. To fit $Y_u$ to the observed data an additional degree of freedom is needed. In our model this is realised through $Z_2\bar Z_2$ enabling additional diagrams contributing to $y_{12}$.
\section{Simultaneous R-symmetry and GUT breaking}
\label{app:SR}
It was shown in the literature that the MSSM with an additional R-symmetry cannot
be obtained from the spontaneous breaking of a four-dimensional (SUSY) GUT
\cite{Fallbacher:2011xg}. On the other hand, in flavour models, R-symmetries
are often used in superpotentials that generate the required VEVs for the family symmetry
breaking Higgs
fields (i.e. the flavons) as well as for spontaneous breaking of CP. In the following we
present some simple examples that show that it is possible - in particular with discrete
R-symmetries - to simultaneously break the GUT gauge group and the R-symmetry.
Without an R-symmetry below the Planck scale, there is no conflict to the statement of
\cite{Fallbacher:2011xg}. We will also illustrate that GUT flavour models can rely on
such a discrete R-symmetry for the flavon VEV alignment, such that our setup can be
used to construct flavour models with non-Abelian family symmetries and spontaneous
CP violation.
\paragraph{Simple Example}
Consider as example a discrete $\mathbb{Z}_{4}^R$ R-symmetry under which
the superfield $S$ is charged and the superfield $H$ is uncharged (for the
beginning we assume them to be SU(5) singlets). The fields
are also charged under an additional conventional $\mathbb{Z}_4$ symmetry
with charges $1$ and $3$ respectively. We consider only the
required lowest order superpotential terms:
\begin{equation}
W_R =\mu_1 S H + \lambda_1 S^3 H^3/M_{\text{Pl}}^3 + \lambda_2 S H^5/M_{\text{Pl}}^3 + \lambda_3 S^5 H/M_{\text{Pl}}^3\,.
\end{equation}
The F-terms for this simple example lead to the conditions
\begin{align}
\mu_1 H + 3 \lambda_1 S^2 H^3/M_{\text{Pl}}^3 + \lambda_2 H^5/M_{\text{Pl}}^3 + 5 \lambda_3 S^4 H/M_{\text{Pl}}^3 =& 0 \\
\mu_1 S + 3 \lambda_1 S^3 H^2/M_{\text{Pl}}^3 + 5 \lambda_2 S H^4/M_{\text{Pl}}^3 + \lambda_3 S^5/M_{\text{Pl}}^3 =& 0 \,.
\end{align}
These equations have several solutions but here we are only interested in the
non-trivial solution $\vev{S}^4 = (\lambda_2/\lambda_3) \langle H \rangle^4$ and
$\langle H \rangle^4 = -\mu_1 M_{\text{Pl}}^3/(3\lambda_1 \sqrt{\lambda_2/\lambda_3} + 6 \lambda_2)$.
If we assign $S$ and $H$ under SU(5) as adjoints (similarly to
superpotential (c) where we had the two fields $H_{24}$ and $H'_{24}$)
$W_R$ is SU(5) invariant (by taking the appropriate contractions) and
importantly, the R-symmetry and SU(5) are simultaneously broken by the
non-trivial VEV configuration. We can achieve a phenomenological viable model by breaking a discrete R-symmetry at the GUT scale.
In general, leaving additional symmetry (or symmetries) unspecified, $S$ may
be a singlet of the GUT symmetry group and we denote the (polynomial) functions of the superfield
$H$ that make the respective terms invariant as $A(H)$, $C(H)$ (which include
the associated couplings and $M_{\text{Pl}}$ suppressions). We write
\begin{equation}
W_R = S A(H) + S^3 C(H) \,,
\end{equation}
leading to the F-term equations
\begin{align}
A(H) + 3 \, C(H) S^2 &= 0 \;,\\
\frac{\text{d}A(H)}{\text{d}H} S + \frac{\text{d}C(H)}{\text{d}H} S^3 &= 0 \,.
\end{align}
The generalised solution breaking the R-symmetry and the GUT gauge group is then
\begin{align}
\vev{S}^2 &= - \VEV{\frac{ A(H)}{ 3 \, C(H)}} \;, \\
\VEV{\frac{\text{d}A(H)}{\text{d}H}} &= \VEV{\frac{A(H)}{ 3 \, C(H)} \frac{ \text{d}C(H)}{\text{d}H}} \,.
\end{align}
Note the second equation, $\tfrac{\text{d}A}{A}=\tfrac{\text{d}C}{3 C}$, constrains the allowed
functions $A(H)$ and $C(H)$ which indirectly imposes conditions on the unspecified
additional symmetries.
\paragraph{Generalisations}
While the R-symmetry must be discrete for the crucial interplay between two
terms, it needs not be a $\mathbb{Z}_4^R$ and generalising to other
$\mathbb{Z}_N^R$ symmetries is straightforward
\begin{equation}
W_R^N = S A(H) + S^{N-1} C(H) \,.
\end{equation}
This can be further generalised to multiple fields. Consider, for instance,
$S$, $R$ and $T$ charged under a $\mathbb{Z}_4^R$, some unspecified
symmetries with fields $H$, $I$, $J$, and generalized functions
$A_{S,R,T}(H,I,J)$ as well as $C(H,I,J)$, such that keeping only the
necessary lowest order terms in $S$, $R$, $T$ we write
\begin{equation}
W_R = S \, A_S(H,I,J) + R \, A_R(H,I,J) + T \, A_T(H,I,J)+ S \, R \, T \,C(H, I, J) \,.
\end{equation}
After some manipulation we find again a non-trivial solution breaking
$\mathbb{Z}_4^R$ and the GUT symmetry
\begin{equation}
\frac{A_S}{\langle RT \rangle} = \frac{A_R}{\langle T S \rangle} = \frac{A_T}{\langle SR \rangle} = - C \,,
\end{equation}
provided the derivatives with respect to each of the GUT superfields $H$, $I$, $J$ fulfill
\begin{equation}
\frac{\partial A_S}{A_S} + \frac{\partial A_R}{A_R} + \frac{\partial A_T}{A_T} + \frac{\partial C}{C}=0 \,.
\end{equation}
The simple examples above illustrate that a simultaneous breaking of
R- and GUT symmetry is possible. VEV alignments required by flavour
models can therefore still be obtained within GUTs by having additional
superfields charged under the R-symmetry.
\paragraph{Flavour Alignment}
As a very simple example
we discuss the ``alignment'' of a GUT singlet flavon $\phi$ charged under a non-Abelian
family symmetry via a driving field $P$. As additional symmetries we impose
$\mathbb{Z}_4^R$ and a conventional $\mathbb{Z}_n$. The allowed
renormalizable superpotential is
\begin{equation}
\label{eq:DVAM}
W = P \left( \frac{ \phi^n }{\Lambda^{n-2}} + M^2 \right) + \kappa P^3\;,
\end{equation}
where $\Lambda$ is a generic messenger scale and $M$ a mass parameter.
Minimising the F-term conditions we find two possible solutions:
\begin{align}
\text{solution A: }& \langle P \rangle = 0 \text{ and } \langle \phi \rangle^n = - M^2 \Lambda^{n-2} \;, \\
\text{solution B: }& \langle P \rangle^2 = - \frac{M^2}{3 \, \kappa} \text{ and } \langle \phi \rangle = 0 \;.
\end{align}
In a flavour model we want the flavon to get a non-vanishing VEV so that
we would adopt solution A there. Furthermore solution A shows how the
discrete vacuum alignment method~\cite{Antusch:2011sx} can be generalised
to models with discrete R-symmetries. This method was invented in the context
of spontaneous CP violation. If CP is promoted to be fundamental all the phases
of the parameters in eq.~\eqref{eq:DVAM} are fixed and hence the phase of the
flavon VEV is fixed as well (up to a discrete choice).
\end{appendix}
|
\section{Introduction}
\label{sec:intro}
Spatio-temporal modeling has received much attention in recent years. Particularly, the rise in global temperature being a major
environmental concern, scientists are now taking keen interest in the study of the dynamics of such climatic spatio-temporal
processes\cite{Furrer:Sain,Jun:Knutti,Sain:Furrer:Cressie,Sang:Jun,Smith:Tebaldi:Nychka,Tebaldi:Sanso}. Another closely
related class of
spatio-temporal processes, that are also of much importance to climatologists, are daily rainfall and precipitation
(like mist, snowfall etc.)
across a region. Apart from climatology, many important spatio-temporal processes are also associated with different subfields
of environmental and
ecological science. To mention a few, studies on ground level concentration of ozone, $\mathrm{SO_{2}}$, $\mathrm{NO_{2}}$ and PM,
species distribution over a region, change in land usage pattern over time, etc. Other than these areas, spatio-temporal models are also
useful in geostatistics, hydrology, astrophysics, social science, archaeology, systems biology, and many more.
So, knowledge of spatio-temporal modeling is becoming increasingly necessary for better understanding of a broad range of subjects.
Hence, it is of no wonder that practitioners from distant fields are working together to develop new and effective spatio-temporal models.
In what follows, we review some such existing models, discuss some issues related to them and then propose a novel
nonparametric model. Then we devote the rest of the article to the development and exploration of the proposed model.
Finally, some simulated and real data analysis results are presented which shows that the proposed model is particularly
useful for analysis of nonstationary spatial time series data.
The proofs of all our mathematical results are deferred to the Appendix.
\section{Existing approaches for spatio-temporal data}
\label{sec:existing_approaches}
The term spatio-temporal data covers many different types of data. But here we consider only one specific type of spatio-temporal data.
We consider that there are some arbitrary spatial locations $\bold{s}_{1}, \bold{s}_{2}, \bold{s}_{3}\cdots, \bold{s}_{n}$ and that
the data have been observed at each of those spatial locations at times $t=1, 2, 3,\cdots, T $. Such data sets are very common in the
context of environmental science. One important example is ground level ozone data.
An early approach for analysis of this type of data was based on stationary Gaussian process.
But this approach to modeling did not turn out to be adequate.
Since time differs intrinsically from space in that time moves only forward, while there may not be any preferred direction
in space, time cannot be treated simply as an additional co-ordinate attached with the spatial co-ordinates. Another drawback of
this approach is stationarity which is seldom satisfied by real, physical processes.
So, alternative approaches to modeling spatio-temporal data were clearly necessary.
Briggs (1968)\cite{Briggs} proposed a simple method for constructing spatio-temporal models, based on a purely spatial model.
Later, Cox and Isham (1988)\cite{Cox:Isham} adopted a similar approach for modeling rainfall pattern. Roughly, their idea was
to start with a purely spatial stationary covariance kernel $C_{s}(\bold{h})$, where $\bold{h}\in \mathbb R^2$ and construct
a spatio-temporal stationary covariance kernel $C(\bold{h},u)=\mathbb EC_{s}(\bold{h}-u\bold{v})$, where $\bold{v}\in \mathbb R^2$
is a random velocity vector and $u$ represents the time lag. In case the velocity vector is nonrandom, this model reduces to the
classical frozen field model (see page 428 of \cite{Gelfand:Diggle:Fuentes:Guttorp}), which are useful for modeling environmental
processes that are under the influence of prevailing winds or ocean currents. Again, this model suffers from the oversimplified
assumption of stationarity.
Sampson and Guttorp (1992)\cite{Sampson:Guttorp} proposed a nonparametric nonstationary model for analysing spatial data,
based on spatial deformation. Later, they extended it to the spatio-temporal setup\cite{Bruno:Guttorp:Sampson:Cocchi,Guttorp:Meiring:Sampson}.
Schmidt and O'Hagan (2003)\cite{Schmidt:O'Hagan} gave a Bayesian formulation of the model in the spatial setup. One basic problem with this
approach is the requirement of replicates of the data, which is rarely available in practice. Only very recently,
Anderes and Stein (2008)\cite{Anderes:Stein} came up with some idea on estimating such deformation model from a single realization.
Still, they considered only the spatial setup and implementing the model in real data problems requires solving a large and quite
difficult optimization problem. (See \cite{Meiring:Monestiez} ).
A completely different approach to modeling spatio-temporal data is the convolution approach proposed by
Higdon et al. (1999)\cite{Higdon:Swall:Kern}. Initially, they proposed it for purely spatial processes and later extended
it to the spatio-temporal setup\cite{Higdon1,Higdon2}. Several extensions and modifications of the convolution approach are adopted
in \cite{Majumdar:Gelfand,Majumdar:Paul:Bautista,Zhu:Wu}. The basic idea is to start with a white noise process in space (space-time)
and create a nonstationary spatial (spatio-temporal) process by convolving it with a spatially varying kernel.
A different type of
kernel convolution based non stationary spatial process was formed by Fuentes and Smith (2001)\cite{Fuentes:Smith}.
They convolved a
family of spatially varying locally stationary processes with a spatial kernel to build a nonstationary process.
A more physically motivated approach to modeling the kind of spatio-temporal data considered here, is based on the dynamic
linear model (DLM) introduced by West and Harrison (1997)\cite{West:Harrison}. Such models are called dynamic spatio-temporal (DSTM)
models. Stroud et al. (2001)\cite{Stroud:Muller:Sanso} proposed a dynamic spatio-temporal model for the analysis of tropical rainfall
and Atlantic ocean temperature. Huerta et al. (2004)\cite{Huerta:Sanso:Stroud} proposed a seasonal dynamic spatio-temporal model
for the analysis of ozone levels. A more flexible dynamic spatio-temporal model with spatially varying state vector is proposed
in Banerjee et al. (2005)\cite{Banerjee:Gamerman:Gelfand}. Several other authors studied these models in the context of dimension
reductions etc.\cite{Lopes:Salazar:Gamerman}. Although DSTMs are used to incorporate the covariate information in
spatio-temporal modeling,
these models can also be used in the absence of covariates.
The basic idea behind the dynamic approach is to model the observed spatio-temporal process as a linear state space model,
where the coefficients associated with different covariates constitute the state vector.
Direct construction of nonstationary covariance functions is yet another approach.
While Paciorek and Schervish (2006)\cite{Paciorek:Schervish} directly construct a nonstationary covariance function in the spatial setup,
Fuentes (2001, 2002)\cite{Fuentes1,Fuentes2} proposed a spectral method.
All the above mentioned approaches mainly focus on achieving nonstationarity. But nonseparability in space-time is also another important
aspect in the context of space time modeling. Roughly, nonseparability in space-time means that space effect and time effect interact
in the formation of the process; to rephrase, space effect depends on time effect and vice versa. Although this is a highly realistic
assumption, most of the spatio-temporal models existing in the literature assume separability. Some works on nonseparable models
can be found in \cite{Cressie:Huang,Gneiting,Fuentes:Chen:Davis,Fuentes:Chen:Davis:Lackmann}. Among them, Cressie and Huang (1999)
\cite{Cressie:Huang} and Gneiting (2002)\cite{Gneiting} considered nonseparability in the stationary setup while the others have
developed spatio-temporal processes, which are both nonseparable and nonstationary.
Apart from the above mentioned approaches there are many more works in the context of spatio-temporal modeling. We attempted to provide
only a broad overview and it is by no way exhaustive.
\section{Our proposed spatio-temporal process}
\label{sec:our_proposal}
First recall, that there are some arbitrary spatial locations $\bold{s}_{1}, \bold{s}_{2}, \bold{s}_{3}\cdots, \bold{s}_{n}$
and the data $y(\bold{s}_{i},t)$ have been observed at those spatial locations at times $t=1, 2, 3,\cdots, T $. We assume that
the observed spatio-temporal process $Y(\bold{s},t)$ is driven by an unobserved spatio-temporal process $X(\bold{s},t)$
which itself is evolving in time in some unknown way. So, we model them via the state space approach but instead of
considering any known form of observational and evolutionary equations we allow them to be unknown.
We propose a nonparametric state space based model, which can even capture the nonlinear evolution flexibly.
We refer to it as a nonparametric state space based spatio-temporal model. Our model has the following form:
\begin{align}
Y(\bold{s},t)&=f(X(\bold{s},t))+\epsilon(\bold{s},t),
\label{eqn:npr1}\\
X(\bold{s},t)&=g(X(\bold{s},t-1))+\eta(\bold{s},t);
\label{eqn:npr2}
\end{align}
where $ \bold{s}\in \mathbb{R}^2 $ and $ t\in \{1,2,3,\ldots\}$;
$X(\cdot,0)$ is a spatial Gaussian process
with appropriate parameters; $\epsilon(\cdot,t)$ and $\eta(\cdot,t)$ are temporally independent and identically distributed
spatial Gaussian processes, and $g(\cdot)$ and $f(\cdot)$ are Gaussian processes on $\mathbb{R}$. They are all independent of each other.
To elaborate, let us consider a hierarchical break up of our model. Suppose, that the observed process depends on the unobserved process
through some unknown continuous function $f(\cdot)$. Then, for the purpose of Bayesian inference we must put some prior on $f(\cdot)$.
Hence, we put a Gaussian process prior on $f(\cdot)$. Similarly we put a Gaussian process prior on $g(\cdot)$.
This approach is able to capture arbitrary functional dependence between the observed and the latent variables, hence the term nonparametric.
Finally, to complete the model specification, we need to describe the parameters associated with the Gaussian processes.
We assume that the Gaussian process $f(x)$ has mean function of the form $\beta_{0f}+\beta_{1f}x$ (where
$\beta_{0f}, \beta_{1f}$ are suitable parameters) and has covariance kernel of the form $c_{f}(x_{1},x_{2})=\gamma(\|x_1-x_2\|)$,
where $\gamma$ is a positive definite function.
It consists of parameters that determine the smoothness of the sample path of $f(\cdot)$. Moreover $\gamma$ is such that the centered Gaussian process associated with it has continuous sample paths. Thus, we consider isotropic covariance functions
with suitable smoothness properties.
Typical examples of $\gamma$ are exponential, powered exponential, Gaussian, Mat\'{e}rn etc. (see, for example, Table 2.1 of
\cite{Banerjee04} for other examples of such covariance kernels).
Similarly, parameters $\beta_{0g}, \beta_{1g}$ and $c_{g}(x_{1},x_{2})$ are associated with the Gaussian process $g(x)$.
The zero mean Gaussian processes associated with the noise variables have covariance kernels $c_{\epsilon}(\bold{s},\bold{s}^{\prime})$
and $c_{\eta}(\bold{s},\bold{s}^{\prime})$ which are also of the form as discussed above.
Regarding the Gaussian process associated with $X(\cdot,0)$, we assume a mean process of the form $\mu_{0}(\cdot)$ and
isotropic covariance kernel $c_{0}(\cdot,\cdot)$. For convenience we introduce separate symbols for the mean vector and the
covariance matrix associated with $(X(\bold{s}_{1},0),X(\bold{s}_{2},0),\cdots X(\bold{s}_{n},0))$ where
$\bold{s}_{1}, \bold{s}_{2}, \bold{s}_{3}\cdots, \bold{s}_{n}$ are the spatial locations where the data are observed.
We denote them by ${\boldsymbol{\mu}}_{0}$ and ${\boldsymbol{\Sigma}}_{0}$ respectively.
From (\ref{eqn:npr1}) and (\ref{eqn:npr2}) it is not difficult to see that our model boils down to simple
DSTM with no covariate if the process means associated with the Gaussian processes $f(\cdot)$ and $g(\cdot)$ are given by
$\beta_{0f}=\beta_{0g}=0$, $\beta_{0f}=\beta_{0g}=1$, and the process variance
(denoted by $\sigma_{f}^{2}$ and $\sigma_{g}^{2}$) become 0. The model equations (\ref{eqn:npr1}) and (\ref{eqn:npr2})
then reduce to the following forms
\begin{align}
Y(\bold{s},t) &=X(\bold{s},t)+\epsilon(\bold{s},t),
\label{eqn:DLM1}\\
X(\bold{s},t) &=X(\bold{s},t-1)+\eta(\bold{s},t);
\label{eqn:DLM2}
\end{align}
where $ \bold{s}\in \mathbb{R}^2 $ and $ t\in\{1,2,3,\ldots\}$.
Although we develop our Gaussian process based spatio-temporal model for equi-spaced time points,
simple modification of this model can handle non equi-spaced time points as well.
However, for the sake of
simplicity and brevity in this article we will confine ourselves only within the framework of equi-spaced time points.
\subsection{Some measurability and existential issues}
Before we proceed to explore the properties of our model, we need to ensure that a family of valid spatio-temporal
stochastic processes is induced by the proposed model. Only then physical processes can be modeled by it.
In general such issues are almost always trivially satisfied and so, never discussed in detail. But in this case we
need to show that $f(X(\bold{s}_{1},t)),f(X(\bold{s}_{2},t)),\cdots,f(X(\bold{s}_{n},t))$ are jointly measurable for any $n$
and any set of spatial locations $\bold{s}_{1},\bold{s}_{2},\cdots,\bold{s}_{n}$, and this is not a trivial problem.
The difficulty is that when $f$ and $X$ both are random, $f(X)$ need not be a measurable or valid random variable.
It is the sample path continuity of $f(\cdot)$, which compels $f(X(\bold{s},t))$ to be measurable. The following
theorem establishes this mathematically.
\begin{theorem}
\label{thm:measurable}
The model defines a family of valid (measurable) spatio-temporal processes on $\mathbb R^{2}\times\mathbb Z^+$.
\end{theorem}
Once it is established that the proposed model induces a family of valid spatio-temporal processes, we look
into some important properties like the joint distributions of state variables and observed variables, covariance structure
of the observed process etc. An important point is that although we develop our model considering the space is
$\mathbb{R}^2$ all the results go through for $\mathbb{R}^d$ $(d>2)$ as well.
\subsection{Joint distribution of the variables}
It is of importance to derive the joint distribution of the observed variables. As our model is based on an
implicit hierarchical structure, the joint distribution of the observed variables is non-Gaussian. But before going
into the details, we need to derive the joint distribution of the state variables, which will also be required
for our MCMC based posterior inference.
\begin{theorem}
\label{thm:state}Suppose that the spatio-temporal process is observed at locations
$\bold{s}_{1}, \bold{s}_{2}, \bold{s}_{3}\cdots, \bold{s}_{n}$ for times $t=1, 2, 3,\cdots, T $.
Then the joint distribution of the state variables is non-Gaussian and has the pdf
\vspace{3mm}
$\mathlarger{\mathlarger{\frac{1}{(2\pi)^\frac{n}{2}}\frac{1}{|{\mathbf{\Sigma}}_{0}|^\frac{1}{2}}}}\exp\left[-\frac{1}{2}{\begin{pmatrix}x(\bold{s}_{1},0)-\mu_{01}\\x(\bold{s}_{2},0)-\mu_{02}\\ \vdots\\x(\bold{s}_{n},0)-\mu_{0n}\end{pmatrix}}^{\prime}{{\mathbf{\Sigma}}_{0}}^{-1}{\begin{pmatrix}x(\bold{s}_{1},0)-\mu_{01}\\x(\bold{s}_{2},0)-\mu_{02}\\ \vdots\\x(\bold{s}_{n},0)-\mu_{0n}\end{pmatrix}}\right] \mathlarger{\mathlarger{\frac{1}{(2\pi)^\frac{nT}{2}}\frac{1}{|\tilde{\mathbf{\Sigma}}|^\frac{1}{2}}}}\times$
\vspace{3mm}
$\exp\left[-\frac{1}{2}{\begin{pmatrix}x(\bold{s}_{1},1)-\beta_{0g}-\beta_{1g}
x(\bold{s}_{1},0)\\x(\bold{s}_{2},1)-\beta_{0g}-\beta_{1g}
x(\bold{s}_{2},0)\\ \vdots\\x(\bold{s}_{n},T)-\beta_{0g}-\beta_{1g}
x(\bold{s}_{n},T-1)\end{pmatrix}}^{\prime}{\tilde{\mathbf{\Sigma}}}^{-1}{\begin{pmatrix}x(\bold{s}_{1},1)-\beta_{0g}-\beta_{1g}
x(\bold{s}_{1},0)\\x(\bold{s}_{2},1)-\beta_{0g}-\beta_{1g}
x(\bold{s}_{2},0)\\ \vdots\\x(\bold{s}_{n},T)-\beta_{0g}-\beta_{1g}
x(\bold{s}_{n},T-1)\end{pmatrix}}\right] $
\vspace{3mm}
where ${\boldsymbol{\mu}}_{0}=(\mu_{01},\mu_{02},\cdots,\mu_{0n})^\prime$ and ${\mathbf{\Sigma}}_{0}$
are already defined to be the mean vector and the covariance matrix of
$(X(\bold{s}_1, 0), X(\bold{s}_2, 0),\ldots, X(\bold{s}_n , 0))$,
and
\vspace{3mm}
$\tilde{\mathbf{\Sigma}}=\begin{pmatrix}1&0&\cdots &0\\0&1&\cdots &0\\ \vdots \\0&0&\cdots &1\\ \end{pmatrix}\bigotimes{\mathbf{\Sigma}}_{\eta}+\mathbf{\Sigma}$
\vspace{3mm}\\
where the elements of ${\mathbf{\Sigma}}_{\eta}$ are obtained from the purely spatial covariance function $c_{\eta}$ and the elements of $\mathbf{\Sigma}$ are obtained from the covariance function $c_{g}$ in the following way\\
the $(i,j)$ th entry of ${\mathbf{\Sigma}}_{\eta}$ is $c_{\eta}(\bold{s}_{i},\bold{s}_{j})$ and the $((t_{1}-1)n+i,(t_{2}-1)n+j)$ th entry of $\mathbf{\Sigma}$ is $c_{g}(x(\bold{s}_{i},t_{1}-1),x(\bold{s}_{j},t_{2}-1))$ where $1 \leq t_{1},t_{2} \leq T$ and $1\leq i,j \leq n$ .
\end{theorem}
Although the above density function has physical resemblance with a Gaussian density, the involvement of
$x(\bold{s}_{i},t)$ in $\tilde{\mathbf{\Sigma}}$ renders it non-Gaussian. In the extreme case when the process
variance $c_{g}(0,0)$ ($=\sigma_{g}^2$) of the Gaussian process $g(\cdot)$ is 0, $\tilde{\mathbf{\Sigma}}$
becomes a block diagonal matrix with identical blocks and the joint density becomes Gaussian.
In the formation of $\tilde{\mathbf{\Sigma}}$ the component
${\mathbf{I}}_{T\times T}\bigotimes\mathbf{\Sigma_{\eta}}$ corresponds to linear evolution and Gaussianity whereas the
component $\mathbf{\Sigma}$ corresponds to nonlinear evolution and non-Gaussianity.
Moreover, it is clear from the form of the density function that the temporal aspect is imposed through both the location function and
the scale function associated with the latent process, making it more flexible.
The interesting property that the joint density of the observed variables is also non-Gaussian,
follows from non-Gaussianity of the joint distribution of the latent states. We have the following
theorem in this regard.
\begin{theorem}
\label{thm:observe}
Suppose that the spatio-temporal process is being observed at locations $\bold{s}_{1}, \bold{s}_{2}, \bold{s}_{3}\cdots, \bold{s}_{n}$ for times $t=1, 2, 3,\cdots, T $.
Then the following hold true:
\vspace{3mm}\\
(a) The joint distribution of the observed variables is a Gaussian mixture and has the following density
\vspace{3mm}
$\mathlarger{\mathlarger{\int}_{\mathbb R^{nT}}\frac{1}{(2\pi)^\frac{nT}{2}}\frac{1}{|{\mathbf\Sigma}_{f,\epsilon}|^\frac{1}{2}}}\times $
\vspace{3mm}
$\exp\left[-\frac{1}{2}{\begin{pmatrix}y(\bold{s}_{1},1)-\beta_{0f}-\beta_{1f}x(\bold{s}_{1},1)\\y(\bold{s}_{2},1)-\beta_{0f}-\beta_{1f}x(\bold{s}_{2},1)\\ \vdots\\y(\bold{s}_{n},T)-\beta_{0f}-\beta_{1f}x(\bold{s}_{n},T)\end{pmatrix}}^{\prime}{{\mathbf\Sigma}_{f,\epsilon}}^{-1}{\begin{pmatrix}y(\bold{s}_{1},1)-\beta_{0f}-\beta_{1f}x(\bold{s}_{1},1)\\y(\bold{s}_{2},1)-\beta_{0f}-\beta_{1f}x(\bold{s}_{2},1)\\ \vdots\\y(\bold{s}_{n},T)-\beta_{0f}-\beta_{1f}x(\bold{s}_{n},T)\end{pmatrix}}\right]\mathlarger{ h(\mathbf{x}) \,d\mathbf{x}} $
\vspace{3mm}\\
where the mixing density $h(\mathbf{x})$ is obtained by marginalizing the pdf derived at Theorem \ref{thm:state} with respect to
the variables $x(\bold{s}_{1},0),x(\bold{s}_{2},0),\cdots,x(\bold{s}_{n},0)$ and the $((t_{1}-1)n+i,(t_{2}-1)n+j)$
th entry of $\mathbf{\Sigma}_{f,\epsilon}$ is given by $c_{f}(x(\bold{s}_i,t_1),x(\bold{s}_j,t_2))+c_{\epsilon}(\bold{s}_i,\bold{s}_j)\delta(t_1-t_2)$ where $1\leq t_{1},t_{2},\leq T$ and $1\leq i,j\leq n$.
\vspace{3mm}\\
(b) In the extreme case, when the process variance $c_{g}(0,0)$ ($=\sigma_{g}^2$) and $c_{f}(0,0)$ ($=\sigma_{f}^2$)
of each of the Gaussian processes $g(\cdot)$ and $f(\cdot)$ are 0,
the joint distribution turns into Gaussian.
\end{theorem}
So, our method is very flexible in the sense that it can yield both Gaussian and non-Gaussian models for the observed data. Non-Gaussianity itself is of independent interest in the context of spatial modelling and some works in this direction are \cite{Kim:Mallick,Oliveira:Kedem:Short,Fonseca:Steel,Palacios:Steel}.
\subsection{Covariance exploration}
Before delving into a deeper study of the covariance function and some related issues like nonstationarity
and nonseparability, a more basic question is whether the process is light tailed, that is, whether
or not all the coordinate variables have finite variance.
From Theorem \ref{thm:observe} we see that the observed variables are distributed as
a Gaussian mixture and Gaussian mixtures sometimes can give rise to heavy tailed distributions.
So, an answer to the above question is not immediately available. In what follows, we first
mathematically establish that the process is light tailed and then we show that the covariance function
is nonstationary and nonseparable.
\begin{theorem}
\label{thm:covariance}
(a) The observed spatio-temporal process $Y(\bold{s},t)$ is light tailed in the sense that all the coordinate variables have finite variance.\\
(b) The covariance function $c_{y}((\bold{s},t),(\bold{s^*},t^*))$ of the observed spatio-temporal process is
nonstationary and nonseparable for any $s,s^*,t,t^*$.
\end{theorem}
Although we are able to show that our method yields nonstationary and nonseparable covariance function,
it has an obvious drawback that no closed form expression for the covariance function is available. This problem,
however exists in many other approaches including the convolution models, DSTM models etc., where a closed form
covariance function is available only in very few special cases. Actually, the direct construction of
nonstationary covariance function is the only approach that yields a closed form covariance function.
Moreover, from the prediction point of view, this is not a serious problem since what we need are the posterior
predictive distributions at un-monitored locations at arbitrary points of time, which do not require
closed form expressions of covariance functions.
However, we still attempted to derive some closed form expression, and are partially successful in the sense
that when our model is approximately linear (in some suitable sense to be described later), the covariance function is approximately a geometric function of time lag.
\subsubsection{Approximate form of the covariance function}
Here, in Theorem \ref{thm:closed_covariance} we show that if the process variances ($\sigma_{g}^2$) and ($\sigma_{f}^2$) are small then under some
additional assumptions, the covariance function is approximately a geometric function of time lag.
This result is mainly of theoretical interest and has nothing to do with the application of our model.
\begin{theorem}
\label{thm:closed_covariance}
Assume that $|\beta_{1g}|<1$. Then for given $\epsilon^{\prime\prime}>0$ arbitrarily small,
$\exists~\delta>0$ such that for $0<\sigma_{g}^{2},\sigma_{f}^{2}<\delta$ the covariance between
$Y(\bold{s},t)$ and $Y(\bold{s^*},t^*)$, denoted by
$c_{y}((\bold{s},t),(\bold{s^*},t^*))$ is of following form
\begin{align*}
\beta_{1g}^{|t-t^*|}\left[c_{0}(\bold{s},\bold{s}^*)+[\frac{1-\beta_{1g}^{2(t^*+1)}}{1-\beta_{1g}^2}]c_{\eta}(\bold{s},\bold{s}^*)\right]-\epsilon^{\prime\prime}
&\leq c_{y}((\bold{s},t),(\bold{s^*},t^*))\\
&\leq \beta_{1g}^{|t-t^*|}\left[c_{0}(\bold{s},\bold{s}^*)+[\frac{1-\beta_{1g}^{2(t^*+1)}}{1-\beta_{1g}^2}]c_{\eta}(\bold{s},\bold{s}^*)\right]+\epsilon^{\prime\prime}.
\end{align*}
\end{theorem}
\subsection{Sample path properties}
Until now we have explored the finite dimensional distributions and properties of the
spatio-temporal process $Y(\bold{s},t)$. But finite dimensional properties alone are not
sufficient to characterize any arbitrary general stochastic process. Two stochastic processes
with completely different sample path behaviour may have identical finite dimensional
distributions and properties. We demonstrate this through the following simple example (see also \cite{Adler81}):
Let us consider two spatio-temporal processes $Y(\bold{s},t)(\omega)$ and $Y^{*}(\bold{s},t)(\omega)$
defined on the same probability space $\Omega=[0,1]^2$ in the following way:
$Y(\bold{s},t)(\omega)=0$ for all $\bold{s},t,\omega$, and
$Y^{*}(\bold{s},t)(\omega)=1$ for all $\bold{s}=\omega$ and $=0$ otherwise.
Then one can show that for any fixed $t$, $Y(\bold{s},t)(\omega)$ has continuous sample path
with probability $1$ whereas $Y^{*}(\bold{s},t)(\omega)$ has discontinuous sample path with probability $1$.
Now, we study the path properties of the proposed spatio-temporal process. The first part of
the following theorem states that the observed spatio-temporal process has continuous
sample path almost surely and in the second part it says that under additional
smoothness assumptions regarding the covariance functions, the observed spatio-temporal process
will have smooth sample paths almost surely.
\begin{theorem}
\label{thm:sample_path}
(a) The spatio-temporal process $Y(\bold{s},t)$ has continuous sample paths with probability $1$.
\vspace{2mm}\\
(b) Assume that the covariance functions $c_{f}(\cdot,\cdot),c_{g}(\cdot,\cdot),
c_{\epsilon}(\cdot,\cdot),c_{\eta}(\cdot,\cdot),c_{0}(\cdot,\cdot)$ satisfy the
additional smoothness assumptions that centered Gaussian processes with these covariance
functions will have almost surely $k$ times differentiable sample paths. Then the
spatio-temporal process $Y(\bold{s},t)$ will also have
almost surely $k$ times differentiable sample paths.
\end{theorem}
The first part of the above theorem states that almost always the spatial surface induced by our model
at any time point $t$ is continuous. So, unless the spatial surface is extremely irregular,
any spatio-temporal data can be modeled reasonably adequately by our proposed spatio-temporal process.
A stronger statement, however, is made in the second part of the theorem. It says that if
the covariance functions $c_{f}(\cdot,\cdot),c_{g}(\cdot,\cdot),c_{\epsilon}(\cdot,\cdot),
c_{\eta}(\cdot,\cdot),c_{0}(\cdot,\cdot)$ are sufficiently regular (smooth) then the sample paths
of the process $Y(\bold{s},t)$ are almost always regular (smooth) and their degrees of smoothnes depend on the degree of smoothness
of the covariance functions. As an illustration consider the situation when all the covariance function
belong to the Mat\'{e}rn family whose smoothenes parameter is $\nu$. Then the sample paths of
the centered Gaussian processes with the respective covariance functions will be almost surely
$\lfloor \nu-1 \rfloor$ times differentiable, where, for any $x$, $\lfloor x\rfloor$ denotes the smallest
integer greater than or equal to $x$. So, the spatio-temporal process
$Y(\bold{s},t)$ will also have almost surely $\lfloor \nu-1 \rfloor$ times differentiable sample paths.
Hence, when we have clear evidence from the data or prior knowledge regarding the degree of smoothness
of the spatio-temporal process we should choose all the covariance functions
$c_{f}(\cdot,\cdot),c_{g}(\cdot,\cdot),c_{\epsilon}(\cdot,\cdot),c_{\eta}(\cdot,\cdot),
c_{0}(\cdot,\cdot)$ from the Mat\'{e}rn family with specific value of $\nu$.
\section{Model fitting and prediction}
Now we illustrate how our model can be used to make prediction at new spatio-temporal coordinates.
Firstly, let us specify the prior structure. We consider bivariate Gaussian priors for each of
$(\beta_{0g},\beta_{1g})$ and $(\beta_{0f},\beta_{1f})$. Although any reasonable isotropic covariance
kernel that satisfy the mild regularity conditions already mentioned in Section \ref{sec:our_proposal} can be used in our model, for the sake of simplicity
we consider the squared exponential covariance kernel with the following representation
$c(\bold{u},\bold{v})=\sigma^2 e^{-\lambda||\bold{u}-\bold{v}||^2}$. Since, we have five covariance
kernels $c_{f}(\cdot,\cdot),c_{g}(\cdot,\cdot),c_{\epsilon}(\cdot,\cdot),c_{\eta}(\cdot,\cdot),
c_{0}(\cdot,\cdot)$ we need to put prior on five scale parameters $\sigma_{f}^2,\sigma_{g}^2,
\sigma_{\epsilon}^2,\sigma_{\eta}^2,\sigma_{0}^2$ and five smoothness parameters $\lambda_{f},\lambda_{g},
\lambda_{\epsilon},\lambda_{\eta},\lambda_{0}$. We consider lognormal priors for all of them. Also, we have taken $N(\boldsymbol{0},\boldsymbol{I})$ prior for the vector parameter ${\boldsymbol{\mu}}_{0}$.
All the priors considered above are mutually independent.
With the prior specification as above and the conditional densities
$[\mathbf{y}|\mathbf{x},\tilde{\theta}]$ and $[\mathbf{x}|\tilde{\theta}]$ being explicitly available
($\tilde{\theta}$ denotes the vector of all the parameters) we design a Gibbs sampler with
Gaussian full conditionals for ${\boldsymbol{\mu}}_{0}$, $(\beta_{0g},\beta_{1g})$ and $(\beta_{0f},\beta_{1f})$ and
Metropolis steps for scale parameters $\sigma_{f}^2,\sigma_{g}^2,\sigma_{\epsilon}^2,\sigma_{\eta}^2,
\sigma_{0}^2$ and smoothness parameters $\lambda_{f},\lambda_{g},\lambda_{\epsilon},\lambda_{\eta},\lambda_{0}$.
We update the latent state vector $\mathbf{x}_{nT}$ using Transformation based Markov Chain Monte
Carlo (TMCMC) introduced by Dutta and Bhattacharya (2014)\cite{Dutta:Bhattacharya}.
In particular, we use the additive transformation which has been shown by
Dutta and Bhattacharya (2014)\cite{Dutta:Bhattacharya} to require less number of ``moves types" compared to
other valid transformations.
The idea of TMCMC is very simple yet a very powerful one.
Here we briefly illustrate the idea of additive TMCMC by contrasting it with the traditional Random Walk Metropolis (RWM)
approach, assuming that we wish to update all the variables simultaneously.
Suppose that we want to simulate from the conditional distribution $[\mathbf{x}_{nT}|\tilde{\theta}]$ using the RWM approach.
Then we have to simulate $nT$ independent Gaussian random variables
$\mathbf{E}_{nT}=(\epsilon_{1},\epsilon_{2},\cdots,\epsilon_{nT})'$; assuming that the current state of the
Markov chain is $\mathbf{x}_{nT}^{(i)}$, we accept the new state $\mathbf{x}_{nT}^{(i)}+\mathbf{E}_{nT}$
with probability $\min\left\{1,\frac{[\mathbf{x}_{nT}^{(i)}+
\mathbf{E}_{nT}|\tilde{\theta}]}{[\mathbf{x}_{nT}^{(i)}|\tilde{\theta}]}\right\}$.
However if $nT$ is large then this acceptance probability will tend to be extremely small.
Hence the RWM chain sticks to a particular state for very long time and therefore the convergence to the posterior
$[\mathbf{x}_{nT}|\tilde{\theta}]$ is very slow. What additive TMCMC does is simulate only one $\epsilon>0$
from some arbitrary distribution left-truncated at zero, and then form the
$nT$ dimensional vector $\mathbf{E}_{nT}$ setting the $k$-th element independently to $-\epsilon$ with probability
$p_k$ and $+\epsilon$ with probability $1-p_k$. For our applications we set $p_k=1/2$ $\forall$ $k=1,2,\ldots,nT$.
Then we accept the new state
$\mathbf{x}_{nT}^{(i)}+\mathbf{E}_{nT}$ with acceptance probability
$\min\left\{1,\frac{[\mathbf{x}_{nT}^{(i)}+\mathbf{E}_{nT}|\tilde{\theta}]}{[\mathbf{x}_{nT}^{(i)}|\tilde{\theta}]}\right\}$.
Dutta and Bhattacharya (2014)\cite{Dutta:Bhattacharya}, Dey and Bhattacharya (2014a)\cite{Dey:Bhattacharya2014a},
Dey and Bhattacharya (2014b)\cite{Dey:Bhattacharya2014b} provide details of many advantages of TMCMC (in particular, additive TMCMC)
as compared to traditional MCMC (in aprticular, RWM).
However, in our setup, we have not used TMCMC directly. Instead, we simulated $T$ independent $\epsilon_{t}$ each to update $n$ latent observations corresponding to time $t$. This may be called a block TMCMC and it improves mixing significantly over ordinary TMCMC. In our model dimension of the state vector is large and updating it using usual RWM
would have been very inefficient. TMCMC saves us from such pitfall.
Using the above sampling-based approach we study the posterior distribution of the unknown quantities and
make inferences regarding the parameters. But our main goal is to predict $y(\bold{s}^*,t^*)$ at some
new spatio-temporal coordinate $(\bold{s}^*,t^*)$. So, we augment $x(\bold{s}^*,t^*)$ with
$\{x(\bold{s}_{i},t)\}$ where $ i=1,\cdots,n $ and $ t=1,\cdots,T $. The conditional distribution of
$[\mathbf{y}|\mathbf{x},x(\bold{s}^*,t^*),\tilde{\theta}]$ remains the same as
$[\mathbf{y}|\mathbf{x},\tilde{\theta}]$ and posterior simulation is done exactly the same way
as in the case of model-fitting. Once the post burn-in posterior samples
$(x^{(B)}(\bold{s}^*,t^*),\tilde{\theta}^{(B)}),(x^{(B+1)}(\bold{s}^*,t^*),\tilde{\theta}^{(B+1)}),\cdots$ from $[x(\bold{s}^*,t^*),\tilde{\theta}|\mathbf{y}]$ are available,
we simulate $y^{(B)}(\bold{s}^*,t^*),y^{(B+1)}(\bold{s}^*,t^*),\cdots$ from
\\ $N\left(\beta_{0f}^{(j)}+\beta_{1f}^{(j)}x^{(j)}(\bold{s}^*,t^*)+{\boldsymbol{\Sigma}}^{(j)}_{12}\left({\boldsymbol{\Sigma}}^{(j)}_{22}\right)^{-1}\boldsymbol{V}^{(j)},(\sigma_{f}^{(j)})^2+(\sigma_{\epsilon}^{(j)})^{2}-{\boldsymbol{\Sigma}}^{(j)}_{12}\left({\boldsymbol{\Sigma}}^{(j)}_{22}\right)^{-1}{\boldsymbol{\Sigma}}^{(j)}_{21}\right)$
where $j=B,B+1,\cdots$ and $\{\beta_{0f}^{(j)},\beta_{1f}^{(j)},\sigma_{f}^{(j)},\sigma_{\epsilon}^{(j)}\}$
are post burn-in posterior samples for the respective parameters. Here ${\boldsymbol{\Sigma}}^{(j)}_{12}$ is the covariance between
$x^{(j)}(\bold{s}^*,t^*)$ and the vector $\{x^{(j)}(\bold{s}_{i},t)\}$ and ${\boldsymbol{\Sigma}}^{(j)}_{22}$ is the covariance matrix of the vector $\{x^{(j)}(\bold{s}_{i},t)\}$. The vector $\boldsymbol{V}^{(j)}$ consists of $ y(\bold{s}_{i},t)-\beta_{0f}^{(j)}-\beta_{1f}^{(j)}x^{(j)}(\bold{s}_{i},t)$ where $ i=1,\cdots,n $ and $ t=1,\cdots,T $.
These $y^{(B)}(\bold{s}^*,t^*),y^{(B+1)}(\bold{s}^*,t^*),\cdots$ are samples from the posterior
predictive distribution $[y(\bold{s}^*,t^*)|\mathbf{y}]$ which is used for prediction at the new
spatio-temporal coordinate $(\bold{s}^*,t^*)$.
\section{Simulation and real data study}
\subsection{Illustration via a simulation example}
We randomly sample $15$ points from a square of length $2$ and generate spatial time series of
length $20$ at those points using our model. The true parameter values for simulation are
$\beta_{0f}=-4.1,\beta_{1f}=0.51,\beta_{0g}=5.1,\beta_{1g}=0.64$ and $\sigma_{f}=\sigma_{g}=1.0,
\sigma_{\epsilon}=4.0,\sigma_{\eta}=4.9,\sigma_{0}=5.8$. The true values of the smoothness parameters
are $\lambda_{f}=4.3,\lambda_{g}=2.4,\lambda_{\epsilon}=\lambda_{\eta}=6.25,\lambda_{0}=4.0$.
We simulate the vector ${\boldsymbol{\mu}}_{0}$ from $N(\boldsymbol{0},\boldsymbol{I})$.
We consider the prior structure as mentioned in Section 4. We use independent diffused normal priors with
mean $0$ and variance $1000$ for each of $\beta_{0g},\beta_{1g},\beta_{0f},\beta_{1f}$. For the vector ${\boldsymbol{\mu}}_{0}$ we consider $N(\boldsymbol{0},\boldsymbol{I})$ prior. For the scale parameters $\sigma_{f}^2,\sigma_{g}^2$ we used lognormal priors with location parameter $0$ and scale parameter $0.7$. For the scale parameters $\sigma_{\epsilon}^2,\sigma_{\eta}^2,\sigma_{0}^2$
associated respectively with the error processes and the spatial process at time $0$ we used lognormal priors
with location parameter $3$ and scale parameter $0.1$. The choice of the hyper parameters associated with
$\sigma_{f}^2,\sigma_{\epsilon}^2$ are made based on the variance of the data. We used lognormal priors
also for the smoothness parameters. However these lognormal priors are concentrated near $0$. This somewhat
concentrated prior provides safeguard against ill-conditioning of the relevant covariance matrices
during MCMC iterations -- indeed,
free movement of the smoothness parameters in their respective parameter spaces for MCMC pusposes
can make the covariance matrices
$\tilde{\mathbf{\Sigma}}$ and $\mathbf{\Sigma}_{f,\epsilon}$ drastically ill-conditioned.
In making the above selection of hyper parameters we have extensively used MCMC pilot runs based on
a smaller subset of the entire dataset. Since, due to smaller sizes of the data sets the pilot runs were much faster,
we were able to experiment with many such pilot MCMC runs with different combinations of the hyper parameters.
We chose that combination with smallest deviation of the posterior predictive median surface from
the true data surface in terms of total absolute deviation.
Then we perform our final MCMC computation based on the entire dataset using those selected values of the hyper parameters. In a standard laptop we have run $200,000$ iterations at a rate almost $70,000$ iterations per day. We have taken first $100,000$ iterations as burn-in and the post burn-in $100,000$ iterations have been used for posterior inference.
Convergence of the MCMC chains are checked using different starting points.
Visual monitoring suggests that the convergence is satisfactory.
Apart from the above set of priors we also experimented with some non informative priors but the results
turned out to be only negligibly different, indicating some degree of robustness with respect to the
prior choices.
We obtain the leave-one-out posterior predictive distribution
for the observed data at each of the 300 space-time coordinates. Out of the 300 data points, at 292 space-time
coordinates the observed data fell within the $95\%$ Bayesian prediction intervals of their respective posterior predictive distributions. The average length of the
$95\%$ Bayesian prediction interval being 20.25 indicates that the results are encouraging.
Although we have used different sets of parameter values to conduct this simulation experiment
we report only one study for space constraint.
The results are shown in Figures 1(a)-1(f).
\begin{figure}[H
\begin{center}
\subfigure[]{%
\label{fig:ourmodelfirst}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{UCV9t7median.pdf}
}%
\subfigure[]{%
\label{fig:ourmodelsecond}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{UCV9t7.pdf}
}\\
\subfigure[]{%
\label{fig:ourmodelthird}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{UCV9t14median.pdf}
}%
\subfigure[]{%
\label{fig:ourmodelfourth}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{UCV9t14.pdf}
}\\
\subfigure[]{%
\label{fig:ourmodelfifth}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{UCV9s6.pdf}
}%
\subfigure[]{%
\label{fig:ourmodelsixth}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{UCV9s9.pdf}
}\\
\end{center}
\caption{Leave-one-out posterior predictive performance of our model when the data are generated from our own model.}
\label{fig:subfigures1}
\end{figure}
A brief of Figure \ref{fig:subfigures1} follows.
Panel (a) displays the leave-one-out 95\% Bayesian prediction interval for all of the $15$ spatial locations at time point $t=7$.
The upper and lower quantiles (which form the intervals) are interpolated to form two surfaces.
The middle surface is obtained by interpolating the leave-one-out point predictions (which is the median in this case)
at the monitoring sites $\bold{s}_{1}, \bold{s}_{2},\cdots, \bold{s}_{15}$ for time point $t=7$. Panel (b) displays
once again the leave-one-out 95\% Bayesian prediction interval for all of the $15$ spatial locations but this time
the middle surface is interpolated through the truly observed data (the truly observed data being represented by black dots).
Panel (c) and Panel (d) display similar plots as (a) and (b) but for time point $t=14$. Panels (e) and (f) show
the leave-one-out 95\% Bayesian prediction intervals for the time series at spatial locations $\bold{s}_{6}$ and $\bold{s}_{9}$
respectively. The starred curve represents the curve interpolated through the truly observed data
(the observed data itself being represented by star) and the plain curves represent respectively the upper,
middle and lower quantiles associated with the posterior predictive distributions.
In other words, Figures 1(a)-1(f) demonstrate the leave-one-out posterior predictive performance of our model.
While figures 1(b) and 1(d) show that our model performs satisfactorily in terms of interval prediction
at purely spatial level, the similarity of the median surface in figure 1(a) with the true data surface in
figure 1(b) indicates that our model performs well also in terms of point prediction. The same conclusion
can be drawn by observing figures 1(c) and 1(d). Figures 1(e) and 1(f) confirm that our model is also
successful in terms of prediction at the temporal level.
\begin{figure}[H
\begin{center}
\subfigure[]{%
\label{fig:corr1}
\includegraphics[width=0.32\textwidth,height=0.20\textheight]{corr5292.pdf}
}%
\subfigure[]{%
\label{fig:corr2}
\includegraphics[width=0.32\textwidth,height=0.20\textheight]{corr37710.pdf}
}
\subfigure[]{%
\label{fig:corr3}
\includegraphics[width=0.32\textwidth,height=0.20\textheight]{corr858.pdf}
}\\% ------- End of the first row ----------------------%
\subfigure[]{%
\label{fig:corr4}
\includegraphics[width=0.32\textwidth,height=0.20\textheight]{corr1623.pdf}
}
\subfigure[]{%
\label{fig:corr5}
\includegraphics[width=0.32\textwidth,height=0.20\textheight]{corr22862.pdf}
}%
\subfigure[]{%
\label{fig:corr6}
\includegraphics[width=0.32\textwidth,height=0.20\textheight]{corr32494.pdf}
}%
\end{center}
\caption{%
Panels (a)-(f) display the 95\% Bayesian credible intervals for the correlation function for some fixed combination of
$(\bold{s}_{i},t_{j}),(\bold{s}_{i^{\prime}},t_{j^{\prime}})$. The coverage region is denoted by the horizontal black line and the true
value is denoted by the vertical black line.
}%
\label{fig:subfigures2}
\end{figure}
While figures 1(a)-1(f) display the performance of our model in terms of prediction, figures 2(a)-2(f) focus on
the issue of capturing the true, underlying correlation structure. The figures display the posterior
distribution of the correlation function $\rho((\bold{s}_{i},t_{j}),(\bold{s}_{i^{\prime}},t_{j^{\prime}}))$
for different choices of $\bold{s}_{i},t_{j},\bold{s}_{i^{\prime}},t_{j^{\prime}}$.
The true value and the $95\%$ Bayesian credible interval are also depicted in the figures.
As is evident from the diagrams, the true values lie well within the respective $95\%$ Bayesian credible intervals,
vindicating reasonable performance of our model in terms of capturing the true correlation structure.
Hence, in summary, this simulation study demonstrates the effectiveness of our model
for the purpose of prediction as well as for learning about the true correlation structure.
\subsection{Simulation from a nonlinear non-Gaussian spatial state space model}
In the previous section we described how our model performs when the data is simulated from our own model.
But, now we describe a simulation study where the data is generated from a nonlinear non-Gaussian state space
model which is completely different from our model.
Let us first describe our data generation method.
We consider a square of length $2$ and generate $20$ random locations inside it. Then we simulate a spatial time series of
length $20$ at those points using the following nonlinear state space model:
\begin{align}
Y(\bold{s},t) =&-4.1+0.7X(\bold{s},t)+\epsilon(\bold{s},t),
\label{eqn:NLDLM1}\\
X(\bold{s},t) =&-1.1+0.5X(\bold{s},t-1)+3\sin(\frac{\pi}{4}X(\bold{s},t-1))-5\sin(\frac{\pi}{5}X(\bold{s},t-1))+\eta(\bold{s},t),
\label{eqn:NLDLM2}\\
(X(\bold{s}_{1},0)&,X(\bold{s}_{2},0),\cdots,X(\bold{s}_{20},0))\sim N({\boldsymbol{0}},{\boldsymbol{\Sigma}}_{0});
\label{eqn:NLDLM3}
\end{align}
where $ \bold{s}\in \mathbb{R}^2 $ and $ t\in\{1,2,3,\ldots\}$. $\epsilon(\cdot,t)$ and $\eta(\cdot,t)$ are temporally independent and identically distributed spatial Gaussian processes with squared exponential covariance kernel. The corresponding scale parameters are $\sigma_{\epsilon}=3.0,\sigma_{\eta}=3.9$ and the smoothness parameters are $\lambda_{\epsilon}=6.25,\lambda_{\eta}=6.25$. The covariance matrix ${\boldsymbol{\Sigma}}_{0}$ is also generated by a squared exponential covariance kernel with scale and smoothnes parameters $3.8$ and $4.0$ respectively.
\begin{figure}[H
\begin{center}
\subfigure[]{%
\label{fig:simlocation}
\includegraphics[width=0.42\textwidth,height=0.30\textheight]{locations.pdf}
}%
\subfigure[]{%
\label{fig:simgfunction}
\includegraphics[width=0.42\textwidth,height=0.30\textheight]{gfunction.pdf}
}\\
\end{center}
\caption{%
Panel (a) shows the distribution of the spatial coordinates where the data have been generated.
Panel (b) displays the plot of the evolutionary transformation $-1.1+0.5x+3\sin(\frac{\pi}{4}x)-5\sin(\frac{\pi}{5}x)$
over the range of $X(\bold{s}_{i},t)$.
}%
\label{fig:subfigures3}
\end{figure}
As is evident from the above figures the evolutionary transformation $-1.1+0.5x+3\sin(\frac{\pi}{4}x)-5\sin(\frac{\pi}{5}x)$
is highly nonlinear over the range of $X(\bold{s}_{i},t)$, which serves our purpose. Our prior specification is very similar to
the previous simulation study, where we have taken diffused normal priors for the location parameters and lognormal
priors for the scale and smoothness parameters. We have taken sufficiently large burn-in and convergence of the chain
is confirmed via visual monitoring. The computation is a bit time consuming with almost 60,000 iterations per day in a standard laptop.
This is however natural for dynamic models, as they are so high-dimensional. In the following set of figures we
describe how our model performs in terms of prediction.
\begin{figure}[H
\begin{center}
\subfigure[]{%
\label{fig:ourmodelfirst2}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{simt3median.pdf}
}%
\subfigure[]{%
\label{fig:ourmodelsecond2}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{simt3.pdf}
}\\
\subfigure[]{%
\label{fig:ourmodelthird2}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{simt11median.pdf}
}%
\subfigure[]{%
\label{fig:ourmodelfourth2}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{simt11.pdf}
}\\
\subfigure[]{%
\label{fig:ourmodelfifth2}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{sims2.pdf}
}%
\subfigure[]{%
\label{fig:ourmodelsixth2}
\includegraphics[width=0.40\textwidth,height=0.25\textheight]{sims8.pdf}
}\\
\end{center}
\caption{Leave-one-out posterior predictive performance of our model when the data are generated from a highly non-linear,
parametric model.}
\label{fig:subfigures4}
\end{figure}
A brief description of Section \ref{fig:subfigures4} follows.
Panel (a) displays the leave-one-out 95\% Bayesian prediction interval for all of the $20$ spatial locations at time point $t=3$.
The upper and lower quantiles (which form the intervals) are interpolated to form two surfaces.
The middle surface is obtained by interpolating the leave-one-out point predictions (which is the median in this case)
at the monitoring sites $\bold{s}_{1}, \bold{s}_{2},\cdots, \bold{s}_{20}$ for time point $t=3$. Panel (b) displays
once again the the leave-one-out 95\% Bayesian prediction interval for all of the $20$ spatial locations but
this time the middle surface is interpolated through the truly observed data (the truly observed data being
represented by black dots). Panel (c) and Panel (d) display plots similar as (a) and (b) but for time point $t=11$.
Panels (e) and (f) depict the leave-one-out 95\% Bayesian prediction interval for the time series at spatial locations
$\bold{s}_{2}$ and $\bold{s}_{8}$, respectively. The starred curve represents the curve interpolated through the
truly observed data (the observed data itself being represented by star) and the plain curves represent respectively
the upper, middle and lower quantiles associated with the posterior predictive distributions.
From the above set of figures it is evident that our model performs very well even though the data arose
from a completely different model.
\subsection{$SO_{2}$ precipitation over Great Britain and France}
Air pollution over large geographical regions is a topic of wide range of studies involving statistics and other disciplines.
Among them statistical modeling of pollution caused by $SO_{2}$ draws considerable attention. Here we consider a $SO_{2}$
precipitation dataset over Great Britain and France. The dataset consists of monthly measurement of sulphur dioxide
precipitation taken at $16$ monitoring stations spread over Great Britain and France, starting from April 1999 to January 2001.
This dataset is a part of data collected through the ‘European monitoring and evaluation programme’ (EMEP) which co-ordinates
the monitoring of airborne pollution over Europe. Further information is available at
{\tt {http://www.emep.int}}, the website from which we have obtained the data set.
First we consider some exploratory analysis based on the dataset. It is important to note that
instead of direct measurement what we have is the measurement of natural logarithm of $SO_{2}$ precipitation for these stations.
\begin{figure}[H
\begin{center}
\subfigure[]{%
\label{fig:realmap}
\includegraphics[width=0.42\textwidth,height=0.30\textheight]{maplatlong.pdf}
}%
\subfigure[]{%
\label{fig:realexploratory}
\includegraphics[width=0.42\textwidth,height=0.30\textheight]{exp.pdf}
}\\
\end{center}
\caption{%
Panel (a) shows the geographical location of the monitoring stations and
Panel (b) displays the time series plots of the data for each of the $16$ monitoring stations.
}%
\label{fig:subfigures5}
\end{figure}
We plot the station wise time series in a single figure and we see clear seasonal component and a very mild decreasing
trend in most of the time series plotes shown in Panel (b) of Figure \ref{fig:subfigures5}.
Then we estimate the trend and seasonal component of each of the time series separately using simple descriptive techniques
and detrend and deseasonalize them. Finally, we model the residual spatio-temporal data using our method.
The prior structure used in the real data analysis is similar to that of the simulation study. We used diffused
bivariate normal priors for $(\beta_{0g},\beta_{1g})$ and $(\beta_{0f},\beta_{1f})$ and lognormal priors for all the other parameters.
Since the monitoring stations are spread over a large geographical region and distance stretches horizontally as latitude
increases, the use of simple longitude and latitude as spatial coordinates would not be appropriate.
The Lambert (or Schmidt) projection addresses this
problem by preserving the area. This projection is defined by the transformation of longitude and latitude,
expressed in radians as $\psi $ and $\phi$, to the new co-ordinate system
$\mathbf{s}=(2\sin(\frac{\pi}{4}-\frac{\phi}{2})\sin(\psi),-2\sin(\frac{\pi}{4}-\frac{\phi}{2})\cos(\psi))$.
However, with respect to the temporal coordinate we simply take one month as one unit of time.
We implemented an MCMC chain with sufficiently large burn-in; visual inspection suggests satisfactory convergence.
Based on the MCMC samples we calculated leave-one-out posterior predictive distributions of $\ln{SO_{2}}$
at each of the monitoring stations for all the months ranging from April 1999 to January 2001.
We calculated the $95\%$ Bayesian prediction interval associated with the leave-one-out posterior
predictive distribution for each of the $352$ space time coordinates. More than $95\%$ of the data fell
within the respective prediction interval.
In Figures 6(a)-6(c) we provide the 95\% Bayesian prediction intervals associated with the leave-one-out
posterior predictive distributions for time series of length $22$ for different spatial locations.
These three plots show that our model performs satisfactorily also in terms of prediction at purely temporal level.
Similarly, in Figure 7(a)-7(i) we display the contourplots of the
leave-one-out posterior point prediction surface, along with the contourplot of the true data surfaces for different months.
The contour plots clearly show that our model performs very well in terms of point prediction.
We also display the interval prediction through the combined plot of the interpolating surfaces of
the upper quantiles, the true data and the lower quantiles, for the corresponding months. Overall, this shows
that our model performs well in terms of both point prediction and interval prediction at purely spatial level.
\begin{figure}[H
\begin{center}
\subfigure[]{%
\label{fig:realtime1}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{GB04timeseries.pdf}
}%
\subfigure[]{%
\label{fig:realtime2}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{FR13timeseries.pdf}
} \subfigure[]{%
\label{fig:realtime3}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{GB07timeseries.pdf}
}\\
\end{center}
\caption{%
Panel (a) shows the time series of the true data, the corresponding point predictions,
the upper and lower quantiles (form the prediction bands) at a particular spatial location.
The starred line represents the true data and the other middle curve denotes the point prediction.
Panel (b) and Panel (c) display similar time series plots for two more monitoring sites
}%
\label{fig:subfigures6}
\end{figure}
\newpage
\begin{figure}[H
\begin{center}
\subfigure[]{%
\label{fig:realcontour1}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{contourjuly1999predict.pdf}
}%
\subfigure[]{%
\label{fig:realcontour3}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{contourapr2000predict.pdf}
}%
\subfigure[]{%
\label{fig:realcontour5}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{contourjan2001predict.pdf}
}\\%
\subfigure[]{%
\label{fig:realcontour2}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{contourjuly1999data.pdf}
}
\subfigure[]{%
\label{fig:realcontour4}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{contourapr2000data.pdf}
}
\subfigure[]{%
\label{fig:realcontour6}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{contourjan2001data.pdf}
}\\
\subfigure[]{%
\label{fig:realsurf1}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{july1999.pdf}
}
\subfigure[]{%
\label{fig:realsurf2}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{apr2000.pdf}
}
\subfigure[]{%
\label{fig:realsurf3}
\includegraphics[width=0.30\textwidth,height=0.20\textheight]{jan2001.pdf}
}
\end{center}
\caption
Panels (a), (b), (c) show the contourplots of surfaces interpolated through leave-one-out point predictions for July 1999, April 2000 and January 2001 respectively. Panels (d), (e), (f) show the contourplots of surfaces
interpolated through the true data for July 1999, April 2000 and January 2001 respectively. Panels (g), (h), (i) display the true data surfaces and the corresponding upper and lower quantile surfaces for the same three months,
to demonstrate the uncertainty in prediction graphically. The red triangles indicate the observed values. The black dots in all the figures indicate the locations of the monitoring stations on the map
}%
\label{fig:subfigures}
\end{figure}
One point to be noted here is that we have fitted our model to the residual space-time dataset (after detrending and deseasonalizing the original dataset), not to the original $\ln{SO_{2}}$ measurements. So, while doing posterior prediction
(or leave-one-out posterior prediction) at a space-time location, we needed to add back the trend and seasonal components to obtain the correct predictions. Only then we can compare the posterior prediction and
the true value at any particular space-time location.
\section{Possible extensions}
A very important aspect of such space-time data is covariate information. Although in this article, we only consider model without any covariate information, it can be easily incorporated in our setup. In that case our model will be of the following form:
\begin{align}
Y(\bold{s},t)&=f(X(\bold{s},t),Z(\bold{s},t))+\epsilon(\bold{s},t),
\label{eqn:npr3}\\
X(\bold{s},t)&=g(X(\bold{s},t-1))+\eta(\bold{s},t),
\label{eqn:npr4}
\end{align}
where $Z(\bold{s},t)$ is the covariate process and $f(\cdot,\cdot)$ is now a Gaussian processes on $\mathbb{R}^2$. The rest of the theory remains the same as before. In fact, we can have $k$ different covariate processes
$Z_{1}(\bold{s},t),Z_{2}(\bold{s},t),\cdots,Z_{k}(\bold{s},t)$, in which case $f(\cdot,\cdot,\cdots,\cdot)$ will be a Gaussian process on $\mathbb{R}^{k+1}$.
Another direction for extension is time varying nonparametric state space model.
In that case we let the evolutionary and observational transformations vary with time, so that our model assumes the form:
\begin{align}
Y(\bold{s},t)&=f_{t}(X(\bold{s},t))+\epsilon(\bold{s},t),
\label{eqn:npr5}\\
X(\bold{s},t)&=g_{t}(X(\bold{s},t-1))+\eta(\bold{s},t);
\label{eqn:npr6}
\end{align}
Such modification is particularly useful when the time interval on which we are observing the data is very wide so that it is unlikely for the functions $f$ and $g$ to remain invariant with respect to time.
In the context of purely temporal nonparametric state space models, \cite{Ghosh14}
consider time-varying functions $f_t(\cdot)$ and $g_t(\cdot)$, which they re-write as $f(\cdot,t)$ and $g(\cdot,t)$,
respectively. In other words, they consider the time component as an argument of their random functions $f$ and $g$, which are modeled by Gaussian processes in the usual manner.
Such ideas can be easily adopted in our spatio-temporal model.
Currently, we are working on these extensions.
\section{Discussion and concluding remarks}
One common problem in the context of space-time data is missing observations. Malfunction of monitoring device, inexperienced handling etc., may lead to such problem. So, any good spatio-temporal model should take care of missing data.
In our case it is very simple. We augment the latent variable (say $x(\bold{s}^*,t^*)$) corresponding to the missing observation with $\{x(\bold{s}_{i},t)\}$ , $ i=1,\cdots,n $ and $ t=1,\cdots,T $. Then we can predict the missing
observation $y(\bold{s}^*,t^*)$ in the same way we make prediction at a new spatio-temporal coordinate (see Section $4$). Thus, prediction at a new spatio-temporal co-ordinate and missing data are taken care of simultaneously in our model.
Although we have developed a new nonstationary model, we have not done any comparative study with respect to to other existing nonstationary models (mentioned in section $2$). Unfortunately, these nonstationary models are built
on entirely different philosophies and hence are not comparable. For example, the nonstationary model developed in \cite{Higdon2} depends heavily on the choice of kernel and may outperform the nonstationary model developed in\cite{Guttorp:Meiring:Sampson} depending on whether some particular kernel is being used. In fact, Dou and Zidek (2010) \cite{Dou:Le:Zidek}, while comparing two different methods for analyzing hourly ozone data expressed similar view (see page 1209).
Similarly, our model, which is built on the idea of state space models is not comparable at all with the nonstationary models that are based on the ideas of kernel convolution or deformation. In fact, our model is not even comparable with
DSTM, another model based on the idea of state space. The reason behind it is that although our model reduces to the no-covariate DSTM model if $\beta_{0f}=\beta_{0g}=0$, $\beta_{1f}=\beta_{1g}=1$, $\sigma_{f}^{2}=\sigma_{g}^{2}=0$,
in spirit it is entirely different from DSTM models, which are generally used to utilize
covariate information in improving the prediction. Indeed, our objective is to utilize unobserved information to enhance prediction. Both of these models fall in the category of state space models, but the motivation behind them are quite different.
DSTM models are much more like spatio-temporal regression modeling where the linear regression structure varies with time (varying coefficient linear regression).
\section*{Acknowledgments} The research of the first author is fully funded by CSIR SPM Fellowship, Govt. of India. The authors thank ‘European monitoring and evaluation programme’ (EMEP) for the $SO_{2}$ dataset. They also thank Moumita Das for many fruitful discussions on this article.
\section*{Appendix}
Before proving the theorems let us make a notational clarification. The notations $[\bold{X}|\bold{Y}],[\bold{x}|\bold{y}]$
and $[\bold{X}=\bold{x}|\bold{Y}=\bold{y}]$ are equivalent and throughout this section they will denote the value of conditional pdf
of $\bold{X}$ given $\bold{Y}=\bold{y}$ at $\bold{X}=\bold{x}$.
\begin{proof}[Proof of Theorem 3.1]
For this proof we assume that there exist continuous modifications of the Gaussian processes that we consider, that is,
there exist processes with sample paths that are continuous everywhere, not just almost everywhere, and
that our Gaussian processes equals such processes with probability one (see, for example, see \cite{Adler81} for details).
Existence of such continous modifications are guaranteed under the correlation structure that we consider
for our Gaussian processes.
Let us first notice that, $\exists$ a probability space $(\Omega,\mathcal{F},P)$ such that
$$(g(x),X(\bold{s}_{1},0),\cdots,X(\bold{s}_{n},0)):
(\Omega,\mathcal{F})\rightarrow (C(\mathbb{R}),\mathcal{A})\bigotimes (\mathbb{R}^n,\mathcal{B}(\mathbb{R}^n))$$
where $C(\mathbb{R})$ is the space of all real valued continuous functions on $\mathbb{R}$ and $\mathcal{A}$ is the
Borel sigma field obtained from the topology of compact convergence on the space $C(\mathbb{R})$. Such a joint measurability
result need not hold unless $g(x)$ and $(X(\bold{s}_{1},0),\cdots,X(\bold{s}_{n},0))$ are independent. Now, we will show
that $g(X(\bold{s}_{1},0))$ is a measurable real valued function or a proper random variable. To show this, first see that
$g(X(\bold{s}_{1},0)):\Omega\rightarrow\mathbb{R}$ can be written as $T(g(\cdot),X(\bold{s}_{1},0))$ where
$T:C(\mathbb{R})\bigotimes\mathbb{R}\rightarrow\mathbb{R}$ is a transformation such that, $T(g,x)=g(x)$ where $g$ is a real valued continuous function on $\mathbb{R}$ and $x$ is a real number.
\begin{lemma}
$T:C(\mathbb{R})\bigotimes\mathbb{R}\rightarrow\mathbb{R}$ is a continuous transformation where the topology associated with
$C(\mathbb{R})$ is the topology of compact convergence and the topology associated with $\mathbb{R}$ is the usual Euclidean
distance based topology on real numbers.
\label{lem:measurable}
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lem:measurable}]
Let us consider the metric $d(g,g^\prime)=\mathlarger{\sum_{i=1}^{\infty}\frac{1}{2^i}
\frac{\sup\limits_{x\in [-i,i]}\abs{g(x)-g^\prime(x)}}{1+\sup\limits_{x\in[-i,i]}\abs{g(x)-g^\prime(x)}}}$. This metric induces the
topology of compact convergence on the space $C(\mathbb{R})$. To prove continuity of $T$ one needs to show that
$d(g_{n},g)\rightarrow 0$ and $\abs{x_{n}-x}\rightarrow 0 \Rightarrow \abs{T(g_{n},x_{n})-T(g,x)}\rightarrow 0$.
\\
Let us assume that $d(g_{n},g)\rightarrow 0$ and $\abs{x_{n}-x}\rightarrow 0$. So, $\exists \ N_{0}$ and $j_{0}$ such that
$\forall \ n\geq N_{0}$, $x_{n}\in [-j_{0},j_{0}]$.
\\
Now, $\mathlarger{\frac{1}{2^{j_{0}}}}\sup\limits_{x\in [-j_{0},j_{0}]}\abs{g_{n}(x)-g(x)}\leq d(g_{n},g) \Rightarrow
\sup\limits_{x\in [-j_{0},j_{0}]}\abs{g_{n}(x)-g(x)}\rightarrow 0$
\\
and $\abs{g(x_{n})-g(x)}\rightarrow 0$ because $g$ is continuous.
\\
But,
\begin{align*}
\abs{g_{n}(x_{n})-g(x)} \leq \abs{g_{n}(x_{n})-g(x_{n})} + \abs{g(x_{n})-g(x)}\\
\text{So,}\ \forall \ n\geq N_{0},\ \abs{g_{n}(x_{n})-g(x)} \leq \sup\limits_{x\in [-j_{0},j_{0}]}\abs{g_{n}(x)-g(x)} + \abs{g(x_{n})-g(x)}\\
\end{align*}
The RHS goes to $0$ as $n\rightarrow\infty$. Hence, $\abs{T(g_{n},x_{n})-T(g,x)}=\abs{g_{n}(x_{n})-g(x)}\rightarrow0$.
\end{proof}
Once continuity of $T$ is proved, note that $T^{-1}(U)$, for any open set $U \subseteq \mathcal{\mathbb{R}}$, is an open set
in the product topology on the space $C(\mathbb{R})\bigotimes\mathbb{R}$. Hence, $T^{-1}(U)$ belongs to the Borel sigma field
generated by this product topology which is in this case equivalent to the
product sigma field $\mathcal{A}\bigotimes\mathcal{B}(\mathbb{R})$ associated with $(C(\mathbb{R}),\mathcal{A})\bigotimes (\mathbb{R},\mathcal{B}(\mathbb{R}))$.
This equivalence holds because both of the spaces $C(\mathbb{R})$ and $\mathbb{R}$ are separable.
But $(g(x),X(\bold{s}_{1},0))$ is measurable with respect to $(\Omega,\mathcal{F})$ and $(C(\mathbb{R}),\mathcal{A})
\bigotimes (\mathbb{R},\mathcal{B}(\mathbb{R}))$. Hence, the inverse image of $T^{-1}(U)$ with respect to $(g(x),X(\bold{s}_{1},0))$
is in $\mathcal{F}$. So, the inverse image of any open set $U \subseteq \mathcal{\mathbb{R}}$ with respect to
$g(X(\bold{s}_{1},0))$ is in $\mathcal{F}$. This proves the measurability of $g(X(\bold{s}_{1},0))$.
Following exactly same argument as above we can further prove that $g(X(\bold{s}_{2},0)),\cdots,g(X(\bold{s}_{n},0))$ are jointly measurable.
Now, as $\eta(\bold{s},t)$ is independent of $g(x)$ and $(X(\bold{s}_{1},0),\cdots,X(\bold{s}_{n},0))$, we have the joint measurability of
$(X(\bold{s}_{1},1),\cdots,X(\bold{s}_{n},1))$ (See \ref{eqn:npr2}). Infact, we can prove that $(g(x),X(\bold{s}_{1},1),\cdots,X(\bold{s}_{n},1))$ are jointly measurable. To do it we consider $T^{\prime}:C(\mathbb{R})\bigotimes\mathbb{R}\rightarrow C(\mathbb{R})\bigotimes\mathbb{R}$ such that $T^{\prime}(g,x)=(g,g(x))$ where $g$ is a real valued continuous function on $\mathbb{R}$ and $x$ is a real number. Then similarly as in case of $T$ we can prove that $T^{\prime}$ is also a continuous map which immediately implies that $(g,g(X(\bold{s}_{1},0)))$ are jointly measurable. Then $\eta(\bold{s},t)$ being independent of $g(x)$ and $(X(\bold{s}_{1},0),\cdots,X(\bold{s}_{n},0))$, implies the joint measurability of
$(g(x),X(\bold{s}_{1},1),\cdots,X(\bold{s}_{n},1))$.
Hence, starting with the joint measurability of
$(g(x),X(\bold{s}_{1},0),\cdots,X(\bold{s}_{n},0))$ we prove the joint measurability of $(X(\bold{s}_{1},1),\cdots,X(\bold{s}_{n},1))$ and $(g(x),X(\bold{s}_{1},1),\cdots,X(\bold{s}_{n},1))$.
Similarly, if we start with joint measurability of $(g(x),X(\bold{s}_{1},1),\cdots,X(\bold{s}_{n},1))$ we can prove the joint measurability
of $(X(\bold{s}_{1},2),\cdots,X(\bold{s}_{n},2))$. Thus, the joint measurability of the whole collection of state variables
${X(\bold{s}_{i},t)}$ $\forall\ i=1,2,\cdots,n;t=0,1,\cdots,T $ is mathematically established.
Now, to prove joint measurability of the collection of observed variables
${Y(\bold{s}_{i},t)}$ $\forall\ i=1,2,\cdots,n;t=1,\cdots,T $ recall the observational equation (1).
Since, $f(x)$ takes values in $(C(\mathbb{R}),\mathcal{A})$ just as $g(x)$,
and also since $\epsilon(\bold{s},t)$ is independent of $f(s)$ just as $\eta(\bold{s},t)$ is independent of $g(x)$,
all the previous arguments go through in this case and joint measurability of ${Y(\bold{s}_{i},t)}$ $\forall\ i=1,2,\cdots,n;t=0,1,\cdots,T $
is established.
Finally, it remains to show that a valid spatio-temporal process is induced by this model. But it is immediate from the application
of Kolmogorov consistency theorem. The consistency conditions of the theorem are trivially satisfied by our construction and hence the result follows.
\end{proof}
\begin{proof}[Alternative proof of Theorem 3.1]
In the previous proof, we needed to assume the continous modification of the underlying Gaussian process.
Here we consider a proof where the underlying Gaussian process need not be continuous. In fact, the alternative proof
that we now provide is valid even if the underlying process admits at most countable number of discontinuities.
Note that it is possible to represent any stochastic process $\{Z(\bold{s});\bold{s}\in T\}$, for fixed
$\bold{s}$ as a random variable $\omega\mapsto Z(\bold{s},\omega)$, where $\omega\in\Omega$;
$\Omega$ being the set of all functions from $T$ into $\mathbb R$.
Also, fixing $\omega\in\Omega$, the function $\bold{s}\mapsto Z(\bold{s},\omega);~\bold{s}\in T$,
represents a path of $Z(\bold{s});\bold{s}\in T$. Indeed, we can identify $\omega$ with the function
$\bold{s}\mapsto Z(\bold{s},\omega)$ from $T$ to $\mathbb R$; see, for example, \cite{Oksendal00}, for
a lucid discussion.
This latter identification will be convenient for our purpose, and we adopt this for proving our result
on measurability.
Note that the $\sigma$-algebra $\mathcal F$ associated with $Z$
is generated by sets of the form
\[
\left\{\omega:\omega(\bold{s}_1)\in B_1,\omega(\bold{s}_2)\in B_2,\ldots,\omega(\bold{s}_k)\in B_k\right\},
\]
where $B_i\subset\mathbb R;i=1,\ldots,k$, are Borel sets in $\mathbb R$.
In our case, the Gaussian process $g(\cdot)$ can be identified with $g(x)(\omega_1)=\omega_1(x)$, for any fixed
$x\in\mathbb R$ and $\omega_1\in\Omega_1$, where $\Omega_1$ is the set of all functions from $\mathbb R$ to $\mathbb R$.
The {\it initial} Gaussian process $X(\cdot,0)$ can be identified with
$X(\bold{s},0)(\omega_2)=\omega_2(\bold{s})$, where $\bold{s}\in\mathbb R^d$ $(d\geq 2)$ and $\omega_2\in\Omega_2$.
Here $\Omega_2$ is the set of all functions from $\mathbb R^d$ to $\mathbb R$. Let $\mathcal F_1$ and $\mathcal F_2$
be the Borel $\sigma$-fields associated with $\Omega_1$ and $\Omega_2$, respectively.
We first show that the composition of $g(\cdot)$ with
$X(\cdot,0)$, given by $g(X(\bold{s},0))$ is a measurable random variable for any $\bold{s}$
is a measurable function. Since $g$ and $X(\cdot,0)$ are independent, we need to consider the
product space $\Omega_1\otimes\Omega_2$, and noting that $g(X(\bold{s},0)(\omega_2))(\omega_1)=\omega_1(\omega_2(\bold{s}))$,
where $(\omega_1,\omega_2)\in\Omega_1\otimes\Omega_2$, need to show that sets of the form
\[
A(\bold{s}_1,\ldots,\bold{s}_k)=\left\{(\omega_1,\omega_2):\omega_1(\omega_2(\bold{s}_1))\in B_1,\omega_1(\omega_2(\bold{s}_2))\in B_2,\ldots,
\omega_1(\omega_2(\bold{s}_k))\in B_k\right\},
\]
where $B_i\subset\mathbb R;i=1,\ldots,k$, are Borel sets in $\mathbb R$, are in $\mathcal F_1\otimes\mathcal F_2$,
the product Borel $\sigma$-field associated with $\Omega_1\otimes\Omega_2$.
For our purpose, we let $B_i$ be of the form $[a_i,b_i]$ for real values $a_i<b_i$.
Now, suppose that $(\omega_1,\omega_2)\in A(\bold{s}_1,\ldots,\bold{s}_k)$.
Then $\omega_1(\omega_2(\bold{s}_i))\in [a_i,b_i]$, which implies that $\omega_2(\bold{s}_i)$ is at most a countable
union of sets of the form $[a^{(i)}_j,b^{(i)}_j];~j\in\mathcal D_i$, where $\mathcal D_i$ is a countable set of indices.
Also, it holds that $\omega_1(x^*)\in [a_i,b_i];~\forall~x^*\in\mathbb Q\cap\left\{\underset{j\in\mathcal D_i}
\cup [a^{(i)}_j,b^{(i)}_j]\right\}$.
Here $\mathbb Q$ is the countable set of rationals in $\mathbb R$.
If necessary, we can envisage a countable set
$\mathcal D^*$ consisting of points of discontinuities of $\omega_1$. If $\xi$ is a point of discontinuity, then
$\omega_1(\xi)$ may be only the left limit of particular sequence $\{\omega_1(\xi_{1,m});m=1,2,\ldots\}$ or only
the right limit of a
particular sequence $\{\omega_1(\xi_{2,m});m=1,2,\ldots\}$, or $\omega_1(\xi)$ may be an isolated point,
not reachable by sequences
of the above forms.
It follows that $(\omega_1,\omega_2)$ must lie in
\[
A^*(\bold{s}_1,\ldots,\bold{s}_k)=\cap_{i=1}^k\left\{(\omega_1,\omega_2):\omega_1(x)\in [a_i,b_i]~\forall~x\in
\left(\mathbb Q\cap\left\{\underset{j\in\mathcal D_i}\cup [a^{(i)}_j,b^{(i)}_j]\right\}\right)\cup\mathcal D^*,
\omega_2(\bold{s}_i)\in\underset{j\in\mathcal D_i}\cup [a^{(i)}_j,b^{(i)}_j]\right\}.
\]
Now, if $(\omega_1,\omega_2)\in A^*(\bold{s}_1,\ldots,\bold{s}_k)$, then, noting that for any point
$x\in \underset{j\in\mathcal D_i}\cup [a^{(i)}_j,b^{(i)}_j]$ of $\omega_1$,
$\omega_1(x)=\underset{m\rightarrow\infty}\lim \omega_1(\xi_m)$, where
$\{\xi_m;m=1,2,\ldots\}\in \mathbb Q\cap\left\{\underset{j\in\mathcal D_i}\cup [a^{(i)}_j,b^{(i)}_j]\right\}$,
it is easily seen that $(\omega_1,\omega_2)\in A(\bold{s}_1,\ldots,\bold{s}_k)$. Hence,
$A(\bold{s}_1,\ldots,\bold{s}_k)=A^*(\bold{s}_1,\ldots,\bold{s}_k)$.
Now observe that $A^*(\bold{s}_1,\ldots,\bold{s}_k)$ is a finite intersection of countable union of measurable sets;
hence, $A^*(\bold{s}_1,\ldots,\bold{s}_k)$ is itself a measurable set.
In other words, we have proved that $g(X(\cdot,0))$ is measurable.
Now, as $\eta(\cdot,t)$ is independent of $g(\cdot)$ and $X(\cdot,0)$, it follows from (\ref{eqn:npr2}) that
$X(\cdot,1)$ is measurable.
To prove measurability of $X(\cdot,2)$, note that
\begin{align}
X(\bold{s},2)&=g(X(\bold{s},1))+\eta(\bold{s},2)\notag\\
&=g(g(X(\bold{s},0))+\eta(\bold{s},1))+\eta(\bold{s},2).
\label{eq:2nd}
\end{align}
The process $\eta(\cdot,1)$ requires introduction an extra sample space $\Omega_3$, so that
we can identify $\eta(\bold{s},1)(\omega_3)$ as $\omega_3(\bold{s})$. With this, we can represent
$g(g(X(\bold{s},0))+\eta(\bold{s},1))$ of (\ref{eq:2nd}) as
$\omega_1(\omega_1(\omega_2(\bold{s}))+\omega_3(\bold{s}))$.
Now, $\omega_1(\omega_1(\omega_2(\bold{s}))+\omega_3(\bold{s}))\in [a_i,b_i]$ implies that
$\omega_1(\omega_2(\bold{s}))+\omega_3(\bold{s})\in\underset{j\in\mathcal D_i}\cup[a^{(j)}_i,b^{(j)}_i]$.
If $\omega_1(\omega_2(\bold{s}))+\omega_3(\bold{s})\in [a^{(k)}_i,b^{(k)}_i]$ for some $k\in\mathcal D_i$,
the the set of solutions is
\begin{equation}
\underset{r\in\mathbb R}\cup\left\{\omega_1(\omega_2(\bold{s}))\in [a^{(k)}_i-r,b^{(k)}_i-r],w_3(\bold{s})=r\right\},
\label{eq:uncountable}
\end{equation}
where $\omega_1(\omega_2(\bold{s}))\in [a^{(k)}_i-r,b^{(k)}_i-r]$ implies, as before, that $\omega_2(\bold{s})$
belongs to a countable union of measurable sets in $\mathbb R$. Although the set (\ref{eq:uncountable})
is an uncountable union, following the technique used for proving measurability of $g(X(\cdot,0))$, we will intersect the set by $\mathbb Q$,
the (countable) set of rationals in $\mathbb R$; this will render the intersection a countable set. The proof of measurability
then follows similarly as before.
Proceeding likewise, we can prove measurability of $X(\cdot,t)$ is measurable for $t=2,3,\ldots$. Proceeding exactly in the same
way, we can also prove that $Y(\cdot,t);~t=1,2,\ldots,T$ are measurable.
Moreover, it can be easily seen that the same methods employed for proving the above results on measurability can be
extended in a straightforward (albeit notationally cumbersome) manner to prove that the sets of the forms
\[
\left\{X(\bold{s}_i,t_i)\in [a_i,b_i];i=1,\ldots,k\right\}\ \ \mbox{and} \ \
\left\{Y(\bold{s}_i,t_i)\in [a_i,b_i];i=1,\ldots,k\right\}
\]
are also measurable. Furthermore, it can be easily verified that $X$ and $Y$ satisfy Kolmogorov's consistency
criteria. In other words, $X$ and $Y$ are well-defined stochastic processes in both space and time.
\end{proof}
\begin{proof}[Proof of Theorem 3.2
Let us first observe that conditional on $g(x)$ our latent process satisfies the Markov property.
That is,
\begin{align}
&[(x(\bold{s}_{1},t),\cdots,x(\bold{s}_{n},t))\mid (g(x(\bold{s}_{1},t-1)),\cdots,g(x(\bold{s}_{n},t-1))),(x(\bold{s}_{1},t-1),\notag\\
&\quad\quad\cdots,x(\bold{s}_{n},t-1)),(x(\bold{s}_{1},t-2),\cdots,x(\bold{s}_{n},t-2)),\cdots,(x(\bold{s}_{1},0),\cdots,x(\bold{s}_{n},0))]\notag\\
&= [(x(\bold{s}_{1},t),\cdots,x(\bold{s}_{n},t))\mid (g(x(\bold{s}_{1},t-1)),\cdots,g(x(\bold{s}_{n},t-1))),(x(\bold{s}_{1},t-1),
\cdots,x(\bold{s}_{n},t-1))]\notag\\
&\sim \mathlarger{\frac{1}{|\mathbf{\Sigma_{\eta}}|^\frac{1}{2}}}
\exp\left[-\frac{1}{2}{\begin{pmatrix}x(\bold{s}_{1},t)-g(x(\bold{s}_{1},t-1))\\
x(\bold{s}_{2},t)-g(x(\bold{s}_{2},t-1))\\
\vdots\\
x(\bold{s}_{n},t)-g(x(\bold{s}_{n},t-1))\end{pmatrix}}^{\prime}{{\mathbf{\Sigma}}_{\eta}}^{-1}
{\begin{pmatrix}x(\bold{s}_{1},t)-g(x(\bold{s}_{1},t-1))\\
x(\bold{s}_{2},t)-g(x(\bold{s}_{2},t-1))\\
\vdots\\
x(\bold{s}_{n},t)-g(x(\bold{s}_{n},t-1))\end{pmatrix}}\right],\notag
\end{align}
\vspace{3mm}
where $[x\mid y]$ denotes the conditional density of $X$ at $x$ given $Y=y$.
Now, let us represent $g(x(\bold{s}_{i},t-1))$ by $u(i,t)$ for all $i=1,\cdots,n$ and $t=1,2,\cdots,T$.
Then repeatedly using the Markov property we have following
\begin{align*}
[x(\bold{s}_{1},T),\cdots,x(\bold{s}_{n},T),\cdots,x(\bold{s}_{1},0),\cdots,x(\bold{s}_{n},0) & \mid g(x(\bold{s}_{1},T-1)),\cdots\\
\cdots,g(x(\bold{s}_{n},T-1)),\cdots,g(x(\bold{s}_{1},0))& ,\cdots,g(x(\bold{s}_{n},0))]\\
\end{align*}
\begin{align*}
\sim
[x(\bold{s}_{1},T)& ,\cdots,x(\bold{s}_{n},T)\mid g(x(\bold{s}_{1},T-1)),\cdots,g(x(\bold{s}_{n},T-1)),
x(\bold{s}_{1},T-1),\cdots,x(\bold{s}_{n},T-1)]\times\\
\dotsm \times &[x(\bold{s}_{1},1),\cdots,x(\bold{s}_{n},1)\mid g(x(\bold{s}_{1},0)),\cdots,g(x(\bold{s}_{n},0)),
x(\bold{s}_{1},0),\cdots,x(\bold{s}_{n},0)]\\
& \times [x(\bold{s}_{1},0),\cdots,x(\bold{s}_{n},0)]\\
\vspace{3mm}
\sim \mathlarger{\frac{1}{(2\pi)^\frac{nT}{2}}}&\mathlarger{\frac{1}{|\mathbf{\Sigma_{\eta}}|^\frac{T}{2}}}\prod_{t=1}^{T}
\exp\left[-\frac{1}{2}{\begin{pmatrix}x(\bold{s}_{1},t)-u(1,t)\\x(\bold{s}_{2},t)-u(2,t)\\
\vdots\\
x(\bold{s}_{n},t)-u(n,t)\end{pmatrix}}^{\prime}{{\mathbf{\Sigma}}_{\eta}}^{-1}
{\begin{pmatrix}x(\bold{s}_{1},t)-u(1,t)\\x(\bold{s}_{2},t)-u(2,t)\\ \vdots\\x(\bold{s}_{n},t)-u(n,t)\end{pmatrix}}\right]\\
\vspace{4mm}
\times\mathlarger{\frac{1}{(2\pi)^\frac{n}{2}}}&\mathlarger{\frac{1}{|\mathbf{\Sigma}_{0}|^\frac{1}{2}}}
\exp\left[-\frac{1}{2}{\begin{pmatrix}x(\bold{s}_{1},0)-\mu_{01}\\
x(\bold{s}_{2},0)-\mu_{02}\\
\vdots\\
x(\bold{s}_{n},0)-\mu_{0n}\end{pmatrix}}^{\prime}{{\mathbf{\Sigma}}_{0}}^{-1}
{\begin{pmatrix}x(\bold{s}_{1},0)-\mu_{01}\\x(\bold{s}_{2},0)-\mu_{02}\\ \vdots\\x(\bold{s}_{n},0)-\mu_{0n}\end{pmatrix}}\right]
\end{align*}
But, this is the joint density of the state variables conditioned on $g(x)$. To obtain the joint density of the state variables
one needs to marginalize it with respect to the Gaussian process $g(\cdot)$.
After marginalization, the joint density takes the following form:
\begin{align*}
\mathlarger{\frac{1}{(2\pi)^\frac{n(T+1)}{2}}}\mathlarger{\frac{1}{|\mathbf{\Sigma}_{0}|^\frac{1}{2}}}
\mathlarger{\frac{1}{|\mathbf{\Sigma_{\eta}}|^\frac{T}{2}}}
\exp\left[-\frac{1}{2}{\begin{pmatrix}x(\bold{s}_{1},0)-\mu_{01}\\x(\bold{s}_{2},0)-\mu_{02}\\
\vdots\\
x(\bold{s}_{n},0)-\mu_{0n}\end{pmatrix}}^{\prime}{{\mathbf{\Sigma}}_{0}}^{-1}
{\begin{pmatrix}x(\bold{s}_{1},0)-\mu_{01}\\x(\bold{s}_{2},0)-\mu_{02}\\ \vdots\\x(\bold{s}_{n},0)-\mu_{0n}\end{pmatrix}}\right]\\
\end{align*}
\begin{align*}
\times \mathlarger{\mathlarger{\int_{\mathbb{R}^{nT}}}}\prod_{t=1}^{T}
\exp\left[-\frac{1}{2}{\begin{pmatrix}x(\bold{s}_{1},t)-u(1,t)\\
x(\bold{s}_{2},t)-u(2,t)\\
\vdots\\
x(\bold{s}_{n},t)-u(n,t)\end{pmatrix}}^{\prime}
{{\mathbf{\Sigma}}_{\eta}}^{-1}
{\begin{pmatrix}x(\bold{s}_{1},t)-u(1,t)\\x(\bold{s}_{2},t)-u(2,t)\\
\vdots\\x(\bold{s}_{n},t)-u(n,t)\end{pmatrix}}\right]
\mathlarger{\frac{1}{(2\pi)^\frac{nT}{2}}}\mathlarger{\frac{1}{|\mathbf{\Sigma}|^\frac{1}{2}}}\\
\exp\left[-\frac{1}{2}{\begin{pmatrix}u(1,1)-\beta_{0g}-\beta_{1g}x(\bold{s}_{1},0)\\u(2,1)-\beta_{0g}-\beta_{1g}x(\bold{s}_{2},0)\\
\vdots\\u(n,T)-\beta_{0g}-\beta_{1g}x(\bold{s}_{n},T-1)\end{pmatrix}}^{\prime}{{\mathbf{\Sigma}}}^{-1}
{\begin{pmatrix}u(1,1)-\beta_{0g}-\beta_{1g}x(\bold{s}_{1},0)\\u(2,1)-\beta_{0g}-\beta_{1g}x(\bold{s}_{2},0)\\
\vdots\\u(n,T)-\beta_{0g}-\beta_{1g}x(\bold{s}_{n},T-1)\end{pmatrix}}\right]d\mathbf{u}\\
\end{align*}
where $\mathbf{\Sigma}$ is as in (3.2).
This is nothing but a convolution of two $\mathbb{R}^{nT}$ dimensional Gaussian densities, one with mean vector $\mathbf{0}$
and covariance matrix ${\mathbf{I}}_{T\times T}\bigotimes\mathbf{\Sigma_{\eta}}$ and the other one with mean vector \\
$(\beta_{0g}+\beta_{1g}x(\bold{s}_{1},0),\cdots,\beta_{0g}+\beta_{1g}x(\bold{s}_{n},T-1))^\prime $ and covariance matrix $\mathbf{\Sigma}$.\\
Hence, the integral boils down to
\begin{align*}
\mathlarger{\frac{1}{(2\pi)^\frac{n}{2}}}\mathlarger{\frac{1}{|\mathbf{\Sigma}_{0}|^\frac{1}{2}}}
\exp\left[-\frac{1}{2}{\begin{pmatrix}x(\bold{s}_{1},0)-\mu_{01}\\x(\bold{s}_{2},0)-\mu_{02}\\
\vdots\\x(\bold{s}_{n},0)-\mu_{0n}\end{pmatrix}}^{\prime}{{\mathbf{\Sigma}}_{0}}^{-1}
{\begin{pmatrix}x(\bold{s}_{1},0)-\mu_{01}\\x(\bold{s}_{2},0)-\mu_{02}\\ \vdots\\x(\bold{s}_{n},0)-\mu_{0n}\end{pmatrix}}\right]
\mathlarger{\mathlarger{\frac{1}{(2\pi)^\frac{nT}{2}}\frac{1}{|\tilde{\mathbf{\Sigma}}|^\frac{1}{2}}}}\\
\times\exp\left[-\frac{1}{2}{\begin{pmatrix}x(\bold{s}_{1},1)-\beta_{0g}-\beta_{1g}
x(\bold{s}_{1},0)\\x(\bold{s}_{2},1)-\beta_{0g}-\beta_{1g}x(\bold{s}_{2},0)\\
\vdots\\x(\bold{s}_{n},T)-\beta_{0g}-\beta_{1g}
x(\bold{s}_{n},T-1)\end{pmatrix}}^{\prime}{\tilde{\mathbf{\Sigma}}}^{-1}{\begin{pmatrix}x(\bold{s}_{1},1)-\beta_{0g}-\beta_{1g}
x(\bold{s}_{1},0)\\x(\bold{s}_{2},1)-\beta_{0g}-\beta_{1g}
x(\bold{s}_{2},0)\\ \vdots\\x(\bold{s}_{n},T)-\beta_{0g}-\beta_{1g}
x(\bold{s}_{n},T-1)\end{pmatrix}}\right],
\end{align*}
where $\tilde{\mathbf{\Sigma}}$ is as in (3.2).
\end{proof}
\begin{proof}[Proof of Theorem 3.3
First, see that for fixed $x(\bold{s}_{i},t_{1})$, $Y(\bold{s}_{i},t_{1})$ is distributed as a Gaussian with mean
$\beta_{0f}+\beta_{1f}x(\bold{s}_{i},t_{1})$ and variance ${\sigma_{f}}^2+{\sigma_{\epsilon}}^2$ where ${\sigma_{f}}^2$ and
${\sigma_{\epsilon}}^2$ are respectively the process variance associated with the isotropic Gaussian processes $\epsilon(\cdot,t)$
and $f(x)$ (see (3) and (1)).
Now, see that for fixed $x(\bold{s}_{i},t_{1})$ and $x(\bold{s}_{j},t_{2})$,
$f(x(\bold{s}_{i},t_{1}))$ and $f(x(\bold{s}_{j},t_{2}))$ has covariance $c_{f}(x(\bold{s}_{i},t_{1}),x(\bold{s}_{j},t_{2}))$.
Also, $\epsilon(\cdot,t_{1})$ and $\epsilon(\cdot,t_{2})$ are mutually independent spatial Gaussian processes for $t_{1}\neq t_{2}$.
Hence, conditional on state variables the covariance between $Y(\bold{s}_{i},t_{1})$ and $Y(\bold{s}_{j},t_{2})$ is
$c_{f}(x(\bold{s}_i,t_1),x(\bold{s}_j,t_2))+c_{\epsilon}(\bold{s}_i,\bold{s}_j)\delta(t_1-t_2)$. Here $\delta(\cdot)$ is the
delta function i.e. $\delta(t)=1$ for $t=0$ and $=0$ otherwise.
So, the joint density of the observed variables, which is denoted by $[y(\bold{s}_{1},1),y(\bold{s}_{2},1),\cdots,y(\bold{s}_{n},T)]$,
is given by
\begin{align*}
&[y(\bold{s}_{1},1),y(\bold{s}_{2},1),\cdots,y(\bold{s}_{n},T)]=\\
&\mathlarger{\int_{\mathbb{R}^{nT}}}[y(\bold{s}_{1},1),y(\bold{s}_{2},1),\cdots,y(\bold{s}_{n},T)\mid x(\bold{s}_{1},1),
x(\bold{s}_{2},1),\cdots,x(\bold{s}_{n},T)][x(\bold{s}_{1},1),x(\bold{s}_{2},1),\cdots,x(\bold{s}_{n},T)]d\mathbf{x}
\end{align*}
Hence, part $(a)$ follows.
For part $(b)$ note that if $\sigma_{f}^{2}=0$, the conditional density\\
$[y(\bold{s}_{1},1),y(\bold{s}_{2},1),\cdots,y(\bold{s}_{n},T)\mid x(\bold{s}_{1},1),x(\bold{s}_{2},1),\cdots,x(\bold{s}_{n},T)]$
is Gaussian with block diagonal covariance matrix ${\mathbf{I}}_{T\times T}\bigotimes\mathbf{\Sigma_{\epsilon}}$. On the other hand,
we have already noted that if $\sigma_{g}^{2}=0$, the joint density of the state variables boils down to Gaussian
(see the discussion following Theorem\ref{thm:state}). Let us consider only the state variables from time $t=1$ onwards.
They jointly follow an $nT$ dimensional Gaussian distribution. It is not difficult to see that the mean vector and the covariance matrix
of the $nT$ dimensional Gaussian distribution are of following forms:
\begin{align*}
\text{the $((t-1)n+i)$ th entry of the mean vector is}
\ \beta_{1g}^{t}\mu_{0i}
\beta_{0g}\frac{\beta_{1g}^{t}-1}{\beta_{1g}-1}\\
\text{where $1\leq t\leq T$}\\
\text{and the $(((t_{1}-1)n+i),((t_{2}-1)n+j))$ th entry of the covariance matrix is}\\
\beta_{1g}^{t_{1}+t_{2}}\sigma_{i,j}^{0}+(\beta_{1g}^{t_{1}+t_{2}-2}+\beta_{1g}^{t_{1}+t_{2}-4}+\cdots+\beta_{1g}^{\abs{t_{1}-t_{2}}})
c_{\eta}(s_{i},s_{j})\ \text{where $1\leq t_{1},t_{2}\leq T$ and $1\leq i,j\leq n$}\\
\sigma_{i,j}^{0} \text{ is the $(i,j)$ th entry of the covariance matrix } \mathbf{\Sigma_{0}}
\end{align*}
Now, using part $(a)$ we see that the joint distribution of $Y(\bold{s}_{1},1),Y(\bold{s}_{2},1),\cdots,Y(\bold{s}_{n},T)$ is nothing but
a convolution of two $\mathbb{R}^{nT}$-dimensional Gaussian densities. Hence, it is a Gaussian distribution whose mean vector
has $((t-1)n+i)$ th entry as $\beta_{0f}+\beta_{1f}\left(\beta_{1g}^{t}\mu_{0i}
+\beta_{0g}\frac{\beta_{1g}^{t}-1}{\beta_{1g}-1}\right)\\
\text{where $1\leq t\leq T$}$ and the $(((t_{1}-1)n+i),((t_{2}-1)n+j))$ th entry of the covariance matrix is \\
$\beta_{1f}^{2}\left(\beta_{1g}^{t_{1}+t_{2}}\sigma_{i,j}^{0}+(\beta_{1g}^{t_{1}+t_{2}-2}+\beta_{1g}^{t_{1}+t_{2}-4}
+\cdots+\beta_{1g}^{\abs{t_{1}-t_{2}}})c_{\eta}(s_{i},s_{j})\right)+c_{\epsilon}(\bold{s}_{i},\bold{s}_{j})\delta(t_{1}-t_{2})$ \\
where $1\leq t_{1},t_{2}\leq T$ and $1\leq i,j\leq n$. So, part $(b)$ is proved.
\end{proof}
\begin{proof}[Proof of Theorem 3.4
part $(a)$ We first show that $Var(X(\bold{s},t))$ is finite. Then by using the formula
\begin{align*}
Var(Y(\bold{s},t))&=Var\left(E(Y(\bold{s},t)|X(\bold{s},t)\right))+E\left(Var(Y(\bold{s},t)|X(\bold{s},t)\right))\\
&=Var(\beta_{0f}+\beta_{1f}X(\bold{s},t))+E(\sigma_{f}^2+\sigma_{\epsilon}^2)\\
&=\beta_{1f}^2Var(X(\bold{s},t))+\sigma_{f}^2+\sigma_{\epsilon}^2
\end{align*}
we establish that $Var(Y(\bold{s},t))$ is finite.
To show $Var(X(\bold{s},t))$ is finite we use principle of mathematical induction, i.e. we first show that $Var(X(\bold{s},0))$ is
finite and then we show that if\\
$Var(X(\bold{s},0)),Var(X(\bold{s},1)),\cdots,Var(X(\bold{s},t-1))$
are finite
then $Var(X(\bold{s},t))$ is finite. These two steps together compel $Var(X(\bold{s},t))$ to be finite.
The first step is trivially shown as $X(\bold{s},0)$ is a Gaussian random variable.
Now we show the second step of mathematical induction, that is, we show that if
\\
$Var(X(\bold{s},0)),Var(X(\bold{s},1)),\cdots,Var(X(\bold{s},t-1))$ are finite then $Var(X(\bold{s},t))$ is finite.
Now consider the following:
\begin{align*}
&Var(X(\bold{s},t)|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0})\\
&=Var(g(X(\bold{s},t-1))+\eta(\bold{s},t)|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0})\\
&=Var(g(X(\bold{s},t-1))|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0})+\sigma_{\eta}^2\\
&=Var(g(x_{t-1})|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0})+\sigma_{\eta}^2\\
&=Var(g(x_{t-1})|g(x_{t-2})+\eta(\bold{s},t-1)=x_{t-1},\cdots,g(x_{0})+\eta(\bold{s},1)=x_{1},X(\bold{s},0)=x_{0})+\sigma_{\eta}^2\\
&=\sigma_{g}^2-\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I})^{-1}\bold{\Sigma}_{g12}+\sigma_{\eta}^2\ \
\text{[see page 16 of \cite{Rasmussen:Williams}]}
\end{align*}
where $\bold{\Sigma}_{g12}^{'}$ is the row vector $(c_{g}(x_{t-1},x_{0})\ c_{g}(x_{t-1}x_{1})\cdots\ c_{g}(x_{t-1},x_{t-2}))$
and $\bold{\Sigma}_{g22}$ is the variance covarince matrix $\begin{pmatrix}c_{g}(x_{0},x_{0})\ c_{g}(x_{0},x_{1})\cdots c_{g}(x_{0},x_{t-2})
\\
c_{g}(x_{1},x_{0})\ c_{g}(x_{1},x_{1})\cdots c_{g}(x_{1},x_{t-2})\\
\vdots\\c_{g}(x_{t-2},x_{0})\ c_{g}(x_{t-2},x_{1})\cdots c_{g}(x_{t-2},x_{t-2})\end{pmatrix}$ induced by covariance function $c_{g}(\cdot,\cdot)$.
Now, we consider $E(Var(X(\bold{s},t)|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0}))$.
We want to show that this quantity is finite. But the problem is that we have to deal with the inverse of a random matrix
$(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I}) $. Fortunately,
the random matrix $(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I}) $ is non-negative definite (nnd). Hence,
$\sigma_{g}^2-\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I})^{-1}\bold{\Sigma}_{g12}
+\sigma_{\eta}^2\leq \sigma_{g}^2+\sigma_{\eta}^2$. On the other hand, this quantity being a conditional variance is always nonnegative.
So, the following inequality holds\\
$$0\leq \sigma_{g}^2-\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I})^{-1}\bold{\Sigma}_{g12}
+\sigma_{\eta}^2\leq \sigma_{g}^2+\sigma_{\eta}^2.$$
Hence, it follows that
$$0\leq E(\sigma_{g}^2-\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I})^{-1}\bold{\Sigma}_{g12}
+\sigma_{\eta}^2)\leq \sigma_{g}^2+\sigma_{\eta}^2.$$ So, the quantity $E(Var(X(\bold{s},t)|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)
=x_{t-2}\cdots,X(\bold{s},0)=x_{0}))$ being equivalent to $E(\sigma_{g}^2-\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}
+\sigma_{\eta}^2\bold{I})^{-1}\bold{\Sigma}_{g12}+\sigma_{\eta}^2)$, is finite.\\
Now we consider the term $E(X(\bold{s},t)|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0})$.
\begin{align*}
&E(X(\bold{s},t)|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0})\\
&=E(g(X(\bold{s},t-1))+\eta(\bold{s},t)|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0})\\
&=E(g(X(\bold{s},t-1))|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0})+0\\
&=E(g(x_{t-1})|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0})\\
&=E(g(x_{t-1})|g(x_{t-2})+\eta(\bold{s},t-1)=x_{t-1},\cdots,g(x_{0})+\eta(\bold{s},1)=x_{1},X(\bold{s},0)=x_{0})\\
&=\beta_{g0}+\beta_{g1}x_{t-1}+\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I})^{-1}\bold{Z(\bold{s})}\ \
\text{[see page 16 of \cite{Rasmussen:Williams}]}
\end{align*}
where $\bold{Z(\bold{s})}^{\prime}$ is the row vector $(x_{1}-\beta_{g0}-\beta_{g1}x_{0}\ \ x_{2}-\beta_{g0}-\beta_{g1}x_{1}\
\cdots \ x_{t-1}-\beta_{g0}-\beta_{g1}x_{t-2})$. We want to show that
$Var(E(X(\bold{s},t)|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0}))$ is finite. Equivalently,
we want to show $Var(\beta_{g0}+\beta_{g1}x_{t-1}+\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I})^{-1}\bold{Z(\bold{s})})$
is finite. For that it is enough to show $Var(\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I})^{-1}\bold{Z(\bold{s})})$
is finite since our induction hypothesis already assume that $Var(X(\bold{S},t-1))$ is finite.
Now we show that $Var(\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I})^{-1}\bold{Z(\bold{s})})$ is finite.
First note that $\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I})^{-1}\bold{Z(\bold{s})}$ can be expressed as a
linear combination of the elements of $\bold{Z(\bold{s})}$ as $w_{1}z_{1}(\bold{s})+w_{2}z_{2}(\bold{s})+\cdots+w_{t-1}z_{t-1}(\bold{s})$.
If the $w_{i}(\bold{S})$ are fixed numbers it is easy to see that $w_{1}z_{1}(\bold{s})+w_{2}z_{2}(\bold{s})+\cdots+w_{t-1}z_{t-1}(\bold{s})$
has finite variance. Unfortunately, $w_{i}(\bold{S})$ are random. However we will show that they are bounded random variables and
then using a lemma we will prove that $Var(w_{1}z_{1}(\bold{s})+w_{2}z_{2}(\bold{s})+\cdots+w_{t-1}z_{t-1}(\bold{s}))$ is finite.
First we show that $w_{i}(\bold{S})$ are bounded random variables. Consider the spectral decomposition of the real symmetric (nnd)
matrix $\bold{\Sigma}_{g22}$. Let us assume that $\bold{\Sigma}_{g22}=\bold{U}\bold{D}\bold{U}^{\prime}$, where $\bold{U}$ is an
orthogonal matrix and $\bold{D}$ is the diagonal matrix whose diagonal elements are eigenvalues. Then
\begin{align*}
\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I})^{-1}\bold{Z(\bold{s})}
&= \bold{\Sigma}_{g12}^{'}(\bold{U}\bold{D}\bold{U}^{\prime}+\sigma_{\eta}^2\bold{I})^{-1}\bold{Z(\bold{s})}\\
&=\bold{\Sigma}_{g12}^{'}(\bold{U}\bold{D}\bold{U}^{\prime}+\sigma_{\eta}^2\bold{U}\bold{U}^{\prime})^{-1}\bold{Z(\bold{s})}\\
&=\bold{\Sigma}_{g12}^{'}(\bold{U}(\bold{D}+\sigma_{\eta}^2\bold{I})\bold{U}^{\prime})^{-1}\bold{Z(\bold{s})}\\
&=\bold{\Sigma}_{g12}^{'}{\bold{U}^{\prime}}^{-1}(\bold{D}+\sigma_{\eta}^2\bold{I})^{-1}\bold{U}^{-1}\bold{Z(\bold{s})}\\
&=\bold{\Sigma}_{g12}^{'}\bold{U}(\bold{D}+\sigma_{\eta}^2\bold{I})^{-1}{\bold{U}^{\prime}}\bold{Z(\bold{s})} \ \
\text{[Since $\bold{U}$ is an orthogonal matrix]}
\end{align*}
Since $\bold{U}$ is a (random) orthogonal matrix its elements are bounded random variables between $-1$ and $1$.
The (random) elements of the row vector $\bold{\Sigma}_{g12}^{'}$ are covariances induced by the isotropic covariance kernel
$c_{g}(\cdot,\cdot)$. Hence, they are bounded random variables between $-\sigma_{g}^{2}$ and $\sigma_{g}^{2}$. Finally, the
(random) elements of $(\bold{D}+\sigma_{\eta}^2\bold{I})^{-1}$ are bounded random variables between $0$ and $\frac{1}{\sigma_{\eta}^{2}}$.
Hence, the (random) row vector $\bold{\Sigma}_{g12}^{'}(\bold{\Sigma}_{g22}+\sigma_{\eta}^2\bold{I})^{-1}$, being a product of some
random matrices whose elements are bounded random variables, is itself composed of bounded random variables. So,
its elements $w_{i}(\bold{S})$, although random, are bounded. Now, we state a crucial lemma.
\begin{lemma}
Let us assume that $X_{1},X_{2},\cdots,X_{n}$ are random variables with finite variance and $W_{1},W_{2},\cdots,W_{n}$ are bounded
random variables all defined on same probability space. Then the random variables $Y=(W_{1}X_{1}+W_{2}X_{2}+\cdots+W_{n}X_{n})$ also
has finite variance.
\begin{proof}
Let us assume that $W_{1},W_{2},\cdots W_{n}$ lie between $[-M,M]$.\\ $E(X_{i}^2)\leq K$ for $i=1,2\cdots,n$.
Now $Var(W_{i}X_{i})= E(Var(W_{i}X_{i}|X_{i}))+Var(E(W_{i}X_{i}|X_{i}))=E(X_{i}^2Var(W_{i}|X_{i}))+Var(X_{i}E(W_{i}|X_{i}))$.
But $Var(W_{i}|X_{i})\leq E(W_{i}^2|X_{i})\leq M^2$. So, $E(X_{i}^2Var(W_{i}|X_{i}))\leq M^2E(X_{i}^2)$.
Similarly, $E(W_{i}|X_{i})$ lies between $[-M,M]$. So, $Var(X_{i}E(W_{i}|X_{i}))\leq E(X_{i}^2(E(W_{i}|X_{i}))^2)\leq M^2E(X_{i}^2)$.
Hence, $Var(W_{i}X_{i})\leq 2M^2E(X_{i}^2)$.
So,
\begin{align*}
|Var(Y)|&=|\sum_{i=1}^{n}Var(W_{i}X_{i})+2\sum_{1\leq i<j\leq n}Cov(W_{i}X_{i},W_{j}X_{j})|\\
&\leq \sum_{i=1}^{n}Var(W_{i}X_{i})+2\sum_{1\leq i<j\leq n}|Cov(W_{i}X_{i},W_{j}X_{j})|\\
&\leq \sum_{i=1}^{n}Var(W_{i}X_{i})+2\sum_{1\leq i<j\leq n}Var^{\frac{1}{2}}(W_{i}X_{i})Var^{\frac{1}{2}}(W_{j}X_{j})\\
&\leq 2nM^2K+n(n-1)2M^2K
\end{align*}
So, $Y$ has finite variance.
\end{proof}
\end{lemma}
Once we apply the lemma to $w_{1}z_{1}(\bold{s})+w_{2}z_{2}(\bold{s})+\cdots+w_{t-1}z_{t-1}(\bold{s})$ the
finiteness of $Var(E(X(\bold{s},t)|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0}))$ is immediate.
Then by the formula $Var(X(\bold{s},t))=E(Var(X(\bold{s},t)|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)
=x_{0}))+Var(E(X(\bold{s},t)|X(\bold{s},t-1)=x_{t-1},X(\bold{s},t-2)=x_{t-2}\cdots,X(\bold{s},0)=x_{0}))$ we get
that $Var(X(\bold{s},t))$ is finite.
\end{proof}
part $(b)$
Since we have already proved in part $(a)$ that the coordinate variables of the observed spatio-temporal process have finite variances,
now we can consider the covariance function associated with the process and study its properties.
Let us denote the covariance between $Y(\bold{s},t)$ and $Y(\bold{s}^*,t^*)$ by $c_{y}((\bold{s},t),(\bold{s}^*,t^*))$.
Then
\begin{align}
c_{y}((\bold{s},t),(\bold{s}^*,t^*))=&E[Cov(Y(\bold{s},t),Y(\bold{s}^*,t^*)\mid
x(\bold{s},t),x(\bold{s}^*,t^*))]\notag\\ &+Cov[E(Y(\bold{s},t)\mid x(\bold{s},t)),
E(Y(\bold{s}^*,t^*)\mid x(\bold{s}^*,t^*))]\notag\\
=E[c_{f}(X(\bold{s},t),X(\bold{s}^*,t^*))]+&c_{\epsilon}(\bold{s},\bold{s}^*)\delta(t-t^*)
+\beta_{1f}^{2}Cov[X(\bold{s},t),X(\bold{s}^*,t^*)]\notag\\
\end{align}
Now, the term $E[c_{f}(X(\bold{s},t),X(\bold{s}^*,t^*))]$ will
be nonstationary and hence $E[c_{f}(X(\bold{s}+\bold{h},t+k),X(\bold{s}^*+\bold{h},t^*+k))]
\neq E[c_{f}(X(\bold{s},t),X(\bold{s}^*,t^*))]$. In fact, $| X(\bold{s}+\bold{h},
t+k)-X(\bold{s}^*+\bold{h},t^*+k)|\neq| X(\bold{s},t)-X(\bold{s}^*,t^*)|$ with probability 1
because $X(\bold{s},t)$ has density with respect to Lebesgue measure and this heuristically justifies our argument. So,
the covariance function $c_{y}(\cdot,\cdot)$ is nonstationary in both space and time.
To prove non separability, first see that $c_{f}(x(\bold{s},t),x(\bold{s}^*,t^*))$ is non separable in space and time,
because both space and time are involved in it through $x(\bold{s},t)$. Hence, $E[c_{f}(X(\bold{s},t),X(\bold{s}^*,t^*))]$
is nonseparable and therefore $c_{y}(\cdot,\cdot)$ is nonseparable in space and time.
\begin{proof}[Proof of Theorem 3.5]
\label{sec:proof5}
First consider $Cov(X(\bold{s},t),X(\bold{s}^{*},t^{*}))$ where WLOG we assume $t>t^*$.
Also, assume that $g^{*}(\cdot)$ is the centered Gaussian process obtained from $g(\cdot)$. Then
\begin{align*}
&Cov(X(\bold{s},t),X(\bold{s}^*,t^*))=Cov(g(X(\bold{s},t-1))+\eta(\bold{s},t),X(\bold{s}^*,t^*))\\
&=Cov(\beta_{0g}+\beta_{1g}
X(\bold{s},t-1)+g^{*}(X(\bold{s},t-1))+\eta(\bold{s},t),X(\bold{s}^*,t^*))\\
&=\beta_{1g}Cov(X(\bold{s},t-1),X(\bold{s}^*,t^*))+Cov(g^{*}(X(\bold{s},t-1)),X(\bold{s}^*,t^*))\\
\end{align*}
Repeatedly expanding the term in the same way we get
\begin{align}
&=\beta_{1g}^{t-t^*}Cov(X(\bold{s},t^*),X(\bold{s}^*,t^*))+\beta_{1g}^{t-t^*-1}
Cov(g^{*}(X(\bold{s},t^*)),X(\bold{s}^*,t^*))+\label{cov}\\
&\cdots+Cov(g^{*}(X(\bold{s},t-1)),X(\bold{s}^*,t^*))\notag
\end{align}
Just as the previous paragraph we can further see that
\begin{align}
&Cov(X(\bold{s},t^*),X(\bold{s}^*,t^*))\notag\\
&=Cov(\beta_{0g}+\beta_{1g}
X(\bold{s},t^*-1)+g^{*}(X(\bold{s},t^*-1))+\eta(\bold{s},t^*),\beta_{0g}\notag\\
&+\beta_{1g}X(\bold{s}^*,t^*-1)+g^{*}(X(\bold{s}^*,t^*-1))+\eta(\bold{s}^*,t^*))\notag\\
&=\beta_{1g}^2Cov(X(\bold{s},t^*-1),X(\bold{s}^*,t^*-1))+\beta_{1g}Cov(X(\bold{s},t^*-1),g^{*}(X(\bold{s}^*,t^*-1)))+\notag\\
&\beta_{1g}Cov(X(\bold{s}^*,t^*-1),g^{*}(X(\bold{s},t^*-1)))+Cov(g^{*}(X(\bold{s},t^*-1)),g^{*}(X(\bold{s}^*,t^*-1)))+c_{\eta}(\bold{s},\bold{s}^*)
\end{align}
Now we plan to show that terms of the types $Cov(g^{*}(X(\bold{s}^*,t^*-1)),X(\bold{s},t^*-1))$ and
$Cov(g^{*}(X(\bold{s},t^*-1)),g^{*}(X(\bold{s}^*,t^*-1)))$ are negligible if $\sigma_{g}^2$ is small enough.
Our next lemma proves it rigorously.
\begin{lemma}
\label{lemma:small_cov}
For arbitrarily small $\epsilon>0$, $\exists\ \delta>0$ such that $Cov(g^{*}(X(\bold{s},t-1)),X(\bold{s}^*,t^*))<\epsilon$
for $0<\sigma_{g}^2<\delta$.
\end{lemma}
See that it is enough to prove that $Var(g^{*}(X(\bold{s},t-1)))$ is arbitrarily small $\forall \bold{s},t$. Then Cauchy-Schwartz inequality implies $Cov^{2}(g^{*}(X(\bold{s},t-1)),g^{*}(X(\bold{s}^*,t^*)))
\leq Var(g^{*}(X(\bold{s},t-1)))Var(g^{*}(X(\bold{s}^*,t^*)))$ is arbitrarily small. Similarly, Cauchy-Schwartz inequality implies $Cov^{2}(g^{*}(X(\bold{s},t-1)),\eta(\bold{s}^*,t^*))
\leq Var(g^{*}(X(\bold{s},t-1)))Var(\eta(\bold{s}^*,t^*))=Var(g^{*}(X(\bold{s},t-1)))\sigma_{\eta}^2$ is arbitrarily small. Then we are done by the expansion
\begin{align*}
Cov(g^{*}(x(\bold{s},t)),x(\bold{s^*},t^*))&=Cov(g^{*}(x(\bold{s},t-1)),g^{*}(x(\bold{s}^*,t^*-1))+\cdots+\beta_{1g}^{t^*}Cov(g^{*}(x(\bold{s},t-1)),g^{*}(x(\bold{s}^*,0)))\\
&+Cov(g^{*}(x(\bold{s},t-1)),\eta(\bold{s}^*,t^*-1))+\cdots+\beta_{1g}^{t^*}Cov(g^{*}(x(\bold{s},t-1)),\eta(\bold{s}^*,0))
\end{align*}
Before proceeding towards the proof we mention two results from Gaussian process (see \cite{Adler07} for details) that will be used subsequently.
\begin{result}[Borell-TIS inequality]
\label{result:Borell-TIS}
Let us assume that $g$ is an almost surely bounded centered Gaussian process on index set
$T\subseteq\mathbb{R}$. Define $\sigma_{T}^{2}=\sup\limits_{t\in T}E(g_{t}^2)$.\\
Then $P(\|g\|>s)\leq \exp(-\frac{(s-E\|g\|)^2}{2\sigma_{T}^2})$ for $s>E(\|g\|)$ where $\|g\|=\sup\limits_{t}g_{t}$.
\end{result}
\begin{result}[Dudley's metric entropy bound]
\label{result:Dudley}
Under the assumption of the Borel-TIS inequality,\\
$E\|g\|\leq K\mathlarger{\int_{0}^{\mbox{diam}(T)}}\sqrt{H(\epsilon)}d\epsilon$,\\
where $\mbox{diam}(T)=\underset{\bold{s}_1,\bold{s}_2\in T}\sup d(\bold{s}_1,\bold{s}_2)$ is the diameter of the index set
$T$ with respect to the canonical pseudo-metric $d$ associated with the Gaussian process $g$
given by $d(\bold{s}_1,\bold{s}_2)=\sqrt{E(g(\bold{s}_1)-g(\bold{s}_2))^2}$,
and
$H(\epsilon)=\ln{N(\epsilon)}$ where $N(\epsilon)$ is the minimum number of $\epsilon$ balls required to cover the index set
$T$ with respect to the canonical pseudo-metric $d$; $K$ is a universal constant.
\end{result}
With the above two results, we are ready to prove Lemma \ref{lemma:small_cov}.
\begin{proof}[Proof of Lemma \ref{lemma:small_cov}]
Consider $Var((g^{*}(X(\bold{s},t-1))))$. Observe that
\begin{align*}
Var((g^{*}(X(\bold{s},t-1))))&\leq E((g^{*}(X(\bold{s},t-1))))^{2}\\
& \leq E(\sup\limits_{x}|g^{*}(x)|^{2})\ \ \ \\
& =\mathlarger{\int_{0}^{\infty}}P(\sup\limits_{x}|g^{*}(x)|^{2}>u)du\ \ \ (\text{by the tail sum formula})\\
&\leq 2\mathlarger{\int_{0}^{\infty}}P(\sup\limits_{x} g^{*}(x)>\sqrt{u})du\\
&= 2\mathlarger{\int_{0}^{L^{2}}}P(\sup\limits_{x} g^{*}(x)>\sqrt{u})du+2\mathlarger{\int_{L^{2}}^{\infty}}
P(\sup\limits_{x} g^{*}(x)>\sqrt{u})du\\
&(\text{where $L=\max{(E(\sup\limits_{x}g^{*}(x)),0)}$}\\
&\leq 2L^{2}+2\mathlarger{\int_{L^{2}}^{\infty}}e^{-\frac{(\sqrt{u}-L)^2}{2\sigma_{g}^{2}}}\ \ \ (\text{using Result \ref{result:Borell-TIS}})
\end{align*}
Now, using the change of variable $\sqrt{u}=z+L$ the integral $\mathlarger{\int_{L^{2}}^{\infty}}e^{-\frac{(\sqrt{u}-L)^2}{2\sigma_{g}^{2}}}$
can be reduced to the form
\begin{align*}
&\mathlarger{\int_{0}^{\infty}}e^{-\frac{z^2}{2\sigma_{g}^{2}}}2zdz+2L\mathlarger{\int_{0}^{\infty}}e^{-\frac{z^2}{2\sigma_{g}^{2}}}dz\\
&=2\sigma_{g}^{2}+L\sigma_{g}(\sqrt{2\pi})
\end{align*}
Hence, $Var((g^{*}(X(\bold{s},t-1))))\leq 2L^{2}+4\sigma_{g}^{2}+2L\sigma_{g}(\sqrt{2\pi})$.
\\But, $0\leq L\leq K\mathlarger{\int_{0}^{diam(T)}}\sqrt{H(\epsilon)}d\epsilon$ by Result \ref{result:Dudley}
and it is not difficult to see that $H(\epsilon)$ is a decreasing function of $\sigma_{g}^2$. The same is true of $diam(T)$
when as a function of $\sigma_{g}^2$. These two facts together permit applicability of the monotone convergence theorem to yield
\begin{align*}
0\leq \lim_{\sigma_{g}^{2}\rightarrow 0^+}L\leq \lim_{\sigma_{g}^{2}\rightarrow 0^+}K\mathlarger{\int_{0}^{diam(T)}}\sqrt{H(\epsilon)}d\epsilon
& \leq \lim_{\sigma_{g}^{2}\rightarrow 0^+}K\mathlarger{\int_{0}^{\infty}}\sqrt{H(\epsilon)}\mathbb{I}(\epsilon\leq diam(T))d\epsilon\\
&\leq K\mathlarger{\int_{0}^{\infty}}\lim_{\sigma_{g}^{2}\rightarrow 0^+}\sqrt{H(\epsilon)}\mathbb{I}(\epsilon\leq diam(T))d\epsilon=0.
\end{align*}
So, $\lim_{\sigma_{g}^{2}\rightarrow 0^+}L=0$ which in turn implies $Var((g^{*}(X(\bold{s},t-1))))$ can be made
arbitrarily small by making $\sigma_{g}^{2}$ small. This proves Lemma \ref{lemma:small_cov}.
\end{proof}
Arguing similarly one can also show that for arbitrarily small $\epsilon>0$, $\exists\ \delta>0$ such that
$Cov(g^{*}(X(\bold{s},t^*-1)),g^{*}(X(\bold{s}^*,t^*-1)))<\epsilon$ for $0<\sigma_{g}^2<\delta$. Moreover,
see that the bound is uniform in $\bold{s}$ and $t$. Since $|\beta_{1g}|<1$, using the bound repeatedly in (15), we obtain
\begin{align*}
&|Cov(X(\bold{s},t),X(\bold{s}^*,t^*))-\beta_{1g}^{t-t^*}Cov(X(\bold{s},t^*),X(\bold{s}^*,t^*))|\\
&\leq \frac{\epsilon}{1-|\beta_{1g}|}
\end{align*}
Similarly, using the bound repeatedly in (16), we obtain
\begin{align*}
&|Cov(X(\bold{s},t^*),X(\bold{s}^*,t^*))-Cov(X(\bold{s},0),X(\bold{s}^*,0))-[\frac{1-\beta_{1g}^{2(t^*+1)}}{1-\beta_{1g}^2}]c_{\eta}(\bold{s},\bold{s}^*)|\\
&\leq \left[\frac{\epsilon}{1-|\beta_{1g}|}+\frac{\epsilon}{1-|\beta_{1g}|}+\frac{\epsilon}{1-|\beta_{1g}|}\right].
\end{align*}
Combining them we get
\begin{align*}
&|Cov(X(\bold{s},t),X(\bold{s}^*,t^*))-\beta_{1g}^{t-t^*}Cov(X(\bold{s},0),X(\bold{s}^*,0))-\beta_{1g}^{t-t^*}[\frac{1-\beta_{1g}^{2(t^*+1)}}{1-\beta_{1g}^2}]c_{\eta}(\bold{s},\bold{s}^*)|\\
&\leq \frac{\epsilon}{1-|\beta_{1g}|}\left[1+3|\beta_{1g}|^{t-t^*}\right] \leq \frac{4\epsilon}{1-|\beta_{1g}|}
\end{align*}
Now plugging in this approximation
in the expression for $c_{y}((\bold{s},t),(\bold{s}^*,t^*))$ we get
the desired result.
So, Theorem 3.5
is finally proved.
\end{proof}
\begin{proof}[Proof of Theorem 3.6]
Part $(a)$: From the condition of the theorem it is clear that $\exists$ a probability space $(\Omega,\mathcal{F},P)$
and a set $A\in\mathcal{F}$ such that $P(A)=1$ and for $\omega\in A$, $X(\bold{s},0)(\omega),\eta(\bold{s},t)(\omega),
\epsilon(\bold{s},t)(\omega)$ are continuous functions in $\bold{s}$ where $t=1,2,3\cdots$ and $g(x)(\omega),f(x)(\omega)$
are continuous functions in $x$. Then by the property of composition of two functions $X(\bold{s},1)(\omega)
=g(X(\bold{s},0)(\omega))(\omega)+\eta(\bold{s},1)(\omega)$ is a continuous function in $\bold{s}$. Proceeding recursively,
one can prove that $X(\bold{s},t)(\omega)$ is a continuous function in $\bold{s}$ for any $t$. Once we show $X(\bold{s},t)(\omega)$
is a continuous function, we prove $Y(\bold{s},t)(\omega)=f(X(\bold{s},t)(\omega))(\omega)+\epsilon(\bold{s},t)(\omega)$
is a continuous function in $\bold{s}$. So, part $(a)$ is proved.
Part $(b)$: Proof of part $(b)$ follows the similar lines of the proof of part $(a)$. Firstly, we state a simple lemma.
\begin{lemma}
Let us consider two real valued functions $u(z)$ and $v(x,y)$ such that both of them are $k$ times differentiable.
Then the composition function $u(v(x,y))$ is also $k$ times differentiable.
\end{lemma}
\begin{proof}
Proof of this lemma is basically a generalization of chain rule for multivariate functions and can be found in advanced
multivariate calculus text books. We give a brief sketch of the proof. First we clarify the term $k$ times differentiable
for the function $v(x,y)$. It means all mixed partial derivatives of $v(x,y)$ of order $k$ exist. We prove the lemma using
mathematical induction. Firstly, We show that the lemma is true for $k=1$ and then we show that if the lemma is true for $k-1$
then it must be true for $k$ as well.
That the lemma is true for $k=1$ easily follows from the chain rule for multivariate functions. Now we prove the second step.
By the induction hypothesis the lemma is true for the $k-1$ case and $u(z)$ and $v(x,y)$ are $k$ times differentiable.
We want to show that $u(v(x,y))$ is also $k$ times differentiable. Without loss of generality, we consider the mixed partial
derivative $\frac{\partial}{\partial x^{k_{1}}\partial y^{k_{2}}}\left(u(v(x,y))\right)$ where $k_{1}+k_{2}=k$ and show that it exists.
Observe that the partial derivative is equivalent to $\frac{\partial}{\partial x^{k_{1}}\partial y^{k_{2}-1}}
\left(u^{\prime}(v(x,y))(\frac{\partial}{\partial y}v(x,y))\right)$ provided it exists. Since, by the induction hypothesis the
lemma is true for the $k-1$ case and $u^{\prime}(z)$ and $v(x,y)$ are $k-1$ times differentiable, the composition of them
$u^{\prime}(v(x,y))$ is also $k-1$ times differentiable. On the other hand, $\frac{\partial}{\partial y}v(x,y)$ is
also $k-1$ times differentiable. So, the product of them $u^{\prime}(v(x,y))(\frac{\partial}{\partial y}v(x,y))$ is also $k-1$
times differentiable. Hence the partial derivative $\frac{\partial}{\partial x^{k_{1}}\partial y^{k_{2}-1}}\left(u^{\prime}
(v(x,y))(\frac{\partial}{\partial y}v(x,y))\right)$ exists. Equivalently, $\frac{\partial}{\partial x^{k_{1}}\partial y^{k_{2}}}
\left(u(v(x,y))\right)$ exists. Similarly one can prove the existence of other mixed partial derivatives of $u(v(x,y))$ of order $k$.
Hence, by induction the proof follows.
\end{proof}
Part $(b)$: From the condition of the theorem it is clear that $\exists$ a probability space $(\Omega,\mathcal{F},P)$ and a set
$A\in\mathcal{F}$ such that $P(A)=1$ and for $\omega\in A$, $X(\bold{s},0)(\omega),\eta(\bold{s},t)(\omega),\epsilon(\bold{s},t)(\omega)$
are $k$ times differentiable functions in $\bold{s}$ where $t=1,2,3\cdots$ and $g(x)(\omega),f(x)(\omega)$ are $k$ times differentiable
functions in $x$. Then by the above lemma $X(\bold{s},1)(\omega)=g(X(\bold{s},0)(\omega))(\omega)+\eta(\bold{s},1)(\omega)$ is
a $k$ times differentiable function in $\bold{s}$. The rest of the proof is exactly similar as in part $(a)$.
\end{proof}
|
\section{Introduction}
$5d$ transition metal oxides have recently attracted considerable interest as
they display unusual properties primarily resulting from the effect of large
spin-orbit coupling
\cite{Kim2008,Moon2008,Kim2009,Pesin2010,Wang2011,KimJW2012,Watanabe2013,
Witczak2014}. Of particular interest is the electronic nature of
Sr$_2$IrO$_4$\cite{Crawford1994} and Sr$_3$Ir$_2$O$_7$\cite{Cao2002}: despite
the large $5d$ bandwidth and weak correlation, both of which favour a metallic
character, these systems are insulators. The opening of an electronic gap has
been explained by means of a Hubbard-like model, in which the effect of
correlation is enhanced by the strong spin-orbit coupling which narrows the
effective $5d$ bandwidth isolating the so-called $j_\mathrm{eff}=1/2$
state\cite{Kim2008,Kim2009}. The $j_\mathrm{eff}=1/2$ state results from a
particular hierarchy of energies at play, most especially the crystal field and
the spin-orbit coupling.
Sr$_2$IrO$_4$ (Sr$_3$Ir$_2$O$_7$) is the $n=1$ ($n=2$) member of the
Ruddlesden-Popper series, Sr$_{n+1}$Ir$_{n}$O$_{3n+1}$, and is built by the
stacking of IrO$_2$ (bi-)layers, in which IrO$_6$ octahedra share the corner
oxygens. The dominant perturbation to the half-filled $5d$ iridium states
in these compounds comes from the cubic component of the crystal field, written
conventionally as $10Dq$. Indeed the $t_{2g}$--$e_g$ splitting, of order
several eV, is often considered to be large enough that the $e_g$ states can be
neglected, allowing the basic electronic structure to be understood in terms of
a single hole occupying the $t_{2g}$ orbitals (tetravalent iridium is $5d^5$).
In order to describe properly the ground state wave function of this hole,
spin-orbit coupling and residual crystal-field effects with symmetry lower than
cubic, such as tetragonal in the post-perovskite
CaIrO$_3$\cite{Hirai2009,Ohgushi2013} or trigonal in pyrochlore R$_2$Ir$_2$O$_7$
(R = rare earth element)\cite{Hozoi2012}, need to be considered. At the
single-ion level, this is achieved by diagonalizing the Hamiltonian $\mathcal{H}
= \zeta \mathbf{L}\cdot\mathbf{S}-\Delta L_z^2$ in the
$t_{2g}$ orbitals
basis-set\cite{Ament2011,Liu2012,Hozoi2012,Ohgushi2013,Moretti2014PRL}, where
$\zeta$ is the spin-orbit coupling and $\Delta$ is the tetragonal (trigonal)
crystal-field splitting.
Strictly speaking, the $j_\mathrm{eff}=1/2$ ground state is realized only for
$\Delta=0$, i.e. for a perfectly cubic symmetry. In real materials this
condition is relaxed to $|\Delta|\ll\zeta$. Estimates of $\Delta$ in
Sr$_2$IrO$_4$ ($\Delta=-0.01$ eV\cite{Boseggia2013JPCM}) and its sister
compound Ba$_2$IrO$_4$($\Delta=0.05$ eV\cite{Moretti2014PRB}) indeed confirm
that the requirement on the relative magnitude of $|\Delta|$ and $\zeta$ is
realized, since the spin-orbit coupling in these materials of order
$\sim$0.5 eV\cite{Kim2008,Boseggia2013PRL,Moretti2014PRB}. One has to keep in
mind, however, that the scenario of the $j_\mathrm{eff}=1/2$ ground state holds
true only when the $e_g$ states do not contribute to the ground state wave
function, i.e. if the cubic component of the crystal field $10Dq$ is much larger
than the spin-orbit coupling, $10Dq\gg\zeta$. Indeed, the contribution of the
$e_g$ states has been invoked as a possible cause of the departure of CaIrO$_3$
from the pure $j_\mathrm{eff}=1/2$ ground state in LDA+SO+U
calculations\cite{Subedi2012}.
Theoretical estimates of $10Dq$ in Sr$_2$IrO$_4$ range from 1.8\cite{Haskel2012}
to 5 eV\cite{Jin2009}. Experimentally, various x-ray techniques have been used
to estimate $10Dq$, including x-ray absorption spectroscopy (XAS),
resonant elastic (REXS) and inelastic (RIXS) x-ray scattering. For example, soft
XAS at the O K edge has been used to probe the empty iridium 5$d$ states through
hybridization with the oxygen 2$p$ orbitals\cite{Kim2008,Moretti2014PRB},
providing values of $10Dq$ for Sr$_2$IrO$_4$\cite{Moon2006} and
Sr$_3$Ir$_2$O$_7$\cite{Park2014} in the range 2.5 eV to 4 eV. However, this
particular technique is highly surface sensitive, especially when performed in
total-electron-yield (TEY) mode, which compromises the reliability of the
extracted value of $10Dq$. The possibility that surface and bulk properties
might be different in iridium oxides was highlighted by Liu \textit{et al.},
who reported the existence of weak metallicity in the
near-surface electronic structure of Sr$_3$Ir$_2$O$_7$ while its bulk is known to be
insulating\cite{Liu2014}. In addition to the surface sensitivity, one has to
deal with self-absorption effects in total-fluorescence-yield (TFY) detected
XAS. As self-absorption is dependent on photon energy and experimental geometry,
extreme caution has to be taken when corrections to the spectra are applied. XAS
at the Ir L$_{2,3}$ edges ensures bulk-sensitivity, but self-absorption equally
affects hard XAS in TFY mode. Moreover, it suffers from the sizeable broadening
of features due to the 2$p$ core-hole lifetime which obscures
details of the electronic structure close to the Fermi energy. This problem can
at least be overcome to a certain degree by measuring partial-fluorescence-yield (PFY)
detected XAS\cite{Hamalainen1991}: this technique provides very
similar information to that of conventional XAS, but with the advantage that a
shallower core-hole is left in the final state of the decay process selected by
energy-discriminating the photons emitted due to radiative decay. For
example, in the case of the $L\alpha_{1}$ ($L\alpha_{2}$) emission line of iridium,
if $\Gamma_{2p}$ is the lifetime broadening of the $2p_{3/2}$ core-hole, and
$\Gamma_{3d}$ is that of the $3d_{5/2}$ ($3d_{3/2}$) core-hole, then the PFY
broadening will be given by
$1/\sqrt{1/\Gamma_{2p}^{2}+1/\Gamma_{3d}^{2}}\approx\Gamma_{3d}$, since
$\Gamma_{3d}\ll\Gamma_{2p}$. However, even if the benefits of PFY XAS are
evident, it is still difficult to extract quantitative information on $10Dq$
from such measurements\cite{Gretarsson2011,Clancy2014}.
Resonant x-ray magnetic scattering
(RXMS)\cite{Kim2009,KimJW2012,Boseggia2012JPCM,Boseggia2013JPCM} and
resonant inelastic x-ray scattering
(RIXS)\cite{Ishii2011,KimJ2012,Liu2012,Moretti2014PRL2} in the hard x-ray regime
also provide rough estimates of the cubic component of the crystal field from
the RXMS/RIXS energy dependence. Indeed, that the intensity of both magnetic
reflections in RXMS and intra-$t_{2g}$ excitations in RIXS are enhanced a few eV
below the main absorption line has been interpreted as a signature of the
$t_{2g}$--$e_g$ splitting. Again, however, both of these techniques suffer from
self-absorption effects due to the proximity of the scattered photon energy to
the Ir L$_{2,3}$ absorption edges.
The present work was designed to provide a reliable, bulk-sensitive probe of the
electronic structure of iridium oxides. We therefore used non-resonant inelastic
x-ray scattering (NIXS) in the hard x-ray energy range, more specifically x-ray
Raman spectroscopy (XRS), to probe the bulk properties of iridium oxides. XRS is
a x-ray scattering technique in which the energy of the incoming and scattered
photons is far from absorption edges of the material, making XRS a
self-absorption-free and bulk-sensitive probe\cite{Schulke2007}. Indeed, the XRS
cross-section in the limit of small momentum transfer $|\mathbf{q}|$ (i.e. in
the dipole limit) is formally identical to that of XAS, with $\mathbf{q}$
playing the role of photon polarization: the XRS cross-sections is then
proportional to $\left| \langle f | \mathbf{q} \cdot \mathbf{r} | i \rangle
\right|^2$, where $|i\rangle$ and $|f\rangle$ are the many-body electronic wave
functions of the initial and final state of the system,
respectively\cite{Schulke2007}. The main drawback of this technique is the low
count-rate, which is partially overcome by collecting the scattered photons over
a large solid angle. In the following we show that XRS allows the precise
determination of the cubic component of the crystal-field splitting in the
compounds Sr$_2$IrO$_4$ and Sr$_3$Ir$_2$O$_7$, thus offering an alternative
spectroscopic tool for the investigation of the electronic structure of iridium
oxides.
\section{Experimental details}
X-ray Raman spectroscopy measurements were performed at the ID20 beam line of
the European Synchrotron Radiation Facility (ESRF), Grenoble. The X-rays
produced by four U26 undulators were monochromatized to an energy-resolution
of $\Delta E_\mathrm{i} \simeq 0.3$ eV by the simultaneous use of a
Si(111) high heat-load liquid-nitrogen cooled monochromator and a Si(311)
post-monochromator. The x-rays were then focused at the sample position by
means of a Kirkpatrick-Baez mirror system down to a spot size of $10 \times 20$
$\mu$m$^2$ (vertical $\times$ horizontal, FWHM). The scattered X-rays were
collected by 12 crystal-analyzers exploiting the Si(660) reflection close to
backscattering geometry (at a fixed Bragg angle of $88.5^\circ$, corresponding
to $E_\mathrm{o}=9670$ eV) and detected by a Maxipix detector\cite{Ponchut2011}
with pixel size of $55\times 55$ $\mu$m$^2$. The resulting energy resolution was
$\Delta E \simeq 0.7$ eV. In order to obtain the XRS spectrum, the incident
photon energy $E_\mathrm{i}$ was varied in the energy range from
$E_\mathrm{i}-E_\mathrm{o}=0$ (the elastic energy) to
$E_\mathrm{i}-E_\mathrm{o}=570$ eV, thus covering the oxygen K edge. The
accumulation time/spectrum was about 2 hours and several spectra were recorded
to improve the counting statistics. XRS spectra were collected in two different
scattering geometries, corresponding to the momentum transfer $\mathbf{q}$ along
the sample $c$-axis and in the $ab$-plane, respectively. In both geometries, the
scattering plane was vertical and the incident X-rays linearly polarized in the
horizontal plane. XAS spectra were recorded at the ID08 beam line of the ESRF in
the TFY mode.
Single crystals of Sr$_2$IrO$_4$ and Sr$_3$Ir$_2$O$_7$, with
dimensions of $\sim 0.5 \times 0.5 \times 0.2$ mm$^3$, were
were grown using
the flux method described in Ref. \onlinecite{Boseggia2012PRB}.
All spectra were recorded at room temperature.
\section{Results and discussion}
\begin{figure}[t]
\centering
\includegraphics[width=.99\columnwidth]{fig1.eps}
\caption{\label{fig:fig1}XRS spectra of (a) Sr$_2$IrO$_4$ and (b)
Sr$_3$Ir$_2$O$_7$ for transferred momenta $\mathbf{q}\parallel(001)$ (black) and
$\mathbf{q}\parallel(100)$ (red dots) with $|\mathbf{q}|\simeq 6$ \AA$^{-1}$
(scattering angle $2\theta = 60^\circ$). Gray triangles in (b) represent the XRS
spectrum of Sr$_3$Ir$_2$O$_7$ with $|\mathbf{q}|\simeq 10$ \AA$^{-1}$
(scattering angle $2\theta = 120^\circ$). XAS spectra at the O K edge of the two
compounds for incoming polarization $\boldsymbol{\epsilon}\parallel(001)$
(black) and $\boldsymbol{\epsilon}\parallel(100)$ (red line) are also shown in
the insets.}
\end{figure}
Figure~\ref{fig:fig1} shows XRS scans for Sr$_2$IrO$_4$ (a) and
Sr$_3$Ir$_2$O$_7$ (b) across the oxygen K edge for $\mathbf{q} \parallel (001)$
(black) and $\mathbf{q}\parallel (100)$ (red dots). The scattering angle was
fixed to $2\theta=60^\circ$, corresponding to a momentum transfer of $|\mathbf{q}|\simeq
6$ \AA$^{-1}$. The background was removed by subtracting a linear fit to the
pre-edge region at energies lower than 528 eV. The spectra were then normalized
to unit area. For both samples, spectra taken in the two geometries are
distinctly different, revealing a very strong orientation dependence of the XRS
signal. In particular, one notes a large change of spectral weight between the
two main features in the 530-535 eV energy range.
In agreement with XAS results\cite{Chen1991,Schmidt1996,Moon2006}, the 528-535
eV energy region is dominated by transitions to the Ir 5$d$ states through the
hybridization with O 2$p$ orbitals, while higher energy features correspond to
excitations involving Ir 6$s$, 6$p$ and Sr 4$d$ states\cite{Mizokawa2001}, as
indicated in Fig.~\ref{fig:fig1}. For comparison, TFY XAS spectra were measured
on the very same samples. These are shown in the insets of Fig.~\ref{fig:fig1}.
Continuous black and red lines correspond to orthogonal directions of the
photon polarization, $\boldsymbol{\epsilon}\parallel(001)$ and
$\boldsymbol{\epsilon}\parallel(100)$, respectively. As expected, the overall
shape is similar to that of the XRS spectra, but the dichroic effect in the XAS
spectra is very small, in stark contrast to the strong orientation dependence of
the XRS measurements performed on the same samples. In order to rule out any
contribution higher than dipolar to the XRS spectra, we investigated the
$|\mathbf{q}|$ dependence of the XRS cross-section in Sr$_3$Ir$_2$O$_7$: by
setting $2\theta=120^\circ$, corresponding to $|\mathbf{q}|\simeq 10$ \AA$^{-1}$
(gray triangles in Fig.~\ref{fig:fig1}(b)), we note that the overall shape of
the spectrum perfectly matches with that acquired for $|\mathbf{q}|\simeq 6$
\AA$^{-1}$, thus implying that the momentum dependence of the XRS is negligible.
We therefore attribute the discrepancy between XRS and XAS measurements to
potential surface and/or self-absorption effects affecting soft x-ray
techniques. This observation underlines the importance of complementing
surface-sensitive techniques with bulk-sensitive probes.
In order to analyse our data we have calculated the number of
peaks expected in the 530-535 eV energy interval and their corresponding spectral weights by
pursuing the analogy between the XRS and XAS cross-sections.
The relevant transitions are those from O $1s$ to $2p$ states with the
latter hybridised with the Ir 5$d$ orbitals\cite{Kim2008,Park2014,Moretti2014PRB}.
The hybridization strength is calculated according to
the orbital overlap model \cite{Slater1954} with the hopping integral $t_{pd\mu}$
written as
\begin{equation}
t_{pd\mu} = V_{pd\mu} r^{-\alpha},
\end{equation}
where $V_{pd\mu}$ is a constant depending on the bond type ($\mu$=$\pi$ or $\sigma$), $r$ is the Ir-O
distance ($r_\mathrm{A} = 2.06$ \AA\ and $r_\mathrm{P} = 1.98$ \AA\ for
apical and in-plane oxygens, respectively, in Sr$_2$IrO$_4$\cite{Crawford1994}; while $r_\mathrm{A} =
2.02$ \AA~and $r_\mathrm{P} = 1.99$ \AA~in
Sr$_3$Ir$_2$O$_7$\cite{Subramanian1994}) and $\alpha = 3.5$\cite{Harrison1989}.
It should be noted that $V_{pd\sigma}$ and $V_{pd\pi}$ are related by
$V_{pd\pi}=-V_{pd\sigma}/\sqrt{3}$\cite{Harrison1989}.
Since the hybridization strength is inversely proportional to the distance
between the atoms involved, we can distinguish the contributions of the
apical (A) and in-plane (P) oxygens. Let us consider the apical oxygens first:
the O $2p_\mathrm{z}$ state hybridizes with the Ir $5d$ $\mathrm{3z^2-r^2}$
states, while the $2p_\mathrm{x}$ ($2p_\mathrm{y}$) mixes with the $\mathrm{zx}$
($\mathrm{yz}$) orbitals. For the in-plane oxygens,
$2p_\mathrm{z}$ hybridizes with the $\mathrm{yz}$ and $\mathrm{zx}$ orbitals,
while $2p_\mathrm{x}$ and $2p_\mathrm{y}$ are mixed with the $\mathrm{xy}$,
$\mathrm{3z^2-r^2}$ and $\mathrm{x^2-y^2}$ orbitals. This
is summarized in Fig.~\ref{fig:fig2}.
\begin{figure}[t]
\centering
\includegraphics[width=.99\columnwidth]{fig2.eps}
\caption{\label{fig:fig2}Sketch of the symmetry of the $t_{2g}$ (top)
and $e_g$ (bottom) orbitals involved in the O $2p$-Ir $5d$ hybridization. The
Eulerian angles $\theta$ and $\varphi$ describing the direction of
$\boldsymbol{\epsilon}$ or $\mathbf{q}$ in the sample reference system are also
shown.}
\end{figure}
It remains to consider the cross-sections associated with transitions to
different orbitals.
In the framework of a single-ion model, these are obtained by
calculating the matrix elements corresponding to the dipolar
$1s\rightarrow 2p_i$ transitions ($i=\mathrm{x, y, z}$)\cite{Moretti2014PRB}.
The cross-section is proportional to
the product of $\left| t_{pd\mu} \right|^2$, $n$ the number of available final $5d$ states and
a polarisation factor.
The polarization (transferred momentum) dependence of the XAS (XRS, in the
dipole limit) cross-sections to the $2p_\mathrm{x}$, $2p_\mathrm{y}$ and
$2p_\mathrm{z}$ states are given by $\sin^2\theta\cos^2\varphi$,
$\sin^2\theta\sin^2\varphi$ and $\cos^2\theta$, respectively, where $\theta$ and
$\varphi$ are the Eulerian angles describing the direction of
$\boldsymbol{\epsilon}$ ($\mathbf{q}$) in the sample reference system, as
sketched in Fig.~\ref{fig:fig2}. Merging the cross-section angular dependence
and the hybridization between Ir $5d$-O $2p$ states, we obtain the polarization
(transferred momentum) dependence of the transitions to the $\mathrm{xy}$,
$\mathrm{yz}$, $\mathrm{zx}$, $\mathrm{3z^2-r^2}$, $\mathrm{x^2-y^2}$ orbitals
as reported in Table \ref{tab:table1}. Note that we have used
$n_\mathrm{xy}=n_\mathrm{yz}=n_\mathrm{zx} = 1/3$ and
$n_\mathrm{3z^2-r^2}=n_\mathrm{x^2-y^2} = 2$ expected for the $j_\mathrm{eff}=1/2$
state.
\begin{table}[t]
\caption{\label{tab:table1}Polarization dependence of the O $1s
\rightarrow$ O $2p$-Ir $5d$ dipolar transitions.}
\begin{ruledtabular}
\begin{tabular}{c c c}
& Apical O & In-plane O \\
\hline
$\mathrm{xy}$ & 0 & $2V_{pd\pi}^2
n_\mathrm{xy}r_\mathrm{P}^{-2\alpha}\sin^2\theta$
\\
$\mathrm{yz}$ &
$2V_{pd\pi}^2 n_\mathrm{yz}r_\mathrm{A}^{-2\alpha}\sin^2\theta\sin^2\varphi$ &
$2V_{pd\pi}^2 n_\mathrm{yz}r_\mathrm{P}^{-2\alpha}\cos^2\theta$ \\
$\mathrm{zx}$ &
$2V_{pd\pi}^2 n_\mathrm{yz}r_\mathrm{A}^{-2\alpha}\sin^2\theta\cos^2\varphi$ &
$2V_{pd\pi}^2 n_\mathrm{yz}r_\mathrm{P}^{-2\alpha}\cos^2\theta$ \\
$\mathrm{3z^2-r^2}$ &
$2V_{pd\sigma}^2 n_\mathrm{3z^2-r^2}r_\mathrm{A}^{-2\alpha}\cos^2\theta$ &
$V_{pd\sigma}^2 n_\mathrm{3z^2-r^2}r_\mathrm{P}^{-2\alpha}\sin^2\theta$ \\
$\mathrm{x^2-y^2}$ & 0 &
$\sqrt{3}V_{pd\sigma}^2 n_\mathrm{x^2-y^2}r_\mathrm{P}^{-2\alpha}\sin^2\theta$
\\
\end{tabular}
\end{ruledtabular}
\end{table}
For the specific geometries used in our experiments,
it transpires that only two transitions are allowed when
$\mathbf{q}\parallel (001)$ ($\theta = 0$) and four when $\mathbf{q}\parallel
(100)$ ($\theta = 90^{\circ}$ and $\varphi = 0$). The appropriate cross-sections
are given in Table \ref{tab:table2}. We therefore performed a fitting of our model
to the data
by adjusting the number of peaks accordingly and constraining their
relative spectral weight to the calculated one. Extra peaks were
introduced in the fit to mimic the high energy features: one for $\mathbf{q}\parallel
(100)$ and two for $\mathbf{q}\parallel (001)$, respectively. The result of the fitting is
shown in Fig.~\ref{fig:fig3} for Sr$_2$IrO$_4$ and in Fig.~\ref{fig:fig4} for
Sr$_3$Ir$_2$O$_7$. We emphasise
that, apart from an overall scale factor for the amplitude, the energy
position and full width at half maximum (FWHM) of the curves are the only free
fitting parameters: their values are summarized in Table~\ref{tab:table2}. The
agreement between the fit and the experimental data is remarkably good in both
scattering geometries, allowing us to unambiguously assign each feature.
In particular, the intense features at 531.4 (531.2) and 534.0 (533.7) eV in
Sr$_2$IrO$_4$ (Sr$_3$Ir$_2$O$_7$) correspond to excitations to the $\mathrm{3z^2-r^2}$ and
$\mathrm{x^2-y^2}$ orbitals via the apical and in-plane oxygens, respectively. This peak
assignment is consistent with the work of Moon \textit{et al.} on
Sr$_2$IrO$_4$\cite{Moon2006}, Schmidt \textit{et al.} on
Sr$_2$RuO$_4$\cite{Schmidt1996} and Park \textit{et al.} on
Sr$_3$Ir$_2$O$_7$\cite{Park2014}.
\begin{figure}[t]
\centering
\includegraphics[width=.99\columnwidth]{fig3.eps}
\caption{\label{fig:fig3}Experimental (open dots) and constrained fit to
the XRS spectra (solid thick line) of Sr$_2$IrO$_4$ for (a) $\mathbf{q}\parallel
(001)$ and (b) $\mathbf{q}\parallel (100)$. The fitting curves are plotted in
solid lines, while the extra-peaks are reported in dashed gray lines.}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=.99\columnwidth]{fig4.eps}
\caption{\label{fig:fig4}Experimental (open dots) and constrained fit to
the XRS spectra (solid thick line) of Sr$_3$Ir$_2$O$_7$ for (a)
$\mathbf{q}\parallel (001)$ and (b) $\mathbf{q}\parallel (100)$. The fitting
curves are plotted in solid lines, while the extra-peaks are reported in dashed
gray lines.}
\end{figure}
\begin{table*}
\caption{\label{tab:table2}Cross-sections, fitted energy positions and FWHM of
electronic dipolar transitions in Sr$_2$IrO$_4$ and Sr$_3$Ir$_2$O$_7$.}
\begin{ruledtabular}
\begin{tabular}{c c c c c c c}
& $\mathbf{q}\parallel (001)$ & $\mathbf{q}\parallel (100)$ & Energy loss (eV) &
FWHM (eV) & Energy loss (eV) & FWHM (eV) \\
& & & Sr$_2$IrO$_4$ & Sr$_2$IrO$_4$ & Sr$_3$Ir$_2$O$_7$ & Sr$_3$Ir$_2$O$_7$ \\
\hline
$\mathrm{xy/yz/zx_A}$ & & $2V_{pd\pi}^2 n_\mathrm{yz}r_\mathrm{A}^{-2\alpha}$ &
$528.9\pm 0.11$ & $0.71\pm 0.35$ & $528.9\pm 0.10$ & $0.78\pm 0.30$ \\
$\mathrm{xy/yz/zx_P}$ & $4V_{pd\pi}^2 n_\mathrm{yz}r_\mathrm{P}^{-2\alpha}$ &
$2V_{pd\pi}^2 n_\mathrm{xy}r_\mathrm{P}^{-2\alpha}$ & $529.8\pm 0.05$ & $1.0\pm
0.17$ & $529.6\pm 0.03$ & $0.87\pm 0.10$ \\
$\mathrm{3z^2-r^2_A}$ & $2V_{pd\sigma}^2
n_\mathrm{3z^2-r^2}r_\mathrm{A}^{-2\alpha}$
& & $531.4\pm 0.05$ & $2.4\pm 0.18$ & $531.2\pm 0.05$ & $2.4\pm 0.15$\\
$\mathrm{3z^2-r^2_P}$ & &
$V_{pd\sigma}^2 n_\mathrm{3z^2-r^2}r_\mathrm{P}^{-2\alpha}$ & $532.4\pm 0.75$ &
$3.0\pm 0.67$ & $531.8\pm 0.12$ & $2.5\pm 0.26$\\
$\mathrm{x^2-y^2_A}$ & & & & & & \\
$\mathrm{x^2-y^2_P}$ & &
$\sqrt{3}V_{pd\sigma}^2 n_\mathrm{x^2-y^2}r_\mathrm{P}^{-2\alpha}$ & $534.0\pm
0.35$ & $2.6\pm 0.46$ & $533.7\pm 0.05$ & $1.9\pm 0.06$\\
\end{tabular}
\end{ruledtabular}
\end{table*}
We are now in a position to extract the cubic component of the crystal field
$10Dq$. This is given by the energy difference between the centres of mass of
the $e_g$ and $t_{2g}$ states for in-plane oxygens. In view of the small
tetragonal crystal field measured in Sr$_2$IrO$_4$ ($|\Delta|=0.01$
eV\cite{Boseggia2013JPCM}), we consider the splitting of the $t_{2g}$ states due
to spin-orbit coupling only in the calculation of $10Dq$. We obtain $3.80\pm
0.82$ eV in Sr$_2$IrO$_4$ and $3.55\pm 0.13$ eV in Sr$_3$Ir$_2$O$_7$, assuming
$\zeta\simeq 0.4$ eV\cite{Kim2008}. Estimates of $10Dq$ extracted from XAS and
RXMS/RIXS measurements are consistent with our results. The cubic component of
the crystal field is thus very large compared to the other energy scales of the
system, namely the spin orbit coupling and the tetragonal crystal field,
therefore validating the initial hypothesis that $10Dq$ is the dominant energy
scale. Finally, in addition to the estimate of the cubic component of the
crystal field, we can deduce the sign of the tetragonal contribution to
the crystal field from the splitting of the $e_g$ states ($1.6\pm 0.82$ eV in
Sr$_2$IrO$_4$ and by $1.9\pm 0.13$ eV in Sr$_3$Ir$_2$O$_7$). Indeed, the fact
that the $\mathrm{x^2-y^2}$ orbital is the highest in energy is consistent with
structural studies indicating an elongation of the IrO$_6$ cage in both
compounds. Note that, for tetragonally distorted octahedra, the description of
$d$ states requires two parameters, $Ds$ and $D t$, in addition to the main
crystal-field parameter $10Dq$. The splitting of $e_g$ and $t_{2g}$ states is
then given by $4Ds+5Dt$ and $3Ds-5Dt$ ($=\Delta$),
respectively\cite{Bersuker2010}. In the absence of spin-orbit coupling,
the $t_{2g}$ states are almost degenerate ($\Delta \approx 0$), implying $3Ds
\approx 5Dt$. A finite splitting of the $e_g$ states is therefore compatible
with the realization of the $j_\mathrm{eff}=1/2$ ground state in Sr$_2$IrO$_4$
and Sr$_3$Ir$_2$O$_7$.
\section{Conclusions}
By exploiting the orientation dependence of oxygen K edge XRS cross-sections in
Sr$_2$IrO$_4$ and Sr$_3$Ir$_2$O$_7$, we have been able to assign spectral
features in the 528-535 eV energy range to specific transitions involving the Ir
$5d$ orbitals. These assignments allow us to extract the value of the cubic
crystal-field splitting $10Dq$ of $3.80\pm 0.82$ and $3.55\pm 0.13$ eV in
Sr$_2$IrO$_4$ and Sr$_3$Ir$_2$O$_7$, respectively. In addition, the tetragonal
crystal field was found to split the $e_g$ states by $1.6\pm 0.82$ eV in
Sr$_2$IrO$_4$ and by $1.9\pm 0.13$ eV in Sr$_3$Ir$_2$O$_7$. It is important to
stress that the reliability of these values of the crystal field splittings
obtained in our study is enhanced by the bulk sensitivity of the XRS technique.
\section{Acknowledgments}
The authors are grateful for technical support by C. Henriquet and R. Verbeni,
and all the colleagues from the ESRF support groups.
\input{XRS_final_DM.bbl}
\end{document}
|
\section*{0. INTRODUCTION}
Braided $T$-categories introduced by Turaev \cite{T2008} are
of interest due to their applications in homotopy quantum field theories,
which are generalizations of ordinary topological quantum field theories.
As such, they are interesting to different research communities
in mathematical physics (see \cite{FY1989, K2004, T1994, VA2001, VA2005}).
Although Yetter-Drinfeld modules over Hopf algebras provide examples of such braided
$T$-categories, these are rather trivial.
The wish to obtain more interesting homotopy quantum field theories provides a strong
motivation to find new examples of braided $T$-categories.
\\
The aim of this article is to construct new examples of braided $T$-categories.
This is achieved by generalizing an existing construction by Panaite and Staic \cite{PS2007}
that twists Yetter-Drinfeld modules over a Hopf algebra $H$ by Hopf algebra automorphisms.
We will generalize this construction to twisted Yetter-Drinfeld modules over so-called
monoidal Hom-Hopf algebras, which are Hopf algebras in the Hom-category of a monoidal category
(see \cite{CG2011}).
We find a suitable generalisation of the notion of a twisted Yetter-Drinfeld
module for this setting and obtain a category of
twisted Yetter-Drinfeld modules that is a braided $T$-category
in the sense of Turaev \cite{T2008}.
\\
The article is organized as follows.
\\
We will present the background material in Section 1.
This section contains the relevant definitions on monoidal Hom-Hopf algebras and braided
$T$-categories necessary for the understanding of the construction.
In Section 2, we define the notion of a Yetter-Drinfeld module over
a monoidal Hom-Hopf algebra that is twisted by two
monoidal Hom-Hopf algebra automorphisms
as well as the notion of a monoidal Hom-entwining structure and
show how such moinoidal Hom-entwining structures are obtained from automorphisms
of monoidal Hom-Hopf algebras.
\\
Section 3 first introduces the tensor product of twisted Yetter-Drinfeld Hom-modules
and then shows that the twisted Yetter-Drinfeld Hom-modules form a
braided $T$-category in the sense of Turaev [13]. At the end of the section,
we give an example of a monoidal Hom-Hopf algebra,
which can be viewed as a generalization of
Sweedler's Hopf algebra. And furthermore, we compute an example of a twisted
Yetter-Drinfeld module over a monoidal Hom-Hopf algebra.
\\
\section*{1. PRELIMINARIES}
\def\theequation{1. \arabic{equation}}
\setcounter{equation} {0} \hskip\parindent
Throughout, let $k$ be a fixed field. Everything is over $k$ unless
otherwise specified. We
refer the readers to the books of Sweedler \cite{S1969}
for the relevant concepts on the general theory of Hopf
algebras. Let $(C, \Delta )$ be a coalgebra. We use the "sigma" notation for
$\Delta $ as follows:
$$
\Delta (c)=\sum c_1\otimes c_2, \,\,\forall c\in C.
$$
\vskip 0.5cm
{\bf 1.1. Braided $T$-categories.}
\vskip 0.5cm
A {\sl monoidal category} ${\cal C}=({\cal C},\mathbb{I},\otimes,a,l,r)$
is a category ${\cal C}$ endowed with a functor
$\otimes: {\cal C}\times{\cal C}\rightarrow{\cal C}$
(the {\sl tensor product}), an object $\mathbb{I}\in {\cal C}$
(the {\sl tensor unit}), and natural isomorphisms $a$
(the {\sl associativity constraint}), where
$a_{U,V,W}:(U\otimes V)\otimes W\rightarrow U\otimes (V\otimes W)$
for all $U,V,W\in {\cal C}$, and $l$ (the {\sl left unit constraint})
where $l_U: \mathbb{I}\otimes U\rightarrow U,\,r$
(the {\sl right unit constraint}) where
$r_{U}:U\otimes{\cal C}\rightarrow U$ for all $U\in {\cal C}$,
such that for all $U,V,W,X\in {\cal C},$
the {\sl associativity pentagon}
$a_{U,V,W\otimes X}\circ a_{U\otimes V,W,X}
=(U\otimes a_{V,W,X})\circ a_{U,V\otimes W,X}\circ
(a_{U,V,W}\otimes X)$ and
$(U\otimes l_V)\circ(r_U\otimes V)=a_{U,I,V}$ are satisfied.
A monoidal categoey ${\cal C}$ is {\sl strict} when all
the constraints are identities.
\\
Let $G$ be a group and let $Aut({\cal C})$ be the group of
invertible strict tensor functors from ${\cal C}$ to itself.
A category ${\cal C}$ over $G$
is called a {\sl crossed category } if it satisfies the following:
\begin{eqnarray*}
&\blacklozenge & {\cal C} \mbox{ is a monoidal category;}\\
&\blacklozenge & {\cal C} \mbox{ is disjoint union of a family
of subcategories }\{{\cal C}_{\a }\}_{\a \in
G},\mbox{ and for any }U\in {\cal C}_{\a },\\
&&V\in {\cal C}_{\b }, U\o V\in {\cal C}_{\a \b }.
\mbox{ The subcategory }{\cal C}_{\a }
\mbox{ is called the }\a\mbox{th component of }{\cal C};\\
&\blacklozenge & \mbox{Consider a group homomorphism }
\vp : G\lr Aut({\cal C}), \b \mapsto \vp _{\b }, \mbox{ and
assume that}\\
&& \vp _{\b }(\vp _{\a })
=\vp _{\b\alpha\beta^{-1}},
\mbox{ for all }\alpha,\beta\in G.\mbox{ The functors }
\vp _{\b } \mbox{ are called conjugation}\\
&&\mbox{ isomorphisms.}
\end{eqnarray*}
Furthermore, $ {\cal C}$ is called strict when it is
strict as a monoidal category.
\\
{\sl Left index notation}: Given $\a \in G$
and an object $V\in {\cal C}_{\a }$, the functor $\vp _{\a }$
will be denoted by ${}^V( \cdot )$, as in Turaev \cite{T2008} or
Zunino \cite{Z2004}, or even ${}^{\a }( \cdot )$.
We use the notation ${}^{\overline{V}}( \cdot )$
for ${}^{\a ^{-1}}( \cdot )$. Then we have
${}^V id_U=id_{{} V^U}$ and
${}^V(g\circ f)={}^Vg\circ {}^Vf$.
Since the conjugation $\vp : G\lr Aut({\cal C})$ is a
group homomorphism, for all $V, W\in {\cal C}$, we have ${}^{V\o W}( \cdot )
={}^V({}^W( \cdot ))$ and ${}^\mathbb{I}( \cdot )={}^V({}^{\overline{V}}( \cdot ))
={}^{\overline{V}}({}^V( \cdot ))=id_{\cal C}$. Since, for all
$V\in {\cal C}$, the functor ${}^V( \cdot )$ is strict, we have
${}^V(f\o g)={}^Vf\o {}^Vg$, for any morphisms $f$ and $g$ in ${\cal C}$,
and ${}^V\mathbb{I}=\mathbb{I}$.
\\
A {\sl braiding} of a crossed category ${\cal C}$ is
a family of isomorphisms $({c=c_{U,V}})_{U,V}\in {\cal C}$,
where $c_{U,V}: U\otimes V\rightarrow {}^UV\otimes U$
satisfying the following conditions:\\
a) For any arrow $f\in {\cal C}_{\a }(U, U')$ and
$g\in {\cal C}(V, V')$,
$$
(({}^{\a }g)\o f)\circ c _{U, V}=c _{U' V'}\circ (f\o g).
$$
b)
For all $ U, V, W\in {\cal C},$ we have
$$
c _{U\o V, W}=a_{{}^{U\o V}W, U, V}\circ (c _{U, {}^VW}\o
id_V)\circ a^{-1}_{U, {}^VW, V}\circ (\i _U\o c _{V, W})
\circ a_{U, V, W},
$$
$$c _{U, V\o W}=a^{-1}_{{}^UV, {}^UW, U}
\circ (\i _{({}^UV)}\o c _{U, W})\circ a_{{}^UV, U, W}\ci
(c _{U, V}\o \i_W)\circ a^{-1}_{U, V, W},$$
where $a$ is the natural isomorphisms in the tensor category
${\cal C}$.\\
c) For all $ U, V\in {\cal C}$ and $\b\in G$,
$$ \vp _{\b }(c
_{U, V})=c _{\vp _{\b }(U), \vp _{\b }(V)}.
$$
A crossed category endowed with a braiding is called
a {\sl braided $T$-category}.
\\
{\bf 1.2. Monoidal Hom-Hopf algebras.}
\vskip 0.5cm
Let $\mathcal{M}_{k}=(\mathcal{M}_{k},\o,k,a,l,r )$
denote the usual monoidal category of $k$-vector spaces and linear maps between them.
Recall from \cite{CG2011}
that there is the {\it monoidal Hom-category} $\widetilde{\mathcal{H}}(\mathcal{M}_{k})=
(\mathcal{H}(\mathcal{M}_{k}),\,\o,\,(k,\,id),
\,\widetilde{a},\,\widetilde{l},\,\widetilde{r })$, a new monoidal category,
associated with $\mathcal {M}_{k}$ as follows:
$\bullet$ The objects of
$ \mathcal{H}(\mathcal{M}_{k})$ are couples
$(M,\mu)$, where $M \in \mathcal {M}_{k}$ and $\mu \in Aut_k(M)$, the set of
all $k$-linear automomorphisms of $M$;
$\bullet$ The morphism $f:(M,\mu)\rightarrow (N,\nu)$ in $ \mathcal{H}(\mathcal{M}_{k})$
is the $k$-linear map $f: M\rightarrow N$ in $\mathcal{M}_{k}$
satisfying $ \nu \circ f = f\ci \mu$, for any two objects
$(M,\mu),(N,\nu)\in \mathcal{H}(\mathcal{M}_{k})$;
$\bullet$ The tensor product is given by
$$
(M,\mu)\o (N,\nu)=(M\o N,\mu\o\nu )
$$
for any $(M,\mu),(N,\nu)\in \mathcal{H}(\mathcal{M}_{k})$.
$\bullet$ The tensor unit is given by $(k, id)$;
$\bullet$ The associativity constraint $\widetilde{a}$
is given by the formula
$$
\widetilde{a}_{M,N,L}=a_{M,N,L}\circ((\mu\o id)\o
\varsigma^{-1})=(\mu\o(id\o\varsigma^{-1}))\circ a_{M,N,L},
$$
for any objects $(M,\mu),(N,\nu),(L,\varsigma)\in \mathcal{H}(\mathcal{M}_{k})$;
$\bullet$ The left and right unit constraint
$\widetilde{l}$ and $\widetilde{r }$ are given by
$$
\widetilde{l}_M=\mu\circ l_M=l_M\circ(id\o\mu),\, \quad
\widetilde{r}_M =\mu\circ r_M=r_M\circ(\mu\o id)
$$
for all $(M,\mu) \in \mathcal{H}(\mathcal{M}_{k})$.
\\
We now recall from \cite{CG2011} the following notions used later.
\\
A {\it unital monoidal Hom-associative algebra} (a monoidal Hom-algebra in
Proposition 2.1 of \cite{CG2011}) is a vector space $A$
together with an element $1_A\in A$ and linear maps
$$m:A\o A\rightarrow A;\,\,a\o b\mapsto ab, \,\,\,\alpha\in Aut_k(A)$$
such that
\begin{equation}
\alpha(a)(bc)=(ab)\alpha(c),
\end{equation}
$$\alpha(ab)=\alpha(a)\alpha(b),$$
\begin{equation}
a1_A=1_Aa=\alpha(a),
\end{equation}
\begin{equation}
\alpha(1_A)=1_A ,
\end{equation}
for all $a,b,c\in A.$
\\
{\bf Remark 1.1.} (1) In the language of Hopf algebras, $m$ is called the
Hom-multiplication, $\alpha$ is the twisting
automorphism and $1_A$ is the unit. Note that Eq.(1.1) can be rewirtten as
$a(b\alpha^{-1}(c)) = (\alpha^{-1}(a)b)c$.
The monoidal Hom-algebra $A$ with $\alpha$ will be denoted by $(A,\alpha)$.
(2) Let $(A,\alpha)$ and $(A',\alpha')$ be two monoidal Hom-algebras.
A monoidal Hom-algebra map
$f:(A,\alpha)\rightarrow (A',\alpha')$ is a linear map such that
$f\circ \alpha=\alpha'\circ f,f(ab)=f(a)f(b)$ and
$f(1_A)=1_{A'}.$
(3) The definition of monoidal Hom-algebras is different
from the unital Hom-associative algebras in \cite{MS2010} and
\cite{MS2009} in the following sense.
The same twisted associativity condition (1.1) holds in both cases.
However, the unitality condition in their notion is the usual untwisted one:
$a1_A=1_Aa =a,$ for any $a\in A,$ and the twisting map $\alpha$
does not need to be monoidal (that is, (1.2) and (1.3) are not required).
\\
{\it A counital monoidal Hom-coassociative coalgebra} is
an object $(C,\gamma)$ in the category
$\tilde{\mathcal{H}}(\mathcal{M}_{k})$
together with linear maps
$\D:C\rightarrow C\o C,\,\D(c)=c_1\o c_2$ and
$\varepsilon:C\rightarrow k$ such that
\begin{equation}
\gamma^{-1}(c_1)\o\D(c_2)=\D(c_1)\o\gamma^{-1}(c_2),
\end{equation}
\begin{equation}
\D(\gamma(c))=\gamma(c_1)\o\gamma(c_2),
\end{equation}
$$c_1\varepsilon(c_2)=\gamma^{-1}(c)=\varepsilon(c_1)c_2,$$
\begin{equation}
\varepsilon(\gamma(c))=\varepsilon(c)
\end{equation}
for all $c\in C.$
\\
{\bf Remark 1.2.} (1) Note that (1.4)is equivalent to
$c_1\o c_{21}\o \gamma(c_{22})=\gamma(c_{11})\o c_{12}\o c_2.$
Analogue to monoidal Hom-algebras, monoidal Hom-coalgebras
will be short for counital monoidal Hom-coassociative coalgebras
without any confusion.
(2) Let $(C,\gamma)$ and $(C',\gamma')$ be two monoidal Hom-coalgebras.
A monoidal Hom-coalgebra map $f:(C,\gamma)\rightarrow(C',\gamma')$
is a linear map such that $f\circ \gamma=\gamma'\circ f, \D\circ f=(f\o f)\circ\D$
and $\varepsilon'\circ f=\varepsilon.$
\\
{\it A monoidal Hom-bialgebra} $H=(H,\alpha,m,1_H,\D,\varepsilon)$
is a bialgebra in the monoidal category
$ \tilde{\mathcal{H}}(\mathcal {M}_{k}).$
This means that $(H,\alpha,m,1_H)$ is a monoidal Hom-algebra and
$(H,\alpha,\D,\varepsilon)$ is
a monoidal Hom-coalgebra such that $\D$ and $\varepsilon$
are morphisms of algebras,
that is, for all $h,g\in H,$
$$\D(hg)=\D(h)\D(g),\, \,\, \D(1_H)=1_H\o1_H,\,\,\,\,\,\,
\varepsilon(hg)=\varepsilon(h)\varepsilon(g), \,\,\,\,\,\varepsilon(1_H)=1.$$
\\
A monoidal Hom-bialgebra $(H,\alpha)$ is called {\it a monoidal Hom-Hopf algebra}
if there exists a morphism (called antipode)
$S: H\rightarrow H$ in $ \tilde{\mathcal{H}}(\mathcal {M}_{k})$
(i.e., $S\ci \alpha=\alpha\ci S$),
which is the convolution inverse of the identity morphism $id_H$
(i.e., $ S*id=1_H\ci \varepsilon=id*S$). Explicitly, for all $h\in H$,
$$
S(h_1)h_2=\varepsilon(h)1_H=h_1S(h_2).
$$
\\
{\bf Remark 1.3.} (1) Note that a monoidal Hom-Hopf algebra is
by definition a Hopf algebra in $ \tilde{\mathcal{H}}(\mathcal {M}_{k})$.
(2) Furthermore, the antipode of monoidal Hom-Hopf algebras has
almost all the properties of antipode of Hopf algebras such as
$$S(hg)=S(g)S(h),\,\,\,\, S(1_H)=1_H,\,\,\,\,
\D(S(h))=S(h_2)\o S(h_1),\,\,\,\,\,\,\varepsilon\ci S=\varepsilon.$$
That is, $S$ is a monoidal Hom-anti-(co)algebra homomorphism.
Since $\alpha$ is bijective and commutes with $S$,
we can also have that the inverse $\alpha^{-1}$ commutes with $S$,
that is, $S\ci \alpha^{-1}= \alpha^{-1}\ci S.$
\\
In the following, we recall the notions of actions on monoidal
Hom-algebras and coactions on monoidal Hom-coalgebras.
\\
Let $(A,\alpha)$ be a monoidal Hom-algebra.
{\it A left $(A,\alpha)$-Hom-module} consists of
an object $(M,\mu)$ in $\tilde{\mathcal{H}}(\mathcal {M}_{k})$
together with a morphism
$\psi:A\o M\rightarrow M,\psi(a\o m)=a\cdot m$ such that
$$\alpha(a)\c(b\c m)=(ab)\c\mu(m),\,\,
\,\,\mu(a\c m)=\alpha(a)\c\mu(m),\,\,
\,\,1_A\c m=\mu(m),$$
for all $a,b\in A$ and $m \in M.$
\\
Monoidal Hom-algebra $(A,\alpha)$ can be
considered as a Hom-module on itself by the Hom-multiplication.
Let $(M,\mu)$ and $(N,\nu)$ be two left $(A,\alpha)$-Hom-modules.
A morphism $f:M\rightarrow N$ is called left
$(A,\alpha)$-linear if
$f(a\c m)=a\c f(m),f\ci \mu= \nu\ci f.$
We denoted the category of left $(A,\alpha)$-Hom modules by
$\tilde{\mathcal{H}}(_{A}\mathcal {M}_{k})$.
\\
Similarly, let $(C,\gamma)$ be a monoidal Hom-coalgebra.
{\it A right $(C,\gamma)$-Hom-comodule} is an object
$(M,\mu)$ in $\tilde{\mathcal{H}}(\mathcal {M}_{k})$
together with a $k$-linear map
$\rho_M:M\rightarrow M\o C,\rho_M(m)=m_{(0)}\o m_{(1)}$ such that
\begin{equation}
\mu^{-1}(m_{(0)})\o \D_C(m_{(1)})
=(m_{(0)(0)}\o m_{(0)(1)})\o \gamma^{-1}(m_{(1)}),
\end{equation}
\begin{equation}
\rho_M(\mu(m))=\mu(m_{(0)})\o\gamma(m_{(1)}),
\ \ \
m_{(0)}\varepsilon(m_{(1)})=\mu^{-1}(m),
\end{equation}
for all $m\in M.$
\\
$(C,\gamma)$ is a Hom-comodule on itself via the Hom-comultiplication.
Let $(M,\mu)$ and $(N,\nu)$ be two right $(C,\gamma)$-Hom-comodules.
A morphism $g:M\rightarrow N$ is called right $(C,\gamma)$-colinear
if $g\ci \mu=\nu\ci g$ and
$g(m_{(0)})\o m_{(1)}=g(m)_{(0)}\o g(m)_{(1)}.$
The category of right
$(C,\gamma)$-Hom-comodules is denoted by
$\tilde{\cal{H}}(\cal {M}^C)$ .
\\
Let $(H, \alpha)$ be a monoidal Hom-bialgebra. We now recall from \cite{CZ2014}
that a monoidal Hom-algebra $(B,\beta)$ is called {\it a left
$H$-Hom-module algebra}, if $(B,\beta)$ is a left
$H$-Hom-module with action $\cdot$ obeying the following axioms:
\begin{equation}
h\c(ab)=(h_1\c a)(h_2\c b),\,\,\,\,\,\,\,\,h\c1_B=\varepsilon(h)1_B,
\end{equation}
for all $a,b\in B,h\in H.$
\\
Recall from \cite{LS2014} that a monoidal Hom-algebra $(B,\beta)$ is called
{\it a left $H$-Hom-comodule algebra}, if $(B,\beta)$ is a left
$H$-Hom-comodule with coaction $\rho$ obeying the following axioms:
$$\rho(ab) = a_{(-1)}b_{(-1)}\otimes a_{(0)}b_{(0)},
\ \ \ \rho_l(1_B)= 1_B\otimes 1_H,$$
for all $a,b\in B,h\in H.$
\\
Let $(H,m,\Delta,\alpha)$ be a monoidal Hom-bialgebra.
Recall from (\cite{CZ2014,LS2014}) that a {\it left-right Yetter-Drinfeld Hom-module}
over $(H,\alpha)$ is the object $(M,\cdot,\rho,\mu)$ which
is both in $\tilde{\cal{H}}(_{H}\cal {M})$ and $\tilde{\cal{H}}(\cal {M}^{H})$
obeying the compatibility condition:
\begin{equation}
h_{1}\c m_{(0)}\o h_{2}m_{(1)}=(\alpha(h_{2})\c m)_{(0)} \o
\alpha^{-1}(\alpha(h_{2})\c m)_{(1)})h_{1}.
\end{equation}
{\bf Remark 1.4.} (1) The category of all left-right Yetter-Drinfeld Hom-modules
is denoted by $\tilde{\cal{H}}(_{H}\cal {YD} ^{H})$ with understanding morphism.
(2) If $(H,\alpha)$ is a monoidal Hom-Hopf algebra with a bijective
antipode $S$, then the above equality is equivalent to
$$\r (h\c m)=\alpha(h_{21})\c m_{(0)}\o (h_{22}\a^{-1}(m_{(1)}))S^{-1}(h_{1}), $$
for all $h\in H$ and $m\in M$.
\\
\section*{2. $(A, B)$-YETTER-DRINFELD HOM-MODULES}
\def\theequation{2. \arabic{equation}}
\setcounter{equation}{0} \hskip\parindent
In this section, we define the notion of a Yetter-Drinfeld module over
a monoidal Hom-Hopf algebra that is twisted by two monoidal
Hom-Hopf algebra automorphisms
as well as the notion of a monoidal Hom-entwining structure and
show how such moinoidal Hom-entwining structures are obtained
from automorphisms of monoidal Hom-Hopf algebras.
\\
In what follows, let $(H,\alpha)$ be a monoidal Hom-Hopf algebra with
the bijective antipode $S$ and let ${\sl Aut}_{mHH}(H)$ denote the set
of all automorphisms of a monoidal Hopf algebra $H$.
\\
{\bf Definition 2.1.} Let $A, B\in {\sl Aut}_{mHH}(H)$.
A left-right {\sl $(A, B)$-Yetter-Drinfeld Hom-module} over $(H,\alpha)$
is a vector space $M$ such that:
\begin{eqnarray*}
&(1)&(M,\cdot,\mu) \mbox{ is a left }H \mbox{-Hom-module;}\\
&(2)& (M,\rho,\mu) \mbox{ is a right }H \mbox{-Hom-comodule;}\\
&(3)& \rho \mbox{ and }\cdot \mbox{ satisfy
the following compatibility condition: }
\end{eqnarray*}
\begin{equation}
\r (h\c m)=\alpha(h_{21})\c m_{(0)}\o (B(h_{22})\a^{-1}(m_{(1)}))A(S^{-1}(h_{1})),
\end{equation}
for all $h\in H$ and $m\in M$. We denote by
$ _{H}\mathcal{MHYD}^{H}(A, B)$
the category of left-right $(A, B)$-Yetter-Drinfeld
Hom-modules, morphisms
being $H$-linear $H$-colinear maps.
\\
{\bf Remark 2.2.} Note that, $A$ and $B$ are bijective,
Hom-algebra morphisms,
Hom-coalgebra morphisms,
and commute with $S$ and $\alpha$.
\\
{\bf Proposition 2.3.} One has that Eq.(2.1) is equivalent to
the following equation:
\begin{equation}
h_{1}\c m_{(0)}\o B(h_{2})m_{(1)}=\mu((h_{2}\c\mu^{-1} (m))_{(0)}) \o (h_{2}\c
\mu^{-1}(m))_{(1)}A(h_{1}).
\end{equation}
{\bf Proof.} Eq.(2.1)$\Longrightarrow$ Eq.(2.2). We compute as follows
\begin{eqnarray*}
&&\mu((h_{2}\c \mu^{-1}(m))_{(0)}) \o (h_{2}\c\mu^{-1}( m))_{(1)}A(h_{1})\\
&\stackrel{(2.1)}{=}& \mu(\a(h_{221})\c \mu^{-1}(m)_{(0)})\o(( B(h_{222})\a^{-2}(m_{(1)}))A(S^{-1}(h_{21})))A(h_{1})\\
&=& \mu(h_{12}\c \mu^{-1}(m)_{(0)})
\o(( B\a^{-2}(h_{2})\a^{-2}(m_{(1)}))A(S^{-1}\a(h_{112})))A\a^{2}(h_{111})\\
&=& \mu(\a^{-1}(h_{1}))\c \mu^{-1}(m)_{(0)})\o B(h_{2})m_{(1)}
=h_{1}\c m_{(0)}\o B(h_{2})m_{(1)}.
\end{eqnarray*}
For Eq.(2.2) $\Longrightarrow$ Eq.(2.1), we have
\begin{eqnarray*}
&& \alpha(h_{21})\c m_{(0)}\o (B(h_{22})\a^{-1}(m_{(1)}))A(S^{-1}(h_{1}))\\
&\stackrel{(2.2)}{=}& \mu((\alpha(h_{22})\c \mu^{-1}(m))_{(0)})\o \a^{-1}((\a(h_{22})\c\mu^{-1}(m))_{(1)}A\a(h_{21}))AS^{-1}(h_{1})\\
&=& \mu((h_{2})\c \mu^{-1}(m))_{(0)})\o \a^{-1}((h_{2}\c\mu^{-1}(m))_{(1)}A\a(h_{12}))AS^{-1}\a(h_{11})\\
&=& (\a(h_{2})\c m)_{(0)}\o (h_{2}\c \mu^{-1}(m))_{1}(A(h_{12})AS^{-1}(h_{11}))
=(h\c m)_{(0)}\o (h\c m)_{(1)}.
\end{eqnarray*}
This finishes the proof. \hfill $\blacksquare$
\\
{\bf Example 2.4.}
For $A=B=id_{H},$ we have
$ _{H}\mathcal{MHYD}^{H}(id, id)
=\mathcal{H}(_{H}\mathcal{YD}^{H})$, the usual monoidal
Yetter-Drinfeld Hom-module category. (see \cite{CZ2014,LS2014}).
\\
{\bf Example 2.5.} (1) Take a non-trivial monoidal Hom-Hopf algebra isomorphism
$B\in {\sl Aut}_{mHH}(H).$ We
define $(H_B,\alpha)=(H,\alpha)$ as $k$-vector spaces,
and we can consider a right $H$-Hom-comodule structure on
$H_B$ via the Hom-comultiplication $\Delta$ and
a left $H$-Hom-module structure on $H_B$ as follows:
$$
h\cdot y=(B(h_2)\alpha^{-1}(y))S^{-1}(\alpha(h_1)).
$$
for all $h\in H,\,y\in H_B.$ Then it is not hard to check that
$H_B\in\,\, _{H}\mathcal{MHYD}^{H}(id, B)$.
More generally, if $A,\,B\in {\sl Aut}_{mHH}(H),$
define $H_{(A,B)}$ as follows: $(H_{(A,B)},\alpha)=(H,\alpha)$ as $k$-vector spaces,
with right $H$-Hom-comodule structure via Hom-comultiplication
and left $H$-Hom-module structure given by:
$$
h\cdot x=(B(h_2)\alpha^{-1}(x))A(S^{-1}(\alpha(h_1))).
$$
for all $h,\,x\in H.$ It is straightforward to check
that $H_{(A,B)}\in\,\, _{H}\mathcal{MHYD}^{H}(A, B)$.
\\
(2) Recall from Example 3.5 in \cite{CWZ2013}
that $( H_{4} =k\{ 1,\ g,\ x, \ gx\,\},\a , \Delta , \varepsilon , S )$
is a monoidal Hom-Hopf algebra, where the algebraic structure are given as follows:
$\bullet$ The multiplication
$"\circ "$ is given by
\begin{eqnarray*}
&&1\circ 1=1,\ \ \ \ \ \ \ \ \ \ 1\circ g = g,
\ \ \ \ \ \ \ \ \ \ \ 1\circ x=cx,
\ \ \ \ \ \ \ \ \ \ 1\circ gx = cgx,\\
&&g\circ 1 =g,\ \ \ \ \ \ \ \ \ \ g\circ g=1,\ \ \ \ \
\ \ \ \ \ \ g\circ x=cgx,\ \ \ \ \ \ \ \ \ g\circ gx=cx,\\
&& x\circ 1=cx,
\ \ \ \ \ \ \ \ x\circ g=-cgx,
\ \ \ \ \ \ \ x\circ x=0,
\ \ \ \ \ \ \ \ \ \ x\circ gx=0,\\
&& gx\circ 1=cgx,\ \ \ \ \
gx\circ g=-cx,\ \ \ \ \ \ \
gx\circ x=0,\ \ \ \ \ \ \ \ x\circ gx=0;
\end{eqnarray*}
$\bullet$ The automorphism $\a$ is given by
$$
\a(1)=1,\ \a(g)=g,\ \a(x)=cx,\ \a(gx)=cgx,
$$
for all $0\neq c\in k;$
$\bullet$ The comultiplication $\Delta$ is defined by
\begin{eqnarray*}
&&\Delta(1)=1\otimes 1,\ \ \ \ \ \ \ \ \ \
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\Delta(g)=g\otimes g,\\
&&\Delta(x)=c^{-1}(x\otimes 1)+ c^{-1}(g\otimes x),
\ \ \Delta(gx)=c^{-1}(gx\otimes g)+c^{-1}(1\otimes gx);
\end{eqnarray*}
$\bullet$ The counit $\varepsilon$ is defined by
$$
\varepsilon(1)=1,\ \
\varepsilon(g)=1,\ \
\varepsilon(x)=0,\ \ \varepsilon(gx)=0,
$$
and
$\bullet$ The antipode $S$ is given by
$$
S(1)=1,\ \ S(g)=g,\ \ S(x)=-gx,
\ \ S(gx)=-x,
$$
\\
Still from Example 3.5 in \cite{CWZ2013} that we have the automorphism group of the monoidal Hom-Hopf algebra
$H_{4}$: ${\sl Aut}_{mHH}(H_{4})=\{\left(
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & \lambda & 0 \\
0 & 0 & 0 & \lambda \\
\end{array}
\right)\mid 0\neq\lambda\in k\}$
\\
\\
In what follows, we will give an explicit describe on the Yetter-Drinfeld
Hom-modules given in Part (1) above for $H_4$.
\\
Let $A=\left(
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & c' & 0 \\
0 & 0 & 0 & c' \\
\end{array}
\right)$.
By Part (1), we can define $H_{4 A}=(H_4,\a)$ as $k$-vector spaces,
but with the right $H_4$-Hom-comodule structure
via $\Delta$ and the left $H_4$-module structures as follows:
\begin{eqnarray*}
&&1\cdot 1=1,\ \ \ \ \ \ \ \ 1\cdot g = g,
\ \ \ \ \ \ \ 1\cdot x=cx, \ \ \ \ \ \ \ \ 1\cdot gx = cgx,\\
&&g\cdot 1 =1,\ \ \ \ \ \ \ \ g\cdot g=g,\ \ \ \ \ \ \
g\cdot x=-cx,\ \ \ \ \ \ g\cdot gx=-cgx,\\
&& x\cdot 1=-c(1+c')gx,
\ \ \ \ \ \
x\cdot g=c(1-c')x,\ \ \ \ \ \
x\cdot x = 0,
\ \ \ \ \ \
x\cdot gx=0,\\
&& gx\cdot 1=c(1-c')gx,\ \ \
\ \ \ gx\cdot g=-c(1+c')x,\ \ \ \
gx\cdot x=0,
\ \ \ \ \ x\cdot gx=0,
\end{eqnarray*}
for any $0\neq c',c\in k$.
\\
Then one can check that $H_{4 A}\in _{H_4}\mathcal {MHYD}^{H_4}(A,id)$,
i.e., a left-right $(A, id)$-Yetter-Drinfeld Hom-module over $(H_4, \a )$.
\\
Let $B=\left(
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & c'' & 0 \\
0 & 0 & 0 & c'' \\
\end{array}
\right)$. Similarly,
define $(H_{4 B},\a)=(H_4,\a)$ as $k$-vector spaces
with the right $H_4$-Hom-comodule structure
via $\Delta$ and the left $H_4$-module structures as follows:
\begin{eqnarray*}
&&1\cdot 1=1,\ \ \ \ \ \ \ \ \ \ 1\cdot g = g,
\ \ \ \ \ \ \ \ \ \ 1\cdot x=cx,
\ \ \ \ \ \ \ \ \ \ 1\cdot gx = cgx\\
&&g\cdot 1 =1,\ \ \ \ \ \ \ \ \ \ g\cdot g=g,\ \ \ \ \
\ \ \ \ \ g\cdot x=-cx,\ \ \ \ \ g\cdot gx=-cgx\\
&& x\cdot 1=-c(1+c'')gx,
\ \ \ \ \ x\cdot g=c(-1+c'')x,
\ \ \ \ \ \ \ \ x\cdot x=0,
\ \ \ \ \ x\cdot gx=0,\\
&& gx\cdot 1=c(-1+c'')gx,\ \ \ \ \
gx\cdot g=-c(1+c'')x,\ \ \ \ \
gx\cdot x=0,\ \ \ \ \ x\cdot gx=0,
\end{eqnarray*}
for any $0\neq c'',c\in k$.
It is straightforward to see that
$H_{4 B}\in _{H_4}\mathcal {MHYD}^{H_4}(id,B)$,
a left-right $(id, B)$-Yetter-Drinfeld Hom-module over $(H_4, \a )$.
\\
Let $A=\left(
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & c' & 0 \\
0 & 0 & 0 & c' \\
\end{array}
\right)$ and $B=\left(
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & c'' & 0 \\
0 & 0 & 0 & c'' \\
\end{array}
\right)$.
Define $(H_{4 A,B},\a)=(H_4,\a)$ as $k$-vector spaces,
with the right $H_4$-Hom-comodule structure
via $\Delta$ and the left $H$-module structures as follows:
\begin{eqnarray*}
&&1\cdot 1=1,\ \ \ \ \ \ \ \ \ \ 1\cdot g = g,
\ \ \ \ \ \ \ \ \ \ 1\cdot x=cx,
\ \ \ \ \ \ \ \ \ \ 1\cdot gx = cgx\\
&&g\cdot 1 =1,\ \ \ \ \ \ \ \ \ \ g\cdot g=g,\ \ \ \ \
\ \ \ \ \ g\cdot x=-cx,\ \ \ \ \ g\cdot gx=-cgx\\
&& x\cdot 1=-c(c'+c'')gx,
\ \ \ \ \ x\cdot g=c(-c'+c'')x,
\ \ \ \ \ \ \ \ x\cdot x=0,
\ \ \ \ \ x\cdot gx=0,\\
&& gx\cdot 1=c(-c'+c'')gx,\ \ \ \ \
gx\cdot g=-c(c'+c'')x,\ \ \ \ \
gx\cdot x=0,\ \ \ \ \ x\cdot gx=0,
\end{eqnarray*}
for any $0\neq c'',c\in k$.
\\
Then it is straightforward to see that $H_{4 (A,B)}$ is
a left-right $(A, B)$-Yetter-Drinfeld Hom-module over $(H_4, \a )$,
i.e.,
$H_{4 (A,B)}\in _{H_4}\mathcal {MHYD}^{H_4}(A,B).$
\\
{\bf Definition 2.6.} A left-right {\sl monoidal Hom-entwining structure} is a triple
$(H, \, C,\, \psi)$, where $(H, \a)$ is a monoidal Hom-algebra
and $(C,\gamma)$ is a monoidal Hom-coalgebra with a linear map
$\psi: H\otimes C\rightarrow H\otimes C,
\,\,\,\,h\otimes c\mapsto _{\psi}h\otimes c^{\psi}$
satisfying the following conditions:
\begin{equation}
_{\psi}(hg)\otimes c^{\psi}= _{\phi}h_{\psi}g\otimes \gamma(\gamma^{-1}(c)^{\psi\phi}),
\end{equation}
\begin{equation}
_{\psi}1\otimes c^{\psi}= 1_A\otimes c,
\end{equation}
\begin{equation}
_{\psi}h\otimes \Delta(c^{\psi})= \a(_{\phi\psi}\a^{-1}(h))\otimes (c^{\phi}_{1}\otimes c^{\psi}_{2}),
\end{equation}
\begin{equation}
\varepsilon(c^{\psi}) _{\psi}h= \varepsilon(c)a,
\end{equation}
Over a monoidal Hom-entwining structure $(H, \, C,\, \psi)$,
a left-right monoidal entwined Hom-module $M$ is both a right
$C$-Hom-comodule and a left $H$-Hom-module such that
$$ \rho^{M}(h\cdot m)=_{\psi}\a^{-1}(h)\cdot m_{(0)}\otimes \a(m_{(1)})^{\psi}$$
for all $h\in H$ and $m\in M$. We denote the category of all
monoidal entwined Hom-modules over $(H,\,C,\,\psi)$
by $_{H}\mathcal{M}^{C}(\psi)$.
\\
Let $(H,\a)$ be a monoidal Hom-Hopf algebra with $S$,
and define a linear map
$$\psi(A,B):H\otimes H\rightarrow H\otimes H,\
\ a\otimes c\mapsto _{\psi}a\otimes c^{\psi}
=\a^{2}(a_{21})\otimes (B(a_{22})\a^{-2}(c))AS^{-1}(a_1),$$
for all $A, B\in {\sl Aut}_{mHH}(H)$.
\\
{\bf Proposition 2.7.} With notations as above,
$(H, H,\psi(A,B))$ is a monoidal Hom-entwining structure
\underline{} for all $A, B\in {\sl Aut}_{mHH}(H)$.
{\bf Proof.} We need to prove that Eqs.(2.3-2.6) hold.
First, it is straightforward to check Eqs.(2.4) and (2.6).
In what follows, we only verify Eqs.(2.3) and (2.5).
In fact, for all $a,b,c\in H$, we have
\begin{eqnarray*}
&&_{\phi}a_{\psi}b\otimes \a(\a^{-1}(c)^{\psi\phi})\\
&=&\a^2(a_{21})_{\psi}b \otimes
\a((B(a_{22})\a^{-2}(\a^{-1}(c)^{\psi}))AS^{-1}(a_1))\\
&=&\a^2(a_{21})\a^2(b_{21})
\otimes\a([B(a_{22})\a^{-2}((B(b_{22})\a^{-3}(c))AS^{-1}(b_1))]AS^{-1}(a_1))\\
&=&\a^2(a_{21}b_{21})\otimes[B\a(a_{22})
((B\a^{-1}(b_{22})\a^{-4}(c))AS^{-1}\a^{-1}(b_1))]AS^{-1}\a(a_1)\\
&=&\a^2(a_{21}b_{21})
\otimes(B(a_{22}b_{22})\a^{-2}(c))(AS^{-1}(b_1)AS^{-1}(a_1))
= _{\psi}(ab)\otimes c^{\psi},
\end{eqnarray*}
and Eq.(2.3) is proven.
For all $a\in H$, we have
\begin{equation}
a_{1}\otimes a_{211}\otimes a_{2121}\otimes a_{2122}\otimes a_{22}
=\a(a_{11})\otimes \a^{-1}(a_{12})\otimes \a^{-2}(a_{21})
\otimes \a^{-1}(a_{221})\otimes \a(a_{222})
\end{equation}
As for Eq.(2.5), we compute:
\begin{eqnarray*}
&& \a(_{\phi\psi}\a^{-1}(a))\otimes (c^{\phi}_{1}\otimes c^{\psi}_{2})\\
&=&\a(\a^{2}((_{\psi}\a^{-1}(a))_{21}))\otimes
((B((_{\psi}\a^{-1}(a))_{22})\a^{-2}(c_{1}))AS^{-1}((_{\psi}\a^{-1}(a))_{1})
\otimes c^{\psi}_{2})\\
&=&\a(\a^{2}(\a^{2}(\a^{-1}(a)_{21})_{21}))\otimes
((B((\a^{2}(\a^{-1}(a)_{21})_{22})\a^{-2}(c_{1}))AS^{-1}((\a^{2}(\a^{-1}(a)_{21})_{1})\\
&&\otimes (B(\a^{-1}(a)_{22})\a^{-2}(c_{2}))AS^{-1}(\a^{-1}(a)_{1}))\\
&=&\a^{4}(a_{2121})\otimes
((B\a(a_{2122})\a^{-2}(c_{1}))AS^{-1}\a(a_{211})
\otimes (B\a^{-1}(a_{22})\a^{-2}(c_{2}))AS^{-1}\a^{-1}(a_{1}))\\
&\stackrel{(2.7)}{=}&\a^{2}(a_{21})\otimes
((B(a_{221})\a^{-2}(c_{1}))AS^{-1}(a_{12})
\otimes (B(a_{222})\a^{-2}(c_{2}))AS^{-1}(a_{11}))\\
&=&\a^{2}(a_{21})\otimes((B(a_{22})\a^{-2}(c))AS^{-1}(a_1))_1
\otimes(B(a_{22})\a^{-2}(c))AS^{-1}(a_1))_2)\\
&=&_{\psi}a\otimes \Delta(c^{\psi}).
\end{eqnarray*}
and Eq.(2.5) is proven.
This finishes the proof. \hfill $\blacksquare$
\\
By Proposition 2.7, we have a monoidal entwined
Hom-module category $ _{H}\mathcal{M}^{H}(\psi(A, B))$
over $(H, H,\psi(A,B))$ with $A,B\in{\sl Aut}_{mHH}(H)$.
In this case, for all $M\in _{H}\mathcal{M}^{H}(\psi(A, B))$,
we have
$$\r (h\c m)=\alpha(h_{21})\c m_{(0)}
\o (B(h_{22})\a^{-1}(m_{(1)}))A(S^{-1}(h_{1})),$$
for all $h\in H,m\in M$. Thus means that
$_{H}\mathcal{M}^{H}(\psi(A, B))=
_{H}\mathcal{MHYD}^{H}(A, B)$
as categories.
\\
{\bf Definition 2.8.} Let $(H,\alpha)$ be a
monoidal Hom-algebra. A monoidal Hom-algebra
$(N,\nu)$ is called an $(H,\alpha)$-Hom-bicomodule
algebra, if $(N,\nu)$ is a left $(H,\alpha)$-Hom-comodule
and a right $(H,\alpha)$-Hom-comodule
with coactions $\rho_r$ and $\rho_l$
obeying the following axioms:
\begin{eqnarray*}
&(1)&\rho_l(n) = n_{[-1]}\otimes n_{[0]},\ \mbox{ and }
\ \rho_r(n)= n_{<0>}\otimes n_{<1>},\\
&(2)& (N,\nu)\mbox{ is a left }H \mbox{-Hom-comodule algebra;}\\
&(3)& (N,\nu)\mbox{ is a right }H \mbox{-Hom-comodule algebra;}\\
&(4)& \rho_l \mbox{ and }\rho_l \mbox{ satisfy
the following compatibility condition: for all } n\in N,\\
&\ \ \ \ \ &n_{<0>[-1]}\otimes n_{<0>[0]}\otimes \alpha^{-1}(n_{<1>})
= \alpha^{-1}(n_{[-1]})\otimes n_{[0]<0>}\otimes n_{[0]<1>}\\
&\ \ \ \ \ & \quad \quad =n_{\{-1\}}\otimes n_{\{0\}}\otimes n_{\{1\}}
\in(H\otimes N)\otimes H = H\otimes( N\otimes H).
\end{eqnarray*}
\\
{\bf Example 2.9.} Let $A,B\in {\sl Aut}_{mHH}(H)$,
and $H_{(A,B)}=H$ as algebra, with $H$-Hom-comodule
structures as follows:
for all $h\in H,$
\begin{eqnarray*}
&& H_{(A,B)}\rightarrow H\otimes H_{(A,B)},\,\,\,
h\mapsto h_{[-1]}\otimes h_{[0]}=A(h_{1})\otimes h_{2},\\
&& H_{(A,B)}\rightarrow H_{(A,B)}\otimes H,\,\,\,
h\mapsto h_{<0>}\otimes h_{<1>}=h_{1}\otimes B(h_{2}).
\end{eqnarray*}
Then on can check $H_{(A,B)}$ is an $H$-Hom-bimodule algebra.
\\
{\bf Definition 2.10.} Let $(H,\alpha)$ be a
monoidal Hom-Hopf algebra and
$(N,\nu)$ be an $H$-Hom-bicomodule algebra,
A left-right Yetter-Drinfeld Hom-module
is a $k$-modules $(M,\mu)$ together with a left
$N$-action (denoted by $n\otimes m\mapsto n\cdot m$)
and a right $H$-coaction (denoted by $ m\mapsto m_{0}\otimes m_{1}$)
satisfying the eqivalent compatibility conditions:
\begin{eqnarray*}
&&(n\cdot m)_{0}\otimes(n\cdot m)_{1}
=\nu(n_{[0]<0>})\cdot m_{0}
\otimes( n_{[0]<1>}\alpha^{-1}(m_{1}))S^{-1}(n_{[-1]}),\\
&&n_{<0>}\cdot m_{0}\otimes n_{<1>} m_{1}
=\mu((n_{[0]}\cdot \mu^{-1}(m))_{0})
\otimes(n_{[0]}\cdot \mu^{-1}(m))_{1}n_{[-1]}.
\end{eqnarray*}
for all $n\in N$ and $m\in M$.
Then we all $(H,N,H)$ a Yetter-Drinfeld Hom-datum
$(H,N,H)$(the second $H$ is regarded as an $H$-Hom-bimodule coalgebra).
Our notion for the category of a left-right Yetter-Drinfeld Hom-modules
and $N$-linear $H$-colinear maps
will be $ _{N}\mathcal{MHYD}^{H}(H).$
\\
{\bf Example 2.11.} Let $H_{(A,B)}$ be an $H$-Hom-cobimodule
algebra, with an $H$-Hom-comodule structures
shown in Example 2.9.
Then we can consider the Yetter-Drinfeld Hom-datum
$(H,H_{(A,B)},H)$ and the Yetter-Drinfeld Hom-modules
over it, $ _{H_{(A,B)}}\mathcal{MHYD}^{H}(H)$.
\\
{\bf Proposition 2.12.}
$_{H}\mathcal{MHYD}^{H}(A, B)= _{H_{(A,B)}}\mathcal{MHYD}^{H}(H)$.
\\
It is easy to see that the compatibility
conditions for the two categories are the same.
The easy proof of this is left to the reader.
\section*{3. A BRAIDED $T$-CATEGORY $\mathcal {MHYD}(H)$}
\def\theequation{3. \arabic{equation}}
\setcounter{equation} {0} \hskip\parindent
In this section, we will construct a class of new braided $T$-categories
$\mathcal {MHYD}(H)$ over any monoidal Hom-Hopf algebra $(H,\alpha)$.
\\
{\bf Proposition 3.1.} If $(M,\mu) \in {_{H}}\mathcal {MHYD}^{H}(A,B)$
and $(N,\nu)\in {_{H}}\mathcal {MHYD}^{H}(C,D)$, with
$A,B,C,D\in {\sl Aut}_{mHH}(H)$, then $(M \o N,\mu\o\nu)
\in {_{H}}\mathcal {MHYD}^{H}(AC, DC^{-1}BC)$ with structures as follows:
\begin{eqnarray*}
&&h\c (m \o n)=C (h_{1})\c m \o C^{-1}BC(h_{2})\c n,\\
&&m\o n \mapsto (m_{(0)}\o n_{(0)})\o n_{(1)}m_{(1)}.
\end{eqnarray*}
for all $m\in M,n\in N$ and $h\in H.$
\\
{\bf Proof.} First, it is easy to get that
$(M \o N,\mu\o\nu)$ is a left $H$-module
and a right $H$-comodule.
Next, we compute the compatibility condition as follows:
\begin{eqnarray*}
&&(h\c (m\o n))_{(0)}\o (h\c (m\o n))_{(1)}\\
&=&((C(h_{1})\c m)_{(0)} \o (C^{-1}BC(h_{2})\c n)_{(0)})\o
(C^{-1}BC(h_{2})\c n)_{(1)} C(h_{1}\c m)_{(1)}\\
&\stackrel{(2.1)}{=}&(C\a (h_{121})\c m_{(0)} \o C^{-1}BC\a(h_{221})\c n_{(0)})\o
[(DC^{-1}BC(h_{222})\a^{-1}(n_{(1)}))\\
&& CS^{-1}C^{-1}BC(h_{21})][(BC(h_{122})\a^{-1}(m_{(1)}))S^{-1}AC(h_{11})]\\
&=&(C(h_{12})\c m_{(0)} \o C^{-1}BC\a(h_{221})\c n_{(0)})\o
[(DC^{-1}BC(h_{222})\a^{-1}(n_{(1)})) \\
&&S^{-1}BC\a(h_{212})][(BC(h_{211})\a^{-1}(m_{(1)}))S^{-1}AC(h_{11})]\\
&=&(C(h_{12})\c m_{(0)} \o C^{-1}BC\a(h_{221})\c n_{(0)})\o
(DC^{-1}BC\a(h_{222})n_{(1)}) \\
&&[(BC\a^{-1}(S^{-1}(h_{212})h_{211})\a^{-1}(m_{(1)}))S^{-1}AC(h_{11}))]\\
&=&(C(h_{12})\c m_{(0)} \o C^{-1}BC\a(h_{221})\c n_{(0)})\o
(DC^{-1}BC\a(h_{222})n_{(1)}) \\
&&[(BC\a^{-1}(\varepsilon(h_{21})1_{H})\a^{-1}(m_{(1)}))S^{-1}AC(h_{11}))]\\
&=&(C(h_{12})\varepsilon(h_{21})\c m_{(0)} \o C^{-1}BC\a(h_{221})\c n_{(0)})\o
(DC^{-1}BC\a(h_{222})n_{(1)}) \\
&&(m_{(1)}S^{-1}AC(h_{11}))\\
&=&(C\a(h_{121})\varepsilon\a(h_{122})\c m_{(0)} \o C^{-1}BC(h_{21})\c n_{(0)})\o
(DC^{-1}BC(h_{22})n_{(1)}) \\
&&(m_{(1)}S^{-1}AC(h_{11}))\\
&=&(C(h_{12})\c m_{(0)} \o C^{-1}BC(h_{21})\c n_{(0)})\o
(DC^{-1}BC(h_{22})(\a^{-1}(n_{(1)}) \a^{-1}(m_{(1)})))\\
&&S^{-1}AC\a(h_{11})\\
&=&(C\a(h_{211})\c m_{(0)} \o C^{-1}BC\a(h_{212})\c n_{(0)})\o
(DC^{-1}BC(h_{22})(\a^{-1}(n_{(1)}m_{(1)}))\\
&&S^{-1}AC(h_{1})\\
&=&\a(h_{21})\c (m\o n)_{(0)}\o DC^{-1}BC(h_{22})\a^{-1}(m\o
n)_{(1)}AC(S^{-1}(h_{1})).
\end{eqnarray*}
for all $m\in M,n\in N$ and $h\in H.$
This completes the proof.
\hfill $\blacksquare$
\\
{\bf Remark. 3.2.} Note that, if $(M,\mu)\in {_{H}}\mathcal
{MHYD}^{H}(A,B),\,\, (N,\nu) \in {_{H}}\mathcal {MHYD}^{H}(C,D)$ \\
and
$(P,\varsigma) \in {_{H}}\mathcal {MHYD}^{H}(E, F)$, then\,\, $(M\o
N)\o P=M \o(N\o P)$
as objects \\
in $_{H}\mathcal {MHYD}^{H}(ACE,
FE^{-1}DC^{-1}BCE)$.
\\
Denote $G={\sl Aut}_{mHH}(H)
\times {\sl Aut}_{mHH}(H)$
a group with multiplication as follows:
for all $A,B,C,D\in {\sl Aut}_{mHH}(H)$,
\begin{equation}
(A,B)\ast (C,D)=(AC, DC^{-1}BC).
\end{equation}
The unit of this group is $(id,id)$ and $(A,B)^{-1}=(A^{-1},
AB^{-1}A^{-1})$.
The above proposition means that if
$M \in {_{H}}\mathcal {MHYD}^{H}(A,B)$
and $N\in {_{H}}\mathcal {MHYD}^{H}(C,D)$, then
$M \o N \in {_{H}}\mathcal {MHYD}^{H}((A,B)\ast (C,D)).$
\\
{\bf Proposition 3.3.} Let $(N,\nu) \in {_{H}}\mathcal {MHYD}^{H}(C,D)$
and $(A,B)\in G$. Define ${}^{(A,B)}N=N$ as
vector space, with structures: for all $n\in N$ and $h\in H.$
$$h\rhd n=C^{-1}BCA^{-1}(h)\c n,$$
\begin{equation}
n\mapsto n_{<0>}\o n_{<1>}=n_{(0)}\o AB^{-1}(n_{(1)}).
\end{equation}
Then
$${}^{(A,B)}N \in {_{H}}\mathcal {MHYD}^{H}(ACA^{-1},AB^{-1}DC^{-1}BCA^{-1})
={_{H}}\mathcal {MHYD}^{H}((A,B)\ast
(C,D)\ast (A,B)^{-1}).$$
{\bf Proof.} Obviously, the equations above define
a module and a comodule action.
In what follows,
we show the compatibility condition:
\begin{eqnarray*}
&&(h\rhd n)_{<0>}\o(h\rhd n)_{<1>}\\
&=&(C^{-1}BCA^{-1}(h)\c n)_{(0)}\o AB^{-1}((C^{-1}BCA^{-1}(h)\c n)_{(1)})\\
&=&C^{-1}BCA^{-1}\a(h_{21})\c n_{(0)}\o AB^{-1}((DC^{-1}BCA^{-1}(h_{22})\a^{-1}(n_{(1)}))\\
&&CC^{-1}BCA^{-1}S^{-1}(h_{1}))\\
&=&C^{-1}BCA^{-1}\a(h_{21})\c n_{(0)}\o (AB^{-1}DC^{-1}BCA^{-1}(h_{22})AB^{-1}\a^{-1}(n_{(1)}))
ACA^{-1}S^{-1}(h_{1}))\\
&=&\a(h_{21})\rhd n_{<0>}\o (AB^{-1}DC^{-1}BCA^{-1}(h_{22})\a^{-1}(n_{<1>}))
ACA^{-1}S^{-1}(h_{1}))
\end{eqnarray*}
for all $n\in N$ and $h\in H,$
that is ${}^{(A,B)}N \in {_{H}}\mathcal {MHYD}^{H}(ACA^{-1},AB^{-1}DC^{-1}BCA^{-1})$
\hfill $\blacksquare$
\\
{\bf Remark. 3.4.} Let $(M,\mu) \in {_{H}}\mathcal {MHYD}^{H}(A, B),
\ \ (N,\nu) \in {_{H}}\mathcal
{MHYD}^{H}(C, D),\,\, \mbox {and}\, (E,F)\in G$. Then by the
above proposition, we have:
$$
{}^{(A, B)\ast (E, F)}N={}^{(A, B)}({}^{(E, F)}N),
$$
as objects in $_{H}\mathcal {MHYD}^{H}(AECE^{-1}A^{-1},
AB^{-1}EF^{-1}DC^{-1}FE^{-1}BECE^{-1}A^{-1})$ and
$$
{}^{(E,F)}(M\o N)= {}^{(E,F)}M \o {}^{(E,F)}N,
$$
as objects in $_{H}\mathcal {MHYD}^{H}(EACE^{-1},
EF^{-1}DC^{-1}BA^{-1}FACE^{-1})$.
\\
{\bf Proposition 3.5.} Let $(M,\mu) \in {_{H}}\mathcal {MHYD}^{H}(A,B)$
and $(N,\nu)\in {_{H}}\mathcal {MHYD}^{H}(C,D)$, take
${}^{M}N={}^{(A,B)}N$ as explained in Subsection 1.2.
Define a map $c_{M, N}: M \o N \rightarrow {}^{M}N \o M$ by
\begin{equation}
c_{M,N}(m\o
n)=\nu(n_{(0)})\o B^{-1}(n_{(1)})\c \mu^{-1}(m).
\end{equation}
for all $m\in M,n\in N.$
Then $c_{M, N}$ is both an $H$-module map and an $H$-comodule map,
and satisfies the following formulae
(for $(P,\varsigma) \in {_{H}}\mathcal {MHYD}^{H}(E,F)$):
\begin{equation}
a^{-1}_{{}^{M\o N}P, M, N}\circ c _{M\o N, P}\circ a^{-1}_{M, N, P}=(c _{M, {}^NP}\o
id_N)\circ a^{-1}_{M, {}^NP, N}\circ (id _M\o c _{N, P}) ,
\end{equation}
\begin{equation}
a_{{}^MN, {}^MP, M}
\circ c _{M, N\o P}\circ a_{M, N, P}=(id _{{}^MN }\o c _{M, P})\circ a_{{}^MN, M, P}\ci
(c _{M, N}\o id_P).
\end{equation}
Furthermore, if $(M,\mu) \in {_{H}}\mathcal {MHYD}^{H}(A,B)$ and
$(N,\nu) \in{_{H}}\mathcal {MHYD}^{H}(C,D)$, then\\
$c_{{}^{(E,F)}M,{}^{(E,F)}N}=c_{M,N},$
for all $(E,F)\in G$.
\smallskip
\\
{\bf Proof.} First, we prove that $c_{M,N}$ is an $H$-module map.
Take $h\cdot(m\o n)=C(h_{1})\c m\o C^{-1}BC(h_{2})\c n$
and $h\cdot(n\o m)= C^{-1}BC(h_{1})\c n\o B^{-1}DC^{-1}BC(h_{2})\c m$
as explained in Proposition 3.1.
\begin{eqnarray*}
&&c_{M,N}(h\cdot(m\o n))\\
&=&\nu((C^{-1}BC(h_{2})\c n)_{(0)})\o
B^{-1}((C^{-1}BC(h_{2})\c n)_{(1)})\c\mu^{-1}(C(h_{1})\c m)\\
&=&\nu(C^{-1}BC\a(h_{221})\c n_{(0)})\o
B^{-1}((DC^{-1}BC(h_{222})\a^{-1}( n_{(1)}))CC^{-1}BCS^{-1}(h_{21}))\\
&&\c\mu^{-1}(C(h_{1})\c m)\\
&=&\nu(C^{-1}BC\a(h_{221})\c n_{(0)})\o
B^{-1}(DC^{-1}BC\a(h_{222}) n_{(1)})\\
&&\c ((CS^{-1}\a^{-1}(h_{21})
C\a^{-2}(h_{1}))\c \mu^{-1}(m))\\
&=&\nu(C^{-1}BC(h_{21})\c n_{(0)})\o
B^{-1}(DC^{-1}BC(h_{22}) n_{(1)})\\
&&\c (C\a^{-1}((S^{-1}(h_{12})
h_{11}))\c \mu^{-1}(m))\\
&=&\nu(C^{-1}BC(h_{21})\c n_{(0)})\o
B^{-1}(DC^{-1}BC(h_{22}) n_{(1)})\c (C\a^{-1}(\varepsilon(h_{1})1_{H})\c \mu^{-1}(m))\\
&=&C^{-1}BC(h_{1})\c \nu (n_{(0)})\o
B^{-1}DC^{-1}BC(h_{2})\c(B^{-1}( n_{(1)})\c \mu^{-1}( m)) \\
&=&\psi_{N\o M}((h\o(\nu(n_{(0)})\o B^{-1}(n_{(1)})\c \mu^{-1}(m)))\\
&=&\psi_{N\o M}\ci(id\o c_{M,N})(h\o(m\o n))
\end{eqnarray*}
Secondly, we check that $c_{M, N}$ is an $H$-comodule map as
follows:
\begin{eqnarray*}
&&\rho_{N\o M}\ci c_{M,N}(m\o n) \\
&=&((\nu(n_{(0)}))_{<0>}\o( B^{-1}(n_{(1)})\c\mu^{-1}(m))_{(0)})
\o( B^{-1}(n_{(1)})\c\mu^{-1}(m))_{(1)}\\
&&(\nu(n_{(0)}))_{<1>}\\
&=&(\nu(n_{(0)(0)})\o( B^{-1}(n_{(1)})\c\mu^{-1}(m))_{(0)})
\o( B^{-1}(n_{(1)})\c\mu^{-1}(m))_{(1)}\\
&&AB^{-1}\a(n_{(0)(1)})\\
&=&(\nu(n_{(0)(0)})\o B^{-1}\a(n_{(1)21})\c\mu^{-1}(m)_{(0)})
\o((B B^{-1}(n_{(1)22})\a^{-2}(m_{(1)}))\\
&&AB^{-1}S^{-1}(n_{(1)1}))AB^{-1}\a(n_{(0)(1)})\\
&\stackrel{(1.7)}{=}&(n_{(0)}\o B^{-1}\a(n_{(1)21})\c\mu^{-1}(m_{(0)}))
\o(\a(n_{(1)22})\a^{-1}(m_{(1)}))\\
&&(AB^{-1}S^{-1}\a(n_{(1)12})AB^{-1}\a(n_{(1)11}))\\
&=&(n_{(0)}\o B^{-1}\a(n_{(1)21})\c\mu^{-1}(m_{(0)}))
\o(\a^2(n_{(1)22})m_{(1)})\varepsilon(n_{(1)1})\\
&\stackrel{(1.7)}{=}&(\nu( n_{(0)(0)})\o B^{-1}(n_{(0)(1)})\c\mu^{-1}(m_{(0)}))
\o n_{(1)}m_{(1)}\\
&=&(c_{M,N}\o id)((m_{(0)}\o n_{(0)})\o n_{(1)}m_{(1)})
=(c_{M,N}\o id)\rho(m\o n).
\end{eqnarray*}
Then we will check Eqs.(3.4) and (3.5). On the one hand,
\begin{eqnarray*}
&&a^{-1}_{{}^{M\o N}P, M, N}\circ c _{M\o N, P}\circ a^{-1}_{M, N, P}(m\o (n\o p))\\
&=&a^{-1}_{{}^{M\o N}P, M, N}\circ c _{M\o N, P}((\mu^{-1}(m)\o n)\o \varsigma(p))\\
&=&a^{-1}_{{}^{M\o N}P, M, N}(\varsigma^{2} ( p_{(0)} )\o C^{-1}B^{-1}CD^{-1}\a( p_{(1)} )
\c(\mu^{-2}(m)\o \nu^{-1}(n)))\\
&=&(\varsigma ( p_{(0)} )\o B^{-1}CD^{-1}\a( p_{(1)1})\c \mu^{-2}(m))\o
D^{-1}\a^{2}( p_{(1)2})\c n\\
&=&(\varsigma^{2}(p_{(0)(0)})\o B^{-1}CD^{-1}\a(p_{(0)(1)})\c\mu^{-2}(m))
\o D^{-1}\a(p_{(1)})\c n\\
&=&(\varsigma(\varsigma(p_{(0)})_{<0>})\o B^{-1}(\varsigma(p_{(0)})_{<1>})\c\mu^{-2}(m))
\o D^{-1}\a(p_{(1)})\c n\\
&=&(c _{M, {}^NP}\o id_N)((\mu^{-1}(m)\o \varsigma(p_{(0)}))
\o D^{-1}\a(p_{(1)})\c n)\\
&=&(c _{M, {}^NP}\o id_N)\circ a^{-1}_{M, {}^NP, N}\circ (id _M\o c _{N, P})(m\o(n\o p)
\end{eqnarray*}
On the another hand,
\begin{eqnarray*}
&&a_{{}^MN, {}^MP, M} \circ c _{M, N\o P}\circ a_{M, N, P}((m\o n)\o p)\\
&=&a_{{}^MN, {}^MP, M} \circ c _{M, N\o P}(\mu(m)\o( n\o\varsigma^{-1} (p)))\\
&=&a_{{}^MN, {}^MP, M} ((\nu\o\varsigma)(n\o \varsigma^{-1}(p))_{(0)}\o
B^{-1}((n\o\varsigma^{-1} (p))_{(1)})\c \mu^{-1}\mu(m)\\
&=&\nu^{2}(n_{(0)})\o( p_{(0)}\o B^{-1}\a^{-1}(\a^{-1} (p_{(1)})n_{(1)})\c \a^{-1}(m))\\
&=&\nu^{2}(n_{(0)})\o \varsigma(\varsigma^{-1}(p)_{(0)})
\o (B^{-1}(\varsigma^{-1}(p)_{(1)})\c(B^{-1}\a^{-1} (n_{(1)})\c \a^{-1}(m)))\\
&=&(id _{{}^MN }\o c _{M, P})
((\nu^{2}(n_{(0)})\o B^{-1}(n_{(1)})\c \mu^{-1}(m))\o p)\\
&=&(id _{{}^MN }\o c _{M, P})\circ a_{{}^MN, M, P}\ci (c _{M, N}\o id_P)
((m\o n)\o p)
\end{eqnarray*}
The proof is completed.
\hfill $\blacksquare$
\\
{\bf Lemma 3.6.} The map $c_{M,N}$
defined by $c_{M,N}(m\o n)=\nu(n_{(0)})
\o B^{-1}(n_{(1)})\c \mu^{-1}(m)$ is bijective; with inverse
$${c}_{M,N}^{-1}(n\o m)=B^{-1}(S(n_{(1)}))\c\mu^{-1}(m) \o \nu (n_{(0)}).$$
{\bf Proof.} First, we prove $c_{M,N}{c}_{M,N}^{-1}=id$. For all $m\in M, n\in
N$, we have
\begin{eqnarray*}
&&c_{M,N}{c}_{M,N}^{-1}(n \o m)\\
&=&c_{M,N}(B^{-1}S(n_{(1)})\c\mu^{-1}(m)\o\nu(n_{(0)}))\\
&=&\nu(\nu(n_{(0)})_{(0)})\o B^{-1}(\nu(n_{(0)})_{(1)})\c \mu^{-1}(B^{-1}S(n_{(1)})\c\mu^{-1}(m))\\
&=&\nu^{2}(n_{(0)(0)})\o B^{-1}((n_{(0)(1)})S\a^{-1}(n_{(1)}))\c\mu^{-1}(m)\\
&=&\nu(n_{(0)})\o B^{-1}((n_{(1)1})S(n_{(1)2}))\c\mu^{-1}(m)\\
&=&\nu(n_{(0)})\o B^{-1}(\varepsilon(n_{(1)})1_{H})\c\mu^{-1}(m)\\
&=&\nu(n_{(0)})\o \varepsilon(n_{(1)})m
= n\o m
\end{eqnarray*}
The fact that $ c_{M,N}^{-1} c_{M,N}=id$ is similar.
This completes the proof. \hfill $\blacksquare$
\\
Let $H$ be a monoidal Hom-Hopf algebra and
$G={\sl Aut}_{mHH}(H)
\times {\sl Aut}_{mHH}(H)$.
Define $\mathcal {MHYD}(H)$ as the
disjoint union of all $_{H}\mathcal {MHYD}^{H}(A,B)$
with $(A,B)\in G$. If we endow $\mathcal {MHYD}(H)$
with tensor product shown in Proposition 3.1,
then $\mathcal {MHYD}(H)$ becomes
a monoidal category with unit $k$.
Define a group homomorphism
$\,\,\varphi: G\rightarrow Aut(\mathcal {MHYD}(H)),
\,\,\,\,\,\,\,\,
(A, B) \,\,\mapsto \,\,\varphi(A,B)\,\,$
on components as follows:
\begin{eqnarray*}
\varphi_{(A,B)}: {_{H}}\mathcal {MHYD}^{H}(C,D)&\rightarrow&
{_{H}}\mathcal {MHYD}^{H}((A,B)\ast
(C,D)\ast (A,B)^{-1}),\\
\quad \quad \quad \quad \quad
\quad \varphi_{(A,B)}(N)&=& {}^{(A,B)} N,
\end{eqnarray*}
and the functor $\varphi_{(A,B)}$ acts as identity on morphisms.
The braiding in $\mathcal {MHYD}(H)$
is given by the family $\{c_{M,N}\}$
in Proposition 3.5.
So we get the following main theorem of this article.
\\
{\bf Theorem 3.7.} $\mathcal {MHYD}(H)$
is a braided $T$-category over G.
\\
We will end this paper by giving some
examples to illustrate the main theorem.
{\bf Example 3.8.}
For $\alpha=id_{H},$ we have
$\mathcal {MHYD}(H)=\mathcal {YD}(H)$, the main constructions
by Panaite and Staic \cite{PS2007}.
\\
{\bf Example 3.9.} Let $( H_{4} =k\{ 1,\ g,\ x, \ gx\,\},\a , \Delta , \varepsilon , S )$
be the monoidal Hom-Hopf algebra (see Example 2.5 (2)).
Let $\mathcal {MHYD}(H_4)$ be the disjoint union of all
categories $_{H_4}\mathcal {MHYD}^{H_4}(A,B)$ of
left-right Hom-$(A,B)$-Yetter-Drinfeld modules
with ${\sl Aut}_{mHH}(H_4) \times {\sl Aut}_{mHH}(H_4).$
Then by Example 2.5, Proposition 3.3,
Proposition 3.5 and Theorem 3.7,
$\mathcal {MHYD}(H_4)$ is a new braided $T$-category
over ${\sl Aut}_{mHH}(H_4) \times {\sl Aut}_{mHH}(H_4).$
\\
Explicitly, it is easily know that we have a group isomorphism:
${\sl Aut}_{mHH}(H_{4})=\{\left(
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & \lambda & 0 \\
0 & 0 & 0 & \lambda \\
\end{array}
\right)| 0\neq\lambda\in k\}\cong \left(k\backslash \{0\},\times\right).$
Furthermore, one has:
${\sl Aut}_{mHH}(H_{4})\times {\sl Aut}_{mHH}(H_{4})
\cong \left(k\backslash \{0\}\oplus k \backslash \{0\},\times\right).$
\\
Let $H_{4A}\in\,\, _{H_4}\mathcal{MHYD}^{H_4}(A, id)$
and $H_{4B}\in\,\, _{H_4}\mathcal{MHYD}^{H_4}(id, B)$,
for all $A,B\in {\sl Aut}_{mHH}(H_4).$
Then $ (H_{4A}\otimes H_{4B},\alpha\otimes \alpha)
\in\,\, _{H_4}\mathcal{MHYD}^{H_4}(A, B)$
with structures as follows:
\begin{eqnarray*}
&&h\c (x\o y)= h_{1}\c x \o h_{2}\c y,
\end{eqnarray*}
for all $h\in H_4,\ x\in H_{4A},\ y\in H_{4B}.$
\\
If $A=\left(
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & c' & 0 \\
0 & 0 & 0 & c' \\
\end{array}
\right)$ and $H_{4 B}\in _{H_4}\mathcal {MHYD}^{H_4}(id,B),$
then ${}^{(A,id)}H_{4 B}=H_{4 B}$ as vector spaces, with action
$\rhd$ given by
\begin{eqnarray*}
&&1\rhd 1=1,\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\ \ \ \ 1\rhd g = g,
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ 1\rhd x=cx,
\ \ \ \ \ \ 1\rhd gx = cgx\\
&&g\rhd 1 =1, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\ \ \ \ \ \ \ g\rhd g=g,\ \ \ \ \
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ g\rhd x=-cx,\ \ \ \ g\rhd gx=-cgx\\
&& x\rhd 1=-c(1+c'')/c'gx,
\ \ \ x\rhd g=c(-1+c'')/c'x,
\ \ \ \ x\rhd x=0,
\ \ \ \ \ \ \ \ x\rhd gx=0,\\
&& gx\rhd 1=c(-1+c'')/c'gx,\
gx\rhd g=-c(1+c'')/c'x,\ \ \
gx\rhd x=0,\ \ \ \ \ \ \ x\rhd gx=0,
\end{eqnarray*}
and coaction $\rho_{r}$ defined by
\begin{eqnarray*}
&&\rho_{r}(1)=1\otimes1, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \rho_{r}(g)=g\otimes g,\\
&&\rho_{r}(x)=c^{-1}(x\otimes 1)+ c^{-1}c'(g\otimes x),
\ \ \ \ \ \ \rho_{r}(gx)=c^{-1}(gx\otimes g)+c^{-1}c'(1\otimes gx),
\end{eqnarray*}
for all $c,c',c''\in k\backslash \{0\}$,
and ${}^{(A,id)}H_{4 B}\in _{H_4}\mathcal {MHYD}^{H_4}(id,B).$
\\
Let $H_{4 A}\in _{H_4}\mathcal {MHYD}^{H_4}(A,id)$
and $H_{4 B}\in _{H_4}\mathcal {MHYD}^{H_4}(id,B).$
Then the braiding
$$
c_{H_{4 A},H_{4 B}}:\
H_{4 A}\otimes H_{4 B}\rightarrow {}^{(A,id)}H_{4 B}
\otimes H_{4 A}
$$
is given by
$$
c_{H_{4 A},H_{4 B}}(m\otimes n)
=\a(n_{1})\otimes n_{2}\cdot \mu^{-1}(m),
$$
for all $m\in H_{4 A}, n\in H_{4 B}.$
\\
If we consider a system of the basis of $H_{4 A}\otimes H_{4 B}$:
$\{1\otimes 1, 1\otimes g ,
1\otimes x, 1\otimes gx ,
g\otimes 1 , g\otimes g ,
g\otimes x , g\otimes gx ,
x\otimes 1, x\otimes g ,
x\otimes x, x\otimes gx ,
gx\otimes 1 , gx\otimes g ,
gx\otimes x, gx\otimes gx \}
$
which is denoted by $\{e_1,e_2, \cdots , e_{16}\}$, then the braiding $c_{H_{4 A},H_{4 B}}$
can be represented by the following matrix:
\\
$D=\left(
\begin{array}{cccccccccccccccc}
1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & m & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & n& 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 \\
0& 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & n & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & m & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & -1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0& 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & -1 & 0 \\
0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & -1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & -1 \\
\end{array}
\right)$\\
i.e. we have
$$
c_{H_{4 A},H_{4 B}}((e_1,e_2,......e_{16})^{T})
=D(e_1,e_2,......e_{16})^{T},
$$
for all $m\neq -1,n\neq 1,$
$m,n\in k.$
Then by Theorem 3.7,
$\mathcal {MHYD}(H_4)$ is a new Braided $T$-category
over $\left(k/\{0\}\oplus k/\{0\},\times\right).$
\\
\section*{ACKNOWLEDGEMENTS}
The authors would like to thank the referee for the valuable comments
even for our English.
This work was supported by the NSF of China (No. 11371088) and the NSF of Jiangsu Province (No. BK2012736).
|
\section{Introduction}
\label{sec:Intro}
Densely deploying small cells is one of the major techniques to address the scarcity of spectrum resources for future cellular networks~\cite{CavAgr05MWC}. By reducing the coverage of each micro/pico base terminal stations (BTS), the capacity of a network can be significantly increased. With overlapping cells of all sizes, a heterogenous network (HetNet) almost always operates in the interference limited regime. Small cells may also lead to more pronounced traffic variations. Hence traditional frequency reuse and cell planning can be inefficient. Resource allocation according to traffic and demand conditions becomes highly desirable.
One conventional spectrum allocation scheme is fractional frequency reuse (FFR). Dynamic FFR can improve the sum-rate and total throughput of a network through effectively mitigating inter-cell interference~\cite{StoVis08INFOCOM,ChaTao09ICC,AliLeu09TransWC}. However these results are based on the assumption that traffic is backlogged at all BTS's. In a dense HetNet with dynamic traffic, physical layer performance metrics are not good indicators of user QoS, especially the delay experienced by a user. In this work, we focus on improving the average packet delay. With dynamic traffic, each BTS alternates between service and vacation periods, which introduces complicated inter-cell interference. A recent study in~\cite{DhiGan2012arXiv} points out that the backlogged traffic assumption, i.e., `always transmitting/interfering', will exaggerate the contribution of each BTS to inter-cell interference. Hence we propose a traffic aware interference model to address this issue. The optimization problems in~\cite{StoVis08INFOCOM,ChaTao09ICC,AliLeu09TransWC} are considered on fast timescales with instantaneous information exchange. Since traffic variation happens on a slower timescale in practice (compared to the timescale of scheduling) and there are limitations on the frequency with which a central controller can acquire traffic and service information from all BTS's, the spectrum allocation problem in this paper is solved by a central controller on a slow timescale as in~\cite{MadBor10JSAC,ZhuGuo2012Allerton} with aggregate traffic and service information.
In this work, we study resource allocation in dense deployed HetNets with traffic variations. The objective is to improve user QoS through spectrum resource adaptation on a slow timescale, as the long-term spatial load changes. The key is to link the spectrum and power resources in the physical layer to the QoS in the network layer. To do this, we use the service rates of the queues at all BTS's as this link. To be specific, a given spectrum allocation across cells induces an interference pattern and corresponding spectral efficiency on each segment of the spectrum. The spectral efficiencies along with the widths of the different segments of spectrum determine the service rate of each BTS, according to Shannon's capacity formula. The average packet delay at a given BTS is then determined by the service rate and the packet arrival rate. Thus an optimization problem can be formulated with the QoS as the objective, the spectrum allocation as desired variable, and the service rates as intermediate variables. Hence we consider the resource allocation problem as a joint physical layer and network layer optimization problem integrating results from both information theory and queueing theory.
The queueing scenario in our proposed system can be modeled by interactive queues. In general, it is difficult to get a closed-form steady state distribution of interactive queues. We use a conservative upper bound as an approximation of the average packet sojourn time. This upper bound is obtained by serving each queue independently at a conservative rate, which can be sustained regardless of the state of the other queues~\cite{BonBor2004ACM}. The spectrum allocation problem is formulated as a convex optimization problem, using the conservative upper bound. The geometric nature of the problem induces a special structure of the optimal solution. An efficient algorithm is provided to solve the optimization problem. Because of the use of conservative upper bound as objective, the optimum can be regarded as a QoS guarantee. According to the numerical results, the optimal solution greatly reduces the average packet delay in the heavy traffic regime compared to both orthogonal and full spectrum reuse allocations. The performance gain is observed mainly because each BTS is driven to allocate the right amount of spectrum to serve its own traffic demands, while leaving enough lightly loaded spectrum for adjacent BTS's.
\section{System Model}
\label{sec:SysMod}
We consider downlink data transmission in a HetNet with $K$ randomly deployed BTS's. Denote the set of all BTS's as $S=\{1,\dots,K\}$. The $K$ BTS's collectively share a unit bandwidth to serve their respective user equipments (UEs). A central controller determines which part of the spectrum is allocated to each BTS. Assuming the frequency resources are interchangeable, the problem is equivalent to deciding the bandwidth shared by each subset of BTS's, denoted by a $2^K$-tuple: $\boldsymbol{x}=\left(x(B)\right)_{B\in 2^S}$, where $2^S=\{B | B\subseteq S\}$ is the power set of $S$ containing all combinations of BTS's (including the empty one), and $x(B)\in[0,1]$ is the fraction of spectrum shared by BTS's in set $B$. Clearly, $\sum_{B\in2^S}x(B)=1$, and any efficient allocation would set $x(\emptyset)=0$. For example, if $K=2$, then there are basically three variables to decide on, $x(\{1\})$, $x(\{2\})$ and $x(\{1,2\})$, which denote the amount of spectrum allocated to BTS 1, BTS 2 exclusively and that shared by the two BTS's.
We next specify the relationship between spectrum allocation and the service rate at each BTS. The actual service rate in each cell depends on its own spectrum usage as well as the interference from other cells. To characterize the interaction among multiple BTS's, we define $A(t)$ as the set of BTS's that are transmitting data to their UEs at time $t$. Since we are interested in the average performance over a slow timescale, the time index $t$ will be omitted. The spectral efficiency of BTS $i$ on a segment of spectrum denoted by $s_i(C)$ is a function of the set $C$, which contains the BTS's actively sharing it. The service rate in cell $i$ when the active set is $\mathcal{A}$ is given by:
\begin{equation}
\label{eq:ServiceRate}
r_i(A)=\sum_{B\in 2^S}s_i(B\cap A)x(B).
\end{equation}
The intersection of sets in~\eqref{eq:ServiceRate} is because among all BTS's in $B$, only those in $B\cap A$ are transmitting. For concreteness, let the spectral efficiency of BTS $i$ with active set $C$ be calculated as:
\begin{equation}
\label{eq:SpeEff}
s_i(C)=\mathbb{1}(i\in C)\log\left(1+\frac{p_i l_{i}}{I_{i}(C)+n_i}\right),
\end{equation}
where $\mathbb{1}$ is the indicator funciton, $p_i$ is the constant transmit power spectral density (PSD), $l_i$ is the signal attenuation from BTS $i$ to its UEs, which includes both pathloss and slow-time fading, $I_{i}(C)$ is the constant interference PSD when BTS's in $C$ are generating interference, and $n_i$ is the noise PSD. The actual values of $s_i(C)$'s depend on transmit PSD at each BTS, the path-loss model and network topology. The receivers of a cell are assumed to be at the same point to simplify the interference model. The model, however, can be refined to address the locations of UEs by considering finer classification of UEs within each cell. Since the optimization will be performed on a slow timescale, this information, i.e., $s_i(C)$'s are assumed known by the central controller \textit{a priori}.
User service requests are modeled as packet arrivals at each BTS following a Poisson process with rate $\lambda_i$ at BTS $i$. All packet lengths are \textit{i.i.d.}\ exponentially distributed with unit mean. The objective is to minimize the average mean packet sojourn time through spectrum allocation, i.e., optimizing $\{x(B),~B\in 2^S\}$ for given $\lambda_i,~s_i(C),~\forall i\in S,~\forall C\in 2^S$. To evaluate the flow-level performance, we assume user requests within a cell are processed according to a `first come first serve' criterion.
The $K$-BTS network described above is a network of $K$ interactive queues, where the service rate of each queue depends on the status of the other queues at the same time. Such an interactive queueing system is also referred to as a \emph{coupled-processors} model. In the special case of two coupled queues, finding the steady-state joint distribution can be formulated as a Riemann-Hilbert problem~\cite{FayIas1979Springer}. Two coupled processors with generally distributed service times have been studied in~\cite{CohBox2000Elsevier}, which shows the joint workload distribution is the solution to a boundary value problem. These results are difficult to use for numerical computation. Also, few results exist for more than two coupled queues. A numerical method for solving the average packet delay using semidefinite program has been proposed in~\cite{RenCar2008}, which is again difficult to incorporate in our optimization. Here we use an upper bound on the true delay as the objective in the proposed optimization problem. The upper bound is achieved by decoupling the interactions among BTS's, which can be written in a simple closed form. Although the approximation is pessimistic, the effect on packet delay due to load variation is preserved.
\section{The Spectrum Allocation Problem}
\label{sec:ProbForm}
In this section, we introduce the upper bound of the mean sojourn time. The key is to assume each BTS always transmits at the worst-case rate, which is achievable regardless of other BTSs' states. This assumption is equivalent to assuming other BTS's are always backlogged and causing interference. Hence each BTS $i$ serves its users with constant rate $r_i(S)$, which is given by~\eqref{eq:ServiceRate} with $A=S$. From now on, we will use $r_i$ to denote $r_i(S)$ for simplicity. Therefore the $K$ interactive queues become $K$ independent $M/M/1$ queues with arrival rate $\lambda_i$ and service rate $r_i$ at BTS $i$. The mean packet sojourn time at BTS $i$ takes a simple form~\cite{Nel95Spinger}:
\begin{equation}
\label{eq:IndDelay}
t_i=\frac{1}{r_i-\lambda_i}.
\end{equation}
\subsection{Optimization Problem}
\label{subsec:Opt}
The spectrum allocation problem using the conservative approximation~\eqref{eq:IndDelay} is formulated as:
\begin{subequations}
\label{eq:SAP-Ind}
\begin{align}
\underset{\{x(B),~B\in 2^S\}}{
\text{minimize}}~&~ \frac{\lambda_i}{\sum_{j=1}^{k}\lambda_j}\sum_{i=1}^k \frac{1}{r_i-\lambda_i} \label{eq:Obj-Ind}\\
\text{subject to}~~&~r_i=\sum_{B\in 2^S}s_i(B)x(B),~\forall i\in S\label{eq:minServiceRate}\\
&~r_i>\lambda_i,~\forall i\in S\label{eq:Con2-Ind}\\
&~ x(B)\geq0,~\forall B\in 2^S\label{eq:Con1-Ind}\\
&~ \sum_{B\in 2^S}x(B)=1\label{eq:Con3-Ind}.
\end{align}
\end{subequations}
The objective~\eqref{eq:Obj-Ind} is the weighted average mean packet sojourn time of the entire network, where the weight $\frac{\lambda_i}{\sum_{j=1}^{k}\lambda_j}$ is the fraction of total traffic served by BTS $i$. Constraints in~\eqref{eq:Con2-Ind} are to guarantee the stability of the queues. The optimization problem~\eqref{eq:SAP-Ind} is a convex optimization problem. To see this, we can take $x(B)$'s and $r_i$'s as the optimization variables. Thus all the constrains are linear, and the objective is a linear combination of convex functions as $t(r)=1/(r-\lambda)$ is convex in $r$ on $(\lambda,\infty)$.
The convex optimization problem~\eqref{eq:SAP-Ind} has a unique globally optimal solution~\cite{BerDim1999nonlinear}. Due to the special geometric structure of~\eqref{eq:SAP-Ind}, the optimal solution has the following property.
\begin{theorem}
\label{thm:KBTS}
The optimal solution of the $K$-BTS spectrum allocation problem divides the spectrum into at most $K$ segments:
\begin{equation}
\nonumber
\left|\{B~|~x(B)>0,~B\in2^S\}\right|\leq K.
\end{equation}
\end{theorem}
\begin{IEEEproof}
\label{pf:KBTS}
Denote the rate vector $\rr$ and spectral efficiency vector of sharing combination $B$, $\ssv(\mathcalB)$, as $\rr=[r_1,\dots,r_K]$ and $\ssv(\mathcalB)=[s_1(\mathcalB), \dots, s_K(\mathcalB)]$. According to~\eqref{eq:minServiceRate} to \eqref{eq:Con3-Ind}, $\rr\in\mathds{R}_+^K$ is a convex combination of the $2^K-1$ points $\{\ssv(\mathcalB)~|~\mathcalB\in2^\mathcalK\}$, with coefficients $\{x(\mathcalB)~|~\mathcalB\in2^\mathcalK\}$, i.e., $\rr=\sum_{\mathcalB\in2^\mathcalK}\ssv(\mathcalB)x(\mathcalB)$. In other words, any $\rr$ given by~\eqref{eq:minServiceRate} is in the convex hull of $\{\ssv(\mathcalB)~|~\mathcalB\in2^\mathcalK\}$. By Carath\'eodory's Theorem~\cite{Egg69Convexity}, $\rr$ lies in a $d$-simplex with vertices in $\{\ssv(\mathcalB)~|~\mathcalB\in2^\mathcalK\}$ and $d\leq K$, i.e., $\rr$ can be written as a convex combination of $x(\mathcalB)$'s with at most $K+1$ nonzero $x(\mathcalB)$'s. This holds for any $\rr$ satisfying~\eqref{eq:minServiceRate} to \eqref{eq:Con3-Ind}. Furthermore, the $\rr^*$ corresponding to the optimal solution to~\eqref{eq:SAP-Ind} must be Pareto optimal in terms of the rate allocation, i.e., there is no spectrum allocation $\boldsymbol{x}$ such that $r_i^*\leq \sum_{\mathcalB\subseteq\mathcalK} s(\mathcalB)x(\mathcalB),~\forall i\in\mathcalK$ and at least one inequality is strict. This is because any spectrum allocation that could increase the rate at any BTS without decreasing the rates at other BTS's would also decrease the objective~\eqref{eq:Obj-Ind}. Hence $\rr^*$ cannot be an interior point of the $d$-simplex, and must lie on some $m$-face of the $d$-simplex with $m<d\leq K$. Therefore $\rr^*$ can be written as a convex combination of $m+1\leq K$ nonzero $x(\mathcalB)$'s.
\end{IEEEproof}
The geometric interpretation can be generalized to any subset of the $K$ BTS's.
\begin{corollary}\label{cor:Sub}
The optimal solution of the $K$-BTS spectrum allocation problem divides the spectrum exclusively used by any subset $\mathcalM\subseteq\mathcalK$ of the $K$ BTS's with $m$ BTS's into at most $m$ segments:
\begin{equation}
\nonumber
\label{eq:Cor}
\left|\{\mathcalB~|~x(\mathcalB)>0,~\mathcalB\in2^\mathcalM\}\right|\leq m.
\end{equation}
\end{corollary}
\begin{IEEEproof}
\label{pf:sub}
If the bandwidths of the spectrum segments not in $2^\mathcalM$ are fixed at their optimal values, then~\eqref{eq:SAP-Ind} becomes an optimization problem of variables $x(\mathcalB),~\mathcalB\in2^\mathcalM$, the service rates at BTS's not in $\mathcalM$ are fixed and the service rates for the $m$ BTS's in $\mathcalM$ are convex combinations of $x(\mathcalB),~\mathcalB\in2^\mathcalM$ plus a constant vector in $\mathds{R}_+^m $. Therefore Corollary~\ref{cor:Sub} can be proved using a similar argument as in the proof of Theorem~\ref{thm:KBTS}.
\end{IEEEproof}
\subsection{An Efficient Algorithm}
\label{subsec:EA}
The structure of the optimal solution given by Theorem~\ref{thm:KBTS} suggests the possibility of solving~\eqref{eq:SAP-Ind} more efficiently in real systems. Originally we needed to determine the sizes of $2^K-1$ spectrum segments. The computational complexity is exponential in $K$ using a standard convex optimization algorithm. By Theorem~\ref{thm:KBTS}, we only need to decide the sizes of the $K$ nonzero segments. The difficulty is that we do not know which $K$ pieces out of the $2^K-1$ possibilities. Algorithm~\ref{alg:SAP-Ind} is an iterative algorithm to efficiently solve the $K$-BTS spectrum allocation problem.
\begin{algorithm}
\caption{Iterative algorithm for solving problem~\eqref{eq:SAP-Ind}}
\label{alg:SAP-Ind}
\begin{algorithmic}[]
\label{alg:SAP-Ind}
\INPUT {$\lambda_i$ and $s_i(C)$ for all $i\in \mathcalK$ and $C\in2^\mathcalK$.}
\OUTPUT{$x(\mathcalB)$ for all $\mathcalB\in2^\mathcalK$}.
\Init{Find a feasible solution $x_0(\mathcalB)$ by solving~\eqref{eq:SAP-Ind} with constant objective.
$\Nc = \{\mathcalB~|~x_0(\mathcalB)>0\}$, $\Nc'=\emptyset$, $\Nc^+ = 2^\mathcalK$.}
\While{$\Nc^+\not\subseteq \Nc'$}
\State 1. $\Nc'=\Nc$;
\State 2. Find $x(\mathcalB)$ by solving~\eqref{eq:SAP-Ind} starting from $x_0(\mathcalB)$ with the additional constraints, $x(\mathcalB)=0,~\forall\mathcalB\notin \Nc$;
\State 3. Compute the partial derivatives of the objective function~\eqref{eq:Obj-Ind} with respect to all $x(\mathcalB)$'s, $\Delta_{x(\mathcalB)}=-\sum_{i\in\mathcalB}\frac{\lambda_is_i(\mathcalB)}{(r_i-\lambda_i)^2}$;
\State 4. $\Nc^+=\{\mathcalB~\text{for}~K~\text{smallest}~\Delta_{x(\mathcalB)}\}$, $\Nc=\Nc\cup \Nc^+$, $x_0(\mathcalB)=x(\mathcalB)$.
\EndWhile
\end{algorithmic}
\end{algorithm}
The algorithm requires to start at a feasible point. A standard method is to solve a modified version of optimization problem~\eqref{eq:SAP-Ind} by replacing the objective function~\eqref{eq:Obj-Ind} with a constant. The resulting optimization problem can be transformed to a linear program in standard form, which can be solved using the simplex method~\cite{BerDim97LP}.
Starting from a feasible point $\left(x_0(\mathcalB)\right)_{\mathcalB\in2^\mathcalK}$, let $\Nc$ be the candidate set, which initially includes the indices of those nonzero spectrum segments of the initial point. In each iteration, the algorithm finds the optimal solution within the candidate set $\Nc$. After each iteration, the partial derivatives with respect to all $x(\mathcalB)$'s are calculated (including those not in $\Nc$). The $K$ segments with the $K$ smallest derivatives are added to the candidate set. (The number of variables added to the candidate set may be less than $K$, since some of the $\mathcalB$'s of the $K$ smallest derivatives may already be in $\Nc$.) The algorithm continues with the solution found in the last iteration as the new initial point and the expanded candidate set, until the candidate set stops growing. At the end of each iteration, if the solution is not optimal, there must be some $\mathcalB$'s outside the candidate set that have smaller partial derivatives. Since we only add more $\mathcalB$'s to the candidate set with each iteration, in the worst-case, the candidate set will eventually include all $2^K-1$ variables. Hence the proposed algorithm is guaranteed to converge to the global optimum.
The algorithm is more efficient when starting at an initial point with fewer nonzero spectrum segments. Therefore we can always use the full-spectrum-reuse allocation, $x(\mathcalK)=1$, as an initial point if it is feasible. Even if it is not, the solution obtained by the initialization method has no more than $K+1$ nonzero spectrum segments according to the properties of basic feasible solution to a linear program~\cite{BerDim97LP}. Examples of delay versus number of iterations are shown in Fig.~\ref{fig:Alg} with $K=7$ and different traffic loads.
\begin{figure}
\centering
\includegraphics[width=2.8in]{EA}
\caption{Approximated average
delay versus number of iterations using Algorithm~\ref{alg:SAP-Ind} with different average packet arrival rates.}
\label{fig:Alg}
\end{figure}
In the simulation, Algorithm 1 starts with the full-spectrum-reuse allocation. The figure shows that the algorithm converges within a few iterations.
\section{Numerical Results}
\label{sec:Sim}
\subsection{Performance of the Spectrum Allocation Method}
\label{subsec:Sim}
In the simulation, we adopt the quantized HetNet model in~\cite{ZhuGuo2012Allerton}. A $100\times100~\text{m}^2$ area is quantized by hexagons with distance between the centers of adjacent hexagons being $20~\text{m}$. Seven BTS's are uniformly randomly dropped at the vertices of the hexagons. UE locations within each hexagon are approximated by the center of the hexagon. UEs are assigned to their respectively nearest BTS's. If there is a tie, UEs in the hexagon are equally distributed to the nearest BTS's. We only consider path-loss in this simulation, although slow fading can be easily incorporated. The average spectral efficiency of BTS $i$ in~\eqref{eq:SpeEff} is calculated as the mean of the spectral efficiencies in the hexagons served by BTS $i$. Other parameters used in the simulation include: path-loss exponent is 3; transmit PSD is 1 watt/Hz for all BTS's; noise PSD is 0.125 micro-watt/Hz.
\begin{figure}
\centering
\includegraphics[width=2.8in]{AprxSimu}
\caption{Comparison of average packet sojourn time versus average packet arrival rate using the approximation and simulation.}
\label{fig:AprxSimu}
\end{figure}
The approximate mean sojourn time in~\eqref{eq:Obj-Ind} and the real mean sojourn time obtained by simulating the interactive queues using the uniformization method are shown in Fig.~\ref{fig:AprxSimu} for different traffic loads. Both the approximated delay and the real delay have similar trends as traffic increases.
\begin{figure}
\centering
\includegraphics[width=2.8in]{Light}
\caption{Simulated average packet sojourn time using orthogonal spectrum allocation, full spectrum reuse allocation and the optimal solution of the spectrum allocation problem in light traffic.}
\label{fig:Light}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=2.8in]{Heavy}
\caption{Simulated average packet sojourn time using orthogonal spectrum allocation, full spectrum reuse allocation and the optimal solution of the spectrum allocation problem in heavy traffic.}
\label{fig:Heavy}
\end{figure}
The optimal orthogonal spectrum allocation and the full spectrum reuse allocation are compared with the optimal solution in Fig.~\ref{fig:Light} and Fig.~\ref{fig:Heavy}.
Fig.~\ref{fig:Light} shows the comparison in the light traffic regime. The average packet delay given by the optimal solution is between those given by the optimal orthogonal spectrum allocation and the full spectrum reuse allocation. Since traffic load is light, BTS's will have a larger fraction of time being empty. Hence the worst case transmit rate assumption exaggerates the harm of inter-cell interference. Due to rare concurrent transmissions among the BTS's, the full spectrum reuse allocation turns out to be more efficient. However, as the average packet arrival rate increases, the optimal solution of~\eqref{eq:SAP-Ind} outperforms the other two spectrum allocation schemes as shown in Fig.~\ref{fig:Heavy}. Due to increasing impact of inter-cell interference, the optimal orthogonal spectrum allocation becomes more efficient than the full spectrum reuse allocation as well. However, after the average packet arrival rate reaches 27 packets/sec, there is no orthogonal spectrum allocation that maintains the stability of the system. On the other hand, the optimal solution of~\eqref{eq:SAP-Ind} remains stable. Since the optimization problem exhaust all spectrum sharing possibilities, traffic and topology driven spectrum reuse is realized in the heavy traffic regime.
\begin{figure}
\centering
\includegraphics[width=2.8in]{SpecAloc}
\caption{Optimal solution of the spectrum allocation problem for different average packet arrival rates.}
\label{fig:OptSol}
\end{figure}
The optimal solutions for different average packet arrival rates are shown in Fig.~\ref{fig:OptSol}. Each subplot in the figure is the spectrum partition for a specific average packet arrival rate, with only the nonzero segments. Each row of a subplot is the spectrum usage of the corresponding BTS. Each BTS only transmits on the shaded pieces of spectrum in its row. By counting the number of pieces in the partition, we can verify Theorem~\ref{thm:KBTS}. Note that in the light traffic regime, the optimal solution still orthogonalizes the spectrum usage among BTS's to some extent, which is why the optimal solution is worse than the full spectrum reuse allocation in this regime. (This is due to the conservative approximation of the true objective.)
\subsection{Power Control}
\label{subsec:Power}
\begin{figure}
\centering
\includegraphics[width=2.8in]{DelayIteration}
\caption{Average packet sojourn time versus number of iterations.}
\label{fig:DelayIteration}
\end{figure}
All the discussions and results so far are based on the assumption that the spectral efficiencies under different partitions are fixed. This assumption helps us simplify the relationship between the spectrum allocation and the service rates. In fact, the service rates are linear functions of the spectrum variables. The assumption is true if all BTS's transmit with fixed power spectral density. However, in practice, we may have a fixed total transmit power constraint at each BTS instead. Hence power control becomes an issue. For example, the spectral efficiencies should be higher for the orthogonal allocation, since each BTS could concentrate all its power on the piece of spectrum exclusively used by itself.
The joint power control and spectrum allocation problem is in general more complicated. Hence we provide a simple solution using alternative spectrum and power updates. We start with fixed $s_i(C)$ assuming each BTS uniformly allocates its maximum transmit power across the whole spectrum. We iterate between the following two updates:
\begin{itemize}
\item[1.] Update the spectrum allocation $x(B),~\forall \mathcal{B}\in2^S$ with the current $s_i(C),~\forall i\in S,~\forall C\in2^S$ by solving the spectrum allocation problem.
\item[2.] Update the spectral efficiencies $s_i(\mathcal{C}),~\forall i\in S,~\forall C\in 2^S$ with the current $x(B),~\forall B\in 2^S$ by letting each BTS $i$ uniformly allocate its maximum transmit power over the spectrum it uses, which includes $x(B)$'s, with $i\in B$.
\end{itemize}
The iteration terminates unti $x(B),~\forall B\in 2^S$ converges (This is not guaranteed). The average packet sojourn time after the spectrum allocation update and the spectral efficiency update at each iteration is shown in Fig.~\ref{fig:DelayIteration} for an average packet arrival rate per BTS of 24 packets/second. The figure shows that the delay performance converges very quickly. The mean sojourn times decrease substantially after the first spectral efficiency update. This is because at this average packet arrival rate both allocations orthogonalize the spectrum use among neighboring BTS's to some extent. This kind of convergence behavior can be expected in general, since spatial reuse will occur in the optimal solution. With each BTS using a fairly large amount of the spectrum, the spectral efficiencies will not change much after several iterations.
Although all BTS's have the same transmit PSD in the simulations, the proposed spectrum allocation problem and algorithm can be directly applied to HetNets with arbitrary power, topology and traffic conditions.
\section{Conclusion}
\label{sec:Con}
We have formulated a joint physical layer and network layer optimization problem to minimize average delay in HetNets using the combination of information theory and queueing theory. The optimization problem takes traffic arrival rates and spectral efficiencies as the input and gives the spectrum partition as the output. Numerical results obtained by simulating the interactive queues suggest the optimal solution of the proposed spectrum allocation achieves a significant delay performance gain in the heavy traffic regime, compared to both orthogonal and full spectrum reuse allocations. For future work, we plan to look for a more accurate approximation of the average packet sojourn time in order to improve the performance in the light and moderate traffic regimes, since the conservative approximation presented has ignored the impact of traffic variations on inter-cell interference.
\bibliographystyle{IEEEtran}
|
\section{Introduction} \label{sec:intro}
When a parser receives an erroneous input,
it should indicate the existence of syntax errors.
However, just informing the user that an error
was encountered (e.g. with a {\tt syntax error} message)
does not help finding and fixing those errors.
Therefore, a parser should
indicate at least the position in the
input where it discovered the error, and ideally
some information about its context.
The LL and LR methods detect syntax errors very efficiently
because they have the \emph{viable prefix} property, that is,
these methods detect a syntax error as soon as a token is read
and cannot be used to form a viable prefix of the language
\cite{aho2006cpt}. LL and LR parsers can use this
property to produce suitable, though generic, error
messages.
A parser has two basic strategies for handling errors:
error reporting and error recovery.
In error reporting, the parser aborts with an informative
message when the first error is found.
In error recovery, the parser is adapted to not abort on the
first error, but to try to process the rest of the input,
collecting and reporting all errors that it finds.
Parsing Expression Grammars (PEGs)~\cite{ford2004peg} are a
formalism for describing the syntax of programming languages.
We can view a PEG as a formal description of a top-down parser
for the language it describes. PEGs have a concrete syntax based on
the syntax of {\em regexes}, or extended regular expressions.
Unlike Context-Free Grammars (CFGs),
PEGs avoid ambiguities in the definition of the grammar's
language due to the use of an {\em ordered choice} operator.
More specifically, a PEG can be interpreted as a the specification
of a recursive
descent parser with restricted (or local) backtracking.
This means that the alternatives of a choice are tried in
order; when the first alternative recognizes an input prefix,
no other alternative of this choice is tried, but when an
alternative fails to recognize an input prefix, the parser
backtracks on the input to try the next alternative.
On the one hand, PEGs are an expressive formalism for describing
top-down parsers~\cite{ford2004peg}; on the other hand,
PEGs cannot use error handling techniques that are often applied
to predictive top-down parsers, because these techniques assume the parser
reads the input without backtracking~\cite{ford2002packrat}.
In top-down parsers without backtracking, it is possible to
signal a syntax error when there is no alternative to continue
reading. In PEGs, it is more complicated to identify the cause
of an error and the position where it happened, because failures
during parsing are not necessarily errors, but just an
indication that the parser should backtrack and try a
different alternative.
Ford~\cite{ford2002packrat} has already identified this
limitation of error reporting in PEGs, and, in his
parser generators for PEGs, included an heuristic for
better error reporting. This heuristic simulates the error
reporting technique that is implemented in top-down parsers
without backtracking. The idea is to track the position
in the input where the farthest failure occurred, as
well as context that tracks what the parser was expecting
at that point, and report this to the user in case of errors.
In this paper, we show how grammar writers can use
this error reporting technique even in PEG implementations
that do not implement it, as long as they have semantic
actions that expose the current position in the input.
We also formalize two versions of the technique,
extending the semantics of PEGs to track failure
positions and to build context information for error
messages.
Tracking the farthest failure position and context
gives us PEGs that can produce error messages
similar to the automatically produced error
messages of other top-down parsers; they tell
the user the position where the error was encountered,
what was found in the input at that position,
and what the parser was expecting to find.
We also propose a complementary approach for error
reporting in PEGs, based on the concept of \emph{labeled
failures}, inspired by the standard exception handling
mechanisms of programming languages. Instead of just
failing, a labeled PEG can produce different kinds of failure
labels using a {\em throw} operator. Each label can be
tied to a more specific error message. PEGs can also
{\em catch} such labeled failures, via a change to
the ordered choice operator. We formalize labeled
failures as an extension of the semantics of regular
PEGs.
With labeled PEGs, we can express some alternative error
reporting techniques from parsers expressed in terms of
parser combinators that, like PEGs, also rely on top-down parsing
with local backtracking. We can also encode
predictive parsing in a PEG, as long as the decision
procedure can be expressed in the PEG formalism, and show
how to do that for $LL(*)$ parsing, a powerful predictive
parsing strategy.
The rest of this paper is organized as follows:
Section~\ref{sec:pegs} contextualizes the problem of error
handling in PEGs, explains in detail the failure tracking
heuristic, and shows how it can be realized in PEG implementations
that do not support it directly; Section~\ref{sec:rel}
discusses related work on error reporting for top-down
parsers with backtracking;
Section~\ref{sec:semfarther} formalizes two
versions of the failure tracking heuristic, and discusses
some of its issues. Section~\ref{sec:lf}
introduces and formalizes the concept of labeled failures,
and shows how to use it for error reporting.
Section~\ref{sec:labelsfft} compares the error messages
generated by a parser based on the failure tracking heuristic
with the ones generated by a parser based on labeled failures.
Section~\ref{sec:labelsrelated} shows how labeled failures
can encode some of the techniques of Section~\ref{sec:rel},
as well as predictive parsing.
Finally, Section~\ref{sec:conc} gives some concluding
remarks.
\section{Handling Syntax Errors with PEGs}
\label{sec:pegs}
\begin{figure}[t]
\begin{align*}
\it{Tiny} & \leftarrow \it{CmdSeq}\\
\it{CmdSeq} & \leftarrow \it{(Cmd \; {\tt SEMICOLON}) \;
(Cmd \; {\tt SEMICOLON})*}\\
\it{Cmd} & \leftarrow \it{IfCmd \; / \; RepeatCmd \; / \;
AssignCmd \; / \; ReadCmd \; / \; WriteCmd}\\
\it{IfCmd} & \leftarrow \it{{\tt IF} \; Exp \; {\tt THEN} \; CmdSeq
\; ({\tt ELSE} \; CmdSeq \; / \; \varepsilon) \; {\tt END}}\\
\it{RepeatCmd} & \leftarrow \it{{\tt REPEAT} \; CmdSeq \;
{\tt UNTIL} \; Exp}\\
\it{AssignCmd} & \leftarrow \it{{\tt NAME} \; {\tt ASSIGNMENT} \; Exp}\\
\it{ReadCmd} & \leftarrow \it{{\tt READ} \; {\tt NAME}}\\
\it{WriteCmd} & \leftarrow \it{{\tt WRITE} \; Exp}\\
\it{Exp} & \leftarrow \it{SimpleExp \;
(({\tt LESS} \; / \; {\tt EQUAL}) \; SimpleExp \; / \; \varepsilon})\\
\it{SimpleExp} & \leftarrow \it{Term \;
(({\tt ADD} \; / \; {\tt SUB}) \; Term)*}\\
\it{Term} & \leftarrow \it{Factor \;
(({\tt MUL} \; / \; {\tt DIV}) \; Factor)*}\\
\it{Factor} & \leftarrow \it{{\tt OPENPAR} \; Exp \; {\tt CLOSEPAR}
\; / \; {\tt NUMBER} \; / \; {\tt NAME}}
\end{align*}
\caption{A PEG for the Tiny language}
\label{fig:tiny}
\end{figure}
In this section, we use examples to present in more detail how a
PEG behaves badly on the presence of syntax errors.
After that, we present a heuristic proposed by Ford
\cite{ford2002packrat} to implement error reporting in PEGs.
Rather than using the original definition of PEGs by Ford~
\cite{ford2004peg}, our examples use the equivalent and more
concise definition proposed by Medeiros et al.
\cite{medeiros2011re2peg,medeiros2012lrp,mascarenhas2014}.
We will extend this definition in later sections to present
semantics for PEGs with built-in error reporting and with
labeled failures.
A PEG $G$ is a tuple $(V,T,P,p_{S})$ where $V$ is a finite set of
non-terminals, $T$ is a finite set of terminals, $P$ is a total
function from non-terminals to \emph{parsing expressions} and $p_{S}$
is the initial parsing expression.
We describe the function $P$ as a set of rules of the form
$A \leftarrow p$, where $A \in V$ and $p$ is a parsing
expression.
A parsing expression, when applied to an input string, either
fails or consumes a prefix of the input resulting in the
remaining suffix. The abstract syntax of parsing expressions is
given as follows, where $a$ is a terminal, $A$ is a non-terminal,
and $p$, $p_1$ and $p_2$ are parsing expressions:
\[\it{
p = \varepsilon \; | \; a \; | \; A \; | \; p_1 p_2 \; | \;
p_1 / p_2 \; | \; p* \; | \; !p
}\]
Intuitively,
$\varepsilon$ successfully matches the empty string, not changing
the input;
$a$ matches and consumes itself or fails otherwise;
$A$ tries to match the expression $P(A)$;
$p_1 p_2$ tries to match $p_1$ followed by $p_2$;
$p_1 / p_2$ tries to match $p_1$;
if $p_1$ fails, then it tries to match $p_2$;
$p*$ repeatedly matches $p$ until $p$ fails, that is, it
consumes as much as it can from the input;
the matching of $!p$ succeeds if the input does not match $p$
and fails when the the input matches $p$, not consuming any input
in both cases;
we call it the negative predicate or the lookahead predicate.
The $\stackrel{\mbox{\tiny{PEG}}}{\leadsto}$ relation formalizes this semantics, relating a
parsing expression grammar, an input string, and a result
that is either {\tt fail}\, or a suffix of the input string.
Figure \ref{fig:tiny} presents a PEG for the Tiny language
\cite{louden1997ccp}.
Tiny is a simple programming language with a syntax that resembles
Pascal's.
We will use this PEG to show how error reporting differs between
top-down parsers without backtracking and PEGs.
PEGs usually express the language syntax down to the character level,
without the need of a separate lexer.
For instance, we can write the lexical rule \texttt{IF} as follows:
\[
{\tt IF} \leftarrow \it{if \; !IDRest \; Skip}
\]
That is, the rule matches the keyword \textit{if} provided that it
is not a prefix of an identifier and then the rule skips trailing
white spaces and comments.
The non-terminal \textit{IDRest} recognizes any character that may
be present on a proper suffix of an identifier while the non-terminal
\textit{Skip} recognizes white spaces and comments.
In the presented PEG, we omitted the lexical rules for brevity.
Now, we present an example of erroneous Tiny code so we can compare
approaches for error reporting.
The program in Figure~\ref{fig:tinyerror} has a missing semicolon
(\texttt{;}) in the assignment in line \texttt{5}.
A hand-written top-down parser without backtracking that aborts on
the first error presents an error message like this:
\begin{verbatim}
factorial.tiny:6:1: syntax error, unexpected 'until', expecting ';'
\end{verbatim}
The error is reported in line \texttt{6} because the parser cannot
complete a valid prefix of the language, since it unexpectedly
finds the token \texttt{until} when it was expecting a command
terminator (\texttt{;}).
\begin{figure}[t]
\begin{verbatim}
01 n := 5;
02 f := 1;
03 repeat
04 f := f * n;
05 n := n - 1
06 until (n < 1);
07 write f;
\end{verbatim}
\caption{Program for the Tiny Language with a Syntax Error}
\label{fig:tinyerror}
\end{figure}
In PEGs, we can try to report errors using the remaining suffix, but
this approach usually does not help the PEG to produce an error
message like the one shown above.
In general, when a PEG finishes parsing the input, a remaining non-empty suffix means that parsing did not reach the
end of file due to a syntax error.
However, this remaining suffix usually does not indicate the position
where the longest parse ends.
This problem occurs because the failure of a parsing expression does
not necessarily mean a failure of the whole parse.
Actually, the failure usually implies that the PEG should backtrack and
try a different alternative.
For this reason, the remaining suffix probably indicates a position
far away from the actual position where the first error happened in case
parsing finishes without consuming the complete input.
In our example, the problem happens when the PEG tries to recognize
the sequence of commands inside the \texttt{repeat} command.
Even though the program has a missing semicolon (\texttt{;}) in the
assignment in line \texttt{5}, making the PEG fail to recognize
the sequence of commands inside the \texttt{repeat} command, this
failure is not treated as an error.
Instead, this failure makes the recognition of the \texttt{repeat}
command also fail.
For this reason, the PEG backtracks the input to line \texttt{3}
to try other to parse other alternatives for {\em CmdSeq},
and since they do not exist, its ancestor {\em Cmd}, that exist in the language.
Since it is not possible to recognize a command other than
\texttt{repeat} at line \texttt{3},
the parsing finishes without consuming all the input.
Hence, if the PEG uses the remaining suffix to produce an error
message, the PEG reports line 3 instead of line 6 as the location of
the error.
There is no
perfect method to identify which information is the most relevant to
report an error. According to Ford \cite{ford2002packrat}, using the information of the farthest position
that the PEG reached in the input is a heuristic that provides
good results.
PEGs implement top-down parsers and try to recognize
the input from left to right, so the position farthest to
the right in the input that a PEG reaches during parsing usually is
close to the real error \cite{ford2002packrat}. The same idea
for error reporting in top-down parsings with backtracking was
also mentioned in Section 16.2 of~\cite{grune2010ptp}.
Ford used this heuristic to add error reporting to his packrat
parsers \cite{ford2002packrat}.
A packrat parser generated by Pappy \cite{ford2002pappy}, Ford's PEG
parser generator, tracks the farthest position and uses this position
to report an error.
In other words, this heuristic helps packrat parsers to simulate the
error reporting technique that is implemented in top-down parsers
without backtracking.
Although Ford only has discussed his heuristic in relation to packrat
parsers, we can use the farthest position heuristic to add error reporting
to any implementation of PEGs that provides semantic actions.
The idea is to annotate the grammar with semantic actions that track
this position.
For instance, in Leg \cite{leg}, a PEG parser generator with
Yacc-style semantic actions, we can annotate the rule \textit{CmdSeq}
as follows:
\begin{verbatim}
CmdSeq = Cmd (";" Skip | &{ updateffp() })
(Cmd (";" Skip | &{ updateffp() }))*
\end{verbatim}
A parsing expression of the form $\&p$ is a predicate that succeeds
when $p$ succeeds and fails when $p$ fails, without consuming the input.
The expression $\&p$ is a syntactic sugar for $!!p$.
The parser defined as above calls the function \texttt{updateffp} when
the matching of a semicolon fails.
The function \texttt{updateffp} is a semantic action that updates
the farthest failure position in a global variable if the current
parsing position is greater than the position that is stored in
this global.
After the update, the semantic action forces another failure
to not interrupt backtracking.
Since this semantic action propagates failures and runs only when
a parsing expression fails, we could annotate all terminals and
non-terminals in the grammar without changing the behavior of the PEG.
In practice, we just need to annotate terminals to implement error
reporting.
However, storing just the farthest failure position does not give
the parser all the information it needs to produce an informative
error message.
That is, the parser has the information about the position where
the error happened, but it lacks the information about what terminals
failed at that position.
Thus, we should include the name of the terminals in the annotations
so the parser can also track these names to compute the set of
expected terminals at a certain position.
Basically, we give an extra argument to each semantic action.
This extra argument is a hard-coded name for the terminal
that we want to keep track along with the farthest failure position.
For instance, now we annotate the \textit{CmdSeq} rule in Leg as
follows:
\begin{verbatim}
CmdSeq = Cmd (";" Skip | &{ updateffp(";") })
(Cmd (";" Skip | &{ updateffp(";") }))*
\end{verbatim}
We then extend the implementation of \texttt{updateffp} to keep,
for a given failure position, the names of all the symbols
expected there.
If the current position is greater than the farthest failure,
{\tt updateffp} initializes this set with just the given name.
If the current position equals the farthest failure, {\tt updateffp}
adds this name to the set.
Parsers generated by Pappy also track the set of expected terminals,
but with limitations.
The error messages include only symbols and keywords that were defined
in the grammar as literal strings.
That is, the error messages do not include terminals that were defined
through character classes.
The approach of naming terminals in the semantic actions avoids the
kind of limitation found in Pappy, though it increases the annotation
burden because the implementor of the PEG is also responsible for
adding one semantic action for each terminal and its respective name.
The annotation burden can be lessened in implementations of PEGs that
treat parsing expressions as first-class objects, as we are able to
define functions to annotate the lexical parts of the grammar to track
errors, record information about the expected terminals to produce
good error messages, and enforce lexical conventions such as the
presence of surrounding white spaces.
For instance, in LPeg \cite{lpeg,ierusalimschy2009lpeg}, a PEG
library for Lua that defines patterns as first-class objects,
we can annotate the rule \textit{CmdSeq} as follows,
where the patterns \verb'V"A"', \verb'p1 * p2', and \verb'p^0'
are respectively equivalent to parsing expressions
$A$, $p_1 p_2$, and $p*$:
\begin{verbatim}
CmdSeq = V"Cmd" * symb(";") * (V"Cmd" * symb(";"))^0;
\end{verbatim}
The function \texttt{symb} works like a parser combinator
\cite{hutton1992hfp}.
It receives a string as its only argument and returns a parser
that is equivalent to the parsing expression that we used in the
Leg example.
That is, \verb'symb(";")' is equivalent to
\verb'";" Skip | &{ updateffp(";") }'.
We implemented error tracking and reporting using semantic actions
as a set of parsing combinators on top of LPeg and used these
combinators to implement the PEG for Tiny.
It produces the following error message for the example we have
been using in this section:
\begin{verbatim}
factorial.tiny:6:1: syntax error, unexpected 'until',
expecting ';', '=', '<', '-', '+', '/', '*'
\end{verbatim}
We tested the PEG for Tiny with other erroneous inputs and in all
cases the PEG identified an error in the same place as a top-down
parser without backtracking.
In addition, the PEG for Tiny produced error messages that are
similar to the error messages produced by packrat parsers generated
by Pappy.
We annotated other grammars too and successfully obtained
similar results.
However, the error messages are still generic, they are not
as specific as the error messages of a hand-written top-down
parser.
\section{Error Reporting in Top-Down Parsers with Backtracking}
\label{sec:rel}
In this section, we discuss alternative approaches for error reporting
in top-down parsers with backtracking other than the heuristic
explained in Section \ref{sec:pegs}.
Mizushima et al. \cite{mizushima2010php} proposed a cut operator
($\uparrow$) to reduce the space consumption of packrat parsers;
the authors claimed that the cut operator can also be used to
implement error reporting in packrat parsers, but the authors did
not give any details on how the cut operator could be used for this
purpose.
The cut operator is borrowed from Prolog to annotate pieces of
a PEG where backtracking should be avoided.
PEGs' ordered choice works in a similar way to Prolog's green cuts,
that is, they limit backtracking to discard unnecessary solutions.
The cut proposed to PEGs is a way to implement Prolog's white cuts,
that is, they prevent backtracking to rules that will certainly fail.
The semantics of cut is similar to the semantics of an
\texttt{if-then-else} control structure and
can be simulated through predicates.
For instance, the PEG (with cut) $A \leftarrow B \uparrow C / D$ is
functionally equivalent to the PEG (without cut)
$A \leftarrow B C / !B D$ that is also functionally equivalent to
the rule $A \leftarrow B[C,D]$ on Generalized Top-Down Parsing
Language (GTDPL), one of the parsing techniques that influenced the
creation of PEGs \cite{ford2002packrat,ford2002pappy,ford2004peg}.
On the three cases, the expression $D$ is tried only if the expression
$B$ fails.
Nevertheless, this translated PEG still backtracks
whenever $B$ successfully matches and $C$ fails.
Thus, it is not trivial to use this translation to implement error
reporting in PEGs.
\emph{Rats!}~\cite{grimm2006rats} is a popular packrat parser
that implements error reporting with a strategy similar to
Ford's, with the change that it always reports error positions
at the start of productions, and pretty-prints non-terminal names
in the error message. For example, an error in a {\em ReturnStatement}
non-terminal becomes {\tt return statement expected}.
Even though error handling is an important task for parsers,
we did not find any other research results about error handling in PEGs,
beyond the heuristic proposed by Ford and the cut operator
proposed by Mizushima et al.
However, parser combinators \cite{hutton1992hfp} present some
similarities with PEGs so we will briefly discuss them for the
rest of this section.
In functional programming it is common to implement recursive
descent parsers using parser combinators \cite{hutton1992hfp}.
A parser is a function that we use to model symbols of the
grammar.
A parser combinator is a higher-order function that we use to
implement grammar constructions such as sequencing and choice.
One kind of parser combinator implements parsers that
return a list of all possible results of a parse, effectively
implementing a recursive descent parser with full backtracking.
Despite being actually deterministic in behavior (parsing the
same input always yields the same list of results), these
combinators are called {\em non-deterministic parser
combinators} due to their use of a non-deterministic choice operator.
We get parser combinators that have the same semantics as PEGs
by changing the return type from list of results to \texttt{Maybe}.
That is, we use deterministic parser combinators that return
\texttt{Maybe} to implement recursive descent parsers with limited
backtracking.
In the rest of this paper, whenever we refer to parser combinators
we mean these parser combinators with limited backtracking.
Like PEGs, parser combinators also use ordered choice and
try to accept input prefixes.
More precisely, parsers implemented using parser combinators also
backtrack in case of failure.
For this reason, when the input string contains syntax errors,
the longest parse usually indicates a position far away from the
position where the error really happened.
Hutton \cite{hutton1992hfp} introduced the \texttt{nofail} combinator
to implement error reporting in a quite simple way:
we just need to distinguish between failure and error during parsing.
More specifically, we can use the \texttt{nofail} combinator to
annotate the grammar's terminals and non-terminals that should
not fail; when they fail, the failure should be transformed into an
error. The difference between an error and a failure is that
an ordered choice just propagates an error in its first alternative
instead of backtracking and trying its second alternative, so any
error aborts the whole parser.
This technique is also called the \emph{three-values} technique~\cite{partridge1996fv}
because the parser finishes with one of the following values:
\texttt{OK}, \texttt{Fail} or \texttt{Error}.
Röjemo \cite{rojemo1995epc} presented a \texttt{cut} combinator
that we can also use to annotate the grammar pieces where
parsing should be aborted on failure, on behalf of efficiency and
error reporting.
The \texttt{cut} combinator is different from the cut operator\footnote{Throughout this paper, we refer to {\em combinators}
of parser combinators and to {\em operators} of PEGs, but these terms are effectively interchangeable.} ($\uparrow$) for PEGs because the combinator is abortive and unary
while the operator is not abortive and nullary.
The \texttt{cut} combinator introduced by Röjemo has the same
semantics as the \texttt{nofail} combinator introduced by Hutton.
Partridge and Wright \cite{partridge1996fv} showed that error detection
can be automated in parser combinators when we assume that the grammar
is LL(1).
Their main idea is:
if one alternative successfully consumes at least one symbol,
no other alternative can successfully consume any symbols.
Their technique is also known as the \emph{four-values} technique
because the parser finishes with one of the following values:
\texttt{Epsn}, when the parser finishes with success without
consuming any input;
\texttt{OK}, when the parser finishes with success consuming
some input;
\texttt{Fail}, when the parser fails without consuming any input; and
\texttt{Error}, when the parser fails consuming some input.
Three values were inspired by Hutton's work \cite{hutton1992hfp},
but with new meanings.
In the four-values technique, we do not need to annotate the grammar
because the authors changed the semantics of the sequence and choice
combinators to automatically generate the \texttt{Error} value
according to the Table \ref{tab:fv}.
In summary, the sequence combinator propagates an error when the
second parse fails after consuming some input while
the choice combinator does not try further alternatives
if the current one consumed at least one symbol from the input.
In case of error, the four-values technique detects the first symbol
following the longest parse of the input and uses this symbol to
report an error.
\begin{table}[t]
\begin{center}
\begin{tabular}{p{1cm}p{2cm}p{2cm}p{2cm}p{2cm}}
\hline
& $p_1$ & $p_2$ & $p_1 p_2$ & $p_1 \; | \; p_2$\\
\hline
& \texttt{Error} & \texttt{Error} & \texttt{Error} & \texttt{Error} \\
& \texttt{Error} & \texttt{Fail} & \texttt{Error} & \texttt{Error} \\
& \texttt{Error} & \texttt{Epsn} & \texttt{Error} & \texttt{Error} \\
& \texttt{Error} & \texttt{OK} ($x$) & \texttt{Error} & \texttt{Error} \\
& \texttt{Fail} & \texttt{Error} & \texttt{Fail} & \texttt{Error} \\
& \texttt{Fail} & \texttt{Fail} & \texttt{Fail} & \texttt{Fail} \\
& \texttt{Fail} & \texttt{Epsn} & \texttt{Fail} & \texttt{Epsn} \\
& \texttt{Fail} & \texttt{OK} ($x$) & \texttt{Fail} & \texttt{OK} ($x$) \\
& \texttt{Epsn} & \texttt{Error} & \texttt{Error} & \texttt{Error} \\
& \texttt{Epsn} & \texttt{Fail} & \texttt{Fail} & \texttt{Epsn} \\
& \texttt{Epsn} & \texttt{Epsn} & \texttt{Epsn} & \texttt{Epsn} \\
& \texttt{Epsn} & \texttt{OK} ($x$) & \texttt{OK} ($x$) & \texttt{OK} ($x$) \\
& \texttt{OK} ($x$) & \texttt{Error} & \texttt{Error} & \texttt{OK} ($x$) \\
& \texttt{OK} ($x$) & \texttt{Fail} & \texttt{Error} & \texttt{OK} ($x$) \\
& \texttt{OK} ($x$) & \texttt{Epsn} & \texttt{OK} ($x$) & \texttt{OK} ($x$) \\
& \texttt{OK} ($x$) & \texttt{OK} ($y$) & \texttt{OK} ($y$) & \texttt{OK} ($x$) \\
\end{tabular}
\end{center}
\caption{Behavior of sequence and choice in the four-values technique}
\label{tab:fv}
\end{table}
The four-values technique assumes that the input is composed by tokens
which are provided by a separate lexer.
However, being restricted to LL(1) grammars can be a limitation
because parser combinators, like PEGs, usually operate on strings of
characters to implement both lexer and parser together.
For instance, a parser for Tiny that is implemented with Parsec
\cite{leijen2001parsec} does not parse the following program:
\verb'read x;'.
That is, the matching of \verb'read' against \verb'repeat' generates
an error.
Such behavior is confirmed in Table \ref{tab:fv} by the third line
from the bottom.
Parsec is a parser combinator library for Haskell that employs a
technique equivalent to the four-values technique for implementing
LL(1) predictive parsers that automatically report errors
\cite{leijen2001parsec}. To overcome the LL(1) limitation,
Parsec introduced the {\tt try} combinator, a dual of Hutton's
{\tt nofail} combinator. The effect of {\tt try}
is to translate an error into a backtrackeable
failure. The idea is
to use {\tt try} to annotate the parts of the grammar where
arbitrary lookahead is needed.
Parsec's restriction to LL(1) grammars made it possible to implement
an error reporting technique similar to the one
use in top-down parsers.
Parsec produces error messages that include the error position,
the character at this position and the \texttt{FIRST} and
\texttt{FOLLOW} sets of the
productions that were expected at this position.
Parsec also implements the error injection combinator (\texttt{<?>})
for naming productions.
This combinator gets two arguments: a parser \texttt{p} and a
string \texttt{exp}.
The string \texttt{exp} replaces the \texttt{FIRST} set of a
parser \texttt{p} when all the alternatives of \texttt{p} failed.
This combinator is useful to name terminals and non-terminals to
get better information about the context of a syntax error.
Swierstra and Duponcheel \cite{swierstra1996dec} showed an
implementation of parser combinators for error recovery,
although most libraries and parser generators that are based
on parser combinators implement only error reporting.
Their work shows an implementation of parser combinators that
repair erroneous inputs, produce an appropriated message,
and continue parsing the rest of the input.
\section{Formalization of Farthest Failure Tracking}
\label{sec:semfarther}
In Section~\ref{sec:pegs}, we saw that a common approach
to report errors in PEGs is to inform the position of the
farthest failure,
and how we can use semantic actions to implement
this approach in common PEG implementations. In this section,
we show a conservative extension of the PEG formalism
that expresses this error reporting approach directly,
making it a part of the semantics of PEGs instead of an
ad-hoc extension to PEG implementations. We also show how
this extension can easily be adapted to build a more
elaborate error object, instead of just returning a position.
The result of a PEG on a given input is either
${\tt fail}$ or a suffix of the input. Returning a suffix of the
input means that the PEG succeeded and consumed a prefix
of the input. PEGs with farthest failure tracking return
the product of the original result and either another
suffix of the input, to
denote the position of the farthest
failure during parsing, or ${\tt nil}$ to denote that
there were no failures. We cannot use an empty string
to denote that there were no failures, as an empty string
already means that a failure has occurred after the last position
of the input.
Figure~\ref{fig:semfarthest} presents a formal semantics of
PEGs with farthest failure tracking as a set of inference rules. The notation
$G[p] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \; \Tup{y}{v}$ means that the
PEG \,$G = (V,T,P,p_{S})$, using a starting expression $p$,
matches the input $xy$, consuming $x$ and leaving the suffix $y$,
with the farthest failure position during this match
denoted by the suffix $v$ of $xy$. The notation $v?$
means a value that is either a suffix $v$ of the input
or ${\tt nil}$. An $X$ denotes a value that is either a
suffix of the input or ${\tt fail}$.
The auxiliary function {\tt smallest}\ that appears on
Figure~\ref{fig:semfarthest} compares two possible
error positions, denoted by a suffix of the
input string, or {\tt nil} if no failure has occurred,
and returns the furthest: any suffix of the input
is a furthest possible error position than {\tt nil}\,
and a shorter suffix is a furthest possible
error position than a longer suffix.
Rule {\bf empty.1} deals with the empty expression. This
expression never fails, so the failure position is always
${\tt nil}$. Rule {\bf var.1} deals with
non-terminals, so it just propagates the result of
matching the right-hand side of the non-terminal.
Rules {\bf char.1}, {\bf char.2}, and {\bf char.3} deal with
terminals. The latter two rules denote failures, so they
return the subject as the failure position.
\begin{figure}[t]
{\small
\begin{align*}
& \textbf{Empty} \;\;\;
{\frac{}{G[\varepsilon] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{x}{{\tt nil}}}} \mylabel{empty.1}
\;\;\;\;\; \textbf{Non-terminal} \;\;\;
{\frac{G[P(A)] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{v?}}
{G[A] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{v?}}} \mylabel{var.1}
\\ \\
& \textbf{Terminal} \fivespaces\fivespaces
{\frac{}{G[a] \; ax \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{x}{{\tt nil}}}} \mylabel{char.1} \\ \\
& \tenspaces\tenspaces
{\frac{}{G[b] \; ax \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{ax}}}, b \ne a \mylabel{char.2}
\;\;\;\;\;
{\frac{}{G[a] \; \varepsilon \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{\varepsilon}}} \mylabel{char.3}
\\ \\
& \textbf{Concatenation} \;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{y}{v?} \;\;\;\;\; G[p_2] \; y \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{w?}}
{G[p_1 \; p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{\Suff{v?}{w?}}}} \mylabel{con.1}
\;\;\;
{\frac{G[p_1] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{v?}}
{G[p_1 \; p_2] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{v?}}} \mylabel{con.2}
\\ \\
& \textbf{Ordered Choice} \;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{y}{v?}}
{G[p_1 \;/\; p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{y}{v?}}} \mylabel{ord.1}
\;\;\;
{\frac{G[p_1] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{v?} \;\;\;\;\; G[p_2] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{w?}}
{G[p_1 \;/\; p_2] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{\Suff{v?}{w?}}}} \mylabel{ord.2}
\\ \\
& \textbf{Repetition} \;\;\;
{\frac{G[p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{v?}}
{G[p*] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{x}{v?}}} \mylabel{rep.1}
\;\;\;
{\frac{G[p] \; xyz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{yz}{v?} \;\;\;\;\; G[p*] \; yz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{z}{w?}}
{G[p*] \; xyz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{z}{\Suff{v?}{w?}}}} \mylabel{rep.2}
\\ \\
& \textbf{Negative Predicate} \;\;\;
{\frac{G[p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{v?}}
{G[!p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{x}{{\tt nil}}}} \mylabel{not.1}
\;\;\;
{\frac{G[p] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{y}{v?}}
{G[!p] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{xy}}} \mylabel{not.2}
\end{align*}
}
\caption{Semantics of PEGs with farthest failure tracking}
\label{fig:semfarthest}
\end{figure}
Rules {\bf con.1} and {\bf con.2} deal with concatenation.
The second rule just propagates the failure position, but
rule {\bf con.1} needs to take the farthest position
between the two parts of the concatenation. The rules
for ordered choice ({\bf ord.1} and {\bf ord.2}) and
repetition ({\bf rep.1} and {\bf rep.2}) work in
a similar way: whenever there are two possible farthest
failure positions, we use {\tt smallest}\ to take the farthest of
them.
Finally, rules {\bf not.1} and {\bf not.2} deal with
the syntactic predicate. Fitting this predicate in
this error reporting framework is subtle. The rules
not only need to make sense for the predicate in
isolation, but also have to make sense for $!!p$, which
works as a ``positive lookahead" predicate that succeeds
if $p$ succeeds but fails if $p$ fails while never
consuming any input.
For {\bf not.1}, the case where $!p$ succeeds, we have two
choices: either propagate $p$'s farthest failure position, or
ignore it (using ${\tt nil}$). The first choice can lead to an
odd case, where the failure that made $!p$ succeed can get
the blame for the overall failure of the PEG, so
{\bf not.1} takes the second choice.
We also have two choices for the
case where $!p$ fails: either propagate the failure position
from $p$ or just use the current position. The first choice
can also lead to an odd cases where the overall failure
of the PEG might be blamed on something in $p$ that, if
corrected, still makes $p$ succeed and $!p$ fail, so
{\bf not.2} also takes the second choice.
The end result is that what happens inside a predicate simply
does not take part in error reporting at all, which is the
simplest approach, and also gives a consistent result for
$!!p$.
As an example that shows the interaction of these
rules, let us consider again the Tiny program in
Figure~\ref{fig:tiny}, reproduced below:
\begin{verbatim}
01 n := 5;
02 f := 1;
03 repeat
04 f := f * n;
05 n := n - 1
06 until (n < 1);
07 write f;
\end{verbatim}
The missing ``{\tt ;}" at the end of line 5
makes the repetition {\it (Cmd SEMICOLON)*}
inside {\it CmdSeq} succeed through rule {\bf rep.1}
and propagate the failure position to the concatenation
inside this same non-terminal. Rule {\bf con.2} then
propagates the same failure position to the concatenation
inside {\it RepeatCmd}.
Another failure occurs in {\it RepeatCmd} when {\tt until}
does not match {\tt n}, but rule {\bf con.2} again propagates
the position of the missing ``{\tt ;}", which is farthest.
This leads to a failure inside the repetition on {\it CmdSeq}
again, which propagates the position of the missing ``{\tt ;}"
through rule {\bf rep.2}. Finally, rules {\bf con.2} and {\bf var.1}
propagate this failure position to the top of the grammar
so the missing semicolon gets the blame for the PEG
not consuming the whole program.
We can translate the failure suffix into a line and column
number inside the original subject, and extract the
first token from the beginning of this suffix to
produce an error message similar to the error messages
in Section~\ref{sec:pegs}:
\begin{verbatim}
factorial.tiny:6:1: syntax error, unexpected 'until'
\end{verbatim}
While this message correctly pinpoints the location
of the error, it could be more informative. We can
extend the semantics of Figure~\ref{fig:semfarthest}
to gather more information than just the farthest
failure position, thus making us able to generate
error messages as informative as the one in the end
of Section~\ref{sec:pegs}.
The new semantics formalizes a strategy similar to
the one used by Ford~\citep{ford2002packrat} in his
PEG implementation. The basic idea is to keep a list of
the simple expressions that the PEG was expecting to match when
a failure occurred, and use this list to build
an error message. Figure~\ref{fig:semfarthestjoin}
gives inference rules for the $\stackrel{\mbox{\tiny{PEG}}}{\leadsto}$ relation of
this new semantics.
The result of the matching of a PEG $G$ against an input $x$
is still a pair, but the second component is now
another pair, the {\em error pair}.
The first component of the error pair is the farthest
error position, same as in the previous semantics.
The second component is a list of parsing expressions
that were expected at this error position.
If the grammar does not have syntactic predicates,
the expressions in this list are just terminals and
non-terminals.
\begin{figure}[p]
{\small
\begin{align*}
& \textbf{Empty} \;\;\;
{\frac{}{G[\varepsilon] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{x}{({\tt nil},\{\})}}} \mylabel{empty.1} \\ \\
& \textbf{Non-terminal} \;\;\;
{\frac{G[P(A)] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{\Tup{v?}{L}}}
{G[A] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{{\tt joinVar}((v?,L),x,A)}}} \mylabel{var.1}
\\ \\
& \textbf{Terminal} \fivespaces\fivespaces
{\frac{}{G[a] \; ax \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{x}{({\tt nil},\{\})}}} \mylabel{char.1} \\ \\
& \fivespaces\fivespaces
{\frac{}{G[b] \; ax \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{\Tup{ax}{\Mon{b}}}}}, b \ne a \mylabel{char.2}
\;\;\;\;\;
{\frac{}{G[a] \; \varepsilon \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{\Tup{\varepsilon}{\Mon{a}}}}} \mylabel{char.3}
\\ \\
& \textbf{Concatenation} \;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{y}{(v?,L_1)}
\;\;\;\;\; G[p_2] \; y \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{(w?,L_2)}}
{G[p_1 \; p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{\Jj{(v?,L_1)}{(w?,L_2)}}}} \mylabel{con.1} \\ \\
&\;\;\;\;\;
{\frac{G[p_1] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{(v?,L)}}
{G[p_1 \; p_2] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{(v?,L)}}} \mylabel{con.2}
\\ \\
& \textbf{Ordered Choice} \;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{y}{(v?,L)}}
{G[p_1 \;/\; p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{y}{(v?,L)}}} \mylabel{ord.1} \\ \\
&\;\;\;\;\;
{\frac{G[p_1] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{(v?,L_1)} \;\;\;\;\; G[p_2] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{(w?,L_2)}}
{G[p_1 \;/\; p_2] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{X}{\Jj{(v?,L_1)}{(w?,L_2)}}}} \mylabel{ord.2}
\\ \\
& \textbf{Repetition} \;\;\;
{\frac{G[p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{(v?,L)}}
{G[p*] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{x}{(v?,L)}}} \mylabel{rep.1}
\\ \\
& \;\;\;\;\; {\frac{G[p] \; xyz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{yz}{(v?,L_1)} \;\;\;\;\; G[p*] \; yz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{z}{(w?,L_2)}}
{G[p*] \; xyz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{z}{\Jj{(v?,L_1)}{(w?,L_2)}}}} \mylabel{rep.2}
\\ \\
& \textbf{Negative Predicate} \;\;\;
{\frac{G[p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{(v?,L)}}
{G[!p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{x}{({\tt nil},\{\})}}} \mylabel{not.1}
\;\;\;
{\frac{G[p] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{y}{(v?,L)}}
{G[!p] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{\Tup{xy}{\Mon{!p}}}}} \mylabel{not.2}
\end{align*}
}
\caption{Semantics of PEGs with farthest failure tracking and error lists}
\label{fig:semfarthestjoin}
\end{figure}
Rules {\bf empty.1} and {\bf char.1} do not change,
and rules {\bf char.2} and {\bf char.3} just
return a list with the terminal that
they tried to match.
Rule {\bf var.1} uses an auxiliary function {\tt joinVar}, defined
as follows:
\begin{align*}
{\tt joinVar}((x,L),x,A) & = (x,\{A\}) \\
{\tt joinVar}(({\tt nil},\{\}),x,A) & = ({\tt nil},\{\}) \\
{\tt joinVar}((v,L),x,A) & = (v,L) & \mathrm{where}\ x \neq v
\end{align*}
The idea behind {\tt joinVar}\ is simple: if the right-hand side
of terminal $A$ has a farthest error position that is the
same as the current position in the input, it means that
we can treat this possible error as an error while
expecting $A$ itself. This is expressed in the first
case of {\tt joinVar}\ by replacing the list of expected expressions
L with just $\{A\}$. The other cases just propagate the
failure information returned by $P(A)$.
To further understand how {\tt joinVar}\ works, let us consider an example,
adapted from the grammar in Figure~\ref{fig:tiny}:
\begin{align*}
\it{Factor} & \leftarrow ``(" \; \it{Exp} \; ``)"
\; / \; \it{Digit} \; \it{Digit}* \\
\it{Digit} & \leftarrow ``0" \;/\; ``1" \;/\; ``2"
\;/\; ``3" \;/\; ``4" \;/\; ``5" \;/\; ``6" \;/\; ``7"
\;/\; ``8" \;/\; ``9"
\end{align*}
When we try to match {\it Factor} against
an input that does not match either $``("$
or {\it Digit}, such as \texttt{id}, both
alternatives of the ordered choice fail and the farthest
failure position of both matches is the same input that
we were trying to match with {\it Factor}.
So, instead of keeping
an error list that would later give us an error message like
the following one:
\begin{verbatim}
Unexpected 'id', expecting '(' or '0' or '1' or '2' or '3' or '4' or '5'
or '6' or '7' or '8' or '9'
\end{verbatim}
Or even an error message like this one:
\begin{verbatim}
Unexpected 'id', expecting '(' or Digit
\end{verbatim}
We replace the error list built during the matching of
the right-hand side of {\it Factor} with just $\{ {\it Factor} \}$.
From this new list we can get the following higher-level
error message:
\begin{verbatim}
Unexpected 'id', expecting Factor
\end{verbatim}
If the failure occurred in the middle of the first
alternative of {\it Factor} (for example, because of a missing ``)',
we would keep the original error list instead of replacing
it with $\{ {\it Factor} \}$.
Using the names of the non-terminals in the error message instead
of a list of terminals that might have started this non-terminal
might not be better if the names of the non-terminals give no
indication to the user of what was expected. It is easy
to replace {\tt joinVar}\ with just $(v?,L)$, propagating the failure
information from the expansion of the non-terminal and keeping
the error message just with symbols that the user can add
to the input. Another possibility is to have two kinds of
non-terminals, and use {\tt joinVar}\ for informative non-terminals and
simple propagation for the others.
In the case of concatenation, a failure in the first
expression of the concatenation means we just need to
propagate the error list, so case {\bf con.2} is
straightforward. If both expressions succeed, we might
need to merge both error lists, if the two parts
of the concatenation have the same failure position.
Rule {\bf con.1} uses the auxiliary {\tt join}\ function,
defined below:
\begin{align*}
{\tt join}((x,L),({\tt nil},\{\})) & = (x,L) \\
{\tt join}(({\tt nil},\{\}),(x,L)) & = (x,L) \\
{\tt join}((xy,L_1),(y,L_2)) & = (y,L_2) & \mathrm{where}\ x \neq \varepsilon \\
{\tt join}((y,L_1),(xy,L_2)) & = (y,L_1) & \mathrm{where}\ x \neq \varepsilon \\
{\tt join}((x,L_1),(x,L_2)) & = (x,L_1 \mathrel{\drawplusplus {12pt}{0.4pt}{8pt}} L_2)
\end{align*}
The first four cases of {\tt join}\ keep the furthest
and associated set of expected expressions. In the first
two cases, {\tt nil} means that no error has occurred,
so any error automatically becomes the farthest
error position. In the next two cases, one of the
furthest error positions is a strict suffix of the
other, so is the furthest of the two and the other
is discarded. The only remaining possibility,
expressed in the last case, is that
the two positions are identical, and we merge their
expected sets.
The rules for ordered choice and repetition use the
same {\tt join}\ function, where applicable. Finally, the rules
for the syntactic predicate $!p$ also ignore the error
information inside $p$. Rule {\bf not.2} blames the
failure on the predicate itself.
Going back to our running example (Figure~\ref{fig:tinyerror}),
our error tracking semantics will give an error list that
lets us generate the following error message:
\begin{verbatim}
factorial.tiny:6:1: syntax error, unexpected 'until',
expecting ';', '=', '<', '-', '+', '/', '*'
\end{verbatim}
The operators also end up in the error list because
their lexical rules all fail in the same position as
the semicolon. This error message is similar to the
error message we get using the error tracking
combinators that we implemented on LPeg, at the end of
Section~\ref{sec:pegs}.
We might try to tweak the error tracking heuristics of
repetition and ordered choice to ignore errors that happen
in the first symbol of the input, which would let us
take out the operators from the error list in our previous
example, and give a more succinct error message:
\begin{verbatim}
factorial.tiny:6:1: syntax error, unexpected 'until', expecting ';'
\end{verbatim}
This heuristic is not sound in the general case, though.
Suppose we replace line 6 of Figure~\ref{fig:tinyerror}
with the following line:
\begin{verbatim}
6 n ; until (n < 1);
\end{verbatim}
The tweaked heuristic would still produce an error list
with just the semicolon, which is clearly wrong, as the
problem is now a missing operator.
It is common in PEGs to mix scanning and parsing in the
same grammar, as syntactic predicates and repetition make
lexical patterns convenient to express. But this can lead
to problems in the automatic generation of error messages
because failures are expected while recognizing a token,
and these failures related to scanning can pollute the
error list that we use to generate the error message.
As an example, let us consider the lexical rule {\tt THEN}
from the PEG for the Tiny language of Section~\ref{sec:pegs}:
\[
{\tt THEN} \leftarrow {\tt then} \; \it{!IDRest \; Skip}
\]
The pattern in the right-hand side fails if any alphanumeric
character follows {\tt then} in the input, putting the
predicate in the error list. The error will be reported to
the user as an unexpected character after {\tt then}
instead of a missing {\tt then} keyword.
One solution is to split the set of non-terminals into
{\em lexical} and {\em non-lexical} non-terminals.
Non-lexical non-terminals follow rule {\bf var.1}, but
lexical terminals follow a pair of new rules:
\begin{align*}
&
{\frac{G[P(A)] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{y}{\Tup{v?}{L}}}
{G[A] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{y}{({\tt nil},\{\})}}} \mylabel{lvar.1}\\ \\
&
{\frac{G[P(A)] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{\Tup{v?}{L}}}
{G[A] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{(x, \{ A \})}}} \mylabel{lvar.2}
\end{align*}
Intuitively, a lexical non-terminal reports errors just like
a terminal. Any failures that happen in a successful match
are ignored, as they are considered to be expected, and a failed
match of the whole non-terminal gets blamed on the
non-terminal itself, at its starting position in the input.
All the extensions to the semantics of PEGs that we discussed
in this section are {\em conservative}: if a PEG fails with
some subject in the original semantics, it will also fail
in all of our extensions, and if a PEG matches some subject
and leaves a suffix, it will also match and leave the same
suffix in all of our extensions. This is stated by the
following lemma, where we use the symbol $\stackrel{\mbox{\tiny{PEG}}}{\leadsto}$ to represent
the regular semantics of PEGs~\cite{mascarenhas2014}, and the symbol $\stackrel{\mbox{\tiny{${\mathrm{PEG}}^\prime$}}}{\leadsto}$ to represent
the extended semantics of PEGs presented in Figure~\ref{fig:semfarthest}:
\begin{lemma}[Conservativeness of farthest failure tracking]
Given a PEG $G$, a parsing expression $p$ and a subject $xy$, we have that
$\Matg{p}{xy} \stackrel{\mbox{\tiny{PEG}}}{\leadsto} y$ \,iff\; $\Matg{p}{xy} \stackrel{\mbox{\tiny{${\mathrm{PEG}}^\prime$}}}{\leadsto} \Tup{y}{v?}$,
and that
$\Matg{p}{xy} \stackrel{\mbox{\tiny{PEG}}}{\leadsto} {\tt fail}$ \,iff\; $\Matg{p}{xy} \stackrel{\mbox{\tiny{${\mathrm{PEG}}^\prime$}}}{\leadsto} \Tup{{\tt fail}}{v?}$.
\end{lemma}
\begin{proof}
By induction on the height of the respective proof trees.
The proof is straightforward, since the only difference
between the rules of $\stackrel{\mbox{\tiny{PEG}}}{\leadsto}$ and $\stackrel{\mbox{\tiny{${\mathrm{PEG}}^\prime$}}}{\leadsto}$ is presence of
the farthest failure position, but this position has
no influence on whether the expression successfully
consumes part of the input or fails.
\end{proof}
A similar Lemma for the semantics with expected expression lists
of Figure~\ref{fig:semfarthestjoin} is also straightforward to
prove, where we use the symbol $\stackrel{\mbox{\tiny{PEGL}}}{\leadsto}$ to represent those extended
semantics:
\begin{lemma}[Conservativeness of farthest failure tracking with expected lists]
Given a PEG $G$, a parsing expression $p$ and a subject $xy$, we have that
$\Matg{p}{xy} \stackrel{\mbox{\tiny{PEG}}}{\leadsto} y$ \,iff\; $\Matg{p}{xy} \stackrel{\mbox{\tiny{PEGL}}}{\leadsto} \Tup{y}{(v?, L)}$,
and that
$\Matg{p}{xy} \stackrel{\mbox{\tiny{PEG}}}{\leadsto} {\tt fail}$ \,iff\; $\Matg{p}{xy} \stackrel{\mbox{\tiny{PEGL}}}{\leadsto} \Tup{{\tt fail}}{(v?,L)}$.
\end{lemma}
\begin{proof}
By induction on the height of the respective proof trees.
The proof is also straightforward, since the only difference
between the rules of $\stackrel{\mbox{\tiny{PEG}}}{\leadsto}$ and $\stackrel{\mbox{\tiny{PEGL}}}{\leadsto}$ is presence of
the farthest failure position and list of expected expressions,
but this extra information has
no influence on whether the expression successfully
consumes part of the input or fails.
\end{proof}
In the next section, we will introduce another approach
for error reporting in PEGs, which can produce more precise
error messages, at the cost of annotating the grammar.
\section{Labeled Failures} \label{sec:lf}
Exceptions are a common mechanism for signaling and handling
errors in programming languages.
Exceptions let programmers classify the different errors their
programs may signal by using distinct types for distinct errors,
and decouple error handling from regular program logic.
In this section we add \emph{labeled failures} to PEGs,
a mechanism akin to exceptions and exception handling,
with the goal of improving error reporting while preserving
the composability of PEGs,
and in the next section we discuss how to use PEGs with labeled
failures to implement some of
the techniques that we have discussed in Section~\ref{sec:rel}:
the \texttt{nofail} combinator \cite{hutton1992hfp},
the \texttt{cut} combinator \cite{rojemo1995epc},
the four-values technique \cite{partridge1996fv} and
the \texttt{try} combinator \cite{leijen2001parsec}.
A labeled PEG $G$ is a tuple $(V,T,P,L,p_{S})$ where
$L$ is a finite set of labels that must include the {\tt fail}
label, and the
expressions in $P$ have been extended with the {\em throw}
operator, explained below. The other parts use the same definitions from Section \ref{sec:pegs}.
The abstract syntax of labeled parsing expressions adds the
\emph{throw} operator $\Uparrow^{l}$, which generates a failure
with label $l$, and adds an extra argument $S$ to the ordered choice
operator, which is the set of labels that the ordered choice should
catch.
$S$ must be a subset of $L$.
\[\it{
p = \varepsilon \; | \; a \; | \; A \; | \; p_1 p_2 \; | \;
p_1 /^{S} p_2 \; | \; p* \; | \; !p \; | \; \Uparrow^{l}
}\]
\begin{figure}
{\small
\begin{align*}
& \textbf{Empty} \;\;\;
{\frac{}{G[\varepsilon] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} x}} \mylabel{empty.1}
\\ \\
& \textbf{Terminal} \;\;\;
{\frac{}{G[a] \; ax \stackrel{\mbox{\tiny{PEG}}}{\leadsto} x}} \mylabel{char.1}
\;\;\;
{\frac{}{G[b] \; ax \stackrel{\mbox{\tiny{PEG}}}{\leadsto} {\tt fail}}}, b \ne a \mylabel{char.2}
\;\;\;
{\frac{}{G[a] \; \varepsilon \stackrel{\mbox{\tiny{PEG}}}{\leadsto} {\tt fail}}} \mylabel{char.3}
\\ \\
& \textbf{Non-terminal} \;\;\;
{\frac{G[P(A)] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}
{G[A] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}} \mylabel{var.1}
\\ \\
& \textbf{Concatenation} \;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} y \;\;\;\;\; G[p_2] \; y \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}
{G[p_1 \; p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}} \mylabel{con.1}
\;\;\;
{\frac{G[p_1] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} l}
{G[p_1 \; p_2] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} l}} \mylabel{con.2}
\\ \\
& \textbf{Ordered Choice} \;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} y}
{G[p_1 \;\slash^S\; p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} y}} \mylabel{ord.1}
\;\;\;
{\frac{G[p_1] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} l}
{G[p_1 \;\slash^S\; p_2] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} l}}, l \not\in S \mylabel{ord.2}
\\ \\
&
\fivespaces\fivespaces \fivespaces\fivespaces \fivespaces\fivespaces \,
{\frac{G[p_1] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} l \;\;\;\;\; G[p_2] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}
{G[p_1 \;\slash^S\; p_2] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}}, l \in S \mylabel{ord.3}
\\ \\
& \textbf{Repetition} \;\;\;
{\frac{G[p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} {\tt fail}}
{G[p*] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} x}} \mylabel{rep.1}
\;\;\;
{\frac{G[p] \; xyz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} yz \;\;\;\;\; G[p*] \; yz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}
{G[p*] \; xyz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}} \mylabel{rep.2}
\\ \\
&
\fivespaces\fivespaces \fivespaces\fivespaces \;\,
{\frac{G[p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} l}
{G[p*] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} l}}, l \neq {\tt fail} \mylabel{rep.3}
\\ \\
& \textbf{Negative Predicate} \;\;\;
{\frac{G[p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} {\tt fail}}
{G[!p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} x}} \mylabel{not.1}
\;\;\;
{\frac{G[p] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} y}
{G[!p] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} {\tt fail}}} \mylabel{not.2}
\\ \\
&
\fivespaces\fivespaces \fivespaces\fivespaces \fivespaces\fivespaces \;\;\;\;\; \;
{\frac{G[p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} l}
{G[!p] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} l}}, l \neq {\tt fail} \mylabel{not.3}
\\ \\
& \textbf{Throw} \;\;\;
{\frac{}{G[\Uparrow^{l}] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} l}} \mylabel{throw.1}
\end{align*}
}
\caption{Semantics of PEGs with labels}
\label{fig:sem}
\end{figure}
The semantics of PEGs with labels is defined by the relation
$\stackrel{\mbox{\tiny{PEG}}}{\leadsto}$ among a parsing expression, an input string and a result.
The result is either a string or a label.
The notation $G[p] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \; y$ means that the expression
$p$ matches the input $xy$, consumes the prefix $x$ and leaves the
suffix $y$ as the output.
The notation $G[p] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \; l$ indicates that the
matching of $p$ fails with label $l$ on the input $xy$.
Figure \ref{fig:sem} presents the semantics of PEGs with labels
as a set of inference rules.
Intuitively,
$\varepsilon$ successfully matches the empty string, not changing
the input;
$a$ matches and consumes itself and fails with label \texttt{fail}
otherwise;
$A$ tries to match the expression $P(A)$;
$p_1 p_2$ tries to match $p_1$, if $p_1$ matches an input prefix,
then it tries to match $p_2$ with the suffix left by $p_1$,
the label $l$ is propagated otherwise;
$p_1 /^{S} p_2$ tries to match $p_1$ in the input and tries to
match $p_2$ in the same input only if $p_1$ fails with a label
$l \in S$, the label $l$ is propagated otherwise;
$p*$ repeatedly matches $p$ until the matching of $p$
silently fails with label {\tt fail}, and propagates a label $l$
when $p$ fails with this label;
$!p$ successfully matches if the input does not match $p$ with
the label {\tt fail}, fails producing the label \texttt{fail} when
the input matches $p$, and propagates a label $l$ when $p$
fails with this label, not consuming the input in all cases;
$\Uparrow^{l}$ produces the label $l$.
We faced some design decisions in our formulation that are worth
discussing. First, we require the presence of a \texttt{fail} label
to maintain
compatibility with the original semantics of PEGs.
For the same reason, we define the expression $p_1 / p_2$ as
syntactic sugar for $p_1 /^{\{{\tt fail}\}} p_2$.
We use a set of labels in the ordered choice as a convenience.
We could have each ordered choice handling a single label,
and it would just lead to duplication:
an expression
$p_1 \; /^{\{l_1,l_2,...,l_n\}} \; p_2$
would become
$( \; ... \; ((p_1 \; /^{l_1} \; p_2) \; /^{l_2} \; p_2) \; ... \; /^{l_n} \; p_2)$.
Another choice was how to handle labels in a repetition.
We chose to have a repetition stop silently only on the
\texttt{fail}
label to maintain the following identity: the expression $p*$
is equivalent to a fresh non-terminal $A$ plus the rule
$A \leftarrow p \; A \; / \; \varepsilon$.
Finally, the negative predicate succeeds only on the
\texttt{fail} label to
allow the implementation of the positive predicate: the
expression
$\&p$ that implements the positive predicate in the original semantics
of PEGs \cite{ford2002packrat,ford2002pappy,ford2004peg}
is equivalent to the expression $!!p$.
Both expressions successfully match if the input matches $p$,
fail producing the label \texttt{fail} when the input does
not match $p$, and propagate a label $l$ when $p$ fails
with this label, not consuming the input in all cases.
Figure \ref{fig:tinylabels} presents a PEG with labels for the Tiny
language from Section \ref{sec:pegs}.
The expression $[p]^{l}$ is syntactic
sugar for $(p \; / \; \Uparrow^{l})$.
The strategy we used to annotate the grammar was the following:
first, annotate every terminal that should not fail, that is,
making the PEG backtrack on failure of that terminal would
be useless, as the whole parse would either fail or not consume
the whole input in that case. For an LL(1) grammar like the
one in the example, that means all terminals in a production except
the one in the very beginning of the production.
After annotating the terminals, we make the same assessment for
whole productions. Productions that can fail but should not get a new
alternative that throws an error label for that production.
If a production starts with a non-terminal, we first assess the
production associated with that non-terminal.
For Tiny,
we end up annotating just two productions, {\em Factor} and {\em CmdSeq}.
Productions {\em Exp}, {\em SimpleExp}, and {\em Term} also
should not fail, but after annotating {\em Factor} they always
either succeed or throw the label {\tt exp}. The {\it Cmd}
production can fail, because it controls whether the repetition inside
{\it CmdSeq} stops or continue.
Notice that this is just an example of how a grammar can be
annotated. More thorough analyses are possible: for example,
we can deduce that {\it Cmd} is not allowed to fail unless
the next token is one of {\tt ELSE}, {\tt END}, {\tt UNTIL},
or the end of the input (the {\tt FOLLOW} set of {\it Cmd}), and instead of $\Uparrow^{cmd}$ add $!({\tt ELSE} \; / \; {\tt END} \; / \;
{\tt UNTIL} \; / \; !.) \Uparrow^{cmd}$ as a new alternative. This would
remove the need for the $\Uparrow^{cmd}$ annotation of {\it CmdSeq}.
\begin{figure}
\begin{align*}
\it{Tiny} & \leftarrow \it{CmdSeq}\\
\it{CmdSeq} & \leftarrow \it{(Cmd \; [{\tt SEMICOLON}]^{\tt sc}) \;
(Cmd \; [{\tt SEMICOLON}]^{\tt sc})*} \; / \; \Uparrow^{cmd}\\
\it{Cmd} & \leftarrow \it{IfCmd \; / \; RepeatCmd \; / \;
AssignCmd \; / \; ReadCmd \; / \; WriteCmd}\\
\it{IfCmd} & \leftarrow \it{{\tt IF} \, Exp \,
[{\tt THEN}]^{\tt then} \, CmdSeq \,
({\tt ELSE} \; CmdSeq \;/\; \varepsilon) \;
[{\tt END}]^{\tt end}}\\
\it{RepeatCmd} & \leftarrow \it{{\tt REPEAT} \; CmdSeq \;
[{\tt UNTIL}]^{\tt until} \; Exp}\\
\it{AssignCmd} & \leftarrow \it{{\tt NAME} \; [{\tt ASSIGNMENT}]^{\tt bind} \;
Exp}\\
\it{ReadCmd} & \leftarrow \it{{\tt READ} \; [{\tt NAME}]^{\tt read}}\\
\it{WriteCmd} & \leftarrow \it{{\tt WRITE} \; Exp}\\
\it{Exp} & \leftarrow \it{SimpleExp \;
(({\tt LESS} \; / \; {\tt EQUAL}) \; SimpleExp \; /
\; \varepsilon})\\
\it{SimpleExp} & \leftarrow \it{Term \;
(({\tt ADD} \; / \; {\tt SUB}) \; Term)*}\\
\it{Term} & \leftarrow \it{Factor \;
(({\tt MUL} \; / \; {\tt DIV}) \; Factor)*}\\
\it{Factor} & \leftarrow \it{{\tt OPENPAR} \; Exp
\; [{\tt CLOSEPAR}]^{\tt cp} \; /
\; {\tt NUMBER} \; / \; {\tt NAME}} \; / \; \Uparrow^{\tt exp}
\end{align*}
\caption{A PEG with labels for the Tiny language}
\label{fig:tinylabels}
\end{figure}
The PEG reports an error when parsing finishes with an
uncaught label. Each label is associated with a
meaningful error message.
For instance, if we use this PEG for Tiny to parse the
code example from Section \ref{sec:pegs}, parsing finishes with the
\texttt{sc} label and the PEG can use it to produce the following
error message:
\begin{verbatim}
factorial.tiny:6:1: syntax error, there is a missing ';'
\end{verbatim}
Note how the semantics of the repetition works with the rule
\textit{CmdSeq}.
Inside the repetition, the \texttt{fail} label means that there are no
more commands to be matched and the repetition should stop while the
\texttt{sc} label means that a semicolon (\texttt{;}) failed to match.
It would not be possible to write the rule \textit{CmdSeq}
using repetition if we had chosen to stop the repetition with any
label, instead of stopping only with the \texttt{fail} label,
because the repetition would accept the \texttt{sc} label as the
end of the repetition whereas it should propagate this label.
\begin{figure}
{\small
\begin{align*}
& \textbf{Empty} \;\;\;
{\frac{}{G[\varepsilon] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} x}} \mylabel{empty.1}
\\ \\
& \textbf{Terminal} \;\;\;
{\frac{}{G[a] \; ax \stackrel{\mbox{\tiny{PEG}}}{\leadsto} x}} \mylabel{char.1}
\;\;\;
{\frac{}{G[b] \; ax \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{ax}}}, b \ne a \mylabel{char.2}
\;\;\;
{\frac{}{G[a] \; \varepsilon \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{\varepsilon}}} \mylabel{char.3}
\\ \\
& \textbf{Non-terminal} \;\;\;
{\frac{G[P(A)] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}
{G[A] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}} \mylabel{var.1}
\\ \\
& \textbf{Concatenation} \;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} y \;\;\;\;\; G[p_2] \; y \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}
{G[p_1 \; p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}} \mylabel{con.1}
\;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{l}{y}}
{G[p_1 \; p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{l}{y}}} \mylabel{con.2}
\\ \\
& \textbf{Ordered Choice} \;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} y}
{G[p_1 \;\slash^S\; p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} y}} \mylabel{ord.1}
\;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{l}{y}}
{G[p_1 \;\slash^S\; p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{l}{y}}}, l \not\in S \mylabel{ord.2}
\\ \\
&
\fivespaces\fivespaces \fivespaces\fivespaces \fivespaces\fivespaces \,
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{l}{y} \;\;\;\;\; G[p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}
{G[p_1 \;\slash^S\; p_2] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}}, l \in S \mylabel{ord.3}
\\ \\
& \textbf{Repetition} \;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{y}}
{G[p_1\!*] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} xy}} \mylabel{rep.1}
\;\;\;
{\frac{G[p_1] \; xyz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} yz \;\;\;\;\; G[p_1\!*] \; yz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}
{G[p_1\!*] \; xyz \stackrel{\mbox{\tiny{PEG}}}{\leadsto} X}} \mylabel{rep.2}
\;\;\;
\\ \\
&
\fivespaces\fivespaces \;\;\;\;\; \;\;\;\;\; \;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{l}{y}}
{G[p_1\!*] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{l}{y}}}, l \neq {\tt fail} \mylabel{rep.3}
\\ \\
& \textbf{Negative Predicate} \;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{y}}
{G[!p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} xy}} \mylabel{not.1}
\;\;\;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} y}
{G[!p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{{\tt fail}}{xy}}} \mylabel{not.2}
\;\;\;
\\ \\
&
\fivespaces\fivespaces \fivespaces\fivespaces \fivespaces\fivespaces \;\;\;\;\; \;
{\frac{G[p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{l}{y}}
{G[!p_1] \; xy \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{l}{y}}}, l \neq {\tt fail} \mylabel{not.3}
\\ \\
& \textbf{Throw} \;\;\;
{\frac{}{G[\Uparrow^{l}] \; x \stackrel{\mbox{\tiny{PEG}}}{\leadsto} \Tup{l}{x}}} \mylabel{throw.1}
\end{align*}
}
\caption{Updated Semantics of PEGs with Labels}
\label{fig:semlabels}
\end{figure}
Although the semantics of PEGs with labels presented in
Figure~\ref{fig:sem} allows us to generate specific error
messages, it does not give us information about the location
where the failure probably is, so it is necessary to use some
extra mechanism (e.g., semantic actions) to get this information.
To avoid this, we can adapt the semantics of PEGs with labels
to give us a tuple $\Tup{l}{y}$ in case of a failure, where
$y$ the suffix of the input that PEG was trying to match
when label $l$ was thrown. Figure~\ref{fig:semlabels} presents
this updated semantics.
In the next section, we try to establish a comparison between
the farthest failure position heuristic and the labeled failure mechanism
by contrasting two different implementations of a parser for a dialect of
the Lua language.
\section{Labeled Failures versus Farthest Failure Position}
\label{sec:labelsfft}
In this section we will compare two parser implementations for
the Typed Lua language, one that uses the farthest failure
position heuristic for error reporting, which was implemented
first, and one based on labeled failures.
Typed Lua~\cite{maidl2014typedlua} is an optionally-typed
extension of the Lua programming language~\cite{lua}.
The Typed Lua parser recognizes plain Lua programs, and
also Lua programs with type annotations. The first version
of the parser was implemented using Ford's
heuristic and the LPeg library~\footnote{The first version of Typed Lua parser is available at
\url{https://github.com/andremm/typedlua/blob/master/typedlua/tlparser.lua}}.
As LPeg does not have a native error reporting mechanism based
on Ford's strategy, the failure tracking heuristic was implemented
following the approach described in Section~\ref{sec:pegs},
which uses semantic actions.
Below we have the example of a Lua statement with a syntax error:
\begin{verbatim}
a = function (a,b,) end
\end{verbatim}
In this case, the parser gives us the following error message,
which is quite precise:
\begin{verbatim}
test.lua:1:19: syntax error, unexpected ')', expecting '...', 'Name'
\end{verbatim}
In the previous case, the list of expected tokens had only two
candidates, but this is not always the case. For example, let us
consider the following Lua program, where there is no expression
after the \textbf{elseif} in line 5:
\begin{verbatim}
01 if a then
02 return x
03 elseif b then
04 return y
05 elseif
06
07 end
\end{verbatim}
The corresponding error message has a lengthy list of tokens,
which does not help much to fix the error:
\begin{verbatim}
test.lua:7:1: syntax error, unexpected 'end', expecting '(', 'Name', '{',
'function', '...', 'true', [9 more tokens]
\end{verbatim}
When using the Typed Lua parser based on Ford's heuristic it is not
uncommon to get a message like this. An analysis of the test cases
available in the parser package shows us that around half of the
expected error messages have a list of at least eight expected tokens
(there were messages with a list of 39 expected tokens).
The second implementation of the Typed Lua parser was
based on labeled failures and used the LPegLabel library~\cite{lpeglabel},
which is an extension of the LPeg library that supports
labeled failures~\footnote{The second version of Typed Lua parser is available at
\url{https://github.com/sqmedeiros/lpeglabel/tree/master/examples/typedlua}}.
The use of labeled failures implies in an annotation burden to
specify when each label should be thrown. In the case of Typed Lua grammar,
we defined almost 50 different labels, using the same basic strategy
that we used to annotate the Tiny grammar of Section~\ref{sec:lf}.
Given the previous Lua program, the error message presented now is
the following one:
\begin{verbatim}
test.lua:7:1: expecting <exp> after 'elseif'
\end{verbatim}
This error message is simpler than the previous one,
which was generated automatically. We analyzed the error
messages generated by the two parsers in 53 examples,
and considered that in more than half of these examples
the parser based on labeled failures produced a better
error message. In about 20\% of the cases we considered
the error messages of both approaches similar, and in other
20\% of the cases the parser based on Ford's heuristic
generated better error messages.
The error location indicated by the two parsers in the examples
analyzed was essentially the same. This seems to indicate that
the main difference in practice between both approaches is related
to the length of the error message generated. By using labeled
failures we can probably get a simple error message at the cost
of annotating the grammar, while by using the farthest failure
tracking heuristic we can automatically generate error messages,
which sometimes may contain a long list of expected tokens.
A point that is worth mentioning about the labeled failure approach
is that is not mandatory to annotate the entire grammar. The grammar
can be annotated incrementally, at the points where the current error
message is not good enough, and when no specific label is thrown, i.e.,
when the label \texttt{fail} is thrown, an error message can be
generated automatically by using the position where the failure
occurred. This means that combining labeled failures with the farthest
failure position reduces the annotation burden, and helps
identify the places in the parser where a label may be necessary.
Like PEGs, parsers written using parser combinators also finish with success
or failure and usually backtrack in case of failure,
making it difficult to implement error reporting. In the next
section we discuss some applications of labeled failures:
we can use PEGs with labeled failures to express the error
reporting techniques that we have discussed in
Section~\ref{sec:rel}~\cite{hutton1992hfp,rojemo1995epc,partridge1996fv,leijen2001parsec}.
We also show how we can use labeled failures to efficiently parse
context-free grammars that can use the
$LL(*)$ parsing strategy~\cite{parr2011llstar} with
a labeled PEG.
\section{Applications of Labeled Failures}
\label{sec:labelsrelated}
This section shows that PEGs with labeled failures can express
several error reporting techniques used in the realm of parsing
combinators. They can also efficiently parse
context-free grammars that
are parseable by the $LL(*)$ top-down parsing strategy.
In Hutton's deterministic parser combinators~\cite{hutton1992hfp}, the \texttt{nofail}
combinator is used to distinguish between failure and error.
We can express the \texttt{nofail} combinators using PEGs with labels
as follows:
\begin{align*}
{\tt nofail} \; p & \; \equiv \; p \; / \; \Uparrow^{\tt error}
\end{align*}
That is, \texttt{nofail} is an expression that transforms the failure
of $p$ into an error to abort backtracking.
Note that the \texttt{error} label should not be caught by any
ordered choice.
Instead, the ordered choice propagates this label and catches solely
the \texttt{fail} label.
The idea is that parsing should finish with one of the following
values: success, \texttt{fail} or \texttt{error}.
The annotation of the Tiny grammar to use \texttt{nofail} is
similar to the annotation we have done using labeled failures.
Basically, we just need to change the grammar to use
\texttt{nofail} instead of $[p]^{l}$.
For instance, we can write the rule \textit{CmdSeq} as follows:
\[
\it{CmdSeq} \leftarrow \it{(Cmd \; ({\tt nofail \; SEMICOLON})) \;
(Cmd \; ({\tt nofail \; SEMICOLON}))*}\\
\]
If we are writing a grammar from scratch, there is no advantage to
use \texttt{nofail} instead of more specific labels,
as the annotation burden is the same and with \texttt{nofail} we
lose more specific error messages.
The \texttt{cut} combinator~\cite{rojemo1995epc} was introduced
to reduce the space inefficiency of backtracking parsers, where
the possibility of backtracking means that any input that has already
been processed must be kept in memory until the end of parsing.
Semantically it is identical to {\tt nofail}, differing only
in the way the combinators are implemented: to implement {\tt cut}
the parser combinators use continuation-passing style, so {\tt cut}
can drop the failure continuation and consequently any pending
backtrack frames. Hutton's {\tt nofail} is implemented in direct style,
and is not able to drop pending backtrack frames.
Expressing a {\tt cut} operator with the same properties
is not possible in our semantics of PEGs.
The four-values technique changed the semantics
of parser combinators to implement predictive parsers for LL(1)
grammars that automatically identify the longest input prefix
in case of error, without needing annotations in the grammar.
We can express this technique using labeled failures by transforming
the original PEG with the following rules:
\begin{align}
\label{fv:e}
\llbracket \varepsilon \rrbracket & \equiv \; \Uparrow^{\tt epsn}\\
\label{fv:a}
\llbracket a \rrbracket & \equiv \; a\\
\label{fv:A}
\llbracket A \rrbracket & \equiv \; A\\
\label{fv:bind}
\llbracket p_1 p_2 \rrbracket & \equiv \;
\llbracket p_1 \rrbracket \;
(\llbracket p_2 \rrbracket \; / \Uparrow^{\tt error} \; /^{\{{\tt epsn}\}} \;
\varepsilon) \; /^{\{{\tt epsn}\}} \; \llbracket p_2 \rrbracket\\
\label{fv:choice}
\llbracket p_1 / p_2 \rrbracket & \equiv \;
\llbracket p_1 \rrbracket \; /^{\{{\tt epsn}\}} \;
(\llbracket p_2 \rrbracket \; / \Uparrow^{\tt epsn}) \; / \;
\llbracket p_2 \rrbracket
\end{align}
This translation is based on three labels:
\texttt{epsn} means that the expression successfully finished
without consuming any input,
\texttt{fail} means that the expression failed without consuming
any input,
and \texttt{error} means that the expression failed after
consuming some input.
In our translation we do not have an \texttt{ok} label because a
resulting suffix means that the expression successfully finished
after consuming some input.
It is straightforward to check that the translated expressions
behave according to the Table \ref{tab:fv} from Section \ref{sec:rel}.
Parsec introduced the \texttt{try} combinator to annotate parts of the
grammar where arbitrary lookahead is needed.
We need arbitrary lookahead because PEGs and parser combinators usually
operate on the character level.
The authors of Parsec also showed a correspondence between the semantics
of Parsec as implemented in their library and Partridge and Wright's
four-valued combinators,
so we can emulate the behavior of Parsec using labeled failures by
building on the five rules above and adding the following rule for
\texttt{try}:
\begin{align}
\label{fv:try}
\llbracket {\tt try} \; p \rrbracket & \equiv
\llbracket p \rrbracket \; /^{\{{\tt error}\}} \; \Uparrow^{\tt fail}
\end{align}
If we take the Tiny grammar of Figure \ref{fig:tiny} from
Section \ref{sec:pegs}, insert \texttt{try} in the necessary places,
and pass this new grammar through the transformation
$\llbracket\;\rrbracket$, then we get a PEG that automatically
identifies errors in the input with the \texttt{error} label.
For instance, we can write the rule \textit{RepeatCmd} as follows:
\begin{align*}
\it{RepeatCmd} & \leftarrow \it{({\tt try \; REPEAT}) \; CmdSeq \;
{\tt UNTIL} \; Exp}
\end{align*}
$LL(*)$~\cite{parr2011llstar} is a parsing strategy used by
the popular parsing tool ANTLR~\cite{parr2013antlr,antlrsite}~\footnote{The
recently released version $4$ of ANTLR uses {\em adaptive}
$LL(*)$
as its parsing strategy.}. An $LL(*)$ parser is a top-down
parser with arbitrary lookahead. The main idea of $LL(*)$
parsing is to build a deterministic finite automata for
each rule in the grammar, and use this automata to predict
which alternative of the rule the parser should follow,
based on the rest of the input. Each final state of
the DFA should correspond to a single alternative, or
we have an $LL(*)$ parsing conflict.
Mascarenhas et al.~\cite{mascarenhas2014} shows how
CFG classes that correspond to top-down predictive parsing
strategies can be encoded with PEGs by using predicates to
encode the lookahead necessary for each alternative.
As translating a Deterministic Finite Automata (DFA) to a
PEG is straightforward~\cite{ierusalimschy2009lpeg,mascarenhas2014},
this gives us one way of encoding an $LL(*)$ parsing strategy in
a PEG, at the cost of encoding a different copy of the
lookahead DFA for each alternative.
Labeled PEGs provide a more straightforward encoding,
where instead of a predicate for each alternative,
we use a single encoding of the lookahead DFA, where each
final state ends with a label corresponding to one
of the alternatives. Each alternative is preceded by
a choice operator that catches its label.
\begin{figure}[t]
\begin{center}
\begin{tikzpicture}[shorten >=1pt,node distance=2.9cm,on grid,auto]
\node[state,initial] (s_0) {$s_0$};
\node[state] (s_1) [above right=of s_0] {$s_1$};
\node[state,accepting] (s_4) [above right=of s_1] {$s_4 \rightarrow 2$};
\node[state,accepting] (s_6) [ right=of s_1] {$s_6 \to 1$};
\node[state] (s_2) [ right=of s_0] {$s_2$};
\node[state,accepting](s_5) [ below=of s_6] {$s_5 \rightarrow 4$};
\node[state,accepting](s_3) [below=of s_5] {$s_3 \rightarrow 3$};
\path[->]
(s_0) edge node {ID} (s_1)
edge[below] node {unsigned} (s_2)
edge[below] node {int} (s_3)
(s_1) edge node {=} (s_4)
edge node {EOF} (s_6)
edge node {ID} (s_5)
(s_2) edge node {ID} (s_5)
edge node {int} (s_3)
edge [loop below] node {unsigned} ();
\end{tikzpicture}
\end{center}
\caption{$LL(*)$ lookahead DFA for rule $S$}
\label{fig:antlrdfa}
\end{figure}
To make the translation clearer, let us consider the following
example, from Parr and Fisher~\cite{parr2011llstar},
where non-terminal $S$ uses non-terminal $Exp$ (omitted) to match
arithmetic expressions:
\[
\it{S} \;\rightarrow\; {\tt ID} \;\,|\,\; \tt{ID} \,`{\tt =}\textrm'\, {\it Exp} \;\,|\;\,
`{\tt unsigned}\textrm' \; `\tt{*}\textrm' \, `{\tt int}\textrm'\; {\tt ID} \;\,|\;\,
`{\tt unsigned}\textrm' \;`\tt{*}\textrm'\, {\tt ID}\; {\tt ID}
\]
After analyzing this grammar, ANTLR produces the DFA
of Figure~\ref{fig:antlrdfa}. When trying to match $S$,
ANTLR runs this DFA on the input until it reaches a
final state that indicates which alternative of the
choice of rule $S$ should be tried. For example,
ANTLR chooses the second alternative if the DFA
reaches state $s_4$.
Figure~\ref{fig:pegllstar} gives a labeled PEG that encodes
the $LL(*)$ parsing strategy for rule $S$. Rules $S_0$,
$S_1$, and $S_2$ encode the lookahead DFA of Figure~\ref{fig:antlrdfa}, and correspond to states
$s_0$, $s_1$, and $s_2$, respectively. The throw
expressions correspond to the final states.
As the throw expressions make the input backtrack to where
it was prior to parsing $S_0$, we do not need to use a
predicate. We can also turn any uncaught failures into
errors.
\begin{figure}[t]
\begin{align*}
\it{S} & \rightarrow \it{S_0} \;/^{1}\; \tt{ID} \;/^{2}\;
\tt{ID} \,`\tt{=}\textrm'\, \it{Exp} \;/^{3}\;
`\tt{unsigned}\textrm' \; `\tt{*}\textrm' \; `\tt{int}\textrm' \; \tt{ID} \;/^{4}\;
`\tt{unsigned}\textrm' \; `\tt{*}\textrm' \; \tt{ID} \; \tt{ID}
\;/\; \Uparrow^{\tt error} \\
S_0 & \rightarrow \tt{ID} \, \it{S_1} \;/\; `\tt{unsigned}\textrm' \, \it{S_2} \;/\; `\tt{int}\textrm' \Uparrow^3 \\
S_1 & \rightarrow \,`\tt{=}\textrm' \Uparrow^2 \;/\; !. \Uparrow^1 \;/\; \tt{ID} \Uparrow^4 \\
S_2 & \rightarrow `\tt{unsigned}\textrm' \, \it{S_2} \;/\; \tt{ID} \Uparrow^4 \;/\; `\tt{int}\textrm' \Uparrow^3
\end{align*}
\caption{PEG with Labels that Simulates the $LL(*)$ Algorithm}
\label{fig:pegllstar}
\end{figure}
\section{Conclusions}
\label{sec:conc}
In this paper, we discussed error reporting strategies for
Parsing Expression Grammars. PEGs behave badly on the presence
of syntax errors, because backtracking often makes the PEG
lose track of the position where the error happened.
This limitation was already known by Ford, and he
tried to fix it in his PEG implementation by having
the implementation track the farthest position in the
input where a failure has happened~\cite{ford2002packrat}.
We took Ford's failure tracking heuristic and showed that
it is not necessary to modify a PEG implementation to
track failure positions as long as the implementation
has mechanisms to execute semantic actions, and the
current parsing position is exposed to these actions.
Nevertheless, we also showed how it is easy to extend
the semantics of PEGs to incorporate failure tracking,
including information that can indicate what the
PEG was expecting when the failure happened.
Tracking the farthest failure position, either by
changing the PEG implementation, using semantic actions,
or redefining the semantics of PEGs, helps PEG parsers
produce error messages that are close to error messages
that predictive top-down parsers are able to produce,
but these are generic error messages, sometimes with
a long list of expected tokens.
As a way of generating more specific error messages,
we introduced a mechanism of labeled failures to PEGs.
This mechanism closely resembles standard exception
handling in programming languages. Instead of a
single kind of failure, we introduced a \emph{throw} operator
$\Uparrow^{l}$ that can throw different kinds of failures,
identified by their labels, and extended the ordered choice
operator to specify the set of labels that it catches.
The implementation of these extensions in parser generator
tools based on PEGs is straightforward.
We showed how labeled failures can be used as a way to
annotate error points in a grammar, and tie them to more
meaningful error messages. Labeled failures are orthogonal
to the failure tracking approach we discussed earlier, so
grammars can be annotated incrementally, at the points
where better error messages are judged necessary.
We also showed that the labeled failures approach can express
several techniques for error reporting used in parsers
based on deterministic parser combinators, as
presented in related
work~\cite{hutton1992hfp,rojemo1995epc,partridge1996fv,leijen2001parsec}.
Labeled failures can also be used as a way of encoding
the decisions made by a predictive top-down parser, as long
as the decision procedure can be encoded as a PEG, and showed
an example of how to encode an $LL(*)$ grammar in this way.
Annotating a grammar with labeled failures demands care:
if we mistakenly annotate expressions that should be able to
fail, this modifies the behavior of the parser beyond error
reporting. In any case, the use of labeled PEGs for error
reporting introduces an annotation burden that is
lesser than the annotation burden introduced by error
productions in LR parsers, which also demand care, as
their introduction usually lead to \textit{reduce-reduce}
conflicts~\cite{jeffery2003lre}.
We showed the error reporting strategies in the context
of a small grammar for a toy language, and we also discussed
the implementation of parsers for the Typed Lua language,
an extension of the Lua language, based on these strategies.
Moreover, we also implemented parsers for other languages,
such as Céu~\cite{ceu}, based on theses approaches,
improving the quality of error reporting either with
generic error messages or with more specific error messages.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
High-energy collisions near black holes attract much interest both from
theoretical and astrophysical viewpoints. This issue was started in \cit
{pir1} - \cite{pir3} and revived after an important observation made in \cit
{ban}. It turned out that the energy $E_{c.m.}$ in the center of mass frame
can grow unbound when two particles moving towards a black hole collide near
the horizon. In doing so, a black hole has to be rotating for the effect to
take place. This is so-called the Ba\~{n}ados - Silk \ - West effect (the
BSW effect, after the names of its authors). Meanwhile, there are
alternative mechanisms connected with the presence of the electromagnetic
field. For the electrically charge nonrotating black hole the similar effect
was obtained in \cite{jl}. The counterpart of the BSW effect due to the
magnetic field in the Schwarzschild background was proposed in \cite{fr}
that gave rise to several works \cite{weak} - \cite{isco} in which more
general metrics were considered. All these works on collisions in the
magnetic field share the common \ feature: for getting large $E_{c.m.}$,
collisions should occur near the horizon similarly to the BSW effect.
From the other hand, there exists one more scenario of getting $E_{c.m.}$
proposed in the paper \cite{ergo} (generalized in \cite{mergo}). It consists
in such collisions when one particle carries a large negative angular
momentum. It turned out that such a scenario can be realized away from the
horizon but inside the ergosphere (or on its boundary) only. The goal of the
present work is to suggest one more mechanism that is in a sense some hybrid
of previous ones. It implies that one of particles has a large negative
angular momentum $\left\vert L\right\vert $ and collisions occurs in the
magnetic field. It turns out that if both $\left\vert L\right\vert $ and the
magnetic field strength are large but their ratio is finite, large $E_{c.m.}$
can be achieved outside the ergoregion. Moreover, under some restrictions,
collisions may happen even in the asymptotically flat region, where
deviation of the black hole metric from the Minkowski one are small.
Actually, this means that effect can be obtained even in the star-like
background when the horizon is absent and is far from the threshold of its
formation. This simplifies greatly the conditions of observation since there
are no objections raised in \cite{mc} (see also discussion in \cite{com})
because of proximity to the horizon and concerning the redshift that "eats"
the gain from large $E_{c.m.}$ when debris of collisions reach a remote
near-flat region.
Throughout the paper, we put fundamental constants $G=c=1,$ except some
cases when we write down $G$ explicitly.
\section{Metric and equations of motion}
Let us consider the metric of the form
\begin{equation}
ds^{2}=-N^{2}dt^{2}+\frac{dr^{2}}{A}+R^{2}(d\phi -\omega dt)^{2}+g_{\theta
}d\theta ^{2}\text{,} \label{met}
\end{equation
where the coefficients do not depend on $t$ and $\phi $. The horizon
corresponds to $N=0$. We also assume that there is an electromagnetic field
with the four-vector $A^{\mu }$ where the only nonvanishing component is
equal t
\begin{equation}
A^{\phi }=\frac{B}{2}\text{.} \label{ab}
\end{equation}
In vacuum, this is an exact solution on the Kerr background with $B=const$
\cite{wald}, \cite{ag}. However, we discuss a more general situation and
only afterwards substitute the quantities for the Kerr metric.
Let us consider motion of test particles in this background. The kinematic
momentum $p^{\mu }=mu^{\mu }$, where $m$ is the particle's mass,
four-velocity $u^{\mu }=\frac{dx^{\mu }}{d\tau }$, where $\tau $ is the
proper time, $x^{\mu }$ are coordinates. Then, the kinematic momentum
p_{\mu }$ and the generalized one $P_{\mu }$ are related according to
\begin{equation}
p_{\mu }=P_{\mu }-qA_{\mu }\text{,}
\end{equation
$q$ is the particle's electric charge. Due to the symmetry of the metric,
P_{0}=-E$ and $P_{\phi }=L$ are conserved, where $E$ is the energy, $L$ is
the angular momentum.
We assume that the equatorial plane is a symmetry one and consider motion
constrained within this plane, so $\theta =\frac{\pi }{2}$. Redefining the
new radial coordinate, we can always achieve that
\begin{equation}
A=N^{2} \label{an}
\end{equation
within this plane. Then, equations of motion give us
\begin{equation}
\dot{t}=\frac{X}{N^{2}m}\text{,}
\end{equation
\begin{equation}
\dot{\phi}=\frac{\beta }{R}+\frac{\omega X}{mN^{2}}\text{,} \label{ft}
\end{equation
\begin{equation}
X=E-\omega L\text{,} \label{X}
\end{equation
\begin{equation}
\beta =\frac{L-L_{0}}{mR}\text{,} \label{b}
\end{equation
\begin{equation}
L_{0}=\frac{qBR^{2}}{2}\text{,} \label{l0}
\end{equation
\begin{equation}
m^{2}\dot{r}^{2}=Z^{2}\text{,} \label{rt}
\end{equation
\begin{equation}
Z^{2}=X^{2}-m^{2}N^{2}(1+\beta ^{2})\text{.} \label{v}
\end{equation
Dot denotes differentiation with respect to the proper time $\tau $. As
usual, we assume the forward in time condition $\dot{t}>0$, so
\begin{equation}
X\geq 0. \label{xt}
\end{equation}
\section{Particle collisions: general formulas}
Let two particles collide. We label their characteristics by indices 1 and
2. Then, in the point of collision, one can define the energy in the centre
of mass frame a
\begin{equation}
E_{c.m.}^{2}=-p_{\mu }p^{\mu }=m_{1}^{2}+m_{2}^{2}+2m_{1}m_{2}\gamma \text{.}
\end{equation}
Here
\begin{equation}
p^{\mu }=m_{1}u_{1}^{\mu }+m_{2}u_{2}^{\mu }
\end{equation
is the total momentum
\begin{equation}
\gamma =-u_{1\mu }u_{2}^{\mu }
\end{equation
is the Lorentz factor of their relative motion.
For motion in the equatorial plane in the external magnetic field (\ref{ab
), one finds from the equations of motion (\ref{ft}), (\ref{rt}) tha
\begin{equation}
\gamma =\frac{X_{1}X_{2}-\varepsilon _{1}\varepsilon _{2}Z_{1}Z_{2}}
m_{1}m_{2}N^{2}}-\beta _{1}\beta _{2}\text{.} \label{ga}
\end{equation}
Here, $\varepsilon =+1$, if the particle moves away from the horizon and
\varepsilon =-1$, if it moves towards it.
To get large $\gamma $ and $E_{c.m.}$, there are two ways - either to
decrease the denominator or to increase the numerator. The first way implies
that collision should occur near the horizon, where $N\ll 1$. The second one
corresponds to collisions with large $\left\vert L\right\vert $, when the
numerator becomes large as well. In the absence of the magnetic field, such
a scenario can be realized inside the ergosphere or on its boundary only
\cite{ergo}, \cite{mergo}. Now, we will show that the magnetic field strong
enough extends this scenario outside the ergosphere.
To make the issue nontrivial, we are interested in the case when large
E_{c.m.}$ can be obtained with finite individual energies $E_{1}$ and $E_{2}$
only.
\section{Large B and L}
Let us try to achieve large $\gamma $ at the expense of large $L$. We assume
that $\omega >0$, like for the Kerr metric. Then, it follows from (\ref{X}),
(\ref{xt}) that if the angular momentum is large it should be negative,
\begin{equation}
L=-\left\vert L\right\vert \text{.}
\end{equation
Let formally $\left\vert L\right\vert \rightarrow \infty $. This means that
in (\ref{X}),
\begin{equation}
\omega \left\vert L\right\vert \gg E\text{.} \label{lx}
\end{equation}
Then, the condition $Z^{2}>0$ in (\ref{v}) gives rise to the inequality \
\begin{equation}
\frac{\omega R}{N}\left\vert L\right\vert >\left\vert \left\vert
L\right\vert +L_{0}\right\vert . \label{neq}
\end{equation}
If $L_{0}>0$, this entails
\begin{equation}
\left\vert L\right\vert (\frac{\omega R}{N}-1)>\left\vert L_{0}\right\vert
\text{,}
\end{equation
whenc
\begin{equation}
\omega R>N\text{.} \label{er}
\end{equation
According to (\ref{met})
\begin{equation}
g_{00}=-N^{2}+\omega ^{2}R^{2}\text{.}
\end{equation
$.$
Thus we have from (\ref{er}) that $g_{00}>0$. In other words, this is just
the ergoregion, so we return to the situation considered in \cite{ergo},
\cite{mergo}.
Let now $L_{0}<0.$
a) $\left\vert L\right\vert >\left\vert L_{0}\right\vert $
If we write
\begin{equation}
\left\vert L\right\vert =\xi \left\vert L_{0}\right\vert , \label{ksi}
\end{equation
we obtain from (\ref{neq}) tha
\begin{equation}
\xi >1\text{, }\xi (1-\frac{\omega R}{N})<1\text{.} \label{a}
\end{equation}
b) $\left\vert L\right\vert \leq \left\vert L_{0}\right\vert $
In a similar way
\begin{equation}
\frac{1}{1+\frac{\omega R}{N}}<\xi \leq 1\text{.} \label{br}
\end{equation}
In both cases a) and b), the sign of the quantity $N-\omega R$ can be
arbitrary, so motion is possible inside or outside the ergoregion.
If the metric is asymptotically flat, $N\rightarrow 1$ and $\omega
\rightarrow 0$ when $R\rightarrow \infty $. We assume that, moreover,
\omega R\rightarrow 0$ (like in the case of the Kerr metric - see below).
Then, it is seen from (\ref{a}), (\ref{br}) that motion with large $L$ (\re
{ksi}) in this region is possible if $\xi \approx 1$ only. This means that
the angular momentum is fine-tuned according to (\ref{l0}). In this sense,
the value $\xi =1$ is the analogue of the critical trajectory in the BSW
effect \cite{ban}, \cite{prd}. However, for finite $r$, any $\xi $ obeying
\ref{a}) or (\ref{br}) is suitable.
\section{Behavior of Lorentz factor}
Let particle 1 have an angular momentum (\ref{ksi}) with $\left\vert
L_{0}\right\vert \rightarrow \infty $. Then, we obtain from (\ref{ga}
\begin{equation}
\gamma \approx \left\vert L_{0}\right\vert [\frac{X_{2}\omega \xi
-\varepsilon _{1}\varepsilon _{2}Z_{2}\sqrt{\omega ^{2}\xi ^{2}-\frac{N^{2}}
R^{2}}(1-\xi )^{2}}}{m_{1}m_{2}N^{2}}-\frac{\beta _{2}(1-\xi )}{Rm}]\text{.}
\end{equation}
As we will be interested in large $r$, we put $\xi =1$. Let us also take for
simplicity $q_{2}=0$, $m_{1}=m_{2}=m=E_{2}\,$, $L_{2}=0$. Then,
\begin{equation}
\gamma \approx \frac{\left\vert L_{0}\right\vert }{m^{2}N^{2}}\omega
(E_{2}-\varepsilon _{1}\varepsilon _{2}\sqrt{E_{2}^{2}-m^{2}N^{2}})
\end{equation}
Assuming that $E_{2}\sim m$ and omitting the factor of the order unity, we
can writ
\begin{equation}
\gamma \sim \frac{\left\vert L_{0}\right\vert \omega }{mN^{2}}\text{.}
\label{gam}
\end{equation}
\section{Kerr metric}
Now, we apply the above results to the Kerr metric using eq. (\ref{gam}).
In terms of the Boyer-Lindquiste coordinate $r$, the coefficients of the
Kerr metric in the form (\ref{met}) rea
\begin{equation}
\frac{\omega }{N^{2}}=\frac{2Ma}{r(r^{2}-2Mr+a^{2})}\text{,}
\end{equation
\begin{equation}
R^{2}=r^{2}+a^{2}+\frac{2M}{r}a^{2}\text{,}
\end{equation
where $M$ is the black hole mass, $a=\frac{J}{M}$, $J$ is its angular
momentum. One should take into account that in the point of observation,
L_{0}$ is given by (\ref{l0}). Considering large $r\gg 2M$, we obtai
\begin{equation}
\gamma \approx \frac{ab}{r}\text{,}
\end{equation
\begin{equation}
b\equiv \frac{\left\vert qB\right\vert GM}{m}\text{,} \label{bq}
\end{equation
where we restored the gravitational constant $G$.
To achieve $\gamma \gg 1$, we must hav
\begin{equation}
2GM\ll r\ll Gab\text{,} \label{abr}
\end{equation
so
\begin{equation}
b\gg 1\text{.} \label{bg}
\end{equation}
One can check that, under the condition (\ref{abr}), the condition (\ref{lx
) is satisfied automatically.
If $b$ is large enough, the second inequality in (\ref{abr}) is not
restrictive. To achieve large $\gamma $, particle 1 should have $\left\vert
L\right\vert \sim \left\vert L_{0}\right\vert $. In terms of $b$,
\begin{equation}
L_{0}=bm\frac{R^{2}}{2M}\text{.}
\end{equation
For $R\gtrsim 2M$, $L_{0}\gtrsim 2mMb$. Thus its value must increase by the
factor $b$ as compared to the value of the order $2mM$ typical of motion of
test particles within the Kerr background. Depending on $M$ and $B$, one can
choose $b$ big enough to get $\gamma \gg 1$ but, at the same time, make
\frac{L_{0}}{2mM}$ not tremendous. Now, we have an additional parameter $B$
that is absent in the case of collisions without the magnetic field. Varying
$B$, one can hope to facilitate realization of the scenario under discussion.
\section{Discussion amd conclusion}
In the most part of works on the BSW effect or its modification, unbound
E_{c.m.}$ were obtained due to the proximity of collision to the event
horizon or (if the horizon is absent) due to proximity of the would-be
horizon to the threshold of its formation. A step away from this requirement
was made in \cite{ergo}, where collision could take place not in the
vicinity of the horizon. However, this implied the region of strong
gravitational field since the corresponding mechanism works in the
ergoregion (or, in the ultimate case, on its boundary). In this sense, the
region of strong gravitational field was still needed. In the present work,
we made, in a sense, one more step away from the horizon and showed that
even the ergosphere is not necessary. Moreover, collision can occur in the
region, where gravitational field is weak and, nonetheless, we can achieve
\gamma \gg 1$. From the other hand, we cannot simply take the limit
G\rightarrow 0$ since the conditions (\ref{abr}), (\ref{bg}) should be
satisfied and, additionally, $G$ enters the definition of $b$ (\ref{bq}). In
other words, the gravitational field causes small (but essentially nonzero)
perturbation of the Minkowski metric but affects motion of test particles
significantly. In a similar way, the magnetic field is also weak (since it
almost does not change the metric) and strong (since it is important for
test particles) simultaneously in different aspects \cite{fr}.
It is worth also noting that in the previous scenarios of high energy
collisions in the magnetic and gravitational fields \cite{fr}, \cite{weak}
either collision occurred near the innermost stable circular orbit (ISCO) or
one of particles had the parameters typical of such an orbit.
Correspondingly, motion on ISCO posed restriction on the relationship
between $E$ and $L$. In our case, we considered collision with arbitrary $E$
and $L$ which are independent quantities. The energy $E$ is finite but $L$
should be negative and large enough with $\left\vert L/B\right\vert $ being
finite. (The situation when $E$ is finite but $L$ grows with $B$ is typical
also for collisions near the magnetized Schwarzschild black hole \cite{fr}
and, in some cases, also for the Kerr black hole \cite{weak}.)
The fact that large $\gamma $ can be achieved far from a black hole,
actually means that instead of a black hole we can take a rotating star, so
the background may look unlike a black hole but, nonetheless, produce the
effect under discussion.
That collision can occur far from the black hole, gives hope that
high-energy collisions in gravitational and magnetic fields (each of which
is strong in one sense and weak in another one) can represent not only
theoretical interest but can be relevant for relativistic astrophysics.
|
\subsection{Notation}
For a graph $G$, $V(G)$ is the vertex set of $G$, and $|G|=|V(G)|$. For a subset $U\subset V(G)$, let $G[U]$ be the subgraph of $G$ induced on the set $U$. Given two disjoint sets $X$ and $Y$, $e_G(X,Y)$ is the number of edges between $X$ and $Y$. For a subset $U\subset V(G)$ and vertices $x,y\in U$, $d_U(x,y)$ is the length of a shortest $x,y$-path in $G[U]$, if such a path exists. The set of neighbours of a vertex $v$ is denoted by $N(v)$, and the neighbourhood of a vertex $v$ in the set $A\subset V(G)$ is $N(v,A)=N(v)\cap A$. The neighbourhood of a vertex set $S\subset V(G)$ is $N(S)=(\cup_{v\in S}N(v))\setminus S$ and $N(S,A)=N(S)\cap A$. Where multiple graphs are involved we use $N_G(v,A)$ to emphasise the graph currently considered. We say a path with $l$ vertices has \emph{length} $l-1$, and call a path $P$ an $x,y$-path if the vertices $x$ and $y$ have degree 1 in $P$.
Given two disjoint vertex sets $A$ and $B$, a \emph{$d$-matching from $A$ into $B$} is a collection of disjoint sets $\{X_a\subset N(a,B):a\in A\}$ so that, for each $a\in A$, $|X_a|=d$. As is well known, such matchings can be found by showing that Hall's generalised matching condition holds. For details on this, and other standard notation, see Bollob\'as~\cite{bollo1}.
\subsection{Graph Expansion and Trees}
We will establish graph expansion properties in the random graph, before using the properties to embed the trees.
\begin{defn} Let $n\in \mathbb{N}$ and $d>0$. A graph $G$ is an \emph{$(n,d)$-expander} if $|G|=n$ and $G$ satisfies the following two conditions.
\begin{enumerate}
\item $|N_G(X)|\geq d|X|$ for all sets $X\subset V(G)$ with $1\leq |X|<\lceil \frac{n}{2d}\rceil$.
\item $e_G(X,Y)>0$ for all disjoint $X,Y\subset V(G)$ with $|X|=|Y|=\lceil\frac{n}{2d}\rceil$.
\end{enumerate}
\end{defn}
\begin{defn}
Let $\mathcal{T}(n,\Delta)$ be the class of all trees on $n$ vertices with maximum degree at most $\Delta$.
\end{defn}
As shown by Balogh, Csaba, Pei and Samotij~\cite{BCPS10}, almost spanning trees can be found in expander graphs using a theorem by Haxell~\cite{PH01}. We will use the following formulation due to Johannsen, Krivelevich and Samotij~\cite{JKS12}.
\begin{theorem}\label{almost}
Let $n,\Delta\in \mathbb{N}$, let $d\in \mathbb{R}^+$ with $d\geq 2\Delta$ and let $G$ be a $(n,d)$-expander. Given any $T\in \mathcal{T}(n-4\Delta\lceil\frac{n}{2d}\rceil,\Delta)$, we can find a copy of $T$ in the graph $G$.
\end{theorem}
For Section \ref{paths} we will also find the following formulation by Balogh, Csaba, Pei and Samotij useful~\cite{BCPS10}.
\begin{theorem}\label{almost2}
Let $\Delta,m,M\in \mathbb{N}$. If $H$ is a non-empty graph such that for every $X\subset V(H)$, if $0<|X|\leq m$, then $N_H(X)|\geq \Delta|X|+1$ and, if $|X|=m$, then $|N_H(X)|\geq 2\Delta m+M$, then $H$ contains every tree in $\mathcal{T}(M,\Delta)$.
\end{theorem}
For our constructions we wish to partition our expander graph so that all vertex subsets expand well into each vertex subset in the partition. The following partition lemma by Johanssen, Krivelevich and Samotij~\cite{JKS12} permits this. For convenience, we state a slightly stronger form of their lemma, using the following definition, but the proof follows identically.
\begin{defn}
For a graph $G$ and a set $W\subset V(G)$, we say $G$ \emph{$d$-expands} into $W$ if
\begin{enumerate}
\item $|N_G(X, W)|\geq d|X|$ for all $X\subset V(G)$ with $1\leq |X|<\lceil \frac{|W|}{2d}\rceil$, and,
\item $e_G(X,Y)>0$ for all disjoint $X,Y\subset V(G)$ with $|X|=|Y|=\lceil\frac{|W|}{2d}\rceil$.
\end{enumerate}
\end{defn}
\begin{lemma}[\cite{JKS12}]\label{splitexpand} There exists an absolute constant $n_0\in \mathbb{N}$ such that the following statement holds. Let $k,n\in \mathbb{N}$ and $d\in \mathbb{R}^+$ satisfy $n\geq n_0$ and $k\leq \log n$. Furthermore, let $m,m_1,\ldots,m_k\in \mathbb{N}$ satisfy $m=m_1+\ldots+m_k$ and let $d_i:=\frac{m_i}{5m}d$ satisfy $d_i\geq 2\log n$ for all $i\in \{1,\ldots,k\}$. Then, for any graph $G$ which $d$-expands into a vertex set $W$, with $|W|=m$, the set $W$ can be partitioned into $k$ disjoint sets $W_1,\ldots, W_k$ of sizes $m_1,\ldots,m_k$ respectively, such that, for each $i$, $G$ $d_i$-expands into $W_i$.
\end{lemma}
In order to find expansion properties in the binomial random graph we will use the following lemma by Johannsen, Krivelevich and Samotij~\cite{JKS12}.
\begin{lemma}\label{randomexpand} Let $d:\mathbb{N}\to\mathbb{R}^+$ satisfy $d\geq 3$. Then $\mathcal{G}(n,7dn^{-1}\log n)$ is almost surely an $(n,d)$-expander.
\end{lemma}
As mentioned in Section \ref{intro}, we will use the following lemma by Krivelevich~\cite{MK10} to divide $\mathcal{T}(n,\Delta)$ into trees with many leaves and trees with many disjoint bare paths.
\begin{lemma}\label{Kleaves} Let $k,l,n>0$ be integers. Let $T$ be a tree on $n$ vertices with at most $l$ leaves. Then $T$ contains a collection of at least $n/(k+1)-(2l-2)$ vertex disjoint bare paths of length $k$ each.
\end{lemma}
We will employ this lemma in the following form, letting $l=n/4k$.
\begin{corollary}\label{split} For any integers $n,k>2$, a tree on $n$ vertices either has at least $n/4k$ leaves or a collection of at least $n/4k$ vertex disjoint bare paths of length $k$ each.
\end{corollary}
We will also use the following lemma by Krivelevich~\cite{MK10} to almost surely attach leaves onto a partial embedding of a tree.
\begin{lemma}\label{matchymatchy} Let $0<d_1,\ldots,d_k$ be integers satisfying: $d_i\leq \Delta$, $\sum^k_{i=1}d_i=l$. Let $A=\{a_1,\ldots,a_k\}$, $B$ be disjoint sets of vertices with $|B|=l$. Let $G$ be a random bipartite graph with sides $A$ and $B$, where each pair $(a,b)$, $a\in A$, $b\in B$, is an edge of $G$ with probability $p$, independent of other pairs. If $p\geq 2\Delta\log l/ l$, then, almost surely as $l\to\infty$, the random graph $G$ contains a collection $S_1,\ldots, S_k$ of vertex disjoint stars such that $S_i$ is centered at $a_i$ and has the remaining $d_i$ vertices in $B$.
\end{lemma}
\subsection{A random graph construction}
For Theorem \ref{mainexpo1}, we require a bipartite graph with a certain `resilient matching' condition. Here we verify that such graphs exist.
\begin{lemma} \label{flexiblematching}
There is a constant $n_0$, such that for every $n\geq n_0$ with $3|n$, there exists a bipartite graph $H$ on vertex classes $X$ and $Y\cup Z$ with $|X|=n$, $|Y|=|Z|=2n/3$, and maximum degree $40$, so that the following is true. Given any subset $Z'\subset Z$ with $|Z'|=n/3$, there is a matching between $X$ and $Y\cup Z'$.
\end{lemma}
\begin{proof}
Let $m=2n/3$, and take two vertex sets $X_1$, $Y$ with $|X_1|=|Y|=2n/3$. Independently, place 20 random matchings between $X_1$ and $Y$ and let the graph $G$ be the union of these matchings. Given two sets $A\subset X_1$ and $B\subset Y$, and a random matching $M$, the probability that $N_M(A)\subset B$ is $\binom{|B|}{|A|}\binom{m}{|A|}^{-1}$. Thus, for each $t\leq m/4$, the probability there is some set $A\subset X_1$, $|A|=t$ with $|N_G(A)|<2t$, is at most
\[
\binom{m}{t}\binom{m}{2t}\left(\binom{2t}{t}\binom{m}{t}^{-1}\right)^{20}
\leq \left(\frac{e^3}{4}\left(\frac mt\right)^{3}\left(\frac{2t}{m}\right)^{20}\right)^t
= \left(\frac{8e^3}{4}\left(\frac{2t}{m}\right)^{17}\right)^t.
\]
Looking separately at the cases $t\leq \log m$, and $\log m\leq t\leq m/4$, we see that the sum of these probabilities tends to 0 as $m\to\infty$. An identical calculation shows that, almost surely, $|N(B)|\geq 2|B|$ for every $B\subset Y$ with $|B|\leq m/4$.
The probability there are no two sets $A\subset X_1$ and $B\subset Y$, with $|A|=|B|=\lceil m/4\rceil$ and $e_G(A,B)=0$ is at most
\[
\binom{m}{m/4}^2\left(\binom{3m/4}{m/4}\binom{m}{m/4}^{-1}\right)^{20}\leq (4e)^{m/2}\left(\frac{3}{4}\right)^{20m/4}\leq \left(\frac12\right)^{m/4}.
\]
Therefore, almost surely, all three of these properties mentioned hold. Let $n$, and hence $m$, be sufficiently large that some graph with these properties must exist, and fix such a graph $G$. Duplicate the vertices in $Y$ to get $Z$, and then duplicate $m/2$ of the vertices in $X_1$ to get the set $X_2$. Consider the bipartite graph $H$ on vertex sets $X=X_1\cup X_2$ and $Y\cup Z$. From its origins in 20 matchings, the maximum degree of $H$ is at most 40, despite the duplication of some vertices.
Now, take any set $Z'\subset Z$ such that $|Z'|=n/3$. We will verify Hall's matching condition between $X$ and $Y\cup Z'$ to show a matching exists, and demonstrate that $H$ satisfies the lemma. For $A\subset X$, let $A'$ be the larger set of $A\cap X_1$ and $A\cap X_2$, so that $|A|\geq|A'|\geq |A|/2$.
Suppose $|A'|\leq m/4$. Then $|N_H(A,Y\cup Z')|\geq|N_H(A',Y)|\geq 2|A'|\geq |A|$, by the properties of $G$.
Suppose $|A'|\geq m/4$ and $|A|\leq n-m/2$. Then there are no edges between $A'$ and $Y\setminus N_H(A')$, so $|Y\setminus N_H(A')|\leq m/4$ and $|Z\setminus N_H(A')|\leq m/4$, using the properties of $G$. Therefore, $|N_H(A',Y\cup Z')|\geq |Y\cup Z'|-m/2\geq |A|$.
Suppose finally then that $|A|\geq n-m/2$. By taking a subset of $A$ of size $n-m/4$ we can see from the previous paragraph that $|N_H(A,Y\cup Z')|\geq n-m/2$. Let $B=(Y\cup Z')\setminus N(A)$, and let $B'$ be the larger set of $B\cap Y$ and $B\cap Z'$, so that $|B'|\leq m/2$. Using similar arguments for $B'$ as we did for $A'$, we can show that $|N_H(B,X)|\geq |N_H(B',X_1)|\geq 2|B'|\geq |B|$. But $N_H(B,X)\subset X\setminus A$, so $|B|\leq|N_H(B,X)|\leq |X\setminus A|=n-|A|$. Therefore $|N_H(A,Y\cup Z')|=n-|B|\geq |A|$, as required.
\end{proof}
\subsection{Between many vertex pairs we can find one path}
\begin{lemma} \label{connect} Let $m,n\in \mathbb{N}$ satisfy $m\leq n/800$, let $d=n/200m$ and let $n$ be sufficiently large. Let the graph $G$ with $n$ vertices have the property that any set $A\subset V(G)$ with $|A|=m$ satifies $|N(A)|\geq (1-1/64)n$. Suppose $G$ contains disjoint vertex sets $X$, $Y$ and $U$, with $X=\{x_1,\ldots,x_{2m}\}$, $Y=\{y_1,\ldots, y_{2m}\}$ and $|U|=\lceil n/8\rceil$. Suppose, in addition, we have integers $k_i$, $i\in[2m]$, satisfying $2\log n/\log d\leq k_i\leq n/40$. Then, for some $i$, there is an $x_i,y_i$-path of length $k_i$ whose internal vertices lie in $U$.
\end{lemma}
\begin{proof}
Divide $U$ into two sets, $U_1$ and $U_2$, each of size at least $n/16$. Pick a largest subset $B\subset U_1$ such that $|B|\leq m$ and $|N(B,U_1)|<2d|B|$. Let $V_1=U_1\setminus B$. Suppose $A\subset V_1$ with $0<|A|\leq m$ has $|N(A,V_1)|<2d|A|$. Then $|N(A\cup B,U_1)|<2d(|A|+|B|)$, so by the choice of $B$ we must have that $|A\cup B|\geq m$. Let $C\subset A\cup B$ have size $|C|=m$. Then $|N(C)|\geq (1-1/64)n$. Since $|A\cup B|\leq 2m$, we have $|N(A\cup B)|\geq (1-1/64)n-m$, so that $|V(G)\setminus (N(A\cup B)\cup A\cup B)|\leq n/64+m$. We thus have,
\[
|N(A,V_1)|\geq |N(A\cup B,U_1)|-|N(B,U_1)|\geq |U_1|-n/64-m-2dm\geq 2dm.
\]
We have, then, that $|V_1|> n/16-m\geq n/20$ and every set $A\subset V_1$ with $|A|\leq m$ satisfies $|N_{V_1}(A)|\geq2d|A|$.
Similarly, find a set $V_2\subset U_2$ with the same expansion property, with $|V_2|\geq n/20$.
Now if $X'\subset X$ is a set of size $m$ then $|N(X',V_1)|\geq |V_1|-n/64\geq n/40$ and hence some vertex $x\in X'$ must have at least $n/40m\geq 2d$ neighbours in the graph $V_1$. Therefore, at least $m+1$ vertices in $X$ have at least $2d$ neighbours in $V_1$.
Similarly at least $m+1$ vertices in $Y$ have at least $2d$ neighbours in $V_2$. Therefore, there is some index $j\in[2m]$ for which $x_j$ and $y_j$ each have at least $2d$ neighbours in $V_1$ and $V_2$ respectively.
The graph $H=G[V_1\cup\{x_j\}]$ then has the property that, given any set $A\subset V(H)$, if $0<|A|\leq m$, then $|N_H(A)|\geq (d+1)|A|+1$, and, if $|A|=m$, then $|N_H(A)|\geq |H|-n/64\geq n/40$. Let $T$ be the $d$-ary tree of depth $l=\lceil \log m/\log d\rceil$. As $k_j\geq 4\log n/\log d$, we have, for sufficiently large $n$, that $\lfloor k_j/2\rfloor-l-1\geq 0$. Attach a path of length $\lfloor k_j/2\rfloor-l-1$ to the root of $T$ to get the tree $T'$ and let the end vertex of the path which is not the root of $T$ be $t_1$. The tree $T'$ has at most $k_j/2+2d^l\leq n/80+2m(d+1)\leq n/40-2md$ vertices. Therefore, by Theorem \ref{almost2}, $H$ contains a copy of $T'$ so that the vertex $t_1$ is embedded onto the vertex $x_j$. Say this copy of $T'$ is $S_1$.
Similarly, $G[V_2\cup\{y_j\}]$ contains a $d$-ary tree with a path of length $\lceil k_j/2 \rceil -l-1$ connecting the root of the regular tree to $y_j$. Call this tree $S_2$.
The set of vertices in the last level of the $d$-regular trees each contain at least $m$ vertices. Suppose these sets are $W_1$ and $W_2$ for the trees $S_1$ and $S_2$ respectively. Let $U'=U\setminus (V(S_1)\cup V(S_2))$, so that $|U'|\geq n/16$. Therefore, $|U'\cap W_1\cap W_2|\geq n/16-2n/64$, so we can pick a vertex $u\in U'$ which is a common neighbour of some vertex $v_1\in W_1$ and some vertex $v_2\in W_2$. Taking the path of length $\lceil k_j/2\rceil-1$ through the tree $S_1$ from $x_j$ to $v_1$, the path $v_1uv_2$ and the path of length $\lfloor k_j/2\rfloor-1$ through the tree $S_2$ from $v_2$ to $y_j$, we get a path from $x_j$ to $y_j$ of length $k_j$, with internal vertices in $U$, as required.
\end{proof}
\subsection{Connecting given vertex pairs}
\begin{lemma} \label{connectexpand} Let $G$ be a graph with $n$ vertices, where $n$ is sufficiently large, and let $d=160\log^2 n/\log\log n$. Suppose $r,k_1,\ldots,k_r$ are integers with $4\lceil\log n/ \log\log n\rceil\leq k_i\leq n/40$, for each $i$, and $\sum_ik_i\leq 3|W|/4$. Suppose $G$ contains the disjoint vertex pairs $(x_i, y_i)$, $1\leq i\leq r$, and let $W\subset V(G)$ be disjoint from these vertex pairs.
If $G$ $d$-expands into $W$, then we can find disjoint paths $P_i$, $1\leq i\leq r$, with interior vertices in $W$, so that each path $P_i$ is an $x_i,y_i$-path with length $k_i$.
\end{lemma}
\begin{proof} Let $k=\lceil\log n/\log\log n\rceil$, and $m=|W|/2d=|W|\log\log n/20\log^2 n$. For any subset $A\subset V(G)$ and $|A|=m$, as $G$ $d$-expands into $W$ and there are no edges between $A$ and $V(G)\setminus (N(A)\cup A)$, we must have that $|N(A)|\geq n-2m\geq (1-1/64)n$.
Using Lemma \ref{splitexpand}, with $n$ sufficiently large, divide $W$ into $k+1$ sets $W_1,\ldots,W_{k}$, and $U$, so that $|W_i|=|W|/16k$, $|U|=15|W|/16$, and $G$ $(2\log n)$-expands into $W_i$, for each $i$. Let $d_0=2\log n$.
We will find as many of the required paths as possible, using vertices from $U$, before finding a $d_0$-matching into $W_1$ from the pairs left to be connected. We can then look at pairs of these neighbours in $W_1$, the connection of any of which will allow us to connect the original pairs. Repeating this with each set $W_i$ in turn will eventually allow us to connect all the pairs of vertices.
Define a \emph{stage-$\alpha$ situation} to consist of an indexing set $I_\alpha\subset[r]$, paths $P_i$, $i\in [r]\setminus I_\alpha$, and sets $T_i,S_i\subset (\cup_{j\leq \alpha} W_j)\cup\{x_i,y_i\}$ , $i\in I_\alpha$, as follows.
\begin{itemize}
\item $|I_\alpha|\leq 2m/(d_0+1)^{\alpha}$, and $|S_i|=|T_i|=(d_0+1)^{\alpha}$,
\item The paths, $P_i$, and subsets, $S_i$, $T_i$, are all disjoint,
\item The path $P_i$ is contained in $U\cup(\cup_{j\leq\alpha}W_j)\cup\{x_i,y_i\}$, has length $k_i$ and end vertices $x_i$ and $y_i$,
\item For each $x\in S_i$, there is an $x,x_i$-path length at most $\alpha$ in $G[S_i]$, and,
\item For each $y\in T_i$, there is a $y,y_i$-path length at most $\alpha$ in $G[T_i]$.
\end{itemize}
We will prove the lemma by creating a stage-$k$ situation. Indeed, in such a situation $|I_k|\leq 2m/d_0^k<1$, so that $I_{k}=\emptyset$ and we must have all the required paths.
We can reach a stage-0 situation as follows. Using vertices in $U$ connect as many of the vertex pairs together with paths of the required length, calling these paths $P_i$. Let $I_0$ index the vertex pairs we have yet to connect. Let $U'$ be the vertices in $U$ not covered by any of these paths, $P_i$. Then $|U'|\geq |U|-\sum_ik_i\geq |W|/8$. By Lemma \ref{connect}, applied to the graph $G$ restricted to the set $W$ and the vertex pairs which still need connected, there can be at most $2m$ pairs needing connected, otherwise we could find another path through $U'$ of the correct length. Therefore $|I_0|\leq 2m$. Setting $S_i=\{x_i\}$ and $T_i=\{y_i\}$, for each $i$, gives a stage-0 situation.
We will prove by induction that, given a stage-$\alpha$ situation with $\alpha\leq k-1$, we can create a stage-$(\alpha+1)$ situation. Thus, by induction we can reach a stage-$k$ situation and complete the required paths.
Suppose then we have a stage-$\alpha$ situation. We have that $|\cup_{i\in I_\alpha}(S_i\cup T_i)|=2(d_0+1)^\alpha|I_\alpha|\leq 4m\leq |W_{\alpha+1}|/2d_0$. As $G$ $d_0$-expands into $W_{\alpha+1}$, we can find a $d_0$-matching from $\cup_{i\in I_\alpha}(S_i\cup T_i)$ into $W_{\alpha+1}$. Let $S_i'$ be the union of $S_i$ and its image under this matching. Form $T'_i$ similarly. Then $|T'_i|=|S'_i|=(d_0+1)|S_i|=(d_0+1)^{\alpha+1}$. The last two properties of a stage-$(\alpha+1)$ situation are satisfied as $S_{i+1}\subset S_i\cup N(S_i)$ and $T_{i+1}\subset T_i\cup N(T_i)$.
For each $i\in I_\alpha$, look for an $x_i$, $y_i$-path of length $k_i$ with interior vertices in $U\cup S_i\cup T_i$ and which is disjoint from all the paths $P_j$ found so far. If such a path exists, take one, and call it $P_i$.
Let $I_{\alpha+1}\subset I_\alpha$ index the pairs which still need to be connected after this process. Let $Z=U\setminus(\cup_i V(P_i))$. From the conditions on the integers $k_i$, we have $|Z|\geq |W|/8$. For each $i\in I_{\alpha+1}$, pair up the vertices in $S_i$ with those in $T_i$. Give each pair, $(s,t)$ say, the integer $k_i-d_{S_i}(s,x_i)-d_{T_i}(t,y_i)\geq k_i-2\alpha\geq 2k$. If we can find a path between the pairs of this length, then, by taking in addition an $x_i,s$-path length $d_{S_i}(s,x_i)$ in $G[S_i]$ and a $t,y_i$-path length $d_{T_i}(t,y_i)$ in $G[T_i]$, we have an $x_i,y_i$-path, a contradiction. By Lemma \ref{connect} then, we must have at most $2m$ pairs. But we have, in total, $|I_{\alpha+1}|d_0^{\alpha+1}$ pairs from the sets $S_i$ and $T_i$, $i\in I_{\alpha+1}$. Therefore, $|I_{\alpha+1}|\leq 2m/(d_0+1)^{\alpha+1}$ as desired.
The indexing set $I_{\alpha+1}$, paths $P_i$, $i\in [r]\setminus I_{\alpha+1}$, and sets $S'_i$, $T'_i$, $i\in I_{\alpha+1}$, form a stage $(\alpha+1)$-situation, as required.
\end{proof}
\subsection{Covering Expanders with paths}
We can find absorbers for each vertex in a given set using the following lemma. It uses Lemma \ref{connectexpand}, which says, roughly, that if $G$ $d$-expands into a set $W$, then, for large enough $d$, we can create paths covering up to $3/4$ of the vertices of $W$.
\begin{lemma}\label{absorbone} Let $n$ be sufficiently large and $d=20\log^2 n$.
Suppose $G$ is a graph containing $W\subset V(G)$, and that $G$ $d$-expands into $W$.
Then, given any set $A\subset V(G)\setminus W$ with $|A|\leq |W|/300\log^2n$, we can find, disjointly in $G[W]$,
40 absorbers size $\log^2n+2$ for each vertex $v\in A$.
\end{lemma}
\begin{proof} Using Lemma \ref{splitexpand}, with $n$ sufficiently large, partition $W$ as $W_1\cup W_2\cup W_3$, so that, for each $i$, $|W_i|=n/3$ and $G$ $(\log^2 n)$-expands into $W_i$. As $|A|\leq |W_1|/160$, we can find an 80-matching from $A$ into $W_1$, by Hall's theorem. Say $v\in A$ is matched to $B_v$ under such a matching, and partition $B_v$ into pairs.
We wish to create $40$ absorbers for each $v\in A$, each using a different pair of vertices from $B_v$. Each absorber will be found by creating paths using $W_2$ and then using $W_3$. As we seek in total to create $40|A|$ absorbers, covering $40|A|(\log^2n+2)\leq 3|W_2|/4=3|W_3|/4$ vertices, we can create all the required absorbers simultaneously and disjointly using Lemma \ref{connectexpand}. To reduce notation, we will describe the construction of one absorber for $v\in A$ using one pair of vertices, $\{x_0,y_1\}\subset B_v$, but, by applying Lemma \ref{connectexpand} to $W_2$ and $W_3$ we can create all the required absorbers at once.
Let $k=\log n$. The graph $G$ $(\log^2n)$-expands into $W_2$. Using Lemma \ref{connectexpand}, we can find a path $Q$ of length $2k+1$ from $x_0$ to $y_1$ with interior vertices in $W_2$. Let $Q$ be the path $x_0x_1x_2\ldots x_{k}y_0y_{k}y_{k-1}\ldots y_2y_1$.
The graph $G[W_3]$ $(\log^2 n)$-expands into $W_3$. By Lemma \ref{connectexpand}, we can find $k$ disjoint paths $P_i$ length $(k-1)$ connecting $x_i$ and $y_i$ pairwise.
\input{absorberpicture}
Let $R=\cup_iV(P_i)\subset W$. When $k$ is even, the following two $x_0,y_0$-paths have vertex sets $R$ and $R\cup\{v\}$ respectively (see Figure \ref{absorberpicture}, where the heavy lines are paths and the light lines are edges).
\[
x_0x_1P_1y_1y_2P_2x_2x_3P_3y_3y_4\ldots y_{k}P_{k}x_{k}y_0
\]
\[
x_0vy_1P_1x_1x_2P_2y_2y_3P_3x_3\ldots x_{k}P_{k}y_{k}y_0
\]
When $k$ is odd the following two $x_0,y_0$-paths have vertex sets $R$ and $R\cup\{v\}$ respectively.
\[
x_0x_1P_1y_1y_2P_2x_2x_3P_3y_3y_4\ldots x_{k}P_{k}y_{k}y_0
\]
\[
x_0vy_1P_1x_1x_2P_2y_2y_3P_3x_3\ldots y_{k}P_{k}x_{k}y_0
\]
Hence, $(R,x_0,y_0)$ is an absorber for $v$ in both cases. Counting the vertices created reveals that $|R|=k^2+2$.
\end{proof}
We use our absorbers for single vertices to build up a system of paths with a global absorption property. Suppose we have two absorbers $(S,s,t)$ and $(S',s',t')$, for the vertices $v$ and $v'$ respectively. Given a $t,s'$-path with interior vertices not in $S\cup S'\cup\{v,v'\}$, suppose we add its vertices to $S\cup S'$ to get the set $S''$. It can be easily seen that $(S'',s,t')$ is an absorber both for $v$ and for $v'$. Similarly, given several absorbers with paths linking their endpoints in some order, we can merge them into a single absorber.
The challenge in our setting is that we are concerned with the precise length of the path eventually found through an absorber. We will take our absorbers from Lemma \ref{absorbone} and divide them into groups of up to 40, connecting them as described above. This creates an absorber capable of absorbing up to 40 vertices. When we come to absorb vertices, each path will absorb exactly one vertex so that we know the length of the resulting path. By controlling precisely which absorbers we group together (using Lemma \ref{flexiblematching} for guidance), we can build up some global flexibility.
\begin{lemma} \label{absorbmany} Let $n$ and $r$ be sufficiently large, and let $l=10^3\log^2n$. Let the graph $G$, with $n$ vertices, contain the disjoint sets $A$, $W$, $X=\{x_1,\ldots,x_{3r}\}$ and $Y=\{y_1,\ldots,y_{3r}\}$. Suppose $r\leq|W|/6l$, $|A|=2r$ and that $G$ $(400\log^2 n)$-expands into $W$. Then, there is a subset $W'\subset W$, with $|W'|=3r(l-2)-r$, as follows. Given any set $A'\subset A$ with $|A'|=r$, there is a set of $3r$ vertex disjoint $x_i,y_i$-paths in $W'\cup A'$ of length $l-1$. In fact, therefore, such paths cover the set $W'\cup A'$.
\end{lemma}
\begin{proof}
Using Lemma \ref{splitexpand}, with $n$ sufficiently large, partition $W$ as $W_1\cup W_2\cup W_3$, such that, for each $i$, $|W_i|=|W|/3$ and $G$ $(20\log^2 n)$-expands into $W_i$.
Pick a set $B\subset W_1$ with $|B|=|A|$. In the set $W_2$, create the disjoint absorbers $(R_{v,j},r_{v,j},s_{v,j})$, $v\in A\cup B$, $1\leq j\leq 40$, using Lemma \ref{absorbone}, so that each absorber $R_{v,j}$ has size $\log^2n+2$ and can absorb $v$. To describe how to route our paths through these absorbers we refer to an ancillary bipartite graph $H$ with the following properties, provided, for $r$ sufficiently large, by Lemma \ref{flexiblematching}. The bipartite graph $H$ has maximum degree $40$ and vertex classes $[3r]$ and $A\cup B$ such that for any subset $A'\subset A$ with $|A'|=r$ there is a matching between $[3r]$ and $A'\cup B$ in $H$. For each vertex $v\in A\cup B$, let $c_v:N_H(v)\to [40]$ be an injective function, listing the neighbours of $v$ in $H$.
Pick an index $j\in[3r]$, and take the absorbers $R_{v,c_v(j)}$, $v\in N_H(j)$. Take the set of vertices $\{r_{v,c_v(j)},s_{v,c_v(j)}:v\in N(j)\}\cup\{x_j,y_j\}$. Using Lemma \ref{connectexpand} and the set $W_3$, join pairs from these vertices by paths of length at least $\log n$ in such a way as to create an absorber $(S_j,x_j,y_j)$, with $|S_j|=l-1$, which is an absorber for each vertex $v\in N(j)$. Using Lemma \ref{connectexpand}, as $3lr\leq 3|W_3|/4$, we can do this for all indices $j\in [3r]$ simultaneously so that the paths created are disjoint. Note that, because the functions $c_v$ are injective, the absorbers $S_i$ will be disjoint.
Let $W'=(\cup_i(S_i\setminus\{x_i,y_i\})\cup B$. We claim this is such a set as required by the lemma. Indeed, let $A'\subset A$ be any set of size $r$. From the property of the graph $H$, we can find a matching $M$ between $A'\cup B$ and $[3r]$. For each $i\in [3r]$, find the vertex $v\in A'\cup B$ matched to $i$ in $M$ and take find an $x_i,y_i$-path length $l-1$ through $S_i\cup\{v\}$. These paths cover $W'\cup A'$, as required.
\end{proof}
\begin{theorem}\label{expomain2} Let $n$ be sufficiently large and let $l\in \mathbb{N}$ satisfy $l\geq 10^3\log^2n$ and $l|n$. Let a graph $G$ contain $n/l$ disjoint vertex pairs $(x_i,y_i)$ and let $W=V(G)\setminus (\cup_i\{x_i,y_i\})$. Suppose $G$ $d$-expands into $W$, where $d=10^{10}\log^4 n/\log\log n$. Then we can cover $G$ with $n/l$ disjoint paths $P_i$, length $l-1$, so that, for each $i$, $P_i$ is an $x_i,y_i$-path.
\end{theorem}
\begin{proof}[Proof of Theorem \ref{expomain2}]
Let $l_0=10^3\log^2n$, and $d_0=10^8\log^4 n/\log\log n$. We first note that it is sufficient to prove the lemma with $l=l_0$ and $d=d_0$, as follows. Suppose, indeed, the lemma holds for $l_0$ and $d_0$. Take the set $W$ and use Lemma \ref{splitexpand}, with $n$ sufficiently large, to partition $W$ as $W_1\cup W_2\cup W_3$, with $|W_1|=3|W|/4$ and $|W_2|=|W_3|=|W|/8$, such that $G$ $(d/40)$-expands into $W_1$, $W_2$ and $W_3$. Let $l+1=ql_0+r$, where $0\leq r< l_0$ and $q\in \mathbb{N}$. Now, take $n/l$ vertices from $W_2$ and label them $z_1,\ldots,z_{n/l}$. Using Lemma \ref{connectexpand} and the set $W_1$, create disjoint paths of length $\max\{r,\log n\}$ connecting $x_i$ to $z_i$ pairwise. As $\max\{r,\log n\}\leq l/2$, these paths cover at most $n/2\leq 3|W_1|/4$ vertices. If $\log n>r$, then delete $(\log n)-r$ vertices from the end of each of these paths, starting with $z_i$. Let $z_i$, again, be the end vertex of this remaining path for each $i\in[n/l]$. We have now have paths of length $r$ from $x_i$ to $z_i$ for each $i$.
Note that between any two disjoint vertex sets of size $n/d_0$ there must be some edge, since $G$ $d_0$-expands into $W$. Hence there is an edge within any set of $2n/d_0$ vertices. Greedily then, we can take $(n/l)(q-1)$ disjoint edges inside the remaining vertices in $W_1\cup W_2$. Label these edges as $y_{i,j}x_{i,j+1}$, $i\in[n/l]$, $1\leq j< q$, and set $x_{i,1}=z_i$ and $y_{i,q}=y_i$ for each $i$. Let $W'$ be the vertices in $W$ not in any of these edges, or in the paths ending in $z_i$. Then the pairs $\{(x_{i,j},y_{i,j}):i\in[n/l],j\in[q]\}$, the set $W'$ and the graph $G[W'\cup\{x_{i,j},y_{i,j}:i\in[n/l],j\in[q]\}]$ satisfy the conditions of the lemma for $l_0$ and $d_0$. Covering the vertices in $W'$ with paths length $l_0-1$ between these pairs produces the required $x_i,y_i$-paths in the general case.
Suppose then that $l=10^3\log^2n$ and $d=10^8\log^4 n/\log\log n$. Let $m=\lceil n/2d\rceil$, so that any two disjoint vertex sets size $m$ in $G$ have some edge between them. Let $s=n/10^5\log^3 n$.
Using Lemma \ref{splitexpand}, for $n$ sufficiently large, partition $W$ into $W_1$, $W_2$ and $W_3$ as follows. We have $|W_1|=s\log n$, $|W_2|=3s\log n$ and $|W_3|=|W|-2s\log n\geq n/2$, such that $G$ $(d/10)$-expands into $W_3$ and $G$ $(ds\log n/5|W|)$-expands into each set $W_1$ and $W_2$. Note that $ds\log n/5|W|\geq ds\log n/5n\geq 160\log^2 n/\log\log n$ and $s\log n\leq |W_3|/12l$.
By Lemma \ref{absorbmany}, taking $r=2s\log n$ and $A=W_1\cup W_2$, we can find a subset $W'\subset W_3$ of size $(6(l-2)-1)s\log n$ so that, given any subset $Z\subset W_1\cup W_2$ with $|Z|=2s\log n$, there is a set of disjoint $x_i,y_i$-paths, $i\in[6s\log n]$, of length $l-1$ that cover the vertices $W'\cup Z$.
Let $Z_1=W_3\setminus W'$ and let $\alpha=l/(2\log n+2)$ be an integer\footnote{We assume this for a clear presentation. Where this is not an integer we make a small adjustment to $l$ so that this is true. Such a small adjustment will easily be tolerated by the proof of Lemma \ref{absorbmany}.}. Take distinct vertices $x_{i,j}\in Z_1$, $6s\log n+1\leq i\leq n/l$, $2\leq j\leq \alpha$, and let $x_{i,1}=x_i$ and $x_{i,\alpha+1}=y_i$, $6s\log n+1\leq i\leq n/l$. Let $Z_2\subset Z_1$ be the vertices in $Z_1$ not among these labelled vertices. Connect as many pairs $x_{i,j}$ and $x_{i,j+1}$ as possible by disjoint paths of length $2\log n+2$ using vertices from $Z_2$, stopping if no more can be connected or until there are $s$ such vertex pairs remaining. If after this there are $t$ vertex pairs remaining, where $t> s$, then take among them $m$ pairs $(a_i,b_i)$, $i\in [m]$, so that the vertices $a_i$ and $b_i$ are all distinct. Let $Z_3\subset Z_2$ be the set of vertices not covered by any of the paths found so far. Now $|Z_3|=(2\log n+1)t-(|W_1|+|W_2|)/2\geq s\geq 1000m$. Given any set $U\subset V(G)$ with $|U|=m$, $|N(U,Z_3)|\geq |Z_3|-2m\geq (1-1/128)|Z_3|$. Therefore, the graph $G[Z_3\cup(\cup_i\{a_i,b_i\})]$ satisfies the conditions of Lemma \ref{connect}, and so, by that lemma, there is a path of length $2\log n+2$ between one of the pairs $(a_i,b_i)$ in this graph, contradicting $t>s$. The process stops, therefore, with only $s$ pairs remaining.
Say these $s$ unconnected pairs are $(c_i,d_i)$, $i\in[s]$. Let $Z_3$ be the remaining vertices left in $Z_2$ not covered by any paths. Then $|Z_3|=(2\log n+1)s-(|W_1|+|W_2|)/2=s$. Label $Z_3$ as $\{e_{i}:i\in[s]\}$. Recalling that $G$ $(160\log^2 n/\log\log n)$-expands into $W_1$, we can (comfortably) find a generalised matching into $W_1$ from these vertices and the vertices $c_i$ and $d_i$, such that, for each $i\in[s]$, $e_i$ is matched to two vertices, say $e'_i$ and $e''_i$, and $c_i$ and $d_i$ are matched to one vertex each, say $c_i'$ and $d_i'$ respectively.
Use $W_2$ and Lemma \ref{connectexpand} to find disjoint paths length $\log n-1$ connecting $c'_i$ to $e'_i$, and $e''_i$ to $d'_i$. For each $i$, these paths and vertices form paths length $2\log n+2$ from $c_i$ to $d_i$. Thus, we have found paths length $l$ from $x_i$ to $y_i$ passing through $x_{i,2},\ldots,x_{i,\alpha}$, for each $i>6s\log n$. The paths used $2s\log n=|W_1\cup W_2|/2$ vertices from $W_1\cup W_2$. Let $Z$ be the vertices of $W_1\cup W_2$ not used in any of the paths. Using the properties of the set $W'$, find a set of $6s\log n$ paths, which connect $x_i$ and $y_i$ pairwise for each $i\in[6s\log n]$, have length $l-1$, and cover the set $W'\cup Z$. This completes all the paths and covers the graph.
\end{proof}
\subsection{Proof of Theorem \ref{mainexpo1}}
\begin{proof} Let $T$ be a tree with $n$ vertices and maximum degree at most $\Delta$. Take $n$ vertices, and reveal edges independently with probability $\Delta\log^5 n/ 2$ to get the graph $G_1$. By Lemma \ref{randomexpand}, $G_1$ is almost surely an $(n,\Delta\log^4 n/ 20)$-expander.
Suppose $T$ has at least $n/(4\cdot 10^3\log^2n)\geq n/\log^3 n$ leaves. Remove $n/\log^3 n$ of these leaves from $T$ to get the tree $T'$. By Theorem \ref{almost}, $G_1$ contains a copy of $T'$. Fix this copy of $T'$ and let $M$ be the set of vertices which need leaves attached to them to complete the embedding of $T$. Let the set of vertices not yet in the embedding be $L$. We have $|L|\geq n/\log^3 n$ and $|M|\geq \log^3 n/ \Delta$. Almost surely, by revealing more edges with probability $\Delta\log^5 n/2n$, we can, by Lemma \ref{matchymatchy} find the required generalised matching between $M$ and $L$ to attach the right number of leaves in $L$ to each vertex in $M$ to complete the embedding of $T$.
If $T$ does not have at least $n/(4\cdot 10^3\log^2n)$ leaves, then by Corollary \ref{split}, it must have at least $n/(4\cdot 10^3\log^2n)$ bare paths length $10^3\log^2n$. Remove the interior vertices of such a set of bare paths to get a forest, $T'$, with at most $5n/6$ vertices. Use Lemma \ref{splitexpand}, with $W=V(G)$ and $n$ sufficiently large, to find disjoint sets $W_1$, $W_2$ with $|W_1|=7n/8$ and $|W_2|=n/8$ so that $G$ $(\log^4 n/200)$-expands into $W_i$, $i=1,2$. By Theorem \ref{almost}, $G[W_1]$ contains a copy of $T'$. Fix such a copy and suppose to complete the embedding we need to find paths length $10^3\log^2n$ between $x_i$ and $y_i$ to cover the remaining vertices not in the embedding, $Z$ say. As $W_2\subset Z$, $G[Z\cup(\cup_i\{x_i,y_i\})]$ $d$-expands into $Z$, and so, by Theorem \ref{expomain2}, we can cover the remaining vertices by disjoint paths of the required length to complete the embedding.
\end{proof}
\subsection{Finding flexible paths}
\begin{proof}
\end{proof}
\section{Introduction}\label{intro}
\input{1introexp}
\section{Preliminaries}\label{tools}
\input{2toolsexpo}
\section{Constructing Paths}\label{paths}
\input{4paths}
\section{Proof of Theorem \ref{mainexpo1}}\label{main}
\input{5absorb}
\section{Paths in directed graphs}\label{technical}
\input{7expanders}
\section{Future work}\label{future}
\input{8future}
\begin{acknowledgements} The author would like to thank Andrew Thomason for his help and suggestions.
\end{acknowledgements}
\bibliographystyle{plain}
|
\section{Introduction}
This paper was motivated by the recent public debates on the safety of nuclear reactors in India. Although, this an old international debate, it has occupied a rather prominent public position in India after the passage of the Indo-US nuclear deal. What is significant is that the debate has not been confined to the technical community, but has seen the active participation of peoples' movements and civil society groups. This work attempts to parse some of the claims made by the Indian nuclear establishment and international nuclear vendors in this debate. Of course, while we have used the Indian discussion to provide motivation and also specialized some of our conclusions to India, the central conclusion that we draw regarding the conflict between the results of probabilistic risk assessment and empirical data are valid more broadly.
In this paper, we analyze a framework called ``probabilistic risk assessment'' (PRA) that is used by the nuclear industry to calculate the expected frequency of nuclear accidents. As we review in section \ref{secprareview} below, the stated results for this expected frequency are often extremely low. Although it likely as we describe in \ref{secliability} that the industry and policy-makers are privately aware of the unreliability of these figures, they are nevertheless used quite commonly in policy-debates. Thus, for
example, the previous chairperson of the Indian Atomic Energy Commission declared that the chance of a nuclear accident in India was ``1-in-infinity'' \cite{banerjeeinfinity}.
Similar statements can be found in the scholarly literature.
For example, in a review of the 1000 MW VVER reactors that were installed in the South Indian city of Kudankulam, officials from the Nuclear Power Corporation of India (NPCIL) pointed out that the Core-Damage Frequency (CDF) of these reactors --- the rate at which the reactor is expected to suffer accidents that damage its core --- was just $10^{-7}$ per reactor-year (ry) \cite{agrawal2006vvers}.
The other two multinational companies that are in line to construct reactors in India --- the French company, Areva, whose European Pressurised Reactors (EPRs) have been selected for Jaitapur (Maharashtra) and the the American company Westinghouse, whose AP1000 reactors have been selected for Mithi Virdi (Gujarat) --- have also put forward similar figures. For example, Areva claims that the CDF of the EPR is
$7.08 \times 10^{-7} \ry$. (See p. 2 of \cite{ukpsaresults}.)
Westinghouse has estimated the CDF of the AP1000 to be $5.09 \times 10^{-7} \ry$. (See Table 5-1 and Table 8-2 of \cite{ukhseap1000}.)
Prima facie, it is prudent to treat such low numbers skeptically, especially when they are provided with a precision of two decimal places. A nuclear reactor is a very complex system, and our understanding of its dynamics, particularly when they are coupled to an uncertain external environment, is not advanced
enough to permit such accurate estimates.
Nevertheless, these numbers are taken seriously by regulators, both
in India and elsewhere. In the United States, the Nuclear Regulatory Commission has set a goal for PRA estimates of both the CDF and the large-release frequency (LRF). The latter involves accidents where the containment fails in addition to core damage. These are expected to be below $10^{-4} \ry$ and $10^{-6} \ry$ respectively. (See \cite{nrcwebgoals} and Appendix D of \cite{nureg1860}.) Similar quantitative criterion have been adopted in other countries \cite{oecdpra}. In India, licensing guidelines suggest that
new plants should have a PRA-estimated CDF smaller than $10^{-5} \ry$ and a LRF lower than $10^{-6} \ry$.
(See Annexure-1 of \cite{aerbconsentpra}.)
The methodology used to carry out PRAs in the nuclear industry, which we review in section \ref{secprareview}, is theoretically suspect. However, the purpose of this article is to point out that, even setting aside theoretical considerations, these extraordinarily low bounds can be ruled out
by using the existing empirical record of nuclear accidents.
There have already been eight core-damage accidents in a little more than fifteen thousand reactor-years of experience as we review in section \ref{secempirical}. Hence, the observed
frequency of accidents
is significantly higher than that predicted by the industry's PRAs. Moreover, and this is the crucial conclusion of this paper: had the true
frequency of accidents been as low as the manufacturers claim, it is exceedingly
unlikely that so many accidents could have occurred. Conversely,
the historical record on accidents implies that, even under rather
conservative assumptions, it is possible to conclude with a very high degree
of confidence that the results of the industry's PRAs are incorrect.
For example, we show in section \ref{frequencypra} that in a simplified model, even with favourable assumptions for the industry, the hypothesis that the frequency of core-damage accidents is smaller than or equal to $10^{-7}\ry$ can be ruled out with a confidence of $1.0-4.0 \times 10^{-24}$. (Said differently, within this model, we can conclude that the core-damage frequency of nuclear reactors is higher than $10^{-7} \ry$ with a confidence of \mbox{99.9999999999999999999996}\%!) The precise value of $4.0 \times 10^{-24}$ is specific to this model but our broader and robust conclusion is that the accident-frequencies suggested by PRA can be ruled out, even with the limited empirical data on accidents, at an extremely high confidence level. We perform a sensitivity analysis in section \ref{secrobust} and show that even by allowing exponential increases in safety standards over time, or significant variations across regions, it is virtually impossible to reconcile the empirical data with the PRA-frequencies. In section \ref{secindia}, we turn briefly to the specific case of India. Here we make the elementary
statistical point that the existing experience of Indian reactor operation is insufficient to derive any strong conclusions about expected accident frequencies in the future. In section \ref{secliability}, we briefly discuss the debate on liability because it shows that the nuclear industry is itself aware that the quantitative results of PRA cannot be taken seriously for commercial questions like liability insurance.
We conclude
in section \ref{secconclude}.
\section{Summary and Methodology \label{secsummary}}
In this section, we provide a short summary of our objectives, methodology and results.
Our central objective in this paper is to use the empirical data to estimate the probability for the hypothesis that the true frequency of accidents is indeed as low as that claimed by PRA. We denote this probability by $\cumprob(\lambda_{\text{pra}}, n_{\text{obs}})$ where $\lambda_{\text{pra}}$ is a frequency predicted by PRA and $n_{\text{obs}}$ is the observed number of accidents of a certain type. Our central result is that
\be
\label{centralclaim}
\cumprob(\lambda_{\text{pra}}, n_{\text{obs}}) = \epsilon \ll 1.
\ee
The precise value of $\epsilon$ depends on many factors: the precise PRA being considered, which leads to variations in $\lambda_{\text{PRA}}$, the way in which accidents are counted which leads to variations in $n_{\text{obs}}$ and, of course, the assumptions that go into modeling nuclear accidents and the variation of safety-standards over time and across regions.
But the point that we want to emphasize in this paper is the following: $\epsilon$ is extraordinarily small, and this fact is robust against virtually any variation of the considerations above. The calculations presented in this paper are all calculations of $\epsilon$ under different assumptions and they all have the feature that $\epsilon$ is negligible.
Note that we are not interested in using the empirical data to compute the true frequency of accidents. Some efforts have been made in this direction \cite{sornette2013exploring,ha2014calculating} but the basic difficulty is that the empirical data is too limited to allow for a reliable computation of this frequency. The importance of focusing on $\epsilon$ is that these intermediate uncertainties do not affect the robust nature of \eqref{centralclaim}.
Second note that to compute $\epsilon$, we are forced to use Bayesian analysis. While a purely frequentist approach might demonstrate that the empirically observed frequency of accidents does not agree with the frequency predicted by PRA this does not, by itself, tell us the confidence level with which we can rule out the correctness of PRA. In fact the question of estimating the confidence-level cannot be posed within the frequentist approach at all. This confidence-level is given by $1 - \epsilon$, and requires Bayesian methods.
We now turn to the main body of the paper.
\section{A Brief Review of Probabilistic Risk Assessment \label{secprareview}}
The systematic use of PRA in the nuclear industry started with the Rasmussen report of 1975 \cite{rasmussen}. This report was commissioned by the U.S. Nuclear Regulatory Commission (NRC) and met with criticism soon after its publication. In 1977, the U.S. NRC commissioned a review of the Rasmussen report through the Lewis Committee, and following this critical review-report, released a statement in 1978 stating that ``the Commission does not regard as reliable the Reactor Safety Study's numerical estimates of the overall risk of reactor accidents.'' (See Pg. 533 of \cite{kok2009nuclear}.) In spite of these early objections, over the past few decades, the framework of probabilistic risk assessment has become influential in propagating the use of quantitative probabilistic techniques for questions of nuclear safety. Moreover, as described above, the use of these techniques is not confined to qualitative safety analyses aided by quantitative probabilistic techniques, but rather the actual numerical results produced by PRA are used by regulators, and in policy debates.
The basic idea used by such studies is simple to describe: one enumerates the possible fault trees that could lead to an accident.
For each individual component in the reactor, one can estimate a frequency of failure. For a serious accident, some combination of these components has to fail simultaneously. The industry advertises a philosophy of ``defense in depth'', which reduces the overall possibility of an accident by building redundancy into a system. The low numbers above result from the fact that several systems have to fail simultaneously before the core is damaged.
For example, Keller and Modarres \cite{keller_historical_2005} report that in 1966, in one of the earliest such assessments, the General Electric company ``showed'' that its reactors ``had a one-in-a-million chance per year for a catastrophic failure because each of the three major subsystems would only fail once-in-one-hundred per years'' {\em (sic)}. The Rasmussen report, which followed a decade later, attempted to refine and formalize this methodology.
The theoretical problem with such estimates is obvious. Consider
the Fukushima nuclear complex, which had 13 backup diesel generators \cite{wnnfukushima}. Assigning a probability of $10^{-1}$ for the failure of each generator per year and assuming that they are independent would lead us to the
naive conclusion that the probability that 12 generators would fail together in any given year is about $13 \times 10^{-12} \times 0.9 \approx 10^{-11}$.
However, the tsunami did precisely this by disabling all but one of the generators at once. The point is that once the obvious fault trees have been eliminated and corrected, we reach a stage where
the dominant contributions to accident-probabilities come from unlikely sequences of events that conspire to cause a failure.
This issue occurs in any suitably complex system. However, a nuclear reactor is also an open system that is coupled
to an external environment. This makes it virtually impossible to foresee all sorts of low probability
pathways to failure. Furthermore, our understanding of the frequency of extreme initiating events, such as tsunamis and earthquakes, is itself rather crude.
These error bars overwhelm
the seemingly accurate predictions that result from elaborate simulations of the reactor.
However, instead of venturing deeper into these theoretical arguments --- which are considered in greater detail elsewhere in the literature (see, for example, \cite{diazmaurinthesis}) --- we will examine how the predictions of PRA stand up against the extant empirical data.
To facilitate this comparison, let us summarize the various kinds of claims that
have been made by the three multinational companies relevant to India.
As we mentioned, Areva has estimated that the CDF of
an EPR accounting for both internal and external hazards is $7.08 \times 10^{-7} \ry$. The LRF of the EPR is estimated to be about 11\% of its CDF: $7.69 \times 10^{-8} \ry.$ (See p. 14 of \cite{ukpsaresults} and p. 222 of \cite{ukpsaresultslevel2}.)
Similarly, Westinghouse has estimated that the CDF of the AP1000 is
$5.09 \times 10^{-7} \ry$.
The LRF of this reactor
is estimated to be $5.94 \times 10^{-8} \ry$. (See Table 8-2 of \cite{ukhseap1000}.)
This includes accidents due to ``external hazards'' including ``external flooding, extreme winds, seismic, and transportation accidents''. In fact Westinghouse concluded that ``conservative bounding assessments show that core damage risk from events listed above is small compared to the core damage risk from at-power and shutdown events.'' (See p. 4-37 of \cite{ukhseap1000}.)
Although
the NPCIL has stated that the CDF of the Kudankulam VVER reactors is $10^{-7} \ry$, we were unable to locate
the details of the PRA that led to this conclusion. So, we will take this figure as is, and consider a LRF of $10^{-8} \ry$ using the common estimate that the LRF ``is generally about ten times less than CDF'' \cite{wnacdflrf}. All these claims are summarized in Table \ref{tablepra}.
\begin{table}[!h]
\begin{center}
\resizebox{\textwidth}{!}{
\begin{tabular}{ccc}
\toprule {\bf Reactor (Manufacturer)} & {\bf Core-Damage Frequency} & {\bf Large-Release Frequency} \\ \midrule
EPR (Areva) & $7.08 \times 10^{-7} \ry $
& $7.69 \times 10^{-8} \ry$
\\
AP1000 (Westinghouse)&$5.09 \times 10^{-7} \ry $
& $5.94 \times 10^{-8} \ry$
\\
VVER (Rosatom) & $10^{-7} \ry$
& $10^{-8} \ry$
\\ \bottomrule
\end{tabular}}
\caption{Predictions of Probabilistic Risk Assessment \label{tablepra}}
\end{center}
\end{table}
\section{Review of the Empirical Experience \label{secempirical}}
We now review the historical record on nuclear accidents. The industry, as a whole, had gathered about $T_{\text{obs}} = 15247$ reactor-years of operating experience by the end of 2012 according
to the latest data put out by the International Atomic Energy Agency (IAEA). (See Table 4 of \cite{npriaea2013}.)
In this time, there have been
several core-damage accidents.
Surprisingly, the IAEA does not maintain a comprehensive historical record
of core-damage accidents. Cochran and McKinzie have compiled a very useful
list of 25 such instances from various sources \cite{cochranmckinziw}. In our analysis, we
will only consider accidents that occurred at commercial reactors, thus excluding
accidents at experimental facilities like Enrico Fermi Unit-1 (1966) or Lucens (1969). Even this enumeration is somewhat subjective, but a conservative approach,
keeping only accidents that involved a significant meltdown of fuel leads
to the list in Table \ref{nuclearacc}. In each case, we have also provided references to more detailed descriptions of these accidents.
\begin{table}
\resizebox{\textwidth}{!}{
\begin{tabular}{p{1.4in}p{0.7in}p{0.5in}p{1in}p{1.9in}}
\toprule
{\bf Reactor} & {\bf Country} & {\bf Year} & {\bf Reference} & {\bf Note} \\
\midrule
Saint-Laurent A-1 & France & 1969 & p. 35 of \cite{selectednuclearacc} & Meltdown of 50 kg of fuel.
\\[10pt]
Three Mile Island Unit 2 & USA & 1977 & \cite{kemenycommissiontrunc} & Severe accident; radiation release.
\\[10pt]
Saint-Laurent A-2 & France &1980 & Table 12 of \cite{inesdescription} and \cite{cochranmckinziw} & Meltdown of one channel of fuel.
\\[10pt]
Chernobyl Unit 4 & Ukraine & 1986 & \cite{_chernobyl_1992} & Severe accident; large radiation release.
\\[10pt]
Greifswald Unit 5 & Germany (GDR) & 1989 & \cite{wnanuclearsafetygreif,wnadecommgreif,widegreif,cochranmckinziw} & Partial core meltdown soon after commissioning.
\\[10pt]
Fukushima Daiichi Units 1,2,3 & Japan & 2011 & \cite{fukushimaindependentreport} & Severe accident; large radiation release. \\[10pt]
\bottomrule
\end{tabular}}
\caption{List of core-damage accidents in commercial reactors \label{nuclearacc}}
\end{table}
Table \ref{nuclearacc} enumerates accidents at $n^{\text{cd}}_{\text{obs}} = 8$ reactors.
Of these 8 accidents, $n^{\text{lr}}_{\text{obs}} = 5$ accidents led to the release of large amounts of radioactive substances into the environment. This list comprises the accidents at Three Mile Island, Chernobyl and the three at Fukushima. Since we have PRA results for
both the CDF and the LRF, we can compare these separately to the historical record.
The reason for counting the three accidents at Fukushima separately in the list above is that we are interested in the rate of accidents per reactor-year of operation, and all these three reactors contribute separately
to the ``total operating experience'' that appears in the denominator.
We comment more on this issue in section \ref{secconclude}. However, to demonstrate the robust nature of our conclusions,
we will present a parallel analysis where the accidents at Fukushima are counted together. By this (incorrect) counting, there have been $n^{\text{cd}}_{\text{low}} = 6$ core-damage accidents and $n^{\text{lr}}_{\text{low}}=3$ accidents with a large release of radioactivity.
\section{Simplified Bayesian Analysis of Empirical Frequencies and PRA \label{frequencypra}}
Table \ref{nuclearacc} leads to the following observed frequency of core-damage and large-release accidents.
\be
\label{observedfreq}
\begin{split}
&\nu^{\text{cd}}_{\text{obs}} = {n^{\text{cd}}_{\text{obs}} \over T_{\text{obs}}} \approx {1 \over 1906} \ry \approx 5.2 \times 10^{-4} \ry , \\
&\nu^{\text{lr}}_{\text{obs}} = {n^{\text{lr}}_{\text{obs}} \over T_{\text{obs}}} \approx {1 \over 3049} \ry \approx 3.3 \times 10^{-4} \ry .
\end{split}
\ee
Counting the Fukushima accidents as a single accident, we find
\be
\begin{split}
&\nu^{\text{cd}}_{\text{low}} = {n^{\text{cd}}_{\text{low}} \over T_{\text{obs}}} \approx {1 \over 2541} \ry \approx 3.9 \times 10^{-4} \ry , \\
&\nu^{\text{lr}}_{\text{low}} = {n^{\text{lr}}_{\text{low}} \over T_{\text{obs}}} \approx {1 \over 5082} \ry \approx 2.0 \times 10^{-4} \ry .
\end{split}
\ee
It is clear that these empirically observed frequencies are far higher than
the predictions of the manufacturers' PRAs. However, we can ask a more detailed question: given the observed rate of accidents, what is the probability that the results of PRA are close to, or smaller than, the ``true frequency'' of accidents.
Bayesian techniques are ideally suited to answer this question, and we return to this issue again below. To answer the question above,
we need
to make a few simplifying assumptions about the probability distribution of nuclear accidents. As a first approximation, we start by assuming that nuclear accidents are independent
events, and that in every small time interval $d t$, each reactor has
a small and constant probability
\be
\label{centralassumption}
d p = \lambda \, d t,
\ee
of suffering an accident. The use of this approximation is {\em not meant} to suggest that it is truly the case that accident frequencies have not changed over time. Rather, as we show in section \ref{secrobust} our results are very {\em robust} against almost any reasonable assumed changes in safety standards over time, and across regions.
Therefore, given this robustness of our central results, we present this simplified model first since it is amenable to simple analysis and already captures the central points that we wish to make. The reader who is interested in a more sophisticated analysis should consult the next section.
Say that we have $m$ reactors functioning simultaneously which we observe for a time period $\teach = N d t$. The probability for $n$ of these to have undergone accidents is
\be
\label{derivprobdistrib}
\begin{split}
p_{\lambda}(n) &= \begin{pmatrix} m \\ n \end{pmatrix} \big(1 - \lambda d t)^{N (m-n)} \Big(1-\big(1 - \lambda d t \big)^N \Big)^n \\ &= \begin{pmatrix} m \\ n \end{pmatrix} \big(1 - {\lambda \teach \over N } \big)^{N(m-n)} \Big(1-\big(1 - {\lambda \teach \over N }\big)^N \Big)^n \\
&\underset{N \rightarrow \infty}{\longrightarrow} {\Gamma(m+1) \over \Gamma(m-n+1) \Gamma(n+1)} e^{-\lambda \teach m (1 - {n \over m})} \big(1 - e^{-\lambda \teach} \big)^n,\\
\end{split}
\ee
where, in the last line, we have taken the continuous limit: $N \rightarrow \infty$, with $\teach$ finite. We remind the reader that $\Gamma(n+1) = n!$ is the standard Gamma function.
We now consider the case where $\lambda \teach \ll 1$, which implies that the chance of any individual reactor undergoing an accident is small, and $n \ll m$, which states that only a small fraction of all reactors undergo an accident. In this limit, the total operating experience gathered becomes $T = m \teach$, and we see that the distribution simplifies to
\be
\label{poisson}
p_{\lambda}(n) =
{1 \over \Gamma(n+1)} (\lambda T)^{n} e^{-\lambda T}.
\ee
This is just a Poisson distribution and we could even have started with this distribution, which is
commonly used to model accidents and other {\em rare events} in various scenarios.
It is
now possible to solve the following Bayesian problem: start with the {\em prior} assumption
that $\lambda$ is uniformly distributed
\be
\label{priorprob}
\probdens(\lambda) = {\theta(\lambda) - \theta(\lambda - \lambda^c)\over \lambda^c},
\ee
where $\lambda^c$ is an irrelevant high frequency cutoff to make the distribution normalizable.
Given the observed
frequency of events above, what is the {\em posterior}
probability distribution for $\lambda$?
Let us try and explain this in simple words. Say we start with no prior bias
for the value of $\lambda$. Given that we have some number $n_{\text{obs}}$ of accidents,
we can use this empirical data to form an estimate as to the value of $\lambda$. In fact, Bayes theorem
tells us that we can actually calculate the probability that the true frequency of
accidents has a value between $\lambda$ and $\lambda + d \lambda$ through the formula
\be
\label{bayestheorem}
\probdens(\lambda|n_{\text{obs}}) = {\probdens(n_{\text{obs}} | \lambda) \probdens(\lambda) \over \probdens(n_{\text{obs}})}.
\ee
This formula is just making precise that intuition that the empirical evidence is giving us
some information about the value of $\lambda$ in our world.
On the right hand side, $\probdens(n_{\text{obs}} | \lambda)=p_{\lambda}(n_{\text{obs}})$, which is given by the Poisson distribution \eqref{poisson}. $\probdens(\lambda)$ is our flat prior probability distribution in \eqref{priorprob}. To fix $\probdens(n_{\text{obs}})$, which
is a $\lambda$ independent constant, we can simply demand that
$\int_0^{\infty} \probdens(\lambda | n_{\text{obs}}) d \lambda = 1$.
After fixing this constant we find
that the posterior
probability distribution for $\lambda$ is given by
\be
\label{posterior}
\probdens(\lambda | n_{\text{obs}}) = {1 \over \Gamma(n_{\text{obs}} + 1) }T_{\text{obs}} (\lambda
T_{\text{obs}})^{n_{\text{obs}}} e^{-\lambda T_{\text{obs}}},
\ee
where we have neglected terms of $\text{O}\big(e^{-\lambda_c T_{\text{obs}}}\big)$, assuming $\lambda_c$ is taken to be large enough.
This function is plotted in figure
\ref{plambda} for the values of $n_{\text{obs}}$ that are relevant to both
the large-release and the core-damage frequency.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.6\textwidth]{plnuplot.eps}
\caption{Posterior probability distribution for the parameter $\lambda$ \label{plambda}}
\end{center}
\end{figure}
We pause to mention a somewhat subtle point. Since the observed number of
events is small, $n_{\text{obs}} \sim \text{O}\left(1\right)$, the curves in figure \ref{plambda} have an appreciable width. This indicates the difficulty with using a {\em frequentist} approach to estimating the true frequency of accidents using empirical data. However, as we emphasized in section \ref{secsummary} if we are interested in estimating the parameter $\epsilon$ in \eqref{centralclaim} instead, then Bayesian methods yield a robust statement. We now turn to this calculation.
To estimate $\epsilon$, we use \eqref{posterior} to calculate the probability that
the probability that $\lambda$ is smaller than any given
$\lambda_0$. This function is given by
\be
\label{cumlambda}
{\cumprob}(\lambda_0, n_{\text{obs}}) = \int_0^{\lambda_0} \probdens(\lambda|n_{\text{obs}}) \, d \lambda= 1 -
{\Gamma(1 + n_{\text{obs}}, T_{\text{obs}} \lambda_0) \over \Gamma(1 +
n_{\text{obs}})}.
\ee
where $\Gamma(k,z)$ is the incomplete gamma function. The function
$\cumprob(\lambda_0, n)$ is shown in figure \ref{cumprob} for all
the relevant values of $n$.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.6\textwidth]{clnuplot.eps}
\caption{Probability for the hypothesis $\lambda < \lambda_0$ \label{cumprob}}
\end{center}
\end{figure}
Since the probability that the true frequency is smaller than
the various results of PRA is so close to zero, it cannot be read off the graph. However, we can use the series expansion
\be
\label{incompgammafunc}
\Gamma(k,z) = \Gamma(k) - {x^k \over k} + {x^{k+1} \over k+1} + \text{O}\left(x^{k+2}\right).
\ee
This leads to the numerical figures given in Table \ref{praright}.
The phrase ``PRA ... is right'' is shorthand for the hypothesis that the true frequency of accidents
is lower than or equal to the frequency predicted by PRA . Therefore, the table lists the values of the function $\cumprob(\lambda_0, n_{\text{obs}})$ from \eqref{cumlambda} with $\lambda_0$ set to the PRA-predicted frequency and $n_{\text{obs}}$ set to the observed number of accidents, which varies depending
on whether the Fukushima accidents are counted together or separately. Note that the values for the probabilities in this table correspond precisely to the parameter $\epsilon$ in \eqref{centralclaim}.
\begin{table}[!h]
\resizebox{\textwidth}{!}{
\begin{tabular}{ccccc}
\toprule
{\bf Reactor}& \multicolumn{2}{c}{\bf Probability PRA CDF is right} & \multicolumn{2}{c}{\bf Probability PRA LRF is right} \\ \cmidrule(r){2-3}
\cmidrule(l){4-5}
&{\bf Fukushima separate} & {\bf Fukushima together} & {\bf Fukushima separate} & {\bf Fukushima together} \\ \midrule
{Kudankulam}&$1 \times10^{-31}$&$4 \times 10^{-24}$&$ 2 \times 10^{-26}$&$2 \times 10^{-17}$ \\
{EPR}& $5 \times 10^{-24}$ & $3 \times 10^{-18}$ & $4 \times 10^{-21}$ & $8 \times 10^{-14}$ \\ {AP1000}&$3 \times 10^{-25}$&$3 \times 10^{-19}$&$8 \times 10^{-22}$& $3 \times 10^{-14}$\\ \bottomrule
\end{tabular}}
\caption{Comparing PRA results with Bayesian estimates from historical observations (Simplified Model) \label{praright}}
\end{table}
One observes immediately that, given the empirical data, the probability
that the industry's PRA-based conclusions are right is astronomically small. As we stated
in the introduction, this implies that with almost perfect certainty
we can conclude that the true frequency of accidents is much larger than
the figures advertised by the manufacturers.
We emphasize that the figures in Table \ref{praright} should not be used as precise numerical bounds on the validity of the PRA results. The precise values of $\epsilon$ listed there suffers from several uncertainties that we have already mentioned. As we show in the next section upon consideration of a more sophisticated model, the numerical values of these probabilities change but the robust statement is that they always remain extremely small.
In words, the results
of Table \ref{praright} can be stated by means of
the following straightforward conclusion: the historical data on nuclear accidents provides overwhelming evidence that the methodology of probabilistic risk assessment is seriously flawed. A corollary is that the observed frequency of accidents contradicts the industry's claim that the probability of an accident is negligible.
\section{Robustness of the Simplified Model \label{secrobust}}
In this section, we model possible improvements in safety standards over time, and variations in accident frequencies across regions. We show that the central results of the simplified Bayesian model above are very robust against any reasonable assumed variations of this kind. However, as the reader will note the analysis of this section is mathematically more involved and therefore the reader who is willing to accept this conclusion may skip this section on a first reading.
\subsection{Modelling Improvements in Safety over Time}
To model possible improvements in safety over time, we now relax the assumption made in \eqref{centralassumption} and allow the probability of an accident to vary with time. We first discuss a general framework to model this possibility and then discuss a concrete model.
\subsubsection{Framework for Time-Variations of Safety}
The probability that an accident occurs between time $t$ and $t + dt$ is given by
\be
d p(t) = \lambda_{\plist}(t) d t,
\ee
where the subscript $\plist$ indicates a set of parameters that control the variation of $\lambda$ in time.
Now, consider a time interval of length $T$, which we divide into $N$ equal parts set off by $0 < t_1 < t_2 \ldots t_{N-1} < T$ with $t_{i+1} - t_{i} = d t$. To specify the pattern of accidents note now that a single number is not enough, but instead we require a {\em function} $N(i)$, which tells us whether an accidents occurred in the interval $[t_i, t_{i+1}]$. This is defined through
\be
N(i) = \left\{\begin{array}{ll}1,&\text{if~an~accident~occurs in~the~interval}~[t_i, t_{i+1}] \\ 0,&\text{otherwise}\end{array}\right.
\ee
Clearly, the probability for any such pattern of accidents is given by
\be
p(\nlist|\plist) = \nconstpni_{\plist} \prod_{N_i = 1} \lambda_{\plist}(t_i) d t \prod_{N_i = 0} (1 - \lambda_{\plist}(t_i)),
\ee
where $\nconstpni_{\plist}$ is a normalization factor that we discuss below. Note that this probability is infinitesimal for any given pattern $N(i)$, and this is not surprising since in the continuum limit, we need to do a {\em path integral} over all possible patterns of accidents. However, as we see below for our purpose of constructing the posterior probability distribution, this constant will not be important.
In the continuum limit, if $n$ accidents are observed at times $t_{i_1} \ldots t_{i_n}$, then we find that
\be
\probdens\left(\nlist | \plist\right) \propto \left[\prod_{j} \lambda_{\plist}(t_{i_j}) \right]e^{-\int_0^T \lambda(t) d t}
\ee
Then, once again using Bayes theorem, we find that
\be
\probdens\left(\plist | \nlist \right) = {\probdens\left(\nlist | \plist \right) \probdens(\plist) \over \probdens(\nlist)}
\ee
We see therefore that the unknown normalization constants drop out and that the {\em posterior probability distribution} for the parameters upon observation of a certain pattern of accidents is given by
\be
\probdens\left(\plist | \nlist \right) = \nconst \left[\prod_{j} \lambda_{\plist}(t_{i_j}) \right]e^{-\int_0^T \lambda(t) d t} \probdens(\plist)
\ee
where $\probdens(\plist)$ is the {\em prior probability distribution} for the parameters and the normalization $\nconst$ is now a simple finite quantity that is simply fixed by
\be
\int \probdens\left(\plist | \nlist \right) d \plist = 1
\ee
\subsubsection{A Concrete Model}
We now turn to a concrete model that shows how the framework above may be utilized. We assume a variation of probability with time as
\be
\label{varylambda}
\lambda(t) = \lambda_i e^{-\gamma t}.
\ee
Here $t$ is the total operating experience accumulated by the industry. We assume that $\gamma > 0$ and this models the frequency of accidents as starting with $\lambda_i$ and decreasing exponentially with time as the industry gains additional experience.
This model of an exponential increase in safety standards constitutes a very favourable assumption for the industry but we will see that even allowing for this, our central conclusions are unchanged.
Of course, the reader can easily generalize these results to more general variations of the frequency.
Furthermore, we assume a flat prior distribution of the form \eqref{priorprob} for both $\lambda_i$ and $\gamma$.
\be
\label{flatpriorvary}
\begin{split}
\probdens(\lambda_i) &= {\theta(\lambda_i) - \theta(\lambda_i - \lambda_i^c)\over \lambda_i^c} \\
\probdens(\gamma) &= {\theta(\gamma) - \theta(\gamma - \gamma^c)\over \gamma^c},
\end{split}
\ee
where $\lambda_i^c$ and $\gamma^c$ are cutoffs. As we described above, the cutoff on $\lambda_i$ is irrelevant, but in this exponential model we have to be somewhat careful about the cutoff $\gamma^c$. This is because if we assume a prior that is flat over very large ranges of $\gamma$, then this allows for a fat-tail in the posterior distribution that represents the scenario where the accident frequencies were very large in the past but have improved rapidly very recently. We will make a specific numerical choice of cutoff below, although we postpone the question for the moment.
We should emphasize two subtleties. First, while the cutoff, $\gamma^c$ is clearly physically important and changes the numerical values of probabilities, it it not strictly required for convergence. Second, while the flat priors in \eqref{flatpriorvary} reflect our ignorance about these parameters this necessarily involves a choice of basis. Note, for example, that assuming a flat prior for the initial frequency $\lambda_i$ is different from assuming a flat prior for the current frequency $\lambda(T)$.
Now consider the case where $n$ accidents have been observed at times $t_{i_1}, \ldots t_{i_n}$ in a total operating time $T$. Define $\tau = \sum_j t_{i_j}$. Then it is clear from the analysis above that the posterior probability distribution for $\lambda_i, \gamma$ is
\be
\label{postligamma}
\probdens\left( \lambda_i, \gamma | \nlist \right) = \nconst \lambda_i^n e^{-\gamma \tau -{\lambda_i \over \gamma}\big(1 - e^{-\gamma T} \big)}
\ee
We can determine the marginal probability distributions for both $\lambda_i$ and $\gamma$ by integrating over the other variable. In particular, the distribution for $\gamma$ can be obtained by doing the easy integral over $\lambda_i$ and is given by
\be
\probdens\left(\gamma | \nlist \right) = \int_0^{\infty} d \lambda_i \probdens\left( \lambda_i, \gamma | \nlist \right) = \nconst e^{-\gamma \tau } \Gamma (n+1) \left(\frac{\gamma}{1-e^{-\gamma T}}\right)^{n+1}
\ee
On the other hand, it does not appear possible to write the distribution for $\lambda_i$ in terms of elementary functions. However, a double-infinite series representation can be obtained as follows by expanding the exponentials.
\be
\begin{split}
\probdens\left(\lambda_i | \nlist \right) &= \int_0^{\gamma^c} d \gamma \probdens\left( \lambda_i, \gamma | \nlist \right) \\
&= \nconst \int_0^{\gamma^c} d \gamma \sum_{m,q=0}^{\infty} \left(\frac{(-1)^q e^{-\frac{\lambda_i}{\alpha }} \lambda_i^{m+n} \alpha ^{q-m} (m T+\tau )^q}{\Gamma (m+1) \Gamma (q+1)} \right) \\
&= \nconst \sum_{m,q=0}^{\infty} \frac{(-1)^q \lambda_i ^{n+q+1} (m T+\tau )^q \Gamma \left(m-q-1,\frac{\lambda_i}{\gamma^c}\right)}{\Gamma (m+1) \Gamma (q+1)}.
\end{split}
\ee
By integrating these distributions, we also obtain the value of the normalization constant $\nconst$ through
\be
\begin{split}
1 &= \int \probdens\left(\gamma | \nlist \right) d \gamma = \int \probdens\left(\lambda_i | \nlist \right) d \lambda_i \\ &= \nconst \sum_{q=0}^{\infty}\frac{\Gamma (n+q+1) (q T+\tau )^{-n-2} (\Gamma (n+2)-\Gamma (n+2,\gamma^c (q T+\tau )))}{\Gamma (q+1)}
\end{split}
\ee
which yields an expression for $\nconst$ as the inverse of an infinite sum.
Another interesting quantity is the posterior probability distribution for the ``current accident frequency'' $\lambda_T = \lambda_i e^{-\gamma T}$. We can change variables in \eqref{postligamma}, after including a Jacobian factor, to obtain
\be
\label{postltgamma}
\probdens\left( \lambda_T, \gamma | \nlist \right) = \nconst \lambda_T^n e^{-\gamma \tau +\frac{\lambda_T \left(1-e^{\gamma T}\right)}{\gamma }+\gamma (n+1) T}
\ee
As usual the probability that $\lambda_i$ is smaller than a given value, $\lambda_0$, or the probability that $\gamma$ is smaller than some $\gamma_0$ is given by integrating the probability distributions above.
\be
\begin{split}
&\cumprob_{\gamma}(\gamma_0 , \nlist) = \int_0^{\gamma_0} \probdens\left(\gamma | \nlist \right) d \gamma \\
&\cumprob_{i}(\lambda_0 , \nlist) = \int_0^{\lambda_0} \probdens\left(\lambda_i | \nlist \right) d \lambda_i \\
&\cumprob_{T}(\lambda_T^0 , \nlist) = \int_0^{\lambda_T^0} d \lambda_T \int_0^{\gamma^c} d \gamma \probdens\left( \lambda_T, \gamma | \nlist \right)
\end{split}
\ee
While it is possible to write these expressions in terms of an infinite series of elementary functions, these forms are too complicated to be useful. It is, of course, possible to evaluate these expressions numerically at any given point as we do below.
We now turn to the numerical values of these expressions at the empirically relevant points. Since the expressions above depend on the specific times at which the accidents occurred, we need some additional information: the reactor-years of operating experience that the industry had accumulated at the time of the accidents listed in Table \ref{nuclearacc}. Using Table 7 of \cite{npriaea2013}, we can estimate these figures for the years in which the accidents occurred, by simply adding the number of operating reactors, as given in that table. This estimate is more than sufficient for our purposes and results in the Table \ref{tableopexp}.
\begin{table}[!h]
\begin{center}
\begin{tabular}{llc}
\toprule
{\bf Accident} & {\bf Year} & {\bf Operating Experience} (ry) \\
\midrule
Saint-Laurent A-1 & 1969 & 391 \\
Three Mile Island & 1977 & 1406 \\
Saint-Laurent A-2 & 1980 & 2048 \\
Chernobyl Unit 4 & 1984 & 3150 \\
Greifswald Unit 5 & 1989 & 5061 \\
Fukushima Units 1,2,3 & 2011 & 14572 \\
\bottomrule
\end{tabular}
\end{center}
\caption{Operating Experience (Reactor-Years) Accumulated by the Industry at the Time of Each Accident \label{tableopexp}}
\end{table}
Finally, to obtain numerical values we also need to choose a value for the cutoff on the rate of improvement. We choose
\be
\gamma^c = {\ln(50) \over T_{\text{obs}}},
\ee
which indicates our prior assumption that safety standards in the industry have improved by, at most, a factor of $50$ from early commercial nuclear reactors. Of course, the reader can consider other values of the cutoff and as we mentioned earlier it is even possible to remove the cutoff altogether.
With these numerical values, it is possible to plot the various probability distribution functions given above. The posterior probability distributions for $\lambda_T$ and $\gamma$ are plotted in Figure \ref{figpostvary}. The alert reader may note that the probability distribution for $n=5$ is peaked to the right of the distribution for $n=6$. This is because the case with $n=5$ represents the analysis for large-release accidents, where the Fukushima accidents are counted separately. This tends to disfavour large values of $\gamma$, since in this counting, $3$ out of $5$ accidents happened at late times. By thus disfavouring rapid recent improvements in safety, this particular case tends to disfavour low values of $\lambda_T$, and this is why the distribution with $n=6$ is peaked to the right of the case with $n=5$ even though the observed number of accidents is larger in this case.
\begin{figure}[!h]
\begin{center}
\begin{subfigure}[t]{0.4\textwidth}
\includegraphics[height=0.3\textheight]{pvaryplot.eps}
\caption{Current Accident Frequency}
\end{subfigure}
\qquad \qquad \qquad
\begin{subfigure}[t]{0.4\textwidth}
\includegraphics[height=0.3\textheight]{avaryplot.eps}
\caption{Rate of Improvement}
\end{subfigure}
\caption{Posterior Probability Distributions for the two parameters of the model. \label{figpostvary} }
\end{center}
\end{figure}
Eventually we are interested in the probability for the hypothesis that the true frequency of accidents is smaller than or equal to the value predicted by PRA. Using these values, we can now obtain the analogue of Table \ref{praright}. Below, we display the relevant values of $\cumprob(\lambda_T^0 | \nlist)$ at the same points as in Table \ref{praright}. This is the probability that the true current frequency of accidents is smaller than the values given by the PRAs considered below.
\begin{table}[!h]
\resizebox{\textwidth}{!}{
\begin{tabular}{ccccc}
\toprule
{\bf Reactor}& \multicolumn{2}{c}{\bf Probability PRA CDF is right} & \multicolumn{2}{c}{\bf Probability PRA LRF is right} \\ \cmidrule(r){2-3}
\cmidrule(l){4-5}
&{\bf Fukushima separate} & {\bf Fukushima together} & {\bf Fukushima separate} & {\bf Fukushima together} \\ \midrule
{Kudankulam}&$7 \times 10^{-24}$&$2 \times 10^{-17}$&$ 3 \times 10^{-22}$&$5 \times 10^{-14}$ \\
{EPR}& $3 \times 10^{-16}$ & $1 \times 10^{-11}$ & $5 \times 10^{-17}$ & $2 \times 10^{-10}$ \\
{AP1000}&$1 \times 10^{-17}$&$1 \times 10^{-12}$&$1 \times 10^{-17}$& $6 \times 10^{-11}$\\ \bottomrule
\end{tabular}}
\caption{Comparing PRA results with Bayesian estimates from historical observations after allowing exponential increase in safety standards \label{prarightsophisticated}}
\end{table}
We have also plotted this probability as a function of $\lambda_T^0$ in Figure \ref{figcumvary}.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.6\textwidth]{cvaryplot.eps}
\caption{Probability for the hypothesis $\lambda_T < \lambda_T^0$ \label{figcumvary}}
\end{center}
\end{figure}
It is sometimes believe that since safety standards have been improving over time, the results of PRA may be valid for newer reactors even if they are inconsistent with the empirical data for older reactors. The calculations above rule out this possibility and reinforce our conclusion that the results of PRA cannot be reconciled with empirical data.
\subsection{Regional Variations \label{secregional}}
It is clear that the safety of nuclear reactors may vary across regions. However, just as in our analysis of possible improvements in safety over time, we now show that no reasonable variation in safety across regions can lend any confidence to the results of PRA-calculations.
In the case of regional variations, it would be inappropriate to proceed along the lines of \eqref{varylambda} since there is no reason to expect that safety will vary monotonically along any spatial parameterization of nuclear reactors.
However, the central point is that for any particular country or region, the experience of nuclear accidents in the rest of the world provides a reasonable {\em prior} estimate of the frequency of accidents in that region. If we have reason to believe, from the record of smaller incidents or from some other information, that safety in that region is better or worse than other parts of the world, we can account for this factor as well in our prior distribution. This prior distribution can then be corrected using empirical data from the region itself to obtain a final posterior distribution for the frequency of accidents.
We now describe, more precisely, how this procedure can be implemented. Consider a particular region, $R$, which may be a country or a group of countries. In this region, we assume that the probability of an accident in the interval $[t, t+dt]$ is given by \eqref{centralassumption}.
\be
d p = \lambda_R d t
\ee
In this subsection, we do not consider additional variations of accident frequencies with time, to avoid complicating the analysis. In addition, we have the complement of the region (the rest of the world), $\widetilde{R}$, and we assume that in $\widetilde{R}$, the probability of an accident in an interval of length $d t$ is given by
\be
d p = \widetilde{\lambda}_{R} d t
\ee
Assuming a flat prior distribution for $\widetilde{\lambda}_R$, we now construct a posterior probability distribution for this parameter. Assuming that $n_{\rcomp}$ accidents have been observed in the rest of the world in a total operating time $T_{\rcomp}$, we use the techniques of section \ref{frequencypra} to construct the posterior probability distribution for $\lcomp$, which is given by
\be
\probdens(\lcomp | \ncomp) = {1 \over \Gamma(\ncomp + 1) } \Tcomp
(\lcomp \Tcomp)^{\ncomp} e^{-\lcomp \Tcomp}.
\ee
Now, we then additionally input the assumption that the prior distribution for $\lambda_R$ is the same as the posterior distribution for $\lcomp$, except for a constant of proportionality $\kappa$ between them. The factor $\kappa$ indicates our prior belief that nuclear safety in region $R$ is better or worse than that in other parts of the world.
For example, we could obtain an estimate for $\kappa$ using a record of less severe incidents (not necessarily core-damage or large-release) that have occurred in $R$ and $\widetilde{R}$
\be
\kappa = {I_{R} \over \widetilde{I}_{R}},
\ee
where $I_{R}, \widetilde{I}_{R}$ are the number of incidents recorded in $R$ and $\widetilde{R}$ respectively. It is, of course, necessary to have a precise criterion to count the incidents above, and we give one example below. Since incidents of lower severity are fairly frequent, therefore a purely frequentist analysis is sufficient to obtain an estimate for $\kappa$.
After fixing $\kappa$ using this technique, or some other, we then take the prior distribution for $\lambda_R$ to be
\be
\label{priorR}
\probdens(\lambda_R) = \left. {1 \over \kappa} \probdens(\lcomp | \ncomp) \right|_{\lcomp = {\lambda_R \over \kappa}} = {1 \over \kappa^{\ncomp + 1} \Gamma(\ncomp + 1) } \Tcomp
(\lambda_R \Tcomp)^{\ncomp} e^{-{\lambda_R \Tcomp \over \kappa}}
\ee
Now, denoting the number of accidents observed inside region $R$ by $n_R$, in a total operating time $T_R$, we can construct a posterior probability distribution for $\lambda_R$ using the same techniques and the prior in \eqref{priorR}. This leads to
\be
\probdens(\lambda_R | n_R) = {T_R + {\Tcomp \over \kappa} \over \Gamma(n_R + \ncomp + 1)} ((T_R + {\Tcomp \over \kappa}) \lambda_R)^{\ncomp + n_R} e^{-\lambda_R (T_R + {\Tcomp \over \kappa})}
\ee
The probability for the hypothesis that $\lambda_R < \lambda_0$ is given by
\be
\cumprob_R(\lambda_0, n_R) = {\Gamma(1 + n_{R} + \ncomp, (T_{R} + {\Tcomp \over \kappa}) \lambda_0) \over \Gamma(1 +
n_{R} + \ncomp)}.
\ee
So, we see that in this model, we obtain a simple result. The rule is that to account for variations in safety across regions, we simply count the total accidents $\ncomp + n_R$ as having occurred in a time
$T_R + {\Tcomp \over \kappa}$. In the situation where we take $\kappa = 1$, we recover the results of section \ref{frequencypra} since $T_R + \Tcomp$ just becomes the total operating experience $T_{\text{obs}}$.
Now the key point is as follows. If we recall the expansion of the incomplete gamma function shown in \eqref{incompgammafunc} we find that for small $\lambda_0 (T_R + {\Tcomp \over \kappa})$ the expressions above are well approximated by
\be
\cumprob_R(\lambda_0, n_R) \approx {\left(\lambda_0 (T_R + {\Tcomp \over \kappa}) \right)^{n_R + \ncomp + 1} \over \Gamma(n_R + \ncomp + 2)}
\ee
If we now take $\lambda_0$ to be one of the values given in Table \ref{tablepra} we see that the expression above necessarily evaluates to an extremely small number.
To take a numerical example, let us take the region under consideration to be France. At the end of 2012, France had accumulated about $T_R = 1874$ reactor-years of operating experience, without a single large-release event. (See Table 4 in \cite{npriaea2013}.) In the same period the rest of the world had accumulated $\Tcomp = T_{\text{obs}} - T_R = 13373$ reactor-years of operating experience. We also take $\kappa = 0.5$, which suggests a prior belief that French reactors are twice as safe as the world-average. We emphasize that this value of $\kappa$ is being taken here just as an example, and not to suggest that this is really the case.
With these parameters, given that there have been five large release events in the rest of the world, and taking the PRA frequency estimate for the French EPR reactor given above, we see that
\be
\cumprob_R(7.69 \times 10^{-8}, 5) = 1.6 \times 10^{-19}.
\ee
Therefore, given the existing empirical experience in the rest of the world, even with an assumption that French reactors are considerably safer, the probability that the EPR reactor genuinely has a true frequency of accidents as small as the predictions of PRA is absurdly low.
It is clear that the models in this section can be extended further to account for more detailed variations. However, the analysis of this section shows that the severe contradiction between the results of PRA and empirical data cannot be resolved in any such manner.
\section{On the Indian Experience \label{secindia}}
We now turn to the Indian experience. The correct procedure to analyze this case would be to use the rest of the world's experience as a prior distribution as was done in section \ref{secregional}. Needless to say, with any reasonable choice of $\kappa$, as defined there, we would end up with the conclusion that $\epsilon \ll 1$ for India. However, in this short subsection, we want to briefly take a separate approach to make an elementary statistical point. Even if one assumes that the existing record of nuclear accidents has absolutely no bearing on the Indian situation, India's operating experience of $T_{\text{ind}} = 394$ reactor years \cite{npcilmain} is too low to provide any statistical confidence in the safety of the Indian nuclear programme.
The significance of this observation pertains to common claims made by the Indian nuclear establishment that Indian reactors are safer since India has not witnessed a major accident in this time period. The point of this section is to point out that such claims are statistically fallacious. While we have phrased our discussion in India's context, it is worth noting that several countries have accumulated similar levels of operating experience. To take a few examples, by the end of 2012, Belgium had accumulated $254$ reactor-years, China had accumulated $141$ reactor-years, Canada had accumulated $634$ reactor-years, the republic of Korea had accumulated $404$. Although none of these countries have seen a major accident, the conclusions of this section apply to all of them: their operating experience is insufficient to suggest that safety levels in these countries are significantly different from the world-average.
In mathematical terms, in this section our objective is somewhat different from the rest of the paper. Here, we are not trying to estimate $\epsilon$ (from \eqref{centralclaim}) and show that it is very small, but rather we would like to show that even if one discards the reasonable prior assumptions made in section \ref{secregional}, there is no reason to suppose that $\epsilon \sim 1$ for these countries. As explained above, the numerical figures that we take below are specific to India, but the reader can easily modify this calculation to the other countries mentioned above.
There have been several minor but no major accidents in India's operating history.
We again start with a
flat prior distribution for the mean frequency of accidents at Indian reactors, which we denote by $\lambda_{\text{ind}}$ to distinguish it from the global frequency. Then the posterior distribution
for $\lambda_{\text{ind}}$ is given by
\be
{\mathcal{Q}}_{\text{ind}} (\lambda_{\text{ind}} ) = T_{\text{ind}} e^{-\lambda_{\text{ind}}
T_{\text{ind}}},
\ee
Note we can obtain this by setting $n_{\text{obs}} \rightarrow 0$ and $T_{\text{obs}} \rightarrow T_{\text{ind}}$ in \eqref{posterior}.
This curve is plotted in figure \ref{pind}.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.6\textwidth]{indpl.eps}
\caption{Posterior probability distribution for
$\lambda_{\text{ind}}$ \label{pind}}
\end{center}
\end{figure}
This figure tapers off quite gently, so India's current operating
experience cannot tell us much about the frequency of accidents in India, especially if the true frequency is around $\nu^{\text{cd}}_{\text{obs}}$ in \eqref{observedfreq}.
To make this sharper, it is useful to
look at the probability for the hypothesis that
$\lambda_{\text{ind}} < \lambda_0$. This probability is given
by the function
\be
{\cumprob}_{\text{ind}} (\lambda_0) = \int_0^{\lambda_0} {\cal
Q}_{\text{ind}} (\lambda) \, d \lambda = 1 - e^{-T_{\text{ind}} \lambda_0}.
\ee
This curve is shown in figure \ref{cumprobind}.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.6\textwidth]{indcl.eps}
\caption{Probability for the hypothesis $\lambda_{\text{ind}} < \lambda_0$ \label{cumprobind}}
\end{center}
\end{figure}
Some relevant numerical figures are
\be
\begin{split}
&{\cumprob}_{\text{ind}}(2.67 \times 10^{-4}) \approx 0.10, \\
&{\cumprob}_{\text{ind}}(1.0 \times 10^{-3}) \approx 0.33.
\end{split}
\ee
In words, this indicates that it is only with a confidence of about 10\%
that we can state that the true frequency of accidents is smaller than
once in 3700 reactor-years.
And it is only with a confidence of about 33\% that we can conclude that
the true frequency of accidents in India is smaller than once per thousand reactor-years.
These results are very simple to understand intuitively. They simply
reflect the fact that even if the expected frequency of accidents in India is once every thousand reactor-years, there is still an excellent chance that one could get through $394$ reactor-years
without an accident. Conversely, the absence of an accident in this time period
is not particularly informative.
It is quite common to find claims like the one made by the Indian Nuclear Society its advertisement for a 2012 conference. ``Having achieved safe and reliable operation of about 360 reactor- years \ldots the Indian nuclear programme has demonstrated a high level of maturity. The safety track record of Indian nuclear power plants has been impeccable'' \cite{insac2012}.
These claims are accepted at the highest levels of government. In 2011, the Prime Minister stated that ``The safety track record of our nuclear power plants over the past 335 reactor-years of operation has been impeccable'' \cite{pmdaespeech}.
Our analysis shows that it is erroneous to draw these complacent conclusions.
India's operating experience is too limited to
provide statistically reliable long term estimates about the efficacy or otherwise of safety practices in the nuclear sector.
In passing, we should point out that, within the nuclear industry, it appears
to be a rather common statistical error to extrapolate from limited experience
to rather strong claims of safety.
For example, the Rasmussen report also stated that ``It is significant that in some 200 reactor-years of commercial operation of reactors of the type considered in the report there have been no fuel melting accidents.'' (See p. 6 of \cite{rasmussen}.) However, this experience had absolutely no significance for the conclusion drawn by the report, which was that the core-damage frequency of reactors was once in 20,000 years (p. 8).
\section{The Debate on Nuclear Liability \label{secliability}}
It is clear that the results of PRA are untenable in the light of empirical data. In this section we provide evidence, using the debate on nuclear liability, that the industry's actions (as opposed to its
public statements) suggest that it has independently reached this conclusion.
Soon after the nuclear deal, multinational nuclear suppliers lobbied
the government to pass a law that would indemnify them in
the event of an accident. Under this pressure, the government
passed a liability law in 2010
that was almost identical
to the annex of a U.S. sponsored
international convention on the subject called the Convention on Supplementary Compensation \cite{suvratramanaliab,suvrat_raju_strange_2011}. However, as a result of various pulls-and-pushes in the legislative process, nuclear
suppliers are largely protected but not completely indemnified by the Indian law. While victims cannot sue the supplier, the law allows the NPCIL, which has the primary
responsibility of compensating victims, to recover some of this compensation
from the supplier
for a disaster caused
by substandard equipment. The refusal of suppliers to accept even this marginal
liability has presented a significant obstacle for contracts for new reactors. (See for example, the recent statement by the CEO of General Electric \cite{gerefusal2015}.)
It is significant that although one of primary advertised benefits of the Indo-US nuclear deal was that India would be able to purchase light water reactors from new suppliers, the deadlock on liability has prevented this entirely. The first new contract for reactors after the nuclear deal was signed in 2014 with the Russian public sector company Rosatom. However, this will just extend the Kudankulam nuclear complex which was covered by an inter-governmental agreement even before the nuclear deal. Negotiations with suppliers from markets that have opened up after the nuclear deal, including companies from France and the United States, have so-far failed to yield a single new contract due to the conflict on liability.
However, this leads directly to the following question: if the chance of a nuclear accident is
indeed as remote as the industry claims, then why are nuclear reactor
manufacturers unwilling to accept liability for an accident?
One of the ostensible reasons given by suppliers is that forcing them to accept liability would cause the cost of power to go up \cite{russiandeputypm}.
To examine the veracity of this claim, consider the expected cost of insurance for suppliers if the results of PRA are taken at face value by the industry and actuaries \cite{suvratramanahindu2012}.
The Indian liability law currently caps the total available compensation for victims at ``the rupee equivalent of three hundred million Special Drawing Rights''(SDRs). (See clause 6 of the text of the law \cite{gazetteliability}.) This includes compensation from the operator, and also the central government, and victims are not legally entitled to any further compensation. Here we consider the worst case scenario for the supplier, where it becomes liable for this entire amount.
Although the rupee to SDR exchange rate fluctuates, this maximum liability is approximately
$\ell_{\text{cap}} = \text{Rs.}~2,500~\text{crores}=\text{Rs.}~2.5
\times 10^{10}$ in rupee terms.
Taking, for example, the claimed frequency of core-damage accidents at the Kudankulam reactors, which we denote by $\mu_{\text{kk}} = 10^{-7} \ry$, then a simple order
of magnitude estimate for the cost of insurance for this amount is
\be
i_p = \mu_{\text{kk}} \times {\ell}_{\text{cap}} = \text{Rs.}~
2,500.
\ee
A reactor with a capacity of 1000 MW, operating at a 80\% load-factor, should produce
$E = 0.8 \times 10^6 \times 365 \times 24~ \text{kWh} \approx 7 \times 10^9~ \text{kWh}$
of electricity each year. So, the cost of insurance above should lead to an
increase in the cost of electricity by
\[
\delta p = {i_p \over E} = 3.6 \times 10^{-7}~ \text{Rs.} /\text{kWh}!
\]
This absurdly small number indicates that something is amiss with the industry's claim that liability
will lead to price increases.
In fact two factors of about about $10^3$ and $10^4$, which are missing
in the calculation above, are required to make sense of the industry's reluctance. The first is that nuclear accidents could lead to damage that is a thousand times more than the cap on liability.
For example, some estimates of the economic damage at Fukushima are as high as $\ell_{\text{real}} \approx \text{USD}~200~\text{billion} \approx ~\text{Rs.}~12~\text{lakh~crores}$. (See Table 4 of \cite{jcerfukushima}.)
In principle, a future Indian government could ignore the liability cap and insist
on recovering larger costs from the supplier.
Even this would not lead to a prohibitive cost of insurance if reactors
were genuinely as safe as manufacturers' claim.
The other crucial missing factor comes from our result above: accidents
affecting the public are likely to happen at a rate that is closer to $\nu^{\text{lr}}_{\text{obs}}$.
Extending our simple linear model with these realistic estimates of damage
and risk leads to the following cost of insurance per unit energy produced
\[
\delta p_{\text{true}} = {\nu^{\text{lr}}_{\text{obs}} \times \ell_{\text{real}} \over E} = 0.56~\text{Rs.}/\text{Kwh}.
\]
This is now a significant fraction (roughly 10\%) of the cost of electricity. However, at this point, corrections to our linear model for the insurance premium become significant.
For example, since the total amount involved, $\ell_{\text{real}}$, is very high, and the
expected rate $\nu^{\text{lr}}_{\text{obs}}$ is non-negligible, financial institutions would evidently
be unwilling to underwrite this risk without additional incentives
in the form of a significantly higher cost of insurance. This helps explain why suppliers insist on legislative indemnity, rather than simply arranging
for the appropriate financial cover.
What the debate on liability shows is that the nuclear industry --- both in the private and the public sector --- has itself taken note of the empirical
rates of accidents, and it is unwilling to take the predictions of its PRAs seriously when its economic
interests are at stake.
\section{Conclusions \label{secconclude}}
In this paper, by means of some simple Bayesian calculations, we have come
to the following conclusion: the historical record contradicts the predictions of probabilistic risk assessment and suggests a significantly higher risk of nuclear accidents. The contradiction between these predictions and data can be quantified in terms of a probability for the hypothesis that true frequencies of accidents are as small as those predicted. This probability can be quantified in various models, and is found to be extraordinarily small in a model-independent fashion, and independent of changes in our detailed assumptions. In particular, in section \ref{secrobust} we showed how, even in models where safety standards improve exponentially with operating experience or where reactors in a given region are assumed to be considerably safer than reactors elsewhere, the conclusions above hold at very high confidence levels.
Second, we also specifically discussed the case of India. Although this is strictly speaking, a subset of the study of worldwide accident frequencies above, we analyzed it separately to show that India's current India's current experience with nuclear reactor operation is far from sufficient to draw any strong conclusions about future reactor safety. The same conclusions hold for other countries with similar levels of operating experience, such as Canada, China, Belgium and Korea.
Therefore it is clear that the methodology and practice of PRA needs to be revised significantly.
In fact, as we showed in section \ref{secliability}, it is clear from the debate on nuclear liability that the nuclear industry already recognizes that the results of PRA are numerically unreliable. Nevertheless, within the technical community, the use of PRA is sometimes justified as a useful tool for safety analysis. For example, the authors of \cite{sornette2013exploring} explained that even though PRA is ``not thought to represent the true risks'' it remains useful as a ``platform for technical exchanges on safety matters between regulators and the industry.'' However, this begs the question: why has PRA failed so badly in achieving the purpose that it was designed for. Indeed, the Rasmussen report \cite{rasmussen} started by explaining that ``the objective of the study was to make a realistic estimate of [the] risks'' that would be ``involved in potential accidents.''
Apart from the theoretical problems mentioned in the introduction, it appears likely that the nuclear industry benefits from the disingenuous suggestion that it can, in fact, accurately predict the frequency of accidents. Although insiders recognize that this is not the case, it is clear that the detailed computer simulations that support a supposedly-scientific calculation of low accident-frequencies computed to several decimal places are useful in public debates. While there have been other critiques of the mismatch between the results of PRA and empirical data, we believe that this study is significant because it emphasizes the extraordinarily high level of confidence with which it is possible to rule out the results of PRA.
Indeed, this would hardly be the only situation in which the nuclear industry has attempted to use the authority of science to dismiss safety concerns. For example, in an attempt to dismiss the history of Chernobyl, the World Nuclear Association declared \cite{wnasafetyarchive2011} in January 2011 that ``In the light of better understanding of the {\em physics and chemistry} of material in a reactor core \ldots it became evident that even a severe core melt coupled with breach of containment could not in fact create a major radiological disaster from any Western reactor design'' (emphasis added). After attempting to initially defend this claim for a few days after Fukushima by claiming that ``clearly there was no major release from the reactors''\cite{wnasafetyarchive2011_afterfuk} and only from the ``fuel pools'', the Association had to reluctantly concede that its claim, seemingly based on rigorous material science, ``did not apply to all'' \cite{wnasafetyarchive2011_concession} Western reactor designs.
It is interesting to note that a similar dynamic operates in other industries as well. For example, in the aviation industry (which inspired the use of
PRA for nuclear reactors), as part of the certification process for its new 787 ``Dreamliner'' aircraft, Boeing estimated that its lithium-ion batteries would vent smoke ``once in every 10 million flight hours.'' In fact this event occurred twice in 52,000 flight hours leading to the grounding of the entire
fleet for inspection \cite{dreamlinerinterim}
To explore the implications of our conclusions, we return to the case of India where the government is planning a large nuclear expansion.
It has announced plans to commence construction on 8 heavy water reactors, with a capacity of 5600 MW in the ``12\textsuperscript{th} Plan'' period (2012--17), and complete work on a separate 2800 MW of installed capacity. In addition, it is also planning to import 8 reactors with a total capacity of 10,500 MW \cite{narayanasamy_setting_2012-1}.
Every reactor site has seen
vigorous local protest movements that have raised issue of land and livelihood but also questions about nuclear safety.
Therefore it is imperative to have a frank conversation on nuclear safety, involving not just the technical community but a far broader cross-section of society. Our results show that
such a debate should start with the acceptance that the ambitious
claims about nuclear safety made on the basis of probabilistic risk assessment
have been conclusively falsified by the empirical data.
\paragraph*{Acknowledgments}
This article is based on a detailed exchange of letters with Mr. Nalinish Nagaich of the NPCIL. I am grateful to M. V. Ramana for comments on a draft of this manuscript.
\bibliographystyle{ieeetrurl}
|
\section{Introduction}
The study of non-perturbative methods for nonlinear differential equations is
of considerable recent interest. Among the various well known singular
perturbation techniques such as multiple scale analysis, method of boundary
layers, WKB method and so on \cite{Jordan Smith, Bender Orszag}, the recently
developed homotopy analysis method (HAM) \cite{Liao HAM, Lopez VDP Amp} and
the Renormalization group method (RGM) \cite{Chen G O RG 1994, Chen G O RG,
DeVille RG, Sarkar Bhattacharjee} appear to be very attractive. The aim of
these new improved methods is to derive in an unified manner uniformly valid
asymptotic quantities of interest for a given nonlinear dynamical problem.
Although formulated almost parallely over the past decades or so, relative
strength and weakness of these two approaches have yet to be investigated
systematically. The purpose of this paper is to undertake a comparative study
of HAM and RGM in the context of the Rayleigh equation%
\begin{equation}
\ddot{y}+\varepsilon\left( \frac{1}{3}\dot{y}^{3}-\dot{y}\right) +y=0
\label{Rayleigh}%
\end{equation}
and the Van der Pol equation%
\begin{equation}
\ddot{x}+\varepsilon~\dot{x}\left( x^{2}-1\right) +x=0 \label{VdP}%
\end{equation}
where the dots are used to designate the derivative with respect to time $t$.
The Rayleigh and the Van der Pol (VdP) equations represent two closely related
nonlinear systems and have found significant applications in the study of self
excited oscillations arising in biology, acoustics, robotics, engineering etc.
\cite{Acoustics, Robotics}. It is easy to observe that differentiating
$\left( \text{\ref{Rayleigh}}\right) $ with respect to time $t$ and putting
$\dot{y}\left( t\right) =x\left( t\right) $ we obtain $\left(
\text{\ref{VdP}}\right) $. Both these systems have unique isolated periodic
orbit (limit cycle). The amplitude of a periodic oscillation $y(t)$ $\left(
\text{or }x\left( t\right) \right) $ is generally defined by $\max
\left\vert {y}\left( {t}\right) \right\vert $ $\left( \text{or }%
\max\left\vert {x}\left( {t}\right) \right\vert \right) $ over the entire
cycle. It is well known that the naive perturbative solutions of these
equations are useful when $0<\varepsilon\ll1$ and yields the asymptotic value
$a(\varepsilon)\approx2$ of the amplitude for the limit cycle correctly. For
$\varepsilon\gg1$, simple analysis based on singular perturbation theory also
yields the asymptotic amplitude for the relaxation oscillation as
$a(\varepsilon)\approx2$ for the VdP equation. However, the conventional
perturbative approaches fail when $\varepsilon$ is finite. One of the aim of
this paper is to determine efficient approximate formulae for the amplitude of
the limit cycle for the above systems by both HAM and RGM. Lopez et al
\cite{Lopez VDP Amp} have reported an efficient formula for the amplitude of
the VdP limit cycle by HAM. We note here that a key difference in Rayleigh and
VdP oscillators is the fact that with increase in input energy (voltage), the
amplitude of the Rayleigh periodic oscillation increases, when that of the VdP
oscillator remains almost constant at the value 2, with possible increase in
the corresponding frequency only. For large $\varepsilon\ (\geq1)$ relaxation
oscillations, on the other hand, the Rayleigh system shows up a rather fast
building up and slow subsequent release of internal energy, when the VdP
models the reverse behaviour, with slow rise and fast drop in the accumulated energy.
As remarked above, HAM and RGM are formulated to determine the uniformly valid
global asymptotic behaviours of relevant dynamical quantities like amplitude,
period, frequency etc. related to periodic solutions of these equations for
finite values of $\varepsilon$, by devising efficient methods in eliminating
divergent secular terms of the naive perturbation theory. HAM seems to have
the advantage of yielding uniformly convergent solutions of very high order in
the nonlinearity parameter $\varepsilon$ utilizing a freedom in the choice of
a free parameter $h$. The computation of higher order term could be
facilitated by symbolic computational algorithms. This method is used to
obtain good approximate solutions for the VdP equation by a number of authors
\cite{Lopez VDP Amp, Liao VDP}. Lopez et al \cite{Lopez VDP Amp} derived
efficient formulae for estimating the amplitude of the limit cycle of the VdP
equation for all values of $\varepsilon>0$. Although, HAM is now considered to
be an efficient method in the study of non-perturbative asymptotic analysis,
it is recently pointed out \cite{Meijer} that this method might fail even in
some innocent looking nonlinear problems.
The RGM, on the other hand, has a rich history, being originally formulated
for managing divergences in the quantum field theory and later having deep
applications in phase transitions and critical phenomena in statistical
mechanics. Subsequently, Chen et al \cite{Chen G O RG 1994, Chen G O RG}
successfully translated the RG formalism into the study of nonlinear
differential equations. It is noted that RGM is more efficient and accurate
than conventional singular perturbative approaches in obtaining global
informations from a naive perturbation series in $\varepsilon$. It is also
recognized that RGM generated expansions yield $\varepsilon$-dependent
space/time scales naturally, when conventional approaches normally require
invoking such scales in an ad hoc manner. The pertubative RGM, however,
appears to have the limitation that the computation of higher order
renormalized solutions could be quite involved and tedious. More serious is
the inability of assuring the convergence of the renormalized expansions for
large nonlinearity parameter. Further, there is still no evidence in the
literature that RGM could be employed successfully to asymptotic estimation of
the amplitude of an isolated periodic orbit for all values of $\varepsilon$ as
was reported for HAM \cite{Lopez VDP Amp}.
Here we report analytic expressions of the amplitude of the periodic solutions
of both the Rayleigh equation $\left( \text{\ref{Rayleigh}}\right) $ and the
VdP equation $\left( \text{\ref{VdP}}\right) $ as functions of $\varepsilon
$. We have made a comparative study of these two sets of formulae using both
HAM and RGM. The HAM contains a control parameter $h=h\left( \varepsilon
\right) $ which controls the convergence of the approximation to the
numerically computed exact value of the amplitude for all values of
$\varepsilon$. Suitable choice of $h$ can control the relative percentage error.
The original RGM gives an approximation to the exact solution for small values
of $\varepsilon$. We report here the RG solution upto order $3$. To the
authors' knowledge this seems to be the first higher order computation other
than second order computations reported so far by various authors \cite{Chen G
O RG, Sarkar Bhattacharjee}. A comparison of the amplitude of the periodic
cycle with the exact computations reveals that even the present higher order
perturbative approximations fails to give accurate estimation for moderate
values of $\varepsilon$. As the higher order RG computations are quite
laborious and inefficient, it is very unlikely that higher order computations
of amplitude would improve the quality of the estimated amplitude of the limit
cycle. Further, in RGM one does not have the resource of a free parameter
equivalent to $h(\varepsilon)$ of HAM to improve the convergence of the RG expansions.
A major contribution in the present study is to propose an \emph{improved} RGM
(IRGM). In IRGM, we advocate the concept of \emph{nonlinear time} \cite{DPD
Raut, DPD, DPD Sen} that extends the original RG idea of eliminating the
divergent secular term of the form $(t-t_{0})\sin t$, where $t_{0}$ is the
initial time, of the naive perturbation series for the solution of the
nonlinear problem, by exploiting the arbitrariness in fixing the initial
moment $t_{0}$. The original prescription rests on introducing new initial
time $\tau$ in the form $(t-\tau+\tau-t_{0})\sin t$ and to allow the
renormalized amplitude $R=R(\tau)$ and phase $\theta=\theta(\tau)$ of the
renormalized solution to depend on the new parameter, viz., $\tau-t_{0}$ so
that the original naive perturbative, constant values of amplitude $R_{0}$ and
phase $\theta_{0}\ (=0)$ (say) `flow' following the RG flow equations of the
form
\begin{equation}
\frac{dR}{d\tau}=f\left( R,\varepsilon\right) ,\ \frac{d\theta}{d\tau
}=g\left( R,\varepsilon\right) \label{rgflow}%
\end{equation}
The functions in the right hand sides of the RG flow equations, in general,
should depend both on $R$ and $\theta$, besides the explicit $\varepsilon$
dependence. We suppress the $\theta$ dependence for simplicity that should
suffice for our present analysis of the Rayleigh and VdP equations $($c.f.
equations $\left( \text{\ref{Amp Eq Order 3}}\right) $, $\left(
\text{\ref{Phase Eq Order 3}}\right) )$. The flow equations are derived from
the consistency condition that the actual renormalized solution $y\left(
t,\tau\right) $ should be independent of the arbitrary initial adjustment
$\tau$: $\frac{\partial y}{\partial\tau}=0$. The final form of the uniformly
valid RG solution $y_{R}\left( t\right) $ is obtained by setting $\tau=t$
that eliminates the secular terms. Let us remark here that the actual
convergence of the RG expansions is not well addressed and should require
further investigations. Moreover, estimation of asymptotic amplitude for a
limit cycle as $t\rightarrow\infty$, for instance, from the perturbation
expansion of $f$ is expected to fail for $\varepsilon>\approx O\left(
1\right) $.
In the framework of nonlinear time, we suppose the arbitrary initial time
$\tau$ to depend explicitly on the nonlinearity parameter (coupling strength)
$\varepsilon$ of the nonlinear equation, so that one can write $\tau
/\varepsilon=\varepsilon^{h}$ where $h=h\left( \varepsilon t\right)
,\ \varepsilon t>1$ is a slowly varying (almost constant), free (asymptotic)
control parameter for $t\rightarrow\infty$ and $\varepsilon\ \rightarrow$
either to 0 or $\infty$, to be utilized judiciously to improve the convergence
and non-perturbative global asymptotic behaviour of the original RG proposal
($h<0$ for $0<\varepsilon<1$). In Appendix, we give an overview, in brief, of
an extended analytic framework that naturally supports nontrivial existence of
such an asymptotic scaling parameter $h(\tilde{\tau})$ as a function of the
rescaled $O(1)$ variable $\tilde{\tau}=\varepsilon t\sim O(1)$, satisfying
what we call the \emph{principle of duality structure}. The secular terms in
the naive perturbation series would now be altered instead as $\left(
t-\tau/\varepsilon+\tau/\varepsilon-t_{0}\right) \sin t$ and we obtain the
new RG flow equations in the form
\begin{equation}
\frac{dR}{d\tau}=f_{0}\left( R\right) \left( 1+O\left( \varepsilon
^{2}\right) \right) ,\ \frac{d\theta}{d\tau}=\varepsilon g_{1}\left(
R\right) \left( 1+O\left( \varepsilon^{3}\right) \right) \label{rgflown}%
\end{equation}
where $f_{0}(R)$ and $g_{1}(R)$ are nonzero, minimal order $R$ dependent terms
in the respective perturbation series. Following the analogy of RG
prescription in annulling secular divergence through corresponding `flowing'
of the renormalized perturbative amplitude and phase, we next make \emph{the
key assumption that there exists, for a given nonlinear oscillation, a set of
right control parameters $h^{i}$ that would absorb any possible secular or
other kind of divergence in the higher order perturbation series}, (see
Appendix for justification), so that in the asymptotic limit $t\rightarrow
\infty$, one obtains the finite, non-perturbative flow equations directly for
the periodic orbit of the nonlinear system
\begin{equation}
\frac{da}{d\tau_{1}}=f_{0}(a),\ \frac{d\theta}{d\tau_{2}}=g_{1}%
(a)\label{rgflownp}%
\end{equation}
where $\tau_{i}=\varepsilon^{ih_{RG}^{i}(\tilde{\tau})}$, and $h_{RG}%
^{i}(\tilde{\tau})$ is a finite \emph{scale} independent control parameter in
the rescaled variable $\tilde{\tau}\sim$ $O(1)$ and $a(\varepsilon
)=\underset{t\rightarrow\infty}{\lim}R(\varepsilon,t)$ is the $\varepsilon$-
dependent amplitude of the limit cycle. A simple quadrature formula should
then relate the control parameter $h_{RG}=h_{RG}^{1}$ with the amplitude
$a(\varepsilon)$. As a consequence, adjusting the control parameter $h_{RG}$
suitably, one can generate an efficient algorithm to estimate the amplitude
$a(\varepsilon)$ that would compare well with the exact values, upto any
desired accuracy. It will transpire that the control parameter $h_{RG}%
(\varepsilon)$ must respect some asymptotic conditions depending on the
characteristic features of a particular relaxation oscillation $($c.f. Section
\ref{Sec M-RG Sol Description}$)$.
Exploiting the rescaling symmetry, one may as well rewrite the above
non-perturbative flow equations (\ref{rgflownp}) in the equivalent
$\tilde{\tau}\sim O(1)$-dependent scaling variable $\tau={\tilde{\tau}%
}^{H_{RG}(\tilde{\tau})}$, $\left( H_{RG}(\tilde{\tau})=h_{RG}(\tilde{\tau
})\frac{\log\tilde{\tau}}{\log\varepsilon}\right) $, for each fixed value of
the nonlinearity parameter $\varepsilon$ that should expose small scale
$\tilde{\tau}\sim O(1)$-dependent variation of the amplitude. As a biproduct
that would allow one to retrieve an efficient approximation of the limit cycle
orbit for the nonlinear oscillator. It turns out that the general framework of
IRGM is quite successful in obtaining excellent fits for the limit cycle orbit
even for relaxation oscillation corresponding to nonlinearity parameters
$\varepsilon\geq1$.
It follows that the application of the of idea of nonlinear time
in RG formalism offers one with a robust formalism
for global asymptotic analysis for a general nonlinear system that might even
be advantageous in many respects compared to HAM. The application of nonlinear
time in HAM will be considered separately.
The paper is organized as follows. In Section \ref{Sec HAM Sol} we have
deduced the solution to the equation $\left( \text{\ref{Rayleigh}}\right) $
by HAM. In Section \ref{Sec RG Sol} we compute the classical RG solution upto
$O(\varepsilon^{3})$ order and compare estimated values of the limit cycle
amplitude with the exact values. The improved RG method is presented in
Section \ref{Sec M-RG Sol Description}. This introduces a control parameter
$h_{RG}$ in the RG analysis. In Subsection \ref{SubSec Amplitude Estimation}
approximate analytic formulae are deduced for the amplitudes of the limit
cycles of the Rayleigh and the VdP equations. Efficient match with the exact
values can be obtained by appropriate choice of $h_{RG}$. In Subsection 4.2 we
present the efficient of approximate limit cycle orbits for the Rayleigh and VdP
oscillators for $\varepsilon=5$. We close our discussions in Sec. 5. In Appendix 1,
we present a brief outline of the formal structure of the analytic formalism presented here.
In Appendix 2, an alternative approach in the derivation of non-perturbative flow equations is presented.
\section{Computation of Amplitude by HAM\label{Sec HAM Sol}}
The Homotopy Analysis method proposed by Liao \cite{Liao HAM, Liao VDP} is
used to obtain the solution of non-linear equation even if the problem does
not contain a small or large parameter. HAM always gives a family of functions
at any given order of approximation. An auxiliary parameter $h$ is introduced
in HAM to control the convergence region of approximating series involved in
this method to the exact solution. HAM is based on the idea of homotopy in
topology. In simple language, it involves continuous deformation of the
solution of a linear ordinary differential equation (ODE) to that of desired
nonlinear ODE. The solution of linear ODE gives a set of functions called
\textit{base functions}. One advantage of HAM is that it can be used to
approximate a nonlinear problem by efficient choice of different sets of base
functions. A suitable choice of the set of base functions and the convergence
control parameter can speed up the convergence process.
In this paper we consider the self-excited system $\left(
\text{\ref{Rayleigh}}\right) $, which can be written as the ODE%
\begin{equation}
\ddot{U}\left( t\right) +\varepsilon\left( \frac{1}{3}\dot{U}^{3}\left(
t\right) -\dot{U}\left( t\right) \right) +U\left( t\right) =0,\quad
t\geq0 \label{RL Eq}%
\end{equation}
where the dot denotes the derivative with respect to the time $t$. A limit
cycle represents an isolated periodic motion of a self-excited system. This is
an isolated closed curve $\Gamma$ $\left( \text{say}\right) $ in the phase
plane so that any path in its suitable small neighbourhood starting from a
point, specified by some given initial condition, ultimately converges to
$($or diverge from$)$ $\Gamma$. Consequently, this periodic motion represented
by limit cycle is independent of initial conditions. It, however, involves the
frequency $\omega$ and the amplitude $a$ of the oscillation. Therefore,
without loss of generality, we consider an initial condition%
\begin{equation}
U\left( 0\right) =a,\qquad\dot{U}\left( 0\right) =0\text{.} \label{IC}%
\end{equation}
In \cite{Chen G O RG}, an alternative initial condition i.e. $U(0)=0,\ \dot
{U}(0)=a$ was considered. Let, with slight abuse of notations,
\[
\tau=\omega t\text{ and }U\left( t\right) =a~u\left( \tau\right) \text{.}%
\]
so that $\left( \text{\ref{RL Eq}}\right) $ and $\left( \text{\ref{IC}%
}\right) $ respectively become%
\begin{equation}
\omega^{2}u^{\prime\prime}\left( \tau\right) +\varepsilon\left( \frac{1}%
{3}a^{2}\omega^{2}u^{\prime2}\left( \tau\right) -1\right) \omega u^{\prime
}\left( \tau\right) +u\left( \tau\right) =0 \label{Normalized RL}%
\end{equation}
and%
\begin{equation}
u\left( 0\right) =1\text{,\quad}u^{\prime}\left( 0\right) =0\text{.}
\label{Normalized-IC}%
\end{equation}
Since the limit cycle represents a periodic motion, so we suppose that the
initial approximation to the solution $u\left( \tau\right) $ to $\left(
\text{\ref{Normalized RL}}\right) $ can be taken as%
\[
u_{0}\left( \tau\right) =\cos\tau
\]
Let, $\omega_{0}$ and $a_{0}$ respectively denote the initial approximations
of the frequency $\omega$ and the amplitude $a$.
We consider a linear operator%
\begin{equation}
\mathcal{L}\left[ \phi\left( \tau,p\right) \right] =\omega_{0}^{2}\left[
\frac{\partial^{2}\phi\left( \tau,p\right) }{\partial\tau^{2}}+\phi\left(
\tau,p\right) \right] \label{L Operator}%
\end{equation}
so that for the coefficients $C_{1}$ and $C_{2}$%
\begin{equation}
\mathcal{L}\left( C_{1}\sin\tau+C_{2}\cos\tau\right) =0 \label{L IC}%
\end{equation}
We further consider a nonlinear operator%
\begin{align}
& \mathcal{N}\left[ \phi\left( \tau,p\right) ,\Omega\left( p\right)
,A\left( p\right) \right] \nonumber\\
& =\Omega^{2}\left( p\right) \frac{\partial^{2}\phi\left( \tau,p\right)
}{\partial\tau^{2}}+\varepsilon\left[ \frac{1}{3}A^{2}\left( p\right)
\Omega^{3}\left( p\right) \left( \frac{\partial\phi\left( \tau,p\right)
}{\partial\tau}\right) ^{3}-\Omega\left( p\right) \left( \frac
{\partial\phi\left( \tau,p\right) }{\partial\tau}\right) \right]
+\phi\left( \tau,p\right) \text{.} \label{NL Operator}%
\end{align}
Next, we construct a homotopy as%
\begin{equation}
\mathcal{H}\left[ \phi\left( \tau,p\right) ,h,p\right] =\left(
1-p\right) \mathcal{L}\left[ \phi\left( \tau,p\right) -u_{0}\left(
\tau\right) \right] -h~p\mathcal{~N}\left[ \phi\left( \tau,p\right)
,\Omega\left( p\right) ,A\left( p\right) \right] \label{Homotopy}%
\end{equation}
where $p\in\left[ 0,1\right] $ is the embedding parameter and $h$ a non-zero
auxiliary (control) parameter used to improve the convergence of series
expansions. Setting $\mathcal{H}\left[ \phi\left( \tau,p\right)
,h,p\right] =0$ we obtain zero-th order deformation equation%
\begin{equation}
\left( 1-p\right) \mathcal{L}\left[ \phi\left( \tau,p\right)
-u_{0}\left( \tau\right) \right] -h~p\mathcal{~N}\left[ \phi\left(
\tau,p\right) ,\Omega\left( p\right) ,A\left( p\right) \right] =0
\label{Zero Deform Eq}%
\end{equation}
subject to the initial conditions%
\begin{equation}
\phi\left( 0,p\right) =1,\quad\left. \frac{\partial\phi\left(
\tau,p\right) }{\partial\tau}\right\vert _{\tau=0}=0\text{.}
\label{Zero Deform IC}%
\end{equation}
Clearly, as $p$ increases from $p=0$ to $p=1$, $\left(
\text{\ref{Zero Deform Eq}}\right) $ changes from $\mathcal{L}\left[
\phi\left( \tau,p\right) -u_{0}\left( \tau\right) \right] =0$ to
$\mathcal{N}\left[ \phi\left( \tau,p\right) ,\Omega\left( p\right)
,A\left( p\right) \right] =0$ and as a consequence $\phi\left(
\tau,p\right) $ varies from the initial guess $\phi\left( \tau,0\right)
=u_{0}\left( \tau\right) =\cos\tau$ to the exact solution $\phi\left(
\tau,1\right) =u\left( \tau\right) $, so does $\Omega\left( p\right) $
from $\omega_{0}$ to exact frequency $\omega$ and $A\left( p\right) $ from
$a_{0}$ to the exact amplitude $a$. It can be shown that assuming $\phi\left(
\tau,p\right) $, $\Omega\left( p\right) $, $A\left( p\right) $ analytic
in $p\in\left[ 0,1\right] $ so that%
\begin{equation}
u_{k}\left( \tau\right) =\frac{1}{k!}\left. \frac{\partial^{k}}{\partial
p^{k}}\phi\left( \tau,p\right) \right\vert _{p=0},\quad\omega_{k}=\frac
{1}{k!}\left. \frac{\partial^{k}}{\partial p^{k}}\Omega\left( p\right)
\right\vert _{p=0},\quad a_{k}=\frac{1}{k!}\left. \frac{\partial^{k}%
}{\partial p^{k}}A\left( p\right) \right\vert _{p=0}
\label{k Deform Derivative}%
\end{equation}
we have,%
\begin{align}
u\left( \tau\right) & =%
{\textstyle\sum\limits_{k=0}^{\infty}}
u_{k}\left( \tau\right) \label{Sol Series}\\
\omega & =%
{\textstyle\sum\limits_{k=0}^{\infty}}
\omega_{k}\label{Freq Series}\\
a & =%
{\textstyle\sum\limits_{k=0}^{\infty}}
a_{k} \label{Amp Series}%
\end{align}
where $u_{k}\left( \tau\right) $ are solutions of the $k$-th order
deformation equation%
\begin{equation}
\mathcal{L}\left[ u_{k}\left( \tau\right) -\chi_{k}u_{k-1}\left(
\tau\right) \right] =h~R_{k}\left( \tau\right) \label{k Deform Eq}%
\end{equation}
subject to the initial conditions%
\begin{equation}
u_{k}\left( 0\right) =0,\qquad u_{k}^{\prime}\left( 0\right) =0
\label{k Deform IC}%
\end{equation}
in which%
\begin{align}
R_{k}\left( \tau\right) & =\frac{1}{\left( k-1\right) !}\left.
\frac{\partial^{k-1}}{\partial p^{k-1}}\mathcal{N}\left[ \phi\left(
\tau,p\right) ,\Omega\left( p\right) ,A\left( p\right) \right]
\right\vert _{p=0}\nonumber\\
& =%
{\textstyle\sum\limits_{n=0}^{k-1}}
u_{k-1-n}^{\prime\prime}\left( \tau\right)
{\textstyle\sum\limits_{j=0}^{n}}
\omega_{j}\omega_{n-j}+u_{k-1}\left( \tau\right) \nonumber\\
& +\frac{\varepsilon}{3}%
{\textstyle\sum\limits_{n=0}^{k-1}}
{\textstyle\sum\limits_{i=0}^{n}}
\left(
{\textstyle\sum\limits_{r=0}^{i}}
a_{r}a_{i-r}\right) \times\left(
{\textstyle\sum\limits_{s=0}^{n-i}}
\omega_{s}%
{\textstyle\sum\limits_{h=0}^{n-i-s}}
\omega_{h}\omega_{n-i-s-h}\right) \nonumber\\
& \times\left(
{\textstyle\sum\limits_{j=0}^{k-1-n}}
u_{j}^{\prime}\left( \tau\right)
{\textstyle\sum\limits_{m=0}^{k-1-n-j}}
u_{m}^{\prime}\left( \tau\right) u_{k-1-n-j-m}^{\prime}\left( \tau\right)
\right) -\varepsilon%
{\textstyle\sum\limits_{n=0}^{k-1}}
\omega_{n}u_{k-1-n}^{\prime}\left( \tau\right) \label{k Deform RHS}%
\end{align}
and%
\begin{equation}
\chi_{k}=\left\{
\begin{array}
[c]{cc}%
0, & k\leq1,\\
1, & k>1.
\end{array}
\right. \label{Chi}%
\end{equation}
To ensure that the solution to the $k$-th order deformation equation $\left(
\text{\ref{k Deform Eq}}\right) $ do not contain the secular terms $\tau
\sin\tau$ and $\tau\cos\tau$ the coefficients of $\sin\tau$ and $\cos\tau$ in
the expressions of $R_{k}$ in $\left( \text{\ref{k Deform RHS}}\right) $
must vanish\ giving successive values of $\omega_{k}$ and $a_{k}$.
The linear equation $\mathcal{L}\left( \phi\left( \tau,p\right) \right)
=0$ represents a simple harmonic motion with frequency $1$. So, we choose the
initial guess of $\omega$ as $\omega_{0}=1$. Again, by perturbation method
\cite{Jordan Smith} we find $a\rightarrow2$ as $\varepsilon\rightarrow0$. So,
we choose the initial guess of $a$ as $a_{0}=2$. Solving the differential
equations given by $\left( \text{\ref{Zero Deform Eq}}\right) $, $\left(
\text{\ref{Zero Deform IC}}\right) $, $\left( \text{\ref{k Deform Eq}%
}\right) $, $\left( \text{\ref{k Deform IC}}\right) $ and avoiding the
generation of secular terms in each iteration we obtain%
\[
u_{1}\left( \tau\right) =-\frac{1}{24}h\varepsilon\sin3\tau+\frac{1}%
{8}h\varepsilon\sin\tau\text{, }\omega_{1}=-\frac{1}{16}h\varepsilon
^{2}\text{, }a_{1}=\frac{1}{8}h\varepsilon^{2}%
\]%
\begin{align*}
u_{2}\left( \tau\right) & =\left( \frac{1}{384}h^{2}\varepsilon^{3}%
-\frac{1}{24}h^{2}\varepsilon-\frac{1}{24}h\varepsilon\right) \sin3\tau
-\frac{1}{64}h^{2}\varepsilon^{2}\cos3\tau\\
& +\frac{1}{64}h^{2}\varepsilon^{2}\cos\tau+\left( \frac{1}{8}%
h^{2}\varepsilon-\frac{1}{128}h^{2}\varepsilon^{3}+\frac{1}{8}h\varepsilon
\right) \sin\tau
\end{align*}
so that%
\begin{align*}
R_{1} & =\frac{1}{3}\varepsilon\sin3\tau\\
R_{2} & =\frac{1}{24}\left[ 3h\varepsilon^{2}\cos3\tau+\left(
8h\varepsilon-\frac{1}{2}h\varepsilon^{3}\right) \sin3\tau\right]
\end{align*}
Computing $R_{k}$ successively, we can find the successive expressions of
$u_{k}\left( \tau\right) $, $\omega_{k}$ and $a_{k}$. The first order
approximation to the amplitude in $\left( \text{\ref{Amp Series}}\right) $
is%
\begin{equation}
a\approx a_{0}+a_{1}=2+\frac{1}{8}h\varepsilon^{2}=a_{E}\left( \varepsilon
\right) \text{ }\left( \text{say}\right) \text{.}
\label{Amp Approx 1 Order}%
\end{equation}
\newline\begin{figure}[ptb]
\begin{center}
\includegraphics[width=3in]{Fig_Exact_Amp_RL.pdf}
\end{center}
\caption{The exact amplitude of Rayleigh Equation $($by solid line$)$ and its
approximation $a_{E}\left( \varepsilon\right) $ given by $\left(
\text{\ref{HAM Amp New}}\right) $ $($by bold points$)$ for $0<\varepsilon
\leq50$.}%
\label{Fig Exact Amp RL}%
\end{figure}
The above first order expression for the amplitude involves as yet arbitrary
control parameter $h$. Lopez et al \cite{Lopez VDP Amp} proposed specific
$\varepsilon$-dependent expressions for $h$ to obtain an efficient formula for
the VdP limit cycle amplitude. They made the proposal that $h$, besides being
continuous, must also vanish in the limits of $\varepsilon\rightarrow0$ and
$\varepsilon\rightarrow\infty$ to reproduce the zeroth order perturbative
solutions. In our application of HAM for the Rayleigh limit cycle amplitude,
we have chosen a different set of base functions and so can weaken the
condition considerably, both on the continuity and the asymptotic limit
$\varepsilon\rightarrow\infty$. From careful inspections of the graph of the
exact amplitude (Fig.1), it turns out that an appropriate ansatz for the
control parameter $h$ is given by
\begin{equation}
h=\frac{1}{0.5+\varepsilon~b\left( \varepsilon\right) }
\label{Control Parameter}%
\end{equation}
where, $b\left( \varepsilon\right) $ is taken as the step function in the
domain $0<\varepsilon\leq50$ as follows%
\[%
\begin{tabular}
[c]{rccccc}%
$\varepsilon:$ & $0<\varepsilon\leq4$ & $4<\varepsilon\leq5$ & $5<\varepsilon
\leq7$ & $7<\varepsilon\leq8$ & $8<\varepsilon\leq9$\\
$b\left( \varepsilon\right) :$ & $0.162$ & $0.165$ & $0.168$ & $0.171$ &
$0.174$\\
$\varepsilon:$ & $9<\varepsilon\leq11$ & $11<\varepsilon\leq15$ &
$15<\varepsilon\leq20$ & $20<\varepsilon\leq30$ & $30<\varepsilon\leq50$\\
$b\left( \varepsilon\right) :$ & $0.176$ & $0.179$ & $0.181$ & $0.183$ &
$0.185$%
\end{tabular}
\
\]
With this particular form of $h$, we are able to find an analytic
approximation $a_{E}\left( \varepsilon\right) $ to the numerically computed
exact value $a=a\left( \varepsilon\right) $ in the domain $0<\varepsilon
\leq50$ with maximum relative percentage error $\left\vert \dfrac{a_{E}\left(
\varepsilon\right) -a\left( \varepsilon\right) }{a\left( \varepsilon
\right) }\times100\right\vert $ less than $1\%$. Obviously, better accuracy
fit can be obtained by considering finer subdivisions in the definition of
$b(\varepsilon)$. We remark that a piece-wise continuous $\varepsilon$
dependence of $h$ as above is admissible in the framework of HAM.
Since the exact graph of $a(\varepsilon)$ is almost a straight line for
sufficiently large $\varepsilon$ $\left( 7<\varepsilon\leq50\right) $, we
can reduce the number of steps to 4 only. Let us choose%
\begin{equation}
h=\frac{8m}{\varepsilon}-\frac{56m}{\varepsilon^{2}}+\frac{8c}{\varepsilon
^{2}}-\frac{16}{\varepsilon^{2}},\qquad7<\varepsilon\leq50
\label{Control Parameter New}%
\end{equation}
so that $\left( \text{\ref{Amp Approx 1 Order}}\right) $ becomes%
\begin{equation}
a_{E}\left( \varepsilon\right) =\left\{
\begin{array}
[c]{lc}%
2+\frac{1}{8}\left( \frac{1}{0.5+0.162~\varepsilon}\right) \varepsilon^{2} &
0<\varepsilon\leq4\\
2+\frac{1}{8}\left( \frac{1}{0.5+0.165~\varepsilon}\right) \varepsilon^{2} &
4<\varepsilon\leq5\\
2+\frac{1}{8}\left( \frac{1}{0.5+0.168~\varepsilon}\right) \varepsilon^{2} &
5<\varepsilon\leq7\\
m\left( \varepsilon-7\right) +c\text{,} & 7<\varepsilon\leq50
\end{array}
\right. \label{HAM Amp New}%
\end{equation}
where $m$ and $c$ are computed from the exact solution as%
\[
m=\dfrac{a\left( 50\right) -a\left( 7\right) }{50-7}=0.657692\text{ and
}c=a\left( 7\right) =5.63108
\]
keeping the maximum relative percentage error $\left\vert \dfrac{a_{E}\left(
\varepsilon\right) -a\left( \varepsilon\right) }{a\left( \varepsilon
\right) }\times100\right\vert $ less than $1\%$. The plot of $a_{E}\left(
\varepsilon\right) $ given by $\left( \text{\ref{HAM Amp New}}\right) $ is
shown by bold points in Figure \ref{Fig Exact Amp RL} (explicit
discontinuities of $h$ at $\varepsilon=4,5$ and $7$ are not visible at the
resolution of the plotted figure) . \begin{figure}[h]
\begin{center}
\includegraphics[width=3in]{Fig_h_Ham_Rayleigh_epsilon_.pdf}
\end{center}
\caption{The graph of $h\left( \varepsilon\right) $ used for approximation
of the amplitude by HAM given by $\left( \text{\ref{HAM Amp New}}\right) $
for $0<\varepsilon\leq50$.}%
\label{Fig h Ham Rayleigh(epsilon)}%
\end{figure}As remarked above, Lopez et. al. \cite{Lopez VDP Amp} proposed
that a reasonable property for $h$ would be to vanish in the limits as
$\varepsilon\rightarrow0$ and $\varepsilon\rightarrow\infty$. However, from
$\left( \text{\ref{Control Parameter}}\right) $ and $\left(
\text{\ref{Control Parameter New}}\right) $ we observe that a suitable
approximation to the amplitude of Rayleigh equation can be obtained even if
$h$ do not follow this property. The graph of $h\left( \varepsilon\right) $
is given in Figure \ref{Fig h Ham Rayleigh(epsilon)} for $0<\varepsilon\leq50$
(discontinuity in $h$ is not visible at the level of resolution in the figure).
To summarize, one can obtain more accurate approximate formula by suitable
choices of the control parameter $h\left( \varepsilon\right) $ upto any
desired level of accuracy. We also note that a piecewise continuous control
parameter $h$ enables us to obtain good approximation by solving only the
first order deformation equation. However, the first order HAM estimated
amplitude $a(\varepsilon)$ is $O\left( \varepsilon^{2}\right) $.
We report the estimation of the amplitude of the limit cycle for the Rayleigh
and VdP equations by the improved RG method in Subsection
\ref{SubSec Amplitude Estimation}. We do not undertake the computation of the
VdP amplitude by HAM separately, as that was already reported by Lopez et al
\cite{Lopez VDP Amp}.
\section{Computation of Amplitude by RG Method\label{Sec RG Sol}}
The Renormalization Group method (RGM) introduced by Chen, Goldenfeld and Oono
(CGO) \cite{Chen G O RG 1994, Chen G O RG} gives a unified formal approach to
derive asymptotic expansions for the solutions of a large class of nonlinear
ODEs. The RG method is used in solid state physics, quantum field theory and
other areas of physics. One advantage of RGM is that it starts from naive
perturbation expansion of a problem and is expected to yield automatically the
gauge functions such as fractional powers of $\varepsilon$ and logarithmic
terms in $\varepsilon$ in the renormalized expansion. One does not require to
have any prior knowledge to prescribe these unexpected gauge functions in an
ad hoc manner. DeVille et. al. \cite{DeVille RG}\ have introduced an
algorithmic approach for RGM which we adopt for the following application. As
it will transpire the RGM appears to be deficient in estimating amplitude of a
periodic orbit because of the absence of any free control parameter. In a
latter section we have improved this RGM to incorporate a control parameter
similar to HAM and derive efficient estimations of amplitudes of both the
Rayleigh and VdP equations. However, before the introduction of the improved
RG method (IRGM), we first discuss the \textit{conventional} RG method, given
by DeVille et. al. and use it to obtain amplitude and phase equations for the
Rayleigh equation $\left( \text{\ref{Rayleigh}}\right) $. These equations
are already obtained in \cite{Chen G O RG, DeVille RG} to the order $O\left(
\varepsilon^{3}\right) $ which agree with the experimental values as
$\varepsilon\rightarrow0$ only. We have extended these results to the order
$O\left( \varepsilon^{4}\right) $ and notice that higher order perturbative
computations of the RG flow equations would fail to obtain good estimation of
the amplitude of the periodic cycle for all values of $\varepsilon$.
Substituting the naive expansion%
\[
y\left( t\right) =y_{0}\left( t\right) +\varepsilon y_{1}\left( t\right)
+\varepsilon^{2}y_{2}\left( t\right) +\varepsilon^{3}y_{3}\left( t\right)
+\cdots
\]
in $\left( \text{\ref{Rayleigh}}\right) $, we find at each order%
\begin{align*}
O\left( 1\right) & :\ddot{y}_{0}+y_{0}=0\\
O\left( \varepsilon\right) & :\ddot{y}_{1}+y_{1}=\dot{y}_{0}-\frac{1}%
{3}\dot{y}_{0}^{3}\\
O\left( \varepsilon^{2}\right) & :\ddot{y}_{2}+y_{2}=\dot{y}_{1}-\dot
{y}_{0}^{2}\dot{y}_{1}\\
O\left( \varepsilon^{3}\right) & :\ddot{y}_{3}+y_{3}=\dot{y}_{2}-\dot
{y}_{0}^{2}\dot{y}_{2}-\dot{y}_{1}^{2}\dot{y}_{0}%
\end{align*}
The solutions are\newline$y_{0}\left( t\right) =Ae^{i\left( t-t_{0}\right)
}+c.c.\bigskip$\newline$y_{1}\left( t\right) =\frac{1}{24}iA^{3}e^{i\left(
t-t_{0}\right) }+\frac{1}{2}A\left( 1-AA^{\ast}\right) \left(
t-t_{0}\right) e^{i\left( t-t_{0}\right) }-\frac{1}{24}iA^{3}e^{3i\left(
t-t_{0}\right) }+c.c\bigskip$\newline$y_{2}\left( t\right) =\left(
\frac{1}{32}A^{3}-\frac{3}{64}A^{4}A^{\ast}\right) e^{i\left( t-t_{0}%
\right) }\bigskip$\newline$\left. {}\right. \hspace{0.5in}+\left(
-\frac{1}{24}iA^{4}A^{\ast}+\frac{1}{16}iA^{3}(A^{\ast})^{2}+\frac{1}%
{48}iA^{3}+\frac{1}{48}iA^{2}(A^{\ast})^{3}-\frac{1}{8}iA\right) \left(
t-t_{0}\right) e^{i\left( t-t_{0}\right) }\bigskip$\newline$\left.
{}\right. \hspace{0.5in}+\left( \frac{3}{8}A^{3}(A^{\ast})^{2}-\frac{1}%
{2}A^{2}A^{\ast}+\frac{1}{8}A\right) \left( t-t_{0}\right) ^{2}e^{i\left(
t-t_{0}\right) }+\left( \frac{3}{64}A^{4}A^{\ast}-\frac{1}{32}A^{3}+\frac
{1}{192}A^{5}\right) e^{3i\left( t-t_{0}\right) }\bigskip$\newline$\left.
{}\right. \hspace{0.5in}-\frac{1}{16}iA^{3}\left( 1-AA^{\ast}\right)
\left( t-t_{0}\right) e^{3i\left( t-t_{0}\right) }-\frac{1}{192}%
A^{5}e^{5i\left( t-t_{0}\right) }+c.c.\bigskip$\newline$y_{3}\left(
t\right) =\left( -\frac{1}{384}iA^{6}A^{\ast}+\frac{37}{1536}iA^{5}(A^{\ast
})^{2}+\frac{1}{2304}iA^{5}+\frac{1}{512}iA^{4}(A^{\ast})^{3}-\frac{7}%
{256}iA^{4}A^{\ast}-\frac{1}{128}iA^{3}\right) e^{i\left( t-t_{0}\right)
}\bigskip$\newline$\left. {}\right. \hspace{0.5in}+\left(
\begin{array}
[c]{c}%
+\frac{1}{1152}A^{6}A^{\ast}+\frac{5}{128}A^{5}(A^{\ast})^{2}-\frac{119}%
{1152}A^{4}(A^{\ast})^{3}+\frac{11}{384}A^{3}(A^{\ast})^{4}\bigskip\\
-\frac{7}{128}A^{4}A^{\ast}+\frac{1}{48}A^{3}+\frac{11}{64}A^{3}(A^{\ast}%
)^{2}-\frac{1}{64}A^{2}(A^{\ast})^{3}%
\end{array}
\right) \left( t-t_{0}\right) e^{i\left( t-t_{0}\right) }\bigskip
$\newline$\left. {}\right. \hspace{0.5in}+\left(
\begin{array}
[c]{c}%
+\frac{3}{64}iA^{5}(A^{\ast})^{2}-\frac{3}{32}iA^{4}(A^{\ast})^{3}-\frac
{1}{24}iA^{4}A^{\ast}-\frac{1}{32}iA^{3}(A^{\ast})^{4}\bigskip\\
+\frac{3}{32}iA^{3}(A^{\ast})^{2}+\frac{1}{192}iA^{3}+\frac{1}{16}%
iA^{2}A^{\ast}+\frac{1}{48}iA^{2}(A^{\ast})^{3}-\frac{1}{16}iA
\end{array}
\right) \left( t-t_{0}\right) ^{2}e^{i\left( t-t_{0}\right) }\bigskip
$\newline$\left. {}\right. \hspace{0.5in}+\left( -\frac{5}{16}A^{4}%
(A^{\ast})^{3}+\frac{9}{16}A^{3}(A^{\ast})^{2}-\frac{13}{48}A^{2}A^{\ast
}+\frac{1}{48}A\right) \left( t-t_{0}\right) ^{3}e^{i\left( t-t_{0}%
\right) }\bigskip$\newline$\left. {}\right. \hspace{0.5in}+\left(
\begin{array}
[c]{c}%
+\frac{1}{4608}iA^{7}+\frac{7}{512}iA^{6}A^{\ast}-\frac{37}{1536}%
iA^{5}(A^{\ast})^{2}-\frac{1}{128}iA^{5}-\frac{1}{512}iA^{4}(A^{\ast}%
)^{3}\bigskip\\
+\frac{1}{128}iA^{3}+\frac{7}{256}iA^{4}A^{\ast}%
\end{array}
\right) e^{3i\left( t-t_{0}\right) }\bigskip$\newline$\left. {}\right.
\hspace{0.5in}+\left(
\begin{array}
[c]{c}%
-\frac{1}{96}A^{6}A^{\ast}-\frac{7}{64}A^{5}(A^{\ast})^{2}+\frac{1}{128}%
A^{5}+\frac{1}{384}A^{4}(A^{\ast})^{3}\bigskip\\
+\frac{21}{128}A^{4}A^{\ast}-\frac{1}{16}A^{3}%
\end{array}
\right) \left( t-t_{0}\right) e^{3i\left( t-t_{0}\right) }\bigskip
$\newline$\left. {}\right. \hspace{0.5in}+\left( -\frac{5}{64}%
iA^{5}(A^{\ast})^{2}+\frac{1}{8}iA^{4}A^{\ast}-\frac{3}{64}iA^{3}\right)
\left( t-t_{0}\right) ^{2}e^{3i\left( t-t_{0}\right) }\bigskip$%
\newline$\left. {}\right. \hspace{0.5in}+\left( \frac{17}{2304}iA^{5}%
-\frac{17}{1536}iA^{6}A^{\ast}-\frac{5}{4608}iA^{7}\right) e^{5i\left(
t-t_{0}\right) }+\left( \frac{5}{384}A^{6}A^{\ast}-\frac{5}{384}%
A^{5}\right) \left( t-t_{0}\right) e^{5i\left( t-t_{0}\right) }\bigskip
$\newline$\left. {}\right. \hspace{0.5in}+\frac{1}{1152}iA^{7}e^{7i\left(
t-t_{0}\right) }+c.c.$
We choose the homogeneous parts to the solutions $y_{1}$, $y_{2}$ and $y_{3}$
in such a manner that the solutions vanish at the initial time $t_{0}$, i.e.
$y_{1}\left( t_{0}\right) =y_{2}\left( t_{0}\right) =y_{3}\left(
t_{0}\right) =0$. Next, we renormalize the integration constant $A$ and
create a new renormalized quantity $\mathcal{A}$ as%
\[
A=\mathcal{A}+a_{1}\varepsilon+a_{2}\varepsilon^{2}+a_{3}\varepsilon
^{3}+O\left( \varepsilon^{4}\right)
\]
where the coefficients $a_{1},a_{2},a_{3},\ldots$ are chosen to absorb the
homogeneous parts of the solutions $y_{1},y_{2},\ldots$. Choosing%
\begin{align*}
a_{1} & =-\frac{i}{24}\mathcal{A}^{3},~a_{2}=-\frac{\mathcal{A}^{3}}%
{32}\left( 1-\frac{3}{2}\mathcal{AA}^{\ast}+\frac{1}{6}\mathcal{A}%
^{2}\right) ,\\
a_{3} & =\frac{1}{1152}i\mathcal{A}^{7}-\frac{17}{1536}i\mathcal{A}%
^{6}\mathcal{A}^{\ast}+\frac{17}{2304}i\mathcal{A}^{5}-\frac{37}%
{1536}i\mathcal{A}^{5}(\mathcal{A}^{\ast})^{2}+\frac{7}{256}i\mathcal{A}%
^{4}\mathcal{A}^{\ast}+\frac{1}{128}i\mathcal{A}^{3}%
\end{align*}
we obtain\newline$y_{0}\left( t\right) =\mathcal{A}e^{i\left(
t-t_{0}\right) }+c.c.\bigskip\newline y_{1}\left( t\right) =\left(
\frac{1}{2}\mathcal{A}\left( 1-\mathcal{AA}^{\ast}\right) \left(
t-t_{0}\right) e^{i\left( t-t_{0}\right) }-\frac{1}{24}i\mathcal{A}%
^{3}e^{3i\left( t-t_{0}\right) }\right) +c.c.\bigskip\newline y_{2}\left(
t\right) =\left( \frac{1}{16}i\mathcal{A}^{3}(\mathcal{A}^{\ast})^{2}%
-\frac{1}{8}i\mathcal{A}\right) \left( t-t_{0}\right) e^{i\left(
t-t_{0}\right) }+\frac{1}{8}\mathcal{A}\left( \mathcal{AA}^{\ast}-1\right)
\left( 3\mathcal{AA}^{\ast}-1\right) ~\left( t-t_{0}\right) ^{2}%
~e^{i\left( t-t_{0}\right) }\bigskip\newline\left. {}\right.
\hspace{0.5in}+\left( \frac{3}{64}\mathcal{A}^{4}\mathcal{A}^{\ast}-\frac
{1}{32}\mathcal{A}^{3}\right) e^{3i\left( t-t_{0}\right) }+\frac{1}%
{16}i\mathcal{A}^{3}\left( \mathcal{AA}^{\ast}-1\right) ~\left(
t-t_{0}\right) ~e^{3i\left( t-t_{0}\right) }\bigskip\newline\left.
{}\right. \hspace{0.5in}-\frac{1}{192}\mathcal{A}^{5}e^{5i\left(
t-t_{0}\right) }+c.c.\bigskip$\newline$y_{3}\left( t\right) =\left(
-\frac{13}{128}\mathcal{A}^{4}(\mathcal{A}^{\ast})^{3}+\frac{11}%
{64}\mathcal{A}^{3}(\mathcal{A}^{\ast})^{2}\right) \left( t-t_{0}\right)
e^{i\left( t-t_{0}\right) }\bigskip\newline\left. {}\right. \hspace
{0.5in}+\left( -\frac{3}{32}i\mathcal{A}^{4}(\mathcal{A}^{\ast})^{3}+\frac
{3}{32}i\mathcal{A}^{3}(\mathcal{A}^{\ast})^{2}+\frac{1}{16}i\mathcal{A}%
^{2}\mathcal{A}^{\ast}-\frac{1}{16}i\mathcal{A}\right) \left( t-t_{0}%
\right) ^{2}e^{i\left( t-t_{0}\right) }\bigskip\newline\left. {}\right.
\hspace{0.5in}+\left( -\frac{5}{16}\mathcal{A}^{4}(\mathcal{A}^{\ast}%
)^{3}+\frac{9}{16}\mathcal{A}^{3}(\mathcal{A}^{\ast})^{2}-\frac{13}%
{48}\mathcal{A}^{2}\mathcal{A}^{\ast}+\frac{1}{48}\mathcal{A}\right) \left(
t-t_{0}\right) ^{3}e^{i\left( t-t_{0}\right) }\bigskip\newline\left.
{}\right. \hspace{0.5in}+\left( -\frac{37}{1536}i\mathcal{A}^{5}%
(\mathcal{A}^{\ast})^{2}+\frac{1}{128}i\mathcal{A}^{3}+\frac{7}{256}%
i\mathcal{A}^{4}\mathcal{A}^{\ast}\right) e^{3i\left( t-t_{0}\right)
}\bigskip\newline\left. {}\right. \hspace{0.5in}+\left( -\frac{7}%
{64}\mathcal{A}^{5}(\mathcal{A}^{\ast})^{2}+\frac{21}{128}\mathcal{A}%
^{4}\mathcal{A}^{\ast}-\frac{1}{16}\mathcal{A}^{3}\right) \left(
t-t_{0}\right) e^{3i\left( t-t_{0}\right) }\bigskip\newline\left.
{}\right. \hspace{0.5in}+\left( -\frac{5}{64}i\mathcal{A}^{5}(\mathcal{A}%
^{\ast})^{2}+\frac{1}{8}i\mathcal{A}^{4}\mathcal{A}^{\ast}-\frac{3}%
{64}i\mathcal{A}^{3}\right) \left( t-t_{0}\right) ^{2}e^{3i\left(
t-t_{0}\right) }\bigskip\newline\left. {}\right. \hspace{0.5in}+\left(
\frac{17}{2304}i\mathcal{A}^{5}-\frac{17}{1536}i\mathcal{A}^{6}\mathcal{A}%
^{\ast}\right) e^{5i\left( t-t_{0}\right) }+\left( \frac{5}{384}%
\mathcal{A}^{6}\mathcal{A}^{\ast}-\frac{5}{384}\mathcal{A}^{5}\right) \left(
t-t_{0}\right) e^{5i\left( t-t_{0}\right) }\bigskip\newline\left.
{}\right. \hspace{0.5in}+\frac{1}{1152}i\mathcal{A}^{7}e^{7i\left(
t-t_{0}\right) }+c.c.$
\begin{remark}
DeVille \cite{DeVille RG} have obtained same result correct upto $O\left(
\varepsilon^{3}\right) $. However, their computed expression of $a_{2}$ is
not correct. We have made the correction in the expression of $a_{2}$.
\end{remark}
We observe that each of $y_{1}\left( t\right) $, $y_{2}\left( t\right) $,
$y_{3}\left( t\right) $ contains secular terms. As a consequence the
solution%
\[
y\left( t\right) =y_{0}\left( t\right) +y_{1}\left( t\right)
\varepsilon+y_{2}\left( t\right) \varepsilon^{2}+y_{3}\left( t\right)
\varepsilon^{3}+O\left( \varepsilon^{4}\right)
\]
becomes divergent as $t\rightarrow\infty$. To regularize the perturbation
series using RGM an arbitrary time $\tau$ is introduced and $t-t_{0}$ is split
as $\left( t-\tau\right) +\left( \tau-t_{0}\right) $. The terms containing
$\tau-t_{0}$ is observed in the renormalized counterpart $\mathcal{A}$ of the
constant of integration $A$. Since the final solution should not depend upon
the choice of the arbitrary time $\tau$, so%
\begin{equation}
\left. \frac{\partial y}{\partial\tau}\right\vert _{\tau=t}=0
\label{CGO RG Cond}%
\end{equation}
for any $t$. However, DeVille et. al. \cite{DeVille RG} have simplified this
condition and proposed an equivalent condition as%
\begin{equation}
\left. \frac{\partial y}{\partial t_{0}}\right\vert _{t_{0}=t}=0
\label{DeVille RG Cond}%
\end{equation}
We note that renormalized counterpart $\mathcal{A}$ is no longer a constant of
motion in RGM. The RG condition $\left( \text{\ref{DeVille RG Cond}}\right)
$ is developed in such a manner that one need to differentiate the terms
containing $e^{i\left( t-t_{0}\right) }$, $e^{-i\left( t-t_{0}\right) }$,
$\left( t-t_{0}\right) e^{i\left( t-t_{0}\right) }$ and $\left(
t-t_{0}\right) e^{-i\left( t-t_{0}\right) }$ and thereafter substituting
$t_{0}=t$ the resultant expression is equated to zero. The other terms related
to higher harmonics are not involved in RG condition. Simplifying RG condition
$\left( \text{\ref{DeVille RG Cond}}\right) $ we get%
\[
\frac{\partial\mathcal{A}}{\partial t_{0}}=\mathcal{A}i-\frac{1}{2}%
\mathcal{A}\left( \mathcal{AA}^{\ast}-1\right) \varepsilon-\frac{1}%
{8}i\mathcal{A}\left( 1-\frac{1}{2}\mathcal{A}^{2}(\mathcal{A}^{\ast}%
)^{2}\right) \varepsilon^{2}-\frac{1}{64}\mathcal{A}^{3}(\mathcal{A}^{\ast
})^{2}\left( \frac{13}{2}\mathcal{AA}^{\ast}-11\right) \varepsilon^{3}
\]
to the order $O\left( \varepsilon^{4}\right) $. Taking $\mathcal{A=}\frac
{R}{2}e^{i(t+\theta)}$ we obtain corresponding amplitude and phase flow
equations to the order $O\left( \varepsilon^{4}\right) $ as%
\begin{align}
\frac{dR}{dt} & =\frac{1}{2}R\left( 1-\frac{R^{2}}{4}\right)
\varepsilon+\frac{1}{1024}R^{5}\left( 11-\frac{13}{8}R^{2}\right)
\varepsilon^{3}+O\left( \varepsilon^{4}\right) \label{Amp Eq Order 3}\\
\frac{d\theta}{dt} & =-\frac{1}{8}\left( 1-\frac{R^{4}}{32}\right)
\varepsilon^{2}+O\left( \varepsilon^{4}\right) \label{Phase Eq Order 3}%
\end{align}
To the authors' knowledge these higher order flow equations are reported for
the first time in the literature. We remark that above flow equations match
exactly with $O(\varepsilon^{3})$ flow equations of the Van der Pol equation
\cite{Sarkar Bhattacharjee}. Although not done explicitly, we expect that the
$O\left( \varepsilon^{4}\right) $ VdP flow equations would also have the
equivalent forms. For latter reference, we also write down the order
$O(\varepsilon^{2})$ solution of the Rayleigh equation \cite{Chen G O RG}
\begin{equation}
\label{Sol1}y(t)= R(t)\cos(t+\theta) + \frac{\varepsilon}{96}R(t)^{3}%
(\sin3(t+\theta)-\sin(t+\theta))
\end{equation}
Solving the amplitude equation $\left( \text{\ref{Amp Eq Order 3}}\right) $
by numerical method and taking the limit as $t\rightarrow\infty$ so that for a
fixed value of $\varepsilon$ we have $R\rightarrow a_{RG}\left(
\varepsilon\right) $, the approximation of the amplitude of limit cycle of
Rayleigh equation $\left( \text{\ref{Rayleigh}}\right) $ by RGM, we obtain
Figure \ref{Fig RG O(3) Compare} representing $\varepsilon$ dependence of the
amplitude $a_{RG}$ by solid lines. \begin{figure}[h]
\begin{center}%
\begin{tabular}
[c]{cc}%
\includegraphics[width=3in]{Fig_RG_O_3_Compare_Small.pdf} &
\includegraphics[
width=3in]{Fig_RG_O_3_Compare_Total.pdf}\\
$\left( a\right) $ & $\left( b\right) $%
\end{tabular}
\end{center}
\caption{Graph of $a_{RG}\left( \varepsilon\right) $ $($by solid lines$)$
correct upto $O\left( \varepsilon^{4}\right) $ and compared with exact graph
of $a\left( \varepsilon\right) $ $($by dotted lines$)$ for $0<\varepsilon
\leq5$ in $\left( a\right) $ and for $0<\varepsilon\leq20$ in $\left(
b\right) $.}%
\label{Fig RG O(3) Compare}%
\end{figure}
Thus we observe that the RG flow equation to the order $O\left(
\varepsilon^{4}\right) $ for the amplitude does not give good approximation
to the exact solution for moderate and large values of $\varepsilon$.
\section{Improved RG Method: Nonlinear Time\label{Sec M-RG Sol Description}}
In RGM an arbitrary time $\tau$ is introduced in between current time $t$ and
the initial time $t_{0}$ so that $t-t_{0}=\left( t-\tau\right) +\left(
\tau-t_{0}\right) $ in order to remove the divergent terms in the naive
perturbation expansion for the solution of the given differential equation.
The solution is renormalized by suitable choice of the constants of
integration to remove the terms containing $\left( \tau-t_{0}\right) $ and
keeping the terms having $\left( t-\tau\right) $. Since the solution should
be independent of the arbitrary time $\tau$, the RG condition%
\[
\left. \frac{\partial y}{\partial\tau}\right\vert _{\tau=t}=0
\]
is applied to the renormalized solution. However, in the previous section we
have seen that the method fails to produce good approximations to the exact
solution for $\varepsilon\sim O(1)$. Our target is not only to remove the
divergent terms in the solution but also to introduce some control parameter
$h(\varepsilon)$ which can control the RG solution in such a manner that this
solution ultimately converges to the exact solution. Moreover, our another
goal is to achieve this accuracy by merely solving the differential equation
to a minimal order of the expansion parameter, viz., upto $O\left(
\varepsilon^{2}\right) $ or less.
Since the basic idea is to split the time difference $t-t_{0}$ by introduction
of an arbitrary time, so we can write $t-t_{0}=\left( t-\dfrac{\tau
}{\varepsilon}\right) +\left( \dfrac{\tau}{\varepsilon}-t_{0}\right) $.
From now on let us assume that $0<<\varepsilon<\approx1$. The case
$\varepsilon>\approx1$ will be commented upon later. The constants of
integration can be renormalized in order to remove the terms containing
$\left( \dfrac{\tau}{\varepsilon}-t_{0}\right) $ from the solution keeping
the terms containing $\left( t-\dfrac{\tau}{\varepsilon}\right) $. Finally
analogous to the classical RG method we put $t=\dfrac{\tau}{\varepsilon}$,
i.e. $\tau=\varepsilon t$, in%
\begin{equation}
\frac{\partial y}{\partial\tau}=0 \label{Pre RG eq}%
\end{equation}
giving rise to an improved form of the RG flow equation to remove secular
terms involving $\left( t-\dfrac{\tau}{\varepsilon}\right) $. So far the
improved method does not produce any qualitative new result compared to the
RGM and so we must get the same phase and amplitude equation as deduced in
Section \ref{Sec RG Sol}.
We next proceed one step further. As stated already in the Introduction, we
now exploit the possibility of extending the original linear $t$ dependence of
$\tau$ viz., $\tau=t$ of RGM in removing the explicit divergences by a
\emph{nonlinear dependence} $\tau=\varepsilon t$ along with the
\emph{additional condition} that $\tau\rightarrow\varepsilon^{-n}\phi
(\tilde\tau)$, where $\phi$ a slowly varying scaling function of the O(1)
rescaled variable $\tilde\tau=\varepsilon\tilde t\sim O(1)$, as the original
linear time $t\rightarrow\infty$ following the\emph{ scales} $t\sim
\varepsilon^{-n}\tilde t,\ n=1,2,\ldots$. (Note that linear time flows with
uniform rate 1 and $\tau$ is nonlinear since rate $\dot\phi(\tilde\tau)<1$).
It follows that for a given nonlinear differential system, such a nonlinear
time dependence always exists and nontrivial, provided one invokes a
\emph{duality principle} transferring nonlinear influences from the far
asymptotic region into the finite observable sector in a {cooperative }manner
\cite{DPD, DPD Sen, DPD New}. In Appendix, we give a brief overview of the
\emph{novel} analytic framework extending the standard classical analysis to
one that supports naturally the above stated \emph{duality structure} and the
emergent nonlinear scaling patterns typical for a given nonlinear system.
In fact, as the linear time $t\rightarrow\infty$ following the above hierarchy
of scales, there exists $\tilde{t}_{n}$ such that $1<<\left( \varepsilon
t\right) ^{n}<\varepsilon^{-n}<\tilde{t}_{n}$ and satisfying \emph{the
inversion law} $\tilde{t}_{n}/\varepsilon^{-n}\propto\varepsilon^{-n}/\left(
\varepsilon t\right) ^{n}$. This inversion law makes a room for transfer of
effective influences, typical for the nonlinear system concerned, from
nonobservable sector $t>\varepsilon^{-n}$ to the observable sector
$t<\varepsilon^{-n}$ bypassing the dynamically generated singular points
denoted by the scales $\varepsilon^{-n}$. Notice the nonlinear connection
between scales of the form $\varepsilon^{n} \tilde t_{n}$ with the scale
$\varepsilon t$ via duality structure (c.f. Appendix). Let $\tilde{t}\left(
t\right) =\underset{n\rightarrow\infty}{\lim}\left( \tilde{t}_{n}\right)
^{1/n}$ so that $\varepsilon t<\varepsilon^{-1}<\tilde{t}\left( t\right) $
and $\tilde{t}/\varepsilon^{-1}\propto\varepsilon^{-1}/\left( \varepsilon
t\right) $. Define
\begin{equation}
\label{visible}h_{0}\left( \tilde\tau\right) =\underset{n\rightarrow\infty
}{\lim}\log_{\varepsilon^{-n}}\tilde{t}_{n}/\varepsilon^{-n}.
\end{equation}
Here, the scaling exponent $h_{0}$ corresponds to the visibility norm
(Appendix), that can access (encode) the non-perturbative region (information)
of the nonlinear system and $\tilde\tau$ is an O(1) rescaled variable. The
exponent $h_{0}$ is \emph{scaling invariant} in the sense that it appears
uniformly for every $n$ as $t\rightarrow\infty$ through the scales $\tilde
t_{n}=\varepsilon^{-n} \varepsilon^{-n h_{0}(\tilde\tau)}$. As a consequence,
a significant amount of asymptotic scaling information in the limit
$t\rightarrow\infty$ could be simply retrieved by considering the scaling
limit instead at $t=\varepsilon^{-1}$.
Exploiting the above insight, one now writes the nonperturbative scaling limit
in the form
\begin{equation}
\tau=\lim\varepsilon t=\varepsilon^{-h_{RG}\left( \tilde\tau\right)
}>1,\ \varepsilon<1 \label{scaling1}%
\end{equation}
as $t\rightarrow\varepsilon^{-1}$. Moreover, $h_{RG}\left( \tilde\tau\right)
=1-h_{0}\left( \tilde\tau\right) $. As noted already, the scaling exponent
$h_{0}\left( \tilde\tau\right) $ here encodes the \emph{effective
}cooperative influence of far asymptotic sector $t>\varepsilon^{-n}$ into the
observable sector $1<t<\varepsilon^{-n}$ by the inversion mediated duality
principle. As pointed out in Appendix, the duality principle \emph{does} allow
asymptotic limiting (non-perturbative) behaviour of the nonlinear system to be
encoded into the scaling exponents of the nonlinear time $\tau$ that, in turn,
offers an efficient handle in uncovering key dynamical information of the said
system. Notice that, in the absence of the said duality the linear time $t$
can in principle attain the scale $\varepsilon^{-1}$ $\left( \text{say}%
\right) $, and as a consequence $h_{0}=0$, retrieving the ordinary scaling of
$\tau=\varepsilon t\sim\varepsilon^{-1}$ as $t\sim\varepsilon^{-2}$. This also
establishes, in retrospect, that the scaling exponent $h_{0}\left(
\varepsilon\right) $ is well defined and can exist nontrivially i.e.
$h_{0}\sim O\left( 1\right) $ in a nonlinear problem (c.f. Appendix). As a
consequence, \emph{the RG control parameter }$\emph{h}_{RG}$\emph{ can be of
both the signs, with relatively small numerical value in fully developed
nonlinear systems }$\varepsilon\gg1$\emph{, but with a possible }$O\left(
1\right) $\emph{ variations for }$\varepsilon\sim O\left( 1\right) $\emph{
or less.}
The above construction actually tells somewhat more. Corresponding to the
first generation scales $\varepsilon^{-n}$, one can, in fact, have the
\emph{second generation } nonlinear scales
\begin{equation}
\tau_{m}=\lim\varepsilon^{m} t=\varepsilon^{-m h^{m}_{RG}\left( \tilde
\tau\right) }>1,\ \varepsilon<1, \ m>1 \label{scaling2}%
\end{equation}
as $t\rightarrow\varepsilon^{-m}$ with $h^{1}_{RG}=h_{RG}$. The
\emph{nonlinear time} $\tau$ now stands for these hierarchy of scales
$\{\tau_{m}\}$. Consequently, as the linear time $t$ approaches $\infty$
through the first generation linear scales, the slowly varying nonlinear time
$\tau$ approaches either to $\infty$ or 0 at slower and slower rates as
represented by the numerically small RG scaling exponents $h^{m}%
_{RG}(\varepsilon)$, each of which remains almost constant over longer and
longer intervals of $\varepsilon^{-1}$ (as $\varepsilon^{-1} \rightarrow
\infty$ ). In the present paper we show how the first two scaling exponents
$h_{i}(\tilde\tau), \ i=1,2$ relate to the nonperturbative properties of the
limit cycle. We expect higher order scaling exponents $h_{m}$ would have vital
role in bifurcation of nonautonomous systems. This problem will be
investigated elsewhere.
Let us remark that for $\varepsilon>1$, we consider instead the first
generation scales as $\varepsilon^{n}$, and the duality is invoked for
variables satisfying $t/\varepsilon<\varepsilon<\tilde{t}(t)$ so that the
asymptotic scaling variables are derived as $\tau_{m}=\varepsilon^{m
h^{m}_{RG}(\tilde\tau)},\ \varepsilon>1$ where $h^{m}_{RG}=1-h^{m}_{0}$.
Moreover, said proliferation of nonlinear scales (\ref{scaling2}) actually
continues ad infinitum. In fact, interpreting each second generation scale
$\tau_{m}$, $m$ fixed, as first generation scale, and iterating above steps
one associates third generation scales $\tau_{m_{k}}, \ k=1,2,\ldots$, and so on.
It now follows, from the above general remarks on the behaviour of $h_{RG}$,
that the nonlinear time $\tau$ actually approaches 0 or $\infty$ as $\tau
\sim(\log\varepsilon)^{-\alpha}$ or $\tau\sim(\log\varepsilon)^{\alpha},
\ \alpha>0$ respectively as $\varepsilon\rightarrow\infty$. However, one must
have $\tau=\varepsilon^{-h_{RG}(\varepsilon)}\rightarrow\infty$ as
$\varepsilon\rightarrow0$. An example of the asymptotic behaviours of $h_{RG}$
is given by $\tau_{m}=\varepsilon^{\pm\alpha_{m}\frac{\log\log\varepsilon
}{\log\varepsilon}}$ for $\varepsilon\rightarrow\infty$, which one expects to
verify explicitly in evaluation of asymptotic quantities, such as amplitude of
a periodic cycle, in a nonlinear system.
In the IRGM, we exploit this duality induced nontrivial scaling information to
rewrite the lowest order perturbative flow equations (\ref{Amp Eq Order 3})
and (\ref{Phase Eq Order 3}) as the asymptotic RG flow equations in the limit
$t\rightarrow\infty$
\begin{align}
\frac{da}{d\tau_{1}} & =\frac{1}{2}a\left( 1-\frac{a^{2}}{4}\right)
\label{Amp}\\
\frac{d\psi}{d\tau_{2}} & =-\frac{1}{8}\left( 1-\frac{a^{4}}{32}\right)
\label{Phase}%
\end{align}
for the amplitude $a=a(\tilde{\tau})$ and the phase $\psi=\psi(\tilde{\tau})$
of the limit cycle of both the Rayleigh and Van der Pol equations, involving
slowly varying nonlinear time scales $\tau_{i},\ i=1,2$. The asymptotic
scaling functions $\tau_{1}=\phi_{1}(\varepsilon t)=\varepsilon^{h_{RG}^{1}}$
and $\tau_{2}=\phi(\varepsilon^{2}t)=\varepsilon^{2h_{RG}^{2}}$ are activated
invoking nonlinear limits as in (\ref{scaling2}) as $t\rightarrow\varepsilon
^{-1}$ and $t\rightarrow\varepsilon^{-2}$ successively in the above equations.
The slowly varying almost constant scaling functions $\phi_{1}$ and $\phi_{2}%
$, satisfying $|\ddot{\phi}_{i}|<<|\dot{\phi}_{i}^{2}|<<1$, are assumed to
have a rhythmic pattern over the cycle: when $\phi_{1}$ varies slowly,
$\phi_{2}$ remains almost constant i.e. $\dot{\phi}_{1}>0,\ \dot{\phi}%
_{2}\approx0$ and vice versa successively on the cycle (c.f. Appendix Sec. B).
Nontrivial ultrametric neighbourhood structure induced asymptotically by
duality principle (c.f. Appendix Sec. A) can indeed support such \emph{locally
constant} nonlinear rhythmic behaviour. The above flow equations may therefore
be considered \emph{exact} and encode non-perturbative information of
the limit cycle variables $a$ and $\psi$ respectively. The conventional
perturbative RG flow equations in the linear time $t$ is now extended into the
non-perturbative flow equations in the nontrivial scaling variable $\tau
_{i}=\varepsilon^{ih_{RG}^{i}(\tilde{\tau})},\ i=1,2$ involving the
nonlinearity parameter $\varepsilon>1$. The perturbative fixed point for the
amplitude equation at $a=2$ for $t\rightarrow\infty$ corresponding to the
periodic oscillation with $\varepsilon<<1$ is superimposed by \emph{small scale
periodic} flow of amplitude $a(\tau_{1})$ over the entire cycle. The
associated phase $\psi(\tau_{2})$ then flow at a slower rate \emph{linearly
with the higher order scale $\tau_{2}$} when $a(\tau_{1})$ remains almost
constant over a relatively small period of time.
The RG estimated approximate formulae for the amplitude $a(\varepsilon)$ for
the Rayleigh and Van der Pol limit cycles are obtained from the equation
(\ref{Amp}) in the Sec.4.1, when appropriate boundary condition, derived
either from exact computation or from perturbative analysis, is used for a
suitable finite value of $\varepsilon$. In the next subsection 4.2, we present
the efficient graphs of the Rayleigh and VdP limit cycle parametrized by the
nonlinear scales $\tau_{i}=\phi_{i}(\tilde\tau), \ \tilde\tau\sim O(1)$ for
fixed values of the nonlinearity parameter $\varepsilon$.
As it turns out, entire onus in the improved RG analysis essentially rests in
proper estimation/identification of the scaling functions $h^{i}_{RG}$ (i.e.
$\phi_{i}(\tilde\tau)$) which should yield correct dynamical properties of a
nonlinear system. We hope to undertake more detailed and systematic analysis
for determining $h^{i}_{RG}$ elsewhere. In this work we limit ourselves only
to show that IRGM can indeed yield correct amplitude and solution for the
Rayleigh and VdP systems provided one makes appropriate choice of $h^i_{RG}$ based on
clues from exact computations and previously known approximate results (for
instance the perturbative RGM). We remark finally that the perturbative RG method is known
to extend the conventional multiple scale method \cite{Chen G O RG}. Nonlinear time formalism
introduces new set of nonlinear scales $\phi_n(\tilde \tau)$ associated
with ordinary scales $\varepsilon^n$.
We study here the nontrivial applications of such nonlinear scaling functions.
\subsection{Approximate Formula for
Amplitude\label{SubSec Amplitude Estimation}}
We shall now use the above asymptotic amplitude flow equation (\ref{Amp}) to
find analytic approximations of the amplitudes of the limit cycle for both the
Rayleigh and Van der Pol equations.
By a direct integration, one obtains from (\ref{Amp})%
\begin{equation}
\ln\left( a^{2}-4\right) -2\ln a=-\varepsilon^{h_{RG}}%
-0.87953\label{Amp Sol RG RL}%
\end{equation}
as the Rayleigh limit cycle amplitude where we use the boundary condition the
value $a=2.17271$ for $\varepsilon=1$ (this choice simplifies calculation). It
follows immediately that for suitable choices of the control parameter
$h_{RG}$ one can achieve efficient matching for the estimated amplitude
$a_{{E}}(\varepsilon)$. For example, using the HAM generated approximate
formula (\ref{HAM Amp New}) for $a_{{E}}(\varepsilon)$, we can determine the
control parameter $h_{RG}(\varepsilon)$ by the formula
\begin{equation}
h_{RG}=\frac{1}{\ln\varepsilon}\ln\left\{ \left\vert \ln\left( \frac{a^{2}%
}{a^{2}-4}\right) -0.87953\right\vert \right\} \label{Cont Parameter RG}%
\end{equation}
In Figure \ref{Fig h IRG Rayleigh(epsilon)}, \begin{figure}[h]
\begin{center}%
\begin{tabular}
[c]{cc}%
\includegraphics[width=3in]{Fig_h_IRG_Rayleigh_epsilon_Small.pdf} &
\includegraphics[
width=3in]{Fig_h_IRG_Rayleigh_epsilon_Total.pdf}\\
$\left( a\right) $ & $\left( b\right) $%
\end{tabular}
\end{center}
\caption{The graph of $h_{RG}\left( \varepsilon\right) $ used for
approximation of the amplitude of the Rayleigh equation $\left(
\text{\ref{Rayleigh}}\right) $ by HAM given by $\left(
\text{\ref{Cont Parameter RG}}\right) $ for $0<\varepsilon\leq5$ in $\left(
a\right) $ and for $0<\varepsilon\leq50$ in $\left( b\right) $.}%
\label{Fig h IRG Rayleigh(epsilon)}%
\end{figure}we display the typical piece-wise smooth form of $h_{RG}%
(\varepsilon)$ given by $\left( \text{\ref{Cont Parameter RG}}\right) $ for
the Rayleigh limit cycle amplitude that would reproduce the HAM generated
amplitude with relative error less that $1\%$. Clearly, the graph reveals
variability of $h_{RG}$ for moderate values of $\varepsilon$, but the
variability dies out fast for larger values $\varepsilon$, as expected.
We recall that the corresponding graph of the exact computed values of VdP
amplitude $a(\varepsilon)$, on the other hand, has a hump like shape with a
maximum roughly at $\varepsilon\approx2.0235$ and having the asymptotic limits
$2$ as $\varepsilon\rightarrow0$ and $\infty$. Lopez et al \cite{Lopez VDP
Amp} obtained HAM generated approximate formula for the VdP amplitude with
relative error less than $0.05\%$ at the order $O\left( \varepsilon
^{4}\right) $. It is interesting to note that the RG generated formula
$\left( \text{\ref{Amp Sol RG RL}}\right) $ can reproduce the exact computed
values of the VdP amplitude with error less than $0.05\%$ directly from only
the first order RG flow equation. To achieve this goal we first intuitively
guess a piecewise smooth formula for the estimated amplitude
$a_{E}$ by
\begin{equation}
a_{E}\left( \varepsilon\right) =\left\{
\begin{array}
[c]{lc}%
1.998+\dfrac{0.015}{8.121~e^{-2.139~\varepsilon}+0.512~e^{0.043~\varepsilon}}
& 0<\varepsilon<3\\
2.0025+\dfrac{0.031}{0.5~e^{-2.033\left( \varepsilon-2.183\right)
}+1.869~e^{0.087\left( \varepsilon-6.376\right) }} & 3\leq\varepsilon\leq50
\end{array}
\right. \label{VdP Amplitude IRGM}%
\end{equation}
keeping the maximum relative percentage error $\left\vert \dfrac{a_{E}\left(
\varepsilon\right) -a\left( \varepsilon\right) }{a\left( \varepsilon
\right) }\times100\right\vert $ less than $0.05\%$. This shows that the
approximation is quite accurate. The graph of $a_{E}\left( \varepsilon
\right) $ is compared with the exact values in Figure \ref{Fig VdP Amp}. One
may as well use a least square fit of the exact data instead of the above fit.
We do not pursue this approach here. \begin{figure}[h]
\begin{center}
\includegraphics[width=3in]{Fig_VdP_Amp.pdf}
\end{center}
\caption{The exact amplitude of Van der Pol Equation $\left( \text{\ref{VdP}%
}\right) $ $($by solid line$)$ and its approximation $a_{E}\left(
\varepsilon\right) $ given by $\left( \text{\ref{VdP Amplitude IRGM}%
}\right) $ $($by bold points$)$ for $0<\varepsilon\leq50$.}%
\label{Fig VdP Amp}%
\end{figure}
Using this efficient formula for the VdP amplitude, we then obtain the RG flow
equation in the form
\begin{equation}
\ln\left( a^{2}-4\right) -2\ln a=-\varepsilon^{h_{RG}}-4.08785
\label{Amp Sol RG VDP}%
\end{equation}
where we use the boundary condition $a=2.0086$ for $\varepsilon=1$ (for
simplicity of calculation) for the VdP amplitude. Inverting this equation, we
finally obtain the corresponding RG control parameter
\begin{equation}
h_{RG}=\frac{1}{\ln\varepsilon}\ln\left\{ \left\vert \ln\left( \frac
{a_{E}^{2}}{a_{E}^{2}-4}\right) -4.08785\right\vert \right\}
\label{Cont Parameter RG2}%
\end{equation}
Figure \ref{Fig h IRG VDP(epsilon)} \begin{figure}[h]
\begin{center}%
\begin{tabular}
[c]{cc}%
\includegraphics[width=3in]{Fig_h_IRG_VDP_epsilon_Small.pdf} &
\includegraphics[
width=3in]{Fig_h_IRG_VDP_epsilon_Total.pdf}\\
$\left( a\right) $ & $\left( b\right) $%
\end{tabular}
\end{center}
\caption{The graph of $h_{RG}\left( \varepsilon\right) $ used for
approximation $\left( \text{\ref{Cont Parameter RG2}}\right) $ of the
amplitude of the Van der Pol equation $\left( \text{\ref{VdP}}\right) $ for
$0<\varepsilon\leq4$ in $\left( a\right) $ and for $0<\varepsilon\leq50$ in
$\left( b\right) $.}%
\label{Fig h IRG VDP(epsilon)}%
\end{figure}displays the piecewise smooth variation of $h_{RG}$ with
$\varepsilon$. The rapid $O\left( 1\right) $ variation for moderate values
of $\varepsilon$ is evident in Fig.6(a). As expected, $h_{RG}$ dies out fast
for larger values of $\varepsilon$. However, a change in sign is noticed here
already for $\varepsilon>20$ $\left( \text{Figure
\ref{Fig h IRG VDP(epsilon)}}\left( \text{b}\right) \right) $. One expects
many more such small scale sign variations as $\varepsilon\rightarrow\infty$.
This particular form of the control parameter $h_{RG}$, in turn, would
reproduce the VdP amplitude with relative error less than $0.05\%$. As this
level of accuracy is achieved only at the order $O\left( \varepsilon\right)
$, the improved RGM may be considered to be more efficient and advantageous
compared to the HAM.
Alternatively, the amplitude equation $\left( \text{\ref{Amp}}\right) $ can
be inverted as
\begin{equation}
a(\tilde\tau)=\frac{a_{0}}{\sqrt{e^{-\tilde\tau}+\frac{a_{0}^{2}}%
{4}(1-e^{-\tilde\tau})}} \label{amp}%
\end{equation}
where $a_{0}$ is estimated from the exact value of amplitude $a(\varepsilon
_{0})$ for a suitably chosen value of $\varepsilon$, for instance $\varepsilon=1$.
Recall that for the VdP
equation $\tau\sim(\log\varepsilon)^{\alpha}$ and for the Rayleigh equation
$\tau\sim(\log\varepsilon)^{-\alpha}$ for $\varepsilon\rightarrow\infty$ and
$\alpha>0$. By adjusting suitably the values of $\alpha$ over appropriate
intervals on $\varepsilon$ one should be able to obtain efficient matchings
with the exact values of $a(\varepsilon)$.
To summarize, the recipe for deriving approximate formula for limit cycle
amplitude of a nonlinear system can be stated as follows: Determine the first
order (perturbative) RG flow equation for amplitude in the nonlinear time
$\tau$. This will yield an explicit formula for amplitude $a$ as a function of
the nonlinearity parameter $\varepsilon$ and the control parameter $h_{RG}$.
Efficient match with the exact amplitude can be achieved by right choice of
the control parameter $h_{RG}$ or $\alpha$. Alternatively, determine an
efficient formula for $a(\varepsilon)$ by inspection (expert guess) or by
appropriate curve fitting methods. Then determine the control parameter
$h_{RG}$ by an inversion of the estimated amplitude $a_{E}(\varepsilon)$ as in
equation $\left( \text{\ref{Cont Parameter RG2}}\right) $ (and Fig.6). Since
the equations concerned form a closed system, this already gives a (numerical)
proof of the unique existence of $h_{RG}$ for a given nonlinear oscillation.
\subsection{Approximate Limit Cycle}
Here we calculate the approximate limit cycle orbit for the Rayleigh and VdP
equations for a sufficiently large $\varepsilon>1$. Perturbative RGM fails to give
correct relaxation oscillation solution. The first order solution given in Sec. 2 by
HAM is also found insufficient. In this work we do not under take the problem of
computing approximate limit cycle by HAM, which has been addressed by Lopez et al
\cite{Lopez VDP Amp} for the VdP equation. Our aim here is to highlight the strength
of IRGM over perturbative RGM.
For a sufficiently large
time $t\rightarrow\varepsilon^{n},\ n$ large, but fixed, the slowly varying
nonlinear scales $\tau_{1}$ and $\tau_{2}$ are activated in a successive
rhythmic manner, as explained in Sec. 4 (see also Appendix Sec.B), so that the
perturbative solution given in (\ref{Sol1}) is extended to the asymptotic
limit cycle (relaxation oscillation) solution
\begin{equation}
y(\tau_{1},\tau_{2})=a(\tau_{1})\cos(\varepsilon^{n}+\psi(\tau_{2}%
))+Y\label{Sol2}%
\end{equation}
where the amplitude $a$ and phase $\psi$ \emph{flow} along the cycle following
the nonperturbative flow equations (\ref{Amp}) and (\ref{Phase}) in the
asymptotic scaling variables $\tau_{1}$ and $\tau_{2}$ respectively. Here, $Y$
encodes all the renormalized perturbative terms depending on higher order,
slowly varying nonlinear scales $\tau_{i},\ i>2$. The corresponding velocity
component $\dot{y}=\frac{\partial y}{\partial t}$ at $t=\varepsilon^{n}$ has
the form
\begin{equation}
\dot{y}(\tau_{1},\tau_{2})=-a(\tau_{1})\sin(\varepsilon^{n}+\psi(\tau
_{2}))+\sum_{i}\dot{\tau}_{i}\frac{\partial \tilde Y}{\partial\tau_{i}}\label{Sol3}%
\end{equation}
where $\tilde Y$ represents possible slow variations in all the nonlinear scales
$\tau_i, \ i\geq 1$.
Equations (\ref{Sol2}) and (\ref{Sol3}) are the parametric equations of the
limit cycle, parametrized by multiple nonlinear scales, when slowly varying
amplitude $a$ and phase $\psi$ are computed from (\ref{Amp}) and (\ref{Phase})
respectively. An alternative derivation of (\ref{Sol2}) and (\ref{Sol3}) based
purely on duality induced nonlinear scales is given in Appendix Sec. B. We
remark that (\ref{Sol2}) and (\ref{Sol3}) actually represent the general form
of the limit cycle for a much larger class of Lienard system having unique
limit cycle. \emph{Typical geometric shape of the periodic cycle of a given
nonlinear system is controlled entirely by the rhythmic cooperative, almost
constant variations of the nonlinear scales $\tau_{i}$.} As explained in Sec.
4 (See also Appendix Sec. A), scale invariance of the scaling functions
$\tau_{1}=\phi_{1}(\tilde{t}_{1}/\varepsilon)$ and $\tau_{2}=\phi_{1}%
(\tilde{t}_{2}/\varepsilon^{2})$ tells that as the linear time $t\rightarrow
\varepsilon^{n}$, $\tau_{1}\rightarrow\phi_{1}(1)=1$ and $\tau_{2}%
\rightarrow\phi_{2}(1)=1$ with $\tilde{t}_{1}\rightarrow\varepsilon$ and
$\tilde{t}_{2}\rightarrow\varepsilon^{2}$. We now set the initial conditions
$a(1)=a_{\mathrm{amp}}(\varepsilon),\ a^{\prime}(1)=\frac{a(1)}{2}%
(1-\frac{a(1)^{2}}{4})$ and $\psi(1)=0$, where $a_{\mathrm{amp}}(\varepsilon)$
is the exact (experimental) value of the amplitude of the limit cycle. Setting
further $\tau_{1}=1+\eta_{1}$ and $\tau_{2}=1+\eta_{2}$ as $t\rightarrow
\varepsilon^{n}$, both amplitude and phase flow equations (\ref{Amp}) and
(\ref{Phase}) now yield linear flow of amplitude and phase relative to the
respective small scale slow, almost constant variables $\eta_{1}$ and
$\eta_{2}$, satisfying $|\eta_{i}^{2}|<<1$ and $|\ddot{\eta}_{i}|<<|\dot{\eta
}_{i}^{2}|<<1$, those vary in a rhythmic manner ( Appendix Sec.B). The exact
(experimental) limit cycle could now be approximated with any desired accuracy
by smooth matching of elementary straight line segments of the form (i)
$z-Z_{0}=k(y-Y_{0})$ and circular arcs of the form (ii) $(y-Y_{0}%
)^{2}+(z-Z_{0})^{2}=a^{2}(\tau_{1})$ over judiciously chosen intervals in $y$
in the $(y,z)$ plane, where $z=\dot{y}$, $k=-\tan(\varepsilon^{n}+\psi
(\tau_{2}))$, $Y_{0}=Y$ and $Z_{0}=\sum_{i}\dot{\tau}_{i}\frac{\partial
\tilde Y}{\partial\tau_{i}}$. Because of the availability of cooperatively evolving
resource of nonlinear scales, such a matching is always possible
theoretically. In the alternative derivation of limit cycle equations in
Appendix Sec. B, we have outlined an approach to gain more analytic
understanding of the rhythmic, cooperative variations of the nonlinear scaling
functions. We hope to address the question of determining the precise analytic
properties of the scaling functions $\phi_{i}(\tilde{\tau})$ corresponding to
a given nonlinear system in future communications.
\begin{figure}[h]
\begin{center}%
\begin{tabular}
[c]{cc}%
\includegraphics[width=3in]{Fig_RL_LC_Non_Pert.pdf} & \includegraphics[
width=3in]{Fig_VDP_LC_Non_Pert.pdf}\\
$\left( a\right) $ & $\left( b\right) $%
\end{tabular}
\end{center}
\caption{Approximate limit cycle for $\left( a\right) $ Rayleigh equation
and $\left( b\right) $ Van der Pol equation,\newline solid $\left(
\text{red}\right) $ line for approximate curve, dotted $\left(
\text{blue}\right) $ line for exact curve $\left( \varepsilon=5\right) $.}%
\label{Fig_APP_LC}%
\end{figure}
In Fig. \ref{Fig_APP_LC}$\left( a\right) $ and Fig. \ref{Fig_APP_LC}$\left(
b\right) $, we display the $(y,z)$ phase plane relaxation oscillation for the
Rayleigh and VdP equations with $\varepsilon=5$. In Appendix Sec. C, the
piece-wise smooth matching curves approximating these cycles are presented in
tabular forms. However, the smoothness at the joining points are achieved at
the level of one decimal only. More accurate approximation may be achieved
with smarter efforts. Judicious choice of slowly varying centres
$(Y_{0},Z_{0})$ and radii $a(\tau_{1})$ of circular arcs of right sizes
(a straight line segment being an arc with sufficiently large radius) should
give better approximations with a given exact (experimental)
cycle that can be obtained on a
symbolic computation platform, Mathematica for instance. The phase plane
dynamics of these slowly varying centres and radii is expected to reveal
interesting new insights into asymptotic properties of the nonlinear
oscillation . It transpires from Appendix Sec. C that radii, for instance,
vary much faster in Rayleigh than that in VdP oscillator,
in which case radii fluctuate between small and large values through intermediate steps.
This might be compared with fast and slow energy build ups in Rayleigh and VdP relaxation
oscillations respectively.
One would like to interpret this phase plane dynamics as
cooperative evolution of multiple nonlinear scales driving amplitude and phase
of the nonlinear oscillation to flow in such a fashion as to generate little
(elementary) circular arcs and linear segments which join smoothly together to form the
complete orbit. Making an accurate plot then boils down to finding right kind
of such arcs and line segments. Intricate dependence of the trajectory itself
into the definition of the nonlinear scales (Appendix Remark 3) tells in
retrospect that one needs to look for a novel iteration scheme that would
allow one to extract the trajectory systematically as a limit process. This
problem will be considered in detail elsewhere.
\section{Concluding Remarks\label{Sec Conclusion}}
In this paper we have presented a comparative study of the homotopy analysis
method and the Renormalization Group method. The approximate formulae for the
amplitudes of the limit cycles of the Rayleigh and the Van der Pol systems are
derived using both the methods and are compared with the exact results. It
turns out that the higher order perturbative calculations based on the
conventional Renormalization group method would fail to give efficient formula
for the limit cycle amplitudes for these nonlinear oscillators. However, an
improved version of the Renormalization group analysis exploiting a novel
concept of nonlinear time is shown to yield efficient amplitude formulae for
all values of $\varepsilon$. Exploiting multiple nonlinear scales of the
associated nonlinear time the improved RG method is also found to yield good
plots for relaxation oscillation orbits for the Rayleigh and VdP systems.
In Appendix we have presented brief review of the nonlinear time formalism and
also given an alternative approach in deriving non-perturbative flow equations
of amplitude and phase of a limit cycle problem. Non-perturbative information
of asymptotic quantities get naturally encoded into nonlinear scales, that can
be exploited judiciously to extract desired asymptotic properties of a
relevant dynamical quantity. More detailed analysis of the nonlinear time
formalism in several other nonlinear systems will be considered in future.
\section*{Acknowledgements}
Authors thank the anonymous reviewers for constructive criticisms improving the
quality of the paper.
\section*{Appendix}
\subsection*{A. Formal Structure}
The idea of nonlinear time can be given a rigorous meaning in a nonclassical
extension of the ordinary analysis \cite{DPD1,DPD New}. Recall that the real
number system $\mathbf{R}$ is generally constructed as the metric completion
of the rational field $\mathbf{Q}$ under the Euclidean metric $|x-y|,
\ x,y\in\mathbf{Q}$. More specifically, let S be the set of all Cauchy
sequences $\{x_{n}\}$ of rational numbers $x_{n}\in\mathbf{Q}$. Then $S$ is a
ring under standard component-wise addition and multiplication of two rational
sequences. Then the real number field $\mathbf{R}$ is the quotient space $S/
S_{0}$, where the set $S_{0}$ is the set of all Cauchy sequences converging to
$0\in Q$ and is a maximal ideal in the ring $S$. Alternatively, $\mathbf{R}$
can be considered as the set $[S]$ of equivalence classes, when two sequences
in $S$ are said to be equivalent if their difference belongs to $S_{0}$.
The nonclassical extension $\mathbf{R}^{*}$ of $\mathbf{R}$ is based on a
\emph{finer} equivalence relation that is defined in $S_{0}$ as follows: let
$\{a_{n}\}\in S_{0}$. Consider an associated family of Cauchy sequences of the
form $S_{0a}:=\{A^{\pm}|A^{\pm}=\{a_{n}\times a_{n}^{\pm a^{\pm}_{m_{n}}}\}\}$
where $a^{\pm}_{m_{n}}\neq0$ is Cauchy for $m_{n}>N$ and $N$ sufficiently
large. Clearly, $S_{0a}\subset S_{0}$, and sequences of $S_{0a}$ also
converges to 0 in the metric $|.|$. As $a$ parametrizes sequences in $S$, it
follows that $\underset{\{a\}}\bigcup S_{0a}=S$. Assume further that
$a_{m_{n}}^{\pm}$ respect the \emph{duality structure} defined by $(a_{m_{n}%
}^{-})^{-1}\propto a_{m_{n}}^{+}$ for $m_{n}>N$. The duality structure extends
also over the limit elements: viz., $\mathbf{R}\ni\ (a^{-})^{-1}\propto a^{+}$
where $a^{\pm}_{m_{n}}\rightarrow a^{\pm}$ as $m_{n}\rightarrow\infty$ such
that $a^{\pm}$ are close to 1 in $\mathbf{R}$.
Next define an equivalence relation in $S_{0a}$ declaring two sequences
$A_{1},A_{2}$ in the set $S_{0a}$ equivalent if the associated exponentiated
sequences $a_{m_{n}}^{1}$ and $a_{m_{n}}^{2}$ differ by an element of $S_{0}$
for $m_{n}>N$. In particular, one may impose the condition that $A_{1}\equiv
A_{2}$ if and only if $\exists M \ \mathrm{such \ that} \ a_{m_{n}}%
^{1}=a_{m_{n}}^{2} \ \forall\ m_{n}>M$. Clearly, the usual metric $|.|$ fails
to distinguish elements belonging to two distinct such finer equivalent
classes. However, the metric defined as the natural logarithmic extension of
the Euclidean norm, generically called the \emph{asymptotically visibility
metric} is introduced by $h(A_{1},A_{2}) =\underset{n\rightarrow\infty}%
\lim|\log_{|A_{0}|^{-1}} |A_{1}-A_{2}|/|A_0||$ where $A_{0}=\{a_{n}\}\in S_{0}$. The
sequence $A_{0}$ is said to define a natural scale relative to which elements
in $S_{0}$ gets nontrivial values and hence become distinguishable. The limit
exists because of concerned sequences $a_{m_{n}}^{\pm}$ being Cauchy. Note
that the mapping $h:S_{0}\rightarrow\mathbf{R^{+}}$ defined by $h(A)=\underset
{n\rightarrow\infty}\lim|\log_{|A_{0}|^{-1}} |A|/|A_{0}||$ is actually a
nontrivial norm \cite{DPD1,DPD New} (for simplicity of notation, we use same
symbol to denote both the norm and metric).
The extended real number system $\mathbf{R}^{*}$ admitting duality induced
\emph{fine structure} is given, by definition, as the equivalence class under
this finer equivalence relation viz., $\mathbf{R}^{*}:=S/S_{0}$ when
convergence is induced naturally by the asymptotically visibility metric
$h(x,y)$. Clearly, under the usual norm $|.|$, $\mathbf{R}^{*}$ reduces to
$\mathbf{R}$ as the exponentiated elements $a^{\pm}$ are essentially
invisible. The natural application of the visibility norm on $\mathbf{R}^{*}$
is activated in the following steps. For any two distinct elements $x,\ y\in
{\bf R}\subset\mathbf{R}^{*}$, \emph{set}, by definition, $h(x,y)=0,\ x\neq y$;
$h(x,y)$ being nontrivial only for $y\in x+S_{0}$. This choice is natural as
for any element $x\in\mathbf{R}$, the corresponding limiting $h$ norm viz,
$h(x)=h(0,x)=\underset{n\rightarrow\infty}\lim\log_{|A_{0}|^{-1}} |x/A_{0}|$=1 and
so, the definition $h(x,y)=0, \ \forall x,y\in\mathbf{R}$ makes sense.
For nontrivial values of $h(x,y),
\ x,\ y\in\mathbf{R}^{*}$, the definition of the visibility metric extends
over to $h(x,y)=\underset{\varepsilon\rightarrow0}\lim|\log_{\varepsilon^{-1}}
|\frac{x-y}{\varepsilon}||$, which exists by construction.
Next, consider the metric $d: \mathbf{R}^{*}\rightarrow {\bf R}^{+}$ by
$d(x,y)=|x-y|+h(x,y)$. Clearly, ${d}(x,y)=|x-y|$ for any $x,\ y\in {\bf R}$ and
${d}(x,y)=h(x,y)$ for $x,\ y\in\mathbf{R}^{*}-{\bf R} $ and hence $(\mathbf{R}%
^{*},{d})$ is a complete metric space. The metric $h(x,y)$ acting nontrivially
on $S_{0}$ is essentially an ultrametric: $h(x,y)\leq\max\{h(x,z),h(z,y)\}$.
This follows immediately from the observation that $h$ maps $\mathbf{R}$ to
the singleton set $\{1\}$. Further, the ultrametric $h$ must be discretely
valued \cite{DPD1} and hence the nontrivial value set of $h$ viz., $h(S_{0})$
is countable. As a consequence, the set $S_{0}$ is totally disconnected and
perfect in the induced topology.
More detailed analytic aspects (including the idea of smooth jump
differentiability and jump derivative) of the extended system $\mathbf{R}^{*}$
equipped with the metric $d$ will be reported elsewhere \cite{DPD New}. Here,
we make a few relevant remarks.
1. Even as the size of a $\delta-$ neighbourhood of a point $x\in\mathbf{R}$
vanishes linearly, the same for $x^{*}\in\mathbf{R^{*}}$ need not vanish at
the same rate and may only vanish at a slower rate $\delta h(\delta)$. The
real number model $\mathbf{R}$ is called the hard or string model when the
space $\mathbf{R^{*}}$ is called the soft or fluid model of real numbers
\cite{DPD Sen}. The ordinary differential measure $dx$ gets extended in
$\mathbf{R^{*}}$ as $d(h(x)x)$.
2. Consider the open interval $(\delta,\delta^{-1})\subset\mathbf{R^{*}}$. In
the asymptotic limit $\delta\rightarrow0^{+}$, the duality structure
identifies the right neighbourhood of $\delta$ with the left neighbourhood of
$\delta^{-1}$ in a nontrivial manner. As a consequence, the linear
(translation) group action on $\mathbf{R}$ is extended to a nonlinear
SL($2,R$) group on $\mathbf{R^{*}}$. Infact, the translation subgroup acts on
$\mathbf{R}$, when the inversion acts nontrivially only on $\mathbf{R^{*}}$ in
the sense that the visibility norm $h$ is invariant under inversion $\hat i:
\ h(\hat i A)=\hat i(h(A))$ where $\hat i(A)=\{a_{n}^{-1}\times a_{n}%
^{(a^{-}_{m_{n}})^{-1}}\}, \ A=\{a_{n}\times a_{n}^{-a^{-}_{m_{n}}}\}$. For a
translation $T_{r}$ by a shift $r$, on the other hand, $h(T_{r}(A))=h(A)$ and
hence $T_{r}(A)=A \ \Rightarrow r=0$ (i.e. $T$ acts trivially on $\bf{R^*-R}$). Above two
\emph{salient} properties of the duality structure are expected to have
significant application in nonlinear problems.
3. To give an example of the intricate nonlinear structure that can get
encoded into a well behaved (smooth) function in $\mathbf{R}$, let us consider
the simplest case of a real variable $x$. In $\mathbf{R^{*}}$ the variable $x$
gets extended to, say, $X=x e^{\phi(\log X)}$. The function $\phi$ exposing
the nonlinear dependence is also assumed to be differentiable. Differentiating
$X$ with respect to $x$ one gets $xX^{\prime}(1-\phi^{\prime})=X$, where
$\prime$ denotes derivation with the argument. We now assume that $\phi(\log
X)$ is vanishingly small (i.e. less than accuracy level $\delta$ in any given
application) for $0<x<\infty$ and O(1) when $\log X>>1$ i.e. $x\rightarrow0$
or $\infty$. As a consequence, existence of $\phi$ is felt only in the
asymptotic neighbourhoods (Remark 2) of 0 or $\infty$. We now make a further
assumption that $\phi^{\prime}=0$ almost everywhere in an asymptotic
neighbourhood, but everywhere in $0<x<\infty$. Then $X$ satisfies
$xX^{\prime}=X$ a.e. in $\mathbf{R^{*}}$. Thus ordinary variable
$x\in\mathbf{R}$ gets extended in $\mathbf{R^{*}}$ as $X$ which has the
intermittent property of a Cantor devil's Stairecase function in an asymptotic
neighbourhood. Since, under duality structure, such a neighbourhood has
ultrametric topology, $X$ in fact satisfies the above scale invariant equation
everywhere in $\mathbf{R^{*}}$, because ordinary non-differentiability at the
points of the associated Cantor set is removed by inversion mediated jump
increments \cite{DPD1, DPD New}. This example tells that an ordinary function
can have nonlinear and nonlocal functional dependence with itself, along with
rhythmic (intermittent) variability that can have significant amplification in
an asymptotic sector.
4. The asymptotic scaling variables $h_{0}(\varepsilon)$ and $H_{RG}%
(\tilde\tau)$ introduced in Sec. 4 correspond to the associated visibility
norm $h(A)$ defined above. A real variable $t\in\mathbf{R}$ approaching
asymptotically either to 0 or $\infty$ has natural images in $\mathbf{R}^{*}$
in the form $\tau_{0}=t\times t^{-h^{-}(\varepsilon t)}, \ h^{-}(\varepsilon
t)<1$ and $\tau_{\infty}=t\times t^{h^{+}(\varepsilon t)}, \ h^{+}(\varepsilon
t)>1$ respectively. The scaling exponents $h^{\pm}$ encode asymptotic scaling
information of a given nonlinear system. Further, $(h^{-}(\varepsilon
t))^{-1}\propto h^{+}(\varepsilon t)$ by duality. In Sec.4, we discuss how
such information can be systematically extracted in the case of a limit cycle
for a nonlinear oscillator (see also below).
5. The fine structures in $\mathbf{R^{*}}$ remain inactive (passive/hidden) in
absence of any stimulus, either intrinsic or external. In presence of an
external input, say, the actions of the nontrivial component of the metric $d$
and the associated duality structure become manifest. The RG analysis makes
a room for direct implementation of the intrinsically realized duality structure
in the context of a nonlinear system in the soft model $\mathbf{R^{*}}$.
\subsection*{B. Application: Limit Cycle}
Consider a general nonlinear oscillator given by
\begin{equation}
\label{NO}\ddot{x}+x= \varepsilon f(x,\dot x)
\end{equation}
We assume $f$ such that the system admits a unique isolated cycle for
$\varepsilon>0$ and other relevant parameter values. For a finite nonlinearity
$\varepsilon>1$ (say), the usual linear time $t$ is extended to one enjoying
\emph{right} asymptotic correction $t\rightarrow\ T_{i}=t\phi_{i}(\tilde
\tau(t))$ as $t\rightarrow\infty$ through linear scales $\varepsilon^{i}$,
where $\phi_{i}(\tilde\tau)$ stands succinctly for the nontrivial
intrinsically generated \emph{slowly varying} scaling components arising from
the associated visibility norm. Here, $\tilde\tau$, as usual denotes an O(1)
rescaled variable in the neighbourhood of linear scales $\varepsilon^{i}$. In
the case of a nonlinear planar autonomous system the relevant dynamical
quantities are only amplitude and phase of the nonlinear oscillation and so we
have only two asymptotic scaling functions $\phi_{i}(\tilde\tau), \ i=1,2$
which get \emph{selected} naturally so as to facilitate direct
non-perturbative calculation of the asymptotic properties i.e. the amplitude
and phase of the limit cycle of the system. An implementation of this
non-pertubative scheme in the perturbative RG formalism is presented in
Sec.4.1 for computation of the amplitude of the concerned oscillators. In
Sec.4.2, we have demonstrated that the computed plot of the limit cycle for
the Rayleigh and VdP oscillators could be matched arbitrarily closely for
appropriate choices of the slowly varying nonlinear time when the amplitude
and phase of the unperturbed periodic solution \emph{flow linearly} in the
appropriately chosen nonlinear scaling time variables.
Here, we give an alternative derivation of the nonperturbative
\emph{relaxation oscillation} flow equations ab-initio from the slowly varying
nonlinear time in the context of the Rayleigh equation (\ref{Rayleigh}) with
$\varepsilon>>1$. It will transpire that the new approach is free of any
divergence problem because of its inbuilt RG cancellations via duality
principle. Since we are interested in the planar limit cycle properties, we
assume that all the relevant quantities e.g. the solution $y$, amplitude $a$
and phase $\psi$ are functions of asymptotic time variable $t\sim
\varepsilon^{n},\ n>>1$ and the associated nontrivial scaling variables
$\tau_{1}=\phi_{1}(\tilde{\tau})$ and $\tau_{2}=\phi_{2}(\tilde{\tau})$ for a
rescaled $\tilde{\tau}\sim O(1)$. Higher order scaling variables $\tau
_{n},\ n>2$ of the nonlinear structure of time variable may become relevant
for a non-planar system. Accordingly, we write the ansatz $y(t,\tau_{1}%
,\tau_{2})=y_{0}(t)+Y_{1}(\tau_{1})+Y_{2}(\tau_{2})$ for the limit cycle
solution involving multiple time scales (only three for the planar system).
Assuming slow variations of nonlinear scales $\tau_{i}=\phi_{i},\ i=1,2$, viz.
$|\phi_{i}^{\prime\prime}|<<|{\phi_{i}^{\prime}}^{2}|<<1,\ \phi_{i}^{\prime
}=\frac{d\phi_{i}}{d\tilde{\tau}}$, as $t\rightarrow\varepsilon^{n}%
,\ n\rightarrow\infty$ and noting that $\frac{dy}{dt}=\frac{\partial y_{0}%
}{\partial t}+\sum_{i}\dot{\tilde{\tau}}\phi_{i}^{\prime}{\frac{\partial
Y_{i}}{\partial\tau_{i}}}$ etc., the Rayleigh equation (\ref{Rayleigh})
simplifies to
\begin{equation}
\frac{\partial^{2}y_{0}}{\partial t^{2}}+y_{0}+\sum_{i}Y_{i}=\varepsilon
(\frac{\partial y_{0}}{\partial t}+\sum_{i}\dot{\tilde{\tau}}\phi_{i}^{\prime
}{\frac{\partial Y_{i}}{\partial\tau_{i}}})-\frac{\varepsilon}{3}%
((\frac{\partial y_{0}}{\partial t})^{3}+3(\frac{\partial y_{0}}{\partial
t})^{2}\sum_{i}\dot{\tilde{\tau}}\phi_{i}^{\prime}{\frac{\partial Y_{i}%
}{\partial\tau_{i}}})\label{mulscale}%
\end{equation}
where we drop all higher order terms involving $\phi_{i}^{\prime\prime}$ and
${\phi_{i}^{\prime}}^{2}$. Assuming $y_{0}(t)=a(\tau_{1},\tau_{2})\cos
(t+\psi(\tau_{1},\tau_{2}))$ with \emph{flowing} amplitude and phase in
scaling times $\tau_{1}$ and $\tau_{2}$ so that
\begin{equation}
\frac{\partial^{2}y_{0}}{\partial t^{2}}+y_{0}=0\label{linear1}%
\end{equation}
we next get a simplified linearized evolution for the nonlinear components of
the asymptotic limit cycle solution in the form
\begin{equation}
(1-(\frac{\partial y_{0}}{\partial t})^{2})\sum\dot{\phi}_{i}\frac{\partial
Y_{i}}{\partial\tau_{i}}=\{\frac{1}{3}(\frac{\partial y_{0}}{\partial t}%
)^{3}-\frac{\partial y_{0}}{\partial t}\}+\varepsilon^{-1}\sum Y_{i}%
\label{linear2}%
\end{equation}
where $\dot{\phi}_{i}=\frac{d\phi_{i}}{dt}$. As a consequence, under the
assumption of slowly varying nonlinear time scales, a second order nonlinear
planar system (\ref{Rayleigh}) would decompose into a linear second order
partial differential equation (\ref{linear1}) for the zero level solution
$y_{0}$ and an associated first order partial differential equation
(\ref{linear2}) for the nonlinear scale dependent components $Y_{i}$. Clearly,
analogous decomposition holds actually for a larger class of planar autonomous
systems (\ref{NO}) having a unique limit cycle solution. Extension of this
result to multiple limit cycles would be considered separately.
We note here that since the system (\ref{linear1}) and (\ref{linear2}) is
\emph{under determined}, there is room for further restrictions to solve the
system self-consistently. To re-derive the RG flow equations (\ref{Amp}) and
(\ref{Phase}) from (\ref{linear2}), we now make following assumptions: we
write (i) $a(\tau_{1},\tau_{2})=a(\tau_{1}),\ \psi(\tau_{1},\tau_{2}%
)=\psi(\tau_{2})$ so that amplitude varies slowly with first order scale
$\tau_{1}$ when the second order scale $\tau_{2}$ and phase $\psi$ remain
almost constant. On the other hand as $a$ stabilizes to an almost constant
value, the phase begins to flow, though slowly with the second order scale
$\tau_{2}$. Such slow, almost constant, rhythmic cooperative variations of
$\tau_{1}$ and $\tau_{2}$ are modeled, depending on the specific problem under
consideration (see below and c.f. Sec. 4.2), to retrieve the RG flow equations
correctly. As shown in the example (Remark 3, Appendix Sec.A) such a rhythmic
nonlinear variation does exist in an ultrametric neighbourhood in
$\mathbf{R^{\ast}}$. To further quantify the slow variation of dynamical
variables, we next impose the condition that (ii) \emph{the total variation of
the exact solution $y(t,\tau_{1},\tau_{2})$ with respect to each slow variable
$\tau_{i}$ along the full periodic cycle $C$ must vanish viz, $\int_{C}%
\frac{\partial y}{\partial\tau_{i}}dt=0$ for each $i$}. To avoid trivialities
i.e. $\int_{C}\cos(t+\psi)dt=0$ etc., we, however, evaluate the concerned
integrals only on the quarter cycle, with the understanding that phase shifts
of $\pi/2$ are absorbed in the definition of $\psi$.
In the sufficiently large $\varepsilon>1$ relaxation oscillation, one can
further simplify (\ref{linear2}) by dropping the $\varepsilon^{-1}$ term to
obtain
\begin{equation}
\label{linear3}\sum\dot\phi_{i}\frac{\partial Y_{i}}{\partial\tau_{i}}%
=\frac{\frac{1}{3}(\frac{\partial y_{0}}{\partial t})^{3}-\frac{\partial
y_{0}}{\partial t}}{1-(\frac{\partial y_{0}}{\partial t})^{2}}\equiv
\Phi(y_{0t}), \ y_{0t}=\frac{\partial y_{0}}{\partial t}%
\end{equation}
To make contact with RG flow equations (\ref{Amp}) and (\ref{Phase}) one now
exploits \emph{the freedom of right choice} in the functional forms of
nonlinear scales. For the Rayleigh equation, we now set for slow,
cooperatively active functional dependence (a) $\dot\phi_{1}=\Phi(y_{0t}%
)S_{1}^{-1}(a,\psi,t), \ \dot\phi_{2}=0$ and (b) $\dot\phi_{1}=0, \ \dot
\phi_{2}=\Phi(y_{0t})S_{2}^{-1}(a,\psi,t)$ for successive slow variations, as
described in (i), of the scales $\tau_{1}$ and $\tau_{2}$ respectively, where,
$S_{1}= \frac{1}{2}a(\frac{a^{2}}{4}-1)\cos(t+\psi)$ and $S_{2}=\frac{1}%
{8}(1-\frac{a^{4}}{32})\sin(t+\psi)$ (recall the example in Remark 3 of Sec.A
above highlighting wide possible choices and intricate functional dependence).
These choices for $S_{1}$ and $S_{2}$ would yield the RG flow equations when
condition (ii) is invoked.
Note that the relations in both (a) and (b) are truly nonlinear; the dynamical
variables $a$ and $\psi$ in $S_{1}$ and $S_{2}$ depend implicitly in $\phi
_{1}$ and $\phi_{2}$ respectively, which, in turn are \emph{slowly varying} as
the linear parameter $t$ is assumed to vary in a neighbourhood of
$\varepsilon^{n}$ for a large but fixed $n$. Invoking the global slow
variation condition (ii) for each $i$, in conjunction with the ansatz (a) and
(b), one finally deduce the amplitude and phase flow equations (\ref{Amp}) and
(\ref{Phase}) in slow variables $\tau_{1}$ and $\tau_{2}$ respectively.
In the present format, the flow equations, however, have got \emph{new}
interpretations: Amplitude and phase must flow in successive rhythmic manner;
phase remains almost constant (i.e. $\frac{\partial\psi}{\partial\tau_{i}}=0$
for each $i$) when amplitude varies slowly with $\tau_{1}$ towards an almost
constant value. Subsequently, the flowing of $a$ is halted temporarily (i.e.
$\frac{\partial a}{\partial\tau_{i}}=0$), initiating flowing of $\psi$ in next
level variable $\tau_{2}$. This rhythmic oscillation would obviously continue
indefinitely over a cycle. The RG flow equations could be treated as
non-perturbative because of \emph{implicit} connections of nonlinear scaling
time functions with amplitude and phase via intrinsically defined duality
principle (c.f. Remark 3 above).
To summarize, based on perturbative RG formalism we have presented an
alternative approach in deriving non-perturbative flow equations of relevant
asymptotic dynamical quantities of a planar autonomous limit cycle problem,
from nonlinear scale invariant time scales, which become available in an
extended analytic framework incorporating duality structure. Non-perturbative
information of asymptotic quantities get naturally encoded into nonlinear
scales, that can be exploited judiciously to extract desired asymptotic
analytic properties of a relevant dynamical quantity. An algorithmic procedure
of extracting such information is explained in estimating both the limit cycle
amplitude and trajectory for Rayleigh and VdP equations.
\subsection*{C. Matching Arcs and Segments}
$\left( a\right) $ Piecewise smooth matching curves for upper half of the
approximate Rayleigh limit cycle for $\varepsilon=5$:
$z\left( y\right) =\left\{
\begin{array}
[c]{cc}%
0.02-\sqrt{1.96-\left( y+3\right) ^{2}} & -4.96<y\leq-4.393\\
10y+43.9 & -4.393<y\leq-4.23\\
1.46+\sqrt{0.35-\left( y+3.65\right) ^{2}} & -4.23<y\leq-3.6\\
-0.1y+1.689 & -3.6<y\leq3.12\\
-0.02+\sqrt{1.96-\left( y-3\right) ^{2}} & 3.12<y\leq4.96
\end{array}
\right. $
$\left( b\right) $ Piecewise smooth matching curves for upper half of the
approximate VdP limit cycle for $\varepsilon=5$:
$z\left( y\right) =\left\{
\begin{array}
[c]{cc}%
-0.388+\sqrt{0.352-\left( y+1.438\right) ^{2}} & -2.05<y\leq-1.7\\
1.8-\sqrt{3.2-\left( y+2.38\right) ^{2}} & -1.7<y\leq-0.633\\
4.5y+4.25 & -0.633<y\leq0.6\\
6.5+\sqrt{2.76-\left( y-2.2\right) ^{2}} & 0.6<y\leq0.9\\
7.325+\sqrt{0.063-\left( y-1.04\right) ^{2}} & 0.90<y\leq1.28\\
0.38+\sqrt{1530-\left( y+37.2\right) ^{2}} & 1.28<y\leq1.8\\
-13y+26.8 & 1.8<y\leq2.033\\
0.388+\sqrt{0.352-\left( y-1.438\right) ^{2}} & 2.033<y\leq2.05
\end{array}
\right. $
|
\section{Introduction \label{sec:intro}}
The search for particles beyond the Standard Model (SM) is one of the key issues of the ATLAS and CMS experiments at LHC. In particular, these experiments can test the theories with extra dimensions, which aim to solve the hierarchy problem by bringing the gravity scale closer to the electroweak scale (For a review see, e.g., \cite{hewett}). Over a decade ago, Arkani-Hamed, Dimopoulos, and Dvali~\cite{add}
proposed a scenario whereby the SM is constrained to the common 3 + 1 space-time dimensions (brane), while the gravity is free to propagate throughout a larger multidimensional space (bulk).
In this Large Extra Dimensions (LED) scenario, $n$ extra spatial dimensions are compactified on a torus with common circumference $R$. Then the four dimensional Planck scale $M_P$ is no longer the relevant scale but is related to the fundamental scale $M_S$ as follows $M_P^2 \approx M_S^{n+2}R^n$ where $M_S$ is ${\cal{O}}$(TeV).
Moreover, the $4+n$ dimensional graviton corresponds to a tower of massive Kaluza-Klein (KK) modes in four dimensions. The interactions of these spin-2 KK gravitons with the SM matter can be described by an effective theory with the Lagrangian given by~\cite{giudice, han}
\begin{eqnarray}
{\cal{L}} = - \frac{\kappa}{2} \sum_{\vec{n}=0}^{\infty} T^{\mu \nu}(x) h_{\mu \nu}^{\vec{n}}(x) \,\,,
\label{lagrangiano}
\end{eqnarray}
where $\kappa = \sqrt{16 \pi}/M_P$, the massive KK gravitons are labelled by a $d$-dimensional vector of positive integers and $T^{\mu \nu}$ denotes the energy-momentum tensor of the SM.
In what follows we assume that
\begin{eqnarray}
\kappa^2R^n = 8 \pi (4\pi)^{n/2} \Gamma(n/2)M_S^{-(n+2)}.
\end{eqnarray}
The Feynman rules that follow from Eq.~(\ref{lagrangiano}) can be found in Refs. \cite{giudice, han}. The individual KK resonances have masses equal to $m_{(\vec{n})} = |\vec{n}|/R$ and thus the mass gap between the neighbouring modes $\Delta m \propto 1/R$ is small for $n$ not too large, which allow us to approximate the discrete mass spectrum by a continuum.
The individual KK mode couples with a strength $\propto 1/M_P$ to the SM fields. However, since there are many KK modes, the total coupling strength is of the order of $1/M_S$ after summing up all of them. Moreover, the excited gravitons preferentially decay into two gauge bosons rather than into two leptons, because the graviton has spin 2, and so fermions cannot be produced in an $s$ wave.
Searches for the large extra dimensions scenario via virtual-graviton effects were performed at HERA, LEP, the Tevatron, and the LHC (For a recent review see, e.g., \cite{review_exp}). Recent experimental results from the ATLAS and CMS Collaborations sets lower limits on $M_S$, being about 3.5 TeV (2.7 TeV) for $n = 2$ ($n = 6$) \cite{refsexp}. From the theoretical side, many studies on the virtual KK graviton effects up to the NLO exist ~\cite{calculos_nlo}. These include the processes of fermion-pair, multijet, and diboson production.
In general the contribution associated to the graviton exchange is mostly undetectable because of the much larger Standard Model background. The lower SM background occurs in the production of $Z$ pairs which have the smallest cross section of all the diboson processes. In contrast, this process have tree-level contributions in the LED scenario, which motivated the study of the $Z$ boson pair production mediated by the Kaluza-Klein graviton in large extra dimensions \cite{koch,gao,agarwal,ravindran_recent}. These studies consider inclusive processes, where the hadrons colliding dissociate after the interaction. In this paper we extend these previous studies for {\it diffractive} processes, in which the hadrons colliding lose only a small fraction of their initial energy and escape the central detectors~\cite{forshaw_review}. In the diffractive interaction the hadron can dissociate (inclusive diffractive process) or remain intact (exclusive diffractive process). In what follows we consider the
$Z$ boson pair production in central exclusive processes (CEP), $h_1 + h_2 \rightarrow h_1 \otimes ZZ \otimes h_2$, which are characterized by empty regions in pseudo-rapidity, called rapidity gaps (denoted by $\otimes$), that separate the intact very forward hadron from the $ZZ$ final state produced in the interaction [See Fig.~\ref{fig1}(a)]. Exclusivity means that nothing else is produced except the leading hadrons and the central object. Moreover, we consider central inclusive processes (CIP), $h_1 + h_2 \rightarrow X \otimes ZZ \otimes Y$, which also exhibit rapidity gaps but the incident hadrons dissociates in the products $X$ and $Y$ [See Fig.~\ref{fig1}(b)]. In general the rapidity gaps present in inclusive processes are smaller than in the exclusive case. A QCD mechanism for the central inclusive and exclusive diffractive production of a heavy central system has been proposed by Khoze, Martin and Ryskin~\cite{kmr_first,kmr_prosp,kmr_review}, denoted Durham model hereafter, with its predictions in
broad agreement with the observed rates measured by CDF Collaboration~\cite{cdf} (For a recent review see Ref.~\cite{forshaw_review}). In this paper the Durham model is applied, for the first time, for the $Z$ boson pair production considering the KK graviton exchange. Our main motivation is associated to the fact that in the Standard Model the $Z$ boson pair is not produced at leading order by the gluon fusion. Consequently, the $Z$ boson pair cannot be produced at leading order in diffractive processes. At next-to-leading-order the $gg \rightarrow ZZ$ subprocess via the quark loop contributes \cite{zzsm}. However, as will be demonstrated below for the case of central inclusive processes, the resulting cross section is one order of magnitude smaller than the LED prediction. It implies a very low QCD background, making its observation a signature of large extra dimensions.
\begin{figure}[t]
\begin{tabular}{cc}
\includegraphics[scale=0.35] {process_zz_excl.eps} & \includegraphics[scale=0.35]{process_zz_incl.eps} \\
(a) & (b)
\end{tabular}
\caption{(color online) $Z$ boson pair production in (a) central exclusive and (b) central inclusive diffractive processes for $h_1 h_2$ collisions, with $h_i = p$ or $Pb$. $\langle {\mathcal S}^2 \rangle$ is the survival probability gap, which gives the probability that secondaries, which are produced by soft rescatterings, do not populate the rapidity gaps.}
\label{fig1}
\end{figure}
This paper is organized as follows. In the next section we present a
brief review of the formalism necessary for the calculation of the $Z$ boson pair production in diffractive interactions in hadron - hadron collisions. Moreover, we discuss the extension of the survival probability gap for nuclear collisions. In Section~\ref{sec:res} we present our results for the $Z$ boson pair production in $pp$, $pPb$ and $PbPb$ collisions at LHC energies. Finally, in Section~\ref{sec:sum} we present a summary of our main conclusions.
\section{The $Z$ boson pair production in diffractive interactions \label{sec:modfor}}
In the Durham model \cite{kmr_prosp} the total cross section for the diffractive production of a $Z$ boson pair reads
\begin{equation}
\sigma(\sqrt{s}) = \int dy \int \frac{dM^2_{ZZ}}{M^2_{ZZ} }{\mathcal L}_{gg} \hat{\sigma}_{gg \rightarrow ZZ}(M^2_{ZZ})
\label{sigma}
\end{equation}
where ${\mathcal L}_{gg}$ is the effective gluon - gluon luminosity for the production of a system with squared invariant mass $M^2_{ZZ}$ and rapidity $y$, and $\hat{\sigma}$ is the cross section for the subprocess $gg \rightarrow ZZ$.
For the $Z$ boson pair production in central exclusive processes (CEP), $h_1 + h_2 \rightarrow h_1 \otimes ZZ \otimes h_2$, the effective gluon - gluon luminosity is given by [See Fig.~\ref{fig1}(a)]
\begin{equation}
{\mathcal L}_{gg}^\mathrm{excl} = \langle {\mathcal S}^2_{excl} \rangle \left[ {\cal{C}} \int \frac{dQ_t^2}{Q_t^4} f_g(x_1,x_1^{\prime},Q_t^2, \mu^2) f_g(x_2,x_2^{\prime},Q_t^2, \mu^2)\right]^2\,\,,
\end{equation}
where $\langle {\mathcal S}^2_{excl} \rangle$ is the survival probability gap for exclusive processes, which gives the probability that secondaries, which are produced by soft rescatterings, do not populate the rapidity gaps and ${\cal{C}} = \pi/[(N_C^2 - 1)b]$, with $b$ the $t$-slope ($b = \unit[4.0]{GeV^{-2}}$ in what follows). Moreover, $Q_t^2$ and $x_i^{\prime}$ are respectively the virtuality and longitudinal momentum of the soft gluon needed for color screening, $x_1$ and $x_2$ are the longitudinal momentum of the gluons which participate of the hard subprocess and the quantities $f_g$ are the skewed unintegrated gluon densities. Since $
(x^{\prime} \approx {Q_t}/{\sqrt{s}}) \ll (x \approx {M_{ZZ}}/{\sqrt{s}}) \ll 1$,
it is possible to express $f_g(x_1,x_1^{\prime},Q_t^2, \mu^2)$, to single log accuracy, in terms of the conventional integral gluon density $g(x)$, together with a known Sudakov suppression $T$ which ensures that the active gluons do not radiate in the evolution from $Q_t$ up to the hard scale $\mu = 2 M_{Z}$, which we assume to be the renormalization and factorization scales of the process. The choice of this scale introduces roughly a factor of two uncertainty when varying the hard scale $\mu$ between $2 M_{ZZ}$ and $M_{ZZ}/2$. Following~\cite{kmr_prosp} we will assume that
\begin{equation}
f_g(x, Q_t^2, \mu^2) = S_g \frac{\partial}{\partial \ln Q_t^2} \left[ \sqrt{T(Q_t,\mu)}\ xg(x,Q_t^2) \right]\,\,,
\end{equation}
where $S_g$ accounts for the single $\log Q^2$ skewed effect and is given by
\[ S_g = \frac{2^{2\lambda+3}}{\sqrt{\pi}} \frac{\Gamma(\lambda + 5/2)}{\Gamma(\lambda + 4)} \]
if one assumes a single powe-law behaviour for the gluon distribution, $xg(x,Q_t^2) \approx x^{-\lambda}$, with $S_g \sim 1.2 $ for LHC. The Sudakov factor is given by
\begin{eqnarray}
T(Q_t,\mu) = \exp \left\{ -\int_{Q_t^2}^{\mu^2} \frac{dk_t^2}{k_t^2} \frac{\alpha_s(k_t^2)}{2\pi} \int_{0}^{1-\Delta} dz \, \left[ zP_{gg}(z) + \sum_{q} P_{qg}(z) \right] \right\},
\end{eqnarray}
with $k_t$ being an intermediate scale between $Q_t$ and $\mu$, $\Delta = k_t/(\mu + k_t)$, and $P_{gg}(z)$ and $P_{qg}(z)$ are the leading order Dokshitzer - Gribov - Lipatov - Altarelli - Parisi (DGLAP) splitting functions~\cite{dglap}.
In this paper we will calculate $f_g$ in the proton case considering that the integrated gluon distribution $xg(x,Q_T^2)$ is described by the MSTW2008lo parametrization~\cite{mstw}. In the nuclear case we will include the shadowing effects in $f_g^A$ considering that the nuclear gluon distribution is given by the EKS98 parametrization~\cite{eks}, where $ x g_A(x,Q_t^2) = A R_g^A(x, Q_t^2) x g_p(x, Q_t^2)$,
with $R_g$ describing the nuclear effects in $xg_A$.
In order to obtain realistic predictions for the exclusive $Z$ boson pair production using the Durham model, it is crucial to use an adequate value for the survival probability gap, $\langle {\mathcal S}^2_{excl} \rangle$. This factor is the probability that secondaries, which are produced by soft rescatterings do not populate the rapidity gaps (For a detailed discussion see~\cite{kmr_review}). In the case of proton-proton collisions, we will assume that $\langle {\mathcal S}^2_{excl} \rangle = 3 \, \%$
for LHC energies~\cite{kmr_prosp}. However, the value of the survival probability for nuclear collisions still is an open question. An estimate of $\langle {\mathcal S}^2_{excl} \rangle$ for nuclear collisions was calculated in \cite{miller} using the Glauber model. Another conservative estimate can be obtained assuming that \cite{vicwerner}: $\langle {\mathcal S}^2_{excl} \rangle_{A_1A_2} = \langle {\mathcal S}^2_{excl} \rangle_{pp}/(A_1 \cdot A_2)$ (For a discussion about nuclear diffraction see~\cite{dif_nuc}). In what follows we will consider the latter model for $\langle {\mathcal S}^2_{excl} \rangle_{A_1A_2}$ when considering the central exclusive $Z$ boson pair production.
On the other hand, the effective gluon - gluon luminosity for the $Z$ boson pair production in central inclusive processes (CIP), $h_1 + h_2 \rightarrow X \otimes ZZ \otimes Y$, is given by~\cite{kmr_prosp}
\begin{eqnarray}
\mathcal{L}_{gg}^\mathrm{incl} = & \langle {\mathcal S}^2_\mathrm{incl} \rangle \frac{\alpha_s^4}{\pi}\left(\frac{N_C^2}{N_C^2 -1}\right)^2 \frac{1}{(Y_1 + Y_ 2)^2}\int \frac{dt_1}{t_1}\frac{dt_2}{t_2}\exp\left(-\frac{3\alpha_s}{\pi}\Delta\eta\left|\ln\frac{t_1}{t_2}\right|\right) \times \nonumber \\
& T(\sqrt{|t_1|}, {\mu}) T(\sqrt{|t_2|}, {\mu}) \int^1_{x_1^\mathrm{min}}\frac{dx_1}{x_1}\mathcal{G}(x_1, k^2_{1\perp}) \int^1_{x_2^\mathrm{min}} \frac{dx_2}{x_2}\mathcal{G}(x_2, k^2_{2\perp}),
\end{eqnarray}
where $\langle {\mathcal S}^2_\mathrm{incl} \rangle$ is the gap survival probability for inclusive processes, $t_i = - k^2_{i\perp}$, $\Delta\eta$ is the rapidity gap size on either side of the $Z$ boson pair and $Y_1 = Y_2 = \frac{3\alpha_s}{2\pi}\Delta\eta$, with $\alpha_s = 0.2$ [See Fig.~\ref{fig1}(b)]. Moreover, $\mathcal{G}(x_i, k^2_{i\perp}) $ are the effective parton distributions given by
\begin{equation}
\mathcal{G}(x_i, k^2_{i\perp}) = x_i g(x_i, k^2_{i\perp}) + \frac{N_c^2}{(N_c^2 -1)^2}\sum_{q} x_i\left[ q(x_i, k^2_{i\perp}) +\bar{q}(x_i, k^2_{i\perp}) \right].
\end{equation}
and
\[ x_i^\mathrm{min} = \frac{e^y}{\sqrt{s}} \left[ M_{ZZ} + k_{i\perp}e^{\Delta\eta} \right]. \]
The basic idea is that the initial hadrons dissociate and the system $ZZ$ with mass $M_{ZZ}$ is produced with rapidity gaps with size $\Delta \eta$ on either side. This process is characterized by two rapidity gaps, the central system $ZZ$ and the presence of secondary particles in the hadron fragmentation regions. An open question is the magnitude of the absorption corrections for the central inclusive processes which determines $\langle \mathcal{S}^2_\mathrm{incl} \rangle$, which is expected to depend on the kinematics of the process~\cite{kmr_prosp}. As it is expected that $\langle {\mathcal S}^2_\mathrm{excl} \rangle \le \langle \mathcal{S}^2_\mathrm{incl} \rangle$, we will assume, for simplicity, in our analysis that $\langle \mathcal{S}^2_\mathrm{incl}\rangle_{pp} = 0.1$. Moreover, we will also assume that $\langle {\mathcal S}^2_\mathrm{incl} \rangle_{A_1A_2} = \langle {\mathcal S}^2_\mathrm{incl} \rangle_{pp}/(A_1\cdot A_2)$. However, this subject deserve more detailed studies in the future.
In contrast to the Standard Model, in the LED scenario, the $gg \rightarrow ZZ$ subprocess contributes at the tree level through the exchange of virtual KK gravitons in the $s$-channel, with the cross section being given by~\cite{gao},
\begin{equation}
\hat{\sigma}_{gg \rightarrow ZZ} = \int_{-1}^{+1} d(\cos \theta) \frac{d\hat{\sigma}_{gg}}{d\cos\theta} \,\,,
\end{equation}
with
\begin{eqnarray}
\frac{d\hat{\sigma}_{gg}}{d\cos \theta}&=&\frac{
\pi\hat{s}^3\sqrt{1-4w}
\mathcal{F}^2(n)\left(\pi^2+4\mathcal{T}^2(n)\right)}{1024}\nonumber\\
&&\times\left\{
3z^4\left(1-4w\right)^2+2z^2\left(1-4w\right)\left(5+12w\right)
+\left(1+12w\right)\left(3+4w\right)\right\},
\end{eqnarray}
and
\begin{equation}
{\mathcal F}(n)=\hat {s}^{n/2-1}/M_S^{n+2}, \ {\mathcal
T}(n)=I(\Lambda/\sqrt{\hat s}),
\end{equation}
where $w=m_Z^2/\hat{s}$ and $z=\cos\theta$.
The function $I(\Lambda/\sqrt{\hat s})$ is directly related with the graviton propagator in the LED scenario~\cite{han}, being given by
\begin{equation}
I(\Lambda/\sqrt{s}) = PV\int_0^{\Lambda/\sqrt{s}} dy \frac{y^{n-1}}{1-y^2} \,\,,
\end{equation}
where $PV$ means the principal value of the integral and $\Lambda$ is identified with the fundamental scale $M_S$ in $4+n$ dimensions~\cite{giudice,han}.
\section{Results \label{sec:res}}
Before to present our predictions for the central exclusive and inclusive $Z$ boson pair production in diffractive processes, a comment is in order. Our calculations are performed at leading order. Next-to-leading-order (NLO) corrections for the inclusive $Z$ boson pair production, $p + p \rightarrow ZZ X$, where both colliding protons dissociate and rapidity gaps are not present in the final state, were estimated in Ref.~\cite{agarwal}
assuming that the interaction is mediated by the Kaluza-Klein graviton in the LED scenario
and are large ($\approx \, 2$). Consequently, our predictions for the inclusive production should be considered a lower bound. In contrast, the magnitude of the NLO corrections for diffractive processes still is an open question (For a recent study, see Ref.~\cite{forshawcou}). Results presented in Refs. \cite{cudell,dechambre} indicate that the predictions for central production are strongly influenced by the choice of the parametrization used for the parton distribution functions and by the treatment of the Sudakov form factor and rapidity gap survival probability, which introduces non-negligible uncertainties in the calculations. Such uncertainties also are present in our analysis. In principle, the uncertainties could be reduced with a possible early LHC measurement of exclusive jets cross section \cite{dechambre}.
\begin{figure}[t]
\begin{center}
\includegraphics*[width=8cm]{inclzz.eps}
\caption{(color online) Dependence on $M_S$ of the cross section for the inclusive $Z$ boson pair production in $pp$ collisions at $\sqrt{s} = 14$ TeV. The SM prediction also is presented for comparison.}
\label{fig2}
\end{center}
\end{figure}
In Fig.~\ref{fig2} we present our predictions for the inclusive production ($p + p \rightarrow ZZ X$) at $\sqrt{s} = \unit[14]{TeV}$. For comparison we also show the SM prediction for this process, which is $\sigma_{tot} = \unit[17]{pb}$ and was obtained at NLO in Ref.~\cite{zzsm} (For recent experimental results see, e.g. Ref.~\cite{Chatrchyan:2013oev}). As already verified in Ref.~\cite{gao}, the $Z$ boson pair production by KK graviton exchange is only competitive for small values of $M_S$ and number of extra dimensions $n$. In particular, if we taken into account the recent lower bounds on $M_S$ set by the ATLAS and CMS Collaborations \cite{refsexp}, we have that the LED predictions are at least a factor 20 smaller than the SM one.
\begin{figure}[t]
\begin{center}
\begin{tabular}{cc}
\includegraphics*[width=6cm]{zz_excl_ms-n.eps} & \includegraphics*[width=6cm]{zz_excl_sqrts-n_REV.eps} \\
\end{tabular}
\caption{(color online) Dependence on $M_S$ (left panel) and on the center-of-mass energy (right panel) of the total cross section for the central {\it exclusive} double $Z$ production in proton-proton collisions for different numbers of extra dimensions.}
\label{fig:exclpp}
\end{center}
\end{figure}
In Fig.~\ref{fig:exclpp} we present our predictions for the central exclusive production of a $Z$ boson pair, $p + p \rightarrow p \otimes ZZ \otimes p$, at LHC energies. In the left panel we present the dependence on $M_S$ of the total cross section for different values of $n$ at fixed center-of-mass energy ($\sqrt{s} = \unit[14]{TeV}$). As expected, it strongly decreases at larger values of $M_S$ and $n$. In the right panel, we present our predictions for the energy dependence of the cross section considering $M_S = \unit[3.5]{TeV}$ and different values of $n$. We predict very small values for the exclusive cross section in the LHC kinematical range, which makes the experimental analysis of this process a very hard task.
\begin{figure}[ht]
\begin{center}
\begin{tabular}{cc}
\includegraphics*[width=6cm]{zz_incl_ms-n.eps} & \includegraphics*[width=6cm]{zz_incl_sqrts-n_REV.eps} \\
\end{tabular}
\caption{(color online) Dependence on $M_S$ (left panel) and on the center-of-mass energy (right panel) of the total cross section for the central {\it inclusive} double $Z$ production in proton-proton collisions for different numbers of extra dimensions.}
\label{fig:incpp}
\end{center}
\end{figure}
In Fig.~\ref{fig:incpp} we present our predictions for the central inclusive production of a $Z$ boson pair, $p + p \rightarrow X \otimes ZZ \otimes Y$, at LHC energies obtained from Eq.~(\ref{sigma}) assuming that $|y| \leq 2$ and $M_{ZZ} \le M_S - \unit[10]{GeV}$. Moreover, initially we assume a fixed value for the rapidity gap size on either side of the $Z$ boson pair: $\Delta \eta = 2$. As in the exclusive case, the cross section decreases at larger values of $M_S$ and $n$. However, the magnitude is a factor $10^3$ larger, which becomes feasible its experimental study. In particular, we predict values of the order of fb at $\sqrt{s} = \unit[14]{TeV}$, $M_S = 3.5$ TeV and $n = 2$. In Fig.~\ref{fig:incpp2} we present the dependence of our predictions on the rapidity size $\Delta \eta$. The cross section decreases by a factor $2.5$ when $\Delta \eta$ increases from 2.0 to 3.0. Consequently, our predictions are not strongly modified by the choice of $\Delta \eta$.
\begin{figure}[t]
\begin{center}
\begin{tabular}{cc}
\includegraphics*[width=6cm]{zz_incl_ms-eta.eps} & \includegraphics*[width=6cm]{zz_incl_sqrts-eta_REV.eps} \\
\end{tabular}
\caption{(color online) Dependence on $M_S$ (left panel) and on the center-of-mass energy (right panel) of the total cross section for the central {\it inclusive} double $Z$ production in proton-proton collisions for different values of the rapidity gap size $\Delta \eta$.}
\label{fig:incpp2}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics*[width=8cm]{dsdmi.eps}
\caption{(color online) $Z$ pair invariant mass $M_{ZZ}$ distribution of the central inclusive cross section in $pp$ collisions at $\sqrt{s} = 14$ TeV. The SM prediction also is presented for comparison.}
\label{fig:mass}
\end{center}
\end{figure}
In Fig. \ref{fig:mass} we present our predictions for the $Z$ pair invariant mass $M_{ZZ}$ distribution of the central inclusive cross section in $pp$ collisions at $\sqrt{s} = 14$ TeV.
For comparison we also present the diffractive SM prediction which consider the next-to-leading-order $gg \rightarrow ZZ$ subprocess via a quark loop diagram \cite{zzsm}. The LED predictions dominate the distribution at larger values of the invariant mass, as expected from previous studies for the inclusive production of different final states \cite{calculos_nlo}. In particular, we obtain that the LED prediction for the total cross section in central inclusive processes is one order of magnitude larger than the SM one for $M_S = 3.5$ TeV and $n = 2$.
Let us now present our predictions for the double $Z$ production in $pPb$ and $PbPb$ collisions at LHC energies. In Fig.~\ref{fig:nuclear} we present our predictions for the energy dependence of the central exclusive (left panel) and inclusive (right panel) cross section assuming $n = 2$, $M_S = \unit[3.5]{TeV}$ and $\Delta \eta = 2$. For comparison the $pp$ predictions also are shown. We have verified that the $M_S$, $n$, $\Delta \eta$ and $M_{ZZ}$ dependences for the nuclear cross sections are very similar to the $pp$ one. In contrast, the magnitude of the cross sections are very different for $pp$, $pPb$ and $PbPb$ collisions. In particular, for central exclusive processes the $PbPb$ cross section is amplified by a factor $\approx 10^4$ in comparison to the $pp$ one. It is directly associated to the strong dependence of the cross section on the nuclear gluon distribution, $\sigma^{A_1A_2}_{excl} \propto [xg_{A_1} \times xg_{A_2}]^2$, where $x g_{A_i}= A_i R_g^{A_i}\cdot x g_p$. Consequently, if we consider
our
conservative model for the nuclear survival probability $\langle {\mathcal S}^2_\mathrm{excl} \rangle_{A_1A_2}$ we obtain that $\sigma^{A_1A_2}_{excl} \propto A_1 R_g^{A_1} \times A_2 R_g^{A_2} \times \sigma^{pp}_{excl}$, with $R_g > 1$ in the kinematical range probed by the double $Z$ production. On the other hand, in central inclusive processes the difference between the predictions is smaller, since in this case $\sigma^{A_1A_2}_{incl} \propto R_g^{A_1} \cdot R_g^{A_2} \times \sigma^{pp}_{incl}$. It implies that the predictions for the central exclusive and inclusive productions are of the same order in $PbPb$ collisions. For $pPb$ collisions, the central inclusive production dominates.
\begin{figure}[t]
\begin{center}
\begin{tabular}{cc}
\includegraphics*[width=6cm]{zzexclnuclear_REV.eps} & \includegraphics*[width=6cm]{zzinclnuclear_REV.eps} \\
\end{tabular}
\caption{Double $Z$ production for central exclusive (left panel) and inclusive (right panel) processes in $pp/pPb/PbPb$ collisions.}
\label{fig:nuclear}
\end{center}
\end{figure}
\begin{table}
\begin{center}
\begin{tabular} {||c|c|c||}
\hline
\hline
& {\bf CEP} & {\bf CIP} \\
\hline
\hline
$pp$ ($\sqrt{s} = \unit[14]{TeV}$) & $8.0 \times 10^{-5}$ & $19$ \\
\hline
$pPb$ ($\sqrt{s} = \unit[8.8]{TeV}$) & $2.7 \times 10^{-8}$ & $3.1 \times 10^{-5}$ \\
& $1.8 \times 10^{-5}$ & 0.02 \\
\hline
$PbPb$ ($\sqrt{s} = \unit[5.5]{TeV}$) & $2 \times 10^{-8}$ & $0.1 \times 10^{-6}$ \\
\hline
\hline
\end{tabular}
\end{center}
\caption{ The events rate/year for the double $Z$ production in $pp/pPb/PbPb$ collisions at LHC energies considering the CEP and CIP mechanisms. We assume $n=2$, $M_S = \unit[3.5]{TeV}$ and $\Delta \eta$ = 2. }
\label{tabZZ}
\end{table}
Finally, now we compute the production rates for LHC energies considering the distinct mechanisms. We assume $n=2$, $M_S = \unit[3.5]{TeV}$ and $\Delta \eta$ = 2. The results are presented in Table~\ref{tabZZ}. At LHC we assume the design luminosities ${\cal L} = 10^7 /\, 150 /\, \unit[0.5]{mb^{-1}s^{-1}}$ for $pp/pPb/PbPb$ collisions at $\sqrt{s} = 14/\,8.8/\,\unit[5.5]{TeV}$ and a run time of $\unit[10^7 \, (10^6)]{s}$ for collisions with protons (ions).
Moreover, we also consider the upgraded $pPb$ scenario proposed in Ref.~\cite{david}, which analyse a potential path to improve the $pPb$ luminosity and the running time. These authors proposed the following scenario for $pPb$ collisions: ${\cal L} = \unit[10^4]{mb^{-1}s^{-1}}$ and a run time of $\unit[10^7]{s}$. The corresponding event rates are presented in the third line of the Table~\ref{tabZZ}.
Our results indicate that for the default settings and running times, the statistics are marginal for $PbPb$ collisions. Consequently, the possibility to carry out a measurement of the double $Z$ production in diffractive interactions is virtually null in these collisions. On the other hand, in $pp$ collisions the event rates for the central inclusive production predicted by the LED scenario is reasonable, with the SM background being one order of magnitude smaller. In comparison to the results for the inclusive double $Z$ production presented in Fig.~\ref{fig1} and Ref.~\cite{gao}, our predictions for diffractive production are a factor $\approx 10^{4}$ smaller. Despite their much smaller cross sections, the clean topology of the diffractive production implies a larger signal to background ratio. Therefore, the experimental detection is in principle feasible. However, the signal is expected to be reduced due to the event pileup which eliminates one of the main advantages of the diffractive processes.
In contrast, in $pA$ collisions it is expected to trigger on and carry out the measurement with almost no pileup~\cite{david}. Therefore, the upgraded $pA$ scenario also provides one possibility to detect the double $Z$ production in central inclusive processes.
\section{Summary \label{sec:sum}}
The search for new physics at the TeV-scale is one of the major tasks for the current and future high-energy physics experiments. In particular, models with extra spatial dimensions and TeV scale gravity are expected to provide a plethora of new and interesting signals. In this paper the double $Z$ production in diffractive processes was considered for the first time. As in the Standard Model the $Z$ boson pair is not produced at leading order by the gluon fusion, this final state cannot be produced at leading order in diffractive processes, which implies that a very low QCD background is expected. We have estimated the cross sections and event rates for the double $Z$ production in central exclusive and central inclusive processes considering $pp/pPb/PbPb$ collisions. Both processes are characterized by the presence of two rapidity gaps in the final state, differing by the dissociation or not of the incident hadrons. Our results indicate that the LED predictions for the central inclusive production are larger than the SM one and that
the experimental analysis of this process in $pp$ collisions is feasible at CERN LHC. This conclusion motivates a more detailed analysis of the gap survival probability for the double $Z$ production, which is one of the main sources of uncertainty in our calculations, as well as to implement our calculations in a Monte Carlo simulation for diffractive processes. Both subjects will be addressed in the future.
\section*{Acknowledgements}
This work was partially financed by the Brazilian funding agencies CNPq, CAPES and FAPERGS.
|
\section{Introduction}
\vskip-0.05in
It's well known that the Reissner-Nordstr\"om metric describes the gravitational field due to a charged mass into an empty background. Despite the importance of this result in the context of general relativity, one must consider the more realistic case where the charged mass is imbedded into a cosmological background. Therefore, is worth to study the properties of the spacetime of a mass charged into a non-empty universe.
McVittie\cite{mcvittie} was the first one who provided a metric for a Schwarzschild mass immersed into a cosmological background. By using isotropic coordinates, McVittie matched the Friedmann-Lemaitre-Robertson-Walker metric with the Schwarzschild metric. From a different point of view, Einstein and Straus\cite{straus} studied the problem, by cutting a spherical region in a cosmological background. In that `hole', they immersed a Schwarzschild mass and discussed the conditions to join both metrics, however they did not provide any specific metric. Considering a spatially flat Robertson-Walker cosmological fluid with pressure, Bona and Stela built their ``Swiss cheese'' model inspired in the early McVittie metric.
By introducing charged black holes, Kastor and Traschen\cite{kastor} found a solution which describes the dynamics of a set of extremal charged black holes immersed in a de Sitter universe. More recently by using isotropic coordinates, Gao and Zhang\cite{gao} extended the McVittie solution considering a Reissner-Nordstr\"om black hole. Posada and Batic\cite{posada} studied the electrovac universe with a cosmological constant.
In two seminal papers, Gautreau\cite{gaut1, gaut2} developed a method for imbedding a Schwarzschild mass into a cosmological background, by working with curvature coordinates $(R,T)$ and geodesic coordinates $(R,\tau)$. The Gautreau's method is based on a splitting of the spacetime in two regions, where the region inside some radius $R_{b}$ is composed of a part of cosmological fluid plus the Schwarzschild mass. Outside $R_{b}$ the energy-momentum tensor is described by a radially moving geodesic cosmological fluid. Gautreau provided examples of his imbedding method for different cosmological scenarios. An important conclusion from his work, is that planetary orbits will spiral when a Schwarzschild mass is immersed into a general universe.
In this report, we extend the Gautreau imbedding model, by considering a Reissner-Nordstr\"om charged mass. In section II we summarize the relevant expressions from references\cite{gaut1,gaut2}, particularly the metric forms and Einstein equations. The main difference will be in the energy-momentum tensor outside the region $R_{b}$ where we now have an additional contribution due to the electromagnetic tensor. In section III, we apply our results to immerse a charged mass into a de Sitter universe. In section IV we apply the Gautreau method to describe a Reissner-Nordstr\"om charged mass into an Einstein-de Sitter universe, which consists of a pressureless spacetime with $\Lambda=0$. Finally we immerse a Reissner-Nordstr\"om metric into a general universe.
\section{Einstein equations for an imbedded mass}
In the following, we will summarize the relevant equations from references\cite{gaut1,gaut2}. In curvature coordinates $(R,T)$, the most general spherically symmetric line element that one can write is
\begin{equation}\label{curvature}
ds^2(R,T)=-A(R,T)f^{2}(R,T)dT^2+\frac{dR^2}{A(R,T)}+R^2d\Omega^{2}
\end{equation}
\noindent The physical meaning of the coordinate time $T$, is that it's recorded by clocks located at points with $R =$ const. On the other hand, one can describe the spacetime by using geodesic coordinates $(R,\tau)$ where the geodesic time $\tau$ is recorded by clocks moving along geodesics. The transformation between $T$ and $\tau$ is given by
\begin{equation}\label{trans1}
\tau_{,_{R}}=-\frac{m}{A}(k^2-A)^{1/2}
\end{equation}
\begin{equation}\label{trans2}
\tau_{,_{T}}=kf
\end{equation}
\noindent where $m=+1$ or $-1$ depending, respectively, whether the particle moves increasing or decreasing $R$. The parameter $k$ is associated to the energy per unit of mass of the particle moving along a geodesic. In terms of the coordinates $(R,\tau)$, the metric (\ref{curvature}) can be written as
\begin{equation}\label{geodesic}
ds^2(R,\tau)=-d\tau^2+\frac{1}{k^2}\left[dR-m(k^2-A)^{1/2}d\tau\right]^2+R^2d\Omega^2
\end{equation}
\noindent We want to solve the Einstein equation
\begin{equation}\label{einstein}
R_{\nu}^{\mu}-\frac{1}{2}(R+2\Lambda)\delta_{\nu}^{\mu}=-8\pi T_{\nu}^{\mu}
\end{equation}
\noindent where we are using geometrized units $G=c=1$. From the metric (\ref{curvature}), the relevant Einstein's equations at first order in the metric components are
\begin{equation}\label{einstein1}
[R(1-A)]_{,_{R}}=-8\pi R^{2}T_{0}^{0}+R^2\Lambda
\end{equation}
\begin{equation}\label{einstein2}
[R(1-A)]_{,_{R}}-2AR\left(\frac{f_{,_{R}}}{f}\right)=-8\pi R^{2}T_{1}^{1}+R^2\Lambda
\end{equation}
\begin{equation}\label{einstein3}
[R(1-A)]_{,_{T}}=8\pi R^2T_{0}^{1}
\end{equation}
\noindent Note that these equations are valid in general for any energy-momentum tensor, which we will specify later. Given the set of equations (\ref{einstein1})-(\ref{einstein3}), Gautreau proposes to solve them by recognizing $T_{\nu}^{\mu}$ as a function of the curvature coordinates $(R,T)$, therefore they can be integrated. From (\ref{einstein1}) we can obtain then
\begin{equation}\label{int1}
A(R,T)=1+\left(\frac{8\pi}{R}\right)\int_{0}^{R}T_{0}^{0}R^2dR-\left(\frac{\Lambda}{3}\right)R^{2}
\end{equation}
\noindent from (\ref{einstein2}) by using (\ref{einstein1}) we can obtain for $f(R,T)$
\begin{equation}\label{int2}
\ln f^{2}(R,T)=8\pi\int_{0}^{R}\left(\frac{R}{A}\right)(T_{1}^{1}-T_{0}^{0})dR
\end{equation}
\noindent The component $T_{0}^{1}$ can be obtained easily from (\ref{einstein3})
\begin{equation}\label{int3}
T_{0}^{1}=-\frac{1}{R^2}\int_{0}^{R}T_{0_{,T}}^{0}R^2dR
\end{equation}
Instead of working with curvature coordinates, Gautreau\cite{gaut3, gaut4} showed that is advantageous to change to geodesic coordinates, which provide new features of the spacetime important in cosmology, which are unrecognized in the curvature coordinates formalism. Using the transformation (\ref{trans1}) and (\ref{trans2}), the Einstein equations (\ref{einstein1})-(\ref{einstein3}) takes the form
\begin{equation}\label{field1}
\frac{\partial}{\partial R}\left[R(1-A)\right]=-8\pi R^2\tau_{0}^{0}+R^2\Lambda
\end{equation}
\begin{equation}\label{field2}
\frac{\partial}{\partial\tau}\left[R(1-A)\right]=8\pi R^2\tau_{0}^{1}
\end{equation}
\begin{eqnarray}\label{field3}
&\frac{\partial}{\partial R}\left[R(1-A)\right]+m(k^2-A)^{-1/2}\frac{\partial}{\partial\tau}\left[R(1-A)\right]\nonumber\\
&+2m\left(\frac{RA}{k}\right)(k^2-A)^{-1/2}\left[\frac{\partial k}{\partial\tau}+m(k^2-A)^{1/2}\frac{\partial k}{\partial R}\right]\nonumber\\
&=-8\pi R^{2}\tau_{1}^{1}+R^2\Lambda
\end{eqnarray}
\noindent where we use $\tau_{\nu}^{\mu}$ to denote the energy-momentum tensor in coordinates $(R,\tau)$. Following the notation of reference \cite{gaut2}, we use partial derivatives explicitly when we use geodesic coordinates $(R,\tau)$, to differentiate from curvature coordinates $(R,T)$ where we use commas to denote derivatives. Using the same procedure as in (\ref{int1})-(\ref{int3}), we obtain the following
\begin{equation}\label{geod1}
A(R,\tau)=1+\frac{8\pi}{R}\int_{0}^{R}R^2\tau_{0}^{0}-\frac{\Lambda}{3}R^2
\end{equation}
\begin{equation}\label{geod22}
\tau_{0}^{1}=-\frac{1}{R^{2}}\int_{0}^{R}\frac{\partial\tau_{0}^{0}}{\partial\tau}R^{2}dR
\end{equation}
\begin{eqnarray}\label{geod3}
&-m2A\left[\frac{\partial k}{\partial\tau}+m(k^2-A)^{1/2}\frac{\partial k}{\partial R}\right]=8\pi kR(k^2-A)^{1/2}\tau_{1}^{1}\nonumber\\
&-8\pi kR(k^2-A)^{1/2}\tau_{0}^{0}-m\left(\frac{8\pi k}{R}\right)\int_{0}^{R}\frac{\partial\tau_{0}^{0}}{\partial R}R^2dR
\end{eqnarray}
Following the Gautreau's method, we break the spacetime in two regions delimited by a radius $R_{b}$. The inner region $R\leq R_{b}$ will be composed of two parts: a Reissner-Nordstr\"om charged mass (RN) and a perfect cosmological fluid
\begin{equation}\label{splitin}
\tau_{\nu}^{\mu}|_{R\leq R_{b}}=\tau_{\nu}^{\mu(RN)}+\tau_{\nu}^{\mu(cosm)}
\end{equation}
\noindent In the region $R>R_{b}$ we will have the contribution of a perfect cosmological fluid plus an electromagnetic field (em)
\begin{equation}\label{splitout}
\tau_{\nu}^{\mu}|_{R>R_{b}}=\tau_{\nu}^{\mu(cosm)}+\tau_{\nu}^{\mu(em)}
\end{equation}
\noindent The cosmological fluid is described by
\begin{equation}\label{fluid}
\tau_{\mu\nu}^{cosm}=(\rho + p)U_{\mu}U_{\nu}+pg_{\mu\nu}
\end{equation}
\noindent where the particles are moving along radial geodesics. The non-zero components of (\ref{fluid}) are (Eqs. (2.9) reference \cite{gaut2})
\begin{equation}\label{fluid0}
\tau_{0}^{0}=-\rho
\end{equation}
\begin{equation}\label{fluid1}
\tau_{1}^{1}=p
\end{equation}
\begin{equation}\label{fluid2}
\tau_{0}^{1}=-m(k^2-A)^{1/2}(\rho+p)
\end{equation}
\begin{equation}\label{fluid3}
\tau_{2}^{2}=\tau_{3}^{3}=p
\end{equation}
\noindent In our description, we have an additional contribution due to the charge of the particle, which will be determined by the electromagnetic tensor
\begin{equation}\label{electro}
\tau^{\mu\nu(em)}=\frac{1}{4\pi}\left(F^{\mu\alpha}F_{\alpha}^{\phantom{a}\nu}+\frac{1}{4}g^{\mu\nu}F_{\alpha\beta}F^{\alpha\beta}\right)
\end{equation}
\noindent Consistent with the symmetry of the problem\cite{ray}, we consider only electrostatic fields, i.e., only $F^{01}\neq 0$, with the rest of components vanishing\footnote{The Maxwell equations $\nabla_{\mu}F^{\nu\mu}=4\pi J^{\nu}$ and $\nabla_{[\mu}F_{\nu\lambda]}=0$ also admits solutions with $F^{23}\neq 0$, which would indicate a magnetic monopole. However, we don't consider this situation in our treatment.}. From the relation $F_{\mu\nu}=A_{\nu_{,\mu}}-A_{\mu_{,\nu}}$ we have
\begin{equation}\label{electro1}
F_{01}=-F_{10}=-\phi_{,_{R}}
\end{equation}
\noindent From (\ref{geodesic}), (\ref{electro}) and (\ref{electro1}), the relevant non-zero components of $\tau_{\nu}^{\mu(em)}$ are
\begin{equation}\label{em0}
\tau_{0}^{0}=\tau_{1}^{1}=\frac{k^2}{8\pi}(\phi_{,_{R}})^2
\end{equation}
A worth point to mention, is that the energy parameter $k(R,\tau)$ appears explicitly in the electromagnetic tensor components. Following the procedure described in \cite{gaut2}, the field equations (\ref{geod1}) - (\ref{geod3}) in the region $R>R_{b}$ are
\begin{eqnarray}\label{important1}
A(R,\tau)&=1+\frac{8\pi}{R}\int_{0}^{R_{b}}\tau_{0}^{0}R^2dR-\frac{8\pi}{R}\int_{R_{b}}^{R}\rho R^2dR\nonumber\\
&+\frac{k^2}{R}\int_{R_{b}}^{R}(\phi_{,_{R}})^2R^2dR-\frac{\Lambda}{3}R^{2}
\end{eqnarray}
\begin{eqnarray}\label{important2}
\tau_{0}^{1}&=-\frac{1}{R^2}\int_{0}^{R_{b}}\left(\frac{\partial\tau_{0}^{0}}{\partial\tau}\right)R^2dR+\frac{1}{R^2}\int_{R_{b}}^{R}\left(\frac{\partial\rho}{\partial\tau}\right)R^2dR\nonumber\\
&-\frac{k^2}{8\pi R^2}\int_{R_{b}}^{R}(\phi_{,_{R}})^2R^2dR
\end{eqnarray}
\begin{eqnarray}\label{important3}
&\rho+p-\frac{m}{R^2}(k^2-A)^{-1/2}\left[\int_{0}^{R_{b}}\left(\frac{\partial\tau_{0}^{0}}{\partial\tau}\right)R^2dR-\int_{R_{b}}^{R}\left(\frac{\partial\rho}{\partial\tau}\right)R^2dR\right]\nonumber\\
&-\frac{m}{R^2}(k^2-A)^{-1/2}\left(\frac{k^2}{8\pi}\int_{R_{b}}^{R}(\phi_{,_{R}})^2R^2dR\right)=0
\end{eqnarray}
Now, we complete our model with the description of the energy-momentum tensor inside $R_{b}$. One contribution comes from $\tau_{0}^{0}=-\rho$, which is the cosmological fluid, and the other contribution is just the imbedded spherical mass M. In the region $R<R_{b}$ we have then
\begin{equation}\label{inner}
\tau_{0}^{0}=\tau_{0}^{0(cosm)}+M_{0}^{0}=-\rho+M_{0}^{0}
\end{equation}
\noindent where the spherical mass M is defined as
\begin{equation}\label{mass}
M=-4\pi\int_{0}^{R_{b}}M_{0}^{0}R^2dR
\end{equation}
\noindent Note that the integral that involves the electric potential $\phi$, can be rewritten in terms of the charge distribution. In that order, we recall that the potential due to a spherical charge is given by $\phi=-\frac{1}{4\pi}\frac{q}{R}$. Using this relation jointly with (\ref{inner}) and (\ref{mass}), Eqs (\ref{important1})-(\ref{important3}) becomes
\begin{equation}\label{imbed1}
A(R,\tau)=1-\frac{2M}{R}+k^2\frac{Q^2}{R^2}-\frac{8\pi}{R}\int_{0}^{R}\rho R^2dR-\frac{\Lambda}{3}R^{2}
\end{equation}
\begin{equation}\label{imbed2}
\tau_{0}^{1}=\frac{1}{R^2}\int_{0}^{R}\left(\frac{\partial\rho}{\partial\tau}\right)R^2dR-\frac{k^2}{8\pi}\frac{Q^2}{R^3}
\end{equation}
\begin{equation}\label{imbed3}
\rho+p+\frac{m}{R^2}(k^2-A)^{-1/2}\left[\int_{0}^{R}\left(\frac{\partial\rho}{\partial\tau}\right)R^2dR-\frac{k^2}{8\pi}\frac{Q^2}{R}\right]=0
\end{equation}
\noindent where we have defined $q\equiv 4\pi Q$. In the next sections, we will apply our results for the case of imbedding a charged mass into different cosmological scenarios.
\section{A Reissner-Nordstr\"om mass imbedded into a de Sitter universe}
A de Sitter spacetime corresponds to a cosmology with no matter, where the dynamics of the universe is driven by a cosmological constant. In this scenario we have $\rho=0$, such that from (\ref{imbed2}) and (\ref{imbed3}) we have
\begin{equation}\label{sitter1}
\tau_{0}^{1}=-\frac{k^2}{2R}\left(\frac{Q^2}{4\pi R^2}\right)
\end{equation}
\begin{equation}\label{sitter2}
p=\frac{mk^2}{2R}(k^2-A)^{-1/2}\left(\frac{Q^2}{4\pi R^2}\right)
\end{equation}
\noindent where (\ref{sitter1}) can be interpreted as flux of energy (due to the electric field) through a surface $R = $const. We also have a non-zero pressure, which can be associated to the electric field produced by the charge. Note that if we have $Q=0$, we recover the results in \cite{gaut2} as must be expected. From (\ref{imbed1}) we have
\begin{equation}\label{asitter}
A(R,\tau)=1-\frac{2M}{R}+k^2\frac{Q^2}{R^2}-\frac{\Lambda}{3}R^{2}
\end{equation}
\noindent and the metric (\ref{geodesic}) in geodesic coordinates $(R,\tau)$ gives
\begin{eqnarray}\label{geositter}
&ds^2(R,\tau)=-d\tau^2+R^2d\Omega^2+\nonumber\\
&\frac{1}{k^2}\left[dR-m\left(\frac{2M}{R}-k^2\frac{Q^2}{R^2}+\frac{\Lambda}{3}R^{2}+k^2-1\right)^{1/2}d\tau\right]^2
\end{eqnarray}
\noindent The transformation relations (\ref{trans1}) and (\ref{trans2}) takes the form
\begin{equation}\label{transitter1}
\tau_{,_{R}}=-m\frac{\left(\frac{2M}{R}-k^2\frac{Q^2}{R^2}+\frac{\Lambda}{3}R^{2}+k^2-1\right)^{1/2}}{1-\frac{2M}{R}+k^2\frac{Q^2}{R^2}-\frac{\Lambda}{3}R^{2}}
\end{equation}
\begin{equation}\label{transitter2}
\tau_{,_{T}}=k
\end{equation}
\noindent and the metric (\ref{curvature}) in curvature coordinates $(R,T)$ gives
\begin{eqnarray}\label{curvasitter}
&ds^2(R,T)=-\left(1-\frac{2M}{R}+k^2\frac{Q^2}{R^2}-\frac{\Lambda}{3}R^{2}\right)dT^2\nonumber\\
&+\left(1-\frac{2M}{R}+k^2\frac{Q^2}{R^2}-\frac{\Lambda}{3}R^{2}\right)^{-1}dR^2+R^2d\Omega^2
\end{eqnarray}
If we set $\Lambda=0$ we have the Reissner-Nordstr\"om metric, while if we set $Q=0$ we recover the Kottler metric\cite{dumin}. Therefore, the metric form (\ref{curvasitter}) can be recognized as the description of a Reissner-Nordstr\"om charged mass imbedded into a de Sitter universe. Note that if we consider the situation $Q=0$ we recover the results in \cite{gaut2}, as is expected. Note also, that the energy parameter $k(R,\tau)$ appears explicitly in \ref{curvasitter}, because we don't have a cosmological fluid to attach clocks. Once we introduce a cosmological fluid, we can consider $k=1$ even though we don't have a homogeneous situation (see argument in \cite{gaut2}). We discuss this situation in the next sections.
\section{Reissner-Nordstr\"om charged mass imbedded into an Einstein-de Sitter universe}
An Einstein-de Sitter universe corresponds to a scenario with $\Lambda=p=0$. Under this condition, (\ref{imbed3}) gives
\begin{equation}\label{eds1}
\rho+\frac{m}{R^2}(1-A)^{-1/2}\left[\int_{0}^{R}\left(\frac{\partial\rho}{\partial\tau}\right)R^2dR-\frac{1}{8\pi}\frac{Q^2}{R}\right]=0
\end{equation}
\noindent where we have set $k=1$ considering that we have a cosmological fluid to attach clocks. After substituting (\ref{imbed1}) in (\ref{eds1}) we obtain
\begin{eqnarray}\label{eds2}
&\left[\left(\frac{-m}{\rho R^2}\right)\left(\int_{0}^{R}(\frac{\partial\rho}{\partial\tau})R^2dR-\frac{1}{8\pi}\frac{Q^2}{R}\right)\right]^2=\frac{8\pi}{R}\int_{0}^{R}\rho R^2dR\nonumber\\
&+\frac{2M}{R}-\frac{Q^2}{R^2}
\end{eqnarray}
\noindent Following \cite{gaut2} we introduce the definition
\begin{equation}\label{zeta}
Z(R,\tau)\equiv 8\pi\int_{0}^{R}\rho R^2dR.
\end{equation}
\noindent Using (\ref{zeta}) in (\ref{eds2}) we have
\begin{equation}\label{eds3}
R^{1/2}\left(\frac{\partial Z}{\partial\tau}-\frac{Q^2}{R}\right)+m\left[Z+\left(2M-\frac{Q^2}{R}\right)\right]^{1/2}\frac{\partial Z}{\partial R}=0
\end{equation}
\noindent and (\ref{imbed1}) takes the form
\begin{equation}\label{eds4}
A(R,\tau)=1-\left(\frac{2M+Z}{R}\right)+\frac{Q^2}{R^2}
\end{equation}
\noindent Note that if we set $Q=0$, results (\ref{eds2})-(\ref{eds4}) reduces to those in \cite{gaut2}. The metric (\ref{geodesic}) in coordinates $(R,\tau)$ takes the form
\begin{eqnarray}\label{geoeds}
ds^2(R,\tau)&=-d\tau^2+\left[dR-m\left(\frac{2M+Z}{R}-\frac{Q^2}{R^2}\right)^{1/2}d\tau\right]^2\nonumber\\
&+R^2d\Omega^2
\end{eqnarray}
The transformation (\ref{trans1}) between the time coordinate $T$ and the geodesic time $\tau$ becomes
\begin{equation}\label{transeds}
\tau_{,_{R}}=-m\frac{\left[\left(\frac{2M+Z}{R}\right)-\frac{Q^2}{R^2}\right]^{1/2}}{1-\left(\frac{2M+Z}{R}\right)+\frac{Q^2}{R^2}}
\end{equation}
\noindent when $R\to 0$ (\ref{transeds}) must satisfy $\tau_{,_{T}}\to 1$ (flatness condition). Finally, the metric (\ref{curvature}) in curvature coordinates takes the form
\begin{eqnarray}\label{curvaeds}
&ds^2(R,T)=-\left[1-\left(\frac{2M+Z}{R}\right)+\frac{Q^2}{R^2}\right]f^2dT^2\nonumber\\
&+\left[1-\left(\frac{2M+Z}{R}\right)+\frac{Q^2}{R^2}\right]^{-1}dR^2+R^2d\Omega^2
\end{eqnarray}
\noindent where $f(R,\tau)$ will be determined from the solution to (\ref{transeds}). In the Einstein-de Sitter universe the density is given by
\begin{equation}\label{rhoeds}
\rho=\frac{1}{6\pi\tau^2}
\end{equation}
\noindent Note that if we substitute (\ref{rhoeds}) in (\ref{zeta}) and we integrate it, after substitution in (\ref{curvaeds}) we obtain the Einstein- de Sitter universe in curvature coordinates \cite{gaut1}. That particular scenario corresponds then to a Reissner-Nordstr\"om charged mass imbedded into an Einstein-de Sitter universe. In the next section we consider a general cosmology.
\section{Imbedding a Reissner-Nordstr\"om charged mass into a general cosmology}
We can extend our methods above, to consider the case of a charged mass imbedded into a general universe with $\rho\neq 0$ and $p\neq 0$. If we substitute (\ref{imbed1}) in (\ref{imbed3}) we obtain
\begin{eqnarray}\label{general1}
&\left[\frac{-m}{R^2}\frac{1}{(\rho+p)}\left(\int_{0}^{R}(\frac{\partial\rho}{\partial\tau})R^2dR-\frac{1}{8\pi}\frac{Q^2}{R}\right)\right]^2=\frac{8\pi}{R}\int_{0}^{R}\rho R^2dR\nonumber\\
&+\frac{2M}{R}-\frac{Q^2}{R^2}+\frac{\Lambda}{3}R^2
\end{eqnarray}
\noindent which provides a relation between density and pressure. Once we know the state equation $p=p(\rho)$, we can solve (\ref{general1}) to find $\rho(R,\tau)$. Using (\ref{zeta}) and (\ref{imbed1}), the geodesic metric (\ref{geodesic}) takes the form
\begin{eqnarray}\label{gengeo}
ds^2(R,\tau)&=\left[dR-m\left(\frac{2M+Z}{R}-\frac{Q^2}{R^2}+\frac{\Lambda}{3}R^2\right)^{1/2}d\tau\right]^2\nonumber\\
&-d\tau^2+R^2d\Omega^2
\end{eqnarray}
From (\ref{imbed2}) the flux of energy through a surface $R = $ const., gives
\begin{equation}
\tau_{0}^{1}=\frac{1}{R^2}\int_{0}^{R}\left(\frac{\partial\rho}{\partial\tau}\right)R^2dR-\frac{Q^2}{8\pi R^3}
\end{equation}
\noindent which corresponds to the energy due to the electric field (Eq. \ref{sitter1}), plus a contribution of the cosmological fluid. The transformation (\ref{trans1}) between curvature coordinates $(R,T)$ and geodesic coordinates $(R,\tau)$ takes the form
\begin{equation}\label{transgen}
\tau_{,_{R}}=-m\frac{\left[\left(\frac{2M+Z}{R}\right)-\frac{Q^2}{R^2}+\frac{\Lambda}{3}R^2\right]^{1/2}}{1-\left(\frac{2M+Z}{R}\right)+\frac{Q^2}{R^2}-\frac{\Lambda}{3}R^2}
\end{equation}
\noindent which must satisfy $\tau_{,_{T}}\to 1$ for small R. Finally, the metric for a Reissner-Nordstr\"om mass imbedded into a general universe in curvature coordinates $(R,T)$ is
\begin{eqnarray}\label{curvagen}
&ds^2(R,T)=-\left[1-\left(\frac{2M+Z}{R}\right)+\frac{Q^2}{R^2}-\frac{\Lambda}{3}R^2\right]f^2dT^2\nonumber\\
&+\left[1-\left(\frac{2M+Z}{R}\right)+\frac{Q^2}{R^2}-\frac{\Lambda}{3}R^2\right]^{-1}dR^2+R^2d\Omega^2
\end{eqnarray}
\noindent where $f(R,\tau)$ is determined from the solution to (\ref{transgen}). Note that if we set $\Lambda=0$, our result coincides with the metric found by Gao and Zhang (Eq. 32 reference\cite{gao}). If we set $\Lambda=\rho=p=0$ we recover the Reissner-Nordstr\"om solution, while if we set $M=Q=0$ we have a cosmological metric for a general function $p=p(\rho)$. If we set $Q=0$ in (\ref{curvagen}) we recover the results of \cite{gaut2} as must be expected.
\section{Equations of motion in geodesic coordinates}
Now that we have extended the Gautreau's imbedding method considering a Reissner-Nordstr\"om charged mass, our last step will be to provide the geodesic equations. The difference of our approach with the one developed by Gautreau\cite{gaut2}, is that we will use the geodesic metric form (\ref{geodesic}), while Gautreau used the curvature form (\ref{curvature}). The geodesic equations are given by
\begin{equation}\label{eom}
\frac{dV^{\mu}}{ds}+\Gamma_{\rho\sigma}^{\mu}V^{\rho}V^{\sigma}=0
\end{equation}
\noindent where $V^{\mu}=\frac{dx^{\mu}}{ds}$, and we take $s$ as the proper time along a particle's trajectory. From (\ref{geodesic}), the relevant affine connections are\footnote{We used the \emph{ctensor} package available in \emph{Maxima} to calculate the Christoffel symbols. Special thanks to Viktor T. Toth.}
\begin{equation}
\Gamma_{00}^{0}=-\frac{m}{2}(1-A)^{1/2}\frac{\partial A}{\partial R}\,\,\,\, ; \,\,\,\, \Gamma_{10}^{1}=-\Gamma_{00}^{0}\nonumber
\end{equation}
\begin{equation}
\Gamma_{00}^{1}=-\frac{1}{2}\left[A\frac{\partial A}{\partial R}-\frac{m}{(1-A)^{1/2}}\frac{\partial A}{\partial\tau}\right]\,\,\,\, ; \,\,\,\,\Gamma_{10}^{0}=\frac{1}{2}\frac{\partial A}{\partial R}\nonumber
\end{equation}
\begin{equation}\label{christo}
\Gamma_{11}^{0}=-\frac{m}{2}(1-A)^{-1/2}\frac{\partial A}{\partial R}\,\,\,\,\, ; \,\,\,\,\, \Gamma_{11}^{1}=-\Gamma_{10}^{0}
\end{equation}
\begin{equation}
\Gamma_{12}^{2}=\Gamma_{13}^{3}=\frac{1}{R}\,\,\,\, ; \,\,\,\, \Gamma_{23}^{3}=\cot\theta\,\,\,\, ; \,\,\,\, \Gamma_{33}^{2}=-\sin\theta\cos\theta\nonumber
\end{equation}
\begin{equation}
\Gamma_{22}^{0}=mR(1-A)^{1/2}\,\,\,\,\, ; \,\,\,\,\, \Gamma_{33}^{0}=\Gamma_{22}^{0}\sin^2\theta\nonumber
\end{equation}
\begin{equation}
\Gamma_{22}^{1}=-RA\nonumber\,\,\,\,\, ; \,\,\,\,\, \Gamma_{33}^{1}=\Gamma_{22}^{1}\sin^2\theta\nonumber
\end{equation}
\noindent We can specialize to the case $\theta=\frac{\pi}{2}$ such that $V^2=0$. Using (\ref{christo}) in (\ref{eom}) we have
\begin{eqnarray}\label{eom1}
&\frac{dV^0}{ds}-\frac{m}{2}(1-A)^{1/2}\frac{\partial A}{\partial R}\left[(V^{0})^2+(V^{1})^2\right]+\frac{\partial A}{\partial R}(V^{0}V^{1})\nonumber\\
&+mR(1-A)^{1/2}(V^{3})^2=0
\end{eqnarray}
\begin{eqnarray}\label{eom2}
&\frac{dV^1}{ds}+\frac{1}{2}\left[A\frac{\partial A}{\partial R}-\frac{m}{(1-A)^{1/2}}\frac{\partial A}{\partial\tau}\right](V^0)^2-\frac{1}{2}\frac{\partial A}{\partial R}(V^1)^2\nonumber\\
&+\frac{m}{2}(1-A)^{1/2}\frac{\partial A}{\partial R}(V^0V^1)-RA(V^3)^2=0
\end{eqnarray}
\begin{eqnarray}\label{eom3}
\frac{dV^3}{ds}+\frac{2}{R}(V^1V^3)=0
\end{eqnarray}
\begin{eqnarray}\label{eom4}
&-A(V^0)^2-2m(1-A)^{1/2}V^0V^1+(V^1)^2\nonumber\\
&+R^2(V^3)^2=-1
\end{eqnarray}
\noindent where (\ref{eom4}) is just the normalization condition $g_{\mu\nu}V^{\mu}V^{\nu}=-1$. We notice that (\ref{eom3}) can be rewritten as
\begin{equation}\label{angular}
R^2\frac{dV^3}{ds}+2RV^3\frac{dR}{ds}=\frac{d}{ds}(R^2V^3)=0
\end{equation}
\noindent which implies
\begin{equation}\label{angmom}
R^2V^3=R^2\frac{d\phi}{ds}=L=const.
\end{equation}
\noindent where $L$ is the angular momentum per unit of mass. Note that (\ref{angmom}) is just the angular momentum conservation. One important result found by Gautreau by using curvature coordinates\cite{gaut2}, is that planetary orbits will spiral when a Schwarzschild mass is imbedded into a cosmological background. In the following we arrive to the same conclusion, with the difference that we work in geodesic coordinates $(R,\tau)$. Let's assume a planetary circular motion, i.e.,
\begin{equation}\label{circ}
R=const. \Rightarrow \frac{dR}{ds}=0
\end{equation}
\noindent We will show that this condition can't be satisfied if we want to satisfy the EOM (\ref{eom1})-(\ref{eom4}). Therefore $R\neq $const., which implies that there is a spiralling of orbits. Using condition (\ref{circ}) we can write
\begin{equation}\label{total}
\frac{dV^0}{ds}=\frac{dV^0}{d\tau}\frac{d\tau}{ds}=V^0\frac{dV^0}{d\tau}
\end{equation}
\noindent using (\ref{total}) in (\ref{eom1}) we have
\begin{eqnarray}\label{circ1}
&V^0R^2\frac{dV^0}{d\tau}-\frac{m}{2}R^2(1-A)^{1/2}\frac{\partial A}{\partial R}(V^{0})^2\nonumber\\
&+mR(1-A)^{1/2}\left[A(V^0)^2-1\right]=0
\end{eqnarray}
\noindent We can rewrite (\ref{eom4}) in terms of $L$ like
\begin{equation}\label{circ2}
A(V^0)^2=1+\left(\frac{L}{R}\right)^2
\end{equation}
\noindent Using (\ref{circ2}) in (\ref{eom2}) we obtain
\begin{eqnarray}\label{orbit}
\frac{\partial}{\partial R}\left[\ln A(R,\tau)\right]-\frac{m}{A}(1-A)^{-1/2}\frac{\partial}{\partial\tau}\left[\ln A(R,\tau)\right]\nonumber\\
=\frac{2/R}{\left[1+(R/L)^2\right]}
\end{eqnarray}
We started with the condition $R = $ const., which makes the right side of (\ref{orbit}) constant, however if $A$ is function of $\tau$, the left side is not going to be constant. This contradiction leads us to conclude that the condition $R = $ constant, is not consistent. Only for the case of a Reissner-Nordstr\"om-de Sitter spacetime (RNS)(\ref{asitter}), where $A$ is not a function of $\tau$, the orbit radius will be constant. However, in the most general case where $A=A(R,\tau)$ the radius of the orbit will change. Note that for the (RNS) case, the second term to the left of (\ref{orbit}) vanishes, which reduces to the expression found by Gautreau (Eq. (6.12), reference \cite{gaut2}). Therefore, we can consider (\ref{orbit}) as the generalization of the equation for the orbit spiralling, in geodesic coordinates. Once we choose a cosmological model, we can determine $A(R,\tau)$ and use (\ref{orbit}) to find the equation for the radius of the orbit.
In order to provide a magnitude of the spiraling of an orbit, Gautreau makes an estimate using Newtonian physics (Eq. (6.26b) reference \cite{gaut2}). From his calculation, Gautreau concludes that changes in the orbital radius at the scale of our solar system are negligible, but these becomes appreciable at the scale of $R\approx 25 Kpc$. Criticism can be raised about his development. In principle, his Newtonian approximation is unrelated to his imbedding formalism, which provides a new way to consider cosmological theory. A more realistic and consistent calculation, would involve to solve the EOM (\ref{eom1})-(\ref{eom4}) in order to obtain an equation for $\frac{dR}{d\tau}$. We see the high complexity of the equations due to the mixed terms $(V^0V^1)$, which makes a hard task to solve them. Instead of that, we can use the curvature coordinates $(R,T)$ which provides more simple EOM (see Eqs. (6.3)-(6.5) reference \cite{gaut2}). However, the metric component $B(R,T)$ depends on the function $\tau_{,_{T}}$, which is not easy to solve in principle. If I am allowed to speculate here, I suggest that using numerical methods might help to find approximate solutions. This problem will be left for future investigation.
\section{Discussion}
We have extended the Gautreau's imbedding method, by considering a Reissner-Nordstr\"om charged mass. We worked in curvature coordinates $(R,T)$ where $T$ is associated to times recorded by clocks at fixed points $R = $ const. However, we can also work cosmological theory by introducing geodesic coordinates $(R,\tau)$, where $\tau$ is the time recorded by clocks radially moving along geodesics with the cosmological fluid. One important consequence of our description, is that when we immerse a mass into a cosmological background, planetary orbits will spiral. We found that only for the (RNS) field, where the metric component $A$ is independent of $\tau$, the orbit radius will be constant. However, in a more general cosmological scenario, the orbit radius will change.
We raised some criticism to the equation for the orbit change found by Gautreau, considering that his approach involved a Newtonian calculation, which is not consistent with the imbedding formalism developed previously. In order to obtain an equation for the orbital change, which will be more consistent with the formalism developed, it's necessary to solve the EOM (\ref{eom1})-(\ref{eom4}). The complexity of these equations, will require perhaps, to use numerical methods in order to find approximate solutions. That problem is left open for future investigation.
\section{Acknowledgments}
I want to thank to the Department of Physics and Astronomy at the University of South Carolina for its support while this work was developed.
|
\section{Introduction}
Natural Language Processing (NLP) aims at developing techniques for processing natural language texts using computers. In order to yield accurate results, NLP requires resources containing various information (sub-categorization frames, semantic roles, selection restrictions, etc.). Unfortunately, such resources are not available for most languages and are very costly to develop manually. A recent trend of research has tried to overcome these limitations through the development of automatic acquisition methods from corpora.
Automatic lexical acquisition is an engineering task aiming at providing compre\-hensive---even if not fully accurate---resources for NLP. As natural languages are complex, lexical acquisition needs to take into account a wide range of parameters and constraints. However, surprisingly, in the acquisition community, relatively few investigations have been done on the structure of the linguistic constraints themselves, beyond the engineering point of view (but note that this work has been extensively done for parsing, see \cite{aarts08}).
In this paper, we want to take another look at some experiments recently done on the automatic acquisition of lexical resources from textual corpora, more specifically on French. In a way, acquisition is converse to parsing: the task consists, from a surface form, in trying to find an abstract lexical-conceptual structure that justify the surface construction (taking into account the relevant set of constraints for the given language). Here, in order to get a tractable model, we limit ourselves to the acquisition of sub-categorization frames from corpora. The task is challenging since surface forms incorporate adverbs, modifiers, interpolated clauses and some flexibility in the ordering of the arguments.
Most approaches, including ours, are based on simple filtering techniques. If a complement appears very rarely associated with a given predicate, the acquisition process will assume that this is an incidental co-occurrence that should be left out. However, as we will see, even if this technique is efficient for high frequency items, it leaves a lot of phenomena aside.
Following these observations, we get interested in Optimality Theory (OT). OT is based on a number of assumptions which are absolutely relevant for the lexical acquisition context \cite{kager99,mccarthy08,prince04}:
\begin{itemize}
\item Linguistic well-formedness is relative, not absolute. Perfect satisfaction of all linguistic
constraints is attained rarely, and perhaps never.
\item Linguistic well-formedness is a matter of comparison or competition among candidate
output forms (none of which is perfect).
\item Linguistic constraints are ranked and violable. Higher ranking constraints can compel
violation of lower ranking constraints. Violation is minimal, however. And even low ranking
constraints can make crucial decisions about the winning output candidate.
\item The grammar of a language is a ranking of constraints. Ranking may differ from
language to language, even if the constraints do not.
\end{itemize}
However, despite these observations, OT has been mainly applied to phonology, more rarely to morphology or syntax \cite{blache08,aarts08}. In this paper, we would like to show, on a precise example, that OT provides a very competitive framework for sub-categorization acquisition.
In order to apply OT to lexical acquisition, we first need to model all the language properties as constraints. The task consists then in identifying the relevant set of constraints that allow one to map a lexical structure to actual (surface) constructions. Note that the task is highly challenging since constraints interact with each other, must be ranked and can be violated.
\section{From Corpus to Resources}
\subsection{OT and Syntax}
OT has been mainly applied to syntax in the framework of the Principles and Parameters (P\&P) theory developed by Chomsky \cite{chomsky95} as part of his Minimalist Program. The central idea of P\&P is that a person's syntactic knowledge can be modeled with two formal mechanisms:
\begin{itemize}
\item A finite set of fundamental principles that are common to all languages; e.g., a sentence must always have a subject, even if it is not overtly pronounced.
\item A finite set of parameters that determine syntactic variability amongst languages; e.g., a binary parameter that determines whether or not the subject of a sentence must be overtly pronounced.
\end{itemize}
Within this framework, the goal of linguistics is to identify all the principles and parameters that are universal to human languages (i.e. what defines the Universal Grammar).
OT provides a nice framework to implement P\&P since the formalism is constraint-based. The input is a set of (universal) abstract candidate forms\footnote{This point, which is much controversial, is based on the assumption that linguistic principles---in P\&P Theory---are supposed to be universal. There is a huge literature on this hypothesis that we will not address in this paper. We do not claim any universal feature in this work; we just use OT as an interesting framework for modeling the constraints used. }. Thus, principles and parameters just have to be translated into constraints (CON); then an evaluation function (EVAL) computes the best output given the input and the set of constraints (the principles and parameters) for a given language.
To summarize, here are the three main components of OT: GEN (+input), CON and EVAL.
\begin{itemize}
\item GEN takes a series of surface forms and generates an infinite number of candidates, or possible realizations of that input. A language's grammar (its ranking of constraints) determines which of the infinite candidates will be assessed as optimal by EVAL.
\item CON includes the set of constraints to be used to determine which of the input candidates is the most likely to be accepted.
\item EVAL determines the best analysis among input candidates, taking into account the set of constraints CON. Given two candidates, A and B, A is better than B on a constraint hierarchy if A incurs fewer violations than B. Candidate A is better than B on an entire constraint hierarchy if A incurs fewer violations of the highest-ranked constraint distinguishing A and B. A is optimal in its candidate set if it is better on the constraint hierarchy than all other candidates.
\end{itemize}
However, the task here is slightly different (converse) since we try to find the best underlying representation from the output (a given utterance), more precisely, we try to learn syntactic frames from data.
\subsection{Learning Syntactic Frames from Raw Data}
As already said, comprehensive and accurate lexical resources are key components of Natural Language
Processing (NLP) systems.
Hand-crafting lexical resources
is difficult and extremely labour-intensive--- particularly as
NLP systems require statistical information about the behavior
of lexical items in context, and this statistical information changes
from one domain to the other. For this reason automatic acquisition
of lexical resources from corpora has become increasingly popular.
One of the most useful lexical information for NLP is that related
to the predicate-argument structure. The sub-categorization frames (SCFs)
of a predicate capture the different combinations
of arguments that a given predicate can take. For example, in French,
the verb ``\textit{acheter}'' (\textit{to buy}) sub-categorizes for a subject,
a direct object and an indirect object (a prepositional phrase governed
by the preposition \textit{``\`a''}). This can be formalized as follows: \textit{N0 acheter N1 \`a N2}.
Sub-categorization lexicons can benefit many NLP applications.
For example, they can be used to enhance tasks such as parsing
\cite{carroll98,arun05} and semantic classification
\cite{schulte02acl} as well as applications such as information
extraction \cite{surdeanu03} and machine translation.
They also make it possible to infer large multilingual semantic classifications
\cite{Sun2010}.
Several sub-categorization lexicons are available for many languages,
but most of them have been built manually. For French these include
the large French dictionary \emph{``Le Lexique Grammaire''}
\cite{gross75} and the more recent \emph{Lefff} \cite{sagot06} and
\emph{Dicovalence} (\url{http://bach.arts.kuleuven.be/dicovalence/})
lexicons.
Some work has been conducted on automatic sub-categorization acquisition,
mostly on English \cite{brent93,manning93,briscoe97,korhonen06} but
also on other languages, from which German is just one example \cite{schulte02lrec}.
This work has shown that although automatically built lexicons are not as accurate
and detailed as manually built ones, they can be useful for real-world tasks.
This is mostly because they provide what manually built resources
do not generally provide: statistical information about the
likelihood of SCFs for individual verbs.
In what follows, we show that statistical information, in order to yield accurate results, must take into consideration a huge number of constraints. First experiments have given interesting results but the nature and the structure of constraints must be further explored in order to strengthen the existing results. We show that OT provides an interesting framework to identify and structure the set of relevant constraints.
\subsection{Introducing Gradience in Lexical Acquisition}
As for most linguistic questions, there is no well-established definition of what to include in a SCF, but everybody agrees that a SCF should minimally include the number and the type of the complements depending on the verb (or more generally on the predicative item considered, since adjectives and nouns can also have a SCF). Most authors agree on the fact that complements should be divided between arguments and adjuncts but the distinction between these two categories is far from obvious. Some linguistic tests exist (can the complement be deleted without changing the meaning of the sentence? Can it be moved easily? Can it be pronominalized? etc.) but none of these tests is sufficient or discriminatory enough.
As outlined by Manning \cite{MANN2003} ``rather than maintaining a categorical argument~/ adjunct distinction and having to make in/out decisions about such cases, we might instead try to represent SCF information as a probability distribution over argument frames, with different verbal dependents expected to occur with a verb with a certain probability''. For example, from the analysis of a large news corpus, one can observe that the French verb \emph{venir (to come)} accepts the frame \emph{PP[de (from)]} with a relative frequency of 59.1\% whereas it accepts the frame \emph{PP[\`a (to)]} with a relative frequency of 5\%. This phenomenon can be seen as a kind of selectional ``preference'' of certain verbs for certain SCFs; the link with more semantic information remains to be done.
It is well known that the evaluation of probability distributions is difficult, since it is by definition dependent on a given corpus. Hand-crafted dictionaries generally do not include any frequency information.
Moreover, very few lexical acquisition frameworks currently integrate an efficient way to deal with various phenomena such as multiword expressions (especially light verb constructions and semi-idiomatic expressions), complement optionality, etc. Therefore, current approaches have a tendency to produce two many SCFs for a given items (semi-idiomatic expressions should be recognized as such and should not be added as new SCFs associated with head verbs, optionality should be handled to reduce the number of partial SCFs).
In the next section, we briefly present a state-of-the art system for French and its limitations; we show that the acquisition model corresponds to OT but does not take into consideration a precise enough set of constraints. We then make some proposals in order to get better results using a finer grain model of constraints.
\section{ASSCI, A State-of-the Art Subcategorization Acquisition System for French}
\label{assci}
A system for the automatic acquisition of sub-categorization frames has recently been implemented for French.
This system called ASSCI is capable of acquiring large scale lexicons from un-annotated
corpora \cite{messiant08b,messiant-lrec08}.
This system is close to other systems developed for example for English \cite{briscoe97,preiss07} in that it extracts SCFs from data parsed using a shallow dependency parser \cite{bourigault05} and is capable of identifying a large number of SCFs. However, unlike most other systems that accept raw corpus data as input, it does not assume a list of predefined SCFs. The system is based on the assumption that the most relevant SCF corresponding to a given surface form will directly emerge from the application of the constraints on the various candidates, as postulated by OT.
\emph{ASSCI} takes raw corpus data as input. Input text is first tagged and syntactically analyzed.
Then, the system generates a list of candidate SCFs for each verb that occurs frequently enough in
data (in the default setting, 200 occurrences of a given verb are necessary).
\emph{ASSCI} consists of three modules: a pattern extractor
which extracts patterns for each target verb; a SCF builder which builds a list of candidate
SCFs per verb (GEN), and a SCF filter (EVAL) which filters out SCFs deemed incorrect according to predefined parameters (CON). They are described
briefly in the following sections. For a more detailed description of \emph{ASSCI}, see \cite{messiant08b}.
\subsection{Preprocessing : Morphosyntactic Tagging and Syntactic Analysis}
The system first tags and lemmatizes corpus data using \emph{TreeTagger}
and then parses it thanks to \emph{Syntex} \cite{bourigault05}.
\emph{Syntex} is a shallow parser for French. It uses a combination of
heuristics and statistics to find dependency relations between tokens
in a sentence. It is a relatively accurate parser, e.g.~it obtained the
best precision and F-measure for written French text in the first EASY
evaluation campaign (2006).
The below example illustrates the dependency relations detected
by \emph{Syntex} \texttt{(2)} for the input sentence in
\texttt{(1)}:\\
\begin{small}
\texttt{(1) La s\'echeresse s' abattit sur le Sahel en 1972-1973 .\\
(The drought came down on Sahel in 1972-1973.)}\\
\end{small}
\begin{scriptsize}
\texttt{(2) DetFS|le|La|1|DET;2|\\
NomFS|s\'echeresse|s\'echeresse|2|SUJ;4|DET;1 \\
Pro|se|s'|3|REF;4| \\
VCONJS|abattre|abattit|4|SUJ;2,REF;3,PREP;5,PREP;8 \\
Prep|sur|sur|5|PREP;4|NOMPREP;7 \\
DetMS|le|le|6|DET;7| \\
NomMS|sahel|Sahel|7|NOMPREP;5|DET;6 \\
Prep|en|en|8|PREP;4|NOMPREP;9 \\
NomXXDate|1972-1973|1972-1973|9|NOMPREP;8| \\
Typo|.|.|10||}
\end{scriptsize}
\bigskip
\emph{Syntex} does not make a distinction between arguments and adjuncts - rather,
each dependency of a verb is attached to the verb.
\subsection{Producing the Input (the Pattern Extractor)}
The pattern extractor collects the dependencies found by the parser for each
occurrence of a target verb.
Some cases receive special treatment in this module.
For example, if the pronoun \textit{``se''} is one of the dependencies of a verb,
the system considers this verb like a new one.
In \texttt{(1)}, the pattern will correspond to \textit{``s'abattre''} and not to \textit{``abattre''}.
If a preposition is the head of one of the dependencies, the module explores the
syntactic analysis to find if it is followed by a noun phrase (\texttt{+SN]})
or an infinitive verb (\texttt{+SINF]}).
\texttt{(3)} shows the output of the
pattern extractor for the input in \texttt{(1)}.\\
\begin{small}
\texttt{(3) VCONJS|s'abattre~:}
\end{small}
\begin{small}
\texttt{Prep+SN|sur|PREP\_\_Prep+SN|en|PREP}
\end{small}
\subsection{GEN (the SCF Builder)}
The SCF builder extracts SCF candidates for each verb from the output of the pattern
extractor and calculates the number of corpus occurrences for each SCF and verb
combination. The syntactic constituents used for building the SCFs are the
following:
\begin{enumerate}
\item \texttt{SN} for nominal phrases;
\item \texttt{SINF} for infinitive clauses;
\item \texttt{SP[}\emph{prep}\texttt{+SN]} for prepositional phrases where the preposition is followed by a noun phrase. \emph{prep} is the head preposition;
\item \texttt{SP[}\emph{prep}\texttt{+SINF]} for prepositional phrases where the preposition is followed by an infinitive verb. \emph{prep} is the head preposition;
\item \texttt{SA} for adjectival phrases;
\item \texttt{COMPL} for subordinate clauses.\\
\end{enumerate}
When a verb has no dependency, its SCF is considered as \texttt{INTRANS}.
\texttt{(4)} shows the output of the SCF builder for \texttt{(1)}.\\
\begin{small}
\texttt{(4) S'ABATTRE+s'abattre ;;; SP[sur+SN]\_SP[en+SN]}
\end{small}
\subsection{CON and EVAL (SCF Filter)}
\label{filter}
Each step of the process is fully automatic, so the output of the SCF builder is noisy
due to tagging, parsing or other processing errors. It is also noisy because
of the difficulty of the argument-adjunct distinction. The latter is difficult
even for humans.
Many criteria that have been defined are not usable in our case because
they either depend on lexical information which the parser cannot make use of
(since the task is to acquire this information) or on semantic information which even the
best parsers cannot yet learn reliably. The approach here is based on the assumption
that true arguments tend to occur in argument positions
more frequently than adjuncts. Thus many frequent SCFs in the system output
are correct.
The strategy is then to filter low frequency entries from the SCF builder output.
This is done using the maximum likelihood estimates \cite{korhonen00}. This
simple method involves calculating the relative frequency
of each SCF (for a verb) and comparing it to an empirically determined threshold.
The relative frequency of the SCF \emph{i} with the verb \emph{j} is calculated as follows:
\begin{center}
\begin{math}
rel\_freq(scf_{i},verb_{j}) = \dfrac{|scf_{i},verb_{j}|}{|verb_{j}|}
\end{math}
\end{center}
\noindent
\begin{math}|scf_{i},verb_{j}|\end{math} is the number of occurrences of the SCF \emph{i} with
the verb \emph{j} and \begin{math}|verb_{j}|\end{math} is the total number of occurrences of
the verb \emph{j} in the corpus.
If, for example, the frequency of the SCF \texttt{SP[sur+SN]\_SP[en+SN]} is below the
empirically defined threshold, the SCF is rejected by the filter. The MLE filter is not
perfect because it is based on rejecting low frequency SCFs. Although relatively more low than
high frequency SCFs are incorrect, sometimes rejected frames are correct.
The filter incorporates special heuristics for cases where this
assumption tends to generate too many errors. With prepositional SCFs involving one PP or more,
the filter determines which one is the less frequent PP.
It then re-assigns the associated frequency to the same SCF without this PP.
For example, \texttt{SP[sur+SN]\_SP[en+SN]} could be split to 2 SCFs : \texttt{SP[sur+SN]}
and \texttt{SP[en+SN]}. In this example, \texttt{SP[en+SN]} is the less frequent prepositional
phrase and the final SCF for the sentence \texttt{(1)} is \texttt{(5)}.\\
\texttt{(5) SP[sur+SN]}
\smallskip
Note that \texttt{SP[en+SN]} is here an adjunct.
\section{Some Limitations of this Approach}
This approach is very efficient to deal with large corpora. However, some issues remain. As the approach is based on automatic tools (especially parsers) that are far from perfect, the obtained resources always contain errors and have to be manually validated. Moreover, the system needs to get enough examples to be able to infer relevant information. Therefore, there is generally a lack of information for a lot of low productivity items (the famous ``sparsity problem'').
More fundamentally, some constructions are difficult to acquire and characterize automatically. On the one hand, idioms are not recognized as such by most acquisition systems. On the other hand, some adjuncts appear frequently with certain verbs (eg. some verbs like \textit{dormir} -- \textit{to sleep }-- frequently appear with location complements). The system then assumes that these are arguments, whereas linguistic theory would say without any doubt that these are adjuncts. Lastly, surface cues are sometimes insufficient to recognize ambiguous constructions (cf. \textit{...manger une glace \`a la vanille...} vs \textit{...manger une glace \`a la terrasse d'un caf\'e...} --- \textit{to eat a vanilla ice-cream} vs \textit{to eat an ice-cream at an outdoor cafe}).
In a traditional architecture, the filtering process incorporates in one modules the set of constraints (CON) and the evaluation function (EVAL). This makes the system less readable than if the constraints were modeled apart from the EVAL function. There is thus a need to refine the set of constraints
\section{A Solution: Provide an Explicit Modeling of the Set of Constraints (CON)}
We have shown in the previous section that a part of the errors produced were due to an over-simplification of the initial model. It is thus necessary to take other parameters into considerations in order to yield better results. This can be done by refining the set of constraints (CON).
\subsection{Refining CON }
The issues we have reported in the previous section do not mean that automatic methods are flawed, but they have a number of drawbacks that should be addressed. The acquisition process, based on an analysis of co-occurrences of the verb with its immediate complements (along with filtering techniques) makes the approach highly functional. It is a good approximation of the problem. However, this model does not take into account external constraints.
The analysis of the co-occurrences of the verb with its complement is meaningful but is not sufficient to fully grasp the problem. The fact that some phrasal complements (with a specific head noun) frequently co-occur with a given verb is most of the time useful, especially to identify idioms \cite{fabre08}, colligations \cite{firth57descriptive} and light verb constructions \cite{butt03}. On the other hand, the fact that a given prepositional phrase appear with a large number of verbs may indicate that the preposition introduces an adjunct rather than an argument.
So, instead of simply capturing the co-occurrences of a verb with its complements, a number of important features should be taken into account:
\begin{itemize}
\item indicator of the dispersion of the prepositional phrases (PP) depending on the nature of the preposition (if a PP with a given preposition appears with a wide range of different verbs, it is more likely to be a modifier);
\item indicator of the co-occurrence of the PP depending on the nature of the head noun (if a verb appears frequently with the same PP frame, it is more likely to form a semi-idiomatic expression);
\item indicator of the complexity of the sentence to be processed (if a sentence is complex, its analysis is less reliable).
\end{itemize}
In order to do this, the pattern extractor has to be modified in order to keep most of the information that were previously rejected as not relevant. These indicators then need to be calculated so as to be taken into account by EVAL.
\subsection{Modifying EVAL}
All the constraints can be evaluated separately, so as to obtain for each of them an ideal evaluation of the parameter. There are two ways of doing this: \textit{i}) by automatically inferring the different weights from a set of annotated data or \textit{ii}) by estimating the results of various manually defined weights. We are currently using this last method since data annotation is very costly. However, the first approach would certainly lead to more accurate results.
The weight and the ranking of the different constraints must then be examined. A linear model can provide a first approximation but there are surely better ways to integrate the different constraints. Some studies provide some cues but they need to be proper evaluated in order to be integrated in this framework \cite{blache08}.
\subsection{Manual Validation}
Lastly, the approach requires a manual validation. Rather than leaving the validation process apart for further examination by a linguist, we propose to integrate it in the acquisition process itself. Taking into consideration the number of examples and the complexity of the sentences used for training, it is possible to associate confidence scores with the different constructions of a given verb: the linguist is then able to quickly focus on the most problematic cases. It is also possible to propose tentative constructions to the linguist, when not enough occurrences are available for training. In the end, when too few examples are available, the linguist can provide relevant information to the machine. However, with a well-designed and dynamic validation process, it is possible to obtain accurate and comprehensive lexicons, using only a small fraction of the time that would be necessary to manually develop a lexicon from scratch.
\section{Conclusion}
Tn this paper, we have proposed a new approach for the automatic acquisition of lexical knowledge from corpora using Optimality Theory.
Using this model, it is possible to represent a large part of the language activity through constraints.
We have shown that the individual evaluation of each constraint yields very accurate and precise results.
An implementation of this model is currently being done for Japanese \cite{marchal2012}. The model provides a better integration of the linguistic contraints within the automatic processing system. First results were competitive with other approaches while providing a more accurate linguistic description.
\bibliographystyle{splncs}
|
\section{\label{I}Introduction}
The subtlety of the theory of swimming at low Reynolds number has not always been fully appreciated. It is important to have simple examples for which calculations can be performed in detail. The first such example was furnished by Taylor \cite{1} in his seminal work on the swimming of an undulating planar sheet immersed in a viscous incompressible fluid. Soon after,
Lighthill \cite{2} studied the swimming of a sphere. He considered a squirming sphere with surface displacements in the spherical surface. His work was extended by Blake \cite{3}, who considered the full class of surface displacements.
The goal of the theory is to calculate the swimming velocity and the rate of dissipation in the fluid for given time-periodic deformations of the body. The rate of dissipation equals the power necessary to achieve the swimming motion. Shapere and Wilczek formulated the problem in terms of a gauge field on the space of shapes \cite{4}. They pointed out \cite{5} that the measure of efficiency of a stroke introduced by Lighthill and Blake is not appropriate. In low Reynolds number swimming, unlike in the problem of Stokes friction, the power is proportional to the speed, rather than the square of the speed. As a measure of efficiency Shapere and Wilczek therefore introduced a dimensionless number measuring the ratio of speed and power, rather than the ratio of speed squared and power.
The theory of swimming at low Reynolds number is based on the Stokes equations \cite{6}. In earlier work we have extended the theory to include the rate of change of fluid momentum, as given by the linearized Navier-Stokes equations \cite{7}. As an example we studied small-amplitude swimming of a deformable sphere \cite{8}, and found the optimum efficiency for the class of swimming motions for which the first order flow velocity is irrotational. Our definition of efficiency was analogous to that of Shapere and Wilczek.
The calculation based on the linearized Navier-Stokes equations was rather elaborate. It turns out that for irrotational flow the inertial effect vanishes, so that for this class of fluid motions it suffices to use the Stokes equations. This allows a simpler formalism and easier calculations. In the following we discuss the theory on the basis of the Stokes equations, and in addition derive some new results. The stroke of maximum efficiency involves a significant contribution of high order multipoles. This leads us to consider an additional measure of swimming performance, allowing minimization of the energy consumption at fixed amplitude of stroke. We provide a numerical estimate of speed and power for optimal swimming via potential flow of a typical bacterium. Customarily the speed is calculated for given power from Stokes drag \cite{9}.
In the first part of the article we restrict attention to axisymmetric irrotational flow. The fluid flow velocity can be derived from a scalar potential which satisfies Laplace's equations. It is therefore natural to introduce multipoles in analogy to electrostatics \cite{10}. To linear order the pressure disturbance vanishes. The swimming speed and the power are bilinear in the surface displacements. The class of potential flows is important because of the connection to inviscid flow theory based on the full set of Navier-Stokes equations, as relevant for swimming at high Reynolds number \cite{11}.
Subsequently we study more general axisymmetric polar flow. This involves modes with vorticity and a non-vanishing pressure disturbance, and requires the use of an additional set of multipoles. It turns out that the more complicated flow with vorticity leads to a significantly higher maximum efficiency than found for potential flow. Again we consider the measure of swimming performance based on energy consumption at fixed amplitude, and provide a numerical estimate for a typical bacterium.
\section{\label{II}Flow equations}
We consider a flexible sphere of radius $a$ immersed in a viscous
incompressible fluid of shear viscosity $\eta$. At low Reynolds
number and on a slow time scale the flow velocity
$\vc{v}(\vc{r},t)$ and the pressure $p(\vc{r},t)$ satisfy the
Stokes equations
\begin{equation}
\label{2.1}\eta\nabla^2\vc{v}-\nabla p=0,\qquad\nabla\cdot\vc{v}=0.
\end{equation}
The fluid is set in motion by time-dependent distortions of the
sphere. We shall study periodic distortions which lead to swimming
motion of the sphere. The surface displacement
$\vc{\xi}(\vc{s},t)$ is defined as the vector distance
\begin{equation}
\label{2.2}\vc{\xi}=\vc{s}'-\vc{s}
\end{equation}
of a point $\vc{s}'$ on the displaced surface $S(t)$ from the
point $\vc{s}$ on the sphere with surface $S_0$. The fluid
velocity $\vc{v}(\vc{r},t)$ is required to satisfy
\begin{equation}
\label{2.3}\vc{v}(\vc{s}+\vc{\xi}(\vc{s},t))=\frac{\partial\vc{\xi}(\vc{s},t)}{\partial t}.
\end{equation}
This amounts to a no-slip boundary condition. The instantaneous translational swimming velocity $\vc{U}(t)$,
the rotational swimming velocity $\vc{\Omega}(t)$, and the flow pattern $(\vc{v},p)$ follow from the condition that no net
force or torque is exerted on the fluid. We evaluate these quantities by a perturbation expansion in powers of the
displacement $\vc{\xi}$.
In the first part of the article we restrict attention to motions for which to first order in the
displacement the flow is irrotational, so that the flow velocity
is the gradient of a scalar potential,
\begin{equation}
\label{2.4}\vc{v}_1=\nabla\phi_1.
\end{equation}
We specify the surface displacement by assuming an expression for
the first order potential. We assume the flow to be symmetric
about the $z$ axis, so that in spherical coordinates
$(r,\theta,\varphi)$, defined with respect to the center of the
sphere in the rest system, the potential takes the form
$\phi_1(r,\theta,t)$. The potential $\phi_1(r,\theta,t)$ tends to
zero at infinity, and can be expressed as the Poisson integral
\begin{equation}
\label{2.5}\phi_1(r,\theta,t)=\int_{r'<a}\frac{1}{|\vc{r}-\vc{r}'|}\;\rho(r',\theta',t)\;d\vc{r}',
\end{equation}
with a source density $\rho(r,\theta,t)$ localized within the sphere of radius $a$. To first order the pressure remains
constant and equal to the ambient pressure $p_0$. We regard the source density $\rho(r,\theta,t)$ as given, and define the
surface displacement from
\begin{equation}
\label{2.6}\frac{\partial\vc{\xi}}{\partial t}=\nabla\phi_1\big|_{r=a}.
\end{equation}
Instead of $\vc{\xi}$ we regard $\rho$ as the expansion parameter.
For given source density $\rho(r,\theta,t)$ one can evaluate the
first order potential $\phi_1(r,\theta,t)$ by use of Eq. (2.5).
Hence one finds the first order flow velocity
$\vc{v}_1(r,\theta,t)$ by use of Eq. (2.4). Since this tends to
zero faster than $1/r$, the force exerted on the fluid and the
swimming velocity $U(t)$ vanish to first order. The rotational
velocity $\Omega(t)$ and the torque vanish automatically by
symmetry.
We consider in particular harmonic time variation at frequency
$\omega$, with source density
\begin{equation}
\label{2.7}\rho(r,\theta,t)=\rho_c(r,\theta)\cos\omega
t+\rho_s(r,\theta)\sin\omega t,
\end{equation}
with suitably chosen functions $\rho_s(r,\theta)$ and
$\rho_c(r,\theta)$. Since the no-slip condition is nonlinear, the
solution of the flow problem involves harmonics with all integer
multiples of $\omega$.
We perform a perturbation expansion in powers of the two-component
source density
$\vc{\rho}(\vc{r})=(\rho_c(\vc{r}),\rho_s(\vc{r}))$. To second
order in $\vc{\rho}$ the flow velocity and the swimming velocity
take the form
\begin{equation}
\label{2.8}\vc{v}(\vc{r},t)=\vc{v}_1(\vc{r},t)+\vc{v}_2(\vc{r},t)+...,\qquad
U(t)=U_2(t)+....
\end{equation}
Both $\vc{v}_1$ and $\vc{\xi}$ vary harmonically with frequency
$\omega$, and can be expressed as
\begin{eqnarray}
\label{2.9}\vc{v}_1(\vc{r},t)&=&\vc{v}_{1c}(\vc{r})\cos\omega t+\vc{v}_{1s}(\vc{r})\sin\omega t,\nonumber\\
\vc{\xi}(\vc{s},t)&=&\vc{\xi}_{c}(\vc{s})\cos\omega
t+\vc{\xi}_{s}(\vc{s})\sin\omega t.
\end{eqnarray}
Expanding the no-slip condition Eq. (2.3) to second order we find
for the flow velocity at the surface
\begin{eqnarray}
\label{2.10}\vc{u}_{1S}(\theta,t)&=&\vc{v}_1\big|_{r=a}=\frac{\partial\vc{\xi}(\theta,t)}{\partial t},\nonumber\\
\vc{u}_{2S}(\theta,t)&=&\vc{v}_2\big|_{r=a}=-\vc{\xi}\cdot\nabla\vc{v}_1\big|_{r=a}.
\end{eqnarray}
Hence the swimming velocity can be evaluated as \cite{7}
\begin{equation}
\label{2.11}
U_2(t)=-\frac{1}{4\pi}\int\vc{u}_{2S}\cdot\vc{e}_z\;d\Omega.
\end{equation}
The time-averaged swimming velocity is given by
\begin{equation}
\label{2.12}\overline{U_2}=-\frac{1}{4\pi}\int\overline{\vc{u}}_{2S}\cdot\vc{e}_z\;d\Omega,
\end{equation}
where the overhead bar indicates a time-average over a period $T=2\pi/\omega$.
The remainder $U_2(t)-\overline{U_2}$ oscillates at frequency $2\omega$.
To second order the rate of dissipation $\mathcal{D}_2(t)$ is
determined entirely by the first order solution. It may be
expressed as a surface integral \cite{7}
\begin{equation}
\label{2.13}\mathcal{D}_2=-2\eta\int_{r=a}\nabla\phi_{1}.(\nabla\nabla\phi_{1}).\vc{e}_r\;dS.
\end{equation}
The rate of dissipation is positive and oscillates in time about a
mean value. The mean rate of dissipation equals the power
necessary to generate the motion.
\section{\label{III}Multipole modulation}
In explicit calculations we expand the source density and the first order potential in spherical
harmonics. We define the solid spherical harmonics $\Phi_l^\pm$ as
\begin{equation}
\label{3.1}\Phi^+_l(r,\theta)=r^lP_l(\cos\theta),\qquad\Phi^-_l(r,\theta)=r^{-l-1}P_l(\cos\theta),
\end{equation}
with Legendre polynomials $P_l$ in the notation of Edmonds \cite{12}. The source density $\rho_l=\Phi^+_l$
inside the sphere generates a potential proportional to $\Phi^-_l$ outside the sphere. It is natural
to extend the potential and the corresponding velocity field inside the sphere. The first order potential
outside the sphere is expanded as
\begin{equation}
\label{3.2}\phi_1(r,\theta)=\omega a^2\sum^\infty_{l=0}\mu_l\bigg(\frac{a}{r}\bigg)^{l+1}P_l(\cos\theta),\qquad r>a,
\end{equation}
with dimensionless multipole coefficients $\{\mu_l\}$. The corresponding first order potential
inside the sphere is given by
\begin{equation}
\label{3.3}\phi_1(r,\theta)=\frac{1}{2}\omega a^2\sum^\infty_{l=0}\mu_l\bigg[(2l+3)\bigg(\frac{r}{a}\bigg)^l
-(2l+1)\bigg(\frac{r}{a}\bigg)^{l+2}\bigg]P_l(\cos\theta),\qquad r<a.
\end{equation}
This has been constructed such that the potential and its radial derivative are continuous at $r=a$. The
corresponding source density is
\begin{equation}
\label{3.4}\rho(r,\theta)=\frac{\omega}{4\pi}\sum^\infty_{l=0}\mu_l(2l+1)(2l+3)\bigg(\frac{r}{a}\bigg)^lP_l(\cos\theta),
\qquad r<a.
\end{equation}
The first order flow outside the sphere is
\begin{equation}
\label{3.5}\vc{v}_1(r,\theta)=-\omega a\sum^\infty_{l=0}\mu_l\vc{u}_l(r,\theta),\qquad r>a,
\end{equation}
with component field
\begin{equation}
\label{3.6}\vc{u}_l(r,\theta)=\bigg(\frac{a}{r}\bigg)^{l+2}\big[(l+1)P_l(\cos\theta)\vc{e}_r
+P^1_l(\cos\theta)\vc{e}_\theta\big],
\end{equation}
with associated Legendre function of the first kind $P^1_l(\cos\theta)$, in the notation of Edmonds \cite{12}. We note that
\begin{equation}
\label{3.7}\vc{u}_l(r,\theta)=-a^{l+2}\nabla\Phi^-_l(r,\theta).
\end{equation}
For the time-dependent source density of the form Eq. (2.7) the
multipole coefficients are time-dependent and can be expressed as
\begin{equation}
\label{3.8}\mu_l(t)=\mu_{lc}\cos\omega t+\mu_{ls}\sin\omega t.
\end{equation}
These generate the first order flow
\begin{equation}
\label{3.9}\vc{v}_1(r,\theta,t)=-\omega a\sum^\infty_{l=0}\mu_l(t)\vc{u}_l(r,\theta),\qquad r>a.
\end{equation}
The corresponding displacement is
\begin{equation}
\label{3.10}\vc{\xi}(\theta,t)=a\sum^\infty_{l=0}\big[\mu_{ls}\cos\omega
t-\mu_{lc}\sin\omega t\big]\vc{u}_l(a,\theta).
\end{equation}
In the calculation of the mean swimming velocity, as given by Eq.
(2.12), we use the identity
\begin{equation}
\label{3.11}\int_{r=a}(\nabla\Phi^-_k)\cdot(\nabla\nabla\Phi^-_l)\cdot\vc{e}_z\;dS=-4\pi
k(k+1)a^{-2k-2}\delta_{k,l+1}.
\end{equation}
This shows that the mean swimming velocity is given by a sum of products of adjacent multipole coefficients,
\begin{equation}
\label{3.12}\overline{U_2}=\frac{1}{2}\omega
a\sum^\infty_{l=0}(l+1)(l+2)\big[\mu_{lc}\mu_{l+1,s}-\mu_{ls}\mu_{l+1,c}\big].
\end{equation}
We define the multipole moment vector $\vc{\mu}$ as the one-dimensional array
\begin{equation}
\label{3.13}\vc{\mu}=(\mu_{0s},\mu_{0c},\mu_{1s},\mu_{1c},....).
\end{equation}
Then $\overline{U_2}$ can be expressed as
\begin{equation}
\label{3.14}\overline{U_2}=\frac{1}{2}\omega
a(\vc{\mu},\du{B}\vc{\mu}),
\end{equation}
with a dimensionless symmetric matrix $\du{B}$. The upper
left-hand corner of the matrix $\du{B}$, truncated at $l=3$, reads
\begin{equation}
\label{3.15}\du{B}_{03}=\left(\begin{array}{cccccccc}0&0&0&-1&0&0&0&0
\\0&0&1&0&0&0&0&0
\\0&1&0&0&0&-3&0&0
\\-1&0&0&0&3&0&0&0
\\0&0&0&3&0&0&0&-6
\\0&0&-3&0&0&0&6&0
\\0&0&0&0&0&6&0&0
\\0&0&0&0&-6&0&0&0\end{array}\right).
\end{equation}
On the cross-diagonals the numbers $\frac{1}{2}(l+1)(l+2)$ appear for $l=0,1,2,...$.
In the calculation of the rate of dissipation, as given by Eq. (2.18), we use the identity
\begin{equation}
\label{3.16}\int_{r=a}(\nabla\Phi^-_k)\cdot(\nabla\nabla\Phi^-_l)\cdot\vc{e}_r\;dS=-4\pi
(k+1)(k+2)a^{-2k-3}\delta_{k,l}.
\end{equation}
Hence the time-averaged rate of dissipation is given by
\begin{equation}
\label{3.17}\overline{\mathcal{D}_2}=4\pi\eta\omega^2a^3\sum^\infty_{l=0}(l+1)(l+2)(\mu_{lc}^2+\mu_{ls}^2).
\end{equation}
This can be expressed as
\begin{equation}
\label{3.18}\overline{\mathcal{D}_2}=8\pi\eta\omega^2a^3(\vc{\mu},\du{A}\vc{\mu}),
\end{equation}
with a dimensionless diagonal matrix $\du{A}$. The upper left-hand
corner of the matrix $\du{A}$, truncated at $l=3$, reads
\begin{equation}
\label{3.19}\du{A}_{03}=\left(\begin{array}{cccccccc}1&0&0&0&0&0&0&0
\\0&1&0&0&0&0&0&0
\\0&0&3&0&0&0&0&0
\\0&0&0&3&0&0&0&0
\\0&0&0&0&6&0&0&0
\\0&0&0&0&0&6&0&0
\\0&0&0&0&0&0&10&0
\\0&0&0&0&0&0&0&10\end{array}\right).
\end{equation}
On the diagonal the numbers $\frac{1}{2}(l+1)(l+2)$ appear for $l=0,1,2,...$. The crucial identities (3.11) and (3.16) are proved by use of the generating function of the Legendre polynomials, or by use of known identities relating the polynomials.
\section{\label{IV}Linear chain problem}
The question arises how to maximize the mean swimming velocity for given mean rate of
dissipation. This leads to an eigenvalue problem for the set of multipole coefficients $\vc{\mu}$,
\begin{equation}
\label{4.1}\du{B}\vc{\mu}_\lambda=\lambda\du{A}\vc{\mu}_\lambda.
\end{equation}
The mathematical discussion is simplified by truncating the matrices at a maximum $l$-value,
say $L$. We call the truncated $2L+2$-dimensional matrices $\du{A}_{0L}$ and $\du{B}_{0L}$.
The truncated matrices correspond to swimmers obeying the constraint that all multipole coefficients for $l>L$ vanish.
It is seen from Eq. (3.12) that there is a degeneracy in the
problem. The sum for the mean velocity consists of a sum of two
interlaced chains. In the one chain the $s$-coefficients for even
$l$ and the $c$-coefficients for odd $l$ appear. In the other
chain the $s$-coefficients for odd $l$ and the $c$-coefficients
for even $l$ appear. It is therefore sufficient to consider the
first type of chain. Eigenvectors of this form with the
coefficients for the second chain put equal to zero can be mapped
onto eigenvectors for the same eigenvalue with the two chains
interchanged. We call eigenvectors of the first type even, and
eigenvectors of the second type odd. The degeneracy corresponds to
invariance under a shift in time by $\pi/2\omega$.
There is also a symmetry under time reversal. Eigenvalues appear in pairs $\pm\lambda_j$.
The even eigenvector for $-\lambda_j$ can be obtained from the even eigenvector for $+\lambda_j$
by the replacement of the $c$-coefficients by their opposites, leaving the $s$-coefficients unchanged.
For the two conjugate eigenvectors the swimming velocity is equal and opposite for the same rate of dissipation.
The first symmetry allows a simplification of the eigenvalue problem by a reduction of the matrix dimension by a factor one half. There is a duplication in the matrices $\du{B}$ and $\du{A}$ which can be removed by use of complex notation. Thus we introduce the complex multipole moment
\begin{equation}
\label{4.2}\mu^c_l=(-i)^l(\mu_{lc}+i\mu_{ls}),
\end{equation}
and correspondingly instead of Eq. (3.13)
\begin{equation}
\label{4.3}\vc{\mu}^c=(\mu^c_{0},\mu^c_{1},\mu^c_{2},....).
\end{equation}
Then $\overline{U}_2$ and $\overline{\mathcal{D}}_2$ can be expressed as
\begin{equation}
\label{4.4}\overline{U_2}=\frac{1}{2}\omega
a(\vc{\mu}^c|\du{B}^c|\vc{\mu}^c),\qquad\overline{\mathcal{D}_2}=8\pi\eta\omega^2a^3
(\vc{\mu}^c|\du{A}^c|\vc{\mu}^c),
\end{equation}
with the notation
\begin{equation}
\label{4.5}(\vc{\mu}^c|\du{B}^c|\vc{\mu}^c)=\sum^\infty_{ll'}\mu^{c*}_lB^c_{ll'}\mu^c_{l'}.
\end{equation}
The truncated matrices $\du{B}^c_{03}$ and $\du{A}^c_{03}$ read
\begin{equation}
\label{4.6}\du{B}^c_{03}=\left(\begin{array}{cccc}
0&1&0&0
\\1&0&3&0
\\0&3&0&6
\\0&0&6&0
\end{array}\right),\qquad\du{A}^c_{03}=\left(\begin{array}{cccc}
1&0&0&0
\\0&3&0&0
\\0&0&6&0
\\0&0&0&10
\end{array}\right).
\end{equation}
The eigenvalue problem now reads
\begin{equation}
\label{4.7}\du{B}^c|\vc{\mu}^c_\lambda)=\lambda\du{A}^c|\vc{\mu}^c_\lambda).
\end{equation}
Since the matrices $\du{B}^c$ and $\du{A}^c$ are real and symmetric, the eigenvectors can be chosen to be real.
With truncation at $l=L$ the eigenvalue problem Eq. (4.7) is identical to that for a linear harmonic chain with masses corresponding to the diagonal elements of the matrix $\du{A}^c_{0L}$ and spring constants corresponding to the off-diagonal elements of the matrix $\du{B}^c_{0L}$. We can simplify further by renormalizing such that the masses are equal. Thus we introduce the modified moments
\begin{equation}
\label{4.8}f_l=\sqrt{(l+1)(l+2)}\mu^c_l.
\end{equation}
With these moments the rate of dissipation is
\begin{equation}
\label{4.9}\overline{\mathcal{D}_2}=4\pi\eta\omega^2a^3\sum^L_{l=0}|f_l|^2
=8\pi\eta\omega^2a^3(\vc{f}|\du{A}^{c\prime}|\vc{f}),
\end{equation}
where $\du{A}^{c\prime}=\frac{1}{2}\du{I}$ with unit matrix
$\du{I}$, and the swimming velocity is
\begin{equation}
\label{4.10}\overline{U_2}=\frac{1}{2}\omega
a\sum^L_{l=0}k_l\mathrm{Re}f_l^*f_{l+1}=\frac{1}{2}\omega
a(\vc{f}|\du{B}^{c\prime}|\vc{f}),
\end{equation}
where $\du{B}^{c\prime}$ is symmetric with non-zero elements
\begin{equation}
\label{4.11}B^{c\prime}_{l,l+1}=B^{c\prime}_{l+1,l}=\frac{1}{2}k_l,\qquad
k_l=\sqrt{\frac{l+1}{l+3}}.
\end{equation}
The coefficients $k_l$ tend to unity for large $l$, so that the eigenvalue problem
\begin{equation}
\label{4.12}
\du{B}^{c\prime}|\vc{f}_\lambda)=\lambda\du{A}^{c\prime}|\vc{f}_\lambda),
\end{equation}
corresponds to a chain of equal masses coupled by spring constants
which become uniform for large $l$.
We impose the constraint that the multipole coefficients for $l=0$
vanish. The coefficients for $l=0$ correspond to uniform spherical
expansion, which is excluded if we impose volume conservation. We
denote the matrices truncated at $L$ and with the first two rows
and columns deleted as $\du{A}_{1L}$ and $\du{B}_{1L}$. These have
dimension $2L$. The corresponding matrices $\du{A}^c_{1L}$ and
$\du{B}^c_{1L}$ have dimension $L$ and the matrices
$\du{A}^{c\prime}_{1L}$ and $\du{B}^{c\prime}_{1L}$ have dimension
$L$.
The eigenvalue problem Eq. (4.12) for the linear chain of $L$ equal masses coupled with equal force constants has eigenvalues
\begin{equation}
\label{4.13}
\lambda_q=2\cos\bigg(\frac{q\pi}{L+1}\bigg),\qquad q=1,...,L,
\end{equation}
and corresponding eigenvectors with components
\begin{equation}
\label{4.14}
f_{k,q}=C_q\sin\bigg(\frac{kq\pi}{L+1}\bigg),\qquad k,q=1,...,L,
\end{equation}
where $C_q$ is a normalization factor. The largest eigenvalue occurs for $q=1$. For this eigenvalue the components of the eigenvector vary slowly with $k$. In the limit $L\rightarrow\infty$ the maximum eigenvalue tends to $2$ and the components of the corresponding eigenvector tend to a constant.
\section{\label{V}Speed, power, efficiency}
As characteristic dimension of the sphere we take the diameter
$2a$. The dimensionless efficiency of translational swimming is
defined as the ratio \cite{7}
\begin{equation}
\label{5.1}E_T=4\eta\omega a^2\frac{|\overline{U_2}|}{\overline{\mathcal{D}_2}}.
\end{equation}
The optimum efficiency is related to the maximum eigenvalue by
\begin{equation}
\label{5.2}E_{T\mathrm{max}}=\lambda_{\mathrm{max}}/(4\pi).
\end{equation}
Due to a different normalization of the matrix $\du{B}$ the eigenvalue is four times that defined earlier \cite{8}.
It follows from Eq. (4.13) that the optimum efficiency is $1/(2\pi)$.
It is therefore of interest to consider the relative efficiency
\begin{equation}
\label{5.3}\eta_{1\mathrm{pot}}=2\pi E_T
\end{equation}
as a measure of efficiency in the space of potential flows. Here we have used the notation of Shapere and Wilczek \cite{5}.
We denote the eigenvector with largest eigenvalue of the truncated eigenvalue problem Eq. (4.7) with matrices $\du{A}^c_{1L}$ and $\du{B}^c_{1L}$ as $\vc{g}_{1L}$,
with normalization $(\vc{g}_{1L}|\vc{g}_{1L})=1$, and define
\begin{equation}
\label{5.4}\hat{U}_{1L}=(\vc{g}_{1L}|\du{B}^c_{1L}|\vc{g}_{1L}),\qquad\hat{\mathcal{D}}_{1L}
=(\vc{g}_{1L}|\du{A}^c_{1L}|\vc{g}_{1L}).
\end{equation}
Then correspondingly
\begin{equation}
\label{5.5}\frac{\hat{U}_{1L}}{\hat{\mathcal{D}}_{1L}}=\lambda_{\mathrm{max}}(1,L).
\end{equation}
The maximum eigenvalue $\lambda_{\mathrm{max}}(1,L)$
increases monotonically with $L$, since with increasing $L$ the space of possible modes gets larger. In Fig. 1 we plot $\frac{1}{2}\lambda_{\mathrm{max}}(1,L)$ for
values $L=2,...,30$. In Fig. 2 we show the components of the eigenvector $\vc{g}_{1L}$
with largest eigenvalue for $L=8$.
As shown in Fig. 1 the efficiency $E_{T\mathrm{max}}(1,L)$
increases monotonically with $L$. This suggests that the limit
$L\rightarrow\infty$ corresponds to the best swimmer. However, it
is worthwhile to consider also the dimensionless speed
$\hat{U}_{1L}$ and power $\hat{\mathcal{D}}_{1L}$ separately.
It is seen numerically that both quantities increase linearly with
$L$ at large $L$. When listing values for different $L$ we are
comparing speed and power for eigenvectors with the same
normalization.
It makes more sense to compare chains with the same amplitude of
motion. It follows from Eqs. (3.6) and (3.10) that for the
eigenvector $\vc{g}_{1L}$ the displacement $\vc{\xi}(t)$ at
$\theta=\pi/2$ describes an ellipse in the $zx$ plane given by the
equation
\begin{equation}
\label{5.6}\frac{\xi_x^2}{A(1,L)^2}+\frac{\xi_z^2}{B(1,L)^2}=a^2,
\end{equation}
with $A(1,L)$ and $B(1,L)$ given by
\begin{equation}
\label{5.7}A(1,L)=|\sum^L_{l=1}(l+1)\mathrm{Re}(i^lg_{1L,l})P_l(0)|,\qquad
B(1,L)=|\sum^L_{l=1}\mathrm{Im}(i^lg_{1L,l})P^1_l(0)|.
\end{equation}
For multipoles given by
$\vc{g}_{1L}/A(1,L)$ the ellipse described by $\vc{\xi}(t)$ will
have vertical semi-axis $a$ and horizontal semi-axis $b=B(1,L)a/A(1,L)$, if we take the $z$ axis to be horizontal. We find that the vertical semi-axis is larger than the horizontal one, except for $L=3$ and $L=5$.
For multipoles $\varepsilon\vc{g}_{1L}/A(1,L)$ the vertical semi-axis
has length $\varepsilon a$, where $\varepsilon$ can be taken to be
independent of $L$. We therefore consider the reduced speed and
power at fixed vertical amplitude of stroke,
\begin{equation}
\label{5.8}\hat{U}^A_{1L}=\frac{\hat{U}_{1L}}{A(1,L)^2},\qquad\hat{\mathcal{D}}^A_{1L}
=\frac{\hat{\mathcal{D}}_{1L}}{A(1,L)^2}.
\end{equation}
In Fig. 3 we plot the reduced speed $\hat{U}^A_{1L}$ as a function
of $L$, and in Fig. 4 we plot the reduced power
$\hat{\mathcal{D}}^A_{1L}$ as a function of $L$. Remarkably, the
reduced power at fixed amplitude shows a minimum at $L=8$, given
by $\hat{\mathcal{D}}^A_{18}=2.761$. An animalcule for which the
amplitude of motion is given by its structure, and for which the
relative amplitude of stroke is fixed, say at $\varepsilon=0.1$,
swims with least power for displacement $\vc{\xi}(\theta,t)$
determined by the set of multipoles
$\varepsilon\vc{g}_{1L}/A(1,L)$ with $L=8$. At $L=8$ the reduced
amplitude is $A(1,8)=2.572$, and the reduced speed is
$\hat{U}^A_{18}=4.380$.
For the set of multipoles $\vc{\mu}^c(1,L)=\varepsilon\vc{g}_{1L}/A(1,L)$ the mean speed and rate of dissipation are
\begin{equation}
\label{5.9}\overline{U}_2=\frac{1}{2}\omega a\varepsilon^2\hat{U}^A_{1L},\qquad\overline{\mathcal{D}}_2
=8\pi\eta\omega^2 a^3\varepsilon^2\hat{\mathcal{D}}^A_{1L}.
\end{equation}
In low Reynolds number swimming the speed is proportional to the
power. It is incorrect to estimate the required power on the basis
of Stokes' law \cite{9}, which corresponds to pulling of the sphere
through the fluid. In the case of pulling the power is
proportional to the square of the speed.
For a bacterium of radius $0.1\;\mu\mathrm{m}$ in water of shear
viscosity $\eta=0.001$ in SI units, the power for $L=8$ is
$P=\overline{\mathcal{D}_2}=6.94\times
10^{-23}\varepsilon^2\omega^2$ watt. The corresponding speed is
$U=2.19\times 10^{-7}\varepsilon^2\omega$ m/sec. The frequency is estimated \cite{14} as $10^4$ sec$^{-1}$. This is to be compared with the viscous time
scale $\tau_v=a^2\rho/\eta=10^{-8}$ sec. The power is calculated from Eq. (5.9) as
$P=6.94\times 10^{-15}\varepsilon^2$ watt and speed
$U=2.19\times 10^{-3}\;\varepsilon^2$ m/sec. The efficiency is $E_T=0.126$, compared with the maximum possible for potential flow $E_{T\mathrm{max}}=1/(2\pi)=0.159$.
The metabolic rate of birds has been
measured as 20.000 watt/m$^3$, of which one quarter is estimated
to be available for mechanical work \cite{15}. Accepting the same
rate for bacteria, we have $P=2.09\times 10^{-17}$ watt, and hence
find relative amplitude $\varepsilon=0.055$ and speed $U=6.6\times
10^{-6}$ m/sec. Therefore the bacterium moves several
diameters per second, in reasonable agreement with experimental
data \cite{3}. The specific energy consumption, defined as the power divided by the product of speed and weight \cite{15}, is about five orders of magnitude larger than that of a Boeing 747. We note that Dusenbery \cite{9} estimates the
available power as only 3 watt/m$^3$, instead of 5000 watt/m$^3$.
In our calculation this low power level would lead to a much too
small speed.
\section{\label{VI}Time-dependent swimming}
It is of interest to study some features of the swimming motion in
more detail. As we have shown above, the mean speed and mean power
to second order in the displacement $\vc{\xi}(t)$ are given by
bilinear expressions derived from the first order flow pattern.
For a chosen characteristic amplitude the latter can be optimized
to provide speed at minimum power. The set of multipoles
$\varepsilon\vc{g}_{1L}/A(1,L)$ with $L=8$ corresponding to the
eigenvector with maximum eigenvalue leads to optimal swimming. In
Fig. 5 we plot the nearly circular motion of the displacement
vector at $\theta=3\pi/12,\theta=5\pi/12,\,\pi/2,\;7\pi/12,\theta=9\pi/12,$ and $\varepsilon=0.1$
for seven-eighth of the period $T=2\pi/\omega$ , starting at
$t=0$. In Fig. 6 we show the radial displacement as a function of
the polar angle $\theta$ at times $t=0,\;t=T/8$ and $t=T/4$. This
demonstrates the running wave character of the surface wave. The
plot for the tangential displacement looks similar.
The second order velocity $U_2(t)$ follows from Eq. (2.11). This
can be evaluated by use of Eq. (3.11), which yields
\begin{eqnarray}
\label{6.1}U_2(t)=\omega
a\sum^\infty_{l=0}(l+1)(l+2)&\big[&\mu_{lc}\mu_{l+1,s}\cos^2\omega
t
-\mu_{ls}\mu_{l+1,c}\sin^2\omega t+\nonumber\\
+&(&\mu_{ls}\mu_{l+1,s}-\mu_{lc}\mu_{l+1,c})\sin\omega t\cos\omega
t\; \big].
\end{eqnarray}
The time-average of this expression equals that given in Eq.
(3.12). In Fig. 7 we plot the ratio $U_2(t)/\overline{U}_2$ for
the optimal stroke with displacement $\vc{\xi}(\theta,t)$
determined by the set of multipoles
$\varepsilon\vc{g}_{1L}/A(1,L)$ with $L=8$. The maximum deviation
from unity is about one percent.
The second order rate of dissipation $\mathcal{D}_2(t)$ follows
from Eq. (2.13). This can be evaluated by use of Eq. (3.16), which
yields
\begin{equation}
\label{6.2}\mathcal{D}_2(t)=8\pi\eta\omega^2a^3
\sum^\infty_{l=0}(l+1)(l+2)\big[\mu_{lc}^2\cos^2\omega t
+\mu_{ls}^2\sin^2\omega t +2\mu_{lc}\mu_{ls}\sin\omega t\cos\omega
t\; \big].
\end{equation}
The time-average of this expression equals that given in Eq.
(3.17). In Fig. 7 we plot also the ratio
$\mathcal{D}_2(t)/\overline{\mathcal{D}}_2$ for the optimal stroke
with displacement $\vc{\xi}(\theta,t)$ determined by the set of
multipoles $\varepsilon\vc{g}_{1L}/A(1,L)$ with $L=8$. It turns out that for this stroke $\mathcal{D}_2(t)/\overline{\mathcal{D}}_2$
equals unity within numerical accuracy.
The second order flow velocity $\vc{v}_2(\vc{r},t)$ follows from
the second order velocity at the surface $\vc{u}_{2S}(\theta,t)$,
as given by Eq. (2.10). The latter can be expanded in terms of a
complete set of outgoing waves $\{\vc{v}^-_{l0\sigma}(\vc{r})\}$,
where $\sigma$ takes the values $0,2$, as indicated
elsewhere \cite{16}. The modes with $\sigma=0$ are accompanied by a
pressure disturbance. The contribution for $l=1,\;\sigma=0$ decays
with a long range flow pattern falling off as $1/r$. This must be
cancelled by a Stokes solution $\vc{v}_2^{St}(\vc{r},t)$ which
vanishes on the sphere of radius $a$ and tends to
$-U_2(t)\vc{e}_z$ as $r\rightarrow\infty$. The procedure can be
performed straightforwardly, but we shall not present the details.
In principle the perturbation expansion in powers of the surface
displacement, as indicated in Eq. (2.8), can be extended to higher
order in similar fashion.
\section{\label{VII}Axisymmetric polar flows}
In the following we extend the analysis to more general flows. We consider motions for which to first order in the
displacement the flow is axisymmetric and polar, so that in spherical coordinates $(r,\theta,\varphi)$ the flow velocity $\vc{v}(\vc{r},t)$ and the pressure $p(\vc{r},t)$
do not depend on $\varphi$, and $\vc{v}$ has vanishing component $v_\varphi$. In general the solutions of the Stokes equations for the flow about a sphere have been classified \cite{16} into three types indexed $\sigma=0,1,2$. The potential flows considered earlier are of type $\sigma=2$. We now consider in addition flows of type $\sigma=0$. For the potential flows the pressure disturbance vanishes, but the flows of type $\sigma=0$ cannot be expressed as the gradient of a scalar potential and there is a pressure disturbance. For an axisymmetric flow of type $\sigma=1$ the flow velocity has only a $v_\varphi$ component, and the pressure disturbance vanishes. Flows of this type do not contribute to the translational velocity of the sphere.
The first order flow outside the sphere is expanded as
\begin{equation}
\label{7.1}\vc{v}_1(r,\theta,t)=-\omega a\sum^\infty_{l=1}\bigg[\mu_l(t)\vc{u}_l(r,\theta)+\kappa_l(t)\vc{v}_l(r,\theta)\bigg],\qquad r>a,
\end{equation}
with component field $\vc{u}_l(r,\theta)$ given by Eq. (3.6), and $\vc{v}_l(r,\theta)$ given by
\begin{equation}
\label{7.2}
\vc{v}_l(r,\theta)=\bigg(\frac{a}{r}\bigg)^{l}\big[(l+1)P_l(\cos\theta)\vc{e}_r
+\frac{l-2}{l}P^1_l(\cos\theta)\vc{e}_\theta\big].
\end{equation}
In the second sum in Eq. (7.1) we must put $\kappa_1(t)=0$, since the term with $l=1$ would correspond to a force $F_1(t)\vc{e}_z$. We have normalized such that at $r=a$ the function $\vc{v}_l$ has the same radial component as $\vc{u}_l$.
The solution $\vc{u}_l(r,\theta)$ is of type $\sigma=2$, the solution $\vc{v}_l(r,\theta)$ is of type $\sigma=0$. The corresponding first order pressure is
\begin{equation}
\label{7.3}p_1(r,\theta,t)=-\omega a\sum^\infty_{l=2}\kappa_l(t)p_l(r,\theta),\qquad r>a
\end{equation}
with component pressure disturbance
\begin{equation}
\label{7.4}p_l(r,\theta)=2\eta (2l-1)a^l\Phi^-_l(r,\theta).
\end{equation}
The multipole coefficients $\mu_l(t)$ and $\kappa_l(t)$ in Eq.
(7.1) can be expressed as
\begin{equation}
\label{7.5}\mu_l(t)=\mu_{lc}\cos\omega t+\mu_{ls}\sin\omega t,\qquad\kappa_l(t)=\kappa_{lc}\cos\omega t+\kappa_{ls}\sin\omega t.
\end{equation}
The corresponding displacement is
\begin{eqnarray}
\label{7.6}\vc{\xi}(\theta,t)=a\sum^\infty_{l=1}&\big[&\big(\mu_{ls}\cos\omega
t-\mu_{lc}\sin\omega t\big)\vc{u}_l(a,\theta)\nonumber\\
&+&\big(\kappa_{ls}\cos\omega
t-\kappa_{lc}\sin\omega t\big)\vc{v}_l(a,\theta)\big].
\end{eqnarray}
In the calculation of the mean swimming velocity, as given by Eq.
(2.12), we use the identities
\begin{eqnarray}
\label{7.7}\int_{r=a}\vc{u}_k\cdot(\nabla\vc{u}_l)\cdot\vc{e}_z\;dS=&-&4\pi
k(k+1)a\delta_{k,l+1},\nonumber\\
\int_{r=a}\vc{u}_k\cdot(\nabla\vc{v}_l)\cdot\vc{e}_z\;dS=&-&8\pi
\frac{(k+1)(k+2)}{2k+3}\;a\delta_{k,l-1}\nonumber\\&-&4\pi
\frac{k(k+1)(2k-3)}{2k+1}\;a\delta_{k,l+1},\nonumber\\
\int_{r=a}\vc{v}_k\cdot(\nabla\vc{u}_l)\cdot\vc{e}_z\;dS=&-&4\pi
\frac{k(k+1(2k-1)}{2k+1})\;a\delta_{k,l+1},\nonumber\\
\int_{r=a}\vc{v}_k\cdot(\nabla\vc{v}_l)\cdot\vc{e}_z\;dS=&-&8\pi
\frac{(k+1)(k+2)(2k-1)}{(2k+1)(2k+3)}\;a\delta_{k,l-1}\nonumber\\&-&4\pi
\frac{k(k+1)(2k-3)^2}{(2k-1)(2k+1)}\;a\delta_{k,l+1}.
\end{eqnarray}
The first one is equivalent to Eq. (3.11). It follows that the mean swimming velocity is again given by a sum of products of adjacent multipole coefficients,
\begin{eqnarray}
\label{7.8}\overline{U_2}&=&\frac{1}{2}\omega
a\sum^\infty_{l=1}\bigg[(l+1)(l+2)\big[\mu_{lc}\mu_{l+1,s}-\mu_{ls}\mu_{l+1,c}\big]\nonumber\\
&+&\frac{(l+1)(l+2)(2l-1)}{2l+3}\big[\kappa_{lc}\mu_{l+1,s}-\kappa_{ls}\mu_{l+1,c}\big]\nonumber\\
&+&\frac{(l+1)(l+2)(2l-1)}{2l+3}\big[\mu_{lc}\kappa_{l+1,s}-\mu_{ls}\kappa_{l+1,c}\big]\nonumber\\
&+&(l+1)(l+2)\frac{(2l-3)(2l-1)}{(2l+1)(2l+3)}\big[\kappa_{lc}\kappa_{l+1,s}-\kappa_{ls}\kappa_{l+1,c}\big]\bigg].
\end{eqnarray}
We define the complex multipole moment vector $\vc{\psi}$ as the one-dimensional array
\begin{equation}
\label{7.9}\vc{\psi}=(\kappa_{1c}+i\kappa_{1s},\mu_{1c}+i\mu_{1s},\kappa_{2c}+i\kappa_{2s},\mu_{2c}+i\mu_{2s},....).
\end{equation}
Then $\overline{U_2}$ can be expressed as
\begin{equation}
\label{7.10}\overline{U_2}=\frac{1}{2}\omega
a(\vc{\psi}|\du{B}|\vc{\psi}),
\end{equation}
with a dimensionless pure imaginary and antisymmetric matrix
$\du{B}$. The upper left-hand corner of the matrix $\du{B}$,
truncated at $l=4$, reads
\begin{equation}
\label{7.11}\du{B}_{14}=i\left(\begin{array}{cccccccc}0&0&\frac{1}{5}&-\frac{3}{5}&0&0&0&0
\\0&0&-\frac{3}{5}&-3&0&0&0&0
\\-\frac{1}{5}&\frac{3}{5}&0&0&-\frac{18}{35}&-\frac{18}{7}&0&0
\\\frac{3}{5}&3&0&0&-\frac{18}{7}&-6&0&0
\\0&0&\frac{18}{35}&\frac{18}{7}&0&0&-\frac{50}{21}&-\frac{50}{9}
\\0&0&\frac{18}{7}&6&0&0&-\frac{50}{9}&-10
\\0&0&0&0&\frac{50}{21}&\frac{50}{9}&0&0
\\0&0&0&0&\frac{50}{9}&10&0&0\end{array}\right).
\end{equation}
We can impose the constraint $\kappa_1=0$ by dropping the first
element of $\vc{\psi}$ and erasing the first row and column of the
matrix $\du{B}$. We denote the corresponding modified vector as
$\hat{\vc{\psi}}$ and the modified matrix as $\hat{\du{B}}$.
The rate of dissipation $\mathcal{D}_2(t)$ is
expressed as a surface integral \cite{7}
\begin{equation}
\label{7.12}\mathcal{D}_2=-\int_{r=a}\vc{v}_{1}.\vc{\sigma}_{1}.\vc{e}_r\;dS,
\end{equation}
where $\vc{\sigma}_1$ is the first order stress tensor, given by
\begin{equation}
\label{7.13}\vc{\sigma}_{1}=\eta(\nabla\vc{v}_1+\widetilde{\nabla\vc{v}_1})-p_1\vc{I}.
\end{equation}
In the calculation of the rate of dissipation we use the identities
\begin{eqnarray}
\label{7.14}\int_{r=a}\vc{u}_k\cdot(\nabla\vc{u}_l)\cdot\vc{e}_r\;dS&=&-4\pi a
(k+1)(k+2)\delta_{kl},\nonumber\\
\int_{r=a}\vc{u}_k\cdot(\nabla\vc{v}_l+\widetilde\nabla\vc{v}_l-p_l)\cdot\vc{e}_r\;dS&=&-8\pi a
\frac{(k+1)(k+2)(2k-1)}{2k+1}\delta_{kl},\nonumber\\
\int_{r=a}\vc{v}_k\cdot(\nabla\vc{u}_l)\cdot\vc{e}_r\;dS&=&-4\pi a
\frac{(k+1)(k+2)(2k-1)}{2k+1}\delta_{kl},\nonumber\\
\int_{r=a}\vc{v}_k\cdot(\nabla\vc{v}_l+\widetilde\nabla\vc{v}_l-p_l)\cdot\vc{e}_r\;dS&=&-8\pi a
\frac{(k+1)(2k^3+k^2-2k+2)}{k(2k+1)}\delta_{kl}.\nonumber\\
\end{eqnarray}
The first one is equivalent to Eq. (3.16).
The time-averaged rate of dissipation is given by
\begin{eqnarray}
\label{7.15}\overline{\mathcal{D}_2}=8\pi\eta\omega^2a^3\sum^\infty_{l=1}&\bigg[&\frac{1}{2}(l+1)(l+2)(\mu_{lc}^2+\mu_{ls}^2)\nonumber\\
&+&\frac{(l+1)(l+2)(2l-1)}{2l+1}(\mu_{lc}\kappa_{lc}+\mu_{ls}\kappa_{ls})\nonumber\\
&+&\frac{(l+1)(2l^3+l^2-2l+2)}{2l(2l+1)}(\kappa_{lc}^2+\kappa_{ls}^2)\bigg].
\end{eqnarray}
This can be expressed as
\begin{equation}
\label{7.16}\overline{\mathcal{D}_2}=8\pi\eta\omega^2a^3(\vc{\psi}|\du{A}|\vc{\psi}),
\end{equation}
with a dimensionless real and symmetric matrix $\du{A}$. We denote
the modified matrix obtained by dropping the first row and column
by $\hat{\du{A}}$. The upper left-hand corner of the matrix
$\du{A}$, truncated at $l=4$, reads
\begin{equation}
\label{7.17}\du{A}_{14}=\left(\begin{array}{cccccccc}1&1&0&0&0&0&0&0
\\1&3&0&0&0&0&0&0
\\0&0&\frac{27}{10}&\frac{18}{5}&0&0&0&0
\\0&0&\frac{18}{5}&6&0&0&0&0
\\0&0&0&0&\frac{118}{21}&\frac{50}{7}&0&0
\\0&0&0&0&\frac{50}{7}&10&0&0
\\0&0&0&0&0&0&\frac{115}{12}&\frac{35}{3}
\\0&0&0&0&0&0&\frac{35}{3}&15\end{array}\right).
\end{equation}
If the elements corresponding to the multipole moments $\{\kappa_l\}$ are omitted, then these results reduce to those obtained earlier for irrotational flows.
\section{\label{VIII}Optimization for axisymmetric polar flows}
We impose the constraint that the force exerted on the fluid vanishes at any time. This requires $\kappa_1(t)=0$. With
this constraint the mean swimming velocity
$\overline{U}_2$ and the mean rate of dissipation $\overline{\mathcal{D}}_2$ can be expressed as
\begin{equation}
\label{8.1}\overline{U_2}=\frac{1}{2}\omega
a(\hat{\vc{\psi}}|\hat{\du{B}}|\hat{\vc{\psi}}),\qquad\overline{\mathcal{D}_2}=8\pi\eta\omega^2a^3
(\hat{\vc{\psi}}|\hat{\du{A}}|\hat{\vc{\psi}}).
\end{equation}
Optimization of the mean swimming velocity for given mean rate of dissipation leads to the eigenvalue problem
\begin{equation}
\label{8.2}\hat{\du{B}}|\hat{\vc{\psi}}_\lambda)=\lambda\hat{\du{A}}|\hat{\vc{\psi}}_\lambda).
\end{equation}
The matrix $\hat{\du{B}}$ is pure imaginary and antisymmetric and
the matrix $\hat{\du{A}}$ is real and symmetric. As in the case of
potential flows we truncate at maximum $l$-value $L$. The
truncated matrices $\hat{\du{A}}_{1L}$ and $\hat{\du{B}}_{1L}$ are $2L-1$-dimensional. The
structure of the eigenvalue equations is such that they can be satisfied for
real eigenvalues by eigenvectors with components which are real for odd $l$
and pure imaginary for even $l$. The complex conjugate of an eigenvector
corresponds to the eigenvalue for the opposite sign. Hence it suffices to consider the positive eigenvalues.
In our plots we have chosen the phase of the eigenvectors such that the first potential multipole moment $\mu^c_1$ is real and positive.
With truncation at $l=L$ the eigenvalue problem is equivalent to
that for two coupled linear harmonic chains with masses
corresponding to the diagonalized form of the matrix
$\hat{\du{A}}_{1L}$. However, it is not necessary to perform this
diagonalization explicitly, and it suffices to discuss Eq. (8.2)
directly. It is of interest to consider the $2\times 2$ matrix
along the diagonal direction of the matrix $\du{A}$ for large $l$.
Diagonalization of this matrix shows that one of its eigenvalues
is of order unity, whereas the second one grows as $l^2$ as $l$
increases. For the eigenvector corresponding to the first
eigenvalue the second component is nearly the opposite of the
first, and for the second eigenvalue the two components are nearly
equal. This suggests that the eigenvector with largest eigenvalue
for the problem Eq. (8.2) for large $L$ is a mixture of flows of
potential and viscous type with nearly equal and opposite
amplitudes. This is confirmed by numerical solution of the
eigenvalue problem for a large value of $L$, say $L=40$. If the
optimal eigenvector is decomposed into potential and viscous
components, corresponding to $\mu$- and $\kappa$-moments
respectively,
\begin{equation}
\label{8.3}\hat{\vc{\psi}}_\lambda=\hat{\vc{\psi}}_{\lambda
p}+\hat{\vc{\psi}}_{\lambda v},
\end{equation}
then the norm of the viscous part is nearly equal to the norm of
the potential part.
It turns out that the inclusion of the viscous part has a dramatic
effect on the maximum eigenvalue. In Fig. 8 we show the maximum
eigenvalue as a function of $L$, in analogy with Fig. 1. This
shows that $\lambda_{max}$ tends to a constant larger than 2 for
large $L$.
We prove that the constant equals $2\sqrt{2}$. The inclusion of viscous flows has led to a qualitative change. It is
no longer sufficient to consider the asymptotically uniform linear chain as in Sec. IV. The asymptotic variation of couplings
and masses along two coupled linear harmonic chains must be taken into account. With modified moments as in Eq. (4.8)
\begin{equation}
\label{8.4}f_l=\sqrt{(l+1)(l+2)}\mu_l,\qquad
g_l=\sqrt{(l+1)(l+2)}\kappa_l,
\end{equation}
the $6\times 6$ matrices along the diagonal of the corresponding
matrices $\du{A}'$ and $\du{B}'$ linking the multipoles of order
$l-1,l$ and $l+1$ in the limit of large $l$ take the form
\begin{equation}
\label{8.5}\du{A}'^{(6)}_0=\frac{1}{2}\left(\begin{array}{cccccc}
1&1&0&0&0&0
\\1&1&0&0&0&0
\\0&0&1&1&0&0
\\0&0&1&1&0&0
\\0&0&0&0&1&1
\\0&0&0&0&1&1
\end{array}\right),\qquad\du{B}'^{(6)}_0=\frac{1}{2}\left(\begin{array}{cccccc}
0&0&-i&-i&0&0
\\0&0&-i&-i&0&0
\\i&i&0&0&-i&-i
\\i&i&0&0&-i&-i
\\0&0&i&i&0&0
\\0&0&i&i&0&0
\end{array}\right).
\end{equation}
In comparison the large $l$ behavior of the $3\times 3$ matrices
along the diagonal of the matrices $\du{A}^{c\prime}$ and
$\du{B}^{c\prime}$ of Sec. 4 is given by
\begin{equation}
\label{8.6}\du{A}^{c\prime(3)}_0=\frac{1}{2}\left(\begin{array}{ccc}
1&0&0
\\0&1&0\\0&0&1
\end{array}\right),\qquad\du{B}^{c\prime(3)}_0=\frac{1}{2}\left(\begin{array}{ccc}
0&1&0\\1&0&1
\\0&1&0
\end{array}\right).
\end{equation}
The eigenvalue problem
$\du{B}^{c\prime(3)}_0|\vc{f}^{(3)})=\lambda\du{A}^{c\prime(3)}_0|\vc{f}^{(3)})$
has eigenvalues $\lambda_{0\pm}=\pm\sqrt{2}$, $\lambda_{00}=0$, and
the eigenvalue problem
$\du{B}'^{(6)}_0|\vc{f}^{(6)})=\lambda\du{A}'^{(6)}_0|\vc{f}^{(6)})$
has the same eigenvalues, each twofold degenerate. However, the result is unstable under small perturbations, and the higher order terms
of the matrix elements to order $1/l^2$ must be considered to
obtain the correct result corresponding to the coupled linear chains.
Thus instead of Eq. (8.5) we consider the asymptotic behavior
\begin{eqnarray}
\label{8.7}\du{A}'^{(6)}(l)&=&\du{A}'^{(6)}_0+\du{A}'^{(6)}_1\frac{1}{l}+\du{A}'^{(6)}_2\frac{1}{l^2}
+O\big(\frac{1}{l^3}\big),\nonumber\\
\du{B}'^{(6)}(l)&=&\du{B}'^{6)}_0+\du{B}'^{(6)}_1\frac{1}{l}+\du{B}'^{(6)}_2\frac{1}{l^2}
+O\big(\frac{1}{l^3}\big).
\end{eqnarray}
From Eqs. (7.8) and (7.15) one finds that the matrices
$\du{A}'^{(6)}_1$ and $\du{B}'^{(6)}_1$ are given by
\begin{eqnarray}
\label{8.8}\du{A}'^{(6)}_1&=&\frac{1}{2}\left(\begin{array}{cccccc}
-2&-1&0&0&0&0
\\-1&0&0&0&0&0
\\0&0&-2&-1&0&0
\\0&0&-1&0&0&0
\\0&0&0&0&-2&-1
\\0&0&0&0&-1&0
\end{array}\right),\nonumber\\
\du{B}'^{(6)}_1&=&\frac{-i}{2}\left(\begin{array}{cccccc}
0&0&-5&-3&0&0
\\0&0&-3&-1&0&0
\\5&3&0&0&-5&-3
\\3&1&0&0&-3&-1
\\0&0&5&3&0&0
\\0&0&3&1&0&0
\end{array}\right).
\end{eqnarray}
The matrices $\du{A}'^{(6)}_2$ and $\du{B}'^{(6)}_2$ are given by
\begin{eqnarray}
\label{8.9}\du{A}'^{(6)}_2&=&\frac{1}{4}\left(\begin{array}{cccccc}
2&-1&0&0&0&0
\\-1&0&0&0&0&0
\\0&0&6&1&0&0
\\0&0&1&0&0&0
\\0&0&0&0&10&3
\\0&0&0&0&3&0
\end{array}\right),\nonumber\\
\du{B}'^{(6)}_2&=&\frac{-i}{4}\left(\begin{array}{cccccc}
0&0&19&9&0&0
\\0&0&9&3&0&0
\\-19&-9&0&0&29&15
\\-9&-3&0&0&15&5
\\0&0&-29&-15&0&0
\\0&0&-15&-5&0&0
\end{array}\right).
\end{eqnarray}
From the eigenvalue equation
$|\du{B}'^{(6)}_0+z\du{B}'^{(6)}_1-\lambda_1(\du{A}'^{(6)}_0+z\du{A}'^{(6)}_1)|=0$ one finds that in the
limit $z\rightarrow 0$ the eigenvalues tend to $\lambda_{1\pm}=\pm 2\sqrt{2},\lambda_{10}=0$, each twofold degenerate.
From the eigenvalue equation
$|\du{B}'^{(6)}_0+z\du{B}'^{(6)}_1+z^2\du{B}'^{(6)}_2-\lambda_2(\du{A}'^{(6)}_0+z\du{A}'^{(6)}_1+z^2\du{A}'^{(6)}_2)|=0$ one finds that in the
limit $z\rightarrow 0$ the eigenvalues tend to $\lambda_{2\pm}=\pm 2,\lambda_{20}=0$, each twofold degenerate. The largest eigenvalue $\lambda_{2+}=2$ is a factor $\sqrt{2}$ larger than $\lambda_{0+}=\sqrt{2}$ given below Eq. (8.6). Hence for the complete problem with matrices $\du{B}_{1L}$ and $\du{A}_{1L}$ the maximum eigenvalue in the limit $L\rightarrow\infty$ is a factor $\sqrt{2}$ larger than obtained from the linear chain problem for potential flows of Sec. IV. The maximum eigenvalue for the present problem therefore tends to $2\sqrt{2}$ in the limit $L\rightarrow\infty$, as suggested by Fig. 8.
Thus with the inclusion of $\sigma=0$ modes the efficiency of translational swimming defined in Eq. (5.1) takes the maximum value
\begin{equation}
\label{8.10}E_{T\mathrm{max}}=\frac{1}{\pi\sqrt{2}}.
\end{equation}
As in the case of potential swimming the optimum value is reached for a set of multipoles decaying in absolute magnitude as $1/l$ at large $l$. This suggests that the maximization of $E_T$ leads to an optimum stroke which is not of physical relevance.
\section{\label{IX}Speed and power}
We denote the eigenvector with maximum eigenvalue corresponding to the truncated matrices $\hat{\du{A}}_{1L}$ and $\hat{\du{B}}_{1L}$ as $\vc{g}_{1L}$ with normalization $(\vc{g}_{1L}|\vc{g}_{1L})=1$. As in Sec. V we look for a different selection criterion for optimization of the stroke.
For the more general axisymmetric flow patterns we find again that for the eigenvector $\vc{g}_{1L}$ the displacement $\vc{\xi}(t)$ at $\theta=\pi/2$ describes an ellipse in the $zx$ plane given by Eq. (5.6), but now with modified expressions for the coefficients $A(1,L)$ and $B(1,L)$. More generally we consider arbitrary values of $\theta$. We then find that in general the vector $\vc{\xi}(t)$ describes an ellipse in the $zx$ plane which is tilted with respect to the $z$ axis. The shape and tilt of the ellipse are described conveniently by Stokes parameters \cite{17}.
The components $\xi_z(\theta,t)$ and $\xi_x(\theta,t)$ can be expressed as
\begin{equation}
\label{9.1}\xi_z(\theta,t)=\mathrm{Im}\big(\alpha_L(\theta) e^{-i\omega t}\big)a,\qquad\xi_x(\theta,t)=\mathrm{Im}\big(\beta_L(\theta) e^{-i\omega t}\big)a,
\end{equation}
with complex amplitudes $\alpha_L(\theta)$ and $\beta_L(\theta)$ given by
\begin{eqnarray}
\label{9.2}\alpha_L(\theta)&=&q_L(\theta)\cos\theta-p_L(\theta)\sin\theta,\nonumber\\
\beta_L(\theta)&=&q_L(\theta)\sin\theta+p_L(\theta)\cos\theta,
\end{eqnarray}
where
\begin{eqnarray}
\label{9.2}p_L(\theta)&=&\sum^L_{l=1}g_{1L,2l-1}P^1_l(\cos\theta)+\sum^{L-1}_{l=1}g_{1L,2l}\frac{l-1}{l+1}P^1_{l+1}(\cos\theta),\nonumber\\
q_L(\theta)&=&\sum^L_{l=1}g_{1L,2l-1}(l+1)P_l(\cos\theta)+\sum^{L-1}_{l=1}g_{1L,2l}(l+2)P_{l+1}(\cos\theta).
\end{eqnarray}
The Stokes parameters of the ellipse at polar angle $\theta$ are defined by \cite{17}
\begin{eqnarray}
\label{9.3}I_S&=&|\alpha|^2+|\beta|^2,\qquad Q_S=|\alpha|^2-|\beta|^2,\qquad \delta=\mathrm{arg}\frac{\alpha}{\beta},\nonumber\\
U_S&=&2|\alpha||\beta|\cos\delta,\qquad V_S=2|\alpha||\beta|\sin\delta,
\end{eqnarray}
where for brevity we have omitted the subscript $L$ and the variable $\theta$.
The tilt angle of the ellipse is given by
\begin{equation}
\label{9.4}\gamma_S=\frac{1}{2}\arctan\frac{U_S}{Q_S},
\end{equation}
and the ellipticity $\varepsilon_S$ follows from
\begin{equation}
\label{9.5}\eta_S=\frac{1}{2}\arctan\frac{V_S}{\sqrt{Q_S^2+V_S^2}},\qquad\varepsilon_S=|\tan\eta_S|.
\end{equation}
The long and short semi-axis of the ellipse are
\begin{equation}
\label{9.6}P(1,L,\theta)=\sqrt{\frac{I_S}{1+\varepsilon_S^2}},\qquad Q(1,L,\theta)=\varepsilon_SP(1,L,\theta).
\end{equation}
We find for each $L$ that the ellipse described by $\vc{\xi}(t)$ at $\theta=\pi/2$ for the stroke with maximum efficiency $E_T$ has its long axis parallel to the $z$ axis. Thus if we represent the ellipse again by Eq. (5.6) then for multipoles given by $\vc{g}_{1L}/B(1,L)$ the ellipse described by $\vc{\xi}(t)$ will have horizontal semi-axis $a$ and vertical semi-axis $b=A(1,L)a/B(1,L)$. We therefore consider the reduced speed and power at fixed horizontal amplitude of stroke,
\begin{equation}
\label{9.7}\hat{U}^B_{1L}=\frac{\hat{U}_{1L}}{B(1,L)^2},\qquad\hat{\mathcal{D}}^B_{1L}
=\frac{\hat{\mathcal{D}}_{1L}}{B(1,L)^2},
\end{equation}
with
\begin{equation}
\label{9.8}B(1,L)=P(1,L,\frac{\pi}{2}).
\end{equation}
In Fig. 9 we
show the plot of $\hat{U}^B_{1L}$ for the optimal eigenvector as a
function of $L$, and in Fig. 10 we show the corresponding plot for
the reduced power $\hat{D}^B_{1L}$. The reduced power shows again
a minimum, this time at $L=7$, given by $\hat{D}^B_{17}=1.529$. At $L=7$ the reduced amplitude is $B(1,7)=2.945$, and the reduced speed is $\hat{U}^B_{17}=3.303$. In Fig. 11 we plot the absolute values of the set of multipole moments $\{\kappa_l,\mu_l\}$ for the optimal eigenvector with $L=7$.
For the set of complex multipoles $\hat{\vc{\psi}}(1,L)=\varepsilon\vc{g}_{1L}/B(1,L)$ the mean speed and rate of dissipation are
\begin{equation}
\label{9.9}\overline{U}_2=\frac{1}{2}\omega a\varepsilon^2\hat{U}^B_{1L},\qquad\overline{\mathcal{D}}_2
=8\pi\eta\omega^2 a^3\varepsilon^2\hat{\mathcal{D}}^B_{1L}.
\end{equation}
Performing the same estimate as at the end of Sec. V for the more general class of flows with the optimum stroke for $L=7$ we find power $P=3.84\times 10^{-15}\varepsilon^2$ watt and speed $U=1.65\times 10^{-3}\varepsilon^2$ m/sec. The efficiency is $E_T=0.172$, compared with the maximum possible for general flow $E_{T\mathrm{max}}=1/(\pi\sqrt{2})=0.225$. For $P=2.09\times 10^{-17}$ watt we find relative amplitude $\varepsilon=0.074$ and speed $U=9.0\times 10^{-6}$ m/sec.
The nature of the optimum stroke for $L=7$ is shown in Fig. 12, in analogy to Fig. 5. The time-dependent swimming velocity $U_2(t)$ and rate of dissipation $D_2(t)$ can be evaluated in analogy to Eqs. (6.1) and (6.2). The dimensionless ratios $U_2(t)/\overline{U}_2$ and $D_2(t)/\overline{D}_2$ for the optimal stroke with $L=7$ vary in time quite similarly to the behavior shown in Fig. 7. Again the ratio $D_2(t)/\overline{D}_2$ equals unity within numerical accuracy.
\section{\label{X}Discussion}
Basing ourselves on the Stokes equations, rather than the linearized Navier-Stokes equations, we have developed a simpler discussion of the swimming of a sphere at low Reynolds number with the restriction to potential flow solutions than was presented before \cite{8}. The identities Eqs. (3.11) and (3.16) play a crucial role. They imply that the representation of the flow in terms of electrostatic multipole potentials is particularly simple. In this representation the matrix $\du{A}^c$, from which the rate of dissipation is calculated, is diagonal, and the matrix $\du{B}^c$, from which the swimming velocity is calculated, is tri-diagonal. Correspondingly, the eigenvalue problem which yields the swimming stroke of maximum efficiency, is relatively simple.
Subsequently we have extended the derivation to the complete set of axisymmetric polar solutions of the Stokes equations. An additional set of multipole moments corresponding to flows with vorticity needs to be introduced. Although this leads to a doubling of dimensionality, the structure of the eigenvalue problem in the chosen representation remains fairly simple.
The additional flow solutions allow a considerable enhancement of efficiency, defined as the dimensionless ratio of speed and power. As in the case of irrotational flow, the maximum efficiency is attained for a stroke characterized by multipoles with a significant weight at high order. This indicates that the efficiency is not the most suitable measure of swimming performance.
Therefore we have considered a measure of performance based on a comparison of energy consumption for strokes with the same amplitude. The measure allows selection of a stroke with minimum energy consumption in a class of possible strokes. The optimal stroke selected in this manner involves multipoles of relatively low order and is expected to be of physical interest.
Although the spherical geometry provides only a crude approximation to the shape of most microorganisms, it has the advantage that the mechanism of swimming can be analyzed in great detail. The analysis shows that it is worthwhile to consider various measures of swimming performance. The mathematical formalism may serve as a guide in the study of more complicated geometry, such as a spheroid or an ellipsoid.
\newpage
|
\section{Introduction}
Compressed sensing (CS) studies sparse signal recovery from far fewer measurements and has brought significant impact on signal processing and information theory in the past decade. Since its development thus far has been focused on signals that can be sparsely represented under a finite discrete dictionary, limitations are present in applications such as array processing, radar and sonar, where the dictionary is typically specified by one or more continuous parameters.
In this paper, we consider the problem of recovering a sinusoidal/frequency-sparse signal, which is a superposition of a few complex sinusoids, from a subset of regularly spaced samples. The problem is referred to as continuous CS as suggested in \cite{tang2012compressed} in the sense that the frequencies of the sinusoids can take any continuous values. A systematic, convex approach is introduced in \cite{tang2012compressed} which works directly in the continuous domain and completely eliminates parameter discretization/gridding of conventional CS methods that causes basis mismatches. In the paper, it is shown that a frequency-sparse signal can be exactly recovered in the noiseless case from far fewer samples provided that the frequencies are appropriately separate. Practical solutions are then provided in \cite{yang2014gridless} in the noisy case. An important, related application is direction of arrival (DOA) estimation \cite{krim1996two}, in which improved performance is typically obtained by receiving multiple frequency-sparse signals in a time interval which share the same frequency components. Note that the method in \cite{tang2012compressed} specified for a single measurement vector (SMV) cannot process the multiple measurement vectors (MMVs) at a single step while the joint processing exploiting the so-called joint sparsity can usually improve the performance \cite{eldar2010average}. Before this paper, the only known continuous/gridless sparse method for MMVs was presented in \cite{yang2014discretization} in the context of DOA estimation based on statistical inference.
In this paper, we study the SMV and MMV continuous CS problems in a unified framework. Based on an $\ell_0$-norm-like formulation we establish a link between continuous CS and a well-studied area of low rank matrix completion (LRMC) \cite{candes2009exact} and provide a sufficient condition for exact recovery. We propose convex optimization methods for signal recovery based on the link and convex and nonconvex relaxations and present computationally efficient algorithms using alternating direction method of multipliers (ADMM) \cite{boyd2011distributed}. Numerical simulations are provided to study their phase transition phenomena and validate their usefulness in DOA estimation.
\section{Signal Recovery via Atomic $\ell_0$ Norm Minimization}
\subsection{Atomic $\ell_0$ Norm Minimization}
Suppose that we observe a number of $L$ sinusoidal signals
\equ{y_{jt}^o=\sum_{k=1}^K s_{kt}e^{i2\pi (j-1) f_k}, \quad \sbra{j,t}\in \mbra{N}\times\mbra{L}, \label{formu:observmodel1}}
denoted by matrix $\m{Y}^o=\mbra{y_{jt}^o}\in\mathbb{C}^{N\times L}$, on the index set $\m{\Omega}\times\mbra{L}$, where $\m{\Omega} \subset \mbra{N}\triangleq\lbra{1,\cdots,N}$ with $M\triangleq\abs{\m{\Omega}}\leq N$ denoting the sample size of each sinusoidal signal.
Here $(j,t)$ indexes the $j$th entry of the $t$th measurement vector (or snapshot data), $i=\sqrt{-1}$, $f_k\in\left[0,1\right)$ denotes the $k$th normalized frequency, and $s_{kt}\in\mathbb{C}$ is the (complex) amplitude of the $k$th component at snapshot $t$. The SMV case where $L=1$ corresponds to line spectral estimation in spectral analysis and the MMV case is common in array processing. In this paper, each column of $\m{Y}^o$ is called a frequency-sparse signal since the number of sinusoids $K$ is typically small. We are interested in the recovery of $\m{Y}^o$ (as well as the parameters $\m{f}$ and $\m{s}$ if possible) under the sparse prior given its partial or compressive measurements on $\m{\Omega}\times\mbra{L}$, denoted by $\m{Y}_{\m{\Omega}}^o$. This problem is called the continuous CS problem to distinguish with the common discrete frequency setting as suggested in \cite{tang2012compressed}. We mainly consider the noiseless case. The general noisy case will be deferred to Subsection \ref{sec:noisycase}.
We exploit sparsity to solve the ill-posed problem of recovering $\m{Y}^o$ from $\m{Y}_{\m{\Omega}}^o$. Following the literature of CS, we seek for the maximally sparse candidate for its recovery. To state it formally, we denote $\m{a}\sbra{f}=\mbra{1,e^{i2\pi f},\cdots,e^{i2\pi\sbra{N-1}f}}^T\in\mathbb{C}^{N}$ and $\m{s}_k=\mbra{s_{k1},\cdots,s_{kL}}\in\mathbb{C}^{1\times L}$. Then (\ref{formu:observmodel1}) can be written as
\equ{\m{Y}^o=\sum_{k=1}^K \m{a}\sbra{f_k}\m{s}_k=\sum_{k=1}^K c_k\m{a}\sbra{f_k}\m{\phi}_k, \label{formu:atomdecomp}}
where $c_k=\twon{\m{s}_k}>0$ and $\m{\phi}_k=c_k^{-1}\m{s}_k$ with $\twon{\m{\phi}_k}=1$. Let $\mathbb{S}^{2L-1}=\lbra{\m{\phi}: \m{\phi}\in\mathbb{C}^{1\times L}, \twon{\m{\phi}}=1}$ denote the unit $2L-1$-sphere. We define the continuous dictionary or the set of atoms
\equ{\mathcal{A}\triangleq \lbra{\m{a}\sbra{f,\m{\phi}}=\m{a}\sbra{f}\m{\phi}: f\in\left[0,1\right), \m{\phi}\in\mathbb{S}^{2L-1}}.\label{formu:atomset}}
It is clear that $\m{Y}^o$ is a linear combination of a number of atoms in $\mathcal{A}$. We define the atomic $\ell_0$ (pseudo-)norm of some $\m{Y}\in\mathbb{C}^{N\times L}$ as the smallest number of atoms that can express it:
\equ{\norm{\m{Y}}_{\mathcal{A},0}
=\inf\lbra{\mathcal{K}: \m{Y}=\sum_{k=1}^{\mathcal{K}} c_k\m{a}_k, \m{a}_k\in\mathcal{A}, c_k>0}. \label{formu:AL0}}
So we propose the following problem for signal recovery:
\equ{\min_{\m{Y}} \norm{\m{Y}}_{\mathcal{A},0}, \text{ subject to } \m{Y}_{\m{\Omega}}=\m{Y}^o_{\m{\Omega}}, \label{formu:AL0min}}
where $\m{Y}_{\m{\Omega}}$ takes the rows of $\m{Y}$ indexed by $\m{\Omega}$.
\subsection{Spark of Continuous Dictionary}
To analyze the atomic $\ell_0$ norm minimization problem in (\ref{formu:AL0min}), we generalize the concept of spark to the case of continuous dictionary. We define the following continuous dictionary with respect to the index set $\m{\Omega}$:
$\mathcal{A}_{\m{\Omega}}^1\triangleq \lbra{\m{a}_{\m{\Omega}}\sbra{f}: f\in\left[0,1\right)}$.
\begin{defi}[Spark of continuous dictionary] Given the continuous dictionary $\mathcal{A}_{\m{\Omega}}^1$, the quantity spark of $\mathcal{A}_{\m{\Omega}}^1$, denoted by $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$, is the smallest number of atoms of $\mathcal{A}_{\m{\Omega}}^1$ which are linearly dependent.
\end{defi}
\begin{thm} We have the following results about $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$:
\begin{enumerate}
\item $2\leq\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}\leq M+1$,
\item $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}= 2$ if and only if the elements of
\equ{\m{\mathcal{D}}\triangleq\lbra{m_1-m_2: m_1,m_2\in\m{\Omega},m_1\geq m_2}}
is not coprime, and
\item $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}= M+1$ if $\m{\Omega}$ consists of $M$ consecutive integers.
\end{enumerate}
\label{thm:fullspark}
\end{thm}
Theorem \ref{thm:fullspark} presents the range of $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$ with respect to the sampling index set $\m{\Omega}$. Readers are referred to \cite{yang2014exact} for its proof and those of the rest results due to the page limit. A sufficient and necessary condition is provided under which $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$ achieves the lower bound 2. Note, however, that such $\m{\Omega}$ is rare. For example, when $\m{\Omega}$ is selected uniformly at random, the probability that the condition holds is 0 whenever $M>\frac{N}{2}$. It is less than $1.2\times10^{-3},\, 1.8\times10^{-7},\, 3.2\times10^{-12}$ when $N=100$ and $M=10,\,20,\,30$ respectively. A sufficient (but unnecessary) condition is also provided under which $\mathcal{A}_{\m{\Omega}}^1$ achieves the upper bound $M+1$.
\subsection{Sufficient Condition for Exact Recovery} \label{sec:l0exactrecovery}
We provide theoretical guarantees of the atomic $\ell_0$ norm minimization in (\ref{formu:AL0min}) for frequency recovery in this subsection. In particular, we have the following result, which can be considered as a continuous version of \cite[Theorem 2.4]{chen2006theoretical}.
\begin{thm} $\m{Y}^o=\sum_{k=1}^K c_k\m{a}\sbra{f_k,\m{\phi}_k}$ is the unique optimizer to (\ref{formu:AL0min}) if
\equ{K< \frac{\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}-1+\rank \sbra{\m{Y}_{\m{\Omega}}^o}}{2}. \label{Kbound}}
Moreover, the atomic decomposition above is the unique one satisfying that $K=\norm{\m{Y}^o}_{\mathcal{A},0}$.
\label{thm:AL0_guanrantee}
\end{thm}
Theorem \ref{thm:AL0_guanrantee} shows that the proposed atomic $\ell_0$ minimization problem can recover a frequency-sparse signal with sparsity $K< \frac{1}{2}\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$ in the SMV case. As we take more measurement vectors, we have a chance to recover more complex signals by increasing $\rank \sbra{\m{Y}_{\m{\Omega}}^o}$, which is practically relevant in array processing applications. In fact, the sparsity $K$ can be as large as $M-1$ with an approximate choice of $\m{\Omega}$.
\subsection{Finite Dimensional Characterization via Rank Minimization}
The optimization problem in (\ref{formu:AL0min}) is computationally infeasible given the infinite dimensional formulation of the atomic $\ell_0$ norm in (\ref{formu:AL0}). We provide a finite dimensional formulation in the following result.
\begin{thm}
$\norm{\m{Y}}_{\mathcal{A},0}$ defined in (\ref{formu:AL0}) equals the optimal value of the following rank minimization problem:
\equ{\min_{\m{W}\in\mathbb{C}^{L\times L},\m{u}\in\mathbb{C}^{N},\m{U}\geq\m{0}} \rank\sbra{\m{U}}, \text{ subject to } \m{U}=\begin{bmatrix}\m{W} & \m{Y}^H \\ \m{Y} & T\sbra{\m{u}}\end{bmatrix}, \label{formu:AL0_rankmin}}
where $T\sbra{\m{u}}$ denotes a (Hermitian) Toeplitz matrix with its first row specified by $\m{u}^T$, and $\m{U}\geq\m{0}$ means that $\m{U}$ is positive semidefinite.
\label{thm:AL0_rankmin}
\end{thm}
Theorem \ref{thm:AL0_rankmin} presents a rank minimization problem to characterize the atomic $\ell_0$ norm. It follows that (\ref{formu:AL0min}) is equivalent to the following LRMC problem:
\equ{\begin{split}
&\min_{\m{Y},\m{W},\m{u},\m{U}\geq\m{0}} \rank\sbra{\m{U}},\\
&\text{ subject to } \m{U}=\begin{bmatrix}\m{W} & \m{Y}^H \\ \m{Y} & T\sbra{\m{u}}\end{bmatrix}, \m{Y}_{\m{\Omega}}=\m{Y}^o_{\m{\Omega}},\end{split} \label{formu:rankmin_CCS}}
where we need to recover a (structured positive semidefinite) low rank matrix $\m{U}$ with partial access to its entries. As a result, we establish a link between continuous CS and the well studied area of LRMC, which enables us to study the continuous CS problem by borrowing ideas in LRMC. Note that a similar rank minimization problem is presented in \cite{tang2012compressed} in the SMV case, where the rank is put on the matrix $T\sbra{\m{u}}$ rather than the full matrix $\m{U}$ in (\ref{formu:rankmin_CCS}). This difference obscures the link between continuous CS and LRMC which, we will see later, plays an important role in this paper.
\section{Signal Recovery via Relaxations}
\subsection{Convex Relaxations} \label{sec:convexrelax}
The atomic $\ell_0$ norm exploits sparsity directly, however, it is nonconvex and the problem in (\ref{formu:rankmin_CCS}) cannot be solved globally in practice. To avoid the nonconvexity and at the same time exploit sparsity, we utilize convex relaxation to relax the atomic $\ell_0$ norm. In particular, it can be relaxed in two ways from two different perspectives. One is to relax the atomic $\ell_0$ norm to the atomic $\ell_1$ norm (or simply the atomic norm) which is defined as the gauge function of $\text{conv}\sbra{\mathcal{A}}$, the convex hull of $\mathcal{A}$ \cite{chandrasekaran2012convex}:
\equ{\begin{split}\atomn{\m{Y}}
&\triangleq\inf\lbra{t>0: \m{Y}\in t\text{conv}\sbra{\mathcal{A}}} \\
&= \inf\lbra{\sum_k c_k: \m{Y}=\sum_k c_k\m{a}_k, c_k\geq0, \m{a}_k\in\mathcal{A}}.\end{split} \label{formu:atomicnorm}}
The atomic norm $\atomn{\cdot}$ is indeed a norm and convex by the property of the gauge function. The other way of convex relaxation is based on a perspective of rank minimization illustrated in (\ref{formu:AL0_rankmin}) and to relax the pseudo rank norm to the nuclear norm or equivalently the trace norm for a positive semidefinite matrix, i.e., to replace $\rank\sbra{\m{U}}$ by $\tr\sbra{\m{U}}$ in (\ref{formu:AL0_rankmin}). Interestingly enough, the two convex relaxations are equivalent, which is shown in the following result.
\begin{thm} $\norm{\m{Y}}_{\mathcal{A}}$ defined in (\ref{formu:atomicnorm}) equals the optimal value of the following semidefinite programming (SDP):
\equ{\min_{\m{W},\m{u},\m{U}\geq\m{0}} \frac{1}{2\sqrt{N}}\tr\sbra{\m{U}}, \text{ subject to } \m{U}=\begin{bmatrix}\m{W} & \m{Y}^H \\ \m{Y} & T\sbra{\m{u}}\end{bmatrix}. \label{formu:AN_SDP}} \label{thm:AN_SDP}
\end{thm}
Theorem \ref{thm:AN_SDP} generalizes the results in \cite{candes2013towards,tang2012compressed} on the SMV case. Consequently, we propose the following atomic norm minimization problem for signal recovery:
\equ{\begin{split}
&\min_{\m{Y},\m{W},\m{u},\m{U}\geq\m{0}} \tr\sbra{\m{U}},\\
&\text{ subject to } \m{U}=\begin{bmatrix}\m{W} & \m{Y}^H \\ \m{Y} & T\sbra{\m{u}}\end{bmatrix}, \m{Y}_{\m{\Omega}}=\m{Y}^o_{\m{\Omega}}.\end{split} \label{formu:tracemin_CCS}}
It is shown in \cite{tang2012compressed} that $\m{Y}^o$ can be exactly recovered in the SMV case with high probability if $M\geq O\sbra{K\log K\log N}$ and the frequencies are separate by at least $\frac{4}{N}$. Note that the condition of frequency separation is introduced by the convex relaxation while it is not required in the atomic $\ell_0$ minimization problem as shown in Theorem \ref{thm:AL0_guanrantee}. After submission of this paper, we have proven in \cite{yang2014exact} that $\m{Y}^o$ can be recovered by (\ref{formu:tracemin_CCS}) in the MMV case under similar conditions.
\subsection{Iterative Reweighted Optimization via Nonconvex Relaxation} \label{sec:nonconvexrelax}
From the perspective of rank minimization, an iterative reweighted trace norm minimization scheme can be implemented to further improve the low-rankness by iteratively minimizing the objective function $\tr\mbra{\sbra{\m{U}_{j-1}+\epsilon\m{I}}^{-1}\m{U}}$ subject to the same constraints, where $\m{U}_{j}$ denotes the solution at the $j$th iteration starting with $\m{U}_0=\m{I}$, and $\epsilon>0$ is a small number. Obviously, the first iteration refers exactly to the convex relaxation. This iterative reweighted optimization scheme corresponds to relaxing $\rank\sbra{\m{U}}$ to the nonconvex objective $\ln\abs{\m{U}+\epsilon\m{I}}$, followed by a majorization-maximization (MM) implementation of the nonconvex optimization which guarantees local convergence of the objective function. We expect that a weaker condition of frequency separation holds for this nonconvex relaxation compared to the convex one since intuitively $\ln\abs{\cdot}$ is a closer approximation of the rank function. In the future, we can also consider other relaxation methods based on the literature of LRMC.
\subsection{Frequency and Amplitude Retrieval}
Given the solution of $\m{Y}$, it is of great importance to retrieve the frequency and amplitude solutions, or equivalently to obtain the atomic decomposition as in (\ref{formu:atomdecomp}), in applications such as line spectral estimation or DOA estimation. In particular, we can firstly obtain the frequency solution by the Vandermonde decomposition of $T\sbra{\m{u}}$ given the solution of $\m{u}$ (see details in \cite{yang2014gridless,yang2014exact}). Then the amplitude can be easily obtained by solving the linear system of equations in (\ref{formu:atomdecomp}).
\subsection{The Noisy Case} \label{sec:noisycase}
Noise is always present in practical scenarios. In this paper we consider only noise with bounded energy. Suppose the noise in the measurements $\m{Y}^o_{\m{\Omega}}$ is upper bounded by $\eta>0$ in the Frobenius norm. Then we can impose the inequality constraint $\frobn{\m{Y}_{\m{\Omega}}-\m{Y}^o_{\m{\Omega}}}\leq\eta$ instead of the equality constraint $\m{Y}_{\m{\Omega}}=\m{Y}^o_{\m{\Omega}}$ in the noiseless case. Note that the latter is a special case with $\eta=0$.
\subsection{Computationally Efficient Algorithms via ADMM}
We present a first-order algorithm based on ADMM to solve the trace minimization problems $\min\tr\sbra{\m{B}\m{U}}$ in Subsections \ref{sec:convexrelax} and \ref{sec:nonconvexrelax}, in particular,
\equ{\begin{split}
&\min_{\m{Y},\m{W},\m{u},\m{U}\geq\m{0}}\tr\sbra{\m{B}_{1}\m{W}} +\tr\sbra{\m{B}_{3}T\sbra{\m{u}}} +\tr\sbra{\m{B}_{2}^H\m{Y}+\m{Y}^H\m{B}_{2}},\\
&\text{ subject to } \m{U}=\begin{bmatrix}\m{W} & \m{Y}^H \\ \m{Y} & T\sbra{\m{u}}\end{bmatrix} \text{ and } \frobn{\m{Y}_{\m{\Omega}}-\m{Y}^o_{\m{\Omega}}}\leq\eta,\end{split} \label{formu:trmin_ADMM}}
where $\m{B}\triangleq\begin{bmatrix}\m{B}_1 &\m{B}_2^H \\\m{B}_2& \m{B}_3\end{bmatrix}\geq\m{0}$ is partitioned as $\m{U}$. (\ref{formu:trmin_ADMM}) can be solved within the framework of ADMM in \cite{boyd2011distributed}, where $\sbra{\m{Y},\m{W},\m{u}}$, $\m{U}$ and the Lagrangian multiplier are iteratively updated with closed-form expressions and converge to the optimal solution (see, e.g., \cite{yang2014gridless}). An eigen-decomposition of a Hermitian matrix of order $N+L$ is required at each iteration. We omit the detailed update rules due to the page limit.
Note that the ADMM converges slowly to an extremely accurate solution while moderate accuracy is typically sufficient in practical applications \cite{boyd2011distributed}.
\section{Numerical Simulations}
We first consider the noiseless case and study the so-called phase transition phenomenon in the $\sbra{M,K}$ plane. In particular, we repeat an experiment in \cite{tang2012compressed} and consider our proposed atomic norm (or trace norm) minimization (ANM) and reweighted trace minimization (RWTM) methods. RWTM is terminated within maximally 3 iterations in our simulation. For achieving high accuracy we solve the SDPs (in fact, their dual problems, see \cite{yang2014exact}) using a standard SDP solver, SDPT3 \cite{toh1999sdpt3}. We fix $N=128$ and vary $M=8,12,\dots,120$ and $K=2,4,\dots,M$. We consider $L=1$ and $5$, where ANM in the SMV case has been studied in \cite{tang2012compressed}. The frequencies $f_k$ are generated randomly with minimal separation $\Delta_f\geq1/N$ which is empirically found in \cite{tang2012compressed} to be the minimal separation required for exact recovery when $L=1$. The amplitudes $s_k\sbra{t}$ are randomly generated as $0.5+w^2$ with random phases, where $w$ is standard normal distributed. The first column of $\m{Y}^o$ is used in the SMV case in each problem generated. The recovery is considered successful if $\frobn{\widehat{\m{Y}}-\m{Y}^o}/\frobn{\m{Y}^o}< 10^{-6}$, where $\widehat{\m{Y}}$ denotes the recovered signal.\footnote{After submission of this paper, we find that this criterion is not strict enough to guarantee exact recovery of the frequencies. See more simulation results in \cite{yang2014exact}.} Simulation results are presented in Fig. \ref{Fig:phasetransition}, where a transition from perfect recovery to complete failure can be observed in each subfigure. By increasing the number of measurement vectors from 1 to 5, the phase of successful recovery is enlarged significantly for both ANM and RWTM. Moreover, RWTM has an enlarged success phase than ANM, especially in the MMV case, due to adoption of nonconvex relaxation. We also notice that the transition boundary of ANM with $L=1$ is not very sharp and failures happen in the area where complete success is expected. Further examination reveals that most of the failures happen when the minimal separation marginally exceeds $1/N$. The situation is better for RWTM with $L=1$. In contrast, sharp phase transitions exhibit for both ANM and RWTM in the MMV case. This implies that the requirement of frequency separation can be relaxed in our considered MMV case where the measurement vectors are statistically independent.
We also plot the line $K=\frac{1}{2}\sbra{M+L}$ in each subfigure which acts as an upper bound of the sufficient condition in Theorem \ref{thm:AL0_guanrantee} for exact recovery using the atomic $\ell_0$ norm minimization. We see that in the MMV case successful recoveries can be obtained even above the line with ANM or RWTM, implying that the sufficient condition is unnecessary.
\begin{figure}
\centering
\includegraphics[width=3.4in]{Fig_phasetransition_sdpt3.pdf}
\caption{Phase transition with minimal frequency separation $\Delta_f\geq\frac{1}{N}$. White means complete success while black means complete failure. The straight lines correspond to $K=\frac{1}{2}\sbra{M+L}$.}
\label{Fig:phasetransition}
\end{figure}
We next consider an example of DOA estimation using a redundancy sparse linear array (SLA) of size $M=10$ with $\m{\Omega}=[1,2, 7,11, 24, 27, 35, 42, 54, 56]^T$ (see, e.g., \cite{yang2014discretization}). By assuming narrowband sources the DOA estimation problem is mathematically equivalent to frequency recovery in continuous CS. We apply the proposed ANM and RWTM methods with the ADMM implementations to the DOA estimation and compare with MUSIC. In the simulation, we consider $K=3$ sources impinging on the array from directions corresponding to frequencies $\m{f}=[0.1, 0.106, 0.3]^T$ with powers 1, 1, and 0.25. Suppose that $L=10$ snapshots (or measurement vectors) are observed with the signal to noise ratio $\text{SNR}=14.2$dB. The DOA estimation results of ANM and RWTM are presented in Fig. \ref{Fig:comparewithMUSIC} compared to MUSIC. It is shown that both ANM and RWTM can separate the first two sources while MUSIC cannot. Note also that ANM produces a few spurious sources with very small powers while RWTM detects exactly 3 sources. Both ANM and RWTM take about 2 seconds (loose convergence criteria are adopted in the first few iterations of RWTM for speed acceleration). Finally, it is worth noting that ANM and RWTM require the knowledge of the noise level while MUSIC needs to know the source number.
\begin{figure}
\centering
\includegraphics[width=3in]{Fig_comparewithMUSIC.pdf}
\caption{DOA estimation using ANM and RWTM compared to MUSIC (shown only on the frequency interval $[0,0.35]$).}
\label{Fig:comparewithMUSIC}
\end{figure}
\section{Conclusion}
In this paper, the SMV and MMV continuous CS problems were studied in a unified framework and linked to low rank matrix completion. We extended existing discrete CS results to the continuous case, introduced computationally efficient algorithms and validated their performances via simulations. We have recently analyzed the proposed atomic norm minimization method in \cite{yang2014exact}, which generalizes \cite{candes2013towards} and \cite{tang2012compressed}.
|
\section{Introduction}
\vspace{-5pt}
The Graph Coloring Problem (GCP) is a well known $NP$-Hard problem which has received many attention from the scientific community because of its large range of real applications and computational difficulty.
Given a simple graph $G = (V, E)$ a $k$-\emph{coloring} is a partition of $V$ into $k$ non-empty stable sets called \emph{color classes}, denoted by $C_1, \ldots, C_k$, such that vertices in $C_i$ are colored with color $i$ for $i \in \{ 1,\ldots,k \}$.
GCP consists of finding the \emph{chromatic number} of $G$, denoted by $\chi(G)$, which is the minimum number $k$ of colors such that $G$ admits a $k$-coloring.
There are many practical situations that can be modeled as a GCP with some additional restrictions. For instance, in scheduling problems, it
would be desirable to assign a balanced workload among employees to avoid unfairness, or to assign balanced task load to prevent further
wear on some machines more than others.
The addition of this extra \emph{equity} constraint gives rise to the \emph{Equitable Coloring Problem} (ECP).
Formally, an \emph{equitable $k$-coloring} (or just $k$-eqcol) of $G$ is a $k$-coloring satisfying the \emph{equity constraint},
i.e.~ $\left| |C_i| - |C_j| \right| \leq 1$, for $i, j \in \{1, \ldots, k \}$ or, equivalently,
$\lfloor n / k \rfloor \leq |C_j| \leq \lceil n / k \rceil$ for each $j \in \{1, \ldots, k \}$, where $n = |V|$.
The \emph{equitable chromatic number} of $G$, $\chi_{eq}(G)$, is the
minimum $k$ for which $G$ admits a $k$-eqcol. The ECP consists of finding $\chi_{eq}(G)$ and it is also an $NP$-hard problem \cite{KUBALE}.
One of the most well known exact algorithms for GCP is Branch-and-Bound \textsc{DSatur}, proposed by Br\'elaz in \cite{DSATUR}
and later improved by Sewell in \cite{SEWELL}. This algorithm is still used by its simplicity, its efficiency in medium-sized graphs and the possibility of applying it at some stage in metaheuristics or in more
complex exact algorithms like Branch-and-Cut. Recently, it was shown that a modification of \textsc{DSatur} performs relatively well compared with many
state-of-the-art Branch-and-Cut algorithms, showing superiority in random instances \cite{PASS}. This fact makes nowadays research on DSatur-based solvers
still important.
The goal of this work is to present a DSatur-based solver for the ECP. In the next sections, we
review \textsc{DSatur} algorithm and we propose new pruning rules specifically derived from the equity constraint. Finally, we report our computational experience.
\vspace{-10pt}
\subsection{Notations and the \textsc{DSatur} algorithm}
\textsc{DSatur} is an implicit enumeration algorithm
where each node of the tree corresponds to a
partial coloration of the graph.
A \emph{partial} $k$-\emph{coloring} of $G$, $\Pi = (k, C_1, \ldots, C_n, U, F)$, is defined by a positive integer $k$, a family of disjoint stable sets $C_1, \ldots, C_n$ of $G$ such that $C_j \neq \varnothing$ if and only if $j \leq k$, \emph{a set of uncolored vertices} $U = V \backslash (\cup_{j=1}^k C_j)$ and
a list $F$ of their \emph{feasible color sets}, i.e.~
for every $u\in U$, $F(u)=\{j\in\{1,\ldots,n\} : \textrm{no vertex of}~C_j~\textrm{is adjacent to}~u \}$.
Clearly, a partial $k$-coloring with $U=\varnothing$ is a $k$-coloring.
Given $\Pi = (k, C_1, \ldots, C_n, U, F)$, $u \in U$ and $j \in \{1,\dots,k+1\}$, we denote by $\langle u,j \rangle \hookrightarrow \Pi$ to the partial coloring of $G$ obtained by adding node $u$ to $C_j$, i.e.~ if
$\langle u,j \rangle \hookrightarrow \Pi = (k', C'_1, \ldots, C'_n, U', F')$ then $k' = \max \{j,k\}$, $C'_j = C_j \cup \{u\}$,
$C'_r = C_r$ for all $r \neq j$, $U' = U \backslash \{u\}$, $F'(v) = F(v) \cap \{j\}$ for each $v \in U'$ adjacent to $u$,
and $F'(v) = F(v)$ otherwise.
Given a maximal clique $Q = \{v_1, v_2$, $\ldots$, $v_q\}$ of $G$, we denote by $\Pi_Q$ the partial $q$-coloring defined by
$C_i = \{v_i\}$ for all $1 \leq i \leq q$ and $U = V \backslash Q$.
\textsc{DSatur} is based on a generic enumerative scheme proposed by Brown \cite{BROWN}, outlined as follows:
\medskip
{ \footnotesize
\noindent \underline{\textsc{Initialization}}: $G$ a graph, $\overline{c}$ an $UB$-coloring of $G$ and $Q$ a maximal clique of $G$.\\
\noindent \underline{\textsc{Node}$(\Pi = (k, C_1, \ldots, C_n, U, F))$}:~~~~~($UB$ and $\overline{c}$ are global variables)\\
\indent \emph{Step 1}. If $U = \varnothing$, set $UB \leftarrow k$, $\overline{c} \leftarrow \Pi$ and return.\\
\indent \emph{Step 2}. Select a vertex $u \in U$.\\
\indent \emph{Step 3}. For each color $1 \leq j \leq \max\{k+1,UB-1\}$ such that $j \in F(u)$, do: \\
\indent \indent $\Pi' \leftarrow \bigl( \langle u,j \rangle \hookrightarrow \Pi \bigr)$\\
\indent \indent If $F'(v) \neq \varnothing$ for all $v \in U'$, execute \textsc{Node}$(\Pi')$.
}
\medskip
It is not hard to see that the recursive execution of \textsc{Node}$(\Pi_Q)$ finally gives the value of $\chi(G)$ into the variable $UB$ and an optimal coloring into $\overline{c}$ \cite{DSATUR}.
Based on this scheme, different algorithms for solving GCP have been evaluated by proposing different \emph{vertex selection strategies} in Step 2.
Two of them are due to Br\'elaz (\textsc{DSatur} algorithm \cite{DSATUR}) and Sewell (\textsc{Celim} algorithm \cite{SEWELL}).
Recently, San Segundo proposed an improvement to Sewell rule (\textsc{Pass} algorithm \cite{PASS}), which gave rise to a competitive solver
with respect to many of the recent exact algorithms in the literature.
\vspace{-10pt}
\section{\textsc{EqDSatur}: An exact algorithm for the ECP} \label{SEQDSATUR}
\vspace{-5pt}
Let us notice that a trivial DSatur-based algorithm for the ECP can be obtained from the previous Brown's scheme, by
changing the $UB$-coloring in the initialization by an $UB$-eqcol and adding the condition ``$\Pi$ is an equitable coloring'' in Step 1.
However, using explicitly the equity property during generation of nodes we can avoid to explore tree regions that will not lead to an equitable coloring and therefore would be needlessly enumerated.
In the following lemma, we present necessary conditions for a partial coloring to be extended to an equitable coloring.
\begin{lemma}
Let $G$ be a graph of $n$ vertices and let $UB$ and $LB$ be, respectively, an upper and a lower bound of $\chi_{eq}(G)$. Let $\Pi$ be
a partial $k$-coloring of $G$ such that $k < UB$
and $M = max\{|C_r|: 1 \leq r\leq k\}$.
Then, if $\Pi$ can be extended to an $r$-eqcol of $G$ with $k \leq r < UB$, the following properties hold:\\
\noindent (P.1) $|U| \geq \sum_{r=1}^k (\max\{M-1, \left\lfloor \frac{n}{UB-1}\right\rfloor\} - |C_r|)^+$
~~~~~~~ (P.2) $M \leq \left\lceil \frac{n}{\max\{k,LB\}} \right\rceil$
\end{lemma}
\vspace{-5pt}
In addition, property P.1 in the previous lemma also gives us a sufficient condition for a partial coloring with $U=\varnothing$ to be an equitable coloring:
\vspace{-5pt}
\begin{lemma}
If $\Pi$ is a partial $k$-coloring satisfying property P.1 and $U=\varnothing$ then $\Pi$ is a $k$-eqcol.
\end{lemma}
\vspace{-5pt}
Our DSatur-based algorithm introduces the previous properties as modifications (written in boldface) into the Brown's scheme, in the following way:
\medskip
{ \footnotesize
\noindent \underline{\textsc{Initialization}}: $G$ a graph, $\overline{c}$ an $UB$-{\bf eqcol} of $G$, {\bf LB a lower bound of $\chi_{eq}(G)$}
and $Q$ a maximal clique of $G$.\\
\noindent \underline{\textsc{Node}$(\Pi = (k, C_1, \ldots, C_n, U, F))$}:~~~~~($UB$ and $\overline{c}$ are global variables)\\
\indent \emph{Step 1}. If $U = \varnothing$, set $UB \leftarrow k$, $\overline{c} \leftarrow \Pi$ and return.\\
\indent \emph{Step 2}. Select a vertex $u \in U$.\\
\indent \emph{Step 3}. For each color $1 \leq j \leq \max\{k+1,UB-1\}$ such that $j \in F(u)$, do: \\
\indent \indent $\Pi' \leftarrow \bigl( \langle u,j \rangle \hookrightarrow \Pi \bigr)$\\
\indent \indent If $F'(v) \neq \varnothing$ for all $v \in U'$ {\bf and $\Pi'$ satisfies P.1 and P.2}, execute \textsc{Node}$(\Pi')$.
}
\medskip
We have the following:
\vspace{-5pt}
\begin{theorem}
The recursive execution of \textsc{Node}$(\Pi_Q)$ gives the value of $\chi_{eq}(G)$ into the variable $UB$
and an optimal equitable coloring into $\overline{c}$.
\end{theorem}
\vspace{-5pt}
Proofs of the previous results are omitted due to lack of space.
In order to initialize our algorithm we use the heuristic \textsc{Naive} given in \cite{KUBALE} for obtaining an initial $UB$-eqcol of $G$. The lower bound $LB$ is computed as in \cite{DANIEL}, i.e.~ as the maximum between the size of the maximal clique computed greedily and a relaxation of a bound given
in \cite{EQTREE}.
Regarding the mentioned vertex selection strategies, we carried out benchmark tests over random instances
and we concluded
that \textsc{Pass} \cite{PASS} is the best choice. We call $\textsc{EqDSatur}$ to our implementation using this strategy.
Another factor we take into account is
in which order the nodes are evaluated in Step 3.
All mentioned DSatur implementations for GCP evaluate
first color $j=1$, then color $j=2$, and so on. We call $\textsc{EqDS}_1$ to
$\textsc{EqDSatur}$ with this criterion. In addition, we consider another strategy based on sorting color classes
according to their size in ascending order: if $|C_{i_1}| \leq |C_{i_2}| \leq \ldots \leq |C_{i_k}|$, we evaluate first $j=i_1$, then $j=i_2$, and so on.
We call $\textsc{EqDS}_2$ to $\textsc{EqDSatur}$ with this strategy.
\vspace{-10pt}
\section{Computational experience}
\vspace{-5pt}
Computational tests were carried out on an Intel i5 2.67Ghz over Linux O.S.
Some details and tables were omitted due to lack of space. Instead, a summary of the most essential details are provided.
Our first experiment compares $\textsc{EqDS}_1$ against the ``trivial'' DSatur-based exact algorithm for the ECP mentioned at Section \ref{SEQDSATUR}, in order to evaluate
whether
the time needed to check properties P.1 and P.2 compensate for
the time wasted in
exploring nodes of the enumeration tree where these properties do not hold. We noticed that $\textsc{EqDS}_1$ really outperforms the trivial implementation.
For instance, in medium-density random graphs of 70 vertices, $\textsc{EqDS}_1$ is in average 25 times faster than the trivial implementation and
is able to solve 21\% more instances within 2 hours of execution.
The second experiment compares $\textsc{EqDS}_1$ and $\textsc{EqDS}_2$ against ``integer linear programming-based'' solvers for ECP, a classical
approach for developing exact algorithms.
We consider the recent Branch-and-Cut proposed in \cite{BYCBRA} (B\&C-$LF_2$) for which the authors report results on random instances up to 70 vertices in a 1.8 Ghz AMD-Athlon platform.
For random instances with 80 vertices, we compare \textsc{EqDSatur} against CPLEX 12.1 solving the formulation for ECP given in \cite{DANIEL}
with the same initial bounds.
The following table
reports the percentage of solved instances, average of relative gap and time elapsed for 100 random instances
(10 per row, except results from \cite{BYCBRA}). An instance is considered \emph{not solved} after the limit of two hours of execution.
A bar ``$-$'' means no instance was solved. Each instance of $d \%$ of density is generated by considering a uniform probability $d$ that two
vertices are adjacent to each other.
\begin{center}
\tiny
\begin{tabular}{c@{\hspace{4pt}}c|c@{\hspace{4pt}}c@{\hspace{4pt}}c|c@{\hspace{4pt}}c@{\hspace{4pt}}c|c@{\hspace{4pt}}c@{\hspace{4pt}}c}
& \%Density & \multicolumn{3}{c|}{\% solved inst.} & \multicolumn{3}{c|}{\% Rel. Gap.} & \multicolumn{3}{c}{Time} \\
Vertices & Graph & $LF_2$ & $\textsc{EqDS}_1$ & $\textsc{EqDS}_2$ & $LF_2$ & $\textsc{EqDS}_1$ & $\textsc{EqDS}_2$ & $LF_2$ & $\textsc{EqDS}_1$ & $\textsc{EqDS}_2$ \\
\hline
70 & 10 & 100 & 100 & 100 & 0 & 0 & 0 & 109 & 0 & 0 \\
70 & 30 & 0 & 100 & 100 & 18 & 0 & 0 & $-$ & 0 & 0 \\
70 & 50 & 0 & 100 & 100 & 8,2 & 0 & 0 & $-$ & 16,2 & 16,1 \\
70 & 70 & 100 & 100 & 100 & 0 & 0 & 0 & 273 & 31,2 & 32,1 \\
70 & 90 & 100 & 100 & 100 & 0 & 0 & 0 & 11 & 0 & 0 \\
\hline
& & CPLEX & $\textsc{EqDS}_1$ & $\textsc{EqDS}_2$ & CPLEX & $\textsc{EqDS}_1$ & $\textsc{EqDS}_2$ & CPLEX & $\textsc{EqDS}_1$ & $\textsc{EqDS}_2$ \\
\hline
80 & 10 & 100 & 100 & 100 & 0 & 0 & 0 & 5,7 & 0 & 0 \\
80 & 30 & 0 & 100 & 100 & 20 & 0 & 0 & $-$ & 23,7 & 17,6 \\
80 & 50 & 0 & 100 & 100 & 24 & 0 & 0 & $-$ & 477 & 424 \\
80 & 70 & 10 & 70 & 70 & 16 & 10 & 10 & 5769 & 1715 & 1653 \\
80 & 90 & 100 & 100 & 100 & 0 & 0 & 0 & 690 & 12,5 & 12,3 \\
\hline
\end{tabular}
\end{center}
\smallskip
Our algorithm is able to solve more instances than CPLEX and, without considering the difference of platforms, B\&C-$LF_2$.
Also, $\textsc{EqDS}_2$ seems to be a little better than $\textsc{EqDS}_1$ in terms of time.
The last experiment compares $\textsc{EqDS}_1$ and $\textsc{EqDS}_2$ against CPLEX and B\&C-$LF_2$ on
24 benchmark instances reported in \cite{BYCBRA}.
16 instances have been solved by CPLEX, $\textsc{EqDS}_1$ and $\textsc{EqDS}_2$ in less than two seconds.
B\&C-$LF_2$ solved these 16 instances with an average of 62 seconds and never outperforms the other three algorithms.
Instance \texttt{queen8\_8} have been solved by CPLEX in 654 sec., by B\&C-$LF_2$ in 441 sec., by $\textsc{EqDS}_1$ in 7.5 sec. and by $\textsc{EqDS}_2$ in 1.1 sec. Instances \texttt{miles1000} and \texttt{miles750} have not been solved by
$\textsc{EqDS}_1$, but have been solved by B\&C-$LF_2$ in 267 and 171 sec. respectively, and by CPLEX and $\textsc{EqDS}_2$ in less than a second.
On the other hand, neither $\textsc{EqDS}_1$ nor $\textsc{EqDS}_2$ could solve 5 instances.
In particular, 3 of these instances (\texttt{3-FullIns\_3}, \texttt{4-FullIns\_3}, \texttt{5-FullIns\_3}) are hard to solve by enumerative
schemes, as reported in \cite{PASS}, so in our opinion, \textsc{EqDSatur} presents the expected behaviour.
We want to remark that \textsc{EqDSatur} has also been able to solve DIMACS benchmark instances which are not mentioned in \cite{BYCBRA}.
For instance, \texttt{queen9\_9} has been solved by $\textsc{EqDS}_1$ and $\textsc{EqDS}_2$ in less than 10 minutes. In contrast, CPLEX
could not solve it in the term of 4 hours of execution. Another example is \texttt{myciel5} which
has been solved in less than a sec.\! by $\textsc{EqDS}_1$ and $\textsc{EqDS}_2$ but CPLEX needed 149 sec.\! to solve it.
From these results, we conclude that $\textsc{EqDS}_2$ is highly competitive with respect to the algorithms available in the literature for the ECP.
\vspace{-10pt}
|
\section{Introduction}
We consider the controlled stochastic differential system in $\R^{n+1}$
\begin{equation}\label{sdeintro} \begin{cases}
dX(s)=f(X(s), Y_\eps(s^-), u(s)) ds+\sigma(X(s),Y_\eps(s^-), u(s))dW(s) \\
dY_\eps(s)= -\frac{1}{\varepsilon }Y_\eps(s^-)ds+dZ\left(\frac{1}{\varepsilon }s\right) \\ \end{cases}
\end{equation} with $s\geq t$ and initial data $X(t)=x\in\R^n$, $Y_\eps(t)=y\in\R$. The function $u(\cdot)$ is the control taking values in a given compact set $U$, $\eps>0$ is a small parameter, $W$ is a standard $r$-dimensional Brownian motion and $Z$ is a $1$-dimensional pure jump L\'evi process, independent of $W$.
We associate to this
system a payoff functional of the form
\[
\E[e^{c(t-T)} g(X(T
)\ |\ X(t)=x, Y_\eps(t)=y] , \qquad 0\leq t\leq T,
\]
where $c\geq 0$ and $g:\R^{n}\to \R$ is a continuous function with quadratic growth, that we wish to maximise among admissible control functions $u(\cdot)$.
The value function of this optimal control problem is defined as
\begin{equation}\label{valueintro}
V^\eps(t,x,y):=\sup_{u(\cdot)}\E[e^{c(t-T)} g(X(T
)].
\end{equation}
We are interested in the analysis of the limit as $\eps \to 0$ of the control problem given by the system \eqref{sdeintro} and by the value function \eqref{valueintro}.
Our main motivation comes from financial models with stochastic volatility. In such models
$X(s)$ represents, for instance,
the log-prices of $n$
assets, or the wealth of a portfolio.
The
volatilities of the assets, collected in the matrix $\sigma$,
are affected by another process
$Y(s)$ that is usually a diffusion driven by another Brownian motion
negatively correlated with the one driving the stock prices.
Fouque, Papanicolaou, and Sircar
argued in the book \cite{FPS} that the bursty behaviour of volatility observed in financial markets can be described by introducing a faster time scale for a mean-reverting diffusion process $Y$.
Several result along these lines were found, mostly for problems without controls, e.g., in \cite{fps1, fps2,
fpss},
see also the references therein. The papers \cite{BCM, bc} by the first two authors introduced viscosity methods to prove the convergence of the singular perturbation for models involving a control variable and therefore associated to a fully nonlinear Hamilton-Jacobi-Bellman
equation. The main example in \cite{BCM} was the classical Merton's portfolio optimization problem with volatility depending on an ergodic diffusion process $Y$.
On the other hand, the work by Barndorff-Nielsen and Shephard \cite{BNS1} showed that processes of Ornstein-Uhlenbeck type driven by a pure-jump L\'evy process are more appropriate models of the volatility than diffusions.
Several authors studied financial problems with such a non-Gaussian mean-reverting stochastic volatility evolving at the same time scale as the prices: Merton's problem in
\cite{BKR} and various option pricing problems in \cite{NV, Hs2, hs}. The novelty of the present paper is combining multiple scales and stochastic volatility with jumps. In particular, we extend to this context the
convergence results for asset pricing and for Merton's problem obtained in \cite{BCM}.
To conclude this brief bibliographical introduction we refer
to the books \cite{SATO} for the theory of L\'evy processes and \cite{CT} for their applications to finance. Let us mention also that the recent paper \cite{LL} deals with a multiscale model whose assumptions are in a sense opposite to ours: the volatility is a diffusion and the slow variable $X$ is driven by a jump process.
We describe now in more details our result.
By dynamic programming
arguments \cite{Ph2, S}, the value function $V^\eps$ in \eqref{valueintro} is a viscosity solution of the integro-differential Hamilton-Jacobi-Bellman equation
\begin{eqnarray}\label{pdeintro} - V^\eps_t &+& H(x,y, D_x V^\eps, D^2_{xx}V^\eps)
-\frac{1}{\eps} y\cdot D_y V^\eps \\&-&\frac{1}{\eps} \int_0^{+\infty} (V^\eps(t,x,z+y)-V^\eps(t,x,y)- D_yV^\eps (t,x,y)\cdot z \,1_{ |z| \leq 1}) d\nu (z)+cV^\eps= 0 \nonumber \end{eqnarray}
in $(0,T)\times\R^n\times\R$, with terminal data $V^\eps(T,x,y)=g(x)$,
where $\nu$ is the L\'evy measure associated to the process $Z$ and $H$ is a standard Hamiltonian associated to stochastic control problems, see the precise definition \eqref{H} in Section \ref{4}.
Letting $\eps\to 0$ in \eqref{pdeintro} is a singular perturbation problem, and we treat it by methods of the theory of viscosity solutions to integro-differential PDEs.
Our main result, see Theorem \ref{ConvTheo}, is the proof of the uniform convergence of $V^\eps$ as $\eps\to 0$ to the unique viscosity solution $V(t,x)$ of the {\em effective PDE}
\begin{equation} \label{effintro}
-V_
+ \int_{\R} H(x, y, D_x V, D^2_{xx} V) d\mu (y) + c
=0
\end{equation}
in $(0,T)\times\R^n$, with terminal data $V(T,x)=g(x)$, where $\mu$ is the unique invariant measure of the process $Y_\eps$ (which is independent of $\eps$, see Proposition \ref{lem:2}), and $H$ is the Hamiltonian appearing in \eqref{pdeintro}.
The second step in the solution of the singular perturbation problem is to interpret the effective equation \eqref{effintro} as the Hamilton-Jacobi-Bellman equation for a limit {\em effective control problem}.
This can be done (Remark \ref{effcontrol}) by a general relaxation procedure taken from \cite{BTer}. However a simpler and explicit
representation of the effective systems can be given for our two main model problems, that is, the asset pricing and the Merton's optimisation problem, see Section \ref{7}. In particular we show that in the limit $\eps\to 0$ the asset pricing problem converges to a new asset pricing with constant volatility
$$\tilde{\sigma}^2:
\int_{\R} \sigma^2(y) \mu(dy)
$$
where $\mu$ is the invariant measure of the process $Y_\eps$, whereas the portfolio optimisation converges to a Merton's problem with
constant volatility given by the harmonic mean of $\sigma(y)$, namely,
$$
\overline{\sigma}^2:
\left(\int_{\R}\frac{1}{\sigma^2(y)}\mu(dy
\right)^{-
} ,
$$
which is smaller than $\tilde{\sigma}$.
The proofs of our results rely on several tools,
among which we mention the exponential ergodicity of the fast process $Y$, proved by Kulik \cite{K}, some properties of viscosity solutions for integrodifferential equations \cite
Sa, Ph2, BI, C}, and the perturbed test function method introduced by Evans for periodic homogenization \cite{E} and extended to singular perturbations in \cite{ab, AB2}. To adapt this method to the current setting of unbounded fast variables $y$, we suitably modify the perturbed test function by means of a Lyapunov function associated to the process $Y$ (this is an improvement also with respect to \cite{BCM, bc}).
Our approach is flexible enough to deal with more general problems, such as integral payoffs and slow variables $X$ depending also on a jump process (under more restrictive growth conditions on the data). Moreover the fast variables $Y$ could be assumed to be vectorial and depending on
a combination of jump and diffusion processes, provided that the resulting process be uniformly ergodic and its generator satisfy the strong maximum principle. The last requirement is a limitation to the applicability of our method: if tested on processes $Z$ whose L\'evy measure $\nu$ has $\supp(\nu)\subseteq [0, +\infty)$ it forces us to assume
\[
\int_{|z|\leq 1} |z|\nu (dz)=+\infty ,
\]
whereas the opposite case is treated, e.g., in \cite{BKR}. On the other hand, all processes $Z$ generated by a fractional Laplacian fit our assumptions.
The paper is organized as follows. Section \ref{2} describes the basic assumptions on the optimal control problem. Section \ref{3} is devoted to the assumptions on the volatility process $Y$ and its properties, in particular the exponential ergodicity and the strong maximum principle and Liouville property of the generator.
Section \ref{4} describes the partial integrodifferential HJB equation associated to $V^\eps$. In Section \ref{5} we study the cell problem that allows to identify the effective Hamiltonian for the limit PDE. Section \ref{6} contains the statement and proof of the convergence theorem. In Section \ref{7} we apply the previous theory to financial models.
\section{Standing assumptions on the control system}
\label{2}
We consider the controlled stochastic differential equation
\begin{equation}\label{sde} \begin{cases}
dX(s)=f(X(s), Y_\eps(s^-), u(s)) ds+\sigma(X(s),Y_\eps(s^-), u(s))dW(s) & X(t)=x
\in \R^n\\
dY_\eps(s)= -\frac{1}{\varepsilon }Y_\eps(s^-)ds+dZ\left(\frac{1}{\varepsilon }s\right) &Y_\eps(t)
\in \R \\ \end{cases}\end{equation}
where $\eps>0$ is a small parameter, $W(t)=(W^1(t),\dots,W^r(t))$ is a r-dimensional Brownian motion and $Z(t)$ is a pure jumps L\`evy process. Furthermore, we assume that $W$ and $Z$ are independent.
We assume the standard conditions on coefficients $f:\R^n\times\R\times U\to\R^n$, $\sigma:\R^n\times\R\times U\to \Mi^{n,r}$, where
$\Mi^{n,r}$ denotes the set of $n\times r$ matrices. In particular we assume that $f,\sigma$ are continuous functions, Lipschitz continuous in $(x,y)$ uniformly w.r.t. $u\in U$, where $U$ is a compact set and that $f(x,\cdot, u)$, $\sigma(x,\cdot, u)$ are bounded for every $(x,u)$.
We say that a process $u(\cdot)$ on the time interval $[0,T]$ is an admissible control function if it takes values in $U$ and is progressively measurable with respect to the filtration generated by $W(\cdot)$ and $Z(\cdot)$, and we set
\[
\mathcal{U}:=\{u(\cdot)\ \text{ admissible control function}\}.
\]
We will not make any non-degeneracy assumption on $\sigma$.
To fit financial models, we assume that the $i$-th component of the velocity of $X$ vanishes when $X_i(s)=0$, i.e.,
\begin{equation}
\label{zero}
x_i=0
\quad \Longrightarrow \quad f_i(x,y,u)=0, \; \sigma_{ij}(x ,y,u)=0, \quad\forall\, j=1,\dots,r, \, y\in\R,\, u\in U .
\end{equation}
We set
\[
\R^n_+:=\{x\in\R^n\,:\,x_i>0 \text{ for every }i=1,\dots,n\}.
\]
Note that under the previous assumption, then the set $\overline{\R^n_+}\times \R$ is invariant under the stochastic process \eqref{sde}. This means that if the initial data satisfies $x_i\geq 0$ for some $i$, then every solution to \eqref{sde} satisfies $X_i(s)\geq 0$ almost surely for every $s\geq t$.
We consider for simplicity a payoff functional depending only on the position of the system at a fixed terminal time $T>0$. The utility function $g:\R^n_+\to\R$ is a continuous function satisfying
\begin{equation}
\label{GrowthC}
\exists\,K>0, \quad \text{such that } \quad |g(x)|\le K(1+|x|^2) \quad \text{for every } x\in\R^n_+ \\
\end{equation}
and the discount factor is $c\geq 0$.
Therefore the value function of the optimal control problem is
\begin{equation}
\label{PO}
V^\eps(t,x,y):=\sup_{u\in\mathcal{U}}\E[e^{c(t-T)}g(X(T))\,|\, X(t)=x,\, Y_\eps(t)=y], \\
\end{equation}
and $(X(\cdot),\,Y_\eps(\cdot))$ satisfy \eqref{sde} with control $u$. This choice of the payoff is sufficiently general for the application to finance models presented in this paper, but we could easily include in the payoff an integral term keeping track of some running costs or earnings.
\section{The fast subsystem}\label{3}
We consider the process $Y$ in \eqref{sde}, putting $\eps=\frac{1}{\lambda}>0$. $Y$ is an Ornstein-Uhlenbeck non-Gaussian process driven by a L\'evy process $Z$, \emph{i.e.}
\begin{equation}
\label{Fast}
d Y_\lambda(s)=-\lambda Y_\lambda(s^-) ds + dZ(\lambda s), \\
\end{equation}
where $\lambda>0$ is the rate of mean reversion.
We assume that the process $Z$ is a
pure-jump
L\'evy process with no drift, i.e., a L\'evy process whose L\'evy-Ito decomposition has null continuous part, and we choose its cadlag (RCLL) version.
We refer to the monograph \cite{SATO} for a general introduction to L\'evy processes.
The process $Y_\lambda(s)$, with initial datum $Y_\lambda(0)=y\in\R$, can be explicitly written as
\[
Y_\lambda(s)=ye^{-\lambda s}+\int_0^se^{\lambda(u-s)} dZ(\lambda u).
\]
We associate to the process $Z(s)$ its L\'evy measure $\nu$. Intuitively speaking, the L\'evy measure describes the expected number of
jumps of a certain height in a time interval of length 1, i.e.
\[
\nu(B)=\E(\#\{s\in[0,1],\,\,Z(s)-Z(s^-)\neq 0,\,\,Z(s)-Z(s^-)\in B \} ). \]
The L\'evy measure has no mass at the origin, while singularities (i.e. infinitely many jumps) can
occur around the origin (i.e. small jumps). Moreover, the mass away from
the origin is bounded (i.e. only a finite number of big jumps can occur).
In particular $\nu$ satisfies the following integrability condition (see \cite{SATO}):
\begin{equation}
\label{int}
\int_{|z|<1}|z|^2\nu(dz)+\int_{|z|\geq 1}\nu(dz) <\infty.
\end{equation}
We will consider L\`evy processes with infinite activity, that is, such that $\nu(\R) = +\infty$. In this case almost all paths of the process $Z$ have an infinite number of jumps on every compact interval.
The infinitesimal generator of this process $Y_\lambda$ (see \cite{SATO}, Thm. 31.5) is given by $\lambda \mathcal{I}$, where $\mathcal{I}$ is defined as follows
\begin{equation}\label{l} \Lu(y, [f] ) = - f'(y)\cdot y + \int_0^{+\infty} (f(z+y)-f(y)-f'(y)z 1_{ z \leq 1}) d\nu (z). \end{equation}
\subsection{Standing assumptions on the L\'evy process}
Besides the integrability condition \eqref{int} we need to assume some other condition on the measure $\nu$.
The first one is a quite standard non degeneracy condition (see \cite{BI}). In particular we assume that the measure is singular at $0$ and we introduce a parameter $p\in (0,2)$ characterizing the singularity of the measure at $0$.
\begin{assumption}
\label{A1}
There exist $C>0$, $p\in(0,2)$ such that for every $0<\delta\le1$
\begin{equation}\label{a1}
\int_{|z|\leq \delta} |z|^2 \nu(dz) \ge C \delta^{2-p}.
\end{equation}
\end{assumption}
\begin{remark}
\upshape An equivalent way to state \eqref{a1} is the following: there exists $r\in(0,2)$ such that
\begin{equation} \label{a1n}
\delta^{-r}\int_{|z|\leq \delta} |z|^2 \nu(dz) \to +\infty \quad\quad \text{as } \delta\to 0^+. \\
\end{equation} Indeed if \eqref{a1} holds then \eqref{a1n} is satisfied for $r\in (0,2-p)$. Viceversa, if \eqref{a1n} holds, then \eqref{a1} is satified with $p\in(0, 2-r)$.
\end{remark}
\begin{remark}\upshape \label{crit}
There is a general difference between the case in which the constant $p$ in \eqref{a1}
is above or below the critical value $p=1$.
In particular if $p\in (0,1)$, using \eqref{a1n}, it can be proved that \[\int_{|z|\leq 1} |z|\nu (dz)<+\infty, \] and then that almost all paths of $Z$ have finite variation. If $p\in (1,2)$, on the other side, \[\int_{|z|\leq 1} |z|\nu (dz)=+\infty, \] and almost all paths of $Z$ have infinite variation (see \cite[Thm. 21.9]{SATO}).
\end{remark}
We assume a stronger integrability condition at infinity than \eqref{int} of the measure $\nu$.
\begin{assumption}
\label{A3}
There exists $q>0$ such that
\begin{equation}\label{a3}
\int_{|z|>1}|z|^q \nu(dz) < +\infty .
\end{equation}\end{assumption}
It can be proved (see \cite[Thm. 25.3]{SATO}) that under Assumption \ref{A3}, the process $Z(t)$ has finite $q$-th moment for every $t$, that is, $\E (|Z_t|^q)<\infty$.
As we will prove in Section \ref{erg} Assumptions \ref{A1},\ref{A3} are sufficient to prove unique ergodicity of the process $Y_\lambda$ defined in \eqref{Fast}.
Nevertheless, we need to add a technical assumption. Indeed the non degeneracy assumption \eqref{a1} is not sufficient to assure the validity of a Strong Maximum Principle for the associated operator $\Lu$ (see Remark \ref{subordinator}). So we assume also the following.
\begin{assumption} \label{A2}
At least one of the following conditions holds:\begin{itemize}
\item[i)] the parameter $p$ in \eqref{a3} satisfies $p\in (1,2)$,
\item[ii)] $\R$ can be covered by translations of the support $\supp (\nu)$ of the measure $\nu
, \emph{i.e.},
\begin{equation}\label{assu}
\R=\bigcup_{n\ge 0}\left(\,\underbrace{\supp(\nu)+\dots+\supp(\nu)}_n\,\right) .
\end{equation}
\end{itemize}
\end{assumption}
\begin{remark}
\upshape \eqref{assu} can be replaced by the requirement that
$0$ belongs to the topological interior of the measure support $\supp(\nu)$.
\end{remark}
The main examples of L\'evy processes satisfying the previous assumptions are the \emph{$\alpha$-stable L\'evy processes}, defined on the whole space or restricted to the half space. We recall that $Z(t)$ is an $\alpha$ stable process if $Z(t) t^{1/\alpha}= Z(1)$ for all $t$, in the sense that the two processes have the same law.
\begin{example}\upshape
Consider the $\alpha$-stable L\'evy process with measure
\begin{equation}
\label{FL}
\nu(dz)=\frac{dz}{|z|^{1+\alpha}}, \\
\end{equation}
where $\alpha\in(0,2)$. The generator of this process is the fractional Laplacian $(-\Delta)^{\alpha/2}$ (see \cite{BI}). Assumption \ref{A1} holds with $p=\alpha$, and Assumption \ref{A3} holds with $q<\alpha$. Finally the case \emph{ii)} \eqref{assu} in Assumption \ref{A2}
holds.
\end{example}
\begin{example}
\upshape Consider the $\alpha$-stable L\'evy process
restricted to the half space, with measure
\begin{equation}
\label{FLhs}
\nu(dz)=1_{\{z\ge0\}}(z)\frac{dz}{z^{1+\alpha}}. \\
\end{equation} In this case the condition \emph{ii)} \eqref{assu} in Assumption \ref{A2}
is no more valid, so we assume that $\alpha\in(1,2)$ to ensure the condition \emph{i)}.
\end{example}
\begin{remark}\label{sub1}\upshape We remark that Assumption \ref{A2} cannot hold if $Z$ is a subordinator.
Subordinators are 1-dimensional L\'evy processes with non-decreasing sample paths.
It's easy to deduce that a 1-dimensional L\'evy process is a subordinator if and only if its associated measure satisfies \begin{equation}\label{sub} \nu((-\infty,0])=0\qquad \int_0^{1}|z|\nu(dz)+\int_1^{+\infty}\nu(dz)<+\infty.\end{equation} This implies in particular that $\nu$ cannot satisfy Assumption \ref{A2} (see Remark \ref{crit}).
Among $\alpha$-stable processes
with associated L\'evy measure $\nu$ defined by \eqref{FLhs}
the subordinators are those with $\alpha\in(0,1)$. A notable example is the inverse Gaussian, corresponding to
$\alpha=\frac12$. It can be defined as $Z(s):=\inf\{t\ |\ W(t)>s\}$, where $W$ is a $r$-dimensional Brownian motion.
(Note moreover that if $Z$ is a $\alpha$-stable subordinator and $W$ is an independent Brownian motion, then $W(Z(t))$ is a stable symmetric L\'evy process of exponent $2\alpha$.)
We will explain in Remark \ref{subordinator} why the proof of convergence presented in this paper does not apply to the subordinators.
\end{remark}
\subsection{Ergodic properties} \label{erg}
In this section, we prove that under the previous Assumptions on the L\'evy process $Z$, the process $Y$ defined in \eqref{Fast} is ergodic. In order to prove ergodicity of a process, two principal features should be checked: recurrence of the process outside some large ball and regularity of the transition probability in some bounded domain.
The first feature can be provided in a quite standard way via an appropriate version of the
Lyapunov criterium (for the relation between existence of a Lyapunov function for the system and recurrence outside compact sets, we refer to the monograph \cite{has}). We show that Assumption \ref{A3} assures the existence of a Lyapunov function.
\begin{lemma}
\label{Lyapunov}
Let
\[
\mathcal{Q} = \{ v\in\mathcal{C}^2(\R) | \,\, \exists \,\, \tilde{v}\,\, \text{locally bounded } : \int_1^{+\infty} v(y+z)d\nu(z) \le \tilde{v}(y) \quad \forall\, y \}. \\
\]
Under Assumption \ref{A3} there exists $\varphi\in\mathcal{Q}$ and constant $a>0$ such that
\begin{equation}
-\Lu[y,\varphi] \geq a\varphi(y) \quad \text{and} \quad \varphi \to +\infty \quad \text{as} \quad |y| \to +\infty \\
\end{equation}
\end{lemma}
\begin{proof} For the proof we refer to \cite[Proposition 4.1]{K}. The idea is to construct $\varphi$ such that $\varphi(y)\leq|y|^q$ for all $y\in\R$ and $\varphi(y)=|y|^q$ for $|y|\geq R$, for $R$ sufficiently large.
\end{proof}
Secondly, Assumption \ref{A1} provides the regularity of the transition probability of the process $Y$.
\begin{lemma}
\label{lem:1}
Under Assumption \ref{A1}, the Ornstein-Uhlenbeck process $Y_\lambda(t)$ has a $\mathcal{C}^{\infty}$ density with all bounded derivatives for all $t$ and every initial data $y$.
\end{lemma}
\begin{proof}
We refer to \cite{P} and \cite{PZ}.
\end{proof}
We prove now exponential ergodicity of the process $Y$.
\begin{proposition}
\label{lem:2}
Let Assumptions \ref{A1}, \ref{A3} hold. Then for any $\lambda>0$, the process $Y_\lambda$ in \eqref{Fast} admits a unique invariant distribution $\mu$, which is independent of $\lambda$. Moreover, it is exponentially ergodic, in the sense that there exists a positive constant $C$ such that for every bounded measurable function $f$ there exists $K>0$ for which
\begin{equation}\label{es}
\left|\frac{1}{t}\int_0^t\E(f(Y_\lambda(s)))ds - \int_{\R}f(z)\mu(dz)\right| \le K(1+|y|^q)e^{-Ct} \quad \text{as} \, t\to+\infty, \\
\end{equation}
where $Y_\lambda$ is the solution to \eqref{Fast} with initial data $Y_\lambda(0)=y$ and $q$ is as in Assumption \ref{A3}.
\end{proposition}
\begin{proof}
In \cite[ Thm. 17.5, Cor. 17.9]{SATO} it is proved that the process $Y_\lambda$ for every $\lambda>0$ has an invariant distribution $\mu_\lambda$. Moreover, $\mu_\lambda$ is the distribution of a L\'evy process with drift
\begin{equation}\notag
\gamma=\frac{1}{\lambda}\int_{|y|>1}\lambda\nu(dy)=\int_{|y|>1}\nu(dy), \\
\end{equation}
and L\'evy measure
\begin{equation}\notag
\mu(B)=\frac{1}{\lambda}\int_{\R}\lambda \int_0^{+\infty}1_{B} (e^{- s}y) ds\nu(dy)= \int_{\R} \int_0^{+\infty}1_{B} (e^{- s}y) ds\nu (dy), \\
\end{equation}
for every $B\in\mathcal{B}(\R)$. So, the time scaling $dZ(\lambda t)$ assures that the invariant distribution $\mu_\lambda$ of $Y_\lambda$ is independent of $\lambda$. The exponential ergodicity \eqref{es} of $Y$ is proved in \cite[Thm. 1.1 and Prop. 0.1]{K}.
\end{proof}
\begin{corollary}\label{Ab}Under the same assumptions and notations of Proposition \ref{lem:2}, for every bounded measurable function $f$ there exists a constant $K>0$ such that \begin{equation} \label{delta}
\left|\delta\int_0^{+\infty}\E f(Y_\lambda(t))e^{-\delta t}dt- \int_{\R} f(z)\mu(dz )\right|\le K(1+|y|^q) \delta\quad \text{as} \, \delta\to 0^+.
\end{equation}
\end{corollary}
\begin{proof}
The property \eqref{delta} can be deduced from \eqref{es} using a result of Abelian-Tauberian type,
see \cite[Thm. 10.2]{Si}. For completeness we give the proof in our case.
First of all note that it is equivalent to prove, instead of \eqref{delta}, that \begin{equation} \label{delta1}
\left|\delta\int_1^{+\infty}\E f(Y_\lambda(t))e^{-\delta t}dt- \int_{\R} f(z)\mu(dz ) \right|\le K(1+|y|^q) \delta\quad \text{as} \, \delta\to 0^+.
\end{equation}
We denote $\int_{\R} f(z)\mu(dz )=M$.
Fix $y\in\R$ and define $F_y(t):= \int_1^t \E f(Y_\lambda(s))ds$, for $t\geq 1$. By integration by parts, we get
\begin{equation}\label{uno}
\delta\int_1^{+\infty}e^{-\delta t}\E(f (Y_\lambda(t)))dt= \delta^2\int_{1}^{+\infty}e^{-\delta t}F_y(t)dt = \delta \int_{\delta}^{+\infty}e^{-s}F_y\left(\frac{s}{\delta}\right)ds.
\end{equation} Note that, for $\delta>0$ fixed, by \eqref{es} \[\frac{\delta}{s} F_y\left(\frac{s}{\delta}\right) \to M \qquad\text{ as } s\to +\infty. \] Therefore $F_0=\max_{s\geq\delta} \left|\frac{\delta}{s} F_y\left(\frac{s}{\delta}\right) \right|<\infty$ and then \[ \delta \ e^{-s}F_y\left(\frac{s}{\delta}\right)= e^{-s}s \frac{\delta}{s}F_y\left(\frac{s}{\delta}\right)\in L^1(\delta, +\infty).\]
By \eqref{es} we obtain
\begin{equation}\label{due}
\left|\frac{\delta}{s} F_y\left(\frac{s}{\delta}\right) -M\right|\le K(1+|y|^q)e^{-C\frac{s}{\delta}}+\frac{\delta}{s}\|f\|_\infty, \\
\end{equation}
for $s$ fixed. Therefore, using \eqref{uno} and \eqref{due} we get
\begin{align*} & \left|\delta\int_1^{+\infty}e^{-\delta t}\E(f (Y_\lambda(t)))dt-M\right| & \\ \leq & \int_{\delta}^{+\infty}
e^{-s}s\left| \frac{\delta}{s}F_y\left(\frac{s}{\delta}\right)-M\right|ds + (1-e^{-\delta}+\delta e^{-\delta})M & \\ \leq &
\int_{\delta}^{+\infty}
e^{-s}s K(1+|y|^q)e^{-C\frac{s}{\delta}}ds+\int_{\delta}^{+\infty}e^{-s}\delta \|f\|_\infty ds + (1-e^{-\delta}+\delta e^{-\delta})M & \\ \leq &
K(1+|y|^q)e^{-C} \delta^2\frac{C+1}{C^2}+ e^{-\delta}\delta \|f\|_\infty + (1-e^{-\delta}+\delta e^{-\delta})M
\end{align*}
which gives the desired result.
\end{proof}
\subsection{Liouville property}
Finally we state and prove a Liouville property for our integro-differential operator $-\Lu$. This is based on the following Strong Maximum Principle.
\begin{theorem}[Strong Maximum Principle] \label{SMPFL}
Assume the measure $\nu$ satisfies Assumptions \ref{A1} and \ref{A2}. Let $u\in USC(\R)$ be a viscosity subsolution of
\begin{equation}
-\Lu[y,u]\leq 0 \qquad \text{in } \R.
\end{equation}
If $u$ attains a global maximum at $y_0\in\R$, then $u$ is constant on $\R$.
\end{theorem}
\begin{proof}
The proof can be found in \cite[Thm. 2]{C} if Assumption \ref{A2} (ii) holds and in in \cite[Thm. 4]{C} if Assumption \ref{A2} (i) holds.
\end{proof}
\begin{remark} \label{subordinator}\upshape
If Assumption \ref{A2} does not hold we cannot expect the Strong Maximum Principle to be true.
We show that this is always the case when the L\'evy process $Z$ is a subordinator (see Remark \ref{sub1}).
If $Z$ is a subordinator, by the properties of the measure $\nu$ \eqref{sub}, the operator $\Lu$ in \eqref{l} can be equivalently written as
\begin{equation}\notag
\Lu[y,f]=-f'(y)\cdot\Big(y+\int_0^1 z \nu(dz)\Big)+\int_0^{+\infty} (f(y+z)-f(y))\nu(dz)=0 \quad \text{in }\R. \\
\end{equation}
We set $c:=\int_0^1 z \nu(dz)$ and we take $f\in\mathcal{C}^2(\R)$, bounded and such that
\begin{equation}\notag
f'(y)>0 \quad \text{for every }y<-c \quad \text{and} \quad f(y)\equiv f(-c) \quad \text{for every }y\ge -c.\\
\end{equation}
We claim such function is a (classical) subsolution of $-\Lu[y,f]=0$.
Indeed,
$$\int_0^{+\infty}(f(y+z)-f(y)) \nu(dz)\ge0$$
for every $y$, since $f$ is nondecreasing, whereas the term $f'\cdot(y+c)$ is negative for $y<-c$ and $0$ for $y\ge-c$. On the other hand the maximum of $f$ at $-c$ propagates to the right but not to the left. Then the Strong Maximum Principle does not hold.
\end{remark}
\begin{theorem}[Liouville Property]
\label{LP} Assume that the L\'evy measure $\nu$ satisfies Assumptions \ref{A1}, \ref{A3}, \ref{A2} and
consider the problem
\begin{equation}
\label{-L}
- \Lu[y,V]=0, \qquad y\in\R, \\
\end{equation}
with $\Lu$ defined as in \eqref{l}. Then the following hold:
\begin{enumerate}
\item\label{Sub} every bounded viscosity subsolution to (\ref{-L}) is constant;
\item\label{Sup} every bounded viscosity supersolution to (\ref{-L}) is constant.
\end{enumerate}
\end{theorem}
\begin{proof}
Let $V$ be a bounded subsolution to (\ref{-L}). We can assume w.l.o.g. that $V\ge0$. Let $\varphi$
be the Lyapunov function of Lemma \ref{Lyapunov} and fix $R>0$ such that $\varphi(y)>0$ for $|y|>R$. Define, for every $\eta>0$,
\begin{equation*}
V_\eta(y)=V(y)-\eta\varphi(y)-\max_{|y|\leq R}V. \\
\end{equation*}
Observe that $V_\eta(y)\to -\infty$ as $|y|\to +\infty$ and moreover $V_\eta$ is USC. Then there exists $\overline{y}$ with $|\overline{y}|\geq R$ such that \[V_\eta(\overline{y})
\geq V_\eta(y) \qquad \forall |y|\geq R.\] Assume that $|\overline{y}|>R$. Since $\varphi\in C^2$ and it is a strict supersolution to \eqref{-L} in $|y|>R$, we get a contradiction to the fact that $V$ is a subsolution to \eqref{-L}. Then $|\overline{y}|=R$. This implies that for all $\eta>0$
\[V(y)\leq \eta \varphi(y)+\max_{|y|\leq R}V, \qquad\forall |y|\geq R\] and then, letting $\eta\to 0$, $V$ attains a maximum in the ball $|y|\leq R$. By the Strong Maximum Principle, theorem \ref{SMPFL}
$V$ is constant.
The proof of (\ref{Sup}) for bounded supersolutions is analogous.
\end{proof}
\section{The Hamilton-Jacobi-Bellman equation}
\label{4}
The HJB equation associated via dynamic programming to the value function \eqref{PO} of the control problem is
\begin{equation}
\label{HJB}
- V^\eps_t+H(x,y,D_x V^\eps,D^2_{xx}V^\eps) -
\frac{1}{\eps}\Lu[y, V^\eps]+cV^\eps= 0 ,\quad \text{in} \quad(0,T) \times \R^n_+ \times \R ,
\end{equation}
where
\begin{equation}
\label{H}
H(x,y,p,X):=\min_{u\in U}\left\{- \frac 12 \tr ( \sigma(x,y,u)\sigma^T(x,y,u)X)- f(x,y,u)\cdot p\right\}
\end{equation}
and $\Lu$ is defined in \eqref{l}. It is a partial integrodifferential equation, briefly, PIDE. The terminal condition associated to it is
\begin{equation}
\label{TC}
V^\eps(T,x,y) = g(x).
\end{equation}
Moreover, there is no natural boundary condition on the space boundary of the domain, \emph{i.e.}
$(0,T)\times\partial\R^n_+\times\R$.
The terminal boundary value problem is well posed without prescribing any boundary condition because the value function is a solution in the set $(0,T)\times\R^n_+\times\R$. The irrelevance of the space boundary $(0,T)\times\partial\R^n_+\times\R$ is essentially due to the fact that $\R^n_+\times\R$ is an invariant set for the system \eqref{sde} for all admissible control functions (almost surely); that is, the state variable cannot exit this closed domain.
\begin{proposition}\label{propvalue}
For every $\eps>0$ the value function $V^\eps$ defined in \eqref{PO} is a continuous viscosity solution to \eqref{HJB} in the set $(0, T)\times\overline{\R^n_+}\times\R$
with terminal condition \eqref{TC}.
Moreover there exists a constant $C_T$ independent of $\eps$ such that
\begin{equation}
\label{growth}
|V^\eps(t,x,y)|\leq C_T(1+|x|^2)\qquad \forall \ x\in\R^n, y\in\R, t\in [0,T].
\end{equation}
\end{proposition}
\begin{proof}
Using the boundedness of $f,\sigma$ with respect to $y$ and classical estimates on the moments of solutions of \eqref{sde} it is easy to show (see \cite[Lemma 3.1]{Ph2}) that
\[\E|X(T)|^2\leq C'_T(1+|x|^2),\] where $(X(s),Y_\eps(s)) $ is the solution to \eqref{sde} with initial data $X(t)=x$, $Y_\eps(t)=y$. Then condition \eqref{GrowthC} implies
\eqref{growth}.
Again, using moment estimates on the solutions to \eqref{sde} and the standing assumptions on the coefficients, it is possible to prove that $V^\eps$ is continuous for every $\eps$ (see \cite[Prop. 3.3]{Ph2}). Moreover, $V^\eps$ satisfies a dynamic programming principle (see \cite[Prop. 3.1 and Prop. 3.2]{Ph2}) and then by standard arguments it is a viscosity solution to \eqref{HJB}.
Finally,
by condition \eqref{zero}, all the points of the boundary of $\R^n_+\times\R$
are irrelevant, according to a Fichera-type classification of boundary points for elliptic problems. In other words, a sub- or supersolution in $(0, T)\times{\R^n_+}\times\R$ is automatically sub- or supersolution in $(0, T)\times\overline{\R^n_+}\times\R$. The argument is detailed in \cite[Prop. 3.1]{BCM} in the case in which the operator $\Lu$ is a local operator and
it adapts without modifications to our case.
\end{proof}
\section{Cell problem}
\label{5}
In this Section we define the candidate limit Cauchy problem of the singular perturbed problem \eqref{HJB} as $\eps\to0$. In particular we provide a formula for the limit Hamiltonian $\overline{H}$ as
\begin{equation}
\label{Heff}
\overline{H}(x,p,X)=\int_{\R} H(x,y,p,X)\mu(dy)
\end{equation}
where $H$ is defined in \eqref{H} and $\mu$ is the invariant measure of the process defined in \eqref{Fast} (see Proposition \ref{lem:2}).
The main tool is the ergodicity of the process $Y_\lambda(t)$ proved in Section \ref{erg}. In the following we will perform such construction in Theorem \ref{CellSold} using mainly PIDE methods.
In principle, for each fixed $(\bar{x},\bar{p},\bar{X})$ one expects the effective Hamiltonian $\overline{H}(\bar{x},\bar{p},\bar{X})$ to be a constant $k\in\R$ such that the \emph{cell problem}
\begin{equation}
\label{CP}
-\Lu[y,\chi]+H(\bar{x},y,\bar{p},\bar{X})=k \quad \text{in} \quad \R, \\
\end{equation}
has a viscosity solution $\chi$, called corrector. Actually, for our approach, it is sufficient to consider an approximate cell problem
\begin{equation}
\label{CPapproxd}
\delta\chi_\delta(y)-\Lu[y,\chi_\delta]+H(\bar{x},y,\bar{p},\bar{X})=0 \quad \text{in } \R, \\
\end{equation}
whose solution $\chi_\delta$ is also called approximate corrector.
\begin{theorem}
\label{CellSold}
For any fixed $(\bar{x},\bar{p},\bar{X})$ and $\delta>0$ the unique bounded continuous viscosity solution $\chi_\delta(y)=\chi_{\delta;\bar{x},\bar{p},\bar{X}}(y)$ to \eqref{CPapproxd} is
\begin{equation}
\chi_\delta(y)=-\E\int_0^{+\infty} H(\bar{x},Y(t),\bar{p},\bar{X})e^{-\delta t}dt, \\
\end{equation}
where $Y(t)$ solves the fast subsystem in \eqref{Fast} with $\lambda=1$ an $Y(0)=y$. Moreover,
\begin{equation}
\lim_{\delta\to 0}\delta\chi_\delta(y)=-\int_{\R} H(\bar{x},y,\bar{p},\bar{X})\mu(dy)=: -\overline{H}(x,p,X), \\
\end{equation}
locally uniformly in $y$, where $H$ is defined in \eqref{H} and $\mu$ is the unique invariant distribution of the process $Y$ defined in Proposition \ref{lem:2}.
\end{theorem}
\begin{proof}
For any fixed $(\bar{x},\bar{p},\bar{X})$, we define a function $h:\R\to\R$ as
\begin{equation} \notag
h(y):=H(\bar{x},y,\bar{p},\bar{X}) \quad \text{for every } y\in\R, \\
\end{equation}
which is bounded and Lipschitz by the standing assumptions. The existence and uniqueness of a bounded viscosity solution $\chi_\delta(y)$ follow from the Perron-Ishii method for PIDEs and the comparison principle in \cite{Sa}. Moreover by dynamic programming principle and standard arguments in Markov processes (see \cite{FS}, Chap. II.3)
we get that \begin{equation}
\chi_\delta(y)=-\E\int_0^{+\infty} h(Y(t))e^{-\delta t}dt
\end{equation}
where $Y(t)$ solves the fast subsystem in \eqref{Fast} with $\lambda=1$.
The second claim follows from Corollary \ref{Ab}.
\end{proof}
\begin{remark}\upshape
If there is no control $u$ in the system the Hamiltonian $H$ is a linear function of $(p,X)$ and $\overline H$ is obtained simply by averaging the coefficients
\[
\overline H (x,p,X)=- \frac 12 \tr \left( \int_{\R} \sigma(x,y)\sigma^T(x,y)\mu(dy) \,X\right) - \int_{\R} f(x,y)\mu(dy) \cdot p .
\]
We will us this observation in the application to asset pricing, Section \ref{uncontrol}.
\end{remark}
\section{The Convergence Theorem}
\label{6}
We state now our main result, namely, the convergence theorem for the singular perturbation problem. We will prove that the value function $V^\eps(t,x,y)$, solution to \eqref{HJB}, converges locally uniformly, as $\eps\to 0$, to a function $V(t,x)$ which can be characterized as the unique solution of the limit problem
\begin{equation}
\label{HJBlim}
\left\{ \begin{array}{ll}
-V_t+\overline{H}\left(x,D_x V,D^2_{xx}V\right)+c
=0 & \text{in} \,\, (0,T)\times\overline{\R^n_+}, \\
\\ V(T,x)=g(x) & \text{in} \,\, \overline{\R^n_+}. \\
\end{array}\right.
\end{equation}
where the Hamiltonian $\overline{H}$ has been defined respectively in \eqref{Heff}.
\begin{theorem}[Convergence Theorem]
\label{ConvTheo}
The value functions $V^\eps$ defined in \eqref{PO} converge as $\eps\to 0$ uniformly on compact subsets of $[0,T]\times\overline{\R^n_+}\times\R$ to the unique continuous viscosity solution to the limit problem \eqref{HJBlim} satisfying a quadratic growth condition in $x$, \emph{i.e.},
\begin{equation}\notag
\label{GrowthC'}
\exists\;K>0\,\,\text{such that for every } (t,x)\in[0,T]\times\R^n_+,\,\,|V(t,x)|\le K(1+|x|^2). \\
\end{equation}
\end{theorem}
\begin{proof}
The proof is divided into several steps.
\begin{step}[Relaxed semilimits] \upshape
Recall that the functions $V^\eps$ are locally equibounded in $[0,T]\times\overline{\R^n_+}\times\R$, uniformly in $\eps$ (see Proposition \ref{propvalue}). We define the half-relaxed semilimits in $[0,T]\times\overline{\R^n_+}\times\R$ (see \cite{BC-D}, Chap. V):
\begin{align*}
& \underline{V}(t,x,y):= \liminf_{\eps\to0,t'\to t,x'\to x,y'\to y} V^\eps(t',x',y'), \\
& \overline{V}(t,x,y):= \limsup_{\eps\to0,t'\to t,x'\to x,y'\to y} V^\eps(t',x',y'),
\end{align*}
for $t\le T$, $x\in\R^n_+$ and $y\in\R$.
It is immediate to get by definitions that also $\underline{V}$ and $\overline{V}$ satisfy a quadratic growth condition in $[0,T]\times\overline{\R^n_+}\times\R$, \emph{i.e.} there exists a positive constant $K_1$ such that
\begin{equation} \notag
\label{GrowthC''}
|\underline{V}(t,x,y)|\le K_1(1+|x|^2) \quad \text{and } |\overline{V}(t,x,y)|\le K_1(1+|x|^2) \\
\end{equation}
for every $(t,x,y)\in[0,T]\times\overline{\R^n_+}\times\R$. \\
\end{step}
\begin{step}[$\underline{V},\overline{V}$ do not depend on $y$] \upshape
We prove the claim only for $\overline{V}$, since the other case is completely analogous.
First of all observe that the function $\overline{V}(t,x,y)$ is a viscosity subsolution to
\begin{equation}
\label{Lu}
-\Lu[y,V]=0 \quad \text{in}\,\R. \\
\end{equation} The detailed argument is in \cite[Thm. 5.1]{BCM}. Then arguing as \cite[Lemma II.5.17]{BC-D}, we get that for every fixed $(\bar{t},\bar{x})$, also the function $y\to\overline{V}(\bar{t},\bar{x},y)$ is a subsolution to \eqref{Lu}.
So we can conclude by the Liouville property Theorem \ref{LP}, since $\overline{V}$ is bounded in $y$, that the function $y\to\overline{V}(\bar{t},\bar{x},y)$ is constant for every $(\bar{t},\bar{x})\in(0,T)\times\R^n_+$. Finally, using the definition it is immediate to see that this implies that also $\overline{V}(T,x,y)$ do not depend on $y$.
\end{step}
\begin{step}[$\underline{V}$ and $\overline{V}$ are super- and subsolutions of the limit PDE] \upshape
First we show
that $\underline{V}$ and $\overline{V}$ are super- and subsolution to \eqref{HJBlim} in $(0,T)\times\R^n_+$. We prove it
only for $\overline{V}$ since the other case is completely analogous. The proof adapts the perturbed test function method introduced by Evans \cite{E} for periodic homogenization and
developed in \cite{AB2, BCM} for singular perturbations.
We fix $(\bar{t},\bar{x})\in(0,T)\times\R^n_+$ and we show that $\overline{V}$ is a viscosity subsolution at $(\bar{t},\bar{x})$ of the limit PIDE. This means that if $\psi$ is a smooth function such that $\psi(\bar{t},\bar{x})=\overline{V}(\bar{t},\bar{x})$ and $\overline{V}-\psi$ has a maximum at $(\bar{t},\bar{x})$ then
\begin{equation}
-\psi_t(\bar{t},\bar{x})+\overline{H}(\bar{x},D_x\psi(\bar{t},D^2_{xx}\psi(\bar{t},\bar{x}))+c\overline{V}(\bar{t},\bar{x})\le 0.
\end{equation}
Without loss of generality we assume that the maximum is strict in $B((\bar{t},\bar{x}),r)$
and that $0<\bar{t}-r<\bar{t}+r<T$ and $\bar{x}_i>r$ for all $i$.
We consider now the Lyapunov function $\phi\in\mathcal{C}^2(\R)$ as in Lemma \ref{Lyapunov}.
By adding a constant to $\phi$ if necessary, we can assume that $-\Lu[y,\phi]\ge0$ in $\R$. We fix $\bar{y}\in\R$, such that $\phi(\bar{y})=\min_{\R} \phi$.
We can also assume that $\phi(\bar{y})<\min_{|y-\bar{y}|\ge
R}\phi$ for $
$ sufficiently large.
Let now $\eta>0$ and take $\delta>0$ sufficiently small such that if $\chi_\delta$ is the solution of \eqref{CPapproxd} at $(\bar{x},D_x\psi(\bar{t},\bar{x}),D^2_{xx}\psi(\bar{t},\bar{x}))$ (see Theorem \ref{CellSold}), then
\begin{equation}
\label{deltabound}
|\delta\chi_\delta(y)+\overline{H}(\bar{x},D_x\psi(\bar{t},\bar{x}),D^2_{xx}\psi(\bar{t},\bar{x}))|\le\eta \quad \text{for any } y\in B(\bar{y},
). \\
\end{equation}
We define the perturbed test function as
\begin{equation}
\psi^\eps(t,x,y):=\psi(t,x)+\eps\chi_\delta(y)+\phi(y). \\
\end{equation}
Observe that
\begin{equation}
\limsup_{\eps\to0,t'\to t,x'\to x, y'\to y} V^\eps(t',x',y')-\psi^\eps(t',x',y')=\overline{V}(t,x)-\psi(t,x)-\phi(y). \\
\end{equation}
Arguing as in \cite[Lemma V.1.6]{BC-D} we get
sequences $\eps_n\to0$ and $(t_n,x_n,y_n)\in B:=B((\bar{t},\bar{x})
r
\times B(\bar y, R)$ such that $(t_n,x_n,y_n)\to(\bar{t},\bar{x},\hat{y})$ for some $\hat{y}\in B(\bar{y},R)$ such that $\phi(\hat{y})=\phi(\bar{y})$, and, as $n\to+\infty$,
\begin{equation} \notag
V^{\eps_n}(t_n,x_n,y_n)-\psi^{\eps_n}(t_n,x_n,y_n)\to\overline{V}(\bar{t},\bar{x})-\psi(\bar{t},\bar{x})-\phi(\hat{y}) \\
\end{equation}
and $(t_n,x_n,y_n)$ is a maximum of $V^{\eps_n}-\psi^{\eps_n}$ in $B$.
Then, using the fact that $V^\eps$ is a subsolution to \eqref{HJB}, we get
\begin{equation}
-\psi_t+H(x_n,y_n,D_x\psi,D^2_{xx}\psi)+cV^\eps-\Lu[y_n,\chi_\delta]-\frac{1}{\eps_n}\Lu[y_n,\phi] \le0 \\
\end{equation}
where $V^\eps$, $\psi$, $\chi_\delta$ and $\phi$ (and their derivatives) are computed respectively in $(t_n,x_n,y_n)$, $(t_n,x_n)$ and in $y_n$. Using the fact that $\phi$ satisfies $-\Lu[\cdot,\phi]\ge0$, we get from the previous inequality that
\begin{equation}
-\psi_t+H(x_n,y_n,D_x\psi,D^2_{xx}\psi)+cV^{\eps_n}-\Lu[y_n,\chi_\delta] \le0. \\
\end{equation}
We now recall that $\chi_\delta$ solves the $\delta$-cell problem \eqref{CPapproxd
, thus
\begin{gather}
-\psi_t(t_n,x_n)+H(x_n,y_n,D_x\psi(t_n,x_n),D^2_{xx}\psi(t_n,x_n)) \notag \\
-H(\bar{x},y_n,D_x\psi(\bar{t},\bar{x}),D^2_{xx}\psi(\bar{t},\bar{x}))-\delta\chi_\delta(y_n)+cV^{\eps_n}(t_n,x_n,y_n)\le0. \notag \\
\end{gather}
By taking the limit as $n\to+\infty$ the second and the third term of the left-hand side of this inequality cancel out. Next we use \eqref{deltabound} to replace $-\delta\chi_\delta$ with $\overline{H}-\eta$ and get that
\begin{equation}
-\psi_t(\bar{t},\bar{x})+\overline{H}(\bar{x},D_x\psi(\bar{t},\bar{x}),D^2_{xx}\psi(\bar{t},\bar{x}))+c\overline{V}(\bar{t},\bar{x})\le\eta. \\
\end{equation}
Finally, since $\eta>0$ is arbitrary, we conclude. \\
To prove that $\underline{V}$ is a supersolution to \eqref{HJBlim} we proceed exactly in the same way, just taking as a perturbed test function
\begin{equation} \notag
\psi^\eps(t,x,y)=\psi(t,x)+\eps\chi_\delta(y)-\phi(y). \\
\end{equation}
Finally, we claim that $\overline{V}$ and $\underline{V}$ are respectively a sub and a supersolution to \eqref{HJBlim}
also at the boundary of $\R^n_+$. In this case it is sufficient to repeat exactly the same
argument as in \cite[Prop. 3.
]{BCM} to get the conclusion, recalling that
the Hamiltonian $\overline{H}$ is defined as
\[
\overline{H}(x,p,X)=\int_{\R} \min_{u\in U}\left\{- \frac 12 \tr ( \sigma(x,y,u)\sigma^T(x,y,u)X)- f(x,y,u)\cdot p\right\} \mu(dy)\]
and $f, \sigma$ satisfy \eqref{zero}.
\end{step}
\begin{step}[Uniform convergence] \upshape
We observe that by definition $\overline{V}\ge\underline{V}$ and that both $\overline{V}$ and $\underline{V}$ satisfy the same quadratic growth condition \eqref{GrowthC''}. Moreover, the Hamiltonian $\overline{H}$ defined in \eqref{Heff} inherits all the regularity properties of $H$ defined in the first section. Therefore, we can apply the comparison result in \cite{DL-L} between sub- and supersolutions to parabolic PDE problems satisfying a polynomial growth condition, to deduce that $\overline{V}\le\underline{V}$. Therefore,
\begin{equation} \notag
\overline{V}=\underline{V}=:V. \\
\end{equation}
In particular, $V$ is continuous, and by Lemma V.1.9 in \cite{BC-D}, this implies that $V^\eps$ converges locally uniformly to $V$.
\end{step}
\end{proof}
\begin{remark}\upshape
The idea of adding or subtracting the Lyapunov function $\phi$ to the perturbed test function $\psi^\epsilon$ seems to be new and is an appropriate tool for extending the methods of \cite{E, AB2} from the periodic case to the present case of unbounded fast variables $y$. It applies also to the proof of Thm. 5.1 of \cite{BCM}, therefore filling in a gap of that proof.
\end{remark}
\begin{remark}
\label{effcontrol}\upshape
The solution $V$ of the limit Cauchy problem \eqref{HJBlim} can be represented as the value function of a new control problem obtained by a relaxation procedure proposed in \cite{BTer} for deterministic systems. Define the extended control set
\[
U^{ex}:= L^1((\R,\mu); U)
\]
and note that it contains a copy of $U$, given by the constant functions.
Extend to $U^{ex}$ the drift and the diffusion of the system \eqref{sde} as follows
\[
\hat\sigma \hat\sigma^T(x,\beta):=\int_{\R} \sigma \sigma^T(x,y,\beta(y)) \mu(dy) , \quad \hat f (x,\beta):= \int_{\R} f((x,y,\beta(y)) \mu(dy), \quad \beta\in U^{ex}.
\]
Then the measurable selection argument in \cite{BTer} allows to prove that
\[
\int_{\R} H(x,y,p,X) \mu(dy) = \inf_{\beta\in U^{ex}}\left\{- \frac 12 \tr (\hat\sigma \hat\sigma^T(x,\beta) X)- \hat f (x,\beta)\cdot p\right\} ,
\]
i.e., $\overline H$ is a Bellman Hamiltonian. By uniqueness of viscosity solutions to \eqref{HJBlim} we can conclude that
\[
V(t,x)=\sup_{\beta\in\mathcal{U}^{ex}}\E[e^{c(t-T)}g(\hat X(T))\,|\, \hat X(t)=
] ,
\]
where $\hat X$ solves
\[
d\hat X(s)=\hat f(\hat X(s),\beta(s)) ds + \hat\sigma(\hat X(s),\beta(s))dW(s)
\]
and $\mathcal{U}^{ex}$ denotes the progressively measurable processes taking values in ${U}^{ex}$. This is an \emph{effective control problem} associated to the general multiscale control problem of Section \ref{2}. In the application to Merton's portfolio optimisation,
Section \ref{effMert}, however, we will find a simpler representation of the limit control problem by exploiting the explicit form of the Hamiltonian in that case.
\end{remark}
\section{Applications and examples
\label{7}
\subsection{Asset pricing} \label{uncontrol}
We consider $
$ underlying risky assets with price $X^i$ evolving according to the standard lognormal model:
\begin{equation}
\label{SDE1}
\left\{ \begin{array}{ll}
dX^i(t)=\alpha^i X^i(t)dt+\sqrt{2}X^i(t)\sigma_i(Y_\eps(t))dW^i(t), \quad i=1,\dots,n, \\
dY_\eps(t)=-\frac{1}{\eps}Y_\eps(t)dt+dZ(\frac{t}{\eps})
\end{array} \right.
\end{equation}
where $X^i(t_0)=x^i\ge0$, $\sigma_i:\R\to\R$ are bounded Lipschitz continuous function,
$\sigma_i(y)\geq 0$ for all $y$, $i=1,\dots,n$, the processes $W=(W^1,\dots,W^n)$ and $Z$ are independent and, respectively, a standard $n
-dimensional Brownian motion and a pure jumps L\'evy process with L\'evy measure satisfying the Assumptions \ref{A1}, \ref{A3}, \ref{A2}.
The problem we consider here is the pricing of an European option given by a non-negative payoff function $g$ depending on the underlying $X^i$ and by a maturity time $T$. According to risk-neutral theory, to define a no-arbitrage derivative price we have to use an equivalent martingale measure $\Q$ under which the discounted stock prices $e^{-rt}X^i(t)$ are martingales, where $r$ is the instantaneous interest rate for lending or borrowing money. Nevertheless, we assume as in \cite{FPS} that the process $Z$ remains unchanged under the equivalent martingale measure $\Q$. Thus the system, under a risk-neutral probability $\Q$, \eqref{SDE1} writes as
\begin{equation}
\label{SDE2}
\left\{ \begin{array}{ll}
dX^i(t)=r X^i(t)dt +\sqrt{2} X^i(t)\sigma_i(Y_\eps(t)
dW^{i,\Q}(t), \quad i=1,\dots,n, \\
dY_\eps(t)=-\frac{1}{\eps}Y_\eps(t)dt+dZ(\frac{t}{\eps}). \\
\end{array} \right.
\end{equation}
In this setting, an European contract has no-arbitrage price given by the formula
\begin{equation}
\label{Veps}
V^\eps(t,x,y):=\E^\Q[e^{c(t-T)}g(X(T))\,|\,X^i(t)=x^i,\,Y_\eps(t)=y], \quad 0\le t \le T
\end{equation}
where $c>0$ and the payoff function $g$ satisfies \eqref{GrowthC}.
The (linear) HJB equation associated with the price function is
\begin{equation}
\label{HJBp}
\left\{ \begin{array}{ll}
-V^\eps_t
-\sum_{i=1}^{n}x_i^2 \sigma_i^2(y) V^\eps_{x_ix_i}
-r x \cdot D_x V^\eps -\frac{1}{\eps}\Lu[y, V^\eps]+cV^\eps=0 \quad &\text{in } (0,T)\times\overline{\R^n_+}\times\R , \\ \\
V^\eps(T,x,y)=g(x) \quad &\text{in } \overline{\R^n_+}\times\R , \\
\end{array} \right.
\end{equation}
where
$\Lu$ is as in \eqref{l}.
Since all the assumptions are satisfied, the convergence theorem holds, and the prices $V^\eps(t,x,y)$ converge locally uniformly, as $\eps\to0$, to the unique viscosity solution $V(t,x)$ of the limit equation
\begin{equation}
\label{HJBeff}
\left\{ \begin{array}{ll}
- V_t- \sum_{i=1}^{n}x_i^2 \int_{\R} \sigma_i^2(y)\mu(dy) \,V^\eps_{x_ix_i
-r x \cdot D_xV+cV=0 \quad &\text{in } (0,T)\times\overline{\R^N_+}, \\ \\
V(T,x)=g(x), \quad &\text{in } \overline{\R^N_+}. \\
\end{array} \right.
\end{equation}
Then $V$ can be represented as
\begin{equation}
V(t,x):=\E^\Q[e^{c(t-T)}g(\overline{X}(T))\,|\,\overline{X}(t)=x], \quad 0 \le t \le T, \\
\end{equation}
where $\overline{X}(t)$ satisfies the averaged \emph{effective system}
\begin{equation}
\label{effSDE}
d\overline{X}^i(t)=r\overline{X}^i(t)dt+\sqrt{2} \overline{\sigma}_i\overline{X}^i(t) dW^{i,\Q}(t), \\
\end{equation}
whose (constant) volatility $\overline{\sigma}_i$ is the so-called mean historical volatility for the $i$-th asset
\begin{equation}
\overline{\sigma}_i:
\left(\int_{\R} \sigma_i^2(y) \mu(dy)\right)^{\frac 12} .
\end{equation}
Therefore the limit of the pricing problem as $\eps\to0$ is a new pricing problem for the effective system \eqref{effSDE}.
\begin{remark}
\label{correlation}
\upshape
The choice of a diagonal matrix $\sigma$ in \eqref{SDE1} is made only for notational simplicity. As in the previous sections and in \cite{BCM} we can replace the term $\sigma_i(Y_\eps(t))dW^i(t)$ in \eqref{SDE1} with the $i$-th component of $\sigma(Y_\eps(t))dW(t)$ for a $n\times r$ matrix $\sigma$, therefore allowing that the components of the Brownian motion acting on different asset prices be correlated. In this case instead of an effective volatility $\overline{\sigma}_i$ for each asset we find an effective matrix $\overline{\sigma}$ such that $\overline{\sigma}\,\overline{\sigma}^T= \int_{\R} \sigma(y) \sigma^T(y)\mu(dy)$.
\end{remark}
\subsection{Merton's portfolio optimization problem}
\subsubsection{The convergence result}
We consider now another classical problem in finance, the Merton's optimal portfolio allocation, under the assumption of fast oscillating stochastic volatility.
We consider a financial market consisting of a nonrisky asset $S$ evolving according to the deterministic equation
\begin{equation} \notag
dS(t)=r S(t) dt, \\
\end{equation}
with $r>0$, and $n$ risky assets $X^i(t)$ evolving according to the stochastic system
\eqref{SDE1}. We denote by $\W$ the wealth of an investor. The investment policy - which will be the control input - is defined by a progressively measurable process $u$ taking values in a compact set $U$, and $u^i_t$ represents the proportion of wealth invested in the asset $X^i(t)$ at time $t$. Then the wealth process evolves according to the following system:
\begin{equation}
\label{SDE3}
\left\{ \begin{array}{ll}
d\W(t)=\W(t) \bigl( r + \sum_{i=1}^n((\alpha^i-r) u^i(t))\bigr) dt + \sqrt{2}\W(t) \bigl(\sum_{i=1}^N u^i(t)\sigma_i(Y_\eps(t))\bigr) dW(t), \\
dY_\eps(t)=-\frac{1}{\eps}Y_\eps(t) dt+dZ(\frac{t}{\eps})
\end{array} \right.
\end{equation}
where $\W(t_0)=w>0$. \\
The Merton's problem consists in choosing a strategy $u_.$ which maximizes a given utility function $g$ at some final time $T$. In particular the problem can be described in terms of the value function
\begin{equation}\notag
V^\eps(t,w,y):=\sup_{u_.\in\mathcal{U}}\E[g(\W(T))\,|\,\W(t)=w,Y_\eps(t)=y]. \\
\end{equation}
Tipically the utility functions in financial applications are chosen in the class of HARA (hyperbolic absolute risk aversion) functions $g(w)=a(bw+c)^\gamma$, where $a,b,c$ are constants, and $\gamma\in(0,1)$ is a given coefficient called the relative risk premium coefficient. The HJB equation associated with the Merton value function is
\begin{equation}\notag
-V^\eps_t+H_M(w,y,D_wV^\eps,D^2_{ww}V^\eps)-\frac{1}{\eps}\Lu[y,V^\eps]=0, \\
\end{equation}
in $(0,T)\times\R_+\times\R$, complemented with the terminal condition $V^\eps(T,w,y)=g(w)$. The integro-differential operator $(1/\eps)\Lu$ is the infinitesimal generator of the process $Y_\eps$, with $\Lu$ as in \eqref{l},
and $H_M(w,y,p,X)$ is defined as
\begin{equation}\notag
H_M(w,y,p,X)=\min_{u\in U}\left\{-\left(\sum_{i=1}^n u^i\sigma_i(y)\right)^2 w^2 X - \left[r+\sum_{i=1}^n (\alpha^i-r)u^i\right]w p\right\}. \\
\end{equation}
Our main theorem applies also in this case and states that the value function $V^\eps$ converges locally uniformly to the unique solution of the limit problem
\begin{equation}
\label{SDE4}
\left\{ \begin{array}{ll}
- V_t+\int_{\R}H_M(w,y,D_wV,D^2_{ww}V)\mu(dy)=0 \quad &\text{for } t\in(0,T), \quad w>0, \\
V(T,w)=g(w) \quad &\text{for } w>0, \\
\end{array} \right.
\end{equation}
where $\mu$ is the invariant distribution associated with the fast subsystem. This convergence result in the framework of L\'evy processes is new in the literature and extends and complements the results stated in \cite{BCM} for fast volatility processes driven by Brownian motion.
\begin{remark}
\upshape
As in the Remark \ref{correlation} of the previous section we can allow the correlation among the noises acting on different assets, as in \cite{BCM}, Sect. 6.2.
\end{remark}
\subsubsection{The effective Merton's problem}
\label{effMert}
Next we want to interpret the effective PDE in \eqref{SDE4} as the HJB equation of a suitable \emph{effective control problem}, simpler than the general one described in Remark \ref{effcontrol}.
For simplicity we restrict ourselves to the case of a single risky asset, \emph{i.e.}, $n=1$. The equation for the wealth becomes
\begin{equation}\notag
d\W(t)=\W(t) (r+(\alpha-r)u(t)) dt + \sqrt{2}\W(t)u(t)\sigma(Y_\eps(t)) dW(t), \quad \alpha>r, \\
\end{equation}
and the HJB equation for $V^\eps$ is
\begin{equation}\label{PIDE5}
-V^\eps_t-\max_{u\in U}\Bigl\{u^2 \sigma^2(y) w^2 V^\eps_{ww} + [r+(\alpha-r)u] w V^\eps_w\Bigr\}=\frac{1}{\eps}\Lu[y,V^\eps]. \\
\end{equation}
Therefore the effective PDE in \eqref{SDE4} is
\begin{equation}\label{braces}
- V_t- r w V_w -\int_{\R}\max_ {u\in U}\Bigl\{u^2\sigma^2(y)w^2 V_{ww} +
(\alpha-r)
w V_w\Bigr\}\mu(dy)=0. \\
\end{equation}
Now we assume that the utility function $g$ is increasing and concave with $g''<0$. Then one expects that the solution $V$ is also increasing in the initial wealth $w$ with $V_{ww}<0$. In this case the expression in braces in \eqref{PIDE5} and \eqref{braces} is a concave parabola in $u$ and the maximum in the Hamiltonian is easily computed if it is attained in the interior of $U$, e.g., if $U$ is very large. Then one gets, at least formally,
\[
H_M(w,y,p,X)=\frac{(\alpha - r)^2 p^2}{4\sigma^2(y) X}
\]
and the effective PDE \eqref{braces} becomes
\begin{equation}\notag
- V_t- r w V_w + \int_{R}\frac{1}{\sigma^2(y)}\mu(dy) \frac{(\alpha - r)^2 V_{w}^2}{4 V_{ww}}=0 .
\end{equation}
This is the HJB equation of the \emph{Merton's problem with constant volatility $\overline{\sigma}>0$}, where the wealth dynamics is
\begin{equation}\label{constant_v}
d\W(t)=\W(t)(r+(\alpha-r)u(t))dt+\sqrt{2}\W(t) u(t)\overline{\sigma} dW(t), \\
\end{equation}
if and only if
\begin{equation}\label{sigma}
\overline{\sigma}:=\Bigl(\int_{\R}\frac{1}{\sigma^2(y)}\mu(dy)\Bigr)^{-\frac{1}{2}}. \\
\end{equation}
Therefore this is the correct parameter to use in a Merton model with constant volatility if we consider it as an approximation of a model with fast and ergodic stochastic volatility. We can call it the \emph{effective Merton problem}.
We point out that the \emph{effective volatility} $\overline{\sigma}$ for the Merton problem is the harmonically averaged long-run volatility, that is smaller than the usual mean historical volatility
\begin{equation}\notag
\tilde{\sigma}:=\Bigl(\int_{\R} \sigma^2(y) \mu(dy)\Bigr)^{\frac{1}{2}}
\end{equation}
in uncontrolled systems, see Section \ref{uncontrol}. Therefore the use of the correct parameter $\overline{\sigma}$ in the model leads to an increase of the value function, \emph{i.e.}, of the optimal expected utility.
\subsubsection{The solution for HARA utility}
In some cases the Cauchy problem for the effective equation\eqref{SDE4} can be solved explicitly, giving also a more rigorous derivation of the formula \eqref{sigma} for the effective volatility. Let us take as a constraint on the control $u(t)$
the interval
\begin{equation} \notag
U:=[R_1,R], \quad \text{with } -R\le R_1\le 0 < R, \\
\end{equation}
and as terminal cost the HARA function
\begin{equation} \notag
\label{UtFunct}
g(w)=a\frac{w^\gamma}{\gamma}, \quad 0<\gamma<1, \quad a>0.
\end{equation}
Since the terminal condition is now
$V(T,w)=
{w^\gamma}/{\gamma},$
we look for solutions of the form
$$V(t,w)=\frac{w^\gamma}{\gamma}v(t) \quad \text{with }\; v(t)\ge 0.
$$
By plugging it into the Cauchy problem we get
\begin{equation} \notag
\dot{v}=-\gamma\overline{h}v, \quad v(T)=a, \quad \overline{h}:=r+\int_{\R}\max_{u\in U}\bigl\{(\alpha-r)u+(\gamma-1)\sigma^2(y)u^2\bigr\}\mu(dy). \\
\end{equation}
Therefore the uniqueness of solution gives
\begin{equation}
\label{Sol}
V(t,w)=a\exp{\Bigl\{\gamma\bar{h}(T-t)\Bigr\}}\frac{w^\gamma}{\gamma}, \quad 0<t<T. \\
\end{equation}
We compute the rate of exponential increase $\overline{h}$ and get
\begin{align*}\notag
\overline{h}=r\,+&\int_{\{y\,:\,2R(1-\gamma)\sigma^2(y)<\alpha-r\}}\bigl[(\alpha-r)R+(\gamma-1)R^2\sigma^2(y)\bigr]\mu(dy) \notag \\
&+\int_{\{y\,:\,2R(1-\gamma)\sigma^2(y)\ge\alpha-r\}}\frac{(\alpha-r)^2}{4(1-\gamma)\sigma^2(y)}\mu(dy). \notag \\
\end{align*}
This formula simplifies considerably if the $\mu$-probability of the set $\{y\,:\,2R(1-\gamma)\sigma^2(y)\ge\alpha-r\}$ is 1, e.g.,
for large upper bound $R$ on the control. In fact we get
\begin{equation}\notag
\overline{h}=r+ \frac{(\alpha-r)^2}{4(1-\gamma)}\int_{\R} \frac{1}{\sigma^2(y)}\mu(dy) . \\
\end{equation}
With this expression for $\overline{h}$ the function $V$ given by \eqref{Sol} coincides with the classical Merton formula solving the
problem with constant volatility $\overline{\sigma}>0$ (for $2R(1-\gamma)\overline{\sigma}\ge\alpha-r$) if and only if $\overline{\sigma}$ is given by \eqref{sigma}, i.e.,
\begin{equation}\notag
V(t,w) = a\exp{\Bigl\{\gamma\left[r+\frac{(\alpha-r)^2}{4(1-\gamma)\overline{\sigma}^2}\right](T-t)\Bigr\}}\frac{w^\gamma}{\gamma}. \\
\end{equation}
\section*{Acknowledgements} We are grateful to Carlo Sgarra for pointing out to us the literature on non-Gaussian models of stochastic volatility and to Tiziano Vargiolu
for several discussions on L\'evy processes. The main results of this research were presented also in the third author's Master Thesis, defended in February 2014.
|
\section{Introduction}
With the static color charges one observes on the lattice
a color-electric flux tube \cite{B}, which, if the distance
between the static quarks is large, can be approximated as a string.
This string is nondynamical (in the sense that its ends are fixed).
In the light quark systems a crucial aspect is a relativistic
motion of quarks at the ends of a possible string. There is no
consistent theory of the dynamical QCD string with quarks at the ends.
It is even apriori unclear whether such a picture has something to do with
reality. In this case the chiral symmetry as well as its
dynamical braking should be relevant.
Both confinement and chiral symmetry breaking dynamics
are important for the hadronic mass generation. Their
interplay is responsible for a rather complicated structure
of the hadron spectra in the light quark sector. A large
degeneracy is seen in the highly excited mesons \cite{G1,G2},
which is, however, absent in the observed spectrum below 1.8 GeV.
The low-lying hadron spectra should be strongly affected by the
chiral symmetry breaking dynamics.
In order to disentangle the confinement physics from the
chiral symmetry breaking dynamics we remove on the lattice
from the valence quark
propagators the lowest-lying quasi-zero modes of the Dirac operator
keeping at the same time the gluon gauge configurations intact
\cite{LS,GLS,DGL}. Indeed,
the quark condensate of the vacuum is related to
a density of the lowest quasi-zero eigenmodes of the
Dirac operator \cite{C}:
\begin{equation}\label{BC}
< 0 | \bar q q | 0 > = - \pi \rho(0).
\end{equation}
\noindent
We subtract from the valence quark propagators their lowest-lying
chiral modes, which are a tiny part of the full amount of modes,
\begin{equation}\label{eq:RD}
S_{RD(k)}=S_{Full}-
\sum_{i=1}^{k}\,\frac{1}{\lambda_i}\,|\lambda_i\rangle \langle \lambda_i| \\,
\end{equation}
\noindent
where $\lambda_i$ and $|\lambda_i\rangle$ are the eigenvalues and
the corresponding eigenvectors of the
Dirac operator. Given these reduced quark propagators we study the existence
of hadrons and
their masses upon reduction of the lowest-lying modes responsible for the
chiral symmetry breaking.
\section{Lattice technology}
In contrast to refs. \cite{LS,GLS}, where a chirally not
invariant Wilson lattice Dirac operator was employed, we adopt now
\cite{DGL}
a manifestly chiral-invariant overlap Dirac operator \cite{N}.
The quark propagators were generously provided by the JLQCD
collaboration \cite{KEK,Aoki:2012pma,Noaki:2008iy}. For some technical details of the work
we refer the reader to ref. \cite{DGL}.
\section{Observations, symmetries and their string interpretation}
At this first stage of the study we have investigated
the ground state and the excited states of all possible $\bar q q$
isovector
$J=1$ mesons, i.e., $\rho (1^{--})$, $a_1 (1^{++})$, $b_1 (1^{+-})$.
An exponential decay of the correlation function is interpreted as
a physical state and its mass is extracted. The quality of the
exponential decay improves with the reduction of the lowest modes.
The evolution of masses of the ground and excited states
upon reduction of the low-lying modes is shown in Fig. 1.
\begin{table}[t]
\caption{The complete set of $q\bar{q}$ $J=1$ states
classified according to $SU(2)_L \times SU(2)_R$.
The symbol $\leftrightarrow$ indicates the states belonging to
the same representation $R$
that must be degenerate in the chirally symmetric world.}
\begin{center}
\begin{tabular}{cc}
$R$ & mesons\\
\hline
$(0,0)$&$\omega(I=0,1^{--}) \leftrightarrow f_1(I=0,1^{++})$\\
$(1/2,1/2)_a$&$\omega(I=0,1^{--}) \leftrightarrow b_1(I=1,1^{+-})$\\
$(1/2,1/2)_b$&$h_1(I=0,1^{+-}) \leftrightarrow \rho(I=1,1^{--}) $\\
$(0,1) \oplus (1,0)$&$a_1(I=1,1^{++} )\leftrightarrow \rho(I=1,1^{--})$\\
\end{tabular}\label{tab:t1}
\end{center}
\end{table}
At the truncation energy about 50 MeV an onset of a degeneracy of the
states , $\rho, \rho', a_1,b_1$, as well as a degeneracy of their excited
states is seen. This degeneracy indicates a symmetry.
All possible multiplets of the $SU(2)_L \times SU(2)_R$ group for the $J=1$
mesons are shown in Table 1 \cite{G1,G2}. In the chirally symmetric world
there
must be two independent $\rho$-mesons that belong to different
chiral representations. The first one is a member of the $(0,1)+(1,0)$
representation. It can be created from the vacuum only by the operators
that have the same chiral structure, e.g. by the vector-isovector current
$\bar q \gamma^i \vec \tau q$.
Its chiral partner is the axial vector meson, $a_1$, that is created by
the axial-vector isovector current. When chiral symmetry is restored this $\rho$-meson
must be degenerate with the $a_1$ state. Another $\rho$-meson, along with
its chiral partner $h_1$, is a member of the
$(1/2,1/2)_b$ representation. It can be created only by the
operators that have the same
chiral structure, e.g. by $\bar q \sigma^{0i} \vec \tau q$.
In the real world (i.e. with broken chiral symmetry)
each $\rho$-state (i.e. $\rho$ and $\rho'$) is a mixture of these two
representations and they are well split. Upon subtraction of 10 lowest
modes we observe two independent degenerate $\rho$-mesons. One of them
couples only to the vector current and does not couple to the
$\bar q \sigma^{0i} \vec \tau q$ operator. The other $\rho$ meson - the
other way around. This means that one of these degenerate $\rho$-states
belongs to the $(0,1)+(1,0)$ multiplet and the other one is a member
of the $(1/2,1/2)_b$ representation.
A degeneracy of the $(0,1) \oplus (1,0)$ $\rho$-meson with the $a_1$ meson
is a clear signal of the chiral $SU(2)_L \times SU(2)_R$ restoration.
Consequently, a similar degeneracy should be observed in all other chiral pairs
from Table 1.
The $U(1)_A$ symmetry transforms the $b_1$ state into the $(1/2,1/2)_b$
$\rho$-meson \cite{G1,G2}. Their degeneracy indicates a restoration of
the $U(1)_A$ symmetry. We conclude
that simultaneously both $SU(2)_L \times SU(2)_R$ and $U(1)_A$ symmetries
get restored.
The restored $ SU(2)_L \times SU(2)_R \times U(1)_A$ symmetry requires a
degeneracy of four mesons that belong to $(1/2,1/2)_a$
and $(1/2,1/2)_b$ chiral multiplets \cite{G1,G2}, see Table \ref{tab:t1}.
This symmetry does not require, however, a degeneracy of these four states with other
mesons, in particular with $a_1$ and its chiral partner $\rho$. We clearly
see the latter degeneracy. This implies that there is some higher symmetry,
that includes
$ SU(2)_L \times SU(2)_R \times U(1)_A $ as a subgroup. This higher symmetry requires a
degeneracy of all eight mesons from the Table \ref{tab:t1}.
All these eight mesons can be combined to a reducible $\bar q q$
representation, which is a product of two fundamental chiral
quark representations \cite{CJ}:
\begin{eqnarray}
[(0,1/2)+(1/2,0)] \times [(0,1/2)+(1/2,0)] =\hspace{15mm} \nonumber\\
(0,0) + (1/2,1/2)_a +(1/2,1/2)_b + [(0,1)+ (1,0)]\;.\hspace{4mm}
\end{eqnarray}
They exhaust all possible chiralities of quarks and antiquarks, i.e.,
their spin orientations,
as well as possible spatial and charge parities for non-exotic mesons.
The observed degeneracy of all these eight mesons suggests that the higher
symmetry, mentioned above, should combine
the $U(1)_A$ and the $ SU(2)_L \times SU(2)_R $ rotations in the isospin
space with the
$SU(2)_S$ spin-symmetry. The latter one is due to independence of the energy
on the orientations
of the quark spins. The higher symmetry group that combines
all eight mesons from the Table 1 into one multiplet
of dimension 16 should be
$ SU(2 \cdot N_f)$.
The quark-spin independence of the energy levels implies that there
are no magnetic interactions in the system, i.e., the spin-orbit
force, the color-magnetic (hyperfine) and tensor interactions
are absent \cite{G3}.
The energy of the system is entirely due to interactions of the color
charges via the color-electric field and due to a relativistic motion
of the system. We interpret (or define) such a system as a dynamical QCD
string. Note a significant qualitative difference with the case of a motion of
an electrically charged fermion in a static electric field or with a relative motion
of two fermions with the electric charge.
In the latter cases there exist
spin-orbit and spin-spin forces that are a manifestation of the magnetic interaction in the system.
In our case such a magnetic interaction is absent, which implies that both
quarks are at rest with respect to the electric field (that
moves together with
quarks). It is this circumstance that suggests to interpret (define) our system as a
dynamical QCD string.
The observed radial levels at the truncation energy 65 MeV,
at which we see the onset of the symmetry,
are approximately equidistant and can be described through
the simple relation
\begin{equation}
E_{n_r} = (n_r +1)\hbar \omega,~~~ n_r=0,1,...
\end{equation}
The extracted value of the radial
string excitation quantum
amounts to $ \hbar \omega = (900\pm70)$ MeV.
At the moment we cannot exclude, however,
the quadratic relation $E_{n_r}^2 \sim (n_r+1)$,
because the excited level can be shifted up due to rather
small finite lattice volume.
There is an interesting aspect of the dynamical
QCD string that crucially distinguishes it from the
Nambu-Goto open string. The energy of the the Nambu-Goto open bosonic
string is determined by its orbital angular momentum $L$, $M^2 \sim L$.
For the dynamical QCD string that contains chiral quarks at the ends
the orbital angular momentum $L$ of the relative motion
is not conserved \cite{GN}.
For instance, two orthogonal $\rho$-mesons at the same energy
level are represented by the mutually orthogonal fixed superpositions of the
$S$- and $D$-waves.
\noindent
\begin{eqnarray}
\displaystyle |(0,1)+(1,0);1 ~ 1^{--}\rangle&=&\sqrt{\tfrac23}\,|1;{}^3S_1\rangle+\sqrt{\tfrac13}\,|1;{}^3D_1\rangle,\nonumber\\
\displaystyle |(1/2,1/2)_b;1 ~ 1^{--}\rangle&=&\sqrt{\tfrac13}\,|1;{}^3S_1\rangle-\sqrt{\tfrac23}\,|1;{}^3D_1\rangle.\nonumber
\end{eqnarray}
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=0.4\textwidth]{meff_fit_RD}
\caption{Evolution of hadron masses under the low-mode truncation.
Both the number $k$ of the removed lowest eigenmodes as well as
the corresponding energy gap $\sigma$ are given.}\label{fig:histo}
\end{center}
\end{figure}
\bigskip
{\bf Acknowledgments}
I am very grateful to M. Denissenya and C. B. Lang
for our common efforts in lattice simulations. The JLQCD collaboration
is acknowledged for their suggestion to use their overlap gauge
configurations and
quark propagators. This work is partially supported by
the Austrian Science
Fund (FWF) through the grant P26627-N16.
|
\section{Introduction}
\label{sec:Intro}
In this note, I am going to address a rather classical problem of the low-temperature physics and the theory of Bose-systems specifically. The excitation spectrum of a Bose-liquid was an important element in the formulation of the theory of superfluidity of liquid helium-4. Phenomenologically formulated by Landau \cite{Landau:1941,Landau:1947}, the spectrum was microscopically derived by Bijl \cite{Bijl:1940}, Bogoliubov \cite{Bogoliubov:1947}, Feynman and Cohen \cite{Feynman:1954,Feynman&Cohen:1956}. In the following decades, numerous works on this subject appeared \cite{Jackson&Feenberg:1962,Sunakawa_etal:1969,Vakarchuk&Yukhnovskii:1980,%
Apaja_etal:1998,Kruglov&Collett:2001,Pashitskii_etal:2002,Adamenko_etal:2003,%
Kruglov&Collett:2008,Fak_etal:2012,Bobrov&Trigger:2013}, just to mention a few.
Experimental observation of the Bogoliubov excitations in exciton--polariton Bose-condensates was reported by Utsunomiya \textit{et al.} \cite{Utsunomiya_etal:2008}.
Such systems \cite{Rubo_etal:2006,Deng_etal:2010} constitute another group for probing Bose-condensation, alongside dilute alkali gases \cite{Pethick&Smith:2001}.
Approaches based on the Hamiltonian of an interacting Bose-system by Bogoliubov and Zubarev \cite{Bogoliubov&Zubarev:1955} were used by the present author to calculate the effective mass of the $^4$He atom and the excitation spectrum of liquid helium-4 \cite{Rovenchak:2003FNT,Rovenchak:2005JLTP}. In this work, yet another method is proposed allowing for a proper treatment of the long-wavelength domain
of the excitation spectrum. This approach is tested both for helium-4 with a self-consistently derived interatomic potential \cite{Vakarchuk_etal:2000} and for a Bose-liquid with the Yukawa potential.
The paper is organized as follows. Section~\ref{sec:Green} contains all the required definitions of the Hamiltonian and two-temperature Green's functions, which are further applied in the calculations. Decouplings for Green's functions are suggested in Section~\ref{sec:Decoupling} and general expressions for the excitation spectrum are obtained there. Results of calculations for two systems (liquid helium-4 and the Yukawa Bose-liquid) are given in Section~\ref{sec:Results}. A short discussion in Section~\ref{sec:Disc} concludes the paper.
\section{Green's functions}
\label{sec:Green}
Let us consider the Hamiltonian of a Bose-system of $N$ particles in the volume $V$ interacting via pairwise $\Phi({\bf r}_1-{\bf r}_2)$ and three-particle $\Phi_3({\bf r}_1,{\bf r}_2,{\bf r}_3)$ potentials
\begin{eqnarray}\label{eq:H}
H = -\frac{\hbar^2}{2m} \sum_{j=1}^N \Delta_j +
\sum_{1\leq j<l\leq N} \Phi({\bf r}_j-{\bf r}_l)+
\sum_{1\leq j<l<s\leq N} \Phi_3({\bf r}_j,{\bf r}_l,{\bf r}_s),
\end{eqnarray}
where $m$ is the mass of a particle and $\Delta_j$ is the Laplace operator with respect to the $j$th coordinate ${\bf r}_j$. In the case of a Bose-system it is possible to pass to the so-called collective variables
\begin{eqnarray}\label{eq:rho_drho}
\rho_{\bf k} = \frac{1}{\sqrt{N}}\sum_{j=1}^N e^{-i{\bf k}{\bf r}_j},
\qquad
\partial_{-{\bf k}} = \frac{\partial}{\partial\rho_{\bf k}},\qquad
\textrm{where}\quad{\bf k}\neq 0.
\end{eqnarray}
Hamiltonian (\ref{eq:H}) becomes as follows \cite{Bogoliubov&Zubarev:1955,Rovenchak:2005CEJP}:
\begin{eqnarray}\label{eq:Hrho}
H &=& \sum_{{\bf k}\neq0}\Big[
\varepsilon_k\left(\rho_{\bf k}\partial_{-{\bf k}}-\partial_{\bf k}\partial_{-{\bf k}}\right) +
\frac{n}{2}\tilde\nu_k\rho_{\bf k}\rho_{-{\bf k}}\Big] \\
&&{}+
\frac{1}{\sqrt{N}}\mathop{\sum_{{\bf k}\neq0}\sum_{{\bf q}\neq0}}\limits_{{\bf k}+{\bf q}\neq0}
\Big[
\frac{\hbar^2}{2m}{\bf k}{\bf q}\,\rho_{{\bf k}+{\bf q}}\partial_{-{\bf k}}\partial_{-{\bf q}}+
\frac{n^2}{6}\nu_3({\bf k},{\bf q})\rho_{{\bf k}+{\bf q}}\rho_{-{\bf k}}\rho_{-{\bf q}}
\Big],
\end{eqnarray}
where $\varepsilon_k = \hbar^2k^2/2m$ is the free-particle energy, $n=N/V$ is the particle density, $\nu_k$ and $\nu_3({\bf k},{\bf q})$ are the Fourier transforms of the pairwise and three-particle potentials, respectively, and
\begin{eqnarray}
\tilde\nu_k = \nu_k +n\nu_3({\bf k},-{\bf k}) -\frac{1}{V}\sum_{{\bf q}\neq0}\nu_3({\bf k},{\bf q}).
\end{eqnarray}
With operators $A$ and $B$ written in the Heisenberg representation, the two-time temperature Green's functions are defined as \cite{Zubarev:1974}
\begin{eqnarray}
\GreenF{A(t)}{B(t')} = i\theta(t-t')\left\langle[A(t),B(t')]\right\rangle,
\end{eqnarray}
where $\theta(t)$ is the Heaviside step function and $[\cdot,\cdot]$ denotes the commutator.
In the case of $t=t'$, the equations of motion in the frequency representation are given by
\begin{eqnarray}
\hbar\omega\GreenF{A}{B} = \frac{1}{2\pi}\langle[A,B]\rangle+
\GreenF{[A,H]}{B},
\end{eqnarray}
which upon simple transformations with Eq.~(\ref{eq:Hrho}) leads to the following set:
\begin{eqnarray}
&&\hbar\omega\GreenF{\rho_{\bf k}}{\rho_{-{\bf k}}} =
-\varepsilon_k\GreenF{\rho_{\bf k}}{\rho_{-{\bf k}}}
+2\varepsilon_k\GreenF{\partial_{\bf k}}{\rho_{-{\bf k}}} \nonumber\\
&&\qquad\qquad\qquad\qquad{}-\frac{2}{\sqrt{N}} \mathop{\sum_{{\bf q}\neq0}}\limits_{{\bf k}+{\bf q}\neq0}
\frac{\hbar^2}{2m}{\bf k}{\bf q}\GreenF{\rho_{{\bf k}+{\bf q}}\partial_{-{\bf q}}}{\rho_{-{\bf k}}}, \nonumber\\ [-6pt]
\label{eq:hwGreen}\\[-6pt]
&&\hbar\omega\GreenF{\partial_{\bf k}}{\rho_{-{\bf k}}} = \frac{1}{2\pi}
+\varepsilon_k\GreenF{\partial_{\bf k}}{\rho_{-{\bf k}}}
+n\nu_k\GreenF{\rho_{\bf k}}{\rho_{-{\bf k}}} \nonumber\\
&&\quad{}+\frac{1}{\sqrt{N}} \mathop{\sum_{{\bf q}\neq0}}\limits_{{\bf k}+{\bf q}\neq0}
\Bigg[
\frac{\hbar^2}{2m}{\bf k}{\bf q}\GreenF{\partial_{{\bf k}+{\bf q}}\partial_{-{\bf q}}}{\rho_{-{\bf k}}}
-\frac{n^2}{2}\nu_3({\bf k},{\bf q})\GreenF{\rho_{{\bf k}+{\bf q}}\rho_{-{\bf q}}}{\rho_{-{\bf k}}}
\Bigg]
. \nonumber
\end{eqnarray}
Dropping off three-operator Green's functions, i.\,e. in the random phase approximation (RPA), one obtains the following expression for one of the solutions of set (\ref{eq:hwGreen}):
\begin{eqnarray}
\GreenF{\rho_{\bf k}}{\rho_{-{\bf k}}} = \frac{\varepsilon_k}{\pi}
\frac{1}{(\hbar\omega)^2-\varepsilon_k^2\alpha_k^2}
\end{eqnarray}
immediately yielding the Bogoliubov spectrum from the poles of Green's functions with respect to $\hbar\omega$:
\begin{eqnarray}\label{eq:EkRPA}
E_k = \varepsilon_k\alpha_k,
\qquad\textrm{where}\quad
\alpha_k=\left(1+2n\nu_k/\varepsilon_k\right)^{1/2}.
\end{eqnarray}
While it is also possible to derive the equations of motion for three-operator functions as well, no closed-form expression for the excitation spectrum can be obtained with them. So, another option leading to easier calculations of the spectrum is considered in the next Section.
\section{Green's function decoupling and excitation spectrum}
\label{sec:Decoupling}
Since Green's functions of the $\GreenF{AB}{C}$ type are used to calculate averages of triple products $\langle CAB \rangle$, the following decoupling can be used
\begin{eqnarray}
\GreenF{A_1B_2}{C_3} =
\lambda(1,2)\GreenF{A_{1+2}}{C_3}+\mu(1,2)\GreenF{B_{1+2}}{C_3}
\end{eqnarray}
with
\begin{eqnarray}\label{eq:lambda_mu}
\lambda(1,2) =
(1-\eta) \frac{\langle C_3A_1B_2\rangle}{\langle C_3A_{1+2}\rangle},
\qquad
\mu(1,2) = \eta \frac{\langle C_3A_1B_2\rangle}{\langle C_3B_{1+2}\rangle},
\end{eqnarray}
where the value of the parameter $\eta=0\div1$ will be fixed on a later stage.
The following two approximation for averages were tested
\begin{eqnarray}
\langle C_3A_1B_2\rangle_s = \frac16\bigg(
\langle CA\rangle_1\langle CB\rangle_2\langle AB\rangle_3+
\textrm{symmetrization over indices}
\bigg)
\end{eqnarray}
with six items in the parentheses, hence the ``$s$'' index, and
\begin{eqnarray}
\langle C_3A_1B_2\rangle_f &=& \frac14\bigg[
\Big(\langle AB\rangle_1\langle CA\rangle_2+
\langle AB\rangle_2\langle CA\rangle_1\Big)\langle CB\rangle_3\\
&&\ {}+\Big(\langle AB\rangle_1\langle CB\rangle_2+
\langle AB\rangle_2\langle CB\rangle_1\Big)\langle CA\rangle_3
\bigg]\nonumber
\end{eqnarray}
with four items in the parentheses.
To decouple the functions entering Eqs.~(\ref{eq:hwGreen}) the following averages are required:
\begin{eqnarray}
\langle\rho_{-{\bf k}}\rho_{{\bf k}+{\bf q}}\rho_{-{\bf q}}\rangle_{s,f} = \frac{1}{\sqrt{N}}
\langle\rho_{-{\bf q}}\rho_{{\bf q}}\rangle
\langle\rho_{-{\bf k}-{\bf q}}\rho_{{\bf k}+{\bf q}}\rangle
\langle\rho_{-{\bf k}}\rho_{{\bf k}}\rangle,
\end{eqnarray}
which corresponds to the so-called convolution approximation and is identical for both the suggested decoupling types,
\begin{eqnarray}
&&\langle\rho_{-{\bf k}}\rho_{{\bf k}+{\bf q}}\partial_{-{\bf q}}\rangle_{s} = \frac{1}{3\sqrt{N}}
\Big(
\langle\rho_{-{\bf q}}\rho_{{\bf q}}\rangle
\langle\rho_{-{\bf k}-{\bf q}}\partial_{{\bf k}+{\bf q}}\rangle
\langle\rho_{-{\bf k}}\partial_{{\bf k}}\rangle
\\
&&\quad{}+
\langle\rho_{-{\bf q}}\partial_{{\bf q}}\rangle
\langle\rho_{-{\bf k}-{\bf q}}\rho_{{\bf k}+{\bf q}}\rangle
\langle\rho_{-{\bf k}}\partial_{{\bf k}}\rangle
+
\langle\rho_{-{\bf q}}\partial_{{\bf q}}\rangle
\langle\rho_{-{\bf k}-{\bf q}}\partial_{{\bf k}+{\bf q}}\rangle
\langle\rho_{-{\bf k}}\rho_{{\bf k}}\rangle
\Big),\nonumber\\[12pt]
&&\langle\rho_{-{\bf k}}\partial_{{\bf k}+{\bf q}}\partial_{-{\bf q}}\rangle_{s} = \frac{1}{3\sqrt{N}}
\Big(
\langle\partial_{-{\bf q}}\partial_{{\bf q}}\rangle
\langle\rho_{-{\bf k}-{\bf q}}\partial_{{\bf k}+{\bf q}}\rangle
\langle\rho_{-{\bf k}}\partial_{{\bf k}}\rangle
\\
&&\quad{}+
\langle\rho_{-{\bf q}}\partial_{{\bf q}}\rangle
\langle\partial_{-{\bf k}-{\bf q}}\partial_{{\bf k}+{\bf q}}\rangle
\langle\rho_{-{\bf k}}\partial_{{\bf k}}\rangle
+
\langle\rho_{-{\bf q}}\partial_{{\bf q}}\rangle
\langle\rho_{-{\bf k}-{\bf q}}\partial_{{\bf k}+{\bf q}}\rangle
\langle\partial_{-{\bf k}}\partial_{{\bf k}}\rangle
\Big)\nonumber
\end{eqnarray}
and
\begin{eqnarray}
&&\langle\rho_{-{\bf k}}\rho_{{\bf k}+{\bf q}}\partial_{-{\bf q}}\rangle_{f} \\
&&\quad=
\frac{1}{2\sqrt{N}}
\Big(
\langle\rho_{-{\bf q}}\rho_{{\bf q}}\rangle
\langle\rho_{-{\bf k}-{\bf q}}\partial_{{\bf k}+{\bf q}}\rangle
\langle\rho_{-{\bf k}}\partial_{{\bf k}}\rangle
+
\langle\rho_{-{\bf q}}\partial_{{\bf q}}\rangle
\langle\rho_{-{\bf k}-{\bf q}}\rho_{{\bf k}+{\bf q}}\rangle
\langle\rho_{-{\bf k}}\partial_{{\bf k}}\rangle
\Big),\nonumber\\[12pt]
&&\langle\rho_{-{\bf k}}\partial_{{\bf k}+{\bf q}}\partial_{-{\bf q}}\rangle_{f} \\
&&\quad=
\frac{1}{2\sqrt{N}}
\Big(
\langle\partial_{-{\bf q}}\partial_{{\bf q}}\rangle
\langle\rho_{-{\bf k}-{\bf q}}\partial_{{\bf k}+{\bf q}}\rangle
\langle\rho_{-{\bf k}}\partial_{{\bf k}}\rangle
+
\langle\rho_{-{\bf q}}\partial_{{\bf q}}\rangle
\langle\partial_{-{\bf k}-{\bf q}}\partial_{{\bf k}+{\bf q}}\rangle
\langle\rho_{-{\bf k}}\partial_{{\bf k}}\rangle
\Big).
\nonumber
\end{eqnarray}
Expressions for pairwise averages are easily obtained from the solutions of Eqs.~(\ref{eq:hwGreen}) in RPA as follows \cite{Rovenchak:2003FNT}:
\begin{eqnarray}
&&\langle\rho_{-{\bf k}}\rho_{{\bf k}}\rangle = \frac{1}{\alpha_k}, \qquad
\langle\rho_{-{\bf k}}\partial_{{\bf k}}\rangle = \frac12\left(\frac{1}{\alpha_k}-1\right),\nonumber\\
&&\langle\partial_{-{\bf k}}\partial_{{\bf k}}\rangle =
\frac14\left(\frac{1}{\alpha_k}-\alpha_k\right).
\end{eqnarray}
With the decouplings applied, equations of motion (\ref{eq:hwGreen}) become:
\begin{eqnarray}
\hbar\omega\GreenF{\rho_{\bf k}}{\rho_{-{\bf k}}} &=&
-\varepsilon_k^{(1)}\GreenF{\rho_{\bf k}}{\rho_{-{\bf k}}}
+2\varepsilon_k^{(2)}\GreenF{\partial_{\bf k}}{\rho_{-{\bf k}}} \nonumber\\
\label{eq:hwGreen_decoupled}\\[-6pt]
\hbar\omega\GreenF{\partial_{\bf k}}{\rho_{-{\bf k}}} &=& \frac{1}{2\pi}+
\varepsilon_k^{(3)}\GreenF{\partial_{\bf k}}{\rho_{-{\bf k}}}
+n\nu_k^{(*)}\GreenF{\rho_{\bf k}}{\rho_{-{\bf k}}}, \nonumber
\end{eqnarray}
where
\begin{eqnarray}\label{eq:eps1}
&&\varepsilon_k^{(1)} = \varepsilon_k+\frac{2}{\sqrt{N}} \mathop{\sum_{{\bf q}\neq0}}\limits_{{\bf k}+{\bf q}\neq0}
\frac{\hbar^2}{2m}{\bf k}{\bf q}\,X({\bf k},{\bf q}),\\
\label{eq:eps2}
&&\varepsilon_k^{(2)} = \varepsilon_k+\frac{1}{\sqrt{N}} \mathop{\sum_{{\bf q}\neq0}}\limits_{{\bf k}+{\bf q}\neq0}
\frac{\hbar^2}{2m}{\bf k}{\bf q}\,Y({\bf k},{\bf q}),
\\
\label{eq:eps3}
&&\varepsilon_k^{(3)} = \varepsilon_k + \frac{2}{\sqrt{N}} \mathop{\sum_{{\bf q}\neq0}}\limits_{{\bf k}+{\bf q}\neq0}
\frac{\hbar^2}{2m}{\bf k}{\bf q}\,Z({\bf k},{\bf q}),
\\
&&\nu_k^{(*)} = \tilde\nu_k+\frac{1}{\sqrt{N}}
\mathop{\sum_{{\bf q}\neq0}}\limits_{{\bf k}+{\bf q}\neq0}\nu_3({\bf k},{\bf q}) T({\bf k},{\bf q}).
\end{eqnarray}
The notations used in the above definitions are as follows:
\begin{eqnarray}
&&X({\bf k},{\bf q}) = (1-\eta)\frac{\langle\rho_{-{\bf k}}\rho_{{\bf k}+{\bf q}}\partial_{-{\bf q}}\rangle}{\langle\rho_{-{\bf k}}\rho_{{\bf k}}\rangle}, \nonumber
\qquad
Y({\bf k},{\bf q}) = \eta \frac{\langle\rho_{-{\bf k}}\rho_{{\bf k}+{\bf q}}\partial_{-{\bf q}}\rangle}{\langle\rho_{-{\bf k}}\partial_{{\bf k}}\rangle}, \nonumber\\[-6pt]
\\
&&Z({\bf k},{\bf q}) = \frac{\langle\rho_{-{\bf k}}\partial_{{\bf k}+{\bf q}}\partial_{-{\bf q}}\rangle}{\langle\rho_{-{\bf k}}\partial_{{\bf k}}\rangle},
\qquad
T({\bf k},{\bf q}) = \frac{\langle\rho_{-{\bf k}}\rho_{{\bf k}+{\bf q}}\rho_{-{\bf q}}\rangle}{\langle\rho_{-{\bf k}}\rho_{{\bf k}}\rangle}\nonumber
\end{eqnarray}
The excitation spectrum is thus given by:
\begin{eqnarray}\label{eq:EkG3}
E_k = \frac12\left\{\varepsilon_k^{(3)}-\varepsilon_k^{(1)} +
\sqrt{\left[\varepsilon_k^{(1)}\right]^2+\left[\varepsilon_k^{(3)}\right]^2
+2\varepsilon_k^{(1)}\varepsilon_k^{(3)}+8\varepsilon_k^{(2)}n\nu_k^{(*)}}\right\}
\end{eqnarray}
With the items containing the summations dropped, Bogoliubov's result (\ref{eq:EkRPA}) immediately follows from this expression.
\section{Results}\label{sec:Results}
The calculations of the excitation spectrum were made for two bosonic systems. The first one is the liquid helium-4 and the second one is the Yukawa Bose-liquid with parameters corresponding to the nuclear matter. As data about the details of three-particle interactions in these systems are rather scarce, the contributions of $\nu_3({\bf k},{\bf q})$ are neglected. It was estimated in \cite{Rovenchak:2005JLTP} that in case of helium-4 such an approach does not influence the long-wavelength limit of the excitation spectrum significantly.
To facilitate the numerical analysis, summation in (\ref{eq:eps1})--(\ref{eq:eps3}) is substituted with integration in the wave-vector space according to such a rule:
\begin{eqnarray}
\frac{1}{N}\sum_{{\bf q}}\ldots = \frac{1}{n}\int d{\bf q}\,\ldots.
\end{eqnarray}
The following set of parameters is used for the calculations of the helium-4 excitation spectrum:
\begin{eqnarray}
m = 4.0026\ \textrm{a.\,m.\,u.},\qquad
n = 0.02185\ \textrm{\AA}^{-3}.
\end{eqnarray}
The data for the interatomic potential $\nu_k$ are taken from \cite{Vakarchuk_etal:2000}.
Results for the excitation spectrum of helium-4 according to Eq.~(\ref{eq:EkG3}) are shown in Fig.~\ref{fig:EkHe} compared to the RPA approximation (Bogoliubov's spectrum) and experimental data. The $\eta$ parameter is set $\eta=1$ as for smaller values the correction to the RPA result appear insufficient to produce the correct slope of the phonon (linear at $k\to 0$) branch. Both of the suggested decoupling types are found suitable to describe the long-wavelength behavior of the helium-4 spectrum without introduction of the effective mass, which is required in the RPA, cf. \cite{Rovenchak:2003FNT}. On the other hand, the proposed method still fails at higher values of the wave vector and further modifications should be sought for to reproduce the maxon and roton domains successfully in this approach.
\begin{figure}[h]
\includegraphics[width=0.48\textwidth]{EkG3.pdf}
\quad
\includegraphics[width=0.48\textwidth]{EkG3-LW.pdf}
\caption{(Color online.) Excitation spectrum of the liquid helium-4. The right graph is the enlarged view of the marked rectangular area of the left graph.
Red solid line -- RPA result (Bogoliubov's spectrum);
green dashed line -- spectrum (\ref{eq:EkG3}) with the $s$-type decoupling, $\eta = 1$;
magenta dotted line -- spectrum (\ref{eq:EkG3}) with the $f$-type decoupling, $\eta = 1$.
Circles (errorbars) are the experimental data from \cite{Cowley&Woods:1971}.}
\label{fig:EkHe}
\end{figure}
Another interesting problem for analysis is the nuclear matter, where different types of Bose-condensation are predicted \cite{Cleymans&vonOertzen:1990,Rezaeian&Pirner:2006}. It is possible to model the nuclear matter as a Bose-liquid interacting via the Yukawa potential $\Phi(r) = \epsilon\,e^{-r/\sigma}/r$ \cite{Ceperley&Chester:1976}. The Fourier transform of this potential reads \cite{Strepparola_etal:1998}
\begin{eqnarray}
\nu_k = \frac{4\pi\epsilon\sigma^3}{1+\sigma^2k^2}.
\end{eqnarray}
with the following values of the parameters \cite{Ceperley&Chester:1976}:
\begin{eqnarray}
\epsilon = 5725\ \textrm{MeV},\qquad
\sigma = 0.244\ \textrm{fm}^{-1}.
\end{eqnarray}
Other quantities used to model the nuclear matter are as follows:
\begin{eqnarray}
\Lambda^* = \frac{2\pi\hbar}{\sigma\sqrt{\epsilon m}} = 1.08,\qquad
n = 0.16\ \textrm{fm}^{-3}.
\end{eqnarray}
Results for the excitation spectrum are given in Fig.~\ref{fig:EkYukawa}. The obtained correction to the RPA changes the shape of the $E_k$ curve leading to a good qualitative agreement with other data \cite{Strepparola_etal:1998,Halinen_etal:2000}. It should be mentioned that the values of the $\eta$ parameter are close to zero in this case as for $\eta\gtrsim0.2$ unphysical divergences in the domain of the minimum ($k=4\div6$~fm$^{-1}$) appear.
\begin{figure}[h]
\includegraphics[width=0.48\textwidth]{EkG3Y.pdf}
\caption{(Color online.) Excitation spectrum of the Yukawa Bose liquid with parameters corresponding to the nuclear matter.
Red solid line -- RPA result (Bogoliubov's spectrum);
green dashed line -- spectrum (\ref{eq:EkG3}) with the $s$-type decoupling, $\eta = 0$;
magenta dotted line -- spectrum (\ref{eq:EkG3}) with the $f$-type decoupling, $\eta = 0$;
blue dashed-dotted line -- spectrum (\ref{eq:EkG3}) with the $s$-type decoupling, $\eta = 0.1$
}
\label{fig:EkYukawa}
\end{figure}
\section{Discussion}\label{sec:Disc}
In summary, an approach was proposed to treat an interacting bosonic system using two-time temperature Green's functions on the collective variables leading to a good description of the elementary excitation spectrum in the long-wavelength limit. For two models considered in the work different values of the Green's function decoupling parameter $\eta$ should be taken: $\eta=1$ for the liquid helium-4 and $\eta=0$ for the Yukawa Bose-liquid.
General expression obtained in the work can be further applied to study other bosonic systems, especially where the Fourier transforms of the interatomic potential are known. This includes dilute Bose-gases, where the $\delta$-function potential is applicable, cf. similar analysis \cite{Vorobets:2013} for a binary bosonic mixture, and charged Bose-systems \cite{Panda&Panda:2010}.
|
\section{}
\label{}
\bibliographystyle{model2-names}
\section{Introduction}
Mangroves are known to provide several socio-ecological and ecosystem services such as timber and fisheries production, nutrient regulation, shoreline protection, etc. [see for example a review by~\cite{Lee}]. Unfortunately, mangroves are being lost worldwide at an alarming rate of 1\% per year due to various natural and anthropogenic causes~\citep{FAO2007}. In the Philippines, the total mangrove forest cover decreased by 51.80\% between 1918 and 2010. Particularly, an annual loss rate of 0.52\% from 1990 to 2010 was mainly attributed to aquaculture development~\citep{Long2013}. The depletion of mangroves shall result in the reduction of ecosystem functionality, and may expose coastal areas to higher vulnerability against natural disasters such as typhoons and storm surges~\citep{Duke2007}.
In November 2013, Super Typhoon Haiyan ravaged the Eastern Visayas region in Central Philippines. It is the strongest in historical records that ever made landfall~\citep{Zhang2013}. In the quest for solutions to mitigate similar future coastal disasters, mangroves along coastal fringes are being considered for their potential to protect against storm surges~\citep{Temmerman2013,Schmitt2013}. Mangroves are known to attenuate waves by as much as $75$\% through its vast underground root networks and high vegetation structural complexity, but only if mangroves have a wide extent of at least one km~\citep{McIvor2012}.
Other than coastal protection, mangrove restoration along coastal fringes also contributes toward climate change adaptation and mitigation strategies~\citep{Duarte2013}; for instance, through the sequestration of atmospheric $\mbox{CO}_2$~\citep{Donato,Masera}. Based on field measurements by~\cite{Salmo}, mangrove forests have been found to increase carbon stock over time through the accumulation of biomass as the forest grows and develops. Carbon stock accumulation demonstrates the capacity of mangroves to absorb atmospheric $\mbox{CO}_2$, and thereby contributes in regulating the impacts of global warming.
But in order to promote carbon stock accumulation, the mangrove forest must increase its total area and tree density (i.e., number of trees per unit area). The effectiveness of a restoration strategy in achieving that goal depends on the growth and survival of planted mangroves. In the Philippines, most restored mangroves are monospecific, oftentimes using the genus \textit{Rhizophora}. The initial mangrove plots are situated in sub-optimal positions that are highly saline and too frequently inundated~\citep{SamsonRollon}.
Optimal restoration strategy usually involves the planting of appropriate species at the right locations [see review by~\cite{Primavera2008}]. The authors criticized the wide use of \textit{Rhizophora} seedlings in restoration programs for convenience even if evidence points that \textit{Rhizophora} sp. planted in coastal fringes could not survive in the long run nor attenuate strong wave action. Similarly,~\cite{Salmo2014} argued that since planted mangroves are less diverse, these plantations (even at $\sim\!\!20\mbox{ years}$) offer less resilience against strong typhoon as documented in the damages brought by Typhoon Chan-hom in Lingayen Gulf (northwestern Philippines) in May 2009.
Due to the long periods of time necessary for the growth and development of mangrove forests, assessment of restoration strategies is challenging. Mathematical models can nevertheless circumvent that difficulty~\citep{BergerReview}. Through model simulations, long-term trajectories of mangrove growth and carbon-stock accumulation can be generated. Simulations can be implemented with settings that mimic the conditions of typical restoration sites in the Philippines.
Here, we propose a mathematical model of mangrove forest growth in order to gain insights on generic features that an optimal restoration strategy must have. By offering a simplified description, the model is made sufficiently tractable upon which mathematical analysis could be performed. Furthermore, it is a substantial innovation from prevailing models of forest dynamics. The biophysical growth is described by a function that considers natural allometry of mangrove trees. Hence, it does not suffer the spurious singularities found in models based on the regression of data in a given sampled site. The present model also considers naturally-observable stochasticity in the scheduling of demographic events such as seedling establishment (birth), and mortality. Most importantly, the model contributes significant insight for determining which restoration strategy is most likely to succeed in the long run. Restored mangroves must flourish with minimal human assistance, such that carbon stock and tree density both rise or stabilize (in contrast to decline) in time.
\section{Materials and Methods}
\subsection{Overview}
\subsubsection{Purpose}
The model is intended to describe and test monospecific mangrove restoration strategies currently being undertaken in the Philippines. Parameterizations of the model are based on Philippine mangrove plantation data reported by~\cite{Salmo}. Focus is particularly given to the genus \emph{Rhizophora}, which is commonly planted in those restorations. Hypothetical strategies are going to be explored through the model. The long-term trajectory of aboveground biomass is particularly considered.
\subsubsection{State variables and scales}
Three hierarchical levels comprise the model: the individual mangroves, the mangrove forest population, and the seashore environment. Trees are characterized mainly by the state variable referred as \emph{diameter at breast height} ($\mbox{dbh}$), which is a proxy variable for age. Asynchronous development is considered; hence, $\mbox{dbh}$ is taken as an age indicator instead of time. The $\mbox{dbh}$ also determines the height $H$, and crown radius $r_{\mbox{\small crown}}$ through scaling arguments based on natural allometry~\citep{Ong2004}. Each mangrove is further characterized by its fixed position on the seashore, which in turn determines its functional response to stress gradients and to resource competition.
A forest population refers to the entire collection of mangrove stands over a certain area of interest. Population is characterized by abundance, and the convex hull (i.e., the widest horizontal convex area enclosing the entire population). The population may further be subdivided according to the subpopulations of three main developmental stages: seedling ($
\mbox{dbh} < 2.5\mbox{ cm}$), sapling ($2.5\leq \mbox{dbh} < 5\mbox{ cm}$), and tree ($\mbox{dbh} > 5\mbox{ cm}$). The ratio of population size against the area $A$ of the convex hull is interpreted as population density (scaled into number of seedlings/saplings/trees per hectare). The forest is further characterized by its aboveground biomass (AGB), which is related by allometry to $\mbox{dbh}$ according to the following equation defined by~\cite{Komiyama}:
\begin{equation}
\mbox{AGB} = 0.235~\mbox{dbh}^{2.42}\quad\mbox{t/ha}
\label{eq:agb}
\end{equation}
The carbon stock is about half of the AGB~\citep{Masera}. The seashore environment and its stress factors are taken to be the highest level in the hierarchy. Nutrient resources in the environment are assumed to be distributed uniformly in space and remain fixed in time. Two stressors relevant to mangrove survival are here considered: salinity and inundation. The salinity and inundation fields increase with seaward gradient. However, the stress fields are assumed not to vary with time. Individual mangroves subjected to stressors will have a general slowdown of growth. The response is nonlinear with respect to salinity, and linear with inundation. Lastly, competition between individuals depends on $\mbox{dbh}$ and the relative position of an individual in the patch. This method is adapted from the \emph{Field of Neighborhood} (FON) approach originally proposed by~\citet{BergerHildenbrandt}.
\subsubsection{Process overview and scheduling}
The model simulates daily time intervals (i.e., $dt = 1\mbox{ day}$). But the length of time step $\tau$ per iteration varies randomly, in accordance with a Poisson process (see Figure~\ref{fig:schedule}). Across each iteration, a randomly chosen tree gives rise to one recruit, or a randomly selected living mangrove dies with a probability related to its developmental stage. The maximum isotropic displacement of the recruit is proportional to $\tau$. Meanwhile, over an iteration, all living mangroves grow proportionately with $\tau$.
\begin{figure}
\centering
\includegraphics[scale=.17]{Figure1.eps}
\caption{Representative time course of a simulation. The $n$-th iteration occurs across a time step $\tau_n$. The sequence $\left\{\tau_n\right\}$ is a random sequence generated by the waiting time distribution of a Poisson process. At the end of each iteration, the population size $N$ changes by a unit due to any of the possible demographic events: recruitment or mortality. The choice of event is random. The time step is assumed short enough so that only one event takes place within it. Meanwhile, individual growth proceeds during an iteration according to the dbh increment $dD/dt$ multiplied by $\tau_n$. Propagule dispersal during an iteration is modelled as a random walk in an annular region, the outer radius of which is the product of a dispersal rate and the time interval $\tau_n$. At the end of the iteration, the propagule establishes and becomes a seedling with $\mbox{dbh}=0.5\mbox{ cm}$.}
\label{fig:schedule}
\end{figure}
\subsection{Design concepts}
\subsubsection{Emergence}
A spatial forest pattern emerges through the asynchronous seedling establishment and growth, the competitive interaction between individuals, and through the constraints imposed by the stressors. The mangrove life cycle is entirely represented by empirical rules describing mortality and dispersal as probabilities, whereas $\mbox{dbh}$ growth is deterministic. While fitness seeking is not modelled explicitly, the competition and stressor fields favor growth and survival in some locations of the forest over other locations.
\subsubsection{Sensing}
Individual mangroves are assumed to sense and respond to the growth constraints imposed by competition and stressors in its immediate surroundings. In particular, the individual mangrove ``perceives" the aggregate effect of the presence of other nearby mangroves.
Mangrove age is not clocked with time, but rather through its developmental stage in terms of its $\mbox{dbh}$. Thus, a mangrove could linger in a stage longer or shorter than the average depending on the extent of growth constraints to which it is subjected at any given moment. For example, when a mangrove's $\mbox{dbh}$ is above a certain empirically known threshold, the mangrove identifies its maturity as a tree and thus only then will initiate seedling production.
\subsubsection{Interaction}
Mangroves interact with one another at a distance from their trunk axis. A ``force-like field" has been proposed by~\cite{BergerHildenbrandt} to describe such interaction. The field decays exponentially with distance from the trunk axis. The existence of the field may be justified by the presence of an extended root system below ground, and crown system above the trunk. Competition ensues from the field interactions: root systems compete for belowground spaces for anchorage, and crown systems compete for sunlight. Competition slows down the growth of mangrove by a factor that is a function of the strength of the aggregate field it encounters from other mangroves. As a consequence of the slow growth, a mangrove could have a higher probability of dying if its size remains small for a longer time as compared to its neighbors with larger sizes.
\subsubsection{Stochasticity}
Randomness plays an essential role in the model. The changes in population size are categorized between recruitment and mortality. A time unit of $1$ day is assumed short enough so that a maximum of only one unit of change in the population size could possibly occur within it. As to which population size-changing event (among those indicated in Fig.~\ref{fig:schedule}) will occur within a time unit, the choice is randomly determined. Also due to stochasticity, an event is not always guaranteed to happen at every time unit. The focal individual is also determined at random. In a recruitment event, the seedling is randomly placed within an annular region surrounding the parent tree, as shown in Figure~\ref{fig:dispersal}.
\begin{figure}
\centering
\includegraphics[scale=.25]{Figure2.eps}
\caption{Radial dispersal of a propagule leading to a seedling recruit. A propagule from a parent tree disperses along a Brownian-like path from any point along the rim of a circle with the crown radius, $r_{\mbox{\small crown}}$. The dispersal radius $r$ of the propagule before it establishes at a fixed position is the radial horizontal length from the vertical axis of the parent tree. The maximum possible displacement over a time interval $\tau$ is $r_{\mbox{\small max}}=r_{\mbox{\small crown}}+\lambda\tau$ where $\lambda$ is an empirically known dispersal rate. In general, $r \leq r_{\mbox{max}}$. Dispersal is assumed to be isotropic so that the seedling could establish at any direction around the parent tree as long as there is available space.}.
\label{fig:dispersal}
\end{figure}
The time course is also stochastic such that the interval between unit changes in the population is random in length, as depicted in Figure~\ref{fig:schedule}. Hence, the population dynamics of the forest can be interpreted as a Poisson process, which is a common model for birth-death processes~\citep{Gardiner}.
\subsection{Details}
\subsubsection{Initialization}
A seashore spanning the middle and part of the upper intertidal zone is simulated as a patch having the simple geometry shown in Figure~\ref{fig:grid}. A point on the patch is described by its coordinates $(x,y)$. Boundaries are closed so that forest growth dynamics are only defined within the patch perimeter.
\begin{figure}
\centering
\includegraphics[scale=.25]{Figure3.eps}
\caption{Top view of the simulated seashore spanning the middle and part of the upper intertidal zone. The mean sea level (MSL) delineates the average coastline. The MSL line separates the shore below the MSL (dark shade) and above it (light shade). The dashed line represents the lowest edge of the middle intertidal just above the lower intertidal. The lower intertidal zone just beyond the dashed boundary may be exposed during low water spring tide. The origin $(0,0)$ of the coordinate system is located at the bottom left corner. Any point on the coordinate system is described by coordinates $(x,y)$, where $x$ and $y$ are the horizontal and vertical distance from origin, respectively. Environment fields are defined as functions of these coordinates. The direction of increasing elevation is toward $(0,0)$.}
\label{fig:grid}
\end{figure}
The seashore is initially occupied by seedlings with $\mbox{dbh} = 0.5\mbox{ cm}$. Positions of individual seedlings are randomly chosen within a designated initial plot on the seashore. Seedlings could not overlap. The seedling density distributed across the initial plot is fixed. Different stochastic realizations of the model are performed for a given initial plot configuration.
\subparagraph*{Salinity field.} A salinity field $S(x,y)$ is attributed to the patch shown in Fig.~\ref{fig:grid}, such that $S(0,0)=0$. The porewater salinity is assumed to be higher at locations nearer to the sea as opposed to the location closer to $(0,0)$. At sea, the salinity stabilizes to about $72\mbox{ ppt}$, which is rather extreme but here considered for the purpose of subjecting mangroves to poor growth (i.e., high salinity) conditions. The salinity field is thus defined as follows:
\begin{equation}
S(x,y) = \min\left[\frac{72}{L}\left(x+y\right), 72\right]
\label{eq:salinity_field}
\end{equation}
where $L=4096\mbox{ cm}$ as depicted in Fig.~\ref{fig:grid}. Consequently, the gradient of the salinity field is about $1.76\mbox{ ppt/m}$.
\subparagraph*{Inundation field.} The tidal inundation field $I(x,y)$ is related to the elevation profile of the patch. The elevation profile is such that $I(0,0)=0$; in other words, the point is high enough that no tide could reach it all year round. The field is defined as follows,
\begin{equation}
I(x,y) = \min\left[\frac{0.8}{L}\left(x+y\right),1\right]
\label{eq:inundation_field}
\end{equation}
where $0.8/L$ corresponds to the slope of the coast of approximately $2\mbox{ cm}$ decrease in elevation for every meter of advance toward the sea, which implies that the seashore advances some 25\% further at low tide (as depicted by the dashed line in Fig.~\ref{fig:grid}). This is consistent with the average tidal fluctuation of about $2\mbox{ m}$ for Philippine coastlines. Beyond the low tide line (dashed line in Fig.~\ref{fig:grid}), land is persistently submerged (except possibly during low water spring tide); hence, $I=1$ beyond the middle intertidal zone.
\subsubsection{Input}
The model requires the initial number of seedlings, and the position and configuration of the initial plot. An individual mangrove responds to stressor gradients with an effective reduction on their growth rates. The salinity and inundation response of a mangrove are denoted as $\sigma = \sigma[S(x,y)] = \sigma(x,y)$, and $\eta = \eta[I(x,y)] = \eta(x,y)$, respectively.
\subparagraph*{Salinity response. } We apply the salinity response proposed by~\cite{BergerHildenbrandt}. Hence, $\sigma(x,y)\in [0,1]$ is defined as a logistic decay function of salinity $S(x,y)$, as follows:
\begin{equation}
\sigma(x,y) = \left[1+\exp\left(\frac{S(x,y)-\tilde{S}}{\Delta S}\right)\right]^{-1}
\label{eq:salinity_response}
\end{equation}
where $\tilde{S}$ is the species specific critical salinity level above which a plant is not likely to grow and survive. $\Delta S$ is a species-specific tolerance to salinity, which indicates a range of salinity levels about $S=\tilde{S}$ for which the response undergoes abrupt transition. In the case of \emph{R. mucronata}, we set $\tilde{S} = 72\mbox{ ppt}$ and $\Delta S = 4\mbox{ ppt}$ in accordance with parameter settings for the genus \textit{Rhizophora} by~\cite{BergerHildenbrandt}.
\subparagraph*{Inundation response.} A mangrove is expected to grow faster if it is subjected to shorter inundation times (or being above water most of the time). Thus, the inundation response $\eta(x,y)\in [0,1]$ is defined by the following function,
\begin{equation}
\eta(x,y) = 1-I(x,y)
\label{eq:inundation_response}
\end{equation}
\subsubsection{Submodels}
\subparagraph*{Competition field.} Plant competition has been made spatially explicit in the FON approach~\citep{BergerHildenbrandt}. Mangroves compete with each other via field interaction. Each individual mangrove is set to give rise to a rotationally symmetric scalar field which decays with radial distance from its trunk axis. Suppose a mangrove is located such that its trunk axis stands on $(x_0,y_0)$ of the horizontal plane. Let $(x,y)$ be the field coordinate so that the radial distance with respect to $(x_0,y_0)$ is $r=\sqrt{\left(x-x_0\right)^2+\left(y-y_0\right)^2}$. Specifically, the field $f_0(r)$ (as a function of $r$) generated by the mangrove at position $(x_0,y_0)$ is defined as follows:
\begin{equation}
f_0(r) =
\begin{cases}
1, & 0\leq r < \dfrac{D}{2}\\
\exp\left[-c\left(r-\dfrac{D}{2}\right)\right], & \dfrac{D}{2}\leq r < r_{\mbox{\small crown}}\\
0, & r > r_{\mbox{\small crown}}
\end{cases}
\label{eq:scalar_field}
\end{equation}
where $D = \mbox{dbh}$ of the focal mangrove, and $r_{\mbox{\small crown}}$ is its crown radius related by allometry to $D$~\citep{BergerHildenbrandt}, as follows:
\begin{equation}
r_{\mbox{\small crown}}= \frac{1}{2}\left[22.2\,D^{0.654}\right]\;\mbox{cm}
\label{eq:crown_radius}
\end{equation}
Note that the parameter $c$ in Eq.~(\ref{eq:scalar_field}) acts as a decay constant which controls the strength of the field beyond the trunk axis. The smaller its value, the more extensive is the field; hence, the longer the interaction range of its generated field. On the contrary, a larger $c$ implies more localized fields; hence, weaker long-range interactions between individuals. A value of $c=0$ means that the crowns are strictly rigid. For the simulation results, the value used is $c=0.1$.
A zone of influence with radius $r_{\mbox{\small crown}}$ denotes a territory around a mangrove wherein it ``perceives" the mean field due to all other mangroves in the forest~\citep{BergerHildenbrandtGrimm}. In Fig.~\ref{fig:dispersal}, the said zone is shown as the inner circle with radius $r_{\mbox{\small crown}}$. The aggregate strength of competition encountered by a focal mangrove is determined by superposition of fields due to other mangroves, which intersect with the focal mangrove's zone of influence. Let this sum be denoted as $F(x,y)$ for a focal mangrove located at $(x,y)$. The individual response to competition, $K = K[F(x,y)] = K(x,y)\in [0,1]$, is defined by the following:
\begin{equation}
K(x,y) = 1-2F(x,y)
\label{eq:competition_response}
\end{equation}
wherein $K$ contributes further to the reduction of growth of each individual mangrove.
\subparagraph*{Individual growth rate.} It is assumed that the relationship of height $H$ and leaf area $La$ to $D=\mbox{dbh}$ can be described in functional form, $H(D)$ and $La(D)$, respectively. A reasonable boundary condition associated with these functions are: $H(0)=0$ and $La(0)=0$, which necessarily mean that in the absence of a trunk, there is no plant.
Forest gap models, which originated with~\cite{Botkin}, emphasize a relationship between $H$ and $D$. This relationship, denoted as $H(D)$, is determined phenomenologically from the nonlinear fitting of available data, which usually consist of a nonzero minimum for $D$. The fundamental flaw in such an imposed relationship is that it does not satisfy the boundary conditions, e.g., $H(0) \neq 0$.
We thus assume that the functional relationships between plant morphological features are expressed by allometry: height, $H(D) = aD^{\alpha}$; and leaf area, $La(D) = bD^{\beta}$. Not only do these functions satisfy the boundary conditions, but also their use has underlying foundations in physiology~\citep{Shingleton}. Allometric scaling is widely applied in forestry literature as reviewed by~\cite{Komiyama}. In expressing the rate of change $dD/dt$ based on these allometric relations, there is no need to assume that $dH/dD=0$ at $D=D_{\mbox{\small max}}$. The underlying problem with such assumption is that $D_{\mbox{\small max}}$ is usually defined empirically through site sampling~\citep{Botkin2011}, rather than as a true biophysical limit. In fact, the existence of maximum height $H_{\mbox{\small max}}\approx 130\mbox{ m}$ for trees on Earth, regardless of species, has sufficient basis on the physics of vertical water transport~\citep{Koch}. $D_{\mbox{\small max}}$ should have a corresponding limit as well, but it is difficult to impose a value for it on the basis of site sampling alone.
Further advantage of using allometry to relate $H$ and $D$ is that the morphologic parameters in the differential equation $dD/dt = f\left(D;D_{\mbox{\small max}}\right)$ are strictly species specific only. In gap models that impose variants of the $H(D)$ function [originally proposed by~\cite{Botkin}], the resulting $dD/dt$ are implicitly site specific as well. A $D_{\mbox{\small max}}$ value set beforehand from a previous sample is strictly not applicable for a new site even for the same species. If in the new site some trees, for instance, have $D>D_{\mbox{\small max}}$, then the entire model would break down. Such problem arises from the existence of zeroes in the quadratic denominator of $f\left(D;D_{\mbox{\small max}}\right)$ for $D>D_{\mbox{\small max}}$. Hence, substantial revisions to the model parameters are required each time a new site is considered, contrary to previous claims that those parameters are merely species specific~\citep{Botkin2011}. The preceding argument precludes the universality of any model based on non-allometric $H(D)$ relationships~\citep{Botkin2011, Lindner, Risch}.
Generally, forests are characterized by non-uniform spatial gradients. It is assumed that these gradients do not intrinsically affect the allometric exponents. Those exponents are assumed to be universal for any given species growing in a particular climate. Consequently, the value of the exponent could be estimated from aggregated data of plants growing in different zones of a forest. Thus, even if resources are not evenly distributed across the forest, the allometric exponent should remain generally valid because the stunted plants would not only be thinner but also shorter. Therefore, the impact of stressors appears merely as correction factors that effectively slow down the growth rate as expressed by $D' \equiv dD/dt$~\citep{BergerReview}.
By putting together the relevant aspects which hypothetically determine a mangrove's growth rate, the following logistic differential equation expresses the time increment of $\mbox{dbh} = D$:
\begin{equation}
\frac{d D(x,y,t)}{d t} = \left(\frac{\Omega}{2+\alpha}\right) D^{\beta-\alpha-1}\left[1-\frac{1}{\gamma}\left(\frac{D}{D_{\mbox{\small max}}}\right)^{1+\alpha}\right]\sigma(x,y)\eta(x,y) K(x,y)
\label{eq:dD_dt}
\end{equation}
where $D_{\mbox{\small max}}$ is calculated from the allometric equation: $H_{\mbox{\small max}}=H(D_{\mbox{\small max}}) = aD_{\mbox{\small max}}^{\alpha}$, where $H_{\mbox{\small max}} = 130\mbox{ m}$~\citep{Koch}, and given the value of $a$ and $\alpha$ for a particular species. Moreover, $0<\gamma\leq 1$ is interpreted as a site-dependent factor such that $\gamma^{1/(1+\alpha)}D_{\mbox{\small max}}$ is the maximum observable $\mbox{dbh}$ in any given site. While there is no attempt to explicitly relate $\gamma$ to salinity, inundation, storm frequency, resource availability and other environmental factors, its value can be determined empirically. Meanwhile, the parameters $\alpha$ and $\beta$ are strictly species specific, and can be evaluated from empirical data for a certain species even if samples originate from different sites. Based on known allometric data on height and leaf area index, the values $\alpha = 0.95$ and $\beta=2$ are used for \textit{R. mucronata}. Lastly, $\Omega$ is a dimensional scaling parameter that converts the units on the right-hand side of Eq.~(\ref{eq:dD_dt}) into $\mbox{cm}\,\mbox{day}^{-1}$ to be consistent with the units on the left-hand side. We used $\Omega=0.25$ without loss of generality.
\subparagraph*{Recruitment.} Although mangroves generally produce propagules throughout their lifetime, not all of these would successfully establish. In describing recruitment, a rate $k_0$ of successful seedling establishment per tree is defined. Within a random time interval $\tau$, one seedling of a randomly selected parent tree would successfully establish a distance away from that tree due to propagule dispersal, as illustrated in Fig.~\ref{fig:dispersal}.
\subparagraph*{Propagule dispersal.} Observations indicate that hydrochorous seedlings can establish a distance away from the parent tree~\citep{Sousa}. Mangrove seedlings are especially mobile because of their capacity to float and be carried by coastal current. Hydrochorous dispersal of propagules has also been found to be wind assisted~\citep{Stocken2013}. Based on empirical measurements of propagule dispersal rates~\citep{Sousa}, a diffusion rate $\lambda$ is defined. When a tree produces a propagule, that propagule may establish and grow anywhere within an annular region as illustrated in Fig.~\ref{fig:dispersal}. The region has inner radius equal to the parent tree's crown radius, $r_{\mbox{\small crown}}$, as defined in Eq.~(\ref{eq:crown_radius}).
On the other hand, the outer radius $r_{\mbox{\small max}}$ of the annulus is determined by the time interval $\tau$ before the propagule establishes, and the dispersal rate $\lambda$. A subroutine checks that the position of the established seedling does not overlap with the trunk area of any existing mangrove. For \textit{R. mucronata}, we use $\lambda = 26.67\mbox{ cm day}^{-1}$ deduced from measurements by~\cite{Sousa}.
\subparagraph*{Mortality.} Mangroves are expected to have Type III survivorship curves as any plant in general~\citep{Schaal}. The mortality rate is therefore highest for seedlings and lowest for trees. Sapling mortality rate is in between those rates. In particular, let $k_1$, $k_2$, and $k_3$ be the seedling, sapling, and tree mortality rate, respectively. Hence, $k_1 > k_2 > k_3$ in accordance with Type III survivorship curves.
Treating mortality as a Poisson process, the death of any mangrove in the forest is probabilistic. Within a randomly determined time interval $\tau$, a randomly chosen mangrove dies. Unlike the approach of~\cite{BergerHildenbrandt}, the mortality event is in no way correlated with the growth rate. Death by weather disturbances is, however, not yet considered in the present model.
\subparagraph*{Stochastic population dynamics.} Let $M$, $S_p$, and $S_d$ represent an individual tree, sapling, and seedling, respectively. The demographic events can be further expressed as reaction processes:
\begin{eqnarray}
\mbox{Recruitment:} & M &\overset{k_0}\longrightarrow M + S_d \label{eq:recruitment}\\
\mbox{Seedling death:} & S_d &\overset{k_1}\longrightarrow \varnothing \label{eq:seedling_death}\\
\mbox{Sapling death:} & S_p &\overset{k_2}\longrightarrow \varnothing \label{eq:sapling_death}\\
\mbox{Tree death:} & M &\overset{k_3}\longrightarrow\varnothing\label{eq:tree_death}
\end{eqnarray}
Equations~(\ref{eq:recruitment}) to~(\ref{eq:tree_death}) can be written in terms of a master equation of the probability density $P = P\left(N_M,N_{S_p},N_{S_d},t\right)$. With the assumption that the population size is $N = N_M + N_{S_p} + N_{S_d} \gg 1$, the master equation yields as a linear approximation the following deterministic system:
\begin{equation}
\begin{bmatrix}
m' \\ s_p' \\ s_d'
\end{bmatrix} \approx \begin{bmatrix}
-k_3 & \bar{D}_5' & 0\\
0 & -k_2 & \bar{D}_{2.5}' \\
k_0 & 0 & -k_1
\end{bmatrix} \begin{bmatrix}
m \\ s_p \\ s_d
\end{bmatrix}
\label{eq:ODEsystem}
\end{equation}
where $m = N_M/A$, $s_p = N_{S_p}/A$, and $s_d = N_{S_d}/A$ (given $A$ is the convex-hull area), and which has the trivial fixed point, $\Gamma^* \equiv \left(m,s_p,s_d\right) = (0,0,0)$. The primes denote time derivatives, and $\bar{D}'$ is a sample average of Eq.~(\ref{eq:dD_dt}). Specifically, $\bar{D}_5'=D'(D=5\mbox{ cm})$ and $\bar{D}_{2.5}'=D'(D=2.5\mbox{ cm})$ denote the mean transition rates from sapling to tree, and from seedling to sapling, respectively.
The value of $\bar{D}'$ could ultimately be computed with respect to the spatial variables, but such an attempt shall be reported in another study. Rather, it suffices to show from the eigenvalues of the square matrix in Eq.~(\ref{eq:ODEsystem}) whether the trivial fixed point, $\Gamma^*$, is either stable or unstable. In particular, we consider $k_0 > 0$ and all other demographic rates are proportional to $k_0$. Barring any human assistance or extreme weather disturbances, a stable $\Gamma^*$ makes forest collapse inevitable, whereas an unstable one implies that the tree population can only increase in time. The latter case is desirable as a dense greenbelt along the coast is preferred for promoting carbon stock accumulation.
\subsection{Simulation Experiments}
\subparagraph*{Gillespie's direct method.} The reaction processes associated with the demographic events denoted in Equations~(\ref{eq:recruitment}) to~(\ref{eq:tree_death}) can be written down as propensities, defined in the following:
\begin{eqnarray}
\mbox{Recruitment:} & T\left[N_M,N_{S_p},N_{S_d}\rightarrow N_{S_d}+1\right] = k_0 N_M \label{eq:propensity_recruitment}\\
\mbox{Seedling death:} & T\left[N_M,N_{S_p},N_{S_d}\rightarrow N_{S_d}-1\right] = k_1 N_{S_d} \label{eq:propensity_seedling_death}\\
\mbox{Sapling death:} & T\left[N_M,S_p\rightarrow N_{S_p}-1,N_{S_d}\right] = k_2N_{S_p} \label{eq:propensity_sapling_death}\\
\mbox{Tree death:} & T\left[N_M\rightarrow N_M-1,N_{S_p},N_{S_d}\right] = k_3 N_M \label{eq:propensity_tree_death}
\end{eqnarray}
The processes associated with the life-stage transitions may also be expressed in a similar manner:
\begin{eqnarray}
\mbox{$S_p \rightarrow M$:} & T\left[N_M\rightarrow N_M+1,N_{S_p}\rightarrow N_{S_p}-1,N_{S_d}\right] = \tilde{D}'_5 N_{S_p}\\
\mbox{$S_d \rightarrow S_p$:} & T\left[N_M,S_p\rightarrow N_{S_p}+1,N_{S_d}\rightarrow N_{S_d}-1\right] = \tilde{D}'_{2.5} N_{S_d}
\end{eqnarray}
However, these are only implicitly included in the Gillespie simulations through the change induced by growth, as described by Eq.~(\ref{eq:dD_dt}). Using the propensities defined in Eqs.~(\ref{eq:propensity_recruitment}) to~(\ref{eq:propensity_tree_death}), standard procedures of carrying out the Gillespie algorithm, detailed elsewhere~\citep{Gillespie}, are applied.
\section{Results}
\subsection{Initial configuration of mangrove forest}
\begin{table}
\centering
\caption{Demographic rates used in the simulations without sufficient loss of generality. The settings are consistent with Type III survivorship curves~\citep{Schaal}, attributed to plants like mangroves (i.e., $k_1 > k_2 > k_3$).}
\begin{tabular}{c|c|c}\hline
\textbf{Demographic rate} & \textbf{Symbol} & \textbf{Value}\\\hline
Seedling establishment rate & $k_0$ & $\dfrac{1}{3650}\mbox{ per day per tree}$\\\hline
Seedling mortality rate & $k_1$ & $2k_0\mbox{ per day per seedling}$\\\hline
Sapling mortality rate & $k_2$ & $k_0\mbox{ per day per sapling}$\\\hline
Tree mortality rate & $k_3$ & $\dfrac{5}{6}k_0\mbox{ per day per tree}$\\\hline
\end{tabular}
\label{table:demographic-rates}
\end{table}
In the simulation results, the demographic rates given in Table~\ref{table:demographic-rates} are used. Two configurations of the initial plot are here considered. For configuration type, the seedling density (i.e., number of seedlings per unit area) is fixed, while the positions of individual seedlings are randomized but do not overlap. The first type of plot configuration is an approximately trapezoidal strip the long axis of which is parallel to the MSL line, as shown in Figures~\ref{fig:initializations}(a) through (c). The initial seedling density $s_d(0)$ is about $42 \mbox{ seedlings per } 100\mbox{ m}^2$. The plot partially includes the zone below MSL in Fig.~\ref{fig:initializations}(a), and the midpoints of its long sides are situated at a distance of $0.8L$ and $1.1L$ from $(0,0)$, respectively. Plots are situated totally above MSL for Fig.~\ref{fig:initializations}(b), such that the long-side midpoints are at $0.6L$ and $0.8L$ from $(0,0)$; and for~\ref{fig:initializations}(c), long-side midpoints at $0.4L$ and $0.6L$ from $(0,0)$.
The second configuration type is an arc, which is situated totally above MSL. The seedling density $s_d(0)$ for initial plots of this type is about $13\mbox{ seedlings per } 100\mbox{ m}^2$. The arc is concave-landward in Fig.~\ref{fig:initializations}(d), and concave-seaward in Fig.~\ref{fig:initializations}(e).
\begin{figure}
\centering
\subfloat[Typical plot: $0.8L\rightarrow 1.1L$
{\includegraphics[scale=.37]{Figure4a.eps}}
\subfloat[Above-MSL plot: $0.6L\!\rightarrow \!0.8L$
{\includegraphics[scale=.37]{Figure4b.eps}}
\subfloat[Above MSL: $0.4L\!\rightarrow\!0.6L$
{\includegraphics[scale=.37]{Figure4c.eps}}\\
\subfloat[Arc concave-landward plot]
{\includegraphics[scale=.37]{Figure4d.eps}}
\subfloat[Arc concave-seaward plot]
{\includegraphics[scale=.37]{Figure4e.eps}}
\caption{The configuration of the initial plot. Strip type, in order of increasing distance from the MSL line toward $(0,0)$: (a), (b), (c); Arc type: (d) concave-landward, (e) concave-seaward. For each type, the seedling density is fixed although seedling positions are randomized. The enclosing curve is the convex hull, the area $A$ of which is used to estimate densities, and the above-ground biomass per unit area.}
\label{fig:initializations}
\end{figure}
\subsection{Long-term configuration, densities, and above-ground biomass}
The performance of the forest development is presented through the seedling, sapling, and tree densities for each iteration time (time unit, $dt = 1\mbox{ day}$). By allowing virtual time to proceed through about $250\mbox{ years}$, the simulation reveals remarkably different outcomes.
The assumptions for the simulation are as follows: human intervention (e.g., replanting, cutting down trees), and damaging environmental disturbances are absent. The above-ground biomass (AGB) is calculated using Eq.~(\ref{eq:agb}), and is also used to assess the development of the forest in time. Ideally, the tree density and AGB must be increasing in time, if not stable. A thicker forest is expected to contain more carbon stock.
The typical plot, which covers zones above and below the MSL [Fig.~\ref{fig:initializations}(a)], thins out progressively until no more plants exist after about 250 years, as illustrated by Fig.~\ref{fig:outcomes1}(a). The AGB also trends down consistently along with the decrease in the tree, sapling, and seedling densities. In other words, the forest becomes extinct, although some intermittent peaks are apparent along the course of the growth trajectory.
Above-MSL plots perform relatively better in the long run, as illustrated by Fig.~\ref{fig:outcomes1}(b) and~\ref{fig:outcomes1}(c). The observation may be explained in part by \textit{Rhizophora}'s less tolerance for salinity, the level of which is assumed to increase seaward. The mangroves are also subjected to more frequent wave action the nearer they are to the MSL line. Due to a combination of both major stressors, and a seaward gradient of both, the \textit{Rhizophora} forest is expected to thrive better at seashore locations farther away from the MSL. Mathematical explanations are provided in a later section.
\begin{figure}
\centering
\subfloat[Typical plot: $0.8L\rightarrow 1.1L$]
{\includegraphics[scale=.4]{Figure5a.eps}} \\
\subfloat[Above-MSL plot: $0.6L\rightarrow 0.8L$
{\includegraphics[scale=.4]{Figure5b.eps}} \\
\subfloat[Above-MSL plot: $0.4L\rightarrow 0.6L$
{\includegraphics[scale=.4]{Figure5c.eps}}
\caption{The long-term forest growth of plots in the order of increasing distance from the MSL line: (a), (b), and (c). For each outcome, the temporal profile of AGB, and densities for tree, sapling, and seedling are shown. Densities are values divided by the area $A$ of the convex hull. Other stochastic realizations of each configuration give rise to more or less the same outcome, even if the specific details are different.}
\label{fig:outcomes1}
\end{figure}
The tree densities display remarkably different trajectories among the various plots considered. Even if the area is bigger for a strip nearest the sea [Fig.~\ref{fig:initializations}(a)], the trajectory of its density and AGB are attracted towards a state of decline. In the long run, and in the absence of human intervention and weather disturbances, the mangrove forest goes extinct. Meanwhile, above-MSL plots have a better chance of survival owing to weaker stressors.
The arc-type configurations depicted in Figures~\ref{fig:initializations}(d) and~(e) also induce growth trajectories which are consistent with the stressor hypothesis. The concave-landward plot includes a substantial portion just above but near the MSL where stressors are stronger. On the other hand, the concave-seaward plot has most of the seedlings situated in inland zones with weaker stressors. Consequently, the trajectory of the concave-landward plot is directed towards extinction whereas that of the concave-seaward plot is not, as illustrated by Fig.~\ref{fig:outcomes2}.
\begin{figure}
\centering
\subfloat[Arc concave-landward plot]
{\includegraphics[scale=.4]{Figure6a.eps}}\\
\subfloat[Arc concave-seaward plot]
{\includegraphics[scale=.4]{Figure6b.eps}}
\caption{The long-term forest growth of arc-type plots: (a) concave-landward; (b) concave-seaward. The initial seedling densities are fixed for both settings. Other trials for the said settings yield similar results, even if the specific details can be different due to the stochasticity inherent in the model.}
\label{fig:outcomes2}
\end{figure}
\subsection{Mathematical analysis: the bifurcation mechanism}
The dependence of the long-term performance of the restored monospecific \emph{Rhizophora} forest to the configuration of the initial plot of seedlings is ultimately related to the dynamical nature of the system. A dynamical control parameter, $\xi$, changes the nature of the system's extinction state (a fixed point). Such qualitative change is referred commonly as a bifurcation. From the linear approximation, it is evident that the only fixed point $\Gamma^*$ of the system is the state of extinction defined by the ordered density-triple: $\left(m,s_p,s_d\right) = (0,0,0) \equiv \Gamma^*$.
The control parameter $\xi$ is implicitly a function of the spatial configuration of the system due to the stressor gradients. The initial configuration, as illustrated in Fig.~\ref{fig:initializations}, sets up the control parameter value. Here, we analyze mathematically the population dynamics of the system through its linear approximation expressed by Eq.~(\ref{eq:ODEsystem}).
First, the parameter $\xi$ is defined in terms of the pertinent rates, as follows:
\begin{equation}
\xi = \frac{k_0\tilde{D}'_5\tilde{D}'_{2.5}}{k_1k_2k_3}
\label{eq:control-parameter-xi}
\end{equation}
which acts as some sort of index the value of which is dictated by the demographic and life-stage transition rates. The spatial dependence of the parameter can be found in the product of the transition rates $\tilde{D}'_5\tilde{D}'_{2.5}$. The spatial dependence reflects the stressor gradients and the competition field.
The characteristic polynomial $p(\lambda)$ of the matrix in the linear approximation given by Eq.~(\ref{eq:ODEsystem}) is written in terms of $\lambda$ and parameterized by $\xi$ as follows:
\begin{equation}
p(\lambda) = \lambda^3 + \left(k_1+k_2+k_3\right)\lambda^2 + \left(k_1k_2+k_2k_3+k_3k_1\right)\lambda + \left(1-\xi\right)k_1k_2k_3
\label{eq:characteristic-polynomial}
\end{equation}
Given that all rates are non-negative, the last term could determine the nature of the roots of $p(\lambda)$, which correspond to the eigenvalues of Eq.~(\ref{eq:ODEsystem}). Particularly, it would depend on whether $\xi < 1$ or $\xi > 1$. In both cases $\Gamma^*$ is a hyperbolic fixed point.
For $\xi < 1$, all terms in $p(\lambda)$ are positive; but for $p(-\lambda)$ there are three sign changes from the first through the last term of Eq.~(\ref{eq:characteristic-polynomial}). Based on Descartes' sign rule, the implication of those sign changes is the existence of either three negative real eigenvalues, or one negative real eigenvalue and a pair of complex eigenvalues with negative real parts. Either case, if $\xi < 1$ then $\Gamma^*$ is a stable fixed point of the linear approximation.
On the other hand, for $\xi > 1$, the last term is the only negative term which means that there is a single sign change in $p(\lambda)$. In other words, a positive real eigenvalue is assured. But for $p(-\lambda)$, two sign changes occur through the terms in Eq.~(\ref{eq:characteristic-polynomial}) which imply the existence of either two negative real eigenvalues (in the case of a saddle point), or a pair of complex eigenvalues with negative real part (in the case of a saddle focus). Both cases nevertheless imply that $\Gamma^*$ is an unstable fixed point.
In order to further specify whether or not the system has a pair of complex eigenvalues, the discriminant $\Delta$, which is expressed in the following equation, is evaluated.
\begin{eqnarray}\nonumber
\Delta &=& \left(k_1+k_2+k_3\right)^2\left(k_1k_2+k_2k_3+k_3k_1\right)^2 - 4\left(k_1+k_2+k_3\right)\left(k_1k_2+k_2k_3+k_3k_1\right)^3\\\nonumber
&& - 4\left(k_1+k_2+k_3\right)^3k_1k_2k_3\left(1-\xi\right) - 27k_1^2k_2^2k_3^2\left(1-\xi\right)^2\\
&&+ 18\left(k_1+k_2+k_3\right)\left(k_1k_2+k_2k_3+k_3k_1\right)k_1k_2k_3\left(1-\xi\right)
\end{eqnarray}
It turns out that with the demographic rates provided in Table~\ref{table:demographic-rates}, $\Delta < 0$ as long as $\xi$ is greater than about $0.11152$. If that is the case, then a pair of complex eigenvalues indeed exist. Since $\Gamma^*$ is hyperbolic, then the complex eigenvalues simply mean that $\Gamma^*$ is either a stable-focus node for $\xi < 1$, whereas it is a saddle-focus for $\xi > 1$. In order to further examine the implication of the above mathematical analyses, on the basis of the results presented in Fig.~\ref{fig:outcomes1} and Fig.~\ref{fig:outcomes2}, we turn our attention to the tree density as a function of time, $m(t)$. Trees represent the life stage that would most likely survive in the long run, as a type--III survivorship curve suggests. Moreover, trees carry the bulk of the AGB, as one can deduce from Figures~\ref{fig:outcomes1} and~\ref{fig:outcomes2}.
Figure~\ref{fig:bifurcation1} shows the behavior of $\xi(t)$ superimposed with the time course of the tree density corresponding to the outcomes for the strip configurations depicted in Fig.~\ref{fig:outcomes1}. The extinction of the simulated forest as shown in Fig.~\ref{fig:outcomes1}(a) can now be justified by association with $\xi(t) < 1$ as presented in Fig.~\ref{fig:bifurcation1}(a). Looking back at the initial configuration in Fig.~\ref{fig:initializations}(a), the seedlings planted at or below MSL line are exposed to high levels of salinity and inundation stress. Thus, $\tilde{D}'_5\tilde{D}'_{2.5}$ is low by virtue of the slow growth rates afforded by high stress, as one can deduce from Eq.~(\ref{eq:dD_dt}). Consequently, $\xi < 1$. In other words, the transition rates toward the tree stage are too slow compared to the mortality rate so that the forest does not have sufficient time to flourish before all trees die out.
\begin{figure}
\centering
\subfloat[Typical plot: $0.8L\rightarrow 1.1L$]
{\includegraphics[scale=.4]{Figure7a.eps}}
\\
\subfloat[Above-MSL plot: $0.6L\rightarrow 0.8L$
{\includegraphics[scale=.4]{Figure7b.eps}}
\\
\subfloat[Above-MSL plot: $0.4L\rightarrow 0.6L$
{\includegraphics[scale=.4]{Figure7c.eps}}
\caption{The bifurcation parameter $\xi(t)$ for plots with strip-type configurations superimposed with the tree density. The dashed, narrow (blue) curve represents $\xi(t)$. The horizontal line corresponds to $\xi = 1$. The solid, thick (red) curve represents $m(t)$.}
\label{fig:bifurcation1}
\end{figure}
For the above-MSL plot shown in Fig.~\ref{fig:initializations}(b), the tree population apparently stabilizes. Fig.~\ref{fig:bifurcation1}(b) correspondingly shows the value of $\xi(t)$ stabilizing above $1$. The value of $\xi$ starts out at a high value, which could be explained by the favorable location of the initial plot. The seedlings are found above and further away from the MSL line, wherein the stressors have less magnitude. Even as the forest spreads in spatial extent as seen in Fig.~\ref{fig:outcomes1}(b), the parameter $\xi$ remains above $1$ even if it is seemingly decreasing in monotonic fashion towards $1$.
As the plots are placed above and further away from the MSL line, the long-term outcome and survival of the plantations considerably improve, as shown in Fig.~\ref{fig:outcomes1}(c). The forest is thicker and wider in extent, which is ideal for the maximization of carbon stock accumulation. At about $t = 150\mbox{ years}$ the tree density (and AGB) increases in breakout fashion. The result is supported by the much higher values of $\xi(t)$ relative to $1$, as depicted in Fig.~\ref{fig:bifurcation1}(c). The increasing value of $m(t)$ as time progresses is a clear manifestation of the instability of $\Gamma^*$ for $\xi > 1$. Although the sapling and seedling densities, $s_p(t)$ and $s_d(t)$, respectively, remain low, the growth rate is amply fast so that the transitions more than compensate for tree mortality. Consequently, the forest proliferates in time with a thickening tree density, which in turn guarantees a continuous supply of seedlings necessary for renewal and succession.
A comparison of the long-term forest growth for the arc-type plot configuration based on the parameter $\xi(t)$ is presented in Fig.~\ref{fig:bifurcation2}. The concave-landward plot is associated with $\xi < 1$, as shown in Fig.~\ref{fig:bifurcation2}(a), which explains the stability of the extinction fixed point $\Gamma^*$. The parts of the plot at or below MSL are exposed to high stress levels, which result in slow growth. Hence, the transition rates $\tilde{D}'_5$ and $\tilde{D}'_{2.5}$ are too low to support a large tree density over a long period of time. This configuration emphasizes the unsuitability of positions at or below MSL in the case of \textit{Rhizophora} species used to reforest a region with seaward gradients of salinity and inundation stresses, as considered in the present study. Although some portions of the plot are situated above MSL, their population is not sufficient to increase the sample average values for the transition rates. Thus, the forest ultimately proceeds toward extinction in the long run.
\begin{figure}
\centering
\subfloat[Arc concave-landward plot]
{\includegraphics[scale=.4]{Figure8a.eps}}\\
\subfloat[Arc concave-seaward plot]
{\includegraphics[scale=.4]{Figure8b.eps}}
\caption{The bifurcation parameter $\xi(t)$ for the arc-type configurations superimposed with the tree density. The dashed, narrow (blue) curve represents $\xi(t)$. The horizontal line corresponds to $\xi = 1$. The solid, thick (red) curve represents $m(t)$.}
\label{fig:bifurcation2}
\end{figure}
On the other hand, the concave-seaward plot has most of its seedling population above the MSL wherein stress levels are relatively weaker. Consequently, the forest is able to achieve and maintain a sufficiently high tree density over a long period of time, as illustrated in Fig.~\ref{fig:outcomes2}(b). The result is confirmed by $\xi > 1$ as time progresses, as shown in Fig.~\ref{fig:bifurcation2}(b). Although $\xi(t)$ started out below $1$, its value crossed above $1$ and stayed there since about time $t = 40\mbox{ years}$. The value of $\xi$ is somewhat only marginally above $1$, however. In some stochastic instantiations of this setting (not shown), the parameter $\xi$ fails to make the cross over to above $1$. The forest goes extinct in those cases.
The dynamic parameter $\xi$ can be interpreted as an index for gauging the success of a particular set of initial conditions of the restored forest. In some cases, certain features of $\xi(t)$ lead the trends in the tree density. For instance, in Fig.~\ref{fig:bifurcation2}(b) the time when $\xi$ crosses from below to above the value $\xi=1$ leads the surge in the tree density by as much as $100\mbox{ years}$. In other words, we can address the question of whether or not a forest will flourish in the long run by evaluating the index within a reasonable time since restoration has commenced. With mere knowledge from field measurements of the rates of seedling establishment, plant mortality, and transition between life stages; and of stress gradients in a site of planned restoration, an estimate of the value of the index $\xi$ could be obtained within $5$ to $25\mbox{ years}$. The index $\xi$ can be used as a practical tool for assessing the long-term growth, development, and carbon stock accumulation of restored mangroves. Other possible restoration strategies as represented by their initial plot configurations may be explored in the same manner as presented in this paper.
\section{Discussion}
We have proposed, simulated, and analyzed a stochastic model for mangrove forest growth in coastal settings with seaward gradient of salinity and inundation stress. The model offers a simple description of the essential features of a coastal setting relevant to mangrove growth such as stress levels, stress gradients, and boundaries. Although we have not incorporated human factors, nor weather disturbances, the model is sufficiently generic to incorporate these factors as additional rates at relevant timescales.
One novel aspect of our approach is the mathematical scheme of expressing the progress and transition of individual mangrove growth from seedling to sapling to tree. Instead of utilizing mathematical functions determined through regression, we instead made use of scaling relations that represent natural allometry. Not only does this method express the growth equation in terms of a single physical variable (i.e., the $\mbox{dbh}$), but also it provides a truly species-specific growth model that is site-independent. Prevailing approaches are site-dependent in terms of determining the growth equation parameters and rates.
We also incorporated a stochastic scheduling procedure for determining which demographic and life-stage transition event would happen to any individual mangrove at any given point in time. The choice is ultimately dictated by random chance owing to our lack of precise and complete information characterizing the state of the system at every instant of time. The importance of stochasticity could not be further emphasized. The advantage of such approach is that we need not explicitly set fixed times during which mangroves die, or when they develop from sapling to tree. Thus, instead of setting a global clock, the Poisson process is used as a natural alternative for scheduling, which also accounts for observable randomness. Our model also considers stochasticity in representing propagule dispersal, although a more accurate description could be made by explicitly considering the influence of currents, tides, and waves especially at the intertidal zone. Lastly, by coupling the demographic processes with the spatial distribution of stress variables (which includes resource competition between individual mangroves), we have added tractability to the model that allows for insightful mathematical analysis.
In the mathematical analysis of the linear approximation of the model, we have discovered a useful index, $\xi$, for pre-determining whether or not a certain restoration strategy would succeed in the long run in terms of promoting stable tree density, and ultimately, carbon stock accumulation.
The index is a simple mathematical ratio between relevant rates that are measurable in the field by means of statistical sampling. The value $\xi = 1$ represents a threshold between success and failure of a restoration strategy. The importance of an index is that we can tell within a few years whether or not a particular restoration strategy would allow a forest to actually flourish and be self-sustaining in the absence of human assistance (e.g., re-planting in forest gaps). Of course, extreme storm events could impose severe damage on restored sites, but under the guidance of a success index like $\xi$, at least one can deduce the ease by which the damage can be repaired through the self-healing ability that a stable forest possesses.
\cite{Donato} have found that the total carbon content of tropical multi-species mangrove forests is in the order of $1000\mbox{ tC/ha}$. We have found that a restored monospecific (\textit{R. mucronata}) forest associated with $\xi > 1$ could achieve half of that carbon stock in about $200$ to $250$~years post planting [see Fig.~\ref{fig:outcomes1}(c)]. On the other hand, forests restored using strategies associated with $\xi < 1$ are bound to exhibit a decreasing amount of carbon stock through the decimation of trees due to natural mortality induced by stressors and competition. Furthermore, as presented in Figures~\ref{fig:bifurcation1} and~\ref{fig:bifurcation2}, the crossing of the threshold $\xi = 1$ occurs within $5$ to $20$ years post planting. Thus, by measuring $\xi$ within that reasonable amount of time, an early indication of whether or not the restored forest shall flourish is obtained.
Current restoration strategies in the Philippines commonly use \textit{R. mucronata} and other related species due to the relative ease by which their seedlings are tended in nurseries. Following the species zonation concept, the species from the \textit{Sonneratia} and \textit{Avicennia} genera are more appropriate to plant at the low intertidal zone~\citep{Primavera2008}. Due to the lack of appreciation for the species zonation concept and of foresight put into restoration efforts (and understanding on the ecological requirements of the chosen species), most of the restored forests are designed in such a way that \textit{Rhizophora} seedlings are planted at the low intertidal areas. However, \textit{R. mucronata} is a stenohaline species that survives better where the fluctuations of salinity are narrower and inundation frequency is low. The low intertidal, where salinity fluctuations are wide in combination with high inundation frequency, is therefore not the most suitable zone for such mangrove species. Based on our simulation results, the long-term chances for survival of restored forests designed in such a manner is between slim to null. Indeed, in the absence of a success index to guide restoration efforts, high costs of replanting are incurred without generating any real long-term positive returns. For example, in plans of building bio-shield greenbelt zones, a thick forest density [of at least $1\mbox{ km}$ width~\citep{McIvor2012}] must be guaranteed to effectively function as carbon sink and protect the coastal areas against natural disasters such as storm surges.
\section*{Acknowledgments}
This research was funded by the Commission on Higher Education (CHED) Philippine Higher Education Research Network (PHERNet). A portion of the study was presented at the PAMS12 conference in Tacloban City, Philippines in October 2013. Support from the 2013 SATU Joint Research Scheme with the International Wave Dynamics Research Center of the National Cheng Kung University, Taiwan is also acknowledged.
|
\section{#1}}
\newcommand{\begin{equation}}{\begin{equation}}
\newcommand{\end{equation}}{\end{equation}}
\newcommand{\begin{equation}}{\begin{equation}}
\newcommand{\end{equation}}{\end{equation}}
\newcommand{\begin{eqnarray}}{\begin{eqnarray}}
\newcommand{\end{eqnarray}}{\end{eqnarray}}
\newcommand{\uparrow}{\uparrow}
\newcommand{\downarrow}{\downarrow}
\newcommand{\langle}{\langle}
\newcommand{\rangle}{\rangle}
\newcommand{\dagger}{\dagger}
\newcommand{{\rm i}}{{\rm i}}
\newcommand{{\rm e}}{{\rm e}}
\newcommand{{\rm d}}{{\rm d}}
\newcommand{\partial}{\partial}
\newcommand \lab {\label}
\newcommand{\varepsilon}{\varepsilon}
\newcommand{\hspace{0.7cm}}{\hspace{0.7cm}}
\newcommand{\sigma}{\sigma}
\newcommand{\longrightarrow}{\rightarrow}
\newcommand{\bar{z}}{\bar{z}}
\def\theta{\theta}
\def\longrightarrow{\longrightarrow}
\setlength{\textwidth}{160mm}
\setlength{\textheight}{230mm}
\setlength{\headsep}{0in}
\setlength{\baselineskip}{0.375in}
\setlength{\oddsidemargin}{0cm}
\setlength{\evensidemargin}{0cm}
\newcommand{\Cblu}[1]{\textcolor{blue}{#1}}
\newcommand{\Cgre}[1]{\textcolor{green}{#1}}
\newcommand{\Cred}[1]{\textcolor{red}{#1}}
\begin{document} \setcounter{page}{0}
\topmargin 0pt
\oddsidemargin 5mm
\renewcommand{\thefootnote}{\arabic{footnote}}
\newpage
\setcounter{page}{0}
\topmargin 0pt
\oddsidemargin 5mm
\renewcommand{\thefootnote}{\arabic{footnote}}
\newpage
\begin{titlepage}
\begin{flushright}
\end{flushright}
\vspace{0.5cm}
\begin{center}
{\large {\bf Quantum quenches with integrable pre-quench dynamics}}\\
\vspace{1.8cm}
{\large Gesualdo Delfino}\\
\vspace{0.5cm}
{\em SISSA -- Via Bonomea 265, 34136 Trieste, Italy}\\
{\em INFN sezione di Trieste}\\
\end{center}
\vspace{1.2cm}
\renewcommand{\thefootnote}{\arabic{footnote}}
\setcounter{footnote}{0}
\begin{abstract}
\noindent
We consider the unitary time evolution of a one-dimensional quantum system which is in a stationary state for negative times and then undergoes a sudden change (quench) of a parameter of its Hamiltonian at $t=0$. For systems possessing a continuum limit described by a massive quantum field theory we investigate in general perturbative quenches for the case in which the theory is integrable before the quench.
\end{abstract}
\end{titlepage}
\newpage
\noindent
Remarkable advances in experiments with cold atomic gases have triggered in the last years a strong interest in non-trivial unitary time evolution of closed quantum systems (see e.g. \cite{PSSV} for a review). A favorite case study is the quantum quench \cite{CC}, in which an eigenstate (normally the ground state) of a closed system with Hamiltonian $H_0$ evolves unitarily for times $t>0$ according to a new Hamiltonian obtained by a sudden change of a parameter of $H_0$ at $t=0$. In particular, integrable one-dimensional systems have attracted special interest due to experimental observations \cite{KWW} and theoretical proposals \cite{RDYO} suggesting that integrability or quasi-integrability may affect in a distinctive way the long time behaviour of the system.
In this paper we consider quenches of one-dimensional quantum systems admitting a continuum limit described by a massive quantum field theory which is integrable before the quench. The theory is specified by the action
\begin{equation}
{\cal A}={\cal A}_0-\lambda\int_0^\infty dt\int_{-\infty}^{\infty} dx\,\Psi(x,t)\,,
\label{action}
\end{equation}
where ${\cal A}_0$ is the action of the integrable pre-quench theory, $\lambda$ the quench parameter, and $\Psi$ its conjugated local operator, which is relevant in the renormalization group sense, i.e. has scaling dimension\footnote{The theory with action ${\cal A}_0$ is an off-critical deformation by relevant operators of the conformal field theory describing a quantum critical point. All scaling dimensions refer to this critical point.} $X_\Psi<2$.
Integrability in (1+1)-dimensional quantum field theory follows from the presence of infinitely many conserved quantities and amounts to two drastic simplifications of the scattering theory \cite{ZZ}, namely complete elasticity (the final state is kinematically identical to the initial one) and complete factorization (the scattering amplitudes of $n$-particle processes factorize into the product
of $n(n-1)/2$ two-body amplitudes). Factorization follows from the fact that conserved quantities other than energy and momentum act as generators of momentum dependent space-time translations of the trajectories of the relativistic particles, allowing to resolve a multi-particle scattering into a sequence of two-particle scatterings separated by arbitrary large distances in space-time. Once added to the general properties of unitarity and crossing symmetry, elasticity and factorization allow the exact determination of the two-particle amplitudes\footnote{In order to simplify the notation, we develop the main discussion referring to theories with a single species of particles. Generalizations will be discussed when relevant.} $S(p_1,p_2)$, and then of the whole $S$-matrix. The theory with action ${\cal A}_0$ is integrable in this sense.
\begin{figure}[t]
\begin{center}
\includegraphics[width=13cm]{quench_fig.eps}
\caption{A particle is transmitted and a pair is created at $t=0$. Pictorial illutstration of the equations following from the requirement of complete factorizabililty of the amplitude.}
\label{factorization}
\end{center}
\end{figure}
We examine first of all under which conditions a similar notion of exact solvability can extend to the quenched theory (\ref{action}). A necessary condition is that the scattering is elastic and factorized also after the quench. Concerning the passage from negative to positive times, requiring elasticity means that particles transmitted through $t=0$ preserve their momentum; on the other hand, the breaking of time translation invariance at $t=0$ allows for creation or destruction of sets of particles with zero total momemtum (momentum is conserved by (\ref{action})). A typical example of a process in the quenched theory is that in which a particle of momentum $p$ is transmitted through $t=0$ and scatters with a pair of particles with momenta $q$ and $-q$ produced at $t=0$. The requirement of complete factorizability implies that the total amplitude must be unaffected by an arbitrary translation of, say, the trajectory of the transmitted particle. Depending on the value of $q$ this gives rise to two equations that are depicted in Fig.~\ref{factorization}. The total amplitudes contain common factors $T(p)$ and $K(q)$, namely the transmission amplitude for the particle with momentum $p$ and the creation amplitude for the pair, respectively. Factorizability then constrains the post-quench amplitudes $S(p_1,p_2)={\cal S}(\theta_1-\theta_2)$, where we introduced the rapidity parameterization $(E_p,p)=(m\cosh\theta,m\sinh\theta)$ for the energy and momentum of a relativistic particle with mass $m$; the dependence of the two-particle amplitude on the rapidity difference follows from relativistic invariance. The pictorial equations of Fig.~\ref{factorization} correspond to
\begin{eqnarray}
& {\cal S}(\theta+\beta)={\cal S}(\beta-\theta)\,,\nonumber\\
& {\cal S}(\theta+\beta){\cal S}(\theta-\beta)=1\,.\nonumber
\end{eqnarray}
The first equation implies that ${\cal S}$ is a constant, and the second (or, equivalently, unitarity) that ${\cal S}=\pm 1$, namely that the post-quench particles are either free bosons or free fermions\footnote{In the context of theories with a defect line a similar factorization argument was used in \cite{DMS} and generalized to more particles species in \cite{CFG}.}. We arrive at the same conclusion for the pre-quench particles if we consider the process in which an initial state with three particles with momenta $p$, $q$ and $-q$ evolves into the final state with a single particle with momentum $p$ through pair desctruction at $t=0$. The conclusion we reach in this way is that the only quenches compatible with factorization are mass quenches for free particles (i.e. ${\cal A}_0$ is a free action and $\lambda\Psi$ is proportional to its mass term). The amplitudes $T(p)$ and $K(p)$ have been determined by Bogoliubov transformations in \cite{CC} for free bosons and in \cite{RSMS} for free fermions.
This argument indicates that the only way to deal analytically with interacting particles in quenched massive field theories is the perturbative one. Here we treat (\ref{action}) perturbatively in $\lambda$, keeping ${\cal A}_0$ integrable. Perturbation theory around integrable quantum field theories in the time translation invariant case was formulated in \cite{nonint}. In the present case the perturbation $\lambda\Psi$ acts only for positive times, but it is still possible to set up a scattering description based on the {\em asymptotic states} $|p_1,\ldots,p_n\rangle_{in\,(out)}$ {\em of the integrable pre-quench theory}. An initial state $|p_1,\ldots,p_n\rangle_{in}$ (which is an eigenstate of the pre-quench Hamiltonian $H_0$ with eigenvalue $\sum_{i=1}^nE_{p_i}$, $E_{p_i}=\sqrt{m^2+p_i^2}$) at $t=-\infty$ evolves at $t=+\infty$ into a final state that we expand on the complete basis of outgoing states of the pre-quench theory. The coefficients of this expansion are
\begin{equation}
{}_{out}\langle q_1,\ldots,q_m|S_\lambda|p_1,\ldots,p_n\rangle_{in}\,,
\label{coeff}
\end{equation}
with
\begin{equation}
S_\lambda=T\,\exp\left(-i\lambda\int_0^\infty dt\int_{-\infty}^\infty dx\,\Psi(x,t)\right)
\label{Slambda}
\end{equation}
($T$ indicates chronological ordering); $S_\lambda$ becomes the identity for $\lambda=0$ and (\ref{coeff}) reduces to the elastic and factorized $S$-matrix of the integrable theory with action ${\cal A}_0$.
We now consider the case in which the system is for $t<0$ in the ground state (i.e. in the vacuum state $|0\rangle$) and $\lambda$ is a small parameter. Then, to lowest order in $\lambda$, the state $|\psi_0\rangle$ of the system after the quench can be written as
\begin{equation}
|\psi_0\rangle=S_\lambda|0\rangle\simeq |0\rangle+\lambda\sum_{n=1}^\infty\frac{2\pi}{n!}\int_{-\infty}^{\infty}\prod_{i=1}^n\frac{dp_i}{2\pi E_{p_i}}\,\delta(\sum_{i=1}^np_i)\,\frac{[F_n^\Psi(p_1,\ldots,p_n)]^*}{\sum_{i=1}^nE_{p_i}}\,|p_1,\ldots,p_n\rangle\,,
\label{psi0}
\end{equation}
\begin{equation}
F_n^\Psi(p_1,\ldots,p_n)=\langle 0|\Psi(0,0)|p_1,\ldots,p_n\rangle\,.
\label{ff}
\end{equation}
The matrix elements (\ref{ff}) are the form factors of the integrable theory with action ${\cal A}_0$ and are exactly computable (see \cite{Smirnov}). The factor $\sum_{i=1}^nE_{p_i}$ in the denominator of (\ref{psi0}) is obtained giving to the energy an infinitesimal imaginary part to make the time integral in (\ref{Slambda}) convergent. In writing (\ref{psi0}) we omitted the $n=0$ term of the sum, which reads $\lambda LT_+\langle 0|\Psi|0\rangle|0\rangle$, where $LT_+$ is the (infinite) post-quench space-time volume; this term corresponds to a renormalization of the vacuum energy density and is canceled introducing a corresponding counterterm in the exponent of (\ref{Slambda}) (see \cite{nonint} for details on an analogous procedure). In (\ref{psi0}) we also passed from the outgoing states $|p_1,\ldots,p_n\rangle_{out}$ and the corresponding integration over momenta restricted by $p_1<p_2<\ldots<p_n$ to the generic states $|p_1,\ldots,p_n\rangle$ and $1/n!$ times unrestricted integration. This is possible because interchanging the position of the particles with momenta $p_i$ and $p_{i+1}$ yields a factor\footnote{Here $S(p_1,p_2)$ denotes the scattering amplitude in the integrable pre-quench theory.} $S(p_i,p_{i+1})$ from $|p_1,\ldots,p_n\rangle$, together with a factor $S^*(p_i,p_{i+1})$ from the form factor, and the two factors cancel by unitarity. Since to lowest order the particles scatter after the quench with the same factorized pre-quench $S$-matrix, the independence on the ordering of the particles means that (\ref{psi0}) describes the state of the system at any positive time, i.e. it can be considered as a state of the Heisenberg picture in which all the time dependence is carried by the operators.
The $n=2$ term of (\ref{psi0}) reads $\frac{1}{2}\int\frac{dp}{2\pi E_p}K^{(1)}(p)|p,-p\rangle$, with $K^{(1)}(p)=\lambda\frac{[F_2^\Psi(p,-p)]^*}{2E^2_p}$. For the mass quench in free theories one can verify that $K^{(1)}(p)$ coincides with the lowest order of the pair creation amplitude obtained by Bogoliubov transformation. In these solvable cases $F_n^\Psi$ vanishes for $n\neq 2$.
The average energy produced in the quench is\footnote{When computing averages on $|\psi_0\rangle$ we should divide by $\langle\psi_0|\psi_0\rangle$. This normalization, however, is $1+O(\lambda^2)$ and can be ignored to lowest order.}
\begin{equation}
\langle\psi_0|H_0|\psi_0\rangle\simeq\lambda^2\,L\sum_{n=1}^\infty\frac{2\pi}{n!}\int_{-\infty}^{\infty}\prod_{i=1}^n\frac{dp_i}{2\pi E_{p_i}}\,\delta(\sum_{i=1}^np_i)\,\frac{|F_n^\Psi(p_1,\ldots,p_n)|^2}{\sum_{i=1}^nE_{p_i}}\,,
\label{energy}
\end{equation}
where $L=2\pi\delta(0)\to\infty$ is the size of system. Since it was shown in \cite{immf} that for a scalar operator $\Phi$ with scaling dimension $X_\Phi$
\begin{equation}
F_n^\Phi(p_1,\ldots,p_n)\sim |p_i|^{y_\Phi}\,,\hspace{.6cm}|p_i|\to\infty\,;\hspace{1.2cm}y_\Phi\leq X_\Phi/2\,,
\label{asymp}
\end{equation}
a sufficient condition for the convergence of the integrals in (\ref{energy}) is\footnote{If this condition is not satisfied one can resort to a more specific asymptotic analysis \cite{D_09}. Divergent integrals at this stage signal operator mixing; see \cite{DSC} for a discussion in the context of a different application.} $X_\Psi<1$. Typically the contribution of a state to a form factor expansion rapidly decreases as the total mass increases.
Using the notation $\delta F(\lambda)=F(\lambda)-F(0)$, we have for one-point functions of hermitian operators
\begin{eqnarray}
\delta\langle\psi_0|\Phi(x,t)|\psi_0\rangle &\simeq &\lambda\sum_{n=1}^\infty\frac{2\pi}{n!}\int_{-\infty}^{\infty}\prod_{j=1}^n\frac{dp_j}{2\pi E_{p_j}}\,\frac{\delta(\sum_{j=1}^np_j)}{\sum_{j=1}^nE_{p_j}}\,\nonumber\\
&\times& 2\mbox{Re}\{[F_n^\Psi(p_1,\ldots,p_n)]^*F_n^\Phi(p_1,\ldots,p_n)\,e^{-i\sum_{j=1}^nE_{p_j}t}\}\,,
\label{1point}
\end{eqnarray}
where the integrals converge for $\Phi$ relevant.
When considering the long time behavior we have to take into account that there are two time scales, i.e. $1/m$ and $t_\lambda=1/\lambda^{1/(2-X_\Psi)}$; the latter goes to infinity as $\lambda$ goes to zero. Within the perturbative approach in $\lambda$, long time means staying in the intermediate regime $1/m\ll t\ll t_\lambda$, so that the limit of small $\lambda$ prevails. At large times the integrand of (\ref{1point}) rapidly oscillates, we replace $e^{-iE_{p_j}t}$ with $e^{-i(m+p_j^2/2m)t}$, and the integrals receive the main contribution from a range of $|p_j|$ with width of order $\sqrt{m/t}$. Since, with the exception of free bosons, the scattering amplitude of two identical particles satisfies $S(p,p)=-1$, the factor $[F_n^\Psi]^*F_n^\Phi$ in (\ref{1point}) evaluated for momenta all tending to zero becomes $\prod_{1\leq i<k\leq n}(p_i-p_k)^2$ times a constant. Hence, following this analysis, the passage to the variables $q_j=\sqrt{\frac{t}{m}}p_j$ gives for the $n$-particle term in (\ref{1point}) a leading intermediate time behavior
\begin{equation}
\frac{\lambda}{t^{(n^2-1)/2}}\,\mbox{Re}(c_n\,e^{-inmt})\,,
\label{leading}
\end{equation}
with coefficients $c_n$ which depend on ${\cal A}_0$, $\Psi$ and $\Phi$. In theories with more particle species, particles of different species with momenta $p_i$ and $p_k$ do not necessarily contribute to the integrand the factor $(p_i-p_k)^2$, and time suppression for $n>1$ may be reduced. In all cases, however, if the form factors\footnote{One-particle form factors of scalar operators are momentum-independent.} $F_{1,a}^\Psi$ and $F_{1,a}^\Phi$ over a single particle of species $a$ are both non-zero, the leading intermediate time behavior is
\begin{equation}
\lambda\sum_a\frac{2}{m_a^2}\mbox{Re}\{[F_{1,a}^\Psi]^*F_{1,a}^\Phi\,e^{-im_at}\}\,.
\label{leading1}
\end{equation}
A convenient illustration is provided by the Ising field theory (see \cite{review} for a review), namely the theory describing the scaling region around the simplest quantum critical point possessing a $Z_2$ symmetry. With reference to the action (\ref{action}), we consider first the case in which ${\cal A}_0$ is massive and $Z_2$-symmetric; this is the continuum limit of the transverse field Ising spin chain and corresponds to the theory of a free neutral fermion. The first possibility is to take $\Psi$ equal to $\varepsilon$, the relevant $Z_2$-invariant operator ($X_\varepsilon=1$), which with the present choice of ${\cal A}_0$ has $F_n^\varepsilon\neq 0$ only for $n=2$. This quench corresponds to one of the exactly solvable cases (mass quench for free fermions, $\lambda=\delta m$). The intermediate time behavior of (\ref{1point}) for the order parameter operator $\sigma$ ($X_\sigma=1/8$) in the ferromagnetic phase (the one-point function vanishes by symmetry in the paramagnetic phase) corresponds to (\ref{leading}) with $n=2$; this agrees with the result of \cite{CEF,SE}. The second possibility is to take $\Psi=\sigma$, a choice which breaks $Z_2$ symmetry after the quench. Now the intermediate time behavior of $\delta\langle\psi_0|\sigma(x,t)|\psi_0\rangle$ should be given by (\ref{leading}) with $n=1$ if we start from the paramagnetic phase, and $n=2$ if we start from the ferromagnetic phase\footnote{The operator $\sigma$ has non-zero form factors on states with odd (resp. even) number of particles in the paramagnetic (resp. ferromagnetic) phase.}.
If instead we move away from the quantum critical point in the direction of the operator $\sigma$ we obtain a theory ${\cal A}_0$ which is no longer $Z_2$ invariant but is still integrable and contains eight species of particles with different masses \cite{Taniguchi} (see \cite{Coldea} for an experimental study). Both for $\Psi=\sigma$ and $\Psi=\varepsilon$ the leading intermediate time behavior of $\delta\langle\psi_0|\sigma(x,t)|\psi_0\rangle$ is given by (\ref{leading1}), with form factors $F_{1,a}^\sigma$ and $F_{1,a}^\varepsilon$ ($a=1,\ldots,8$) which are also known \cite{immf,DS,review}. The case $\Psi=\sigma$ now corresponds to the quench of the mass scale in an interacting theory.
Two-point correlators $\langle\psi_0|\Phi_1(x_1,t_1)\Phi_2(x_2,t_2)|\psi_0\rangle$ can also be expanded over form factors inserting a complete set of states of the pre-quench theory in between the two operators. This results in a double sum over particle states that we do not try to analyze here.
\vspace{.3cm}
Summarizing, we considered quantum quenches in massive (1+1)-dimensional field theory. After arguing that exactly solvable cases restrict to mass quenches in free theories, we formulated the perturbative theory in the quench parameter for the case of integrable pre-quench dynamics. In particular, we determined to lowest order the average energy per unit length produced in the quench and analyzed the behavior of one-point functions of local operators for times much larger than the pre-quench time scale but much smaller than the time scale associated to the quench parameter.
|
\section{Introduction}
In the age of Web 2.0, OSN have gained pervasive interests in both research communities and industries. There is a trend to model OSN using Semantic Web technologies, especially the vocabularies from
FOAF\footnote{\url{http://www.foaf-project.org/}} and SIOC\footnote{\url{http://rdfs.org/sioc/spec/}} project. RDF, originally designed for the Semantic Web, have been wildly adopted for representing such kind of linked data. SPARQL\footnote{\url{http://www.w3.org/TR/rdf-sparql-query/}} as the de-facto RDF query language is used for Basic Graph Pattern(BGP) query with bounded or unbounded variables. The scalable graph representation and flexible query capability make RDF model suited for large-scale complex OSN management and analysis.
\begin{figure}
\centering
\includegraphics[width=3 in]{fig1}
\caption{Examples for a fraction of social network. }
\label{fig:graph-pattern}
\end{figure}
Figure \ref{fig:graph-pattern} illustrated a snippet of large social graph representing relations between four users and three User Generated Contents(UGC). Query which \textit{Find pair of users in a path of friend relationship which user2 has a job and like the documents created by user1 } is expressed in SPARQL as:
\begin{lstlisting}[
language=SQL,
showspaces=false,
basicstyle=\ttfamily,
commentstyle=\color{gray}
]
SELECT DISTINCT ?user1, ?user2 WHERE {
?user1 knows* ?user2 .
?user1 creatorOf ?doc1 .
?user2 worksFor ?organization .
?doc1 likedBy ?user2 }
\end{lstlisting}
For RDF graph in Figure \ref{fig:graph-pattern}, this query returns the result set $R(q) =\{<P1, P3>\}$. The pattern \emph{?user1 knows* ?user2} states the path consist of zero or more \emph{knows} predicates, it is a property path query pattern.
\vspace{6pt} \noindent \textbf{Challenges.} Path queries are of common interest in OSN analysis for discover complex relations among entities. Despite the scalability and flexibility provided by RDF model, Path queries performs poorly and lack of efficient implementation in existing RDF management. Current researches in graph analysis domain are mainly based on pre-constructed reachability indices. Building such indices are both time and memory consuming, especially when dealing with large complex graph. From RDF management point of view, due to costly join for triple pattern matching, there lacks efficient implementation of property path query. Although property path query is recommended by newest SPARQL 1.1 standard, to the best of our knowledge Jena\footnote{\url{https://jena.apache.org/}}, Virtuoso\footnote{\url{http://virtuoso.openlinksw.com/}} and Sesame\footnote{\url{http://www.openrdf.org/}} are the only three off-the-shelf RDF stores that support standardized path query. They all suffer from performance problems when dealing with path query on large-scale data.
\vspace{6pt} \noindent \textbf{Overview of Our approach.} Our approach is motivated by three observations. First, current
off-the-shelf RDF store performs well only on join-based star queries. Second, the search space of property path query is restricted to triples representing relations among entities. Third, due to the rich semantics in OSN, most triples are for entities attributes not for relations. Based on these observations, we argue that manage graph topology-related triples into main memory is feasible, and this will greatly enhance the online query performance of property path related sophisticate SPARQL queries. We propose a hybrid framework that has in-memory management of graph topology-related RDF data along with disk-based Jena TDB\footnote{\url{https://svn.apache.org/repos/asf/jena/}} native triple store. In our approach, while loading and indexing triples into TDB, graph topology-related triples are identified and duplicated in main memory. For an online query, we implement a graph traversal based algebra operators for property path pattern, which is more efficient than traditional join-based operator.
\vspace{6pt} \noindent \textbf{Contributions.} We summarized our contribution in this paper as:
\begin{enumerate}
\item We propose a main memory/disk based hybrid framework for efficient property path query. While leverage the functionality of existing well-established RDF store for BGP query, it specialized in property path pattern query through manage graph topology related data in main memory.
\item We present an algebra operator for property path pattern query. Its realization based on in-memory graph traversal instead of costly join. Using the characteristics of OSN, heuristics for estimating the execution cost of this operator is given that can be used for cost-base optimization.
\end{enumerate}
\vspace{6pt} \noindent \textbf{Organization of the paper.} The rest of the paper is organized as follows. Section \ref{sec:hybrid} presents the basic design of
our hybrid RDF management framework. Section \ref{sec:evaluation} shows the evaluation results. We conclude in Section \ref{sec:conclusion}.
\section{Hybrid RDF Data Management}\label{sec:hybrid}
Given vocabulary $\Sigma = V_{E} \cup V_{A} \cup E_{EE} \cup E_{EA} \cup L_{E} \cup L_{A} $ defined in Table \ref{table:notations}, an OSN is represented as a triple set $T_{OSN} = T_G \cup T_A $, where the graph-topology set $T_G \subseteq V_{E} \times L_{E} \times V_{E}$ holds triples representing social entities and relations among them, and the attributes set $T_A \subseteq V_{E} \times L_{A} \times V_{A}$ holds triples representing attributes and theirs relations to social entities.
\begin{table}[ht]
\label{table:notations}
\caption{Notations for social graph representation.}
\centering
\begin{tabular}{ | c | c | c |}
\hline
Notations & Refers to the set of & Instance of Figure 1 \\[0.5ex]
\hline
$V_{E}$ & Nodes for social entities. & $\{P1,D1,\ldots\}$ \\
$V_{A}$ & Nodes for attributes values. & $\{"John", "London", "abcde", \ldots\}$ \\
$E_{EE}$ & Edges among $V_{E}$. & $\{(P1,P2),(P1,D1), \ldots\}$ \\
$E_{EA}$ & Edges between $V_{E}$ and $V_{A}$. & $\{(P1,"Sam"),(D1,"abcde"), \ldots\}$ \\
$L_{E}$ & Labels for $E_{EE}$. & $\{<knows>,<likeBy>,\ldots\}$ \\
$L_{A}$ & Labels for $E_{EA}$. & $\{<hasName>, <ns\#type>, \ldots\}$ \\[1ex]
\hline
\end{tabular}
\label{table:nonlin
\end{table}
\subsection{Architectural Design}\label{subsec:architecture}
In this paper, we focus on efficient implementation of property path query. Consider that path query only related to Triples in $T_G$ (In Figure \ref{fig:graph-pattern}, $T_G$ is represented as dashline-encircled part), one direct motivation is that manage $T_G$ in memory will greatly enhance the overall query performance. Based on this motivation, we implement a hybrid RDF data management architecture that manage different query-prone data respectively. This architecture is shown in Figure \ref{fig:architecture}.
\begin{figure}
\centering
\includegraphics[width=3.2in]{fig2}
\caption{Hybird RDF management architecture of our approach }
\label{fig:architecture}
\end{figure}
Our approach can be thought as a plugin component which override the functionality of corresponding part in TDB. At data loading stage, $\forall t_i \in T_{OSN}$ is loaded into Jena TDB (step \textcircled{1} in Figure \ref{fig:architecture}). At the same time, $\exists t_i \in T_G$ is filtered out based on two kinds of rules:
\begin{enumerate}
\item The type of \textit{Objects}. If \textit{Objects} for $t_i$ is a literal, then $t_i \in T_A$ .
\item The semantic meaning of \textit{Predicate}. Such as \textit{foaf:knows} defined in FOAF project which state the relation that \emph{Subject User} know \emph{Object User}.
\end{enumerate}
$T_G$ is stored in main memory with subject index(PSO) for forward traversal and object index(POS)for backward traversal(step \textcircled{2}).These indices can be constructed incrementally when topology-related data is extracted and loaded into main memory.
When an online query is submitted, SPARQL parser translates query strings into patterns based on standard abstract syntax tree recommended by W3C (step \textcircled{3}). Analyzer translates these patterns into predefined SPARQL algebra operators and construct execution plan (step \textcircled{4}).An algebra operator generates designated result set from given input (step \textcircled{5}). In our approach we implemented a special operator named \textit{OpPath} which only uses in-memory data as input. If a query string in \textit{WHERE} clause is analyzed as property path pattern, \textit{OpPath} operator is added to the query plan. (step \textcircled{6}) We explain the design of \textit{OpPath} operator in detail in Section \ref{subsec:operator}. Result set of algebra operator is joined to get the final result set. The join order of operators is optimized using cost and selectivity estimation (step \textcircled{7}).
\subsection{Property Path Algebra Operator}\label{subsec:operator}
\begin{definition}[\textit{OpPath} Operator]
\textit{OpPath} is a ternary algebra operator that can be defined as \textit{OpPath(O,S,$P_P$)}. $S,O \subseteq V_{EE}$ can be either bounded or unbounded variables, $|S| =s$, $|O| =o$ . $P_P$ is a regular pattern expression defining the property path. \textit{OpPath(O,S,$P_P$)} operator find existing path from set $S$ to set $O$, and return all triple sets that each paths is consist of as result set.
\end{definition}
Based on research\cite{zeng2013distributed} which have testified that graph explorations is extremely efficient and more easy to implement than costly joins, the \textit{OpPath} operator is realized as in-memory Breadth-First Search(BFS). The \textit{OpPath} operator has time complexity $O(|V_E|+|E_{EE}|)$ and space complexity $O(|T_G|)$. It is much less than tradition nested-loop join that has time complexity $O(|V_E| \cdot |E_{EE}|)$.
The cost of \textit{OpPath} operator is the cardinality of result set $R(q)$ for path query pattern $q$. Existing researches such as G-SPARQL\cite{sakr2012g} using predefined heuristics which always take $|R(q)|$ as the largest, this is far from optimal. Sparqling Kleene\cite{gubichev2013sparqling} using pre-computed reachability path indices which affects data load efficiency. We consider three factors that affects $|R(q)|$, the average nodes out-degree, the path length $l$ and the pathes that fits for the given pattern. In our approach, we leverages the graph generation model \cite{leskovec2007graph} which expects the average out-degree as $d_{out}=|V_{EE}|^{1-\ln c}, 1<c\leq 2$. We also assumes that nodes in $T_G$ has the same probability of being added to the path, thus the modifying factor of out-degree follows the binomial distribution. For all considerations above, $|R(q)|$ can be approximately estimated as:
\begin{equation}
\label{equ:cost}
|R_q| = s\cdot o \cdot \sum_{i=1}^{l}(|V_{EE}|^{(1-\ln c) \cdot i} \cdot (\sum_{i=1}^{l}\frac{l!}{i!(l-i)!} \cdot p^i \cdot(1-p)^{l-i} ))
\end{equation}
where $p=\frac{|E_{EE}|-|V_{EE}|}{|V_{EE}|}$. $|V_{EE}|$ and $|E_{EE}|$ can be got from the metadata in RDF store. We performs preliminary testing to measure the accuracy of Equation \ref{equ:cost} with real all-pair cardinality of dataset in Table \ref{table:datasets}. For SNIB $T_G$ with average $d_{out}=12$, $c=1.75$, $relative\ error = \frac{max(real\ cardinality ,estimate\ cardinality)}{min(real\ cardinality ,estimate\ cardinality)}-1$ is about $27\%$. For DBLP $T_G$ with $d_{out}=7$, $c=1.81$, and $error = 32\%$. This preliminary testing shows that the heuristic defined in Equation \ref{equ:cost} is with acceptable cardinality estimation error.
\section{Evaluation}\label{sec:evaluation}
We used one machine with Debian 7.4 in 64-bit Linux kernel, two Intel Xeon E5-2640 2.0GHz processor and 64 GB RAM for our evaluation. Our approach is compared with Jena(version 2.11.1), Sesame(version 2.7.10) and competitive research G-SPARQL\cite{stocker2008sparql}. Jena implements path query based on join while Sesame is based on graph traversal. We also implements our approach with no cost estimation and treats path query as the most costly(denoted as \textit{NoCE}). All evaluations were done 10 times and the results are the averages.
We adopted two datasets, the Social Network Intelligence Benchmark (SNIB)\footnote{Social Network Intelligence Benchmark(SNIB), \url{http://www.w3.org/wiki/Social_Network_Intelligence_BenchMark/}.} as synthetic dataset, and DBLP as real dataset, all in RDF N-Triples format\footnote{\url{http://www.w3.org/TR/n-triples/}}. SNIB dataset is generated using S3G2\cite{pham2013s3g2}. Considering G-SPARQL uses ACM digital library dataset which is not publicly available, we uses the DBLP dataset instead \footnote{DBLP dataset can be download from \url{http://sw.deri.org/~aharth/2004/07/dblp/dblp-2006-02-06.rdf}. Same as stated in G-SPARQL, we manually created the co-author relationships between author nodes, which originally recorded as $<creator>$ tag in raw dataset.}. Statistics of datasets are shown in Table \ref{table:datasets}. The DBLP dataset has approximately the same characteristics as the \textit{Large Graph Size} experiment in G-SPARQL.
\begin{table}[ht]
\caption{Statistics for SNIB and DBLP datasets.}
\label{table:datasets}
\centering
\begin{tabular}{ | c | c | c | c | c |}
\hline
Dataset & Vertices($|V_{EE}|$) & Edges ($|E_{EE}|$) & Attributes ($|T_{A}|$) & $|T_G|/|T_{OSN}|$ \\[0.5ex]
\hline
SNIB & $566,472$ & 2,001,333 & 7,273,177 & $26\%$ \\
DBLP & $900,440$ & 2,243,827 & 9,363,166 & $25\%$ \\
\hline
\end{tabular}
\label{table:dataset}
\end{table}
For offline data loading time and memory expenses, we compare our approach with four competitors, Sesame and Jena in-memory store which store and index data only in memory, Sesame native store and Jena TDB which use disk as triple storage. Results of data loading time is represented in Figure \ref{subfig:loadtime},and disk usage in Figure \ref{subfig:diskusage}, memory usage in Figure\ref{subfig:memusage}. For our approach only load graph topological data into main memory, it need fewer memory than that of Jena and Sesame in-memory store, but with a little overhead of the data loading time.
\begin{figure}
\centering
\subfigure[Data loading time. ]{
\label{subfig:loadtime}
\includegraphics[width=1.4in]{fig3a}}
\subfigure[Disk usage. ]{
\label{subfig:diskusage}
\includegraphics[width=1.4in]{fig3b}}
\subfigure[Memory usage. ]{
\label{subfig:memusage}
\includegraphics[width=1.4in]{fig3c}}
\caption{\textit{Offline performance evaluation.} }
\label{fig:offline}
\end{figure}
For online query performance, in SNIB benchmark only $Q3$ and $Q5$ are path query related, while for DBLP we chosen 7 out of total 12 queries in G-SPARQL experiments(denoted as $Q_\_g$). In order for comparison, we had to rewrite these queries on DBLP dataset. Figure \ref{subfig:snib} shows that our approach achieved the best performance for SNIB $Q3$. As for SNIB $Q5$, the 3-HOP which expressed in UNION clause is explicitly parsed into six joins. This causes an expensive join expenses. Results in Figure \ref{subfig:gsparql} show that our approach works better than G-SPARQL(got directly from the \textit{Large Graph Size} result in \cite{sakr2012g}), while \textit{NoCE} has approximately the same performance as that of G-SPARQL. This shows that cost estimation for optimal join order can enhance the overall query performance.
\begin{figure}
\centering
\subfigure[\textit{SNIB query.}]{
\label{subfig:snib}
\includegraphics[height=1.5in, width=1.2 in]{fig4a}}
\subfigure[\textit{DBLP query performance.}]{
\label{subfig:gsparql}
\includegraphics[height=1.5in,width=2.8 in]{fig4b}}
\caption{\textit{Online Query Performance.} }
\label{fig:online}
\end{figure}
\section{Related Works}
Most existing RDF stores uses a relational model to manage data, either in a traditional RDBMS or using a native triple store. They all processes SPARQL queries as sets of join operations using disk-based indices, which are costly for sophisticated joins. Some researches have focused on compressing and managing RDFs in main memory, Trinity RDF\cite{zeng2013distributed} is the most prominent among them. It uses graph exploration instead of join operations and greatly boosts SPARQL query performance. But manages all data in memory is not trivial, distributed shared memory increases the complexity for maintenance.
From graph analysis perspective, Property path can be viewed as the label-constraint reachability problem on labeled graph. Though reachability on graph is a further investigated problem, few literatures\cite{fan2011adding, zou2014efficient} considered its usage in property graph, which is more common in nature. These researches are mainly focus on building reachability indices in advance, which is time or space consuming and not adequate for large-scale data management.
Besides Jena and Sesame, several frameworks and prototypes have been proposed for path queries\cite{janik2005brahms, sakr2012g, gubichev2013sparqling}. BRAHMS\cite{janik2005brahms} only supports query on paths with predefined length. Sparqling Kleene\cite{gubichev2013sparqling} realizes join based on pre-constructed reachability indices which are space consuming. Our work is mainly motivated by G-SPARQL\cite{sakr2012g} which use the same hybrid storage and manage graph topological data in memory. Our work different from G-SPARQL in that, first, we are not design a new query language but uses standard SPARQL 1.1 instead, this makes our work more general. Second, G-SPARQL uses index-free pointer-based data structure to representing the graph topological data in memory, in order to suit for most graph algorithms. Our work only cares about path patterns which BFS algorithm is used to answer such reachability queries. We build simple indices only for facilitating BFS.
\section{Conclusion}\label{sec:conclusion}
In this paper we addressed the problem of property path query in RDF data, presented a step towards
incorporation of in-memory storage. In our approach we are not trying to invent new wheels, but managed to combine existing effective approaches as well as some technical enhancements. Contrast to traditional RDF management and graph query method, we used in-memory graph traversal instead of costly join to realize path query operator, used simple graph indices other than RDF permutation indices and complex graph reachability indices for efficient graph traversal. Evaluations have shown that our approach is feasible and efficient for process SPARQL property path queries.
|
\section{Introduction}\label{s1}
Nowadays the thermodynamic description of the efficient performance regimes in heat devices has
been becoming of special relevance, due to the growing importance
of saving energy resources in any operation of energy conversion. The Carnot's theorem states
the upper bounds of heat-energy conversion processes between two heat reservoirs at
temperatures $T_{c}$ and $T_{h}$ ($T_{c}<T_{h})$: for a heat device working as a heat engine (HE) the maximum
efficiency is $\eta_{\mathrm{C}}=1-\tau$ ($\tau \equiv T_{c}/T_{h}$), while working as a refrigerator (RE) the maximum coefficient of performance (COP) is $\varepsilon_{\mathrm{C}}=\tau/(1-\tau)$. However,
these upper bounds are of no practical use since they
refer to reversible cycles with zero power output for HE and zero cooling power for RE. For the
thermodynamic analysis of real heat devices (working at non-zero
rate along irreversible paths), different models and different
figures of merit (based on thermodynamic, economic, compromised, and
sustainable considerations) have been proposed~\cite{bejan96, lchen99,berry00,wu04,durmayaz04}.
A part of such models is founded on finite-time thermodynamics (FTT) considerations. FTT focuses on
irreversibilities caused by finite-rate heat transfers
between the working fluid, and the external heat reservoirs,
internal dissipation of the working fluid and heat leaks between the heat
reservoirs. The optimization procedure in FTT, carried out under a fixed cycle time,
usually assumes two degrees of freedom, that is, the inner temperatures of
the isothermal steps of the working system~\cite{lchen99,wu04,durmayaz04,jchen01,luisarias13}.
In spite of their analytical simplicity, these models can reproduce, at least qualitatively,
the power-efficiency and the cooling power-COP behaviors observed in real HE~\cite{jchen01} and
RE~\cite{gordon00,chen97}. More simplified FTT models assume the so-called
endoreversible approximation~\cite{lchen99,wu04,jchen01,luisarias13,gordon00},
where the heat leaks and the internal dissipations are neglected. In this case,
if the linear heat transfer laws are additionally assumed for the external heat exchange, the resulting efficiency at maximum power becomes the well-known Curzon-Ahlborn (CA) efficiency $\eta_{\mathrm{CA}}=1-\sqrt{\tau}=1-\sqrt{1-\eta_{\rm
C}}$~\cite{curzon75}.
In the recently proposed low-dissipation (LD) model~\cite{esposito10a}, the basic starting
point is that the entropy production in the isothermal hot
(cold) heat exchange process is assumed to behave as $\Sigma_h/t_h$
($\Sigma_c/t_c$), with $t_h$ and $t_c$ denoting the corresponding
time durations, and $\Sigma_h$ and $\Sigma_c$ being coefficients
containing information on the irreversibility sources.
Considering $t_h$ and $t_c$ as degrees of freedom for optimization, the LD model for HE provides the
upper ($\eta_{\rm {maxP}}^+=\frac{\eta_{\rm C}}{2-\eta_{\rm C}}$)
and lower ($\eta_{\rm {maxP}}^-=\frac{\eta_{\rm C}}{2}$) bounds for
the efficiency at maximum power $\eta_{\rm maxP}$ under extremely asymmetric
dissipation limits $\Sigma_h/\Sigma_c\rightarrow \infty$
and $\Sigma_h/\Sigma_c \rightarrow 0$, respectively.
Indeed, the LD model allows us to recover the Curzon-Ahlborn value $\eta_{\rm CA}$ when
symmetric dissipation is considered ($\Sigma_h=\Sigma_c$), in this case without assuming any specific heat transfer law.
An important result of the LD model for HE is the linking of the
CA efficiency to symmetry considerations, providing a unified
framework to understand the quasi-universal behavior shown by the
efficiency at maximum power of many different kinds of heat engines.
An extension of the LD model to RE was also reported~\cite{tu12d,carla12a,hu13},
but in this case considering, as a figure of merit, the $\chi$ function defined
as $\chi=\varepsilon \dot Q_c$, being $\varepsilon$ and $\dot Q_c$ the COP and the cooling power, respectively. The lower
($\varepsilon_{{\rm max} \chi}^-=0$) and upper ($\varepsilon_{{\rm max}
\chi}^+=(\sqrt{9+8\varepsilon_{\rm C}}-3)/2$) bounds of the COP at maximum
$\chi$ were obtained under extremely asymmetric dissipation limits
$\Sigma_h/\Sigma_c \rightarrow \infty$ and $\Sigma_h/\Sigma_c \rightarrow 0$, respectively~\cite{tu12d}. Under the symmetric condition ($\Sigma_h=\Sigma_c$), the optimized COP was found to be $\varepsilon_{{\rm
max\chi}}=\sqrt{1+\varepsilon_{\rm C}}-1\equiv \varepsilon_{\rm CA}$~\cite{carla12a}.
This result, obtained previously in different contexts~\cite{yan90,velasco97a,allahv10a},
could be viewed as a counterpart of the CA efficiency for HE, though this point is a current issue of discussion~\cite{apertet13a} (see Sec.~V).
Linear irreversible thermodynamics (LIT) is a well-founded formalism, which is focused on the irreversible evolution of macroscopic
systems allowing us to extend the scope of the equilibrium thermodynamics. LIT assumes systems in local equilibrium and defines thermodynamic
forces and fluxes both interlinked by means of linear relationships governing the macroscopic evolution. From the
LIT principles, it is possible to construct models of heat devices~\cite{vbroeck05,yuki09,arias08,apertet13b,cisneros06,cisneros08,tu13b,yuki14a}, from which
the optimization procedure considers the thermodynamic force as the sole relevant degree of freedom. In LIT the endoreversible features of the simplified FTT models (including the CA model~\cite{curzon75}) of heat devices are recovered under the so-called tight-coupling condition~\cite{vbroeck05,cisneros06,cisneros08}.
Minimally nonlinear irreversible thermodynamic (MNLIT) models have also
been proposed for cyclic and steady-state heat devices, in order
to account for possible thermal dissipation effects in the interaction between the working
system and the external heat reservoirs~\cite{yuki12a,yuki13a}.
These models incorporate the Onsager relations with an additional nonlinear dissipation term.
As in LIT, the optimization procedure also involves only one degree of freedom (the thermodynamic flux),
and the outstanding results provided by the low-dissipation models~\cite{esposito10a,tu12d,carla12a} are recovered by the MNLIT model as a particular case~\cite{yuki12a} and~\cite{yuki13a} (see Appendix A).
We stress that a numerical validation of the LD-model and thus the MNLIT model has been confirmed in a computer simulation of a Carnot cycle with a single particle~\cite{hoppenau13a}.
However, the MNLIT models are subject to some criticism. In Ref.~\cite{apertet13b} it is claimed that the MNLIT models are misleading, since dissipations should naturally appear in the LIT models when the local Onsager relations are extended to a global scale. Indeed, for the thermoelectric devices these authors of Ref.~\cite{apertet13b} showed that the Joule dissipation is well-founded in LIT based on this argument. However, for any generic heat device, the relation between the local and global scales may be too complicated or it simply has not been obtained yet. In such cases, the addition of the dissipative terms adopted in the MNLIT models could be considered as a reasonable and natural assumption in order to explain the thermodynamic (macroscopic) influence of the local dissipative effects on a generic heat device.
The present paper is aimed to present results obtained by the MNLIT models for both HE and RE, and it has two main goals: (i) to present a unified description of non-isothermal heat devices for HE and RE, making emphasis on the global performance characteristics of the behaviors of power-efficiency and cooling power-COP curves, including the non-tight-coupling case; and (ii) to analyze the maximum efficiency and the efficiency at maximum power regimes for HE and the maximum COP and the COP at maximum cooling power regimes for RE. Especially, we reveal the impacts of irreversibilities by heat leaks (degrees of the coupling strength), internal dissipations, and their interplay on the performance of the heat devices, by considering their various limits.
To attain these goals, after giving a brief theoretical background in Sec.~II, we present in Sec.~III
a detailed analysis of the main global performance characteristics of HE and RE in terms of the degrees of the coupling strength and the dissipation effects. Then Sec.~IV focuses on the optimum performance regimes, and, finally, we discuss our main results in Sec.~V.
The original idea of the MNLIT model, which was proposed with the motivation to explain and extend the LD model~\cite{yuki12a}, is also given in Appendix A.
\section{Theoretical background}\label{s2}
Although the main theoretical aspects for both HE and RE have been reported previously~\cite{yuki12a,yuki13a}, here for the sake of completeness,
we give a brief theoretical background emphasizing the unified framework for HE and RE.
\begin{figure}[h]
\includegraphics[scale=0.3]{heat_device4.eps}
\caption{Set up of minimally nonlinear irreversible heat devices: (a) heat engine and (b) refrigerator.
The thin arrows inside the bold arrows show the direction of the dissipation terms included in the heat fluxes.
}\label{heat_device}
\end{figure}
\subsection{Minimally nonlinear irreversible model for HE}\label{s2a}
We start from the entropy production rate $\dot{\sigma}$ of the total system (that is, the heat engine and the heat reservoirs).
Hereafter the dot denotes a quantity divided by cycle period for cyclic heat engines or a quantity per unit time for steady-state heat engines.
Because the internal state of the heat engine comes back to the original state after one cycle for cyclic heat engines
or it remains unchanged for steady-state heat engines,
$\dot{\sigma}$ can be written by the sum of the entropy-change rate of the heat reservoirs as
$\dot{\sigma}=\dot{S}_h+\dot{S}_c$ according to the local equilibrium hypothesis~\cite{callen}, where $S_i$ ($i=h, c$) denotes the equilibrium entropy of the heat reservoir.
It can be written as
\begin{eqnarray}
\dot{\sigma}=\dot{S}_h+\dot{S}_c=\dot{U}_h\frac{\partial S_h}{\partial U_h}+\dot{U}_c\frac{\partial S_c}{\partial U_c}=-\frac{\dot{Q}_h}{T_h}+\frac{\dot{Q}_c}{T_c},\label{eq.sigma_dot_he_1}
\end{eqnarray}
where we have used $\frac{\partial S_i}{\partial U_i}=\frac{1}{T_i}$ with $U_i$ being the equilibrium internal-energy of the heat reservoir, and
for HE, we denote by $\dot{Q}_h\equiv -\dot{U}_h$ the heat flux from the hot heat reservoir and by $\dot{Q}_c\equiv \dot{U}_c$ the heat flux into the cold heat reservoir, respectively.
We also denote by $\dot{W}\equiv P$ the power output.
Then, from the relation $\dot{Q}_c=\dot{Q}_h-P=\dot{Q}_h-F\dot{x}$ with $F$ and $x$ a generalized external force and its conjugate variable, respectively,
Eq.~(\ref{eq.sigma_dot_he_1}) can be rewritten as
\begin{eqnarray}
\dot{\sigma}=-\frac{F\dot{x}}{T_c}+\dot{Q}_h\left(\frac{1}{T_c}-\frac{1}{T_h}\right)\equiv J_1X_1+J_2X_2.\label{eq.sigma_dot_he_2}
\end{eqnarray}
It naturally leads us to define the thermodynamic flux $J_1\equiv \dot{x}$
(the motion speed of the heat engine)
conjugate to the thermodynamic force $X_1 \equiv -F/T_c$, and the other thermodynamic flux
$J_2\equiv \dot{Q}_h$ (the heat flux from the hot heat reservoir)
conjugate to the other thermodynamic force $X_2\equiv 1/T_c-1/T_h$, where these quantities are expressed by using the thermodynamic extensive and intensive parameters of the equilibrium heat reservoirs~\cite{callen}.
By expanding the thermodynamic flux $J_i$ by the thermodynamic force $X_i$ around the equilibrium point $X_1=X_2=0$ as
$J_i=\sum_{j=1}^2 L_{ij}X_j+\sum_{j,k=1}^2 M_{ijk}X_jX_k+\sum_{j,k,m=1}^2 N_{ijkm}X_jX_kX_m+\cdots$
with $L_{ij}$'s, $M_{ijk}$'s, and $N_{ijkm}$'s being the expansion coefficients of each order, we obtain a full description of the evolution of the entropy production rate of the nonequilibrium heat engine.
The LIT model assumes the following linear Onsager relations between the thermodynamic fluxes and forces~\cite{vbroeck05}:
\begin{eqnarray}
&&J_1=\dot{x}=L_{11}X_1+L_{12}X_2,\label{eq.onsager_J1_he}\\
&&J_2=\dot{Q}_h=L_{21}X_1+L_{22}X_2,\label{eq.onsager_J2_he}
\end{eqnarray}
where the coefficients $L_{ij}$'s are the Onsager coefficients satisfying the reciprocal relation $L_{12}=L_{21}$.
In the LIT model,
the entropy production rate $\dot{\sigma}=J_1X_1+J_2X_2$ becomes a quadratic form in $X_i$'s and its non-negativity leads to the following restriction on the Onsager coefficients $L_{ij}$'s as
\begin{eqnarray}
L_{11}\ge 0, \ L_{22}\ge 0, \ L_{11}L_{22}-L_{12}^2\ge0.\label{eq.restrict_onsager_coeffi_he}
\end{eqnarray}
By changing the variable from $X_1$ to $J_1$ by using Eq.~(\ref{eq.onsager_J1_he}), we can write
the Onsager relation Eq.~(\ref{eq.onsager_J2_he}) and $\dot{Q}_c=J_2-P\equiv J_3$ by using $J_1$ as
\begin{eqnarray}
&&J_2=\frac{L_{21}}{L_{11}}J_1+L_{22}(1-q^2)X_2,\label{eq.lin_J2_by_J1_he}\\
&&J_3=\frac{L_{21}T_c}{L_{11}T_h}J_1+L_{22}(1-q^2)X_2+\frac{T_c}{L_{11}} {J_1}^2, \label{eq.lin_J3_by_J1_he}
\end{eqnarray}
respectively, where $q$ is the coupling strength parameter defined as
\begin{eqnarray}
q\equiv \frac{L_{12}}{\sqrt{L_{11}L_{22}}} \ \ (|q|\le 1).\label{eq.q_he}
\end{eqnarray}
From Eqs.~(\ref{eq.lin_J2_by_J1_he}) and (\ref{eq.lin_J3_by_J1_he}), we immediately notice that the nonlinear term $\frac{T_c}{L_{11}}J_1^2$ appears only in $J_3$ in an ``asymmetric" way.
By using Eqs.~(\ref{eq.lin_J2_by_J1_he}) and (\ref{eq.lin_J3_by_J1_he}), the entropy production rate $\dot{\sigma}=-\frac{J_2}{T_h}+\frac{J_3}{T_c}$ is given by
\begin{eqnarray}
\dot{\sigma}=L_{22}(1-q^2)X_2^2+\frac{J_1^2}{L_{11}},
\end{eqnarray}
where it turns out that the nonlinear term expresses the dissipation effect contributing to the entropy production rate.
Our minimally nonlinear irreversible heat engine assumes the following extended Onsager relations
such that the dissipation terms in both sides of the heat fluxes are equally taken into account~\cite{yuki12a} [see Fig.~\ref{heat_device} (a)]:
\begin{eqnarray}
&&J_1=\dot{x}=L_{11}X_1+L_{12}X_2,\label{eq.exonsager_J1_he}\\
&&J_2=\dot{Q}_h=L_{21}X_1+L_{22}X_2-\gamma_h J_1^2.\label{eq.exonsager_J2_he}
\end{eqnarray}
The nonlinear term in $J_2$ expresses the dissipation into the hot heat reservoir
caused by the finite-time motion of the heat engine $J_1\ne 0$ and $\gamma_h$ ($>0$) is a constant meaning the strength of the dissipation.
This choice of the specific form is motivated by the low-dissipation Carnot cycle model, which is proved to be equivalent with
Eqs.~(\ref{eq.exonsager_J1_he}) and (\ref{eq.exonsager_J2_he}) under the tight-coupling condition~\cite{yuki12a}
(see Appendix A). In the absence of the nonlinear term, Eqs.~(\ref{eq.exonsager_J1_he}) and (\ref{eq.exonsager_J2_he}) recover the usual linear Onsager relations Eqs.~(\ref{eq.onsager_J1_he}) and (\ref{eq.onsager_J2_he}).
Although our extended Onsager relation Eq.~(\ref{eq.exonsager_J2_he}) includes the additional nonlinear term,
we assume that the Onsager reciprocity $L_{12}=L_{21}$ and the restriction Eq.~(\ref{eq.restrict_onsager_coeffi_he}) still holds for our $L_{ij}$'s (Appendix A).
By changing the variable from $X_1$ to $J_1$ by using Eq.~(\ref{eq.exonsager_J1_he}), we can write
the extended Onsager relation Eq.~(\ref{eq.exonsager_J2_he}) and $\dot{Q}_c=J_2-P=J_3$ by using $J_1$ as
\begin{eqnarray}
&&J_2=\frac{L_{21}}{L_{11}}J_1+L_{22}(1-q^2)X_2-\gamma_h {J_1}^2,\label{eq.J2_by_J1_he}\\
&&J_3=\frac{L_{21}T_c}{L_{11}T_h}J_1+L_{22}(1-q^2)X_2+\gamma_c {J_1}^2,
\label{eq.J3_by_J1_he}
\end{eqnarray}
respectively, where we define a constant meaning the strength of the dissipation into the cold heat reservoir $\gamma_c$ as
\begin{eqnarray}
\gamma_c \equiv \frac{T_c}{L_{11}}-\gamma_h >0,
\end{eqnarray}
assuming its positivity.
The meaning of each term in Eqs.~(\ref{eq.J2_by_J1_he}) and (\ref{eq.J3_by_J1_he}) can be highlighted by
expressing the entropy production rate $\dot{\sigma}=-\frac{J_2}{T_h}+\frac{J_3}{T_c}$ by using them~\cite{yuki12a}:
\begin{eqnarray}
\dot{\sigma}=L_{22}(1-q^2)X_2^2+\frac{\gamma_h}{T_h}J_1^2+\frac{\gamma_c}{T_c}J_1^2,\label{eq.entropy_production_he}
\end{eqnarray}
where it is always assured to be non-negative from Eqs.~(\ref{eq.restrict_onsager_coeffi_he}) and (\ref{eq.q_he}), and the non-negativity of $\gamma_h>0$ and $\gamma_c>0$.
From Eq.~(\ref{eq.entropy_production_he}), we find that the first terms in Eqs.~(\ref{eq.J2_by_J1_he}) and (\ref{eq.J3_by_J1_he})
mean the reversible heat transports that do not contribute to the entropy production rate.
The second terms mean the steady heat leaks from the hot heat reservoir to the cold heat reservoir, which vanish under $|q|=1$ called the tight coupling.
This tight-coupling condition in the MNLIT model assures that the heat fluxes $J_2$ and $J_3$ vanish simultaneously in the quasistatic limit $J_1\to 0$, playing a similar role in the LIT model.
The third terms mean the dissipations into the heat reservoirs due to the finite-time operation of the heat engine.
The power output $P=F\dot{x}=-J_1X_1T_c$ is also expressed by using $J_1$ as
\begin{eqnarray}
P=\frac{L_{12}}{L_{11}}\eta_{\rm C}J_1-\frac{T_c}{L_{11}}J_1^2.\label{eq.P_he}
\end{eqnarray}
Instead of the extended Onsager relations Eqs.~(\ref{eq.exonsager_J1_he}) and (\ref{eq.exonsager_J2_he}),
we can describe the heat engine by using Eqs.~(\ref{eq.J2_by_J1_he}) and (\ref{eq.J3_by_J1_he}).
Under given Onsager coefficients $L_{ij}$'s and the thermodynamic force $X_2$,
the working regime of the heat engines depends on $J_1$.
For the requirement of the positive power $P>0$, $J_1$ should be located in the following range:
\begin{eqnarray}
\begin{cases}
0<J_1< L_{12}X_2 & (0<L_{12}), \\
L_{12}X_2 <J_1<0 & (L_{12}<0). \label{eq.J1_range_he}
\end{cases}
\end{eqnarray}
The efficiency $\eta$ of the heat engine in our minimally nonlinear irreversible model is expressed as
\begin{eqnarray}
\eta=\frac{P}{J_2}=\frac{\frac{L_{12}}{L_{11}}\eta_{\rm C}J_1-\frac{T_c}{L_{11}}J_1^2}{\frac{L_{21}}{L_{11}}J_1+L_{22}(1-q^2)X_2-\gamma_h {J_1}^2},\label{eq.def_eta}
\end{eqnarray}
by using Eqs.~(\ref{eq.J2_by_J1_he}) and (\ref{eq.P_he}).
\subsection{Minimally nonlinear irreversible model for RE}\label{s2b}
As well as in the case of HE,
we start from the entropy production rate $\dot{\sigma}$ of the total system (the refrigerator and the heat reservoirs).
Because the internal state of the refrigerator comes back to the original state after one cycle for cyclic refrigerators
or it remains unchanged for steady-state refrigerators,
$\dot{\sigma}$ can be written by the sum of the entropy-change rate of the heat reservoirs as
$\dot{\sigma}=\dot{S}_h+\dot{S}_c$ according to the local equilibrium hypothesis~\cite{callen}, where $S_i$ ($i=h, c$) denotes the equilibrium entropy of the heat reservoir.
It can be written as
\begin{eqnarray}
\dot{\sigma}=\dot{S}_h+\dot{S}_c=\dot{U}_h\frac{\partial S_h}{\partial U_h}+\dot{U}_c\frac{\partial S_c}{\partial U_c}=\frac{\dot{Q}_h}{T_h}-\frac{\dot{Q}_c}{T_c},\label{eq.sigma_dot_re_1}
\end{eqnarray}
where we used $\frac{\partial S_i}{\partial U_i}=\frac{1}{T_i}$ with $U_i$ being the equilibrium internal-energy of the heat reservoir, and
for RE, we denote by $\dot{Q}_c\equiv -\dot{U}_c$ the heat flux from the cold heat reservoir and by $\dot{Q}_h\equiv \dot{U}_h$ the heat flux into the hot heat reservoir, respectively, as is opposite to HE.
We also denote by $\dot{W}\equiv P$ the power injection.
Then, from $\dot{Q}_h=P+\dot{Q}_c=F\dot{x}+\dot{Q}_c$
with $F$ and $x$ a generalized external force and its conjugate variable, respectively,
Eq.~(\ref{eq.sigma_dot_re_1}) can be rewritten as
\begin{eqnarray}
\dot{\sigma}=\frac{F\dot{x}}{T_h}+\dot{Q}_c\left(\frac{1}{T_h}-\frac{1}{T_c}\right)\equiv J_1X_1+J_2X_2.
\end{eqnarray}
It naturally leads us to define the thermodynamic flux $J_1\equiv \dot{x}$
(the motion speed of the refrigerator)
conjugate to the thermodynamic force $X_1 \equiv F/T_h$, and the other thermodynamic flux $J_2\equiv \dot{Q}_c$ (the heat flux from the cold heat reservoir) conjugate to the other thermodynamic force $X_2\equiv 1/T_h-1/T_c$,
where these quantities are expressed by using the thermodynamic extensive and intensive parameters of the equilibrium heat reservoirs~\cite{callen}.
To establish our election of the thermodynamic fluxes and forces for refrigerators,
we write the entropy production rate $\dot{\sigma}$ of the refrigerator as a function of $\dot Q_c$ and $\dot W$, thus incorporating in the formalism in a natural way the specific job of the refrigerator system (the extracted cooling power of the low-temperature reservoir $\dot Q_c$ as a consequence of the input of an external power $\dot W$).
While it is more intuitive from a thermodynamic point of view that $\dot{\sigma}$ is written in terms of the specific job of each thermodynamic device in this way, an alternative starting point may be to express $\dot{\sigma}$ of the refrigerator in terms of $\dot W$ and $\dot Q_h$, as for heat engines.
If this is done the thermodynamic fluxes and forces are exactly the same as those obtained for a heat engine, but it does not change the main results.
Then, as is similar to the heat engines in Sec.~II A, our minimally nonlinear irreversible refrigerator assumes the following extended Onsager relations between the thermodynamic fluxes $J_i$'s and forces $X_i$'s~\cite{yuki13a} [see Fig.~\ref{heat_device} (b)]:
\begin{eqnarray}
&&J_1=\dot{x}=L_{11}X_1+L_{12}X_2,\label{eq.exonsager_J1_re}\\
&&J_2=\dot{Q}_c=L_{21}X_1+L_{22}X_2-\gamma_c J_1^2,\label{eq.exonsager_J2_re}
\end{eqnarray}
where $L_{ij}$'s are the Onsager coefficients satisfying the reciprocal relation $L_{12}=L_{21}$.
The nonlinear term in $J_2$ expresses the dissipation into the cold heat reservoir caused by the finite-time motion of the refrigerator
and $\gamma_c$ ($>0$) is a constant meaning the strength of the dissipation.
In the absence of the nonlinear term, Eqs.~(\ref{eq.exonsager_J1_re}) and (\ref{eq.exonsager_J2_re}) recover the usual linear Onsager relations in LIT~\cite{vbroeck05}.
In LIT, the entropy production rate $\dot{\sigma}=J_1X_1+J_2X_2$ becomes the quadratic form in $X_i$'s and its non-negativity leads to the following restriction on the Onsager coefficients $L_{ij}$'s as
\begin{eqnarray}
L_{11}\ge 0, \ L_{22}\ge 0, \ L_{11}L_{22}-L_{12}^2\ge0.\label{eq.restrict_onsager_coeffi_re}
\end{eqnarray}
Although our extended Onsager relation Eq.~(\ref{eq.exonsager_J2_re}) includes the additional nonlinear term,
we assume that the restriction Eq.~(\ref{eq.restrict_onsager_coeffi_re}) still holds for our $L_{ij}$'s.
By changing the variable from $X_1$ to $J_1$ by using Eq.~(\ref{eq.exonsager_J1_re}), we can write the extended Onsager relation Eq.~(\ref{eq.exonsager_J2_re}) and $\dot{Q}_h=J_2+P\equiv J_3$ by using $J_1$ as
\begin{eqnarray}
&&J_2=\frac{L_{21}}{L_{11}}J_1+L_{22}(1-q^2)X_2-\gamma_c
{J_1}^2,\label{eq.J2_by_J1_re}\\
&&J_3=\frac{L_{21}T_h}{L_{11}T_c}J_1+L_{22}(1-q^2)X_2+\gamma_h {J_1}^2,
\label{eq.J3_by_J1_re}
\end{eqnarray}
respectively, where we define a constant meaning the strength of the dissipation into the hot heat reservoir $\gamma_h$ as
\begin{eqnarray}
\gamma_h \equiv \frac{T_h}{L_{11}}-\gamma_c >0,\label{eq.restrict_gh_ref}
\end{eqnarray}
assuming its positivity and the coupling strength parameter $q$ as
\begin{eqnarray}
q\equiv \frac{L_{12}}{\sqrt{L_{11}L_{22}}}\ \ (|q|\le 1).\label{eq.q_re}
\end{eqnarray}
Meaning of each term in Eqs.~(\ref{eq.J2_by_J1_re}) and (\ref{eq.J3_by_J1_re}) can be highlighted by
expressing the entropy production rate $\dot{\sigma}=\frac{J_3}{T_h}-\frac{J_2}{T_c}$ by using them~\cite{yuki13a}:
\begin{eqnarray}
\dot{\sigma}=L_{22}(1-q^2)X_2^2+\frac{\gamma_h}{T_h}J_1^2+\frac{\gamma_c}{T_c}J_1^2,\label{eq.entropy_production_re}
\end{eqnarray}
where it is always assured to be non-negative from Eqs.~(\ref{eq.restrict_onsager_coeffi_re}) and (\ref{eq.q_re}), and the non-negativity of $\gamma_h>0$ and $\gamma_c>0$.
The expression of the entropy production rate Eq.~(\ref{eq.entropy_production_re}) for the refrigerator agrees with Eq.~(\ref{eq.entropy_production_he})
for the heat engines, presenting a unified description of heat devices in our MNLIT models.
From Eq.~(\ref{eq.entropy_production_re}), we find that the first terms in Eqs.~(\ref{eq.J2_by_J1_re}) and (\ref{eq.J3_by_J1_re}) mean the reversible heat transports that do not contribute to the entropy production rate.
The second terms mean the steady heat-leaks from the hot heat reservoir to the cold heat reservoir, which vanish under the tight-coupling condition $|q|=1$.
Under this tight-coupling condition in the MNLIT model, the heat fluxes $J_2$ and $J_3$ vanish simultaneously in the quasistatic limit $J_1\to 0$,
as it happens in the LIT model.
The third terms mean the dissipations into the heat reservoirs due to the finite-time operation of the refrigerator.
The power injection $P=F\dot{x}=J_1X_1T_h$ is also expressed by using $J_1$ as
\begin{eqnarray}
P=\frac{L_{12}}{L_{11}\varepsilon_{\mathrm{C}}}J_1+\frac{T_h}{L_{11}}J_1^2.\label{eq.P_re}
\end{eqnarray}
Instead of the extended Onsager relations Eqs.~(\ref{eq.exonsager_J1_re}) and (\ref{eq.exonsager_J2_re}),
we can describe the refrigerator by using Eqs.~(\ref{eq.J2_by_J1_re}) and (\ref{eq.J3_by_J1_re}).
Under given Onsager coefficients $L_{ij}$'s and the thermodynamic force $X_2$, the working regime of the refrigerators depends on $J_1$.
For the requirement of the positive cooling power $J_2>0$, $J_1$ should be located in the following range:
\begin{eqnarray}
\frac{L_{21}-\sqrt{D}}{2L_{11}\gamma_c}<J_1< \frac{L_{21}+\sqrt{D}}{2L_{11}\gamma_c}, \label{eq.J1_range_re}
\end{eqnarray}
where the discriminant $D$, which should be positive, is given by
\begin{eqnarray}
D\equiv L_{21}^2+4L_{11}^2L_{22}\gamma_c (1-q^2)X_2>0.\label{eq.discriminant}
\end{eqnarray}
Under the tight-coupling condition $|q|=1$, Eq.~(\ref{eq.J1_range_re}) reduces to
\begin{eqnarray}
\begin{cases}
0< J_1 <\frac{L_{21}}{\gamma_c L_{11}} & (L_{12}>0),\\
\frac{L_{21}}{\gamma_c L_{11}} < J_1 <0 & (L_{12}<0),
\end{cases}
\end{eqnarray}
showing that the quasistatic limit $J_1\to 0$ is included.
In contrast, under the nontight-coupling condition $|q|<1$, such quasistatic limit is not included and $J_1$ must be a finite value as in Eq.~(\ref{eq.J1_range_re}) for $J_2$ to be positive.
Intuitively, this constraint is necessary for the cooling effect to overcome the steady heat-leak effect.
In addition, from Eq.~(\ref{eq.discriminant}), we also have a constraint on $\gamma_c$ under the non-tight-coupling condition $|q|<1$ as
\begin{eqnarray}
\gamma_c < -\frac{q^2}{4(1-q^2)L_{11}X_2}\equiv \gamma_c^+.\label{eq.gcplus}
\end{eqnarray}
This constraint is also related to the positivity of the cooling power under $|q|<1$:
even when the refrigerator operates at a finite rate,
large enough dissipation into the cold heat reservoir can also violate the positive cooling power.
Combining Eq.~(\ref{eq.gcplus}) with Eq.~(\ref{eq.restrict_gh_ref}), we obtain the following constraint on $\gamma_c$ depending on the parameter values as
\begin{eqnarray}
\begin{cases}
0< \gamma_c <\frac{T_h}{L_{11}} & \left(\gamma_c^+ \ge \frac{T_h}{L_{11}}, {\rm i.e.,} \ \tau \ge \frac{1}{\frac{q^2}{4(1-q^2)}+1}\right), \\
0<\gamma_c<\gamma_c^+ & \left(\gamma_c^+ < \frac{T_h}{L_{11}}, {\rm i.e.,} \ \tau < \frac{1}{\frac{q^2}{4(1-q^2)}+1}\right).\label{eq.gc_range}
\end{cases}
\end{eqnarray}
As is clear from Eq.~(\ref{eq.J1_range_he}), such restriction does not exist in the heat engines. However it plays a key role in the behavior of the optimum performance regimes of RE (see Sec. IV. B).
The COP $\varepsilon$ of the refrigerator in our minimally nonlinear irreversible model is expressed as
\begin{eqnarray}
\varepsilon=\frac{J_2}{P}=\frac{\frac{L_{21}}{L_{11}}J_1+L_{22}(1-q^2)X_2-\gamma_c
{J_1}^2}{\frac{L_{12}}{L_{11}\varepsilon_{\mathrm{C}}}J_1+\frac{T_h}{L_{11}}J_1^2},\label{eq.def_cop}
\end{eqnarray}
by using Eqs.~(\ref{eq.J2_by_J1_re}) and (\ref{eq.P_re}).
\section{Global performance characteristics}
We consider global performance characteristics of the minimally nonlinear irreversible heat devices
described by Eqs.~(\ref{eq.J2_by_J1_he}) and (\ref{eq.J3_by_J1_he})
for HE or Eqs.~(\ref{eq.J2_by_J1_re}) and (\ref{eq.J3_by_J1_re}) for RE.
\subsection{HE: Performance characteristics}
\begin{figure*}
\includegraphics[scale=0.85]{global_performance_regime_heat_engine_latest.eps}
\caption{Power-efficiency ($P$--$\eta$) curve for the minimally nonlinear irreversible heat engine under various coupling strengths (solid line for $q=1$, dashed line for $q=0.95$, and dotted line for $q=0.85$):
(a) asymmetric dissipation ($\gamma_h=0.001$ and $\gamma_c=0.699$), (b) symmetric dissipation ($\gamma_h=\gamma_c=0.35$), and (c) asymmetric dissipation ($\gamma_h=0.699$ and $\gamma_c=0.001$).
We used $L_{11}=L_{22}=1$, $T_h=1$, and $T_c=0.7$.
The Carnot efficiency is $\eta_{\rm C}=0.3$.}\label{gpc_p_eta_g}
\end{figure*}
\begin{figure*}
\includegraphics[scale=0.85]{global_performance_regime_heat_engine_ver2_latest.eps}
\caption{Power-efficiency ($P$--$\eta$) curve for the minimally nonlinear irreversible heat engine under various dissipation regimes (solid line for $\gamma_h=0.001$, dashed line for $\gamma_h=0.35$, and dotted line for $\gamma_h=0.699$):
(a) $q=1$, (b) $q=0.95$, and (c) $q=0.85$.
We used $L_{11}=L_{22}=1$, $T_h=1$, and $T_c=0.7$.
The Carnot efficiency is $\eta_{\rm C}=0.3$.}\label{gpc_p_eta_q}
\end{figure*}
In our model, $|q|=1$ means a perfect, tight-coupling condition for the internal degrees of freedom, so that the heat-leak terms proportional to the thermodynamic force $X_2$ in Eqs.~(\ref{eq.J2_by_J1_he}) and (\ref{eq.J3_by_J1_he}) do not play any role. Thus, the
thermodynamic fluxes $J_{2}$ and $J_{3}$ depend only on the thermodynamic flux $J_{1\,}$ and the dissipation constants $\gamma_i$'s.
The resulting power-efficiency plots for HE [Figs.~\ref{gpc_p_eta_g}(a)--\ref{gpc_p_eta_g}(c) for $|q|=1$] are parabolic shaped defined between the two null-power states corresponding to a stalled state~\cite{vbroeck05,seifert08}
where the power vanishes due to too quick operation of the heat engine and the Carnot efficiency state realized in the quasistatic limit $J_1\to 0$, independently of the dissipation constants.
These figures exhibit the characteristics common to general Carnot-like models when the heat leak is absent, and the irreversibilities are thus limited to external coupling between the working system and external heat reservoirs with adequate heat transfer laws (endoreversible limit~\cite{lchen99,wu04,jchen01,luisarias13,gordon00}). Only if additionally the heat transfer law is linear, the CA efficiency emerges in FTT~\cite{curzon75}.
However, this is not the case in our model, where this particular value is realized only as a limiting case. Later we will come back to this particular point in Sec.~IV.
For $|q|<1$, the thermodynamic fluxes $J_2$ and $J_3$
also depend on an additional direct heat transfer between the hot and cold heat reservoirs, which is proportional to $X_2$ as in Eqs.~(\ref{eq.J2_by_J1_he}) and (\ref{eq.J3_by_J1_he}). Now, the resulting power-efficiency ($P$--$\eta$) plots are loop shaped (as those obtained in the irreversible Carnot-like FTT models~\cite{jchen01}) with near but non-coincident maximum power and maximum efficiency points. As the degree of the heat leak increases (i.e., as $|q|$ decreases) [Figs.~\ref{gpc_p_eta_g}(a)--\ref{gpc_p_eta_g}(c)], the distance between maximum power and maximum efficiency points becomes smaller determining more closed loops. This loop-shaped behavior is explained as follows: when $|q|< 1$,
the heat-leak effect steadily remains even in the quasistatic limit $J_1 \to 0$, which implies $\eta \to 0$ from Eq.~(\ref{eq.def_eta}).
In contrast, even when the magnitude of $J_1$ becomes too large,
the efficiency (as well as the power) becomes $0$ at the stalled state $J_1=L_{12}X_2$~\cite{vbroeck05,seifert08}.
Therefore the optimum $J_1$ that maximizes $\eta$ must exist somewhere between these extreme points.
The explicit influence of the dissipation constants $\gamma_i$'s on the global performance characteristics is better visualized from Figs.~\ref{gpc_p_eta_q}(a)--\ref{gpc_p_eta_q}(c). It is clearly observed that at any fixed power (including the maximum power point) and any $q$ value, the efficiency increases as $\gamma_{h}$ increases. This result generalizes the result reported by Apertet {\it et al.}~\cite{apertet12a,apertet12b} for a tightly coupled thermoelectric generator to any coupling case. As these authors of Refs.~\cite{apertet12a,apertet12b} argue, the heat released at the hot heat reservoir can eventually be recycled by the hot isothermal step, and thus a preferential dissipation into this side provokes increase of the efficiency at any fixed power.
Actually, this mechanism works for any $q$ value: we can obtain the $\eta=\eta(P)$ curve explicitly by combining Eqs.~(\ref{eq.P_he}) and (\ref{eq.def_eta}) as
\begin{eqnarray}
\eta(P)=\frac{P}{\frac{q^2L_{22}X_2C_h}{2}+L_{22}(1-q^2)X_2-\gamma_h \frac{L_{12}^2X_2^2C_h^2}{4}},\label{eq.eta_p_curve}
\end{eqnarray}
where $C_h$ is defined as
\begin{eqnarray}
C_h \equiv 1\pm \sqrt{1-\frac{4PT_c}{q^2L_{22}\eta_{\rm C}^2}},
\end{eqnarray}
and the sign $+$ ($-$) corresponds to a branch for the working regimes from the stalled-state point (quasistatic limit) to the maximum power point.
From the curve Eq.~(\ref{eq.eta_p_curve}), it is obvious that the efficiency increases as $\gamma_h$ increases at any fixed power and for any $q$ value.
\subsection{RE: Performance characteristics}
\begin{figure*}
\includegraphics[scale=0.72]{global_performance_regime_refrigerator_latest.eps}
\caption{Cooling power-COP ($J_2$--$\varepsilon$) curve for the minimally nonlinear irreversible refrigerator under various
coupling strengths (solid line for $q=1$, dashed line for $q=0.95$, and dotted line for $q=0.85$): (a) asymmetric dissipation ($\gamma_h=0.001$ and $\gamma_c=0.999$), (b) symmetric dissipation ($\gamma_h=\gamma_c=0.5$), and (c) asymmetric dissipation ($\gamma_h=0.999$ and $\gamma_c=0.001$). (d) shows a closer inspection of (c) in the small $\varepsilon$-range for better understanding of the shape of $J_2$--$\varepsilon$ curve (open curve for $q=1$ and closed curve for $q=0.95, 0.85$). Diverging behavior of $J_2$ in the limit of $\gamma_c \to 0$ is confirmed.
We used $L_{11}=L_{22}=1$, $T_h=1$, and $T_c=0.7$.
The Carnot COP is $\varepsilon_{\rm C}\simeq 2.33$.}
\label{gpc_j2_cop_g}
\end{figure*}
\begin{figure*}
\includegraphics[scale=0.85]{global_performance_regime_refrigerator_ver2_latest.eps}
\caption{Cooling power-COP ($J_2$--$\varepsilon$) curve for the minimally nonlinear irreversible refrigerator under various dissipation regimes (solid line for $\gamma_c=0.001$, dashed line for $\gamma_c=0.5$,
and dotted line for $\gamma_c=0.999$):
(a) $q=1$, (b) $q=0.95$, and (c) $q=0.85$.
We used $L_{11}=L_{22}=1$, $T_h=1$, and $T_c=0.7$.
The Carnot COP is $\varepsilon_{\rm C}\simeq 2.33$.}\label{gpc_j2_cop_q}
\end{figure*}
Following the same methodology as for HE above, we will analyze the
cooling power-COP ($J_2$--$\varepsilon$) plots. See Figs.~\ref{gpc_j2_cop_g}(a)--\ref{gpc_j2_cop_g}(d).
Again, we clearly observe open curves under the tight-coupling condition $|q|=1$, which are similar to the characteristics of the endoreversible models~\cite{lchen99,wu04,jchen01,luisarias13,gordon00}, where the maximum cooling power is realized at a finite rate while the maximum COP is realized at the zero cooling power (quasistatic limit).
In contrast, under the non-tight-coupling condition $|q|<1$, the maximum COP is no longer realized at the zero cooling power but at a finite cooling power because in this case the COP at the zero cooling power (quasistatic limit) becomes zero due to a steady heat leak.
Then we observe loop-shaped curves with near but non-coincident optimum values for both of the cooling power and COP.
As $|q|$ progressively decreases, this tendency becomes prominent and
both of the cooling power and COP become smaller and their behaviors are also strongly
modulated by the values of the dissipation constants $\gamma_i$'s.
This loop-shaped behavior is explained as in the HE case above: when $|q|< 1$, $J_1$ takes values in a bounded interval as in Eq.~(\ref{eq.J1_range_re}) for the cooling power $J_2$ to be positive.
At both ends in Eq.~(\ref{eq.J1_range_re}), $\varepsilon$ becomes $0$ because $J_2$ vanishes there [see Eq.~(\ref{eq.def_cop})].
Therefore the optimum $J_1$ that maximizes $\varepsilon$ must exist somewhere between these extreme points. We also see that, in the limit of $\gamma_c \to 0$, the maximum cooling power shows diverging behavior for all $q$'s,
which has previously been pointed out by Apertet {\it et al.} for the tight-coupling case $|q|=1$ in the thermoelectric generator~\cite{apertet13a}.
Similar behavior has also been reported for the cooling power at the maximum $\chi$ condition for minimally nonlinear irreversible refrigerators~\cite{yuki13a}.
The explicit influence of the dissipation constants $\gamma_i$'s on the global performance characteristics can be analyzed more clearly from Figs.~\ref{gpc_j2_cop_q}(a)--\ref{gpc_j2_cop_q}(c).
For $|q|=1$ in Fig.~\ref{gpc_j2_cop_q}(a), a preferential dissipation into the cold heat reservoir ($\gamma_c\rightarrow \frac{T_h}{L_{11}}$) provokes a decreasing of the cooling power at any fixed COP, but the COP at maximum cooling power monotonically increases.
This counterintuitive behavior was also reported in Ref.~\cite{apertet13a} for the tight-coupling case $|q|=1$ in the thermoelectric generator.
For more realistic situations with $|q|<1$ as in Figs.~\ref{gpc_j2_cop_q}(b) and \ref{gpc_j2_cop_q}(c),
we can see that the preferential dissipation into the cold heat reservoir generally induces the smaller loop.
Interestingly, we can also see that the COP at maximum cooling power can show a non-monotonic behavior with respect to the strength of the dissipation as in Fig.~\ref{gpc_j2_cop_q}(c) ($q=0.85$) in contrast to the monotonic behavior as in \ref{gpc_j2_cop_q}(a) ($|q|=1$) and \ref{gpc_j2_cop_q}(b) ($q=0.95$).
This behavior of the COP at maximum cooling power will be discussed in detail in Sec.~IV B.
\section{Optimized regimes}
We analyze in detail the optimized performance regimes of the maximum efficiency and the efficiency at maximum power for HE and those of the maximum COP and the COP at maximum cooling power for RE found in the analysis of the global performance regimes in Sec.~III.
\subsection{Optimized regimes: HE}
\begin{figure*}
\includegraphics[scale=1.0]{upper_lower_bounds_for_eta_max_latest.eps}
\caption{Normalized upper bounds $\eta_{\rm max}^+(q)/\eta_{\rm C}$ in Eq.~(\ref{eq.etamaxup}) and $\eta_{\rm maxP}^+(q)/\eta_{\rm C}$ in Eq.~(\ref{eq.eta_maxpow_upper}), and lower bounds $\eta_{\rm max}^-(q)/\eta_{\rm C}$ in Eq.~(\ref{eq.etamaxlow}) and $\eta_{\rm maxP}^-(q)/\eta_{\rm C}$ in Eq.~(\ref{eq.eta_maxpow_lower}) as a function of $\eta_{\rm C}$: (a) $q=1$, (b) $q=0.95$, and (c) $q=0.85$. $\eta_{\rm max}^+(1)=\eta_{\rm max}^-(1)=\eta_{\rm C}$
[see Eqs.~(\ref{eq.etamaxlow}) and (\ref{eq.etamaxup})].
}\label{lower_and_upper_bound_eta_max}
\end{figure*}
By using the definition of $\eta$ in Eq.~(\ref{eq.def_eta}) and solving $\partial \eta/\partial J_1=0$, we obtain
the maximum efficiency $\eta_{\rm max}(q, \gamma_h/\gamma_c)$ explicitly as:
\begin{widetext}
\begin{eqnarray}
\eta_{\rm max}\left(q, \frac{\gamma_h}{\gamma_c} \right)=\frac{\eta_{\rm C}+\frac{\left(1+\frac{\gamma_c}{\gamma_h}\right)(1-q^2)\eta_{\rm C}}{q^2\left(1-\eta_{\rm C}+\frac{\gamma_c}{\gamma_h}\right)}\left(1-\sqrt{1+\frac{q^2\left(1-\eta_{\rm C}+\frac{\gamma_c}{\gamma_h}\right)}{(1-q^2)\left(1+\frac{\gamma_c}{\gamma_h}\right)}}\right)}{1-\frac{1-\eta_{\rm C}+\frac{\gamma_c}{\gamma_h}}{1+\frac{\gamma_c}{\gamma_h}}\frac{1}{\left(1-\sqrt{1+\frac{q^2\left(1-\eta_{\rm C}+\frac{\gamma_c}{\gamma_h}\right)}{(1-q^2)\left(1+\frac{\gamma_c}{\gamma_h}\right)}}\right)}+\frac{(1-q^2)\eta_{\rm C}}{q^2\left(1+\frac{\gamma_c}{\gamma_h}\right)}\left(1-\sqrt{1+\frac{q^2\left(1-\eta_{\rm C}+\frac{\gamma_c}{\gamma_h}\right)}{(1-q^2)\left(1+\frac{\gamma_c}{\gamma_h}\right)}}\right)}.\label{eq.eta_max_nontight}
\end{eqnarray}
\end{widetext}
This is a monotonically increasing function of $|q|$ and $\gamma_h/\gamma_c$~\cite{monotonicity}.
In the limit of the small temperature difference $\Delta T\to 0$, Eq.~(\ref{eq.eta_max_nontight}) can be expanded as
\begin{eqnarray}
\eta_{\rm max}\left(q, \frac{\gamma_h}{\gamma_c} \right)=\frac{\left(1-\sqrt{1-q^2}\right)^2}{q^2}\frac{\Delta T}{T}+O({\Delta T}^2),
\end{eqnarray}
where $T\equiv (T_h+T_c)/2$. Up to the first order of $\Delta T$, we have no $\gamma_h/\gamma_c$ dependence.
This expression has been previously obtained in the framework of LIT~\cite{cisneros08}.
In asymmetric dissipation limits $\gamma_h/\gamma_c \to 0$ and $\gamma_h/\gamma_c \to \infty$,
we find that Eq.~(\ref{eq.eta_max_nontight}) is bounded from the lower side by $\eta_{\rm max}^-(q)$ and the upper side by $\eta_{\rm max}^+(q)$,
which are given as
\begin{eqnarray}
\eta_{\rm max}^-(q) &&\equiv \eta_{\rm max}\left(q,\frac{\gamma_h}{\gamma_c} \rightarrow 0\right)\nonumber\\
&&=\frac{\left(1-\sqrt{1-q^2}\right)^2}{q^2}\eta_{\rm C},\label{eq.etamaxlow}
\end{eqnarray}
and
\begin{eqnarray}
&&\eta_{\rm max}^+(q)\equiv \eta_{\rm max}\left(q,\frac{\gamma_h}{\gamma_c} \rightarrow \infty \right)=\nonumber\\
&&\frac{\sqrt{\frac{1-q^2\eta_{\rm C}}{1-q^2}}\left(q^2\eta_{\rm C}^2+(q^2-2)\eta_{\rm C}\right)+2\eta_{\rm C}(1-q^2\eta_{\rm C})} {\sqrt{\frac{1-q^2\eta_{\rm C}}{1-q^2}}\left((3q^2-2)\eta_{\rm C}-q^2\right)+2\eta_{\rm C}(1-q^2\eta_{\rm C})},\label{eq.etamaxup}
\end{eqnarray}
respectively.
In the limit of $|q|\to 1$, $\eta^{\pm}_{\rm max}(q) \to \eta_{\rm C}$ as expected,
but as the coupling strength decreases, these values drastically decrease showing the behavior plotted in Figs.~\ref{lower_and_upper_bound_eta_max}(b) and \ref{lower_and_upper_bound_eta_max}(c).
In the case of the symmetric dissipation $\gamma_h=\gamma_c$, we obtain
\begin{eqnarray}
\eta_{\rm max}^{\rm sym}(q)&&\equiv \eta_{\rm max}\left(q,\frac{\gamma_h}{\gamma_c}=1\right)\\\nonumber
&&=\frac{\eta_{\rm C}+\frac{2(1-q^2)\eta_{\rm C}}{q^2(2-\eta_{\rm C})}H_1}{1-\frac{2-\eta_{\rm C}}{2}\frac{1}{H_1}+\frac{(1-q^2)\eta_{\rm C}}{2q^2}H_1},
\end{eqnarray}
where $H_1$ is defined as
\begin{eqnarray}
H_1\equiv 1-\sqrt{1+\frac{q^2(2-\eta_{\rm C})}{2(1-q^2)}}.
\end{eqnarray}
Among different optimization regimes, the efficiency at maximum power $\eta_{\rm maxP}$ has been playing an important role
for studies of traditional~\cite{esposito10a, apertet12a, apertet12b,tu12abc, jwang12a,bizarro,ito,jchen13a,broeck13a,norma10a}, stochastic~\cite{bekele04,tu08,seifert08,gaveau10,tu12a,seifert12,broeck12a,he13a}, and quantum~\cite{esposito09a,jwang12b,hyan12a,bekele12,rwang13a,jchen13b,allahv13} HE.
The maximum power and the efficiency at the maximum power $\eta_{\rm maxP}$ of the present model, which were studied in Ref.~\cite{yuki12a} previously, are obtained by solving $\partial P/\partial J_1=0$:
\begin{eqnarray}
P_{\rm max}&&=\frac{q^2L_{22}\eta_{\rm C}^2}{4T_c},\\
\eta_{\rm maxP}\left(q, \frac{\gamma_h}{\gamma_c}\right)&&=
\frac{\eta_{\rm C}}{2} \frac{q^2}{2-q^2\left(1+\frac{\eta_{\rm C}}{2\left(1+\frac{\gamma_c}{\gamma_h}\right)}\right)},\label{remchig}
\end{eqnarray}
where we used the definition of $P$ in Eq.~(\ref{eq.P_he}) and $\eta$ in Eq.~(\ref{eq.def_eta}).
Equation.~(\ref{remchig}) is a monotonically increasing function of $|q|$ and $\gamma_h/\gamma_c$~\cite{yuki12a}.
The corresponding lower and upper bounds, and symmetric case are easily obtained in the limit of
$\gamma_h/\gamma_c \to 0$,
$\gamma_h/\gamma_c \to \infty$, and $\gamma_c/\gamma_h=1$, respectively:
\begin{eqnarray}
\eta^-_{\rm maxP}(q)\equiv \eta_{\rm maxP}\left(q,\frac{\gamma_h}{\gamma_c}\rightarrow 0\right)=\frac{\eta_{\rm C}}{2}\frac{q^2}{2-q^2},\label{eq.eta_maxpow_lower}
\end{eqnarray}
\begin{eqnarray}
\eta^{\rm sym}_{\rm maxP}(q)&&\equiv \eta_{\rm maxP}\left(q,\frac{\gamma_h}{\gamma_c}=1\right)\nonumber\\
&&=\frac{\eta_{\rm C}}{2}\frac{q^2}{2-q^2(1+\frac{\eta_{\rm C}}{4})},
\end{eqnarray}
\begin{eqnarray}
\eta^+_{\rm maxP}(q)&&\equiv \eta_{\rm maxP}\left(q,\frac{\gamma_h}{\gamma_c}\rightarrow \infty \right)\nonumber\\
&&=\frac{\eta_{\rm C}}{2}\frac{q^2}{2-q^2(1+\frac{\eta_{\rm C}}{2})}.\label{eq.eta_maxpow_upper}
\end{eqnarray}
The bounds $\eta^-_{\rm maxP}(q)$ and $\eta^+_{\rm maxP}(q)$ are also plotted in Figs.~\ref{lower_and_upper_bound_eta_max}(a)--\ref{lower_and_upper_bound_eta_max}(c)
for the sake of comparison with the lower and upper bounds $\eta^-_{\rm max}(q)$ and $\eta^+_{\rm max}(q)$ of the maximum efficiency.
Note that as the coupling strength progressively decreases due to the heat-leak increase [Fig.~\ref{lower_and_upper_bound_eta_max}(c)],
the maximum efficiency and efficiency at the maximum power regimes tend to collapse in a unique inefficient performance regime as we have seen, for instance, in Fig.~\ref{gpc_p_eta_q}.
The following limits of Eq.~(\ref{remchig}) under the tight-coupling condition $|q|=1$ are especially interesting:
\begin{eqnarray}
\eta_{\rm maxP}^{-}=\frac{\eta_{\rm C}}{2}\,\,\left(\frac{\gamma_h}{\gamma_c}\to 0, \ |q|=1 \right),\label{remchilb}
\end{eqnarray}
\begin{eqnarray}
\eta_{\rm maxP}^{\rm sym}=\frac{2\eta_{\rm C}}{4-\eta_{\rm C}}\,\,\left(\frac{\gamma_h}{\gamma_c}=1, \ |q|=1 \right),\label{remchisyma}
\end{eqnarray}
\begin{eqnarray}
\eta_{\rm maxP}^{+}=\frac{\eta_{\rm C}}{2-\eta_{\rm C}}\,\,\left(\frac{\gamma_h}{\gamma_c}\to \infty, \ |q|=1 \right),\label{remchiub}
\end{eqnarray}
where Eqs.~(\ref{remchilb}) and (\ref{remchiub}) are previously obtained in the LD models~\cite{esposito10a}, and Eq.~(\ref{remchisyma}) may be comparable to the previous result obtained for a heat engine model
under a left-right (spatially) symmetric condition in Ref.~\cite{esposito09a}.
We note that when $\eta_{\rm C}\rightarrow 0$ ($T_c \sim T_h$), $\eta_{\rm maxP}^{\rm sym}$ in Eq.~(\ref{remchisyma}) is expanded as
\begin{eqnarray}\label{remchisymb}
\eta_{\rm maxP}^{\rm sym}=\frac{2\eta_{\rm C}}{4-\eta_{\rm C}}\approx \frac{\eta_{\rm C}}{2}+\frac{\eta_{\rm C}^2}{8}+\frac{\eta_{\rm C}^3}{32}+\cdots,
\end{eqnarray}
which reproduces $\eta_{\rm CA}$~\cite{curzon75} up to the second order of $\eta_{\rm C}$:
\begin{equation}\label{remca}
\eta_{\rm CA}=1-\sqrt{1-\eta_{\rm C}}\approx \frac{\eta_{\rm C}}{2}+\frac{\eta_{\rm C}^2}{8}+\frac{\eta_{\rm C}^3}{16}+\cdots.
\end{equation}
Consequently, under the maximum power condition, our model optimized with respect to only one-parameter $J_1$ reproduces the lower and upper bounds of the low-dissipation HE model optimized with respect to two parameters in~\cite{esposito10a} and also reproduces the CA efficiency up to second order in the limit of the small temperature difference~\cite{esposito09a}.
\subsection {Optimized regimes: RE}
\begin{figure*}
\includegraphics[scale=1.0]{upper_lower_bounds_for_cop_max_latest.eps}
\caption{Normalized upper bounds $\varepsilon_{\rm max}^+(q)/\varepsilon_{\rm C}$ in Eq.~(\ref{eq.maxcop_upper}) and $\varepsilon_{\rm maxJ_2}^+(q)/\varepsilon_{\rm C}$ in Eq.~(\ref{eq.copmaxj2_upper}), and lower bound $\varepsilon_{\rm max}^-(q)/\varepsilon_{\rm C}$ in Eq.~(\ref{eq.maxcop_lower}) as a function of $\tau$:
(a) $q=1$, (b) $q=0.95$, and (c) $q=0.85$.
$\varepsilon_{\rm max}^+(1)/\varepsilon_{\rm C}=\varepsilon_{\rm max}^-(1)/\varepsilon_{\rm C}=1$ in all $\tau$-range (see Eqs.~(\ref{eq.maxcop_upper}) and (\ref{eq.maxcop_lower})).
}\label{lower_and_upper_bound_cop_max}
\end{figure*}
By using the definition of $\varepsilon$ in Eq.~(\ref{eq.def_cop}) and solving $\partial \varepsilon/\partial J_1=0$,
the maximum COP under the nontight-coupling condition $|q|<1$ is given as follows:
\begin{eqnarray}
\varepsilon_{\rm max}\left(q, \frac{\gamma_h}{\gamma_c}\right)=\frac{-q^2R_1 R_2+q^2R_1^2-\frac{(1-q^2)(1-\frac{1}{\tau})R_2^2}{1+\frac{\gamma_h}{\gamma_c}}}{q^2(1-\frac{1}{\tau})R_1R_2+(1-q^2)(1-\frac{1}{\tau})R_2^2},\label{eq.maxcop_re}
\end{eqnarray}
where $R_1$ and $R_2$ are defined as
\begin{eqnarray}
&&R_1 \equiv 1-\frac{1-\frac{1}{\tau}}{1+\frac{\gamma_h}{\gamma_c}},\\
&&R_2 \equiv 1+\sqrt{1+\frac{q^2R_1}{1-q^2}},
\end{eqnarray}
respectively.
Eq.~(\ref{eq.maxcop_re}) is a monotonically increasing function of $|q|$ and $\gamma_h/\gamma_c$~\cite{monotonicity}.
In the asymmetric dissipation limit $\gamma_h/\gamma_c \to \infty$, we obtain $R_1 \to 1$, $R_2 \to 1+\sqrt{1+q^2/(1-q^2)}$, and the upper bound:
\begin{eqnarray}
\varepsilon_{\rm max}^+(q)=\frac{q^2}{(1+\sqrt{1-q^2})^2}\varepsilon_{\rm C}.\label{eq.maxcop_upper}
\end{eqnarray}
We also obtain the lower bound as
\begin{widetext}
\begin{eqnarray}
\varepsilon_{\rm max}^-(q)=\label{eq.maxcop_lower}
\begin{cases}
\frac{\frac{q^2}{{\tau}^2}-\frac{q^2}{\tau}\left(1+\sqrt{1+\frac{q^2}{(1-q^2)\tau}}\right)-(1-q^2)(1-\frac{1}{\tau})\left(1+\sqrt{1+\frac{q^2}{(1-q^2)\tau}}\right)^2}{(1-\frac{1}{\tau})\frac{q^2}{\tau}\left(1+\sqrt{1+\frac{q^2}{(1-q^2)\tau}}\right)+(1-q^2)(1-\frac{1}{\tau})\left(1+\sqrt{1+\frac{q^2}{(1-q^2)\tau}}\right)^2} & \left(\frac{\gamma_h}{\gamma_c}\to 0 \ {\rm for} \ \tau \ge \frac{1}{\frac{q^2}{4(1-q^2)}+1}\right), \\
0 & \left(\frac{\gamma_h}{\gamma_c}\to \frac{\gamma_h}{\gamma_c^+} \ {\rm for} \ \tau < \frac{1}{\frac{q^2}{4(1-q^2)}+1}\right),
\end{cases}
\end{eqnarray}
\end{widetext}
depending on the parameter values corresponding to each case in Eq.~(\ref{eq.gc_range}).
$\varepsilon_{\rm max}^+(q)$ and $\varepsilon_{\rm max}^-(q)$ normalized by $\varepsilon_{\rm C}$ are plotted in Figs.~\ref{lower_and_upper_bound_cop_max}(b) and \ref{lower_and_upper_bound_cop_max}(c).
In the case of the symmetric dissipation $\gamma_h=\gamma_c$, we obtain
\begin{eqnarray}
\varepsilon_{\rm max}^{\rm sym}(q)=\frac{q^2(1-\frac{1}{\tau})-q^2R_3-(1-q^2)R_3^2}{q^2(1-\frac{1}{\tau})R_3+2(1-q^2)R_3^2},
\end{eqnarray}
where
\begin{eqnarray}
R_3 \equiv 1+\sqrt{1+\frac{q^2(1-\frac{1}{\tau})}{2(1-q^2)}}.
\end{eqnarray}
It is evident from Figs.~\ref{gpc_j2_cop_g} and \ref{gpc_j2_cop_q} that the COP at the maximum cooling power $\varepsilon_{\rm maxJ_2}$ is a well-defined optimum performance regime as is also suggested in~\cite{apertet13a}.
Analytical derivation of the maximum cooling power $J_{2, {\rm max}}$ and the COP at the maximum cooling power $\varepsilon_{\rm maxJ_2}$ are easily done as $\partial J_2/\partial J_1=0$ by using Eq.~(\ref{eq.J2_by_J1_re}) and they read as
\begin{eqnarray}
J_{2, {\rm max}}&&=\frac{q^2L_{22}}{4\gamma_c L_{11}}+L_{22}(1-q^2)X_2,\label{eq.maxJ2} \\
\varepsilon_{\rm max J_2}\left(q, \frac{\gamma_h}{\gamma_c}\right)&&=\frac{q^2\varepsilon_{\rm C}-\frac{4(1-q^2)}{(1+\frac{\gamma_h}{\gamma_c})}}{2q^2+q^2\varepsilon_{\rm C}\left(1+\frac{\gamma_h}{\gamma_c}\right)}.\label{eq.cop_maxJ2}
\end{eqnarray}
Eq.~(\ref{eq.cop_maxJ2}) is a monotonically increasing function of $|q|$, and a monotonically decreasing function of $\gamma_h/\gamma_c$ for $|q|=1$ while for $|q|\ne 1$ it depends as follows.
By solving $\partial \varepsilon_{\rm maxJ_2}/\partial (\gamma_h/\gamma_c)=0$, we obtain the optimum dissipation ratio:
\begin{eqnarray}
\left(\frac{\gamma_h}{\gamma_c}\right)_{\rm opt}=-1+\frac{4(1-q^2)}{q^2\varepsilon_{\rm C}} \left(1+\sqrt{1+\frac{q^2}{2(1-q^2)}}\right).\label{eq.opt_dissipation}
\end{eqnarray}
Because of the requirement $\left(\frac{\gamma_h}{\gamma_c}\right)_{\rm opt}>0$, we obtain a parameter range
\begin{eqnarray}
\varepsilon_{\rm C} < \frac{4(1-q^2)\left(1+\sqrt{1+\frac{q^2}{2(1-q^2)}}\right)}{q^2}
\end{eqnarray}
for this optimum value to exist. A non-monotonic behavior corresponding to this case can be seen in
Fig.~\ref{gpc_j2_cop_q} (c) ($\varepsilon_{\rm maxJ_2}$ of $\gamma_c=0.5$ is the maximum in the
three $\gamma_c$'s.), as mentioned in the last part of Sec. III B.
If this condition does not hold [e.g., the parameter values used in Fig.~\ref{gpc_j2_cop_q} (b)], Eq.~(\ref{eq.cop_maxJ2}) is a monotonically decreasing function of $\gamma_h/\gamma_c$~\cite{monotonicity}
as is similar to the case of $|q|=1$.
Then the upper bound of $\varepsilon_{\rm maxJ_2}(q)$ depending on the parameter values is given as follows:
\begin{widetext}
\begin{eqnarray}
\varepsilon_{\rm maxJ_2}^+(q) &&=\label{eq.cop_maxJ2_optimum}
\begin{cases}
\frac{q^2\varepsilon_{\rm C}R_4}{2(1+R_4)\left(q^2+2(1-q^2)(1+R_4)\right)} \ \left(\frac{\gamma_h}{\gamma_c}=\left(\frac{\gamma_h}{\gamma_c}\right)_{\rm opt} \ {\rm for} \ \varepsilon_{\rm C} < \frac{4(1-q^2)\left(1+\sqrt{1+\frac{q^2}{2(1-q^2)}}\right)}{q^2} \right),\\
\frac{q^2\varepsilon_{\rm C}-4(1-q^2)}{q^2(2+\varepsilon_{\rm C})}\ \left(\frac{\gamma_h}{\gamma_c}\to 0 \ {\rm for} \ \varepsilon_{\rm C} \ge \frac{4(1-q^2)\left(1+\sqrt{1+\frac{q^2}{2(1-q^2)}}\right)}{q^2} \right), \label{eq.copmaxj2_upper}
\end{cases}
\end{eqnarray}
\end{widetext}
where
\begin{eqnarray}
R_4\equiv \sqrt{1+\frac{q^2}{2(1-q^2)}}.
\end{eqnarray}
We obtain the lower bound of $\varepsilon_{\rm maxJ_2}(q)$ in the asymmetric dissipation limit of
$\gamma_h/\gamma_c \to \infty$ for any $|q|$-value:
\begin{eqnarray}
\varepsilon_{\rm maxJ_2}^-(q)=0.\label{eq.cop_maxJ2_lower}
\end{eqnarray}
These bounds are compared with those of the maximum COP in Fig.~\ref{lower_and_upper_bound_cop_max}.
In this figure, we stress that as $q$ is decreased from unity, as is similar to the behavior of the heat engine in Fig.~\ref{gpc_p_eta_q}, the allowed range of the optimized COPs rapidly becomes smaller and the maximum COP and the COP at maximum cooling power regimes tend to collapse in a unique inefficient performance regime (smaller loop) as in Fig.~\ref{gpc_j2_cop_q}.
For the symmetric dissipation $\gamma_h=\gamma_c$, we obtain
\begin{equation}
\varepsilon_{\rm maxJ_2}^{\rm sym}(q)=\frac{q^2\varepsilon_{\rm C}-2(1-q^2)}
{2q^2 \left(1+\varepsilon_{\rm C}\right)}.
\end{equation}
Additionally, if the tight-coupling condition $|q|=1$ is fulfilled, we reproduce the results in~\cite{apertet13a}:
\begin{equation}\label{copmaxj2a}
\varepsilon_{\rm maxJ_2}^-=0 \,\,\left(\frac{\gamma_h}{\gamma_c}\rightarrow \infty, \ |q|=1 \right),
\end{equation}
\begin{equation}\label{copmaxj2b}
\varepsilon_{\rm maxJ_2}^{\rm sym}=\frac{\varepsilon_{\rm C}}{2(\varepsilon_{\rm C}+1)} \,\,\left(\frac{\gamma_h}{\gamma_c}\rightarrow 1, \ |q|=1 \right),
\end{equation}
\begin{equation}\label{copmaxj2c}
\varepsilon_{\rm maxJ_2}^+=\frac{\varepsilon_{\rm C}}{\varepsilon_{\rm C}+2} \,\,\left(\frac{\gamma_h}{\gamma_c}\rightarrow 0, \ |q|=1 \right).
\end{equation}
\section{Discussion and conclusions}
We have reported unified results for both HE and RE obtained by MNLIT models making emphasis on the influence of irreversibilities by the heat leaks, internal dissipations, and their interplay. The results in Sec.~III clearly show that in order to obtain realistic (loop-shaped) power-efficiency and cooling power-COP performance characteristics, the irreversibilities by the heat leaks are a necessary ingredient. On the other hand, the internal dissipations into the heat reservoirs
account for quantitative variations in the involved energetic magnitudes but without affecting the qualitative behaviors.
In Sec.~IV, we have presented explicit calculations for some optimized figures of merit and their bounds in terms of dissipation to heat reservoirs and of the coupling parameter.
In particular, the limiting results for $|q|=1$ deserve some comments.
In this limit, the efficiency at maximum power Eq.~(\ref{remchig}) can be rewritten as
\begin{equation}
\eta_{\rm maxP}=\frac{\eta_{\rm C}}{2-\gamma \eta_{\rm C}},\label{eq.eta_ss}
\end{equation}
with $\gamma \equiv \frac{1}{1+\gamma_c/\gamma_h}=\frac{\gamma_h}{\gamma_c+\gamma_h}$, i.e., the ratio of the dissipation strength in the hot heat reservoir to the overall strength.
This result was also previously reported by Schmiedl and Seifert in a stochastic heat engine model~\cite{seifert08} and then reinterpreted
by Apertet et al.~\cite{apertet12b} as the characteristic efficiency at maximum power for exoreversible HE models where the only irreversibility comes from the internal dissipations.
This formula may also be connected to the LD models by interpreting the coefficients of dissipation-strength as the coefficients of heat-transfer between the working substance and the heat reservoir in
the FTT framework~\cite{jchen13a}.
Indeed, if we choose $\gamma=0$ and $1$ in Eq.~(\ref{eq.eta_ss}), we reproduce the bounds given by Eqs.~(\ref{remchilb}) and (\ref{remchiub}), respectively, while for symmetric dissipation $\gamma=1/2$ we reproduce
Eq.~(\ref{remchisyma}).
For RE, the COP at maximum cooling power Eq.~(\ref{eq.cop_maxJ2}) under $|q|=1$ is given as
\begin{eqnarray}
\varepsilon_{\rm maxJ_2}=\frac{\varepsilon_{\rm C}}
{2+\frac{\varepsilon_{\rm C}}{1-\gamma}},\label{eq.cop_mp}
\end{eqnarray}
by using $\gamma$.
For $\gamma=0$ and $1$ we reproduce the bounds in Eqs.~(\ref{copmaxj2c}) and (\ref{copmaxj2a}), respectively, while for symmetric dissipation $\gamma=1/2$ we reproduce Eq.~(\ref{copmaxj2b}),
which is already reported in~\cite{apertet13a} as a particular case of a thermoelectric refrigerator.
Therefore, the MNLIT model under the tight-coupling condition also reproduces correctly the results of this exoreversible model.
A special comment is merited by the maximum cooling power condition for RE expressed in Eq.~(\ref{eq.cop_mp}) for which Apertet et al.~\cite{apertet13a} have proposed that it should be considered as the only genuine counterpart of the Schmiedl-Seifert efficiency Eq.~(\ref{eq.eta_ss}) for HE. Both results are obtained under the same exoreversible conditions optimizing the efficiency (COP) under maximum useful benefit (power output for HE and cooling power for RE). On the other hand, the original CA value was obtained under quite different assumption of endoreversibility (without internal dissipations and heat leaks), which when reversed does not allow the optimization of the cooling power~\cite{apertet13a}.
Exoreversible and endoreversible models are indeed two (extreme) different models which define different specific coupling to the external heat reservoirs, thus it is not surprising that they lead to different expressions. Conversely, the LD models provide a unified framework for HE and RE where the exact CA-value emerges linked to a certain symmetric condition. In this context (with the same model, same symmetric condition, and same optimization criterion),
$\varepsilon_{{\rm max\chi}}=\sqrt{1+\varepsilon_{\rm C}}-1\equiv \varepsilon_{\rm CA}$ for RE
was proposed as the CA counterpart~\cite{carla12a,yuki13a}.
This value gives a better comparison with experimental results~\cite{carla13} than the COP at maximum cooling power Eq.~(\ref{eq.cop_mp}) whose maximum possible value is unity even in the limit of the small temperature difference as $\varepsilon_{\rm maxJ_2}\approx 1-\gamma\leq 1$ while the Carnot COP diverges as $\varepsilon_{\rm C}\to \infty$ in this limit.
In closing, all of the above clearly illustrates that the minimally nonlinear irreversible model succeeds in reproducing
various results derived by previous studies.
In particular, the MNLIT model provides a clear interpretation of the global performance characteristics of generic heat devices in terms of the interplay between the heat leaks and the internal dissipations, and it reproduces the figures of merit optimized under some performance criteria and some conditions for both HE and RE.
Additionally, further studies are needed to establish clearer connections between the MNLIT models~\cite{yuki12a,yuki13a}, FTT frameworks~\cite{curzon75}, LD models~\cite{esposito10a}, and LIT models based on the local force-flux relationships~\cite{apertet13b}.
Related to this, we note that a recent work~\cite{sheng13a} reports a complementary idea of the nonlinear dissipation terms
by introducing the concepts of weighted reciprocal temperatures and weighted thermal fluxes.
In~\cite{bizarro,B}, we also note that dissipation effects by the friction on the heat devices have been discussed. Although such friction effects as a cause of the dissipations into the heat reservoirs have not been taken into account in our present model,
an extension of our model along this line would be interesting in terms of explaining behaviors of actual heat devices.
\begin{acknowledgements}
Y. I. acknowledges the financial support from a Grant-in-Aid for JSPS Fellows (Grant No. 25-9748).
J. M. M. R. and A. C. H. acknowledge Ministerio de Economia y Competetividad of Spain, MINECO (ENE2013-40644-R).
\end{acknowledgements}
|
\section{Alternative test statistics} \label{Sec1}
The covariance test statistic $T_k$ associated with covariate $X_j$ is
defined as the covariance between the response vector $y$ and the net
contribution of covariate $X_j$ toward the mean vector $X \beta$ at
the $(k+1)$th step of the Lasso solution path with regularization
parameter $\lambda= \lambda_{k + 1}$, scaled by the inverse of the
error variance $\sigma^2$. Since the Lasso solution is gauged by the
regularization parameter $\lambda$, a key ingredient in the definition
of $T_k$ is a refitting of the Lasso on the previous support at the
$k$th step right before the inclusion of covariate $X_j$ with the
reduced regularization parameter $\lambda= \lambda_{k + 1}$. This
alignment of the regularization parameter yields a more accurate
account of the contribution of covariate $X_j$ conditional on
previously selected covariates before the next step occurs (either an
addition or a deletion of a covariate).
In view of the geometrical representation of the Lasso solution path,
the choice of a common regularization parameter amounts to that of a
common correlation in magnitude between selected covariates and the
residual vector, provided that all covariate vectors are aligned to a
common scale. In this sense, the covariance test statistic $T_k$ bears
some similarity to the conventional chi-squared test statistic, in
terms of the reduction of the residual sum of squares, for evaluating
the significance of the contribution of a newly added covariate. A main
difference is that the above correlation is constrained as a fixed
number zero in the latter, while it is adaptively
determined at the $(k+1)$th step in the Lasso.
From the estimation point of view, it seems appealing to impose a
smaller regularization to reduce the bias issue incurred by shrinkage.
The bias issue can also affect the significance of relatively weak
covariates. Therefore, a natural extension of the covariance test
statistic $T_k$ can be
\begin{equation}
\label{001} T_{k, c} = \bigl(\bigl\langle y, X \hat{\beta}(c
\lambda_{k + 1})\bigr\rangle- \bigl\langle y, X_A \tilde{
\beta}_A(c \lambda_{k + 1})\bigr\rangle\bigr)/
\sigma^2
\end{equation}
for some constant $0 \leq c \leq1$, where $\hat{\beta}(c \lambda_{k
+ 1})$ and $\tilde{\beta}_A(c \lambda_{k + 1})$ are the Lasso
estimators with regularization parameter $c \lambda_{k + 1}$
constrained on sets of covariates $A \cup\{j\}$ and $A$, respectively.
Clearly, $T_{k, c}$ with $c = 1$ reduces to $T_k$. An interesting
question is whether a suitable choice of the constant $c$ may lead to
an exact $\Exp(1)$ asymptotic null distribution for the test statistic
$T_{k, c}$ in the case of general design matrix, as opposed to a
conservative exponential limit for $T_k$.
The significance test with the covariance test statistic $T_k$ is a
sequential procedure which evaluates the significance of any newly
selected covariate in the current Lasso\vadjust{\goodbreak} model. It would also be
appealing to test the significance of each active covariate $X_\ell$
conditional on the set of all remaining active covariates $A_\ell= (A
\cup\{j\}) \setminus\{\ell\}$, since some previously significant
covariates may no longer be significant as new covariates enter the
model. A natural procedure seems to be applying the covariance test
statistic $T_k$ or $T_{k, c}$ to each covariate $X_\ell$ with set $A$
replaced by $A_\ell$. This may also be used to test the significance
of covariates in a general Lasso model given by a prespecified
regularization parameter $\lambda$.
\section{The event $B$ and generalized irrepresentable conditions}\label{Sec2}
There seem to be two key conditions for establishing the conservative
exponential limit for the covariance test statistic $T_k$ in the case
of general design matrix: the event $B$ and generalized irrepresentable
conditions. It would be appealing to verify whether these conditions
hold for a given design matrix. The event $B$ condition assumes that
there exist an integer $k_0 \geq0$ and a fixed set $A_0$ of covariates
containing the set $A^* = \supp(\beta^*)$ of all true active
covariates such that with asymptotic probability one, the Lasso model
$A$ at step $k_0$ of the Lasso solution path is identical to $A_0$. In
other words, this condition requires the sure screening property to
hold for the Lasso.
To provide some insights into the event $B$ condition, let us consider
the setting of linear regression model
\begin{equation}
\label{002} y = X \beta^* + \varepsilon
\end{equation}
as in \citet{LF09}. Set $p = 1000$ with true regression
coefficient vector $\beta^* = (1, -0.5, 0.7, -1.2, -0.9, 0.3, 0.55, 0, \ldots, 0)^T$, sample\vspace*{1pt} the rows of the design matrix $X$ as i.i.d.
copies from $N(0, \Sigma)$ with $\Sigma= (0.5^{|i-j|})_{i, j = 1,
\ldots, p}$, and generate error vector $\varepsilon$ independently
from $N(0, \sigma^2 I_n)$ with $\sigma= 0.15$ or $0.3$. Note that the
minimum signal strength is the same as or twice the error standard
deviation. We generated 200 data sets from this model with sample size
$n$ ranging from 80 to 120 and applied the Lasso with the LARS
algorithm [\citet{EHJT04}] to generate the solution path. Figure~\ref{fig1} depicts the sure screening probability curves as a function
of sparse model size for Lasso with different sample size and error
level. The sure screening probability provides an upper bound on the
probability of event $B$. We see that both the signal strength (in
terms of the sample size and error level) and sparse model size are
crucial for the sure screening property of the Lasso estimator. As the
sparse model size and sample size become larger, the Lasso can have
significant sure screening probability. Such a probability can drop as
the noise level increases. It would be interesting to provide some
theoretical understandings on the impacts of these factors on both sure
screening probability and the probability of event $B$ for Lasso.
\begin{figure}
\includegraphics{1175df01.eps}
\caption{Sure screening probability curves as a function of sparse
model size for Lasso with $n = 80$ (thin solid), $90$ (dashed), $100$
(dotted), $110$ (dash--dot) and $120$ (thick solid). Left panel for
$\sigma= 0.15$ and right panel for $\sigma= 0.3$.}
\label{fig1
\end{figure}
The generalized irrepresentable condition introduced in the paper
extends the irrepresentable condition [\citet{ZY06}] for
characterizing the model selection consistency of the Lasso, which
means that the true underlying sparse model $A^*$\vadjust{\goodbreak} is exactly recovered
with asymptotic probability one. It involves the fixed set $A_0 \supset
A^*$ introduced in the event $B$ condition. Intuitively, this condition
puts some constraint on the correlation between the noise covariates
and true ones. See, for example, \citet{LF09} and \citet
{FL11}, for examples, and more discussions on these types of conditions
for characterizing the model selection consistency of a wide class of
regularization methods including Lasso.
\section{Model misspecification} \label{Sec3}
The event $B$ condition makes an implicit assumption on the minimum
signal strength. It would be interesting to investigate the more
general case of strong and weak covariates, in which some covariates
may have relatively weak contributions to the response. In such a case,
the true underlying sparse model may no long be contained somewhere on
the solution path given by a regularization method. In other words, the
true model may not be included in the sequence of sparse candidate
models, leading to model misspecification. Apart from missing some true
covariates, model misspecification can generally occur when one
misspecifies the family of distributions. Since model misspecification
is unavoidable in practice, it would be helpful to understand its
impact on statistical inference. For example, \citet{LL13}
recently revealed that the covariance contrast matrix between the
covariance structures in the misspecified model and in the true model
plays a pivotal role in characterizing the impact of model
misspecification on the problem of model selection. It would be
interesting to study the effects of model misspecification in the
context of significance testing.\looseness=-1
\section{More general regularization methods} \label{Sec4}
A key ingredient that makes the covariance test statistic $T_k$ admit
the nice $\Exp(1)$ asymptotic null distribution is the shrinkage
effect induced by the $L_1$-penalty in Lasso which offsets the
inflated\vadjust{\goodbreak}
stochastic variability due to the adaptivity in variable selection.
Many other regularization methods, including concave ones such as the
SCAD [\citet{FL01}], MCP [\citet{Zhang10}] and SICA [\citet{LF09}],
have been proposed for variable selection. A natural and important
question is whether a similar test statistic can be constructed for
testing the significance of covariates for the class of concave
regularization methods. Since these methods are generally nonconvex, it
would be crucial to study the regularized estimate as the global or
computable solution. Consider a fixed regularization parameter $\lambda
$ associated with a penalty function $p_\lambda(t)$ defined on $[0,
\infty)$. One possible test statistic is to extend $T_k$ or $T_{k, c}$
by replacing the constrained Lasso estimators with the corresponding
constrained regularized estimators with the same regularization
parameter. An interesting open question is whether such a generalized
covariance test statistic would have a similar asymptotic null
distribution as for Lasso or a different asymptotic limit may appear.
To gain some insights into these questions, let us consider a natural
extension of the Lasso, the combined $L_1$ and concave regularization
method introduced in \citet{FL13} and defined as the following
regularization problem:
\begin{equation}
\label{003} \min_{\beta\in\mathbb{R}^p} \bigl\{(2n)^{-1} \|y - X
\beta\|_2^2 + \lambda_0\|\beta
\|_1 + \bigl\|p_\lambda(\beta)\bigr\|_1 \bigr\},
\end{equation}
where $\lambda_0 = \tilde{c} \{(\log p)/n\}^{1/2}$ for some positive
constant $\tilde{c}$, $p_\lambda(\beta)$ is a compact notation
denoting $p_\lambda(|\beta|) = (p_\lambda(|\beta_1|), \ldots,
p_\lambda(|\beta_p|))^T$ with $|\beta| = (|\beta_1|, \ldots,\break
|\beta_p|)^T$, and $p_\lambda(t)$, $t \in[0, \infty)$, is a penalty
function indexed by the regularization parameter $\lambda\geq0$. This
approach combines the strengths of both Lasso and concave methods in
prediction and variable selection and has enhanced stability compared
with using concave methods alone. They proved that under mild
regularity conditions, the global and computable solutions can enjoy
oracle inequalities under various prediction and estimation losses in
parallel to those in \citet{BRT09} established for
Lasso and Dantzig selector [\citet{CandesTao07}], but with improved
sparsity. In particular, the combined regularization method admits an
explicit bound on the false sign rate, which can be asymptotically vanishing.
\begin{figure
\includegraphics{1175df02.eps}
\caption{Quantile--quantile plots of the covariance test statistic
$T_{k, c}$ with $c = 1$ and $0.1$ versus the $\Exp(1)$ distribution
for the combined $L_1$ and concave regularization with $c_0 = 0.6$ and
$0.25$, respectively.}
\label{fig2
\end{figure}
\begin{figure
\includegraphics{1175df03.eps}
\caption{Quantile--quantile plots of the covariance test statistic
$T_{k, c}$ with $c = 1$ and $0.1$ for the combined $L_1$ and concave
regularization with $c_0 = 0.1$. Top panel is versus the $\Exp(1)$
distribution and bottom panel is versus the chi-squared distribution
with 1 df.}
\label{fig3
\end{figure}
Consider the same example as in Section~\ref{Sec2}, with $n = 100$ and
$\sigma= 0.3$, and use the SICA penalty $p_\lambda(t) = \lambda(a +
1) t/(a + t)$, $t \in[0, \infty)$, with a small shape parameter $a$
for the concave component. For each data set, we applied the combined
$L_1$ and SICA method to generate a sequence of sparse candidate models
with positive constant $\tilde{c} = c_0 \sigma$ and $c_0$ chosen to
be $0.1$, $0.25$ and $0.6$. With tighter control of the false sign
rate, the sure screening property can hold for this method with smaller
sparse model size. Since the true underlying sparse model has seven
variables, we are interested in testing the significance of the eighth
covariate to enter the model. To this end, we used the generalized
covariance test statistics $T_k$ and $T_{k, c}$ (with a slight abuse of
notation) as suggested before, where $T_{k, c}$ with $c = 1$ reduces to
$T_k$ and both $c \lambda_0$ and $c \lambda$ are used in place of
$\lambda_0$ and $\lambda$ in the constrained refittings for~$T_{k,
c}$. The cases of $c = 1$ and $0.1$ were considered.
Figures~\ref{fig2} and \ref{fig3} compare the distributions of the
generalized covariance test statistic $T_{k, c}$ with the $\Exp(1)$
distribution or chi-squared distribution with 1~df over different
combinations of $c_0$ and $c$. Note that as $c_0$ increases, the
combined regularization becomes closer to the Lasso, which is reflected
in Figure~\ref{fig2}. The top panel of Figure~\ref{fig2} for the case
of $c_0 = 0.6$ shows that $\Exp(1)$ fits the distributions of $T_{k,
c}$ well and the bottom panel for the case of $c_0 = 0.25$ suggests
that $\Exp(1)$ is a conservative fit. It is interesting to observe
that the choice of $c = 0.1$ seems to make the distribution of $T_{k,
c}$ closer to the $\Exp(1)$ distribution compared to that of $c = 1$\vadjust{\goodbreak}
in the latter. In the case of $c_0 = 0.1$, the combined regularization
becomes closer to concave regularization. We observe an interesting
transition phenomenon for the distribution of the generalized
covariance test statistic $T_{k, c}$. As demonstrated in Figure~\ref{fig3}, it is now more light-tailed than the $\Exp(1)$ distribution
and interestingly the chi-squared distribution with 1~df provides a
nice fit. It would be interesting to provide theoretical understandings
on such a phenomenon.
\section{Concluding remarks} \label{Sec5}
The covariance test statistic proposed in this paper provides a new
general framework for testing the significance of covariates for the
Lasso and related sparse modeling methods in high dimensions. There are
many interesting and important questions that remain to be answered in
high-dimensional inference. This paper initiates a new area and will
definitely stimulate new ideas and developments in the future. We thank
the authors for their clear and imaginative work.
\section*{Acknowledgment}
We sincerely thank the Co-Editor, Professor Peter Hall, for his kind
invitation to comment on this discussion paper.
|
\section{Introduction}
Strong interactions between single atoms and single photons, engineered through the use of optical resonators or cavities (i.e., cavity quantum electrodynamics, or cavity QED), provide a means of realizing quantum dynamics and phenomena that are of both fundamental and practical interest \cite{Mabuchi02,Miller05,Parkins13}. On the fundamental side, for example, issues of quantum coherence, entanglement, and measurement can be addressed, while, on the practical side, capabilities for quantum state preparation and manipulation are of immediate relevance to quantum information processing. Indeed, in this latter context, cavity QED systems comprised of single atoms in optical Fabry-P\'erot cavities have already contributed an impressive range of experimental demonstrations, including deterministic single photon generation, state transfer between atoms and light, and conditional quantum dynamics \cite{Turchette95,Kuhn02,McKeever04,Keller04,Birnbaum05,Boozer07,Wilk07,Hijlkema07,Kubanek08,Barros09,Specht11,Ritter12,Reiserer13,Reiserer14}. These operations would be key building blocks for a quantum network in which atoms are used for storage of quantum information at local nodes of the network, photons (``flying qubits'') provide a means of distributing quantum information between distant nodes, and atom-photon interactions enable transfer between matter and light, and also manipulation, of the quantum information \cite{Kimble08}.
Despite this remarkable progress in experimental cavity QED, the ``conventional'' Fabry-P\'erot cavities used in the above-mentioned demonstrations do not lend themselves well to large scale networking. So, in the pursuit of scalable architectures for quantum networks based on effects in cavity QED and on propagating light fields, a variety of microfabricated (i.e., microchip-based) systems are now under active investigation; these include semiconductor quantum dots in micropillar, microdisk, or photonic crystal cavities (for recent reviews see, e.g., \cite{Reitzenstein12,Lodahl13}), cold atoms with monolithic WGM resonators \cite{Spillane05,Barclay06,Aoki06,Dayan08,Aoki09,Alton10,Junge13,OShea13,Shomroni14}, fiber-Fabry-P\'erot resonators \cite{Steinmetz06,Colombe07,Trupke07}, or photonic crystal waveguides \cite{Thompson13,Tiecke14,Yu14,Goban14}, and nanocrystals with microsphere cavities \cite{Thomas06,Park06}.
Key properties of ultra-low intrinsic cavity mode losses and high efficiency transfer of light fields into and out of the cavity modes are well established features of, in particular, microspheres, microdisks and microtoroids coupled via evanescent fields to tapered optical fibers \cite{Cai00,Armani03,Spillane03}. Additionally, the ultra-small volumes of the WGM's supported by these monolithic structures lead to very large single photon electric fields and, consequently, very large atom-field coupling strengths, $g$, which may exceed dissipative rates in the system.
In fact, this regime of strong coupling cavity QED has now been demonstrated experimentally for both a single alkali atom in the evanescent field of a microtoroidal \cite{Aoki06} or microbottle \cite{Junge13} resonator and for a single quantum dot embedded in a microdisk resonator \cite{Srinivasan07a,Srinivasan08}.
An important aspect of these particular systems, related to the input-output coupling efficiency of photons, is the ability to ``tune'' the external coupling rate, $\kappa_{\rm ex}$, of the cavity modes to the optical fiber by adjusting the distance between the taper and the WGM microresonator. By setting $\kappa_{\rm ex}$ to be much larger than the rates for intrinsic cavity losses ($\ki$) and spontaneous emission of the atom or quantum dot ($\gamma$), one ensures that the fiber is indeed the dominant input-output channel in the system.
Moreover, depending on the precise size of $\kappa_{\rm ex}$ relative to certain other coupling parameters, the microresonator cavity QED system can exhibit quite distinct regimes of operation with regards to effect on a light field propagating along an evanescently-coupled optical fiber.
In particular, much interest has to date been focussed on operation under the condition of {\em critical coupling} between the fiber and the WGM microresonator \cite{Cai00,Aoki06,Junge13,OShea13}, i.e., $\kex=\kex^{\rm cr}$, whereby, in the absence of an atom, destructive interference prevents transmission of resonant light along the fiber, with, instead, light either reflected back in the opposite (backward) direction, dissipated through the intrinsic WGM losses, or transferred to a second evanescently-coupled fiber \cite{OShea13} (or, in general, some combination of these possibilities). A strongly coupled atom ($g>\{\kex,\ki,\gamma\}$) induces vacuum Rabi splitting of the cavity resonance and breaks the critical coupling condition, causing a finite transmission in the forward direction. However, such WGM microresonators support degenerate, counterpropagating modes, both of which couple to the atom. Identically-polarized, counterpropagating TE modes give rise to a pair of orthogonal standing-wave normal modes and the phase of one of these modes can always be chosen so that it has a node at the position of the atom. The presence of this uncoupled normal mode means that for TE fields the incident field can never be fully transmitted; in fact, the transmission is limited to 25\% of the incident field \cite{Aoki06b,Junge13} for the strong coupling regime $g>\{\kex,\ki,\gamma\}$.
In the so-called ``bad cavity'' regime, for which $\kappa=\kex+\ki\gg g$ (but $g\gg\gamma$ so that $g^2/\kappa\gamma>1$), it is actually possible for this limit to be increased towards 100\% \cite{Dayan08}, but only for very weak incident fields, as the atom is typically saturated with extremely small photon numbers. Alternatively, it has been pointed out, and demonstrated experimentally, that for nontransversally-polarized WGM's (i.e., TM rather than TE fields) it is possible to realize a situation where the atom effectively couples to only a single, unidirectional WGM, such that strong vacuum Rabi splitting of this mode by the atom can in principle enable 100\% transmission along the fiber of light resonant with the bare WGM frequency \cite{Junge13,OShea13}.
In this paper, we describe a new and promising scheme which, under ideal conditions, also enables 100\% contrast in transmission along the fiber between the no-atom and single-atom cases, but which works despite the presence of an uncoupled, standing-wave normal mode of the microresonator (i.e., it works with transversally-polarized TE WGM's) and also enables control of the propagation of relatively strong light fields or, alternatively, of light pulses of significant photon number. The scheme requires the strong coupling regime of cavity QED, i.e., $g>\kappa,\gamma$, but, with regards to the fiber-microresonator coupling, operates in the {\em strongly over-coupled regime}, i.e., $\kex\gg\kex^{\rm cr}$. Under this condition, and in the absence of an atom, incident light is simply transmitted along the fiber. However, as we shall show, the effect of a single, strongly-coupled atom is now to effectively restore the condition of critical coupling and block the transmission of resonant light past the microresonator.
The present scheme can be viewed as complementing a previously studied, single-atom setup in the bad-cavity, over-coupled regime, demonstrated experimentally in \cite{Aoki09}, which exhibits the same contrast in transmission, but works only for weak incident light fields; in particular, in this regime the (two-level) atom can only reflect incident light efficiently provided that no more than one photon ever arrives within the effective atomic lifetime \cite{Carmichael93,Chang07,Rosenblum11}. This mechanism for reflection also yields strongly antibunched light in the reflected field, in contrast to the present scheme, which preserves coherent fields in reflection.
We begin in Section II with a description of our theoretical model, then in Section III present a linearized analysis, which allows a quite straightforward and transparent explanation of the present scheme. In Section IV we move beyond the linearized analysis to numerical solutions of the master equation for the system, enabling us to examine dependence on the atom-field coupling strength and on the coherent driving field strength. We connect the saturation behavior of the system with the semiclassical description of optical bistability of the cavity field, which allows us to quantify simply the regime of validity of the linear analysis. Given this knowledge, in Section V we consider the case of incident coherent-state pulses and give a conservative upper estimate for the mean photon number of a pulse that could in principle be controlled by the single-atom switch. In Section VI we then outline a scheme, involving an auxiliary (uncoupled) internal atomic level, which could allow the preparation of an entangled state of the atom and coherent state pulses propagating in opposite directions along the fiber. Using a linearized analysis, we evaluate the fidelity of preparation of such an entangled state for realistic parameters.
\section{Theoretical model}
A schematic of the system is shown in Fig.~\ref{fig:microtoroid}. The internal, counter-propagating
cavity modes are described
in terms of the annihilation operators $a$ and $b$, while the external input and output fields are
described by the operators $\{ a_{\rm in,ex},a_{\rm out,ex},b_{\rm in,ex},b_{\rm out,ex}\}$. The internal fields
suffer an intrinsic loss at the rate $\kappa_{\rm i}$ and an extrinsic loss
at the rate $\kappa_{\rm ex}$ due to the fiber coupling.
An atom is assumed to couple to the evanescent fields of the intracavity (traveling-wave) modes
with a strength of magnitude $|g_{\rm tw}| = g_0^{\rm tw}\, f(r)$, where $r$ is the radial distance of the atom from the surface of the toroid and a characteristic form for the (real) function $f(r)$ is $f(r)={\rm e}^{-\alpha r}$ with $\alpha\sim 2\pi /\lambda$.
\begin{figure}
\includegraphics[scale=0.48]{fig1.eps}
\caption{\label{fig:microtoroid} (Color online). Schematic of the microtoroid, atom, and tapered-fiber coupler. Whispering gallery modes, $a$ and $b$, couple with strength $g$ to an atom through their evanescent fields. The modes are coupled via intrinsic scattering at rate $h$ and suffer intrinsic loss at rate $\ki$ into the output fields $a_{\rm out,i}$ and $b_{\rm out,i}$. Evanescent coupling of the modes to the tapered fiber corresponds to an extrinsic loss at rate $\kex$, while atomic spontaneous emission into free space (output field $c_{\rm out}$) occurs at rate $\gamma$. Fields propagating along the fiber towards and away from the microtoroid are denoted by $\{ a_{\rm in,ex},b_{\rm in,ex}\}$ and $\{a_{\rm out,ex},b_{\rm out,ex}\}$, respectively. }
\end{figure}
\subsection{Hamiltonian and master equation}
We consider a two-level atom with transition frequency $\omega_{\rm A}$ and described by the
raising and lowering operators $\sigma^\pm$. The ``bare'' cavity mode (WGM)
frequency is $\omega_{\rm C}$, and the two counterpropagating modes are assumed to be coupled
(due, e.g., to scattering off imperfections) with a strength $h$. A coherent probe field of frequency
$\omega_{\rm p}$ in the input field $a_{\rm in,ex}$ drives the mode $a$ with strength $\Ep$.
In a frame rotating at the probe frequency, the Hamiltonian for the system can be written in the form \cite{Aoki06b,Dayan08b,Domokos00,Rosenblit04,Srinivasan07b}
\begin{eqnarray}
H &=& \Delta_{\rm A}\sigma^+\sigma^- + \Delta_{\rm C} \left( a^\dag a + b^\dag b \right)
+
h \left( a^\dag b + b^\dag a \right) \nonumber
\\
&& + \left( \Ep^\ast a + \Ep a^\dag \right) + \left( g_{\rm tw}^\ast a^\dag\sigma^- + g_{\rm tw} \sigma^+a \right)
\nonumber
\\
&& + \left( g_{\rm tw} b^\dag\sigma^- + g_{\rm tw}^\ast \sigma^+b \right) ,
\end{eqnarray}
where, specifically, $g_{\rm tw}=g_0^{\rm tw}\, f(r) {\rm e}^{ikx}$, with $x$ the position of the atom around the circumference of the toroid, and
\begin{equation}
\Delta_{\rm A} = \omega_{\rm A} - \omega_{\rm p} \, , ~~~~
\Delta_{\rm C} = \omega_{\rm C} - \omega_{\rm p} .
\end{equation}
Introducing dissipation, the system can be described by the master equation
\begin{eqnarray}
\dot{\rho} &=& -{\rm i}\left[ H ,\rho \right]
+ \kappa {\cal D}[a]\rho + \kappa {\cal D}[b]\rho + \frac{\gamma}{2} {\cal D}[\sigma^-]\rho ,
\label{eq:ME}
\end{eqnarray}
where $\rho$ is the density operator for the atom-cavity system, $\kappa = \kappa_{\rm i} + \kappa_{\rm ex}$ is the field decay rate for the cavity modes, $\gamma$ is the atomic spontaneous emission
rate, and ${\cal D}[{\cal O}]\rho\equiv 2{\cal O}\rho{\cal O}^\dagger - {\cal O}^\dagger{\cal O}\rho - \rho {\cal O}^\dagger{\cal O}$. For a fully quantum mechanical treatment of the system we can compute numerical solutions to the master equation (\ref{eq:ME}) using truncated number state bases for the cavity modes.
\subsection{Normal mode representation}
The Hamiltonian and master equation for the atom-cavity system can also be usefully expressed in terms of the normal modes of the cavity, which are defined by the operator combinations
\begin{equation}
A = \frac{a+b}{\sqrt{2}} \, , ~~~~~ B = \frac{a-b}{\sqrt{2}} \, .
\end{equation}
In particular, one can show that
\begin{eqnarray}
\dot{\rho} &=& -{\rm i}\left[ H^\prime ,\rho \right]
+ \kappa {\cal D}[A]\rho + \kappa {\cal D}[B]\rho + \frac{\gamma}{2} {\cal D}[\sigma^-]\rho ,
\label{eq:ME_AB}
\end{eqnarray}
where
\begin{eqnarray}
H^\prime &=& \Delta_{\rm A}\sigma^+\sigma^- + (\Delta_{\rm C}+h) A^\dag A
+ (\Delta_{\rm C}-h) B^\dag B
\nonumber
\\
&& + \frac{1}{\sqrt{2}} \left( \Ep^\ast A
+ \Ep A^\dag \right) + \frac{1}{\sqrt{2}} \left( \Ep^\ast B
+ \Ep B^\dag \right) \nonumber
\\
&& + g_A \left( A^\dag\sigma^- + \sigma^+A \right)
- {\rm i}g_B \left( B^\dag\sigma^- - \sigma^+B \right) ,
\label{eq:H_normalmodes}
\end{eqnarray}
with
\begin{eqnarray}
g_A &=& \sqrt{2}\,{\rm Re}\{ g_{\rm tw}\} = \sqrt{2}\, g_0^{\rm tw}f(r)\cos(kx) ,
\\
g_B &=& \sqrt{2}\,{\rm Im}\{ g_{\rm tw}\} = \sqrt{2}\, g_0^{\rm tw} f(r)\sin(kx) .
\end{eqnarray}
So, depending on the position of the atom, coupling may occur predominantly
(or even exclusively) to only one of the two normal modes.
\subsection{Input and output fields}
From the theory of inputs and outputs in optical cavities \cite{CollettGardiner84_85,QuantumNoise}, the operators for the fields output from the two cavity modes into the fiber are given in terms of the input and intracavity field operators as
\begin{eqnarray}
a_{\rm out,ex}(t) &=& - a_{\rm in,ex}(t) + \sqrt{2\kappa_{\rm ex}}\, a(t) ,
\\
b_{\rm out,ex}(t) &=& - b_{\rm in,ex}(t) + \sqrt{2\kappa_{\rm ex}}\, b(t) ,
\end{eqnarray}
where $[a_{\rm in,ex}(t),a_{\rm in,ex}^\dagger (t^\prime)]=[a_{\rm out,ex}(t),a_{\rm out,ex}^\dagger (t^\prime)]=\delta (t-t^\prime )$ (and similarly for $b_{\rm in,ex}(t)$ and $b_{\rm out,ex}(t))$.
The particular form of the probe driving in the Hamiltonian corresponds to
coherent amplitudes of the input fields given by
\begin{equation}
\langle a_{\rm in,ex} \rangle = - \frac{{\rm i}\Ep}{\sqrt{2\kappa_{\rm ex}}} , ~~~~~
\langle b_{\rm in,ex} \rangle = 0 ,
\end{equation}
and corresponding input photon fluxes of
\begin{equation}
\left|\langle a_{\rm in,ex}\rangle\right|^2 =
\frac{\left|\Ep\right|^2}{2\kappa_{\rm ex}} \, , ~~~~~
\left|\langle b_{\rm in,ex}\rangle\right|^2 = 0 \, .
\end{equation}
The normalized forward and backward photon fluxes are therefore defined as
\begin{eqnarray}
T_{\rm F} = \frac{\langle a_{\rm out,ex}^\dag a_{\rm out,ex}\rangle_{\rm ss}}{\left|\Ep\right|^2/(2\kappa_{\rm ex})} \, , ~~~~
T_{\rm B} = \frac{\langle b_{\rm out,ex}^\dag b_{\rm out,ex}\rangle_{\rm ss}}{\left|\Ep\right|^2/(2\kappa_{\rm ex})}
\end{eqnarray}
and the corresponding photon correlation functions as
\begin{eqnarray}
g_{\rm FF}^{(2)}(0) &=& \frac{\langle a_{\rm out,ex}^\dag a_{\rm out,ex}^\dag a_{\rm out,ex} a_{\rm out,ex}\rangle_{\rm ss}}{\left(\langle a_{\rm out,ex}^\dag a_{\rm out,ex}\rangle_{\rm ss}\right)^2} \, ,
\\
g_{\rm BB}^{(2)}(0) &=& \frac{\langle b_{\rm out,ex}^\dag b_{\rm out,ex}^\dag b_{\rm out,ex} b_{\rm out,ex}\rangle_{\rm ss}}{\left(\langle b_{\rm out,ex}^\dag b_{\rm out,ex}\rangle_{\rm ss}\right)^2} \, .
\end{eqnarray}
\section{Linearized analysis}
For sufficiently weak driving by a coherent probe field we can assume that the atom is
only very weakly excited and spends most of its time in the ground state, such that $\left[\sigma^-,\sigma^+\right]\rho \simeq \sigma^-\sigma^+\rho \simeq \rho$, or, effectively, $\left[\sigma^-,\sigma^+\right]\simeq 1$, and hence the atom behaves essentially like another harmonic oscillator. In this limit, all of the fields are coherent and we need only consider the mean field amplitudes, which obey the following equations of
motion:
\begin{widetext}
\begin{eqnarray}
\dot{\langle a\rangle} &=& -\left( \kappa + i\Delta_{\rm C} \right) \langle a\rangle -
ih\langle b\rangle - i\Ep - ig_{\rm tw}^\ast \langle \sigma^-\rangle ,
\\
\dot{\langle b\rangle} &=& -\left( \kappa + i\Delta_{\rm C} \right) \langle b\rangle -
ih\langle a\rangle - ig_{\rm tw} \langle \sigma^-\rangle ,
\\
\dot{\langle \sigma^-\rangle} &=&
-\left( \gamma/2 + i\Delta_{\rm A} \right) \langle \sigma^-\rangle -
ig_{\rm tw}\langle a\rangle - ig_{\rm tw}^\ast \langle b\rangle .
\end{eqnarray}
The steady state solutions for the cavity field amplitudes are
\begin{eqnarray}
\langle a\rangle_{\rm ss} &=& i\Ep\,
\frac{\left(\gamma/2+i\Delta_{\rm A}\right)\left[\left(\kappa+i\Delta_{\rm C}\right)\left(\gamma/2+i\Delta_{\rm A}\right)
+|g_{\rm tw}|^2\right]}
{\left[ ih\left(\gamma/2+i\Delta_{\rm A}\right)+{g_{\rm tw}^\ast}^2\right]
\left[ ih\left(\gamma/2+i\Delta_{\rm A}\right)+g_{\rm tw}^2\right]-
\left[\left(\kappa+i\Delta_{\rm C}\right)\left(\gamma/2+i\Delta_{\rm A}\right)+|g_{\rm tw}|^2\right]^2} \, ,
\label{eq:a_ss}
\\
\langle b\rangle_{\rm ss} &=&
- \frac{ih\left(\gamma/2+i\Delta_{\rm A}\right)+g_{\rm tw}^2}
{\left(\kappa+i\Delta_{\rm C}\right)\left(\gamma/2+i\Delta_{\rm A}\right)+|g_{\rm tw}|^2}\, \langle a\rangle_{\rm ss} ,
\label{eq:b_ss}
\end{eqnarray}
\end{widetext}
and for the amplitudes of the cavity output fields we have
\begin{eqnarray}
\langle a_{\rm out,ex}\rangle_{\rm ss} &=& \frac{i\Ep}{\sqrt{2\kappa_{\rm ex}}} +
\sqrt{2\kappa_{\rm ex}}\,\langle a\rangle_{\rm ss} \, ,
\\
\langle b_{\rm out,ex}\rangle_{\rm ss} &=& \sqrt{2\kappa_{\rm ex}}\,\langle b\rangle_{\rm ss} \, .
\end{eqnarray}
\subsection{No atom}
In the absence of an atom ($g_{\rm tw}=0$) linear analysis is exact (all fields are coherent) and
\begin{eqnarray}
\braket{a}_{\rm ss} &=& -\frac{i\Ep (\kappa +i\Dc)}{[\kappa+i(\Dc +h)][\kappa+i(\Dc -h)]} ,
\\
\braket{b}_{\rm ss} &=& - \frac{ih}{\kappa +i\Dc} \braket{a}_{\rm ss} .
\end{eqnarray}
For $\Delta_{\rm C}=\pm h$ one can then show that
\begin{eqnarray}
T_{\rm F} &=& \frac{(1-2\kappa_{\rm ex}/\kappa)^2+(4h^2/\kappa^2)(1-\kappa_{\rm ex}/\kappa)^2}{1+4h^2/\kappa^2} ,
\\
T_{\rm B} &=& \left( \frac{\kappa_{\rm ex}}{\kappa}\right)^2 \frac{4h^2/\kappa^2}{1+4h^2/\kappa^2} .
\end{eqnarray}
For $h=0$ (or $h\ll\kappa$) these reduce to (or are approximated by)
\begin{eqnarray}
T_{\rm F} &=& \left( 1 - 2\kappa_{\rm ex}/\kappa\right)^2 = \left( \frac{\kappa_{\rm i}-\kappa_{\rm ex}}{\kappa_{\rm i}+\kappa_{\rm ex}} \right)^2 ,
\\
T_{\rm B} &=& 0 .
\end{eqnarray}
Two cases are of particular interest to us. Firstly, critical coupling, for which $\kappa_{\rm ex}=\kex^{\rm cr}\equiv\sqrt{\kappa_{\rm i}^2+h^2}=\ki$ and hence $T_{\rm F}=0$. In this case, all of the incident light is dissipated through intrinsic loss in the microtoroid. Secondly, over-coupling, for which $\kappa_{\rm ex}\gg\{\kappa_{\rm i},h\}$, so $\kappa\simeq\kappa_{\rm ex}$ and $T_{\rm F}\simeq 1$; that is, virtually all of the light is simply transmitted along the fiber, past the microtoroid.
\subsection{Strongly-coupled atom}
Now consider the case in which an atom couples strongly and resonantly (i.e., $\Da =\Dc$) to the microtoroid field modes, such that $|\gtw |\gg\{\kappa,\gamma,h,|\Dc |\}$. Then from (\ref{eq:a_ss}) and (\ref{eq:b_ss}) one finds
\begin{eqnarray}
\braket{a}_{\rm ss} &\simeq& -\frac{i\Ep/2}{\kappa+i[\Dc -h\cos(2kx)]} ,
\\
\braket{b}_{\rm ss} &\simeq& - {\rm e}^{2ikx} \braket{a}_{\rm ss} ,
\end{eqnarray}
which in turn give
\begin{eqnarray}
\braket{a_{\rm out,ex}}_{\rm ss} &\simeq& \frac{i\Ep}{\sqrt{2\kex}} \left\{ 1 - \frac{\kex}{\kappa+i[\Dc -h\cos(2kx)]} \right\},
\\
\braket{b_{\rm out,ex}}_{\rm ss} &\simeq& \frac{i\Ep}{\sqrt{2\kex}} {\rm e}^{2ikx} \frac{\kex}{\kappa+i[\Dc -h\cos(2kx)]} .
\end{eqnarray}
So, for $\Dc =h\cos(2kx)$ one has
\begin{eqnarray}
T_{\rm F} &\simeq & (1-\kex /\kappa)^2 = \left( \frac{\kappa_{\rm i}}{\kappa_{\rm i}+\kappa_{\rm ex}} \right)^2,
\label{eq:TFatom}
\\
T_{\rm B} &\simeq & (\kex /\kappa)^2 = \left( \frac{\kappa_{\rm ex}}{\kappa_{\rm i}+\kappa_{\rm ex}} \right)^2.
\label{eq:TBatom}
\end{eqnarray}
Now, for critical coupling, and with $h\ll\ki$ (so $\kex\simeq\ki$), we find that $T_{\rm F}\simeq T_{\rm B}\simeq 1/4$, and the increase (from zero) in forward and backward fluxes is limited to 25\% of the incident flux \cite{Aoki06b,Junge13}. This situation is illustrated in Fig.~\ref{fig:spec_cc}, where $T_{\rm F}$ and $T_{\rm B}$ are plotted as a function of probe detuning $\Dc$ under the conditions of critical coupling and with either $\gtw=0$ or $\gtw$ large. For the example shown ($\gtw$ real) the atom couples only to normal mode $A$ and the consequent normal mode splitting leads to a substantial reduction in $T_{\rm F}$ at $\Dc\simeq\pm\sqrt{2}\gtw$, i.e., at the vacuum Rabi sidebands. However, the transmission spectrum around $\Dc=0$ is dominated by the resonance of the uncoupled normal mode $B$, and $T_{\rm F}(\Dc=0)$ increases from zero by only 0.25 in the presence of the atom. Under more general conditions (i.e., $\gtw$ complex) the system, in the linear regime, amounts to three coupled oscillators, but the three resonances remain at $\Dc\simeq\pm\sqrt{2}\gtw$and $\Dc\simeq 0$ (for $h\simeq 0$), and one still finds $T_{\rm F}(\Dc\simeq 0)\simeq 0.25$.
However, for strong over-coupling, $\kex\gg\ki$ (but, of course, with $\kex$ still much smaller than $|\gtw |$), it is apparent from (\ref{eq:TFatom}) and (\ref{eq:TBatom}) that $T_{\rm F}\simeq 0$ and $T_{\rm B}\simeq 1$ (cf. $T_{\rm F}\simeq 1$ and $T_{\rm B}\simeq 0$ with no atom). Hence, in this regime a single atom can act as a near-ideal switch for resonant or near-resonant light incident along the fiber. This is despite the presence of an intracavity normal mode that does not couple to the atom, which, as pointed out above, prevents this possibility in the case of critical coupling. In fact, if, for example (and without loss of generality), we consider the case in which $x=0$, so that the atom couples only to normal mode $A$, then the minimum in $T_{\rm F}$ actually occurs at $\Delta_{\rm C}=h$ (i.e., at $\omega_{\rm p}=\omega_{\rm C}-h$), which is the frequency of the {\em uncoupled} normal mode $B$.
The overcoupled case is illustrated in Fig.~\ref{fig:spec_oc} and, to help understand the behavior on resonance ($\Dc=0$) for the case considered ($h=0$), we also give in Table~\ref{tabl:ampls} a summary of the approximate intracavity and output field amplitudes when $\sin (kx)=0$ and $\Delta_{\rm C}=h$. We observe that the presence of a strongly-coupled atom ($|\gtw |$ large) causes a redistribution of the intracavity field between the counterpropagating modes $a$ and $b$, in particular reducing by a factor of 2 the amplitude $\braket{a}_{\rm ss}$. This reduction now yields near-ideal destructive interference in the output field $a_{\rm out,ex}$ between the components $-a_{\rm in,ex}$ and $\sqrt{2\kex} a$, and consequently the incident light field is reflected along the fiber.
\begin{figure}
\centering
\includegraphics[scale=0.45]{fig2.eps}
\caption{\label{fig:spec_cc} (Color online). Normalised power transmission $T_{\rm F}$ (blue) and reflection $T_{\rm B}$ (red) as a function of probe detuning in the weak-driving limit (using the linearised theory) for critical coupling ($\kex =\sqrt{\ki^2+h^2}$) and $\omega_{\rm A}=\omega_{\rm C}$. Parameters are $\{ \kex,\ki,h,\gamma\}/2\pi=\{10,10,0,5.2\}$~MHz. Solid lines are for $g_{\rm tw}/2\pi=100$~MHz ($\sin(kx)=0$) and dashed lines are for $g_{\rm tw}=0$. For large $\gtw$, the maximum increase in $T_{\rm F}(\Dc=0)$ is 0.25.}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.45]{fig3.eps}
\caption{\label{fig:spec_oc} (Color online). Normalised power transmission $T_{\rm F}$ (blue) and reflection $T_{\rm B}$ (red) as a function of probe detuning in the weak-driving limit (using the linearised theory) for strong overcoupling ($\kex\gg\ki,h$) and $\omega_{\rm A}=\omega_{\rm C}$. Parameters are $\{ \kex,\ki,h,\gamma\}/2\pi=\{20,0.2,0,5.2\}$~MHz. Solid lines are for $g_{\rm tw}/2\pi=100$~MHz ($\sin(kx)=0$) and dashed lines are for $g_{\rm tw}=0$. For large $\gtw$, the maximum change in $T_{\rm F}(\Dc=0)$ is close to 1.}
\end{figure}
\begin{table}
\begin{tabular}{c}
\begin{tabular}{c c c c c c c}
\multicolumn{7}{c}{Field amplitudes} \\
\hline
~ & ~ & ~ & $\gtw =0$ & ~ & ~ & $|\gtw |$ large \\
\hline
$\braket{a}_{\rm ss}$ & ~ & ~ & $-i\Ep /\kex$ & ~ & ~ & $-i\Ep /2\kex$ \\
$\braket{b}_{\rm ss}$ & ~ & ~ & $0$ & ~ & ~ & $i\Ep /2\kex$ \\
\hline
$\braket{A}_{\rm ss}$ & ~ & ~ & $-i\Ep /\sqrt{2}\kex$ & ~ & ~ & $0$ \\
$\braket{B}_{\rm ss}$ & ~ & ~ & $-i\Ep /\sqrt{2}\kex$ & ~ & ~ & $-i\Ep /\sqrt{2}\kex$ \\
\hline
$\braket{a_{\rm out,ex}}_{\rm ss}$ & ~ & ~ & $\braket{a_{\rm in,ex}}$ & ~ & ~ & $0$ \\
$\braket{b_{\rm out,ex}}_{\rm ss}$ & ~ & ~ & $0$ & ~ & ~ & $-\braket{a_{\rm in,ex}}$ \\
\hline
\end{tabular}
\end{tabular}
\caption{Summary of approximate intracavity and output field amplitudes for operation in the overcoupled regime, $\kex\gg\{\ki ,h\}$, with $\gtw=0$ or $|\gtw |\gg\{\kappa,\gamma,h\}$. We take $\omega_{\rm A}=\omega_{\rm C}$, $\sin (kx)=0$, and $\Dc=h$ (i.e., probe driving at the frequency of the uncoupled normal mode). Note that $\braket{a_{\rm in,ex}}=-i\Ep /\sqrt{2\kex}$.}
\label{tabl:ampls}
\end{table}
\section{Numerical solutions of the master equation}
In this section we present results obtained from numerical solutions of the full quantum master equation (\ref{eq:ME}) in the regime of strong overcoupling ($\kex\gg\{\ki ,h\}$). Our main aim is to determine requirements on the system, in particular with regards to atom-field coupling strength and probe driving strength, for efficient operation as a switch for incident light fields.
\subsection{Variation with atom-field coupling strength}
First, we examine the behavior on resonance ($\Dc=\Da=0$) as a function of the atom-field coupling strength $\gtw$. Results for the power transmission and reflection in the fiber, the intracavity field amplitudes, and the atomic excited state population are presented in Fig.~\ref{fig:varying_gtw} for relatively weak driving field strength. Good operation of the atom-microtoroid system as a ``mirror'' for incident light is already observed once $\gtw >\kex$, with $T_{\rm F}(\Dc=0)\simeq 0$, $T_{\rm B}(\Dc=0)>0.9$ (Fig.~\ref{fig:varying_gtw}(a)), and $g_{\rm BB}^{(2)}(0)\simeq 1$ (Fig.~\ref{fig:varying_gtw}(b)), indicating that the reflected light remains essentially coherent. The atomic excited state population (Fig.~\ref{fig:varying_gtw}(e)) is seen to be very small for sufficiently large $\gtw$, hence keeping the effects of atomic spontaneous emission small, while the intracavity field amplitudes (Fig.~\ref{fig:varying_gtw}(c,d)) for $\gtw >\kex$ closely match the approximate results of Table \ref{tabl:ampls}.
We note from Fig.~\ref{fig:varying_gtw}(b) that the transmitted field $a_{\rm out,ex}$, although generally very weak, exhibits strong bunching over much of the range of $\gtw$. In terms of intracavity normal modes, this field is given by $a_{\rm out,ex}=-a_{\rm in,ex}+\sqrt{\kex}(A+B)$. For $\gtw$ real, the field component $\sqrt{\kex}B$ does not couple to the atom and is coherent, with an amplitude that, for $\kex\simeq\kappa$, essentially cancels with that of the coherent input field component, $-a_{\rm in,ex}$. The field component $\sqrt{\kex}A$, for $\gtw\gg\{\kex,\gamma\}$, typically has an amplitude even smaller than $-a_{\rm in,ex}+\sqrt{\kex}B$, but for resonant driving as considered here ($\Dc =\Da =0$) is strongly bunched; in particular, while absorption of a single photon by mode $A$ is very small in this regime, if one is absorbed then the probability of a second absorption is significantly enhanced, and, consequently, extreme bunching occurs in the photon statistics \cite{Kubanek08,Faraon08}, i.e., $\braket{A^\dag A^\dag AA}\gg\braket{A^\dag A}^2$.
\begin{figure}
\centering
\includegraphics[scale=0.42]{fig4.eps}
\caption{\label{fig:varying_gtw} (Color online). (a) Normalised power transmission $T_{\rm F}$ (blue) and reflection $T_{\rm B}$ (red), (b) output photon correlation functions $g_{\rm FF}^{(2)}(0)$ (blue) and $g_{\rm BB}^{(2)}(0)$ (red), (c) intracavity field amplitudes $\braket{a}$ (blue) and $\braket{b}$ (red), (d) normal mode amplitudes $\braket{A}$ (blue) and $\braket{B}$ (red), and (e) atomic excited state population as a function of atom-field coupling strength $\gtw$ for strong overcoupling ($\kex\gg\ki,h$) and $\omega_{\rm A}=\omega_{\rm C}=\omega_{\rm p}$ ($\Dc =\Da =0$). Parameters are $\{\kex,\ki,h,\gamma,E_{\rm p}\}/2\pi=\{30,0.5,0,5.2,10\}$~MHz. The horizontal dashed lines in (c) show the approximations to $\braket{a}$ and $\braket{b}$ from Table~\ref{tabl:ampls} for the limit of large $\gtw$. The vertical dashed line in each plot indicates the value of $\kex$.}
\end{figure}
\subsection{Variation with probe driving strength: saturation behavior}
For sufficiently strong driving field strength the atom begins to saturate and the desired behavior just discussed starts to break down. This is illustrated in Fig.~\ref{fig:saturation}, where the key quantities of interest are plotted now as a function of probe strength $E_{\rm p}$ for several values of the atom-field coupling strength $\gtw$. With the onset of saturation, $T_{\rm F}(\Dc=0)$ ($T_{\rm B}(\Dc=0)$) (Fig.~\ref{fig:saturation}(a)) increases (decreases) steadily from zero (one). The probe strength at which this onset occurs increases approximately linearly with the value of $\gtw$; we investigate this further below.
As an aside, we note the interesting result that a small steady-state atomic inversion ($\braket{\sigma_+\sigma_-}>0.5$ or $\braket{\sigma_z}>0$) occurs over a range of $E_{\rm p}$ depending on the magnitude of $\gtw$ (Fig.~\ref{fig:saturation}(e)).
\begin{figure}
\centering
\includegraphics[scale=0.42]{fig5.eps}
\caption{\label{fig:saturation} (Color online). (a) Normalised power transmission $T_{\rm F}$ (blue) and reflection $T_{\rm B}$ (red), (b) output photon correlation functions $g_{\rm FF}^{(2)}(0)$ (blue) and $g_{\rm BB}^{(2)}(0)$ (red), (c) intracavity field amplitudes $\braket{a}$ (blue) and $\braket{b}$ (red), (d) normal mode amplitudes $\braket{A}$ (blue) and $\braket{B}$ (red), and (e) atomic excited state population as a function of probe strength $E_{\rm p}$ for strong overcoupling ($\kex\gg\ki,h$) and $\omega_{\rm A}=\omega_{\rm C}=\omega_{\rm p}$ ($\Dc =\Da =0$), with $\gtw /2\pi=50$~MHz (solid), $100$~MHz (dashed), and $150$~MHz (dot-dashed). Other parameters are $\{\kex,\ki,h,\gamma\}/2\pi=\{30,0.5,0,5.2\}$~MHz.}
\end{figure}
If we consider the case $\sin (kx)=0$ (as in Fig.~\ref{fig:saturation}), then the atom couples only to the normal mode $A=(a+b)/\sqrt{2}$. If we further assume that $h=0$ and $\Dc=\Da=0$, then one can derive in a semiclassical approximation (whereby, for example, we set $\braket{\sigma_+a}=\braket{\sigma_+}\braket{a}$ in the equations of motion for $\braket{\sigma_+}$, $\braket{\sigma_-}$, $\braket{a^\dagger}$, $\braket{a}$, and $\braket{\sigma_z}$) the following equation relating the driving field strength, $\Ep$, to the normal mode amplitude, $\braket{A}$:
\begin{align}
|Y| = |X| \left( 1 + \frac{4C}{1+2|X|^2} \right) .
\end{align}
Here,
\begin{align}
Y = \frac{i\Ep}{\kappa\sqrt{2n}} , ~~~ X = \frac{\braket{A}}{\sqrt{n}} , ~~~ n=\frac{\gamma^2}{8g_{\rm tw}^2} , ~~~ C=\frac{g_{\rm tw}^2}{\kappa\gamma} .
\end{align}
This is simply the equation of optical bistability for the normal mode $A$. Bistability curves, in terms of the normal mode amplitude, $|\braket{A}|$, and the driving field amplitude, $\Ep$, are shown in Fig.~\ref{fig:bistability}.
By setting $d|Y|/d|X|=0$ one can find the turning points of these curves. For $C\gg 1$ (as is the case for the situation we consider), these turning points are located at $|X|\simeq\sqrt{2C}$ and $|X|\simeq\sqrt{1/2}$, with corresponding driving field strengths
\begin{align}
\Ep = \sqrt{2\kappa\gamma} ~~~ {\rm and} ~~~ \Ep=g_{\rm tw} \sqrt{\frac{1}{8}}\, \frac{1+2C}{C} \simeq \frac{g_{\rm tw}}{\sqrt{2}} \, .
\end{align}
The value of $\Ep$ at the lower turning point, $\Ep\simeq g_{\rm tw}/\sqrt{2}$, agrees well with the value, observed in the numerical simulations of the full quantum model, at which the onset of saturation occurs, i.e., at which $\braket{A}$ (Fig.~\ref{fig:saturation}(d)) and $T_{\rm F}(\Dc=0)$ start to deviate significantly from zero.
\begin{figure}
\centering
\includegraphics[scale=0.42]{fig6.eps}
\caption{\label{fig:bistability} (Color online). Optical bistability curves for $\gtw /2\pi=50$~MHz (solid), $100$~MHz (dashed), and $150$~MHz (dot-dashed) with strong overcoupling ($\kex\gg\ki,h$) and $\omega_{\rm A}=\omega_{\rm C}=\omega_{\rm p}$ ($\Dc =\Da =0$). The inset shows the case $\gtw /2\pi=100$~MHz on a larger vertical scale. Other parameters are $\{\kex,\ki,h,\gamma\}/2\pi=\{30,0.5,0,5.2\}$~MHz.}
\end{figure}
Finally, to further investigate the effects of saturation, in Fig.~\ref{fig:spec_Ep} we plot the transmission and reflection as a function of probe detuning (Figs.~\ref{fig:spec_Ep}(a,b)), as well as the output photon correlation functions (Figs.~\ref{fig:spec_Ep}(c,d)), for several values of the probe strength, with $\gtw/2\pi = 150$~MHz. While properties associated with the vacuum Rabi sidebands around $\Dc\simeq\sqrt{2}\gtw$ are drastically affected by saturation, the resonance feature centered at $\Dc=0$ (and central to the operation of the system as a switch for resonant light) is very robust to increasing probe strength.
\begin{figure}
\centering
\includegraphics[scale=0.46]{fig7.eps}
\caption{\label{fig:spec_Ep} (Color online). Normalised power (a) transmission $T_{\rm F}$ and (b) reflection $T_{\rm B}$, and output photon correlation functions (c) $g_{\rm FF}^{(2)}(0)$ and (d) $g_{\rm BB}^{(2)}(0)$, for $\gtw /2\pi=150$~MHz with strong overcoupling ($\kex\gg\ki,h$), $\omega_{\rm A}=\omega_{\rm C}$, and $\Ep /2\pi =10$~MHz (solid), $50$~MHz (dashed), and $100$~MHz (dot-dashed). Other parameters are $\{\kex,\ki,h,\gamma\}/2\pi=\{30,0.5,0,5.2\}$~MHz.}
\end{figure}
\section{Coherent-state pulse reflection}
As shown above, for strong atom-field coupling (i.e., for $C\gg 1$), we find a maximum driving field strength of $\Ep\simeq g_{\rm tw}/\sqrt{2}$ before saturation effects start to occur. The incident photon flux,
\begin{align}
{\cal F} = \frac{\Ep^2}{2\kappa_{\rm ex}} ,
\end{align}
is therefore restricted to a maximum value
\begin{align}
{\cal F}_{\rm sat} \simeq \frac{g_{\rm tw}^2}{4\kappa_{\rm ex}} ,
\end{align}
if saturation effects are to be avoided. If we take, for example, $g_{\rm tw}=2\pi\cdot 100$~MHz and $\kappa_{\rm ex}=2\pi\cdot 30$~MHz, then ${\cal F}_{\rm sat}\simeq 520~(\mu {\rm s})^{-1}$, corresponding to a maximum input power, at 852~nm, of ${\cal P}_{\rm sat}\simeq 0.12$~nW.
Now consider an incident Gaussian pulse with a maximum photon flux of ${\cal F}_{\rm max}$, i.e., an incident flux profile of the form
\begin{align}
{\cal F}(t) = {\cal F}_{\rm max} \exp \left( -\frac{t^2}{2t_{\rm p}^2} \right) ,
\end{align}
where $2.35t_{\rm p}$ gives the FWHM of the pulse. If we consider the transmission and reflection spectra of Fig.~\ref{fig:spec_Ep}, then the width of the central resonance is approximately $2\kappa\simeq 2\kappa_{\rm ex}$. The frequency bandwidth of the incident pulse should be much less than this width; that is, we require
\begin{align}
\frac{2.35}{t_{\rm p}} \ll 2\kappa_{\rm ex} ,
\end{align}
which, for $\kappa_{\rm ex}=2\pi\cdot 30$~MHz, gives $t_{\rm p}\gg 6.2~{\rm ns}$. Choosing $t_{\rm p}=310~{\rm ns}$ and ${\cal F}_{\rm max}={\cal F}_{\rm sat}/2\simeq 260~(\mu {\rm s})^{-1}$, the (average) total photon number in the pulse is
\begin{align}
\bar{N}_{\rm p} = \int_{-\infty}^\infty {\cal F}(t)\, dt = {\cal F}_{\rm max} \sqrt{2\pi}\, t_{\rm p} \simeq 200 .
\end{align}
This suggests that our single-atom system could act effectively as a ``switch'' for an incident light pulse of significant photon number.
This in turn raises interesting possibilities for the generation, via the intrinsically quantum nature of the single-atom switch, of entangled states of an atom and propagating, coherent-state light fields, which we consider in the following section.
Before doing so, our scheme for pulse reflection should be contrasted with schemes based upon a single two-level emitter or atom coupled strongly to light fields propagating in a waveguide (e.g., nanowire or fiber) \cite{Chang07,Aoki09}. In these cases, the effective output field in the forward direction takes the form $a_{\rm out,ex}=-a_{\rm in,ex}+\sqrt{\Gamma}\sigma_-$, where $\Gamma$ is the Purcell-effect-enhanced spontaneous emission rate of the emitter. Given the two-level nature of the emitter, efficient reflection of light can only occur ``one photon at a time,'' and saturation of the emitter, marking the onset of an increase in transmission, corresponds to a single photon within a pulse of duration $\Gamma^{-1}$
(or, alternatively, within the effective bandwidth of the device) \cite{Carmichael93,Chang07}.
In the present scheme, however, saturation corresponds to a photon number $\sqrt{2\pi} {\cal F}_{\rm sat}/\kappa_{\rm ex}\sim (\gtw /\kappa_{\rm ex})^2$ within a (Gaussian) pulse of duration $\kappa_{\rm ex}^{-1}$ (the inverse of the effective bandwidth of our device), which may clearly be much larger than one and is not limited by the two-level nature of the atom.
That this is possible is evident from consideration of the output field for our scheme, written in the form $a_{\rm out,ex}=-a_{\rm in,ex}+\sqrt{\kex}(A+B)$. An atom coupled strongly to normal mode $A$ takes this mode far out of resonance, reducing its amplitude essentially to zero and the output field to $a_{\rm out,ex}\simeq -a_{\rm in,ex}+\sqrt{\kex}B$, where $B$ is the orthogonal, uncoupled normal mode, which for coherent input fields is itself simply a coherent field. The amplitude of $B$ is not constrained and is such as to give near ideal destructive interference in the output field $a_{\rm out,ex}$ (assuming, of course, operation within the saturation restrictions of the system).
\section{Entangled-path coherent-state preparation}
Let $\ket{\rm g}$ and $\ket{\rm e}$ denote the atomic states coupled via the microtoroid field modes, and let $\ket{\rm g'}$ denote an auxiliary internal atomic state (for example, in another hyperfine level) that does not couple to the field modes. If the atom is prepared in a superposition of $\ket{\rm g}$ and $\ket{\rm g'}$, and a coherent state pulse, denoted by $\ket{\alpha}_{a_{\rm in,ex}}$, is incident along the fiber (in input field $a_{\rm in,ex}$), then under ideal operation of the quantum switch, the following transformation would be implemented,
\begin{align}
&\ket{\alpha}_{a_{\rm in,ex}} \ket{0}_{b_{\rm in,ex}} \frac{1}{\sqrt{2}}\left( \ket{\rm g}+\ket{\rm g'} \right) \nonumber
\\
& ~~~\longrightarrow \ket{\Phi} = \frac{1}{\sqrt{2}} \ket{0}_{a_{\rm out,ex}} \ket{-\alpha}_{b_{\rm out,ex}} \ket{\rm g} \nonumber
\\
& ~~~~~~~~~~~~~~~~~~~+\frac{1}{\sqrt{2}} \ket{\alpha}_{a_{\rm out,ex}} \ket{0}_{b_{\rm out,ex}} \ket{\rm g'} ,
\end{align}
where $\ket{0}_{b_{\rm in}}$ denotes a vacuum state in $b_{\rm in}$. That is, the atomic state becomes entangled with a coherent state pulse (of potentially quite large photon number, as shown in the previous section) propagating in either the forward or backward direction along the fiber. With a suitable (unitary) rotation and projective measurement of the atomic state, this system could therefore be used to prepare entangled coherent states, which are of significant interest in the contexts of, e.g., quantum information processing and quantum metrology
\cite{Ourjoumtsev09,Sanders12,Zhang13,Knott14}. However, such states are by nature extremely sensitive (i.e., fragile) to uncontrolled losses, which in the present system arise from intrinsic losses (i.e., absorption) in the resonator modes and from atomic spontaneous emission into free space. More detailed investigation of the effect of these losses, as well as the influence of key parameter values, on entangled-state preparation is therefore important and now follows.
\subsection{Input pulse}
We consider a coherent-state, Gaussian optical pulse incident along the fiber, which we describe by \cite{Wang05}
\begin{widetext}
\begin{align}
\ket{\alpha}_{a_{\rm in,ex}} = \exp \left( -|\alpha|^2/2\right) \exp\left[ \int_{-\infty}^\infty \braket{a_{\rm in,ex}(t)} a_{\rm in,ex}^\dagger (t) dt\right] \ket{0}_{a_{\rm in,ex}} ,
\end{align}
\end{widetext}
where
\begin{align}
\braket{a_{\rm in,ex}(t)} &= - \frac{i\Ep}{\sqrt{2\kex}} \exp \left( - \frac{t^2}{4\tp^2} \right),
\end{align}
or, equivalently,
\begin{align}
\braket{\tilde{a}_{\rm in,ex}(\omega)} &= \frac{1}{\sqrt{2\pi}} \int_{-\infty}^\infty {\rm e}^{-i\omega t} \braket{a_{\rm in,ex}(t)} dt \nonumber
\\
&= - \frac{i\Ep\tp}{\sqrt{\kex}} \exp \left( -\omega^2\tp^2 \right) ,
\end{align}
and
\begin{align}
\int_{-\infty}^\infty |\braket{a_{\rm in,ex}(t)}|^2 dt &= \int_{-\infty}^\infty |\braket{\tilde{a}_{\rm in,ex}(\omega)}|^2 d\omega \nonumber
\\
&= \bar{N}_{\rm p} \equiv |\alpha|^2 .
\end{align}
\subsection{Linear (coherent state) analysis}
If we assume operation of the system in the linear regime, i.e., that the driving strength is sufficiently small ($\Ep <\gtw/\sqrt{2}$), then we can solve operator equations of motion for the resonator and atomic modes exactly and derive expressions for all of the output field operators (in frequency space) in terms of $\tilde{a}_{\rm in,ex}(\omega)$ (Appendix A). Furthermore, in a linear system all fields remain coherent, so, for the atom in state $\ket{\rm g}$, we can write the output state of our system as a product of coherent states in each of the possible output channels \cite{CohStateApprox}, i.e.,
\begin{align}
\ket{\psi} = \ket{\alpha_{\rm ex}}_{a_{\rm out,ex}} \ket{\alpha_{\rm i}}_{a_{\rm out,i}} \ket{\beta_{\rm ex}}_{b_{\rm out,ex}} \ket{\beta_{\rm i}}_{b_{\rm out,i}} \ket{\eta}_{c_{\rm out}} ,
\end{align}
with, for example (see Appendix A),
\begin{widetext}
\begin{align}
\ket{\alpha_{\rm ex}}_{a_{\rm out,ex}} = \exp \left( -|\alpha_{\rm ex}|^2/2\right) \exp\left[ \int_{-\infty}^\infty \braket{a_{\rm out,ex}(t)} a_{\rm out,ex}^\dagger (t) dt\right] \ket{0}_{a_{\rm out,ex}} ,
\end{align}
\end{widetext}
where
\begin{align}
\braket{a_{\rm out,ex}(t)} &= \frac{1}{\sqrt{2\pi}} \int_{-\infty}^\infty {\rm e}^{i\omega t} \braket{\tilde{a}_{\rm out,ex}(\omega)} d\omega
\\
&= \frac{1}{\sqrt{2\pi}} \int_{-\infty}^\infty {\rm e}^{i\omega t} t_{\rm ex}(\omega)\braket{\tilde{a}_{\rm in,ex}(\omega)} d\omega ,
\end{align}
\begin{align}
|\alpha_{\rm ex}|^2 = \int_{-\infty}^\infty |t_{\rm ex}(\omega)|^2 |\braket{\tilde{a}_{\rm in,ex}(\omega)}|^2 d\omega ,
\end{align}
and $t_{\rm ex}(\omega)$ is also given in Appendix A.
Note that $c_{\rm out}$ denotes the output field from the atom into free space (spontaneous emission).
Similarly, for the atom in state $\ket{\rm g'}$ we can write
\begin{align}
\ket{\psi^\prime} = \ket{\alpha_{\rm ex}^{(0)}}_{a_{\rm out,ex}} \ket{\alpha_{\rm i}^{(0)}}_{a_{\rm out,i}} \ket{\beta_{\rm ex}^{(0)}}_{b_{\rm out,ex}} \ket{\beta_{\rm i}^{(0)}}_{b_{\rm out,i}} \ket{\eta^{(0)}}_{c_{\rm out}} ,
\end{align}
where now, for example,
\begin{align}
|\alpha_{\rm ex}^{(0)}|^2 = \int_{-\infty}^\infty |t_{\rm ex}^{(0)}(\omega)|^2 |\braket{\tilde{a}_{\rm in,ex}(\omega)}|^2 d\omega ,
\end{align}
with
\begin{align}
t_{\rm ex}^{(0)}(\omega) = \left. t_{\rm ex}(\omega)\right|_{\gtw=0} .
\end{align}
Hence, for the atom in an equal superposition of $\ket{\rm g}$ and $\ket{\rm g'}$, the output state of the system is given by
\begin{align}
\ket{\Psi} = \frac{1}{\sqrt{2}} \left( \ket{\rm g}\ket{\psi} + \ket{\rm g^\prime}\ket{\psi^\prime} \right) .
\end{align}
\subsection{Reduced density operator}
Tracing over the output fields $a_{\rm out,i}$, $b_{\rm out,i}$, and $c_{\rm out}$, we obtain a reduced density operator for the fiber output fields (and atom) only, which takes the form
\begin{align}
\rho &=\frac{1}{2} \ket{\rm g} \ket{\alpha_{\rm ex}} \ket{\beta_{\rm ex}} \bra{\beta_{\rm ex}} \bra{\alpha_{\rm ex}} \bra{\rm g} \nonumber
\\
&+\,\frac{1}{2} \ket{\rm g^\prime} \ket{\alpha_{\rm ex}^{(0)}} \ket{\beta_{\rm ex}^{(0)}} \bra{\beta_{\rm ex}^{(0)}} \bra{\alpha_{\rm ex}^{(0)}} \bra{\rm g^\prime} \nonumber
\\
&+\, \frac{\xi}{2} \ket{\rm g} \ket{\alpha_{\rm ex}} \ket{\beta_{\rm ex}} \bra{\beta_{\rm ex}^{(0)}} \bra{\alpha_{\rm ex}^{(0)}} \bra{\rm g^\prime} \nonumber
\\
&+\, \frac{\xi^\ast}{2} \ket{\rm g^\prime} \ket{\alpha_{\rm ex}^{(0)}} \ket{\beta_{\rm ex}^{(0)}} \bra{\beta_{\rm ex}} \bra{\alpha_{\rm ex}} \bra{\rm g} ,
\end{align}
where, for brevity, we have dropped the output field subscripts; states associated with output channels $a_{\rm out,ex}$ and $b_{\rm out,ex}$ can be distinguished by the use of $\{\alpha_{\rm ex},\alpha_{\rm ex}^{(0)}\}$ or $\{\beta_{\rm ex},\beta_{\rm ex}^{(0)}\}$, respectively. The parameter $\xi$ appearing in the reduced density operator is given by
\begin{align}
\xi = \braket{\alpha_{\rm i}^{(0)}|\alpha_{\rm i}}_{a_{\rm out,i}} \braket{\beta_{\rm i}^{(0)}|\beta_{\rm i}}_{b_{\rm out,i}} \braket{\eta^{(0)}|\eta}_{c_{\rm out}} ,
\end{align}
where one can show (noting that, in fact, $\beta_{\rm i}^{(0)} =0$ for $h=0$, and $\eta^{(0)}=0$) that
\begin{widetext}
\begin{align}
\braket{\alpha_{\rm i}^{(0)}|\alpha_{\rm i}}_{a_{\rm out,i}} &= \exp \left[ -\frac{1}{2} \int_{-\infty}^\infty \left( |t_{\rm i}(\omega)|^2 + |t_{\rm i}^{(0)}(\omega)|^2 - 2t_{\rm i}^{(0)}(\omega)^\ast t_{\rm i}(\omega) \right) |\braket{\tilde{a}_{\rm in,ex}(\omega)}|^2 d\omega \right] ,
\\
\braket{\beta_{\rm i}^{(0)}|\beta_{\rm i}}_{b_{\rm out,i}} &=\exp \left[ -\frac{1}{2} \int_{-\infty}^\infty |r_{\rm i}(\omega)|^2 |\braket{\tilde{a}_{\rm in,ex}(\omega)}|^2 d\omega \right] \equiv \exp \left[ -\frac{1}{2} |\beta_{\rm i}|^2 \right] ,
\\
\braket{\eta^{(0)}|\eta}_{c_{\rm out}} &= \exp \left[ -\frac{1}{2} \int_{-\infty}^\infty |s(\omega)|^2 |\braket{\tilde{a}_{\rm in,ex}(\omega)}|^2 d\omega \right] \equiv \exp \left[ -\frac{1}{2} |\eta|^2 \right] ,
\end{align}
\end{widetext}
with $t_{\rm i}(\omega)$, $r_{\rm i}(\omega)$, and $s(\omega)$ given in Appendix A. Note that these are overlap factors between the states of these output channels with the atom either in state $\ket{\rm g}$ or state $\ket{\rm g'}$. The smaller these factors are, the more ``distinguishable'' are the components of the desired entangled state from information carried by these channels, and the worse the expected fidelity of preparation of the entangled coherent state (i.e., the reduced density operator approaches a mixed state in the limit $\xi\rightarrow 0$).
\subsection{Fidelity of entangled state preparation}
Finally, the fidelity of preparation of the ideal output state $\ket{\Phi}$ is given by
\begin{align}
F &= \bra{\Phi} \rho \ket{\Phi} \nonumber
\\
&= \frac{1}{4} |\braket{0|\alpha_{\rm ex}}|^2 |\braket{-\alpha |\beta_{\rm ex}}|^2 + \frac{1}{4} |\braket{\alpha |\alpha_{\rm ex}^{(0)}}|^2 |\braket{0 |\beta_{\rm ex}^{(0)}}|^2 \nonumber
\\
& +\, \frac{\xi}{4} \braket{0|\alpha_{\rm ex}} \braket{-\alpha |\beta_{\rm ex}} \braket{\alpha_{\rm ex}^{(0)} |\alpha} \braket{\beta_{\rm ex}^{(0)}|0} \nonumber
\\
& +\, \frac{\xi^\ast}{4} \braket{\alpha_{\rm ex}|0} \braket{\beta_{\rm ex}|-\alpha} \braket{\alpha |\alpha_{\rm ex}^{(0)}} \braket{0|\beta_{\rm ex}^{(0)}} ,
\label{eq:F}
\end{align}
where
\begin{align}
\braket{0|\alpha_{\rm ex}} &= \exp\left( -\frac{1}{2} |\alpha_{\rm ex}|^2 \right) ,
\\
\braket{0 |\beta_{\rm ex}^{(0)}} &= \braket{0|0} =1 ,
\end{align}
and
\begin{widetext}
\begin{align}
\braket{-\alpha |\beta_{\rm ex}} &= \exp \left[ -\frac{1}{2} \int_{-\infty}^\infty \left( 1 + |r_{\rm ex}(\omega)|^2 + 2r_{\rm ex}(\omega) \right) |\braket{\tilde{a}_{\rm in,ex}(\omega)}|^2 d\omega \right] ,
\\
\braket{\alpha |\alpha_{\rm ex}^{(0)}} &= \exp \left[ -\frac{1}{2} \int_{-\infty}^\infty \left( 1 + |t_{\rm ex}^{(0)}(\omega)|^2 - 2t_{\rm ex}^{(0)}(\omega) \right) |\braket{\tilde{a}_{\rm in,ex}(\omega)}|^2 d\omega \right] ,
\end{align}
\end{widetext}
with $r_{\rm ex}(\omega)$ given in Appendix A.
\begin{figure}[t]
\centering
\includegraphics[scale=0.46]{fig8.eps}
\caption{\label{fig:F_g} (Color online). Fidelity, $F$, of entangled-path coherent state preparation as a function of the mean photon number, $|\alpha|^2$, in the incident pulse for $\gtw /2\pi =\{50,100,150\}$~MHz (curves are labelled by these values in units of MHz). Dashed lines are the approximate expression (\ref{eq:F_approx}). Other parameters are $\{\kex,\ki,h,\gamma\}/2\pi=\{30,0.5,0,5.2\}$~MHz, $t_{\rm p}=318$~ns ($\kappa t_{\rm p}\simeq 60$), and $\omega_{\rm A}=\omega_{\rm C}$.}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.46]{fig9.eps}
\caption{\label{fig:F_ki} (Color online). Fidelity, $F$, of entangled-path coherent state preparation as a function of the mean photon number, $|\alpha|^2$, in the incident pulse for $\gtw /2\pi =100$~MHz, with $\ki /2\pi =\{ 0.25,0.5,1.0,2.0\}$~MHz (curves are labelled by these values in units of MHz). Dashed lines are the approximate expression (\ref{eq:F_approx}). Other parameters are $\{\kex,h,\gamma\}/2\pi=\{50,0,5.2\}$~MHz, $t_{\rm p}=159$~ns ($\kappa t_{\rm p}\simeq 50$), and $\omega_{\rm A}=\omega_{\rm C}$.}
\end{figure}
We evaluate the fidelity, (\ref{eq:F}), numerically and present some results in Figs.~\ref{fig:F_g} and \ref{fig:F_ki}. In particular, we plot $F$ as a function of the mean photon number in the input pulse, $|\alpha|^2$, firstly in Fig.~\ref{fig:F_g} for several values of the atom-field coupling strength, $\gtw$, and secondly in Fig.~\ref{fig:F_ki} for a range of values of the intrinsic cavity mode loss rate, $\ki$. Considering Fig.~\ref{fig:F_g}, we see that an entangled coherent state with pulses of mean photon number $|\alpha|^2\simeq 10$ could in principle be achieved with fidelity in excess of 85\% for $\gtw /2\pi \gtrsim 50$~MHz and $\ki /2\pi \simeq 0.5$~MHz. With larger values of $\gtw$ and/or smaller values of $\ki$, similar fidelities appear possible with pulses of even larger mean photon numbers, as illustrated in Figs.~\ref{fig:F_g} and \ref{fig:F_ki}.
Implicit in the achievement of high fidelity state preparation is good matching, or overlap, between input and output pulses. An example of these pulses is shown in Fig.~\ref{fig:pulses}, where we plot the photon fluxes $|\braket{a_{\rm in,ex}(t)}|^2$, $|\braket{a_{\rm out,ex}^{(0)}(t)}|^2$, and $|\braket{b_{\rm out,ex}(t)}|^2$, where $a_{\rm out,ex}^{(0)}(t)$ denotes the output field for the atom in state $\ket{{\rm g}'}$. For the parameters used, the output pulses overlap extremely well with eachother and with the input pulse; the fidelity $F\simeq 0.85$ for this case. Note that on the scale used in Fig.~\ref{fig:pulses}, $|\braket{a_{\rm out,ex}(t)}|^2$ (not shown) is virtually indistinguishable from zero for all $t$ (and $|\alpha_{\rm ex}|^2=0.0017$).
\begin{figure}[t]
\centering
\includegraphics[scale=0.46]{fig10.eps}
\caption{\label{fig:pulses} (Color online). Input photon flux, $|\braket{a_{\rm in,ex}(t)}|^2$ (blue, short-dashed curve), and output photon fluxes, $|\braket{a_{\rm out,ex}^{(0)}(t)}|^2$ (blue, long-dashed curve) and $|\braket{b_{\rm out,ex}(t)}|^2$ (red solid curve) as a function of time, for
parameters $\{\gtw,\kex,\ki,h,\gamma\}/2\pi=\{100,50,0.5,0,5.2\}$~MHz, $\Ep /2\pi =28$~MHz, and $t_{\rm p}=159$~ns ($\kappa t_{\rm p}\simeq 50$), with $\omega_{\rm A}=\omega_{\rm C}$. The mean photon number in the input pulse (i.e., the area under the dashed curve) is $|\alpha|^2=20$. The output pulses have mean photon numbers $|\alpha_{\rm ex}^{(0)}|^2=19.2$ and $|\beta_{\rm ex}|^2=19.3$, respectively.}
\end{figure}
Provided the bandwidth of the incident pulse is sufficiently narrow, i.e., $\kappa\tp\gg 1$, then the various integrals appearing in the expression for $F$ can be simplified; for example, we may set $|\alpha_{\rm ex}|^2 \simeq |t_{\rm ex}(0)|^2 \int_{-\infty}^\infty |\braket{\tilde{a}_{\rm in,ex}(\omega)}|^2 d\omega = |t_{\rm ex}(0)|^2|\alpha|^2$, and it is possible to derive the following approximate result for the fidelity,
\begin{align}
F & \simeq \frac{1}{4} \exp\left( -\Gamma_1 |\alpha|^2\right) + \frac{1}{4} \exp\left( -\Gamma_2 |\alpha|^2\right) \nonumber
\\
&\, +\, \frac{1}{2} \exp\left[ -\frac{1}{2}(\Gamma_1 +\Gamma_2 + 2\Gamma_3)|\alpha|^2 \right] ,
\label{eq:F_approx}
\end{align}
with
\begin{align}
\Gamma_1 &= \left( 1 - \frac{\kex}{\kappa}\frac{4C+2}{4C+1} \right)^2 + \left( 1 - \frac{\kex}{\kappa}\frac{4C}{4C+1} \right)^2 ,
\\
\Gamma_2 &= 4\left(\frac{\ki}{\kappa}\right)^2 ,
\\
\Gamma_3 &= \frac{\kex}{\kappa} \left( \frac{4C}{4C+1} \right)^2 \left( \frac{\ki}{\kappa} + \frac{1}{4C} \right) .
\end{align}
The result (\ref{eq:F_approx}) is also plotted in Figs.~\ref{fig:F_g} and \ref{fig:F_ki} and shows good agreement with the full expression for the chosen values of the pulse length $\tp$.
Note that $\xi\simeq\exp (-\Gamma_3|\alpha |^2)$ is typically the key factor in determining $F$ for the entangled state preparation since, for $C\gg 1$, one has $\Gamma_1\simeq 2(\ki /\kappa)^2$ and $\Gamma_2\simeq 4(\ki /\kappa)^2$, while $\Gamma_3\simeq\kex\ki /\kappa^2\simeq\ki /\kappa\gg\Gamma_1,\Gamma_2$ for strong overcoupling ($\ki /\kappa\ll 1$). Hence, if $\Gamma_{1,2}|\alpha|^2\ll 1$ then one can write $F\simeq\frac{1}{2}(1+{\rm e}^{-\Gamma_3|\alpha|^2})$, which, e.g., agrees reasonably well with the curve for $\gtw/2\pi =150$~MHz in Fig.~\ref{fig:F_g}.
In comparison, the fidelity for merely ideal reflection of a coherent state pulse by an atom in state $\ket{\rm g}$,
\begin{align}
F_{\rm refl} = |\braket{0|\alpha_{\rm ex}}|^2 |\braket{-\alpha |\beta_{\rm ex}}|^2 ,
\end{align}
is (in the same limits) given by $\exp (-\Gamma_1|\alpha |^2)$. Taking $\Gamma_1\simeq 2(\ki /\kappa)^2$ and the parameters of Fig.~\ref{fig:F_g} ($\ki/\kappa=0.016$), this gives $F_{\rm refl}\gtrsim 0.97$ for $|\alpha|^2=50$. This highlights the relative fragility of the entangled-path coherent state and makes explicit the requirements on the operating regime of the system for high-fidelity preparation of such states; in particular, one requires $\ki/\kex\ll 1$ and $C=\gtw^2/\kappa\gamma\gg 1$. These conditions should indeed be achievable with microtoroid cavity QED.
\section{Conclusion}
To summarize, we have presented a scheme for a single-atom quantum switch for coherent light fields based upon microtoroid cavity QED in a regime not previously considered, at least not in this particular context. It complements recent schemes employing TM-polarized fields by also enabling, in principle, 100\% contrast in transmission and reflection of TE-polarized fields depending on the presence or absence of a strongly-coupled atom. The scheme can potentially work for quite strong incident fields, allowing, for example, control of a coherent state pulse of significant photon number and high-fidelity preparation of entangled-path coherent states.
\begin{acknowledgments}
SP acknowledges support from the International Central Networks Fund of the University of Auckland and from the Marsden Fund of the Royal Society of New Zealand. TA acknowledges support from JSPS KAKENHI Grant Number 26707022, MATSUO Foundation, and The Murata Science Foundation.
\end{acknowledgments}
|
\section{Introduction}
As it is well known (see, e.\,g., \cite{jac,scha}), the Cayley-Dickson algebras represent a nested sequence $A_0, A_1, A_2,\ldots, A_N, \ldots$ of $2^N$-dimensional (in general non-associative) $\mathbb{R}$-algebras with $A_N \subset A_{N+1}$, where $A_0 = \mathbb{R}$ and where for any $N \geq 0$ $A_{N+1}$ comprises all ordered pairs of elements from $A_{N}$ with conjugation defined by
\begin{equation}
(x,y)^{\ast} = (x^{\ast}, -y)
\label{eq1}
\end{equation}
and multiplication usually by
\begin{equation}
(x,y)(X,Y) = (xX - Yy^{\ast}, x^{\ast}Y + Xy).
\label{eq2}
\end{equation}
Every finite-dimensional algebra (see, e.\,g., \cite{scha}) is basically defined by the multiplication rule of its basis.
The basis elements (or units) $e_0, e_1, e_2, \ldots, e_{2^{N+1} -1}$ of $A_{N+1}$, $e_0$ being the real basis element (identity), can be chosen in various ways. Our preference is the {\it canonical} basis
\begin{align*}
e_0 &= (e_0,0), & e_1 &= (e_1,0), & e_2 &= (e_2,0), && \ldots, && e_{2^N - 1} &= (e_{2^N - 1},0), &~~~~~& \\
e_{2^N} &= (0,e_0), & e_{2^N+1} &= (0,e_1), & e_{2^N+2} &= (0,e_2), && \ldots, && e_{2^{N+1} - 1} &= (0,e_{2^{N} - 1}),&~~~~~~&
\end{align*}
where, by abuse of notation (see, e.\,g., \cite{mor}), the same symbols are also used for the basis elements of $A_N$.
This is because the paper essentially focuses on {\it multiplication} properties of basis elements and the canonical basis seems to
display most naturally the inherent symmetry of this operation. For,
in addition to revealing the nature of the Cayley-Dickson recursive process, it also implies that for both $a$ and $b$ being non-zero we have
$e_a e_b = \pm e_{a \oplus b},$ where the symbol `$\oplus$' denotes `{\it exclusive or}' of the binary representations of $a$ and $b$ (see, e.\,g., \cite{arndt}).
From the above expressions and Eqs.~(\ref{eq1}) and (\ref{eq2}) one can readily find the product of any two distinct units of $A_{N+1}$
if the multiplication properties of those of $A_{N}$ are given.
Such products are usually expressed/presented in a tabular form, and we shall also follow this tradition here. All the multiplication tables we made use of were computed for us by J\"org Arndt; the computer program is named {\tt cayley-dickson-demo.cc} and is freely available from his web-site {\tt http://jjj.de/fxt/demo/arith/index.html}.
Employing such a multiplication table of $A_N$, $N \geq 2$, it can be verified that the $2^N-1$ imaginaries $e_a$, $1 \leq a \leq 2^N -1$, form ${2^N-1 \choose 2}/3$ distinguished sets each of which comprises three different units $\{e_a, e_b, e_c\}$ that satisfy equation
\begin{equation}
e_a e_b = \pm e_c,
\label{eq4}
\end{equation}
and where each unit is found to belong to $2^{N-1} - 1$ such sets.\footnote{Although these properties will explicitly be illustrated only for the cases $3 \leq N \leq 6$, due to the nested structure of the algebras they must be exhibited by any other higher-dimensional $A_N$.}
Regarding the imaginaries as points and their distinguished triples as lines, one gets a point-line incidence geometry where every line has three points and through each point there pass $2^{N-1} - 1$ lines and which is isomorphic to PG$(N-1,2)$, the $(N-1)$-dimensional projective space over the smallest Galois field $GF(2)$ (see also \cite{baez} for $N=3$ and \cite{culb} for $N=4$).
Let us assume, without loss of generality, that the elements in any distinguished triple $\{e_a, e_b, e_c\}$ of $A_N$ are ordered in such a way that $a < b < c$. Then, for $N \geq 3$, we can naturally speak about two different kinds of triples and, hence, two distinct kinds of lines of the associated $2^N$-nionic PG$(N-1,2)$, according as $a+b = c$ or $a+b \neq c$; in what follows a line of the former/latter kind will be called ordinary/defective. This stratification of the line-set of the PG$(N-1,2)$ induces a similar partition of the point-set of the latter space into several types, where a point of a given type is characterized by the same number of lines of either kind that pass through it.
Obviously, if our projective space PG$(N-1,2)$ is regarded as an {\it abstract} geometry {\it per se}, every point and/or every line in it has the same footing.
So, to account for the above-described `refinement' of the structure of our $2^N$-nionic PG$(N-1,2)$, it turns out to be necessary to find a representation of this space where each point/line is ascribed a certain `internal' structure, which at first sight may seem to be quite a challenging task.
To tackle this task successfully, we need to introduce a few notions/concepts from the realm of finite geometry.
We start with a finite {\it point-line incidence structure} $\mathcal{C} = (\mathcal{P},\mathcal{L},I)$ where $\mathcal{P}$ and $\mathcal{L}$ are, respectively, finite sets of points and lines and where incidence $I \subseteq \mathcal{P} \times \mathcal{L}$ is a binary relation indicating which point-line pairs are incident (see, e.\,g., \cite{shult}).
Here, we shall only be concerned with specific point-line incidence structures called {\it configurations} \cite{grun}. A $(v_r,b_k)$-configuration is a $\mathcal{C}$ where: 1) $v = \vert \mathcal{P} \vert$ and $b = \vert \mathcal{L}\vert$, 2) every line has $k$ points and every point is on $r$ lines, and 3) two distinct lines intersect in at most one point and every two distinct points are joined by at most one line; a configuration where $v=b$ and $r=k$ is called symmetric (or balanced), and usually denoted as a $(v_r)$-configuration.
A $(v_r,b_k)$-configuration with $v = {r+k-1 \choose r}$ and $b = {r+k-1 \choose k}$ is called a $binomial$ configuration.
Next it comes a {\it geometric hyperplane} of $\mathcal{C} = (\mathcal{P},\mathcal{L},I)$, which is a proper subset of $\mathcal{P}$ such that a line from $\mathcal{C}$ either lies fully in the subset, or shares with it only one point.
If $\mathcal{C}$ possesses geometric hyperplanes, then one can define the {\it Veldkamp space} of $\mathcal{C}$, $\mathcal{V}(\mathcal{C})$, as follows \cite{buec}: (i) a point of $\mathcal{V}(\mathcal{C})$ is a geometric hyperplane of $\mathcal{C}$
and (ii) a line of $\mathcal{V}(\mathcal{C})$ is the collection $H'H''$ of all geometric hyperplanes $H$ of $\mathcal{C}$ such that $H' \cap H'' = H' \cap H = H'' \cap H$ or $H = H', H''$, where $H'$ and $H''$ are distinct geometric hyperplanes. If each line of $\mathcal{C}$ has three points and $\mathcal{C}$ `behaves well,' a line of $\mathcal{V}(\mathcal{C})$ is also of size three and can equivalently be defined as $\{H', H'', \overline{H' \Delta H''}\}$, where the symbol $\Delta$ stands for the symmetric difference of the two geometric hyperplanes and an overbar denotes the complement of the object indicated.
From its definition it is obvious that $\mathcal{V}(\mathcal{C})$ is well suitable for our needs because its points, being themselves {\it sets} of points, have different `internal' structure and so, in general, they can no longer be on the same par; clearly, the same applies to the lines $\mathcal{V}(\mathcal{C})$. Our task thus basically boils down to finding such $\mathcal{C}_N$ whose $\mathcal{V}(\mathcal{C}_N)$ is isomorphic to PG$(N-1,2)$ and completely reproduces its $2^N$-nionic fine structure. This will be carried out in great detail for the first four non-trivial cases, $3 \leq N \leq 6$, which, when combined with the two trivial cases ($N=1,2$), will provide us with sufficient amount of information to guess a general pattern.
The paper is organized as follows. In Sect.~2 it is shown that ${\cal C}_3$ (octonions) is isomorphic to the Pasch $(6_2,4_3)$-configuration, which plays a key role in classifying Steiner triple systems. In Sect.~3 one demonstrates that ${\cal C}_4$ (sedenions) is nothing but the famous Desargues $(10_3)$-configuration. In Sect.~4 our ${\cal C}_5$ (32-nions) is shown to be identical with the Cayley-Salmon $(15_4,20_3)$-configuration found in the well-known Pascal mystic hexagram. In Sect.~5 we find that ${\cal C}_6$ corresponds to a particular $(21_5,35_3)$-configuration encompassing seven distinct copies of the Cayley-Salmon $(15_4,20_3)$-configuration as geometric hyperplanes. In Sect.~6 some rudimentary properties of the generic ${\cal C}_N \cong \left({N+1 \choose 2}_{N-1}, {N+1 \choose 3}_{3}\right)$-configuration are outlined and its isomorphism to a combinatorial Grassmannian of type $G_2(N+1)$ is conjectured. Finally, Sect.~7 is reserved for concluding remarks.
\section{Octonions and the Pasch $(6_2,4_3)$-configuration}
From the nesting property of the Cayley-Dickson construction of $A_N$ it is obvious that the smallest non-trivial case to be addressed is
$A_3$, the algebra of octonions, whose multiplication table is presented in Table 1.
\begin{table}[h]
\begin{center}
\caption{The multiplication table of the imaginary unit octonions $e_a$, $1 \leq a \leq 7$. For the sake of simplicity, in what follows we shall employ a short-hand notation $e_a \equiv a$; likewise for the real unit $e_0 \equiv 0$. There are also delineated multiplication tables corresponding to the distinguished nested sequence of sub-algebras of complex numbers ($a=1$, the upper left corner) and quaternions ($1 \leq a \leq 3$, the upper left $3 \times 3$ square).}
\vspace*{0.4cm}
\begin{tabular}{||r|r|rr|rrrr||}
\hline \hline
$*$ & 1 & 2 & 3 & 4 & 5 & 6 & 7 \\
\hline
1 & $-$0 & $-$3 & +2 & $-$5 & +4 & +7 & $-$6 \\
\hline
2 & +3 & $-$0 & $-$1 & $-$6 & $-$7 & +4 & +5 \\
3 & $-$2 & +1 & $-$0 & $-$7 & +6 & $-$5 & +4 \\
\hline
4 & +5 & +6 & +7 & $-$0 & $-$1 & $-$2 & $-$3 \\
5 & $-$4 & +7 & $-$6 & +1 & $-$0 & +3 & $-$2 \\
6 & $-$7 & $-$4 & +5 & +2 & $-$3 & $-$0 & +1 \\
7 & +6 & $-$5 & $-$4 & +3 & +2 & $-$1 & $-$0 \\
\hline \hline
\end{tabular}
\end{center}
\end{table}
\noindent
The above-given multiplication table implies the existence of the following seven {\it distinguished} trios of imaginary units:
$$
\{1,2,3\}, \{1,4,5\}, \{1,6,7\},
$$
$$
\{2,4,6\}, \{2,5,7\},
$$
$$
\{3,4,7\}, \{3,5,6\}.
$$
\noindent
Regarding the seven imaginary units as points and the seven distinguished triples of them as lines, we obtain a point-line incidence structure where each line has three points and, dually, each point is on three lines, and which is isomorphic to the smallest projective plane PG$(2,2)$, often called the Fano plane, depicted in Figure \ref{f1}.
\begin{figure}[t]
\centering
\includegraphics[width=4.0truecm]{f1.eps}
\caption{An illustration of the structure of PG$(2,2)$, the Fano plane, that provides the multiplication law for octonions (see, e.\,g., \cite{baez}). The points of the plane are seven small circles. The lines are represented by triples of circles located on the sides of the triangle, on its altitudes, and by the triple lying on the big circle. The three imaginaries lying on the same line satisfy Eq.\,(\ref{eq4}).}
\label{f1}
\end{figure}
It is then readily seen that we have six ordinary lines, namely
$$
\{1,2,3\}, \{1,4,5\}, \{1,6,7\},
$$
$$
\{2,4,6\}, \{2,5,7\},
$$
$$
\{3,4,7\},
$$
and only single defective one, viz.
$$
\{3,5,6\}.
$$
Similarly, our octonionic PG$(2,2)$ features two distinct types of points.
A type-one point is such that two lines passing through it are ordinary, the remaining one being defective; such a point lies in the set
$$\{3,5,6\} \equiv \alpha.$$
A type-two point is such that every line passing through it is ordinary; such a point belongs to the set
$$\{1,2,4,7\} \equiv \beta,$$
which is highlighted by gray color in Figure \ref{f1}.
A configuration $\mathcal{C}_3$ whose Veldkamp space reproduces the above-described partitions of points and lines of PG$(2,2)$ is, as we will soon see, nothing but the well-known {\it Pasch} $(6_2,4_3)$-configuration, $\mathcal{P}$. This configuration, which plays a very important role in classifying Steiner triple systems (see, e.\,g., \cite{enc}), is depicted in Figure \ref{f2} in a form showing an automorphism of order three; it also lives in the Fano plane and, as it is readily seen by comparing Figures \ref{f1} and \ref{f2}, it can be obtained from the latter by removal of any of its seven points and all the three lines passing through it.
In order to see that $\mathcal{V}(\mathcal{P}) \cong$ PG$(2,2)$ we shall first show, using our diagrammatical representation of $\mathcal{P}$, all seven geometric hyperplanes of $\mathcal{P}$ --- Figure \ref{f3}. We see that they are indeed of two different forms, of cardinality three and four. A member of the former set comprises two points at maximum distance from each other. Such geometric hyperplane corresponds to a type-one (or $\alpha$-) point of PG$(2,2)$. A member of the latter set features three points on a common line; such a geometric hyperplane of $\mathcal{P}$ corresponds to a type-two (or $\beta$-) point of our PG$(2,2)$.
The seven lines of $\mathcal{V}(\mathcal{P})$ are illustrated in a compact diagrammatic form in Figure \ref{f4}; as it is easily discernible, each of the six ordinary lines is of the form $\{\alpha, \beta, \beta\}$, whilst the remaining defective one has the $\{\alpha, \alpha,
\alpha\}$ shape.
\begin{figure}[pth!]
\centering
\includegraphics[width=2.5truecm]{f2.eps}
\caption{An illustrative portrayal of the Pasch configuration: circles stand for its points, whereas its lines are represented by triples of points on common straight segments (three) and the triple lying on a big circle.}
\label{f2}
\end{figure}
\begin{figure}[pth!]
\centering
\includegraphics[width=5truecm]{f3.eps}
\caption{The seven geometric hyperplanes of the Pasch configuration. The hyperplanes are labelled by imaginary units of octonions in such a way that --- as it is obvious from the next figure --- the seven lines of the Veldkamp space of the Pasch configuration are identical with the seven distinguished triples of units, that is with the seven lines of the PG$(2,2)$ shown in Figure \ref{f1}.}
\label{f3}
\end{figure}
\begin{figure}[pth!]
\centering
\includegraphics[width=7.5truecm]{f4.eps}
\caption{A unified view of the seven Veldkamp lines of the Pasch configuration. The reader can readily verify that for any three geometric hyperplanes lying on a given line of the Fano plane, one is the complement of the symmetric difference of the other two.}
\label{f4}
\end{figure}
\section{Sedenions and the Desargues $(10_3)$-configuration}
Our next focus is on $A_4$, the sedenions, whose basic multiplication properties are summarized in Table 2.
\begin{table}[h]
\centering
\caption{The multiplication table of the imaginary unit sedenions $e_a$, $1 \leq a \leq 15$. As in the previous section, we shall employ a short-hand notation $e_a \equiv a$; likewise for the real unit $e_0 \equiv 0$. There are also shown multiplication tables corresponding to the distinguished nested sequence of sub-algebras starting with complex numbers ($a=1$), quaternions ($1 \leq a \leq 3$) and octonions ($1 \leq a \leq 7$).}
\vspace*{0.4cm}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{||r|r|rr|rrrr|rrrrrrrr||}
\hline \hline
$*$ & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 \\
\hline
1 & $-$0 & $-$3 & +2 & $-$5 & +4 & +7 & $-$6 & $-$9 & +8 & +11 & $-$10 & +13 & $-$12 & $-$15 & +14 \\
\hline
2 & +3 & $-$0 & $-$1 & $-$6 & $-$7 & +4 & +5 & $-$10 & $-$11 & +8 & +9 & +14 & +15 & $-$12 & $-$13 \\
3 & $-$2 & +1 & $-$0 & $-$7 & +6 & $-$5 & +4 & $-$11 & +10 & $-$9 & +8 & +15 & $-$14 & +13 & $-$12 \\
\hline
4 & +5 & +6 & +7 & $-$0 & $-$1 & $-$2 & $-$3 & $-$12 & $-$13 & $-$14 & $-$15 & +8 & +9 & +10 & +11 \\
5 & $-$4 & +7 & $-$6 & +1 & $-$0 & +3 & $-$2 & $-$13 & +12 & $-$15 & +14 & $-$9 & +8 & $-$11 & +10 \\
6 & $-$7 & $-$4 & +5 & +2 & $-$3 & $-$0 & +1 & $-$14 & +15 & +12 & $-$13 & $-$10 & +11 & +8 & $-$9 \\
7 & +6 & $-$5 & $-$4 & +3 & +2 & $-$1 & $-$0 & $-$15 & $-$14 & +13 & +12 & $-$11 & $-$10 & +9 & +8 \\
\hline
8 & +9 & +10 & +11 & +12 & +13 & +14 & +15 & $-$0 & $-$1 & $-$2 & $-$3 & $-$4 & $-$5 & $-$6 & $-$7 \\
9 & $-$8 & +11 & $-$10 & +13 & $-$12 & $-$15 & +14 & +1 & $-$0 & +3 & $-$2 & +5 & $-$4 & $-$7 & +6 \\
10 & $-$11 & $-$8 & +9 & +14 & +15 & $-$12 & $-$13 & +2 & $-$3 & $-$0 & +1 & +6 & +7 & $-$4 & $-$5 \\
11 & +10 & $-$9 & $-$8 & +15 & $-$14 & +13 & $-$12 & +3 & +2 & $-$1 & $-$0 & +7 & $-$6 & +5 & $-$4 \\
12 & $-$13 & $-$14 & $-$15 & $-$8 & +9 & +10 & +11 & +4 & $-$5 & $-$6 & $-$7 & $-$0 & +1 & +2 & +3 \\
13 & +12 & $-$15 & +14 & $-$9 & $-$8 & $-$11 & +10 & +5 & +4 & $-$7 & +6 & $-$1 & $-$0 & $-$3 & +2 \\
14 & +15 & +12 & $-$13 & $-$10 & +11 & $-$8 & $-$9 & +6 & +7 & +4 & $-$5 & $-$2 & +3 & $-$0 & $-$1 \\
15 & $-$14 & +13 & +12 & $-$11 & $-$10 & +9 & $-$8 & +7 & $-$6 & +5 & +4 & $-$3 & $-$2 & +1 & $-$0 \\
\hline \hline
\end{tabular}%
}
\end{table}
\noindent
An inspection of this table yields as many as 35 distinguished triples, namely:
$$
\{1,2,3\}, \{1,4,5\}, \{1,6,7\}, \{1,8,9\}, \{1,10,11\}, \{1,12,13\}, \{1,14,15\},
$$
$$
\{2,4,6\}, \{2,5,7\}, \{2,8,10\}, \{2,9,11\}, \{2,12,14\}, \{2,13,15\},
$$
$$
\{3,4,7\}, \{3,5,6\}, \{3,8,11\}, \{3,9,10\}, \{3,12,15\}, \{3,13,14\},
$$
$$
\{4,8,12\}, \{4,9,13\}, \{4,10,14\}, \{4,11,15\},
$$
$$
\{5,8,13\}, \{5,9,12\}, \{5,10,15\}, \{5,11,14\},
$$
$$
\{6,8,14\}, \{6,9,15\}, \{6,10,12\}, \{6,11,13\},
$$
$$
\{7,8,15\}, \{7,9,14\}, \{7,10,13\}, \{7,11,12\}.
$$
\noindent
Regarding the 15 imaginary units as points and the 35 distinguished trios of them as lines, we obtain a point-line incidence structure where each line has three points and each point is on seven lines, and which is isomorphic to PG$(3,2)$, the smallest projective space --- as depicted in Figure \ref{fgr1}. The latter figure employs a diagrammatical model of PG$(3,2)$ built, after Polster \cite{pol}, around the pentagonal model of the generalized quadrangle of type GQ$(2,2)$ whose 15 lines are illustrated by triples of points lying on black line-segments (10 of them) and/or black arcs of circles (5). The remaining 20 lines of PG$(3,2)$ comprise four distinct orbits: the yellow, red, blue and green one consisting, respectively, of the yellow ($\{1,10,11\}$), red
($\{1,8,9\}$), blue ($\{3,13,14\}$) and green ($\{3,12,15\}$) line and other four lines we get from each by rotation through 72 degrees around the center of the pentagon.
\begin{figure}[pth!]
\centering
\includegraphics[width=9truecm]{fgr1.eps}
\caption{An illustration of the structure of PG$(3,2)$ that provides the multiplication law for sedenions. As in the previous case, the three imaginaries lying on the same line are such that the product of two of them yields the third one, sign disregarded.}
\label{fgr1}
\end{figure}
It is not difficult to verify that we have 25 ordinary lines, namely
$$
\{1,2,3\}, \{1,4,5\}, \{1,6,7\}, \{1,8,9\}, \{1,10,11\}, \{1,12,13\}, \{1,14,15\},
$$
$$
\{2,4,6\}, \{2,5,7\}, \{2,8,10\}, \{2,9,11\}, \{2,12,14\}, \{2,13,15\},
$$
$$
\{3,4,7\}, \{4,8,12\}, \{4,9,13\}, \{4,10,14\}, \{4,11,15\},
$$
$$
\{3,8,11\}, \{5,8,13\}, \{6,8,14\}, \{7,8,15\},
$$
$$
\{3,12,15\}, \{5,10,15\}, \{6,9,15\},
$$
and 10 defective ones, namely
$$
\{3,5,6\}, \{3,9,10\}, \{3,13,14\},
$$
$$
\{5,9,12\}, \{5,11,14\},
$$
$$
\{6,10,12\}, \{6,11,13\},
$$
$$
\{7,9,14\}, \{7,10,13\}, \{7,11,12\}.
$$
Similarly, our sedenionic PG$(3,2)$ features two distinct types of points.
A type-one point is such that four lines passing through it are ordinary, the remaining three being defective; such a point lies in the set
$$\{3,5,6,7,9,10,11,12,13,14\} \equiv \alpha.$$
A type-two point is such that every line passing through it is ordinary; such a point belongs to the set
$$\{1,2,4,8,15\} \equiv \beta,$$
being illustrated by gray shading in Figure \ref{fgr1}.
We see that all defective lines are of the same form, namely $\{\alpha, \alpha, \alpha \}$. The 25 ordinary lines split into two distinct families. Ten of them are of the form $\{\alpha, \beta, \beta\}$, namely
$$
\{1,2,3\}, \{1,4,5\}, \{1,8,9\}, \{1,14,15\},
$$
$$
\{2,4,6\}, \{2,8,10\}, \{2,13,15\},
$$
$$
\{4,8,12\},\{4,11,15\},
$$
$$
\{7,8,15\},
$$
and the remaining 15 are of the form $\{\alpha, \alpha, \beta\}$, namely
$$
\{1,6,7\}, \{1,10,11\}, \{1,12,13\},
$$
$$
\{2,5,7\}, \{2,9,11\}, \{2,12,14\},
$$
$$
\{3,4,7\}, \{4,9,13\}, \{4,10,14\},
$$
$$
\{3,8,11\}, \{5,8,13\}, \{6,8,14\},
$$
$$
\{3,12,15\}, \{5,10,15\}, \{6,9,15\}.
$$
A configuration $\mathcal{C}_4$ whose Veldkamp space reproduces the above-described partitions of points and lines of PG$(3,2)$ is, as demonstrated below, the famous {\it Desargues} $(10_3)$-configuration, $\mathcal{D}$, which is one of the most prominent point-line incidence structures (see, e.\,g., \cite{piser}). Up to isomorphism, there exist altogether ten $(10_3)$-configurations. The Desargues configuration is, unlike the others, flag-transitive and the only one where for {\it each} of its points the three points that are not collinear with it lie on a line. This configuration, depicted in Figure \ref{fgr2} in a form showing its automorphism of order three, also lives in our sedenionic PG$(3,2)$; here, its points are the ten $\alpha$-points and its lines are all the defective lines.
In order to see that $\mathcal{V}(\mathcal{D}) \cong$ PG$(3,2)$ we shall first introduce, using our diagrammatical representation of $\mathcal{D}$, all 15 geometric hyperplanes of $\mathcal{D}$ --- Figure \ref{fgr3}. We see that they are indeed of two different forms, and of required cardinality ten and five. A member of the former set comprises a point and three points not collinear with it. Such geometric hyperplane corresponds to a type-one (or $\alpha$-) point of PG$(3,2)$. A member of the latter set features six points located on four lines, with two lines per each point; this is nothing but the Pasch configuration we introduced in the previous section. Such a geometric hyperplane of $\mathcal{D}$ corresponds to a type-two (or $\beta$-) point of our PG$(3,2)$. It is also a straightforward task to verify that $\mathcal{V}(\mathcal{D})$ is endowed with 35 lines splitting into the required three families; those that correspond to defective lines of our sedenionic PG$(3,2)$ are shown in Figure \ref{fgr4}, while those that correspond to ordinary lines are depicted in Figure \ref{fgr5} (of type $\{\alpha, \beta, \beta\}$) and Figure \ref{fgr6} (of type $\{\alpha, \alpha, \beta\}$). Figure \ref{fgr7} offers a `condensed' view of the isomorphism $\mathcal{V}(\mathcal{D}) \cong$ PG$(3,2)$.
We shall finalize this section by pointing out that the existence of two different kinds of geometric hyperplanes of the Desargues configuration is closely connected with two well-known views of this configuration. The first one is as a pair of triangles that are in perspective from both a point and a line (Desargues' theorem), the point and the line forming a geometric hyperplane. The other view is as the incidence sum of a complete quadrangle (i.\,e., a $(4_3,6_2)$-configuration) and a Pasch $(6_2,4_3)$-configuration \cite{bogepi}.
\begin{figure}[pth!]
\centering
\includegraphics[width=4.5truecm]{fgr2.eps}
\caption{An illustrative portrayal of the Desargues configuration, built around the model of the Pasch configuration shown in Figure \ref{f2}: circles stand for its points, whereas its lines are represented by triples of points on common straight segments (six), arcs of circles (three) and a big circle.}
\label{fgr2}
\end{figure}
\begin{figure}[pth!]
\centering
\psfrag{1}{\small 14}
\psfrag{2}{\small 13}
\psfrag{3}{\small 11}
\psfrag{4}{\small 9}
\psfrag{5}{\small 10}
\psfrag{6}{\small 12}
\psfrag{7}{\small 6}
\psfrag{8}{\small 5}
\psfrag{9}{\small 3}
\psfrag{10}{\small ~7}
\psfrag{11}{\small ~1}
\psfrag{12}{\small ~2}
\psfrag{13}{\small 4}
\psfrag{14}{\small ~8}
\psfrag{15}{\small 15}
\includegraphics[width=12truecm]{fgr3.eps}
\caption{The fifteen geometric hyperplanes of the Desargues configuration. The hyperplanes are labelled by imaginary units of sedenions in such a way that --- as we shall verify in the next three figures --- the 35 lines of the Veldkamp space of the Desargues configuration are identical with the 35 distinguished triples of units, that is with the 35 lines of the PG$(3,2)$ shown in Figure \ref{fgr1}.}
\label{fgr3}
\end{figure}
\begin{figure}[pth!]
\centering
\psfrag{1-4-10}{~7-9-14}
\psfrag{2-5-10}{7-10-13}
\psfrag{3-6-10}{7-11-12}
\psfrag{1-2-9}{3-13-14}
\psfrag{1-3-8}{5-11-14}
\psfrag{2-3-7}{6-11-13}
\psfrag{4-5-9}{3-9-10}
\psfrag{4-6-8}{5-9-12}
\psfrag{5-6-7}{6-10-12}
\psfrag{7-8-9}{~3-5-6}
\includegraphics[width=14truecm,clip=]{fgr4.eps}
\caption{The ten Veldkamp lines of the Desargues configuration that represent the ten defective lines
of the sedenionic PG$(3,2)$. Here, as well as in the next two figures, the three geometric hyperplanes
comprising a given Veldkamp line
are distinguished by different colors, with their common elements (here just a single point) being colored black.
For each Veldkamp line we also explicitly indicate its composition.}
\label{fgr4}
\end{figure}
\begin{figure}[pth!]
\centering
\psfrag{7-12-13}{~~2-4-6}
\psfrag{8-11-13}{~~1-4-5}
\psfrag{9-11-12}{~~1-2-3}
\psfrag{6-13-14}{~~4-8-12}
\psfrag{5-12-14}{~~2-8-10}
\psfrag{4-11-14}{~~1-8-9}
\psfrag{3-13-15}{~~4-11-15}
\psfrag{2-12-15}{~~2-13-15}
\psfrag{1-11-15}{~~1-14-15}
\psfrag{10-14-15}{~~7-8-15}
\includegraphics[width=14truecm,clip=]{fgr5.eps}
\caption{The ten Veldkamp lines of the Desargues configuration that represent the ten ordinary lines
of the sedenionic PG$(3,2)$ of type $\{\alpha, \beta, \beta\}$.}
\label{fgr5}
\end{figure}
\begin{figure}[pth!]
\centering
\psfrag{4-7-15}{~~6-9-15}
\psfrag{5-8-15}{~~5-10-15}
\psfrag{6-9-15}{~~3-12-15}
\psfrag{1-7-14}{~~6-8-14}
\psfrag{2-8-14}{~~5-8-13}
\psfrag{3-9-14}{~~3-8-11}
\psfrag{7-10-11}{~~1-6-7}
\psfrag{8-10-12}{~~2-5-7}
\psfrag{9-10-13}{~~3-4-7}
\psfrag{1-6-12}{~2-12-14}
\psfrag{2-4-13}{~~4-9-13}
\psfrag{3-5-11}{~1-10-11}
\psfrag{3-4-12}{~~2-9-11}
\psfrag{2-6-11}{~1-12-13}
\psfrag{1-5-13}{~4-10-14}
\includegraphics[width=13truecm,clip=]{fgr6.eps}
\caption{The fifteen Veldkamp lines of the Desargues configuration that represent the fifteen ordinary lines
of the sedenionic PG$(3,2)$ of type $\{\alpha, \alpha, \beta\}$.}
\label{fgr6}
\end{figure}
\begin{figure}[t]
\centering
\psfrag{1}{\tiny 1}
\psfrag{2}{\tiny 2}
\psfrag{3}{\tiny 3}
\psfrag{4}{\tiny 4}
\psfrag{5}{\tiny 5}
\psfrag{6}{\tiny 6}
\psfrag{7}{\tiny 7}
\psfrag{8}{\tiny 8}
\psfrag{9}{\tiny 9}
\psfrag{10}{\tiny 10}
\psfrag{11}{\tiny 11}
\psfrag{12}{\tiny 12}
\psfrag{13}{\tiny 13}
\psfrag{14}{\tiny 14}
\psfrag{15}{\tiny 15}
\includegraphics[width=14truecm]{fgr7.eps}
\caption{A compact graphical view of illustrating the bijection between 15 imaginary unit sedenions and 15 geometric hyperplanes of the Desargues configuration, as well as between 35 distinguished triples of units and 35 Veldkamp lines of the Desargues configuration. }
\label{fgr7}
\end{figure}
\newpage
\section{32-nions and the Cayley-Salmon $(15_4,20_3)$-configuration}
Our next case is $A_5$, or the 32-nions, whose multiplication properties are encoded in Table 3.
\begin{table}[pth!]
\centering
\caption{The multiplication table --- for better readability split into two parts --- of the imaginary unit 32-nions $e_a$, $1 \leq a \leq 31$, in short-handed notation $e_a \equiv a$ (and $e_0 \equiv 0$). }
\vspace*{0.1cm}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{||c|r|rr|rrrr|rrrrrrrrr||}
\hline \hline
$*$ & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 & ~~~~~~ \\
\hline
1 & $-$0 & $-$3 & +2 & $-$5 & +4 & +7 & $-$6 & $-$9 & +8 & +11 & $-$10 & +13 & $-$12 & $-$15 & +14 & \\
\hline
2 & +3 & $-$0 & $-$1 & $-$6 & $-$7 & +4 & +5 & $-$10 & $-$11 & +8 & +9 & +14 & +15 & $-$12 & $-$13 &\\
3 & $-$2 & +1 & $-$0 & $-$7 & +6 & $-$5 & +4 & $-$11 & +10 & $-$9 & +8 & +15 & $-$14 & +13 & $-$12 &\\
\hline
4 & +5 & +6 & +7 & $-$0 & $-$1 & $-$2 & $-$3 & $-$12 & $-$13 & $-$14 & $-$15 & +8 & +9 & +10 & +11 &\\
5 & $-$4 & +7 & $-$6 & +1 & $-$0 & +3 & $-$2 & $-$13 & +12 & $-$15 & +14 & $-$9 & +8 & $-$11 & +10 &\\
6 & $-$7 & $-$4 & +5 & +2 & $-$3 & $-$0 & +1 & $-$14 & +15 & +12 & $-$13 & $-$10 & +11 & +8 & $-$9 &\\
7 & +6 & $-$5 & $-$4 & +3 & +2 & $-$1 & $-$0 & $-$15 & $-$14 & +13 & +12 & $-$11 & $-$10 & +9 & +8 &\\
\hline
8 & +9 & +10 & +11 & +12 & +13 & +14 & +15 & $-$0 & $-$1 & $-$2 & $-$3 & $-$4 & $-$5 & $-$6 & $-$7 &\\
9 & $-$8 & +11 & $-$10 & +13 & $-$12 & $-$15 & +14 & +1 & $-$0 & +3 & $-$2 & +5 & $-$4 & $-$7 & +6 &\\
10 & $-$11 & $-$8 & +9 & +14 & +15 & $-$12 & $-$13 & +2 & $-$3 & $-$0 & +1 & +6 & +7 & $-$4 & $-$5 &\\
11 & +10 & $-$9 & $-$8 & +15 & $-$14 & +13 & $-$12 & +3 & +2 & $-$1 & $-$0 & +7 & $-$6 & +5 & $-$4 &\\
12 & $-$13 & $-$14 & $-$15 & $-$8 & +9 & +10 & +11 & +4 & $-$5 & $-$6 & $-$7 & $-$0 & +1 & +2 & +3 &\\
13 & +12 & $-$15 & +14 & $-$9 & $-$8 & $-$11 & +10 & +5 & +4 & $-$7 & +6 & $-$1 & $-$0 & $-$3 & +2 &\\
14 & +15 & +12 & $-$13 & $-$10 & +11 & $-$8 & $-$9 & +6 & +7 & +4 & $-$5 & $-$2 & +3 & $-$0 & $-$1 &\\
15 & $-$14 & +13 & +12 & $-$11 & $-$10 & +9 & $-$8 & +7 & $-$6 & +5 & +4 & $-$3 & $-$2 & +1 & $-$0 &\\
\hline
16 & +17 & +18 & +19 & +20 & +21 & +22 & +23 & +24 & +25 & +26 & +27 & +28 & +29 & +30 & +31 &\\
17 & $-$16 & +19 & $-$18 & +21 & $-$20 & $-$23 & +22 & +25 & $-$24 & $-$27 & +26 & $-$29 & +28 & +31 & $-$30 &\\
18 & $-$19 & $-$16 & +17 & +22 & +23 & $-$20 & $-$21 & +26 & +27 & $-$24 & $-$25 & $-$30 & $-$31 & +28 & +29 &\\
19 & +18 & $-$17 & $-$16 & +23 & $-$22 & +21 & $-$20 & +27 & $-$26 & +25 & $-$24 & $-$31 & +30 & $-$29 & +28 &\\
20 & $-$21 & $-$22 & $-$23 & $-$16 & +17 & +18 & +19 & +28 & +29 & +30 & +31 & $-$24 & $-$25 & $-$26 & $-$27 & \\
21 & +20 & $-$23 & +22 & $-$17 & $-$16 & $-$19 & +18 & +29 & $-$28 & +31 & $-$30 & +25 & $-$24 & +27 & $-$26 &\\
22 & +23 & +20 & $-$21 & $-$18 & +19 & $-$16 & $-$17 & +30 & $-$31 & $-$28 & +29 & +26 & $-$27 & $-$24 & +25 &\\
23 & $-$22 & +21 & +20 & $-$19 & $-$18 & +17 & $-$16 & +31 & +30 & $-$29 & $-$28 & +27 & +26 & $-$25 & $-$24 &\\
24 & $-$25 & $-$26 & $-$27 & $-$28 & $-$29 & $-$30 & $-$31 & $-$16 & +17 & +18 & +19 & +20 & +21 & +22 & +23 &\\
25 & +24 & $-$27 & +26 & $-$29 & +28 & +31 & $-$30 & $-$17 & $-$16 & $-$19 & +18 & $-$21 & +20 & +23 & $-$22 &\\
26 & +27 & +24 & $-$25 & $-$30 & $-$31 & +28 & +29 & $-$18 & +19 & $-$16 & $-$17 & $-$22 & $-$23 & +20 & +21 &\\
27 & $-$26 & +25 & +24 & $-$31 & +30 & $-$29 & +28 & $-$19 & $-$18 & +17 & $-$16 & $-$23 & +22 & $-$21 & +20 & \\
28 & +29 & +30 & +31 & +24 & $-$25 & $-$26 & $-$27 & $-$20 & +21 & +22 & +23 & $-$16 & $-$17 & $-$18 & $-$19 & \\
29 & $-$28 & +31 & $-$30 & +25 & +24 & +27 & $-$26 & $-$21 & $-$20 & +23 & $-$22 & +17 & $-$16 & +19 & $-$18 &\\
30 & $-$31 & $-$28 & +29 & +26 & $-$27 & +24 & +25 & $-$22 & $-$23 & $-$20 & +21 & +18 & $-$19 & $-$16 & +17 &\\
31 & +30 & $-$29 & $-$28 & +27 & +26 & $-$25 & +24 & $-$23 & +22 & $-$21 & $-$20 & +19 & +18 & $-$17 & $-$16 &\\
\hline
\end{tabular}%
}
\begin{tabular}{cc}
& \\
\end{tabular}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{||c|rrrrrrrrrrrrrrrr||}
\hline
$*$ & 16 & 17 & 18 & 19 & 20 & 21 & 22 & 23 & 24 & 25 & 26 & 27 & 28 & 29 & 30 & 31 \\
\hline
1 & $-$17 & +16 & +19 & $-$18 & +21 & $-$20 & $-$23 & +22 & +25 & $-$24 & $-$27 & +26 & $-$29 & +28 & +31 & $-$30 \\
\hline
2 & $-$18 & $-$19 & +16 & +17 & +22 & +23 & $-$20 & $-$21 & +26 & +27 & $-$24 & $-$25 & $-$30 & $-$31 & +28 & +29 \\
3 & $-$19 & +18 & $-$17 & +16 & +23 & $-$22 & +21 & $-$20 & +27 & $-$26 & +25 & $-$24 & $-$31 & +30 & $-$29 & +28 \\
\hline
4 & $-$20 & $-$21 & $-$22 & $-$23 & +16 & +17 & +18 & +19 & +28 & +29 & +30 & +31 & $-$24 & $-$25 & $-$26 & $-$27 \\
5 & $-$21 & +20 & $-$23 & +22 & $-$17 & +16 & $-$19 & +18 & +29 & $-$28 & +31 & $-$30 & +25 & $-$24 & +27 & $-$26 \\
6 & $-$22 & +23 & +20 & $-$21 & $-$18 & +19 & +16 & $-$17 & +30 & $-$31 & $-$28 & +29 & +26 & $-$27 & $-$24 & +25 \\
7 & $-$23 & $-$22 & +21 & +20 & $-$19 & $-$18 & +17 & +16 & +31 & +30 & $-$29 & $-$28 & +27 & +26 & $-$25 & $-$24 \\
\hline
8 & $-$24 & $-$25 & $-$26 & $-$27 & $-$28 & $-$29 & $-$30 & $-$31 & +16 & +17 & +18 & +19 & +20 & +21 & +22 & +23 \\
9 & $-$25 & +24 & $-$27 & +26 & $-$29 & +28 & +31 & $-$30 & $-$17 & +16 & $-$19 & +18 & $-$21 & +20 & +23 & $-$ 22 \\
10 & $-$26 & +27 & +24 & $-$25 & $-$30 & $-$31 & +28 & +29 & $-$18 & +19 & +16 & $-$17 & $-$22 & $-$23 & +20 & +21 \\
11 & $-$27 & $-$26 & +25 & +24 & $-$31 & +30 & $-$29 & +28 & $-$19 & $-$18 & +17 & +16 & $-$23 & +22 & $-$21 & +20
\\
12 & $-$28 & +29 & +30 & +31 & +24 & $-$25 & $-$26 & $-$27 & $-$20 & +21 & +22 & +23 & +16 & $-$17 & $-$18 & $-$19 \\
13 & $-$29 & $-$28 & +31 & $-$30 & +25 & +24 & +27 & $-$26 & $-$21 & $-$20 & +23 & $-$22 & +17 & +16 & +19 & $-$18 \\
14 & $-$30 & $-$31 & $-$28 & +29 & +26 & $-$27 & +24 & +25 & $-$22 & $-$23 & $-$20 & +21 & +18 & $-$19 & +16 & +17 \\
15 & $-$31 & +30 & $-$29 & $-$28 & +27 & +26 & $-$25 & +24 & $-$23 & +22 & $-$21 & $-$20 & +19 & +18 & $-$17 & +16
\\
\hline
16 & $-$0 & $-$1 & $-$2 & $-$3 & $-$4 & $-$5 & $-$6 & $-$7 & $-$8 & $-$9 & $-$10 & $-$11 & $-$12 & $-$13 & $-$14 & $-$15 \\
17 & +1 & $-$0 & +3 & $-$2 & +5 & $-$4 & $-$7 & +6 & +9 & $-$8 & $-$11 & +10 & $-$13 & +12 & +15 & $-$14 \\
18 & +2 & $-$3 & $-$0 & +1 & +6 & +7 & $-$4 & $-$5 & +10 & +11 & $-$8 & $-$9 & $-$14 & $-$15 & +12 & +13 \\
19 & +3 & +2 & $-$1 & $-$0 & +7 & $-$6 & +5 & $-$4 & +11 & $-$10 & +9 & $-$8 & $-$15 & +14 & $-$13 & +12 \\
20 & +4 & $-$5 & $-$6 & $-$7 & $-$0 & +1 & +2 & +3 & +12 & +13 & +14 & +15 & $-$8 & $-$9 & $-$10 & $-$11 \\
21 & +5 & +4 & $-$7 & +6 & $-$1 & $-$0 & $-$3 & +2 & +13 & $-$12 & +15 & $-$14 & +9 & $-$8 & +11 & $-$10 \\
22 & +6 & +7 & +4 & $-$5 & $-$2 & +3 & $-$0 & $-$1 & +14 & $-$15 & $-$12 & +13 & +10 & $-$11 & $-$8 & +9 \\
23 & +7 & $-$6 & +5 & +4 & $-$3 & $-$2 & +1 & $-$0 & +15 & +14 & $-$13 & $-$12 & +11 & +10 & $-$9 & $-$8\\
24 & +8 & $-$9 & $-$10 & $-$11 & $-$12 & $-$13 & $-$14 & $-$15 & $-$0 & +1 & +2 & +3 & +4 & +5 & +6 & +7 \\
25 & +9 & +8 & $-$11 & +10 & $-$13 & +12 & +15 & $-$14 & $-$1 & $-$0 & $-$3 & +2 & $-$5 & +4 & +7 & $-$6\\
26 & +10 & +11 & +8 & $-$9 & $-$14 & $-$15 & +12 & +13 & $-$2 & +3 & $-$0 & $-$1 & $-$6 & $-$7 & +4 & +5 \\
27 & +11 & $-$10 & +9 & +8 & $-$15 & +14 & $-$13 & +12 & $-$3 & $-$2 & +1 & $-$0 & $-$7 & +6 & $-$5 & +4 \\
28 & +12 & +13 & +14 & +15 & +8 & $-$9 & $-$10 & $-$11 & $-$4 & +5 & +6 & +7 & $-$0 & $-$1 & $-$2 & $-$3 \\
29 & +13 & $-$12 & +15 & $-$14 & +9 & +8 & +11 & $-$10 & $-$5 & $-$4 & +7 & $-$6 & +1 & $-$0 & +3 & $-$2 \\
30 & +14 & $-$15 & $-$12 & +13 & +10 & $-$11 & +8 & +9 & $-$6 & $-$7 & $-$4 & +5 & +2 & $-$3 & $-$0 & +1 \\
31 & +15 & +14 & $-$13 & $-$12 & +11 & +10 & $-$9 & +8 & $-$7 & +6 & $-$5 & $-$4 & +3 & +2 & $-$1 & $-$0 \\
\hline \hline
\end{tabular}%
}
\end{table}
From this table we infer the existence of 155 distinguished triples of imaginary units. Regarding the 31 imaginary units of $A_5$ as points and the 155 distinguished triples of them as lines, we obtain a point-line incidence structure where each line has three points and each point is on 15 lines, and which is isomorphic to PG$(4,2)$. We next find that 65 lines of this space are defective and 90 ordinary. However, unlike the preceding two cases, there are {\it three} different types of points in our 32-nionic PG$(4,2)$:
ten $\alpha$-points, $$\alpha \equiv \{7,11,13,14,19,21,22,25,26,28\},$$ each of which is on nine defective and six ordinary lines; fifteen $\beta$-points, $$\beta \equiv \{3,5,6,9,10,12,15,17,18,20,23,24,27,29,30\},$$ each of which is on seven defective and eight ordinary lines; and six $\gamma$-points, $$\gamma \equiv \{1,2,4,8,16,31\},$$ each of them being on fifteen ordinary (and, hence, on zero defective) lines. This stratification of the point-set of PG$(4,2)$ leads, in turn, to two different kinds of defective lines and three distinct kinds of ordinary lines.
Concerning the former, there are 45 of them of type $\{\alpha,\alpha,\beta\}$, namely
$$
\{3,13,14\}, \{3,21,22\}, \{3,25,26\},
$$
$$
\{5,11,14\}, \{5,19,22\}, \{5,25,28\},
$$
$$
\{6,11,13\}, \{6,19,21\}, \{6,26,28\},
$$
$$
\{7,9,14\}, \{7,10,13\}, \{7,11,12\}, \{7,17,22\}, \{7,18,21\}, \{7,19,20\},
$$ $$
\{7,25,30\}, \{7,26,29\}, \{7,27,28\},
$$
$$
\{9,19,26\}, \{9,21,28\},
$$
$$
\{10,19,25\}, \{10,22,28\},
$$
$$
\{11,17,26\}, \{11,18,25\}, \{11,19,24\}, \{11,21,30\}, \{11,22,29\}, \{11,23,28\},
$$
$$
\{12,21,25\}, \{12,22,26\},
$$
$$
\{13,17,28\}, \{13,19,30\}, \{13,20,25\}, \{13,21,24\}, \{13,22,27\}, \{13,23,26\},
$$
$$
\{14,18,28\}, \{14,19,29\}, \{14,20,26\}, \{14,21,27\}, \{14,22,24\}, \{14,23,25\},
$$
$$
\{15,19,28\}, \{15,21,26\}, \{15,22,25\},
$$
and 20 of type $\{\beta,\beta,\beta\}$, namely
$$
\{3,5,6\}, \{3,9,10\}, \{3,17,18\}, \{3,29,30\},
$$
$$
\{5,9,12\}, \{5,17,20\}, \{5,27,30\},
$$
$$
\{6,10,12\}, \{6,18,20\}, \{6,27,29\},
$$
$$
\{9,17,24\}, \{9,23,30\},
$$
$$
\{10,18,24\}, \{10,23,29\},
$$
$$
\{12,20,24\}, \{12,23,27\},
$$
$$
\{15,17,30\}, \{15,18,29\}, \{15,20,27\}, \{15,23,24\}.
$$
As per the latter, one finds 15 of them of type $\{\beta,\beta,\beta\}$, namely
$$
\{3,12,15\}, \{3,20,23\}, \{3,24,27\},
$$
$$
\{5,10,15\}, \{5,18,23\}, \{5,24,29\},
$$
$$
\{6,9,15\}, \{6,17,23\}, \{6,24,30\},
$$
$$
\{9,18,27\}, \{9,20,29\},
$$
$$
\{10,17,27\}, \{10,20,30\},
$$
$$
\{12,17,29\}, \{12,18,30\},
$$
60 of type $\{\alpha,\beta,\gamma \}$, namely
$$
\{1, 6, 7\}, \{1, 10, 11\}, \{1, 12, 13\}, \{1, 14, 15\}, \{1, 18, 19\}, \{1, 20, 21\},
$$ $$
\{1, 22, 23\}, \{1, 24, 25\},
\{1, 26, 27\}, \{1, 28, 29\},
$$
$$
\{2, 5, 7\}, \{2, 9, 11\}, \{2, 12, 14\}, \{2, 13, 15\}, \{2, 17, 19\}, \{2, 20, 22\},
$$ $$
\{2, 21, 23\}, \{2, 24, 26\},
\{2, 25, 27\}, \{2, 28, 30\},
$$
$$
\{3, 4, 7\}, \{3, 8, 11\}, \{3, 16, 19\}, \{3, 28, 31\},
$$ $$
\{4, 9, 13\}, \{4, 10, 14\}, \{4, 11, 15\}, \{4, 17, 21\}, \{4, 18, 22\},
$$ $$
\{4, 19, 23\},
\{4, 24, 28\}, \{4, 25, 29\},
\{4, 26, 30\},
$$
$$
\{5, 8, 13\}, \{5, 16, 21\}, \{5, 26, 31\},
$$ $$
\{6, 8, 14\}, \{6, 16, 22\}, \{6, 25, 31\},
$$ $$
\{7, 8, 15\}, \{7, 16, 23\}, \{7, 24, 31\},
$$ $$
\{8, 17, 25\}, \{8, 18, 26\}, \{8, 19, 27\}, \{8, 20, 28\}, \{8, 21, 29\}, \{8, 22, 30\},
$$ $$
\{9, 16, 25\}, \{9, 22, 31\},
$$ $$
\{10, 16, 26\}, \{10, 21, 31\},
$$ $$
\{11, 16, 27\}, \{11, 20, 31\},
$$ $$
\{12, 16, 28\}, \{12, 19, 31\},
$$ $$
\{13, 16, 29\}, \{13, 18, 31\},
$$ $$
\{14, 16, 30\}, \{14, 17, 31\},
$$
and, finally, 15 of type $\{\beta, \gamma, \gamma \}$, namely
$$
\{1, 2, 3\}, \{1, 4, 5\}, \{1, 8, 9\}, \{1, 16, 17\}, \{1, 30, 31\},
$$ $$
\{2, 4, 6\}, \{2, 8, 10\}, \{2, 16, 18\}, \{2, 29, 31\},
$$ $$
\{4, 8, 12\}, \{4, 16, 20\}, \{4, 27, 31\},
$$ $$
\{8, 16, 24\}, \{8, 23, 31\},
$$ $$
\{5, 16, 31\}.
$$
A point-line configuration $\mathcal{C}_5$ whose Veldkamp space accounts for these stratifications of both the point- and line-set of our 32-nionic PG$(4,2)$ is of type $(15_4,20_3)$.\footnote{It is worth mentioning here that there exists another remarkable configuration whose Veldkamp space does the same job for us, namely the generalized quadrangle of order two, also known as the Cremona-Richmond $(15_3)$-configuration (see, e.\,g., \cite{twoq}).
However, this configuration is {\it triangle-free} and so it can$not$ contain the Desargues configuration as dictated by the nesting property of the Cayley-Dickson algebras.} This configuration is formed within our PG$(4,2)$ by 15 $\beta$-points and 20 defective lines of $\{\beta,\beta,\beta\}$ type and its structure is sketched in Figure \ref{fig1}. It is a rather easy task to verify that this particular $(15_4,20_3)$-configuration possesses 31 distinct geometric hyperplanes that fall into three different types. A type-one hyperplane consists of a pair of skew lines at maximum distance from each other;
there are, as depicted in Figure \ref{fig2}, ten hyperplanes of this type and they correspond to $\alpha$-points of PG$(4,2)$. A type-two hyperplane features a point and all the points not collinear with it, the latter forming --- not surprisingly --- the Pasch configuration; there are, as shown in
Figure \ref{fig3}, fifteen hyperplanes of this type and their counterparts are $\beta$-points of PG$(4,2)$. A type-three hyperplane is identical with the Desargues configuration; we find, as portrayed in Figure \ref{fig4}, altogether six guys of this type, each standing for a $\gamma$-point of PG$(4,2)$.
\begin{figure}[pth!]
\centering
\includegraphics[width=9truecm]{fig1.eps}
\caption{An illustration of the structure of the $(15_4,20_3)$-configuration, built around the model of the Desargues configuration shown in Figure \ref{fgr2}. The five points added to the Desargues configuration are the three peripheral points and the red and blue point in the center. The ten lines added are three lines denoted by red color, three blue lines, three lines joining pairwise the three peripheral points and the line that comprises the three points in the center of the figure, that is the ones represented by a bigger red circle, a smaller blue circle and a medium-sized black one.}
\label{fig1}
\end{figure}
\begin{figure}[pth!]
\centering
\includegraphics[width=14truecm]{fig2.eps}
\caption{The ten geometric hyperplanes of the $(15_4,20_3)$-configuration of type one; the number below a subfigure indicates how many hyperplane's copies we get by rotating
the particular subfigure through 120 degrees around its center.}
\label{fig2}
\end{figure}
\begin{figure}[pth!]
\centering
\includegraphics[width=14truecm]{fig3.eps}
\caption{The fifteen geometric hyperplanes of the $(15_4,20_3)$-configuration of type two.}
\label{fig3}
\end{figure}
\begin{figure}[pth!]
\centering
\includegraphics[width=14truecm]{fig4.eps}
\caption{The six geometric hyperplanes of the $(15_4,20_3)$-configuration of type three.}
\label{fig4}
\end{figure}
\begin{figure}[pth!]
\centering
\includegraphics[width=14truecm]{fig5.eps}
\caption{The five types of Veldkamp lines of the $(15_4,20_3)$-configuration. Here, unlike Figures 8 to 10, each representative of a geometric hyperplane is drawn separately and different colors are used to distinguish between different hyperplane types: red is reserved for type one, yellow for type two and blue for type three hyperplanes. As before, black color denotes the core of a Veldkamp line, that is the elements common to all the three hyperplanes comprising it.}
\label{fig5}
\end{figure}
We also find that our $(15_4,20_3)$-configuration yields 155 Veldkamp lines that are, as expected,
of five different types. A type-I Veldkamp line, shown in Figure \ref{fig5}a, features two hyperplanes of type one and a type-two hyperplane and its core consists of two points that are at maximum distance from each other; there are ${10 \choose 2} = 15 \times 6/2 = 45$ Veldkamp lines of this type and they correspond to defective lines of PG$(4,2)$ of type $\{\alpha,\alpha,\beta\}$. A type-II Veldkamp line, featured in Figure \ref{fig5}b, is composed of three hyperplanes of type two that share three points on a common line; there are, obviously, 20 Veldkamp lines of this type, having for their counterparts defective lines of PG$(4,2)$ of type $\{\beta, \beta, \beta\}$. A type-III Veldkamp line, portrayed in Figure \ref{fig5}c, also consists of three hyperplanes of type two, but in this case the three common points are pairwise at maximum distance from each other; a quick count leads to 15 Veldkamp lines of this type, these being in a bijection with 15 ordinary lines of PG$(4,2)$ of type $\{\beta, \beta, \beta\}$. Next, it comes a type-IV Veldkamp line, depicted in Figure \ref{fig5}d, which exhibits a hyperplane of each type and whose core is composed of a line and a point at the maximum distance from it;
since for each line of our $(15_4,20_3)$-configuration there are three points at maximum distance from it, there are $20 \times 3 = 60$ Veldkamp lines of this type, having their twins in ordinary lines of PG$(4,2)$ of type $\{\alpha,\beta,\gamma \}$. Finally, we meet a type-V Veldkamp line, sketched in Figure \ref{fig5}e, which is endowed with two hyperplanes of type three and a single one of type two, and whose core is isomorphic to the Pasch configuration; hence, we have ${6 \choose 2} = 15$ Veldkamp lines of this type, being all representatives of ordinary lines of PG$(4,2)$ of type $\{\beta, \gamma, \gamma \}$.
Before embarking on the final case to be dealt with in detail, it is worth having a closer look at our $(15_4,20_3)$-configuration and pointing out its intimate relation with famous Pascal's Mystic Hexagram. If six arbitrary points are chosen on a conic section and joined by line segments in any order to form a hexagon, then the three pairs of opposite sides of the hexagon meet in three points that lie on a straight line, the latter being called the Pascal line. Taking the permutations of the six points, one obtains 60 different hexagons. Thus, the so-called complete Pascal hexagon determines altogether 60 Pascal lines, which generate a remarkable configuration of 146 points and 110 lines called the {\it hexagrammum mysticum}, or the complete Pascal figure
(for the most comprehensive, applet-based representation of this remarkable geometrical object, see \cite{norma}).
Both the points and lines of the complete Pascal figure split into several distinct families, usually named after their discoverers in the first half of the 19-th century. We are concerned here with the 15 Salmon points and the 20 Cayley lines (see, e.\,g. \cite{lord,cory}) which form a $(15_4,20_3)$-configuration. This configuration is discussed in some detail in \cite{bogepi}, where it is also depicted (Figure 6) and called the {\it Cayley-Salmon} $(15_4,20_3)$-configuration. And it is precisely this Cayley-Salmon $(15_4,20_3)$-configuration which our 32-nionic $(15_4,20_3)$-configuration is isomorphic to. The same configuration is also portrayed in Figure 8 of \cite{norma}. In the latter work, two different views/interpretations of the configuration are also mentioned. One is as three pairwise-disjoint triangles that are in perspective from a line, in which case the centers of perspectivity are guaranteed by Desargues' theorem to also lie on a line; we just stress here that these two lines form a geometric hyperplane (of type one, see Figure \ref{fig2}). The other view of the figure takes any point of the configuration to be the center of perspectivity of two quadrangles whose six pairs of corresponding sides meet necessarily in the points of a Pasch configuration; again, the point and the associated Pasch configuration form a geometric hyperplane (of type two, see Figure \ref{fig3}). Obviously, we can offer one more view of the configuration, that stemming from the existence of type-three hyperplanes, namely as the incidence sum of a Desargues configuration and three triangles on a commmon side (see Figure \ref{fig4}).
\section{64-nions and a $(21_5,35_3)$-configuration}
The final algebra we shall treat in sufficient detail is $A_6$, or the 64-nions.
From the corresponding multiplication table, which due to its size we do not show here but which is freely available at {\tt http://jjj.de/tmp-zero-divisors/mult-table-64-ions.txt}, we infer the existence of 651 distinguished triples of imaginary units. Regarding the 63 imaginary units of 64-nions as points and the 651 distinguished triples of them as lines, we obtain a point-line incidence structure where each line has three points and each point is on 31 lines, and which is isomorphic to PG$(5,2)$. Following the usual procedure, we find that 350 lines of this space are defective and 301 ordinary. Likewise the preceding case, we encounter {\it three} different types of points in our 64-nionic PG$(5,2)$:
35 $\alpha$-points, each of which is on 21 defective and 10 ordinary lines; 21 $\beta$-points, each of which is on 15 defective and 16 ordinary lines; and seven $\gamma$-points, each of them being on 31 ordinary (and, hence, on zero defective) lines. This stratification of the point-set of PG$(5,2)$ leads, in turn, to three different kinds of defective lines and four distinct kinds of ordinary lines. Out of 350 defective lines, we find 105 of type $\{\alpha, \alpha, \alpha\}$, 210 of type $\{\alpha, \alpha, \beta\}$ and 35 of type $\{\beta, \beta, \beta\}$. On the other hand, 301 ordinary lines are partitioned into 105 guys of type $\{\alpha, \beta, \beta\}$, 70 of type $\{\alpha, \alpha, \gamma\}$, 105 of type $\{\alpha, \beta, \gamma\}$ and 21 of type $\{\beta, \gamma, \gamma\}$.
\pagebreak
The Veldkamp space mimicking such a fine structure of PG$(5,2)$ is that of a particular
$(21_5,35_3)$-configuration, $\mathcal{C}_6$, that also lives in our PG$(5,2)$ and whose points are the 21 $\beta$-points and whose lines are the 35 defective lines of $\{\beta, \beta, \beta\}$ type. To visualise this configuration, we build it around the model of the Cayley-Salmon $(15_4,20_3)$-configuration of 32-nions shown in Figure \ref{fig1}. Given the Cayley-Salmon configuration, there are six points and 15 lines to be added to yield our $(21_5,35_3)$-configuration, and this is to be done in such a way that the configuration we started with forms a geometric hyperplane in it. As putting all the lines into a single figure would make the latter look rather messy, in Figure \ref{f64} we briefly illustrate this construction by drawing six different figures, each featuring all six additional points (gray) but only five out of 15 additional lines (these lines being also drawn in gray color), namely those passing through a selected additional point (represented by a doubled circle).
Employing this handy diagrammatical representation, one can verify that our $(21_5,35_3)$-configuration exhibits 63 geometric hyperplanes that fall into three distinct types. A type-one hyperplane consists of a line and its complement, which is the Pasch configuration; there are 35 distinct hyperplanes of this form, each corresponding to an $\alpha$-point of our PG$(5,2)$. A type-two hyperplane comprises a point and its complement, which is the Desargues configuration; there are 21 hyperplanes of this form, each having a $\beta$-point for its PG$(5,2)$ counterpart. Finally,
a type-three hyperplane is isomorphic to the Cayley-Salmon configuration; there are seven distinct guys of this type, each answering to a
$\gamma$-point of the PG$(5,2)$. We leave it with the interested reader to verify by themselves that the Veldkamp space of our $(21_5,35_3)$-configuration indeed features 651 lines that do fall into the above-mentioned seven distinct kinds.
\begin{figure}[t]
\centerline{\includegraphics[width=4truecm]{f64a.eps}\hspace*{.6cm}\includegraphics[width=4truecm]{f64b.eps}\hspace*{.6cm}\includegraphics[width=4truecm]{f64c.eps}}
\vspace*{.5cm}
\centerline{\includegraphics[width=4.4truecm]{f64d.eps}\hspace*{.6cm}\includegraphics[width=4truecm]{f64e.eps}\hspace*{.6cm}\includegraphics[width=4truecm]{f64f.eps}}
\caption{An illustration of the structure of the $(21_5,35_3)$-configuration, built around the model of the Cayley-Salmon $(15_4,20_3)$-configuration shown in Figure \ref{fig1}.}
\label{f64}
\end{figure}
As in the previous two cases, we shall briefly describe a couple of interesting views of our $(21_5,35_3)$-configuration, both related to type-one hyperplanes.
The first one is as four triangles in perspective from a line where the points of perspectivity of six pairs of
them form a Pasch configuration, the line and the Pasch configuration comprising a geometric hyperplane (compare with the first view of both the Desargues and the Cayley-Salmon configuration). This is sketched in Figure
\ref{f64view}, where the four triangles are denoted, in boldfacing, by green, red, yellow and blue color, the
line of perspectivity by boldfaced gray color, and the points of perspectivity of
pairs of triangles (together with the corresponding lines they lie on and that are also boldfaced) by black color.
The other view is as three complete quadrangles that are pairwise in perspective in such a way that the three points of perspectivity lie on a line and where the six triples of their corresponding sides meet at points located on a Pasch configuration, again the line and the Pasch configuration forming a geometric hyperplane (compare with the second view of both the Desargues and the Cayley-Salmon configuration).
\begin{figure}[t]
\centering
\includegraphics[width=6truecm, angle=90]{f64view.eps}
\caption{A `generalized Desargues' view of the $(21_5,35_3)$-configuration.}
\label{f64view}
\end{figure}
\section{$2^N$-nions and a $\left({N+1 \choose 2}_{N-1}, {N+1 \choose 3}_{3}\right)$-configuration}
At this point it is quite easy to spot the general pattern. If one also includes
the trivial cases of complex numbers ($N=1$), where the relevant geometry is just a single point i.\,e. the $(1_0,0_3)$-configuration, and quaternions ($N=2$), whose geometry is a single line i.\,e. the $(3_1,1_3)$-configuration, we obtain the following nested sequence of configurations
whose Veldkamp spaces capture the stratification/partition of the point- and line-sets of the $2^N$-nionic PG$(N-1,2)$, $N$ being a positive integer,
$$(1_0,0_3),$$ $$(3_1,1_3),$$ $$(6_2,4_3),$$ $$(10_3,10_3),$$ $$(15_4,20_3),$$ $$(21_5,35_3),$$
$$\ldots,$$
$$\left({N+1 \choose 2}_{N-1}, {N+1 \choose 3}_{3}\right),$$
$$\ldots.$$
It is curious to notice that the first entry represents a {\it triangular} number, while the second one is a {\it tetrahedral} number, or triangular pyramidal number. In other words, we get a nested sequence of {\it binomial} $({r+k-1 \choose r}_r, {r+k-1 \choose k}_k)$-configurations with $r = N-1$ and $k=3$, whose properties have very recently been discussed in a couple of interesting papers \cite{prap,pepr}. The first few configurations are shown, in a form where the configurations are {\it nested} inside each other, in Figure \ref{nested2}.
\begin{figure}[t]
\centering
\includegraphics[width=14truecm]{nested2.eps}
\caption{A nested hierarchy of finite $\left({N+1 \choose 2}_{N-1}, {N+1 \choose 3}_{3}\right)$-configurations of $2^N$-nions for $1 \leq N \leq 5$ when embedded in the Cayley-Salmon configuration ($N=5$).}
\label{nested2}
\end{figure}
A particular character of this nesting is reflected in the structure of geometric hyperplanes. Denoting our generic $\left({N+1 \choose 2}_{N-1}, {N+1 \choose 3}_{3}\right)$-configuration by ${\cal C}_N$, we can express the types of geometric hyperplanes of the above-discussed cases in a compact form as follows
\bigskip
\begin{tabular}{llllll}
${\cal C}_1$: & $\varnothing$,& & & & \\
${\cal C}_2$: & ${\cal C}_1$, & & & & \\
${\cal C}_3$: & ${\cal C}_2$, & ${\cal C}_1 \sqcup {\cal C}_1$, & & & \\
${\cal C}_4$: & ${\cal C}_3$, & ${\cal C}_2 \sqcup {\cal C}_1$, & & & \\
${\cal C}_5$: & ${\cal C}_4$, & ${\cal C}_3 \sqcup {\cal C}_1$, & ${\cal C}_2 \sqcup {\cal C}_2$, & & \\
${\cal C}_6$: & ${\cal C}_5$, & ${\cal C}_4 \sqcup {\cal C}_1$, & ${\cal C}_3 \sqcup {\cal C}_2$, & &
\end{tabular}
\bigskip
\noindent
which implies the following generic hyperplane compositions
\bigskip
\begin{tabular}{llllll}
${\cal C}_N$: & ${\cal C}_{N-1}$, & ${\cal C}_{N-2} \sqcup {\cal C}_1$, & ${\cal C}_{N-3} \sqcup {\cal C}_2$, & \ldots, & ${\cal C}_{\frac{N}{2}} \sqcup {\cal C}_{\frac{N}{2}-1}$,
\end{tabular}
\bigskip
\noindent
or
\bigskip
\begin{tabular}{llllll}
${\cal C}_N$: & ${\cal C}_{N-1}$, & ${\cal C}_{N-2} \sqcup {\cal C}_1$, & ${\cal C}_{N-3} \sqcup {\cal C}_2$, & \ldots, & ${\cal C}_{\lfloor\frac{N}{2}\rfloor} \sqcup {\cal C}_{\lfloor\frac{N}{2}\rfloor}$,
\end{tabular}
\bigskip
\noindent
according as $N$ is even or odd, respectively; here, the symbol `$\sqcup$' stands for a disjoint union of two sets.
In the spirit of previous sections, let us also have a closer look at the nature of our generic ${\cal C}_N$. To this end, we first recall the following observations. ${\cal C}_4$, the Desargues configuration, can be viewed as ($4-2=$) two triangles in perspective from a line which are also perspective from a point, that is ${\cal C}_1$; the line and the point form a geometric hyperplane of ${\cal C}_4$. Next, ${\cal C}_5$, the Cayley-Salmon configuration, admits a view as ($5-2=$) three triangles in perspective from a line where the points of perspectivity of three pairs of them are on a line, {\it aka} ${\cal C}_2$; the two lines form a geometric hyperplane of ${\cal C}_5$. Finally, ${\cal C}_6$, our $(21_5,35_3)$-configuration, can be treated as ($6-2=$) four triangles in perspective from a line where the points of perspectivity of six pairs of them lie on a Pasch configuration, {\it alias} ${\cal C}_3$; the line and the Pasch configuration form a geometric hyperplane of ${\cal C}_6$. Generalizing these observations, we conjecture that for {\it any} $N \geq 4$, ${\cal C}_N$ can be regarded as $N-2$ triangles that are in perspective from a line in such a way that the points of perspectivity of ${N-2 \choose 2}$ pairs of them form the configuration isomorphic to ${\cal C}_{N-3}$, where the latter and the axis of perspectivity form a geometric hyperplane of ${\cal C}_N$.
Next, we invoke the concept of combinatorial Grassmannian (see, e.\,g., \cite{pra1,pra2}). Briefly, a combinatorial Grassmannian $G_k(|X|)$, where $k$ is a positive
integer and $X$ is a finite set, is a point-line incidence structure whose points are $k$-element subsets of $X$ and whose lines are $(k + 1)$-element subsets of $X$, incidence being inclusion. It is known \cite{pra1} that if $|X| =
N+1$ and $k=2$,
$G_2(N+1)$ is a binomial $\left({N+1 \choose 2}_{N-1}, {N+1 \choose
3}_{3}\right)$-configuration; in particular, $G_2(4)$ is the Pasch configuration,
$G_2(5)$ is the Desargues configuration and $G_2(N+1)$'s with $N \geq 5$ are called {\it
generalized} Desargues configurations. Now, from our detailed examination of the four cases it follows that ${\cal C}_3$, ${\cal C}_4$, ${\cal C}_5$, ${\cal C}_6$, \ldots, ${\cal C}_N$ are endowed with 1, 5, 15, 35, \ldots, ${N+1 \choose 4}$ Pasch configurations. And as ${N+1 \choose 4}$ is also the number of Pasch configurations in $G_2(N+1)$, $N \geq 3$, we are also naturally led to conjecture that ${\cal C}_N \cong G_2(N+1)$.
From what we have found in the previous sections it follows that this property indeed holds for $1 \leq N \leq 6$, being illustrated for
$N=5$ and $N=6$ in Figure \ref{grass}.
\begin{figure}[pth!]
\centerline{\includegraphics[width=7.0truecm]{grassl.eps}\hspace*{.6cm}\includegraphics[width=6truecm]{grassr.eps}}
\caption{{\it Left}: -- A diagrammatical proof of the isomorphism between ${\cal C}_5$ and $G_2(6)$.
The points of ${\cal C}_5$ are labeled by pairs of elements from the set $\{1,2,\ldots, 6\}$ in such a way that each line of the configuration is indeed of the form $\{\{a,b\}, \{a,c\}, \{b,c\}\}$, $a \neq b \neq c \neq a$. {\it Right}: -- A pictorial illustration of ${\cal C}_6 \cong G_2(7)$. Here, the labels of six additional points are only depicted, the rest of the labeling being identical to that shown in the left-hand side figure.}
\label{grass}
\end{figure}
\section{Conclusion}
An intriguing finite-geometrical underpinning of the multiplication tables of Cayley-Dickson algebras $A_N$, $3 \leq N \leq 6$, has been found that admits generalization to any higher-dimensional $A_N$. This started with an observation that the multiplication properties of imaginary units of the algebra $A_N$ are encoded in the structure of the projective space PG$(N-1,2)$. Next, this space was shown to possess a refined structure stemming from particular properties of triples of imaginary units forming its lines. To account for this refinement, we employed the concept of Veldkamp space of point-line incidence structure and found out the latter to be a binomial $\left({N+1 \choose 2}_{N-1}, {N+1 \choose 3}_{3}\right)$-configuration ${\cal C}_N$; in particular, ${\cal C}_3$ (octonions) was found to be isomorphic to the Pasch $(6_2,4_3)$-configuration, ${\cal C}_4$ (sedenions) to the famous Desargues $(10_3)$-configuration, ${\cal C}_5$ (32-nions) to the Cayley-Salmon $(15_4,20_3)$-configuration found in the well-known Pascal mystic hexagram and ${\cal C}_6$ (64-nions) was shown to be identical with a particular
$(21_5,35_3)$-configuration that can be viewed as four triangles in perspective from a line where the points of perspectivity of six pairs of
them form a Pasch configuration.
These configurations are seen to form a remarkable nested pattern, where ${\cal C}_{N-1}$ is embedded in ${\cal C}_N$ as its geometric hyperplane, that naturally reflects the spirit of the Cayley-Dickson recursive construction of corresponding algebras.
It is a well-known fact that the only first four algebras $A_N$, $0 \leq N \leq 3$, are `well-behaving' in the sense of being normed, alternative and devoid of zero-divisors --- the facts that are frequently offered as an explanation why a relatively little attention has been paid so far to their higher-dimensional cousins, these latter being even regarded by some scholars as `pathological.' It may well be that our finite-geometric, Veldkamp-space-based approach will be able to shed a novel, unexpected light at this issue as it is only starting with $N=4$ when ${\cal C}_N$ is found to feature a `generalized Desargues property' in the sense that it can be interpreted as $N-2$ triangles that are in perspective from a line in such a way that the points of perspectivity of ${N-2 \choose 2}$ pairs of them form the configuration isomorphic to ${\cal C}_{N-3}$. Or, in a slightly different form, it is only for $N \geq 4$ when ${\cal C}_N$ contains {\it Desargues} configurations, these occurring as components of its geometric hyperplanes at that.
\section*{Acknowledgments}
This work was partially supported by the VEGA Grant Agency, Project 2/0003/13, as well as by the Austrian Science Fund (Fonds zur F\"orderung der Wissenschaftlichen Forschung (FWF)), Research Project M1564--N27. We would like to express our sincere gratitude to Hans Havlicek (Vienna University of Technology) for a number of critical remarks on the first draft of the paper.
Our special thanks go to J\"org Arndt who computed for us multiplication tables of octonions, sedenions, 32-nions and 64-nions. We are also very grateful to Boris Odehnal (University of Applied Arts, Vienna) for creating nice computer versions of several figures.
\vspace*{-.1cm}
|
\section{Introduction}
\subsection{Aim and scope}
The aim of this article is to define and to study properties and estimation procedures for Bregman extension of the superquantile defined in \cite{rocury00} or in \cite{ref24} (see also \cite{rocroy13}, \cite{rocroy13a} and references therein). In the introduction we first recall the necessary conditions for a measure of risk to be coherent. Further in Section 2 we present the superquantile as a partial response to this problem. We also introduce the Bregman superquantile and study axioms of a coherent measure of risk for this quantity. In Section 3 we seek to estimate this Bregman superquantile, we introduce a plug-in estimator and study its convergence and its asymptotic normality. Some numerical simulations are shown in Section 4. An application on real data of radiological exposure is given in Section 5. All the proofs are postponed to Section 6.
\subsection{Coherent measures of risk}
Let $X$ be a real-valued random variable and let $F_X$ be its cumulative distribution function. We define for $u\in]0,1[$, the quantile function
$$F^{-1}_X(u):=\inf\{x: F_X(x)\geq u\}.$$
A usual way to quantify the risk associated with $X$ is to consider, for a given number $\alpha\in]0,1[$ close to 1, its lower quantile $q_\alpha^X:=F^{-1}_X(\alpha)$.
Nevertheless, the quantile is not a subadditive function of $X$, a major property in some applications (e.g. finance, see \cite{coherent}). Moreover, the quantile does not give any information about what is happening in the distribution-tail above the quantile (which can be dangerous when we deal with insurance premiums for example). In \cite{def} a new quantity called superquantile satisfying the property of subadditivity and giving more information about the distribution-tail is introduced. The superquantile is defined by
$$Q_{\alpha}:=Q_{\alpha}(X)= \mathbb{E}(X | X \geq q_{\alpha}^X)=\mathbb{E}(X | X \geq F_X^{-1}(\alpha))$$
We can notice that $Q^{\alpha}$ is always well defined as an element of $\bar{\mathbb{R}}= \mathbb{R} \cup \{ + \infty\}$. Indeed, if the expectation is not finite we may set it to $+ \infty$. We indeed have
$$Q_{\alpha} = \frac{\mathbb{E} \left(X \mathbf{1}_{X \geq F_X^{-1}(\alpha)} \right)}{\mathbb{P}\left(X \geq F_X^{-1}(\alpha)\right)} \geq F_X^{-1}(\alpha) \frac{\mathbb{E}\left(\mathbf{1}_{X \geq F_X^{-1}(\alpha)}\right)}{\mathbb{P}\left(X \geq F_X^{-1}(\alpha)\right)}= F_X^{-1}(\alpha).$$
\begin{remark}
In particular, when $F_X\left(F_X^{-1} (\alpha)\right)=\alpha$, the Bayes formula gives us
$$Q_{\alpha} =\mathbb{E} \left(X | X \geq F_X^{-1}(\alpha)\right)= \mathbb{E}\left(\frac{X \mathbf{1}_{ X \geq F_X^{-1}(\alpha)}}{1-\alpha}\right).$$
\end{remark}
\vspace{0.3cm}
From now, we always deal with random variables $X$ which distribution are continuous (that is $F_X$ is continuous and then $F_X\left(F_X^{-1} (\alpha)\right)=\alpha$).
\vspace{0.3cm}
Notice that this quantity is also called conditional value at risk in other references (\cite{rocury00}, \cite{rocury13}, \cite{rocroy13a}). Further, (when $F$ is conitnuous), it is also a distortion measure of risk studied for example in \cite{ref1}, \cite{ref13}, \cite{ref14}, \cite{ref15}, \cite{ref16}, \cite{ref24}, \cite{ref25}, \cite{ref26} and \cite{lui}. In these papers, a distortion measure of risk is the quantity
$$\rho_g(X) := - \int_{\mathbb{R}} x dg\left( F_X(x) \right),$$
where $g$, called the distortion function, is a map from $[0, 1]$ to $[0, 1]$. $g$ is assumed to be nondecreasing and such that $g(0)=0$ and $g(1)=1$. Then, taking $$g(x)=\alpha \left( \dfrac{x}{\alpha} \wedge 1\right),$$ we get
$$\rho_g(X)= \alpha Q_{1-\alpha}(-X).$$
\vspace{0.3cm}
Sub-additivity is not the sole interesting property for a measure of risk (for example for financial applications). Following \cite{def} we define:
\begin{definition}
\label{deco}
Let $\mathcal{R}$ be a measure of risk that is a numerical function defined on random variables. Let $X$ and $X'$ be two real-valued random variables.
We say that $\mathcal{R}$ is coherent if, and only if, it satisfies the five following properties :
\begin{itemize}
\item[i)] Constant invariance : let $C\in\mathbb{R}$, if $X=C$ (a.s.) then $\mathcal{R}(C)=C$.
\item[ii)] Positive homogeneity : $\forall \lambda > 0$, $\mathcal{R}(\lambda X) = \lambda \mathcal{R}(X)$.
\item[iii)] Subaddidivity : $\mathcal{R}(X + X') \leq \mathcal{R}(X) + \mathcal{R}(X')$.
\item[iv)] Monotonicity : If $X \leq X'$ (a.s.) then $\mathcal{R}(X) \leq \mathcal{R}(X')$.
\item[v)] Closeness : Let $(X_h)_{h\in\mathbb{R}}$ be a collection of random variables.\\
If $\mathcal{R}(X_h) \leq 0$ and ${\displaystyle \lim_{h\rightarrow 0}||X_{h}-X||_{2}=0}$ then $\mathcal{R}(X) \leq 0$.
\end{itemize}
\end{definition}
The superquantile is a coherent measure of risk (see \cite{archiv}, \cite{croissance}, \cite{On} for direct proofs). More generally, Wang and Dhaene show in \cite{ref26} that a distortion risk measure is coherent if and only if the distortion function is concave (which holds in our case).
\begin{remark}
In the litterature, alternative set of axioms for coherent measure of risks have been considered (in \cite{cono1}, \cite{ref26} or \cite{coherent} for example, additivity for a particular class of risk (comonotonic risks) is also studied). In this paper we will only focus on the Rockafellar's definition of a coherent measure of risk (see \cite{def}).
Besides, theoretical results have also been shown for coherent measures of risk. These measures can indeed be represented by suprema of linear functionals (see for example \cite{kusu} for the Kusuoka representation or \cite{coherent} for the scenarios set representation). We will not be interested in such representations here.
\end{remark}
\section{Bregman superquantiles}
In this section the aim is to build a general measure of risk that satisfies some of the regularity axioms stated in Definition \ref{deco}. These quantities will be built by using a dissimilarity measure beetween real numbers, the Bregman divergence (see \cite{bre67}).
\subsection{Bregman divergence, mean and superquantile}
In this section we first recall the definition of the Bregman mean of a probability measure $\mu$ (see \cite{bencha89}) and define the measure of risk that we will study. To begin with, we recall the definition of the Bregman divergence that will be used to build the Bregman mean.
Let $\gamma$ be a strictly convex function, $\overline{\mathbb{R}}$-valued on $\mathbb{R}$. As usual we set
$$\dom_{\gamma} :=\{x\in\mathbb{R}:\gamma(x)<+\infty\}.$$
For sake of simplicity we assume that $\dom_{\gamma}$ is a non empty open set and that $\gamma$ is a closed proper differentiable function on the interior of $\dom_{\gamma}$ (see \cite{Rockaconvex}). From now we always consider function $\gamma$ satisfying this assumption. The Bregman divergence $d_{\gamma}$ associated to $\gamma$ (see \cite{bre67}) is a function defined on $\dom_{\gamma}\times \dom_{\gamma}$ by
$$d_\gamma(x,x'):= \gamma(x)-\gamma(x')-\gamma'(x')(x-x')\;\;, \; \; (x,x'\in \dom_{\gamma}).$$
The Bregman divergence is not a distance as it is not symmetric. Nevertheless, as it is non negative and vanishes, if and only if, the two arguments are equal, it quantifies the proximity of points in $\dom_{\gamma}$. Let us recall some classical examples of such a divergence.
\begin{itemize}
\item Euclidean. $\gamma(x)=x^2$ on $\mathbb{R}$, we obviously obtain, for $x,x'\in\mathbb{R}$,
$$d_\gamma(x,x')=(x-x')^2.$$
\item Geometric. $\gamma(x)=x\ln(x)-x+1$ on $\mathbb{R}_+^*$ we obtain, for $x,x'\in\mathbb{R}_+^*$,
$$d_\gamma(x,x')=x\ln\frac{x}{x'}+x'-x.$$
\item Harmonic. $\gamma(x)=-\ln(x)+x-1$ on $\mathbb{R}_+^*$ we obtain, for $x,x'\in\mathbb{R}_+^*$,
$$d_\gamma(x,x')=-\ln\frac{x}{x'}+\frac{x}{x'}-1.$$
\end{itemize}
Let $\mu$ be a probability measure whose support is included in $\dom_{\gamma}$ and that does not weight the boundary of $\dom_{\gamma}$. Assume further that $\gamma'$ is integrable with respect to $\mu$. Following \cite{bencha89}, we first define the Bregman mean as the unique point $b$ in the support of $\mu$ satisfying
\begin{equation}
\int d_\gamma(b, x)\mu(dx)=\min_{m\in\dom_{\gamma}}\int d_\gamma(m, x)\mu(dx).
\label{eq1}
\end{equation}
In fact, we replace the $L^{2}$ minimization in the definition of the mathematical classical expectation by the minimization of the Bregman divergence.
Existence and uniqueness come from the convexity properties of $d_\gamma$ with respect to its first argument. By differentiating it is easy to see that
$$b=\gamma^{'-1}\left[\int\gamma'(x)\mu(dx)\right].$$
Hence, coming back to our three previous examples, we obtain the classical mean in the first example (Euclidean case), the geometric mean ($\exp\int\ln(x)\mu(dx)$), in the second one and the harmonic mean ($[\int x^{-1}\mu(dx)]^{-1}$), in the third one.
Notice that, as the Bregman divergence is not symmetric,
we have to pay attention to the definition of the Bregman mean. Indeed, we have
$$\int d_\gamma(x, \mathbb{E}(X) )\mu(dx)=\min_{m\in\dom_{\gamma}}\int d_\gamma(x,m)\mu(dx).$$
We turn now to the definition of our new measure of risk.
\begin{definition}
Let $\alpha\in]0,1[$, the Bregman superquantile $Q_{\alpha}^{d_\gamma}$ is defined by
$$Q_{\alpha}^{d_\gamma}:= \gamma'^{-1}\Big(\mathbb{E}(\gamma'(X)|X \geq F_{X}^{-1}(\alpha))\Big)=\gamma'^{-1}\left[\mathbb{E}\left(\frac{\gamma'(X)\mathbf{1}_{X \geq F_{X}^{-1}(\alpha)}}{1-\alpha}\right)\right] $$
where the second egality holds because $F_X$ is continuous.
In words $Q_{\alpha}^{d_\gamma}$ satisfies (\ref{eq1}) taking for $\mu$ the distribution of $X$ conditionally to
$X \geq F_X^{-1}(\alpha)$. We now denote $Q_{\alpha}^{d \gamma}$ the Bregman superquantile of the random variable $X$ when there is no ambiguity and $Q_{\alpha}^{d \gamma}(X)$ if we need to distinguish Bregman superquantile of different distributions.
\label{desupbreg}
\end{definition}
\vspace{0.3cm}
For the same reasons as before, the Bregman superquantile is always well-defined as an element of $\bar{\mathbb{R}}$. Moreover, we can already see an advantage of the Bregman superquantile over the classical superquantile. Indeed, some real random variables are not integrable (and so the superquantile is egal to $+ \infty$), but thanks to a choice of a very regular function $\gamma$, the Bregman superquantile can be finite. For example, let us introduce $X$ from the one side Cauchy distribution, that is having density function
$$f(x)= \frac{2}{\pi(1+x^2)} \mathbf{1}_{x \geq 0}.$$
Since $X$ is not integrable, its classical superquantile is egal to $+ \infty$. Nevertheless, considering the Bregman superquantile associated to the strictly convex fonction
$\gamma(x)= x\ln(x)-x+1$, we have
$$\mathbb{E} \left( \gamma'(X) \mathbf{1}_{X \geq F^{-1}_X(\alpha)}\right) < + \infty$$
because the function $x \mapsto \frac{\ln(x)}{1+x^2}$ is integrable on $[0,+\infty[$.
\vspace{0.3cm}
\textbf{Interpretation of the Bregman superquantile :} As a matter of fact we have
\begin{equation}
Q_{\alpha}^{d \gamma}(X)=\gamma'^{-1}\left(Q_{\alpha}(\gamma'(X)\right).
\label{link}
\end{equation}
Indeed, denoting $Z=\gamma'(X)$, as $\gamma'\left(F^{-1}_{X}(\alpha)\right) = F_{Z}^{-1}(\alpha)$, so that
$$\mathbb{E}\left(\gamma'(X) |X > F_{X}^{-1}(\alpha)\right) = \mathbb{E}\left(Z |Z > F_{Z}^{-1}(\alpha)\right).$$
Thus, the Bregman superquantile can be interpreted in the same way that a superquantile under a change of scale. In other words : fix a threshold $\alpha$ and compute the corresponding quantile. Further, change the scale $X \mapsto \gamma'(X)$ and compute the corresponding mean. At last, apply the inverse change of scale to come back to the true space.
\vspace{0.3cm}
This natural idea has already been used in economy. For example, noticing that the classical Gini index does not satisfy properties that are essential to ensure a reliable modelisation, Satya et al. introduced in \cite{gini} generalized Gini index thanks to a similar change of scale allowing the index to satisfy these properties.
\vspace{0.3cm}
In our case, the main interest of this new measure of risk is also in the change of scale. Indeed, choosing a slowing varying convex function $\gamma$ leads to a more \textit{robust} risk allowing a statistical esimation with better statistical properties (we show for example in Section 3 that empirical estimator for classical superquantile is not always consistent when $X$ has a Pareto distribution, whereas it is always the case with the Bregman superquantile).
\begin{remark}
The Bregman superquantile has a close link with the weighted allocation functional in the capital allocation fields. Indeed, in \cite{eco1}, this quantity is defined as :
$$A_{w}[U, V]:= \frac{E(Uw(V))}{w(V)}$$
where $U$ and $V$ are two real random variables and $w$ is a given map from $\mathbb{R}^+$ to $\mathbb{R}^+$. Choosing $U=X$, $V=\gamma'(X)$ and $w(V)=\mathbf{1}_{V \geq Q_{\alpha}^V}$, we obtain
$$A_{w}[X, \gamma'(X)]=\gamma'\left(Q_{\alpha}^{d\gamma}(X)\right).$$
\end{remark}
\subsection{Coherence of Bregman superquantile}
The following proposition gives some conditions under which the Bregman superquantile is a coherent measure of risk.
\begin{proposition}
\label{coherence}
Fix $\alpha$ in $]0, 1[$.
$\;$
\begin{itemize}
\item[i)] Any Bregman superquantile always satisfies the properties of constant invariance and monotonicity.
\item[ii)] The Bregman superquantile associated to the function $\gamma$ is homogeneous, if and only if,
$\gamma''(x) = \beta x^{\delta}$ for some real numbers $\beta>0$ and $\delta$ (as $\gamma$ is convex, if the support of $\gamma$ is strictly included in $\mathbb{R}^{+}_*:=]0, + \infty[$ there is no condition on $\delta$ but if not, $\delta$ is an even number).
\item[iii)] If $\gamma'$ is concave and sub-additive, then subadditivity and closeness axioms both hold.
\end{itemize}
\end{proposition}
The proof of this proposition, like all the others, is differed to Section 5.
To conclude, under some regularity assumptions on $\gamma$, the Bregman superquantile is a coherent measure of risk. Let us take some examples.
\subsubsection{Examples and counter-examples}
\begin{itemize}
\item \textbf{Example 1 :} $x \mapsto x^{2}$ satisfies all the hypothesis but it is already known that the classical superquantile is subaddtive.
\item \textbf{Example 2 :} The Bregman geometric and harmonic functions satisfies the assumptions i) and ii). Moreover, their derivatives are, respectively, $x \mapsto \gamma'(x) = \ln(x)$ and $x \mapsto \gamma'(x)= \frac{x-1}{x}$ which are concave but subadditive only on $[1, + \infty[$. Then the harmonic and geometric functions satisfy iii) not for all pairs of random variables but only for pairs $(X, X')$ such that, denoting $Z:=X+X'$ we have
$$\min \left(q_{\alpha}^{X}(\alpha), q_{\alpha}^{X'}(\alpha), q_{\alpha}^{Z}(\alpha) \right) >1.$$
\item \textbf{Example 3 :} A classical strictly convex function in economy (for example for the computations of the extended Gini index in \cite{gini}) is the function $\gamma(x)=x^{\alpha}$ with $\alpha >1$ when considering random variables which support is included in $R^+:=[0, + \infty[$. This convex function satisfies axiom ii) of our proposition, so the associated Bregman superquantile is homogeneous. Moreover, $\gamma'(x)=\alpha x^{\alpha-1}$ is concave if, and only if, $\alpha < 2$. In this case, it is subadditive on $[0, + \infty[ $ as a concave function such that $f(0) \geq 0$. Finally, the Bregman superquantile associated to the function $\gamma(x)=x^{\alpha}$, $1< \alpha <2$ is a coherent measure of risk when considering non-negativ random variables.
\item \textbf{Counter-example 4 :} The subadditivity is not true in the general case. Indeed, let $\gamma(x)=\exp(x)$ and assume that $X \sim \mathcal{U}\left([0, 1]\right)$.
$$\mathbb{E}\left(\gamma'(X) \mathbf{1}_{X \geq F_{X}^{-1}( \alpha)}\right) = \int_{\alpha}^{1} \exp( x) dx = e - \exp(\alpha).$$
Then, denoting $\mathcal{R}(V)=Q_{\alpha}^{d\gamma}(V)$ for $V$ a random variable, we get
$$ \mathcal{R}( X) = \ln\left( \frac{e - \exp(\alpha)}{1-\alpha}\right).$$
Moreover,
$$\mathcal{R}(\lambda X)= \ln \left(\int_{\alpha}^{1} \exp( \lambda x) dx\right)=\ln \left( \frac{\exp(\lambda) - \exp((\alpha) \lambda)}{\lambda(1-\alpha)} \right).$$
For $\alpha=0.95$ and $\lambda =2$, we obtain
$$\mathcal{R}(2X) - 2\mathcal{R}(X) = \mathcal{R}(X+X) - \left(\mathcal{R}(X)+\mathcal{R}(X) \right) = 0.000107 >0,$$
and subadditivity fails.
We can also notice that for $\lambda=4$
$$\frac{\mathcal{R}(4 X)}{4 \mathcal{R}(X)} = 1, 000321,$$
and the positive homogeneity is not true which is coherent with the Proposition \ref{coherence} since the derivative of $\gamma$ does not fulfill the assumption.
\end{itemize}
\subsection{Remarks toward other natural properties}
We study the Bregman superquantile as a measure of risk. It is then natural to wonder if other classical properties of measure of risk are satisfied by this new quantity. Let us make some remarks.
\begin{itemize}
\item[1)] First, we can study the continuity of the Bregman superquantile. A condition for classical superquantile to be continuous, that is to have
$$X_n \underset{a.s}{\longrightarrow} X \, \, \Rightarrow Q_{\alpha}(X_n) \rightarrow Q_{\alpha}(X_n)$$
is that the sequence $(X_n)$ is equi-integrable. Then the continuity of the Bregman superquantile is true when the sequence $\left(\gamma'(X_n)\right)_n$ is equi-integrable.
We thus put forward an other advantage of the Bregman superquantile over the classical superquantile, because the transformation throught $\gamma$ can regularized the sequence and make it equi-integrable. Indeed, let us consider a sample $(X_n)$ of independent copies of $X$ where $X$ has the truncated Cauchy distribution. We have already seen that $X$ it not integrable. Then the sequence $(X_n)$ is not bounded in $L^1$ and so not equi-integrable. But, with the function $\gamma(x)=x\ln(x)-x+1$, the random variable $\gamma'(X)$ is integrable. Then the independent sample $\left(\gamma'(X_n) \right)_n$ is equi-integrable.
\item[2)] The relation exhibited in Equation (\ref{link}) allows us to deduce some properties for the Bregman superquantile from the classical superquantile. For example, Gneiting et al. show in \cite{sqeli} that the classical superquantile is not elicitable (notion introduced in \cite{eli}). Then, an easy proof by reduction show that, the Bregman superquantile is not elicitable. Likewise, Cont et al. show in \cite{robu} that the classical superquantile is not robust (in particular because it is not subadditive). A direct consequence (because the fonction $\gamma'$ is continuous and the Levy distance is the distance associated to the weak convergence) is that the Bregman superquantile is not robust either. The Bregman superquantile is another example of what Cont et al. calls the conflict between subadditivity and robustness.
\end{itemize}
\section{Estimation of the Bregman superquantile}
In this section the aim is to estimate the Bregman superquantile from a sample. We introduce a Monte Carlo estimator and study its asymptotics properties. Under regularity assumptions on the functions $\gamma$ and $F_X^{-1}$, the Bregman superquantile is consistent and asymptotically Gaussian. All along this section, we consider a function $\gamma$ satisfying our usual properties and a real-valued random variable $X$ such that $\gamma'(X)\textbf{1}_{X \geq 0}$ is integrable.
\subsection{Monte Carlo estimator}
Assume that we have at hand $(X_{1}, \dots, X_{n})$ an i.i.d sample with same distribution as $X$. If we wish to estimate $Q^{d_{\gamma}}_{\alpha}$, we may use the following empirical estimator :
\begin{equation}
\begin{aligned}
\label{estimateurSQB}
\widehat{Q_{\alpha}^{d_{\gamma}}}&= \gamma^{'-1}\left[ \frac{1}{1-\alpha} \left( \frac{1}{n} \sum_{i= \lfloor n \alpha \rfloor +1} ^{n} \gamma'(X_{(i)})\right) \right],
\end{aligned}
\end{equation}
where $X_{(1)} \leq X_{(2)} \leq \dots \leq X_{(n)}$ is the re-ordered sample built with $(X_{1}, \dots, X_{n})$.
\subsection{Asymptotics}
We give a theorem which gives the asymptotic behaviour of the Bregman superquantile. The following assumptions are usefull for our next theorem.
\begin{itemize}
\item [\textbf{H1)}] The function $\gamma' \circ F_X^{-1}$ is of class $\mathcal{C}^1$ on $]0, 1[$ and its derivative that we denote by $l_{\gamma}$ satisfies $l_{\gamma}(t)=O \left( (1-t)^{-2+\epsilon_{l_{\gamma}}} \right)$ for an $\epsilon_{l_{\gamma}}>0$ when $t$ goes to $1^-$.
\item[\textbf{H2)}] The function $\gamma' \circ F_X^{-1}$ is of class $\mathcal{C}^2$ on $]0, 1[$ and its second derivative that we denote by $L_{\gamma}$ satisfies $L_{\gamma}(t)=O \left( (1-t)^{-\frac{5}{2}+\epsilon_{L_{\gamma}}} \right)$ for an $\epsilon_{L_{\gamma}}>0$ when $t$ goes to $1^-$.
\begin{remark}
\label{petito}
Assumption \textbf{H1} implies that $X$ is absolutely continuous of density $f_X$ which is continuous and positive and that $\gamma$ is of class $\mathcal{C}^2$. It also implies that $l_{\gamma}=o \left( (1-t)^{-2} \right)$ around 1. Assumption \textbf{H2} implies that $f_X$ is also of class $\mathcal{C}^1$ and $\gamma$ of class $\mathcal{C}^3$. It also implies that $L_{\gamma}(t) = o\left((1-t)^{- \frac{5}{2}} \right)$.
\end{remark}
\end{itemize}
\begin{theorem}
\label{bregasympto}
Let $\frac{1}{2} < \alpha < 1$ and $X$ be a real-valued random variable. Let $(X_1, \dots X_n)$ be an independent sample with the same distribution as $X$.
\begin{itemize}
\item [i)] Under assumption \textbf{H1}, the estimator $\widehat{Q_{\alpha}^{d_{\gamma}}}$ is consistent in probability :
$$ \widehat{Q_{\alpha}^{d_{\gamma}}} \overset{\mathbb{P}}{\underset{n \to + \infty}{\longrightarrow}} Q_{\alpha}^{d\gamma}.$$
\item[ii)] Under assumption \textbf{H2}, the estimator $\widehat{Q_{\alpha}^{d_{\gamma}}}$ is asymptotically normal :
$$\sqrt{n} \left( \widehat{Q_{\alpha}^{d_{\gamma}}} - Q_{\alpha}^{d \gamma} \right) \underset{n \to \infty}{\Longrightarrow} \mathcal{N}\left(0, \frac{\sigma^{2}_{\gamma}}{ \left(\gamma'' \left(Q_{\alpha}^{d \gamma}(X) \right)\right)^2 (1-\alpha)^2}\right), $$
where $$\sigma^2_{\gamma} := \int_{\alpha}^1 \int_{\alpha}^1 \frac{(\min(x, y) - xy)}{f_Z(F_Z^{-1}(x))f_Z(F_Z^{-1}(y))} dx dy.$$
and $Z:=\gamma'(X)$.
\end{itemize}
\end{theorem}
\begin{remark}
Easy calculations show that we have the following equalities $$l_{\gamma}:= \frac{\gamma'' \circ F_X^{-1}}{f_X \circ F_X^{-1}},$$ $$L_{\gamma}:= \frac{\left( \gamma''' \circ F_X^{-1}) \times \left( f_X \circ F_X^{-1} \right)- \left( f_{X}' \circ F_X^{-1} \right) \times (\gamma'' \circ F_X^{-1} \right)}{\left(f_X \circ F_X^{-1}\right)^3}$$ and $$f_Z=\frac{f_X \circ \gamma'^{-1}}{\gamma'' \circ \gamma'^{-1}}.$$
\end{remark}
\begin{remark}
The second part of the theorem shows the asymptotic normality of the Bregman superquantile empirical estimator. We can then use the Slutsky's lemma to find confidence intervals. Indeed, since our estimator (\ref{estimateurSQB}) is consistent, we also have
$$\frac{\sqrt{n}}{\left( \gamma'' \circ \widehat{Q_{\alpha}^{d \gamma}}\right)} \left( \widehat{Q_{\alpha}^{d \gamma}}- Q_{\alpha}^{d \gamma}(X) \right) \Longrightarrow \mathcal{N}\left(0, \frac{\sigma^{2}_{\gamma}}{(1-\alpha)}\right).$$
\end{remark}
To prove Theorem \ref{bregasympto} we use the following results on the asymptotic properties of the superquantile (which is equivalent to deal with the Bregman superquantile when the function $\gamma$ equals to identity), thanks to the link established in Equation (\ref{link}). Asymptotic behaviour of plug-in estimator for general distortion risk measure has already been studied (see for example \cite{lui} for strong consistency and \cite{ref6} for a central limit theorem). Nevertheless, for sake of completeness we propose in this paper a simpler and self-contained proof of these results for the particular case of the superquantile. Further, our proof also allows to exhibit the explicit asymptotic variance which makes the study of Bregman superquantile estimator easier. Indeed, we can then apply these results to the sample $(Z_1, \dots, Z_n)$ where $Z_i:=\gamma'(X_i)$. We conclude by applying the continuous mapping theorem with $\gamma'^{-1}$ which is continuous thanks to assumption \textbf{H1} for the consistency, and by applying the delta-method (see for example \cite{Vander}), with the function $\gamma'^{-1}$ which is differentiable thanks to assumption \textbf{H2} and of positive derivative since $\gamma$ is strictly convex.
\vspace{0.3cm}
Then, let us consider a real valued random variable $X$ such that $X\mathbf{1}_{X \geq 0}$ is integrable. With the previous notations, if we wish to estimate the classical superquantile $Q_{\alpha}$ we may use the following empirical estimator
\begin{equation}
\label{estimateurSQ}
\begin{aligned}
Q_{\alpha}:= \frac{1}{(1-\alpha)n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^n X_{(i)}
\end{aligned}
\end{equation}
\vspace{0.3cm}
For the next proposition we need the two following assumptions.
\begin{itemize}
\item [\textbf{H3)}] The quantile function $F_X^{-1}$ is of class $\mathcal{C}^1$ on $]0, 1[$ and its derivative, denoting $l$, satisfies $l=O((1-t)^{-2+ \epsilon_l})$ for an $\epsilon_l>0$ when $t$ goes to $1^-$.
\item[\textbf{H4)}] The quantile function is $\mathcal{C}^2$ on $]0, 1[$ and its second derivative that we denote $L$ satisfies $L=O \left( (1-t)^{-\frac{5}{2}+ \epsilon_L} \right)$ for an $\epsilon_L >0$ when $t$ goes to $1^-$.
\end{itemize}
\begin{proposition}
\label{superasympto}
Let $\frac{1}{2} < \alpha < 1$, and $X$ be a real-valued random variable. Let $(X_1, \dots X_n)$ be an independent sample with the same distribution as $X$.
\begin{itemize}
\item [i)] Under \textbf{H3}, the empirical estimator (\ref{estimateurSQ}) is consistent in probability
$$ \widehat{Q_{\alpha}} \overset{\mathbb{P}}{\underset{n \to + \infty}{\longrightarrow}} Q_{\alpha}.$$
\item[ii)] Under \textbf{H4}, the empirical estimator (\ref{estimateurSQ}) is asymptotically Gaussian
$$\sqrt{n} \left( \frac{1}{n(1-\alpha)} \displaystyle\sum_{i= \lfloor n \alpha \rfloor +1}^{n} X_{(i)} - Q_{\alpha}\right) \Longrightarrow \mathcal{N}\left(0, \frac{\sigma^{2}}{(1- \alpha)^2}\right),$$
where $$\sigma^2 := \int_{\alpha}^1 \int_{\alpha}^1 \frac{(\min(x, y) - xy)}{f(F^{-1}(x))f(F^{-1}(y))}.$$
\end{itemize}
\begin{remark}
\label{appelle}
Remark \ref{petito} still holds for $l$ instead of $l_{\gamma}$ and $L$ instead of $L_{\gamma}$.
\end{remark}
\end{proposition}
\subsection{Examples of asymptotic behaviors for the classical superquantile}\label{sec:examples}
Our assumptions are easy to check in practice. Let us show some examples of the asymptotic behaviour of the estimateur of the superquantile (\ref{estimateurSQ}) by using the exponential distribution of parameter 1 and the Pareto distribution.
\subsubsection{Exponential distribution}
In this case, we have on $\mathbb{R}^{+}_{*} \, \, f(t)= \exp(-t)$,
$F(t)=1- \exp(-x)$. Then $F^{-1}(t)= - \ln(1-t)$.
\begin{itemize}
\item Consistency :
$l(t)= (1-t)^{-1} = O\left((1-t)^{-2+ \frac{1}{2}} \right)$ (when $t \mapsto 1^-$) so that the estimator (\ref{estimateurSQ}) is consistent.
\item Asymptotic normality :
$L(t)= (1-t)^{-2} = O \left((1-t)^{-\frac{5}{2}+ \frac{1}{3}} \right)$ (when $t \mapsto 1^-$). So the estimator (\ref{estimateurSQ}) is asymptotically Gaussian.
\end{itemize}
\subsubsection{Pareto distribution}
Here, we consider the Pareto distribution of parameter $a>0$ : on $\mathbb{R}^{+}_{*}$,
$F(t)= 1 - x^{-a}$, $f(t)= a x^{-a-1}$, and $F^{-1}(t) = (1-t)^{\frac{-1}{a}}$.
\begin{itemize}
\item Consistency :
$l(t)=(a (1-t)^{-1 - \frac{1}{a}})$ thus, $l(t)=O\left( (1-t)^{-2+ \epsilon_l} \right)$ (when $t \mapsto 1^-$) as soon as $a>1$ (for exemple for $\epsilon_l= \frac{1- \frac{1}{a}}{2}$). The consistency for estimator (\ref{estimateurSQ}) is true.
\item Asymptotic normality :
$L(t) = C(a) (1-x)^{-\frac{1}{a}-2}$ thus, as soon as $a>2$, $L(t)= O \left( (1-t)^{- \frac{5}{2}+ \epsilon_L} \right)$ (for example $\epsilon_L= \frac{\frac{1}{2}- \frac{1}{a}}{2}$), when $t \mapsto 1^-$. The estimator (\ref{estimateurSQ}) is asymptotically gaussian if and only if $a>2$.
\end{itemize}
\subsection{Examples of asymptotic behaviour of the Bregman superquantile}
Let us now study the same examples in the case of the Bregman superquantile and its empirical estimator (\ref{estimateurSQB}). For the exponential distribution, the conclusion is the same as in the previous case of classical superquantile. However, for the Pareto distribution, we can find a function $\gamma$ such that the estimator of the Bregman superquantile is asymptotically normal without any condition on the exponent $a$ involved in the distribution. So, the Bregman superquantile is more interesting than the superquantile in this example.
\subsubsection{Exponential distribution}
Let us show the example of the exponential distribution (for $X$) and the harmonic Bregman function (for $\gamma$). We have $\gamma'(x) = (x-1)x^{-1}$ and $F^{-1}_X(t)=- \ln(1-t)$. So that, denoting $Z= \gamma'(X)$ as in Theorem \ref{bregasympto}, $$F_Z^{-1}(t) = 1 + \frac{1}{\ln(1-t)},$$
\begin{itemize}
\item Consistency. In this case, we have, $$l_{\gamma}(t)= \frac{1}{(1-t)(\ln(1-t))^2}. $$
So, $l_{\gamma}$ is a $O \left( (1-t)^{-2+ \frac{1}{2}} \right)$ (when $t \mapsto 1^-$). The estimator (\ref{estimateurSQB}) is consistent.
\item Asymptotical normality.
$$L_{\gamma}(t) = \frac{(\ln(1-t))^2 + 2 \ln(1-t)}{(1-t)^2(\ln(1-t))^4},$$
Then $L_{\gamma}$ is $O \left( (1-t)^{-\frac{5}{2}+ \frac{1}{3}}\right)$ (when $t \mapsto 1^-$). The estimator (\ref{estimateurSQB}) is asymptotically Gaussian.
\end{itemize}
\subsubsection{Pareto distribution}
Let us now study the case of the Pareto distribution with the geometric Bregman function. We have $F^{-1}_X(t)=(1-t)^{\frac{-1}{a}}$ and $\gamma'(t)=\ln(t)$. Then $$F_Z^{-1}(t)=- \frac{1}{a} \ln(1-t).$$
\begin{itemize}
\item Consistency.
$$l_{\gamma}(t)= \frac{1}{a} \frac{1}{1-t} = O \left(\frac{1}{(1-t)^{2- \frac{1}{2}}} \right),$$
and the monotonicity is true. The estimator (\ref{estimateurSQB}) is consistent.
\item Asymptotic normality.
$$L_{\gamma}(t) = \frac{1}{a} \frac{1}{(1-t)^2}= O \left( (1-t)^{-\frac{5}{2}+ \frac{1}{3}}\right).$$
The estimator (\ref{estimateurSQB}) is consistent and asymptotically gaussian for every $a > 0$.
\end{itemize}
See next part for an illustration of these results by simulations and a summary.
\section{Numerical simulations}
In our numerical tests we simulate values from a known theoretical distribution and computing the $0.95$-quantiles and superquantiles.
For each estimated quantity, the reference value is given via a $10^6$-size random sample and a convergence study is performed from a $1
000$-size sample to a $10^5$ size sample (with a step of $500$).
In order to annihilate the effect of randomness, $50$ repetitions of each numerical experiment are made. Then, we compute
\begin{itemize}
\item The mean value of the $50$ estimations to be compared to the reference value,
\item The standard deviation of the $50$ estimations. It allows to compute an experimental $95\%$-confidence interval (CI) to be compared to the theoretical $95\%$-CI (given by the central limit theorem).
\end{itemize}
Each is composed of four plots of convergence for the following quantities: quantile (up left), classical superquantile (up right), geometrical superquantile (bottom left) and harmonic superquantile (bottom right).
Each superquantile convergence plot is composed of the following curves: Reference value (dotted black line), mean estimated values (red circles), theoretical $95\%$-CI (dashed black line) and experimental $95\%$-CI (solid blue line).
Figure \ref{fig:SQexp1} gives the results for an exponential distribution of parameter $\lambda=1$.
As predicted by the theory (see Section \ref{sec:examples}), for the three different superquantiles, the consistency is verified while the experimental CI perfectly fits the theoretical CI (given by the central limit theorem).
\begin{figure}[!ht]
$$\includegraphics[height=8cm]{exp1.png}$$
\caption{Numerical convergence test for the exponential distribution.}\label{fig:SQexp1}
\end{figure}
We then test the Pareto distribution (see Section \ref{sec:examples}) with three different shape parameters: $a=0.5$, $a=1.5$ and $a=2.5$.
Figures \ref{fig:SQpareto0.5} ($a=0.5$), \ref{fig:SQpareto1.5} ($a=1.5$) and \ref{fig:SQpareto2.5} ($a=2.5$) give the convergence results.
For the geometrical and harmonic superquantiles, as predicted by the theory (see Section \ref{sec:examples}), the consistency of the Monte Carlo estimation is verified while the experimental CI perfectly fits the theoretical CI (asymptotic normality).
For the classical superquantile, we distinguish three different behaviors:
\begin{itemize}
\item No consistency for $a=0.5$ (Figure \ref{fig:SQpareto0.5}) (theory predicts consistency only if $a>1$),
\item Consistency but no asymptotic normality for $a=1.5$ (Figure \ref{fig:SQpareto1.5}) (theory predicts asymptotic normality only if $a>2$),
\item Consistency and asymptotic normality for $a=2.5$ (Figure \ref{fig:SQpareto2.5}),
\end{itemize}
\begin{figure}[!ht]
$$\includegraphics[height=8cm]{pareto0,5.png}$$
\caption{Numerical convergence test for the Pareto distribution ($a=0.5$).}\label{fig:SQpareto0.5}
\end{figure}
\begin{figure}[!ht]
$$\includegraphics[height=8cm]{pareto1,5.png}$$
\caption{Numerical convergence test for the Pareto distribution ($a=1.5$).}\label{fig:SQpareto1.5}
\end{figure}
\begin{figure}[!ht]
$$\includegraphics[height=7cm]{pareto2,5.png}$$
\caption{Numerical convergence test for the Pareto distribution ($a=2.5$).}\label{fig:SQpareto2.5}
\end{figure}
To sum up the two previous parts, the plug in estimators (\ref{estimateurSQ}) and (\ref{estimateurSQB}) seem to have the same asymptotic behaviour when considering distribution with not too heavy tail-distribution, like the exponential distribution. Nevertheless, the estimator of the Bregman superquantile (\ref{estimateurSQB}) has better statistical properties when we deal with heavy tail-distribution. A typical example is the Pareto distribution. Indeed, with Parato distribution of parameter $a>2$, the tail is not too heavy and the both estimator have good asymptotic properties. This not the case any more when we choose parameter $a<2$. When $1<a<2$, the estimator of the classical superquantile (\ref{estimateurSQ}) is no more asymptotically gaussian and when $0<a<1$ it is even not consistent, whereas the estimator for Bregman superquantile (\ref{estimateurSQB}) keeps good asymptotic properties in each case.
\section{Applications to a nuclear safety exercise}
GASCON is a software developed by CEA (French Atomic Energy Commission) to study the potential chronological
atmospheric releases and dosimetric impact for nuclear facilities safety assessment \cite{ioovan06}.
It evaluates, from a fictitious radioactive release,
the doses received by a population exposed to the cloud of radionuclides and through the food chains. It takes into account the interactions that exist between humans, plants and animals, the different pathways of transfer (wind, rain, \ldots), the distance between emission and observation, and the time from emission.
As GASCON is relatively costly in computational time, in \cite{ioovan06}, the authors have built metamodels (of polynomial form) of GASCON outputs in order to perform uncertainty and sensitivity analysis.
As in \cite{jaclav06}, we focus on one output of GASCON, the annual effective dose in $^{129}$I received in one year by an adult who lives in the neighborhood of a particular nuclear facility.
Instead of the GASCON software, we will use here the metamodel of this output which depends on $10$ input variables, each one modelled by a log-uniform random variable (bounds are defined in \cite{ioovan06}).
The range of the model output stands on several decades ($10^{-14}$ to $10^{-11}$ Sv/year) as shown by Figure \ref{fig:gascon-hist} which represents the histogram (in logarithmic scale) of $10^6$ simulated values.
\begin{figure}[!ht]
$$\includegraphics[height=7cm]{densGascon_logunif.png}$$
\caption{Distribution of the GASCON output variable.}\label{fig:gascon-hist}
\end{figure}
For this kind of numerical simulation exercises, we can be typically interested by safety criteria as $95\%$-quantile and its associated superquantiles.
The idea is to compare these values to regulatory limits or to results coming from other scenarios or from other tools.
In practice, the number of simulations performed with the GASCON model is several hundreds.
Table \ref{tab:gascon-SQ} gives the estimated values of the quantile and superquantiles for $1000$ metamodel simulations.
Figure \ref{fig:gascon-SQ} shows the relative errors (computed by averaging $1000$ different estimations) which are made when estimating the superquantiles using $3$ different Bregman divergences and with different sampling sizes.
We observe that geometrical and harmonic superquantiles are clearly more precise than the classical one. Using such measures is therefore more relevant when performing comparisons.
\begin{table}[!ht]
\caption{Estimated values of $95\%$-quantile and $95\%$-superquantiles for $1000$ simulations.}\label{tab:gascon-SQ}
\begin{tabular}{c|c|c|c}
Quantile & Classical & Geometrical & Harmonic \\
& superquantile & superquantile & superquantile\\
\hline
$1.304\times 10^{-13}$ & $4.769\times 10^{-13}$ & $3.316\times 10^{-13}$ & $2.637\times 10^{-13}$\\
\end{tabular}
\end{table}
\begin{figure}[!ht]
$$\includegraphics[height=7cm]{errorGascon-SQ.png}$$
\caption{Evolution of the relative errors (mean square error divided by the reference value) on the estimated superquantiles in function of the sample size.
In black: classical superquantile; in red: geometrical superquantile; in blue: harmonic superquantile.
The reference value has been calculated with $10^7$ simulations.}\label{fig:gascon-SQ}
\end{figure}
\section{Proofs}
\subsection{Proof of the Proposition \ref{coherence} : coherence of the Bregman superquantile}
\begin{proof}
\textbf{Proof of i) :}
First we obviously have $ Q_{\alpha}^{d\gamma}(C) = \gamma'^{-1}( \gamma'(C))=C$.
The monotonicity property is well-known for the superquantile (see for example \cite{archiv}, \cite{croissance} or \cite{On}). Then, \eqref{link} and the monotonicity of $\gamma'^{-1}$ and $\gamma'$ (since $\gamma$ is strictly convex, its derivative $\gamma'$ is strictly non-decreasing and so is its inverse function $\gamma'^{-1}$) allow us to conclude.
\vspace{0.3cm}
\textbf{Proof of ii) :}
Let us reformulate the problem. For every (measurable) function $f$ and for every random variable $X$, we denote $$\mathbb{E}_\alpha[f(X)] = \mathbb{E}[f(X)|X\geq F_X^{-1}(\alpha)].$$ Let $X$ and $X'$ be two real-valued random variable.
The Bregman superquantile of $X$ associated to $\gamma$ is
\[\gamma'^{-1}\left(\mathbb{E}_\alpha[\gamma'( X)]\right)\,.\]
According to Definition \ref{deco}, the Bregman superquantile associated $\gamma$ is \emph{homogeneous} if, for every random variable $X$ and every $\lambda>0$,
\begin{equation}
\label{preuveaurelien}
\begin{aligned}
\gamma'^{-1}\left(\mathbb{E}_\alpha[\gamma'(\lambda X)]\right) &= \lambda \gamma'^{-1}\left(\mathbb{E}_\alpha[\gamma'(X)]\right)\\
\end{aligned}
\end{equation}
As $\gamma'$ and $x\mapsto (\gamma'(x)-\gamma'(1))/\gamma''(1)$ yield the same superquantiles, one may assume without loss of generality that $\gamma'(1)=0$ and that $\gamma''(1) = 1$
First, it is easy to check that the condition given is sufficient. For simplicity, we write $\phi = \gamma'$. Let us show that (\ref{preuveaurelien}) holds for each possible form of $\Phi$ given in the proposition.
If $\phi(x) = (x^\beta-1)/\beta$, then $\phi^{-1}(y) = (1+\beta y)^{1/\beta}$ and
\begin{align*}
\phi^{-1}\left(\mathbb{E}_\alpha[\phi(\lambda X)]\right) &=\left(1+\beta \mathbb{E}_\alpha\left[\frac{(\lambda X)^\beta -1}{\beta}\right]\right)^{\frac{1}{\beta}}\\
&= \lambda \left(\mathbb{E}_\alpha\left[X^\beta\right] \right)^{1/\beta}\\
&= \lambda\left(1+\beta \mathbb{E}_\alpha\left[\frac{X^\beta -1}{\beta}\right]\right)^{\frac{1}{\beta}}\\
&= \lambda \phi^{-1}\left(\mathbb{E}_\alpha[\phi( X)]\right) \,.
\end{align*}
If $\phi(x)= \ln(x)$, then $\phi^{-1}(y)=\exp(y)$ and
\begin{align*}
\phi^{-1}\left(\mathbb{E}_\alpha[\phi(\lambda X)]\right) &=\exp\left(\mathbb{E}_\alpha\left( \ln( \lambda X) \right) \right)\\
&= \exp\left(\mathbb{E}_\alpha\left( \ln( \lambda) \right) + \mathbb{E}_\alpha\left( \ln( X) \right) \right)\\
&= \lambda \exp \left( \mathbb{E}_\alpha \left( \ln(X) \right) \right)\\
&= \lambda \phi^{-1}\left(\mathbb{E}_\alpha[\phi( X)]\right) \,.
\end{align*}
Let us now show that the condition on $\Phi$ is necessary for (\ref{preuveaurelien}) holds. Let $y>0$. Let $\gamma$ be a strictly convex function such that its associated Bregman superquantile is positively homogeneous : for avery random variable $X$ and for every $\lambda >0$, \ref{preuveaurelien} holds. In particular, let $Y$ be a random variable with distribution $\P$ such that, denoting $a = y\wedge 1$, $\P(du) = \alpha a ^{-1}1_{[0,a]}(u)du + (1-\alpha)p\delta_y + (1-\alpha)(1-p) \delta_1$. Its quantile of order $\alpha$ is $F_Y^{-1}(\alpha) = a$. The conditional distribution of $Y$ given $Y\geq F_Y^{-1}(\alpha)$ is $(1-p)\delta_1 + p\delta_y$, and $\mathbb{E}_\alpha[\phi(Y)] = (1-p)\phi(1) + p \phi(y)$.
The positive homogeneity property (\ref{preuveaurelien}) and the assumption $\phi(1)=0$ imply that
\[\phi^{-1}\left((1-p) \phi(\lambda) + p\phi(\lambda y)\right) = \lambda \phi^{-1}\left( p\phi(y)\right)\,.\]
By assumption, the expressions on both sides are smooth in $p$ and $y$. Taking the derivative in $p$ at $p=0$ yields \[\frac{\phi(\lambda y) -\phi(\lambda)}{\phi'(\lambda)} = \lambda \frac{\phi(y)}{\phi'(1)}\,,\]
and hence, as $\phi'(1)=1$,
\[\phi(\lambda y) -\phi(\lambda) = \lambda \phi'(\lambda) \phi(y)\,.\]
By differentiating with respect to $y$, one gets
\begin{equation}\phi'(\lambda y) = \phi'(\lambda)\phi'(y)\,.\label{eq:phi}\end{equation}
Let $\psi$ be defined on $\mathbb{R}$ by $\psi(z) = \ln \left(\phi'(\exp( z))\right)$.
One readily checks that Equation~\eqref{eq:phi} yields
\[\psi\left(\ln(y) + \ln(\lambda)\right) = \psi\left(\ln(y)\right) + \psi\left(\ln(\lambda)\right)\,. \]
This equation holds for every $y,\lambda>0$. This is well known to imply the linearity of $\psi$: there exists a real number $\beta$ (which will be positive since $\gamma$ is convex) such that for all $z\in\mathbb{R}$,
\[\ln\left(\phi'(\exp(z))\right) = \psi(z) = \beta z\,.\]
Thus, $\phi'(\exp(z)) = \exp(\beta z)$, that is
$\phi'(y) = y^{\beta}$ for all $y>0$.
For $\beta = -1$, one obtains $\phi(y) = \ln(y)$. Otherwise, taking into account the constraint $\phi(1)=0$, this yields
\[\phi(y) = \frac{y^{1+\beta}-1}{1+\beta}\,.\]
\vspace{0.3cm}
\textbf{Proof of iii) : }
Let $X$ and $X'$ be two real-valued random variable. Still denoting, $\mathcal{R}(x)=Q_{\alpha}^{d \gamma}(X)$, we want to show that
$$\mathcal{R}(X+X') \leq \mathcal{R}(X)+ \mathcal{R}(X').$$
Since $\gamma$ is convex, $\gamma'$ is non-decreasing and this is the same thing to show that
$$\gamma'\left(\frac{\mathcal{R}(X + X')}{2}\right) \leq \gamma'\left(\frac{\mathcal{R}(X) + \mathcal{R}(X')}{2} \right).$$ We set $S:=X+X'$. Using the concavity of $\gamma$, we have
\begin{alignat*}{1}
\gamma' \left(\frac{\mathcal{R}(X) + \mathcal{R}(X')}{2} \right) & \leq \frac{1}{2} \left[ \gamma' \Big(\mathcal{R}(X) \Big) + \gamma' \left(\mathcal{R}(X') \right) \right]
\\
& = \frac{1}{2(1-\alpha)} \mathbb{E} \left(\gamma'(X) \mathbf{1}_{X \geq q_{\alpha}^{X}} + \gamma'(X') \mathbf{1}_{X' \geq q_{\alpha}^{X'}} \right)\\
\end{alignat*}
where we still denote $\mathcal{R}(V)=Q_{\alpha}^{d\gamma}(V)$ when $V$ is a random variable.
But
$$ \gamma' \left(\frac{\mathcal{R}(S)}{2} \right)= \frac{1}{2(1-\alpha)} \mathbb{E} \left(\gamma'(S) \mathbf{1}_{S \geq q_{\alpha}^{S}}\right). $$
So we want to show that
\[\mathbb{E} \left(\gamma'(X) \mathbf{1}_{X \geq q_{\alpha}^{X}} + \gamma'(X') \mathbf{1}_{X' \geq q_{\alpha}^{X'}} - \gamma'(S) \mathbf{1}_{S \geq q_{\alpha}^{S}} \right) \geq 0.\]
The sud-additivity hypothesis on $\gamma$ allows us to use the same argument as in \cite{coherent} for the classical superquantile :
\begin{alignat*}{1}
& \mathbb{E} \Big(\gamma'(X) \mathbf{1}_{X \geq q_{\alpha}^{X}} + \gamma'(X') \mathbf{1}_{X' \geq q_{\alpha}^{X'}} - \gamma'(S) \mathbf{1}_{S \geq q_{\alpha}^{S}} \Big)\\
& \geq \mathbb{E} \Big(\gamma'(X) \mathbf{1}_{X \geq q_{\alpha}^{X}} + \gamma'(X') \mathbf{1}_{X' \geq q_{\alpha}^{X'}} - \gamma'(X) \mathbf{1}_{S \geq q_{\alpha}^{S}} - \gamma'(X') \mathbf{1}_{S \geq q_{\alpha}^{S}} \Big)\\
& \geq \gamma'(q_{\alpha}^{X}) \mathbb{E} \Big(\mathbf{1}_{X \geq q_{\alpha}^{X}} - \mathbf{1}_{S \geq q_{\alpha}^{S}} \Big) + \gamma'(q_{\alpha}^{X'} ) \mathbb{E} \Big(\mathbf{1}_{X' \geq q_{\alpha}^{X'}} - \mathbf{1}_{S \geq q_{\alpha}^{S}} \Big) \\
& = 0. \\
\end{alignat*}
\vspace{0.3cm}
Finally, we show the closeness under the same assumption as just before. Let be $(X_{h})_{h>0}$ satisfying the hypothesis. By subadditivity we have
$$\mathcal{R}(X) \leq \mathcal{R}(X_h) + \mathcal{R}(X_{h}-X) \leq 0 + \mathcal{R}(X_{h}-X).$$
Then denoting $Y_h=X_h-X$, it is enough to show that $$Y_{h} \underset{L^{2}, \, n \to + \infty}{\longrightarrow} 0 \Longrightarrow \mathcal{R}(Y_{h}) \underset{n \to + \infty}{\longrightarrow} 0 $$ to conclude. Thanks to the concavity of $\gamma'$ we can use Jensen inequality for conditional expectation
$$ \gamma'^{-1} \left[\mathbb{E} \Big( \gamma'(Y_{h}) | Y_{h}\geq F_{Y_h}^{-1} (\alpha) \Big) \right] \leq \mathbb{E}\left(Y_{h} |Y_{h} \geq F_{Y_{h}}^{-1}(\alpha)\right).$$
We conclude with Cauchy-Schwartz inequality
$$\frac{\mathbb{E}\left(Y_{h} \mathbf{1}_{Y_{h} \geq F_{Y_{h}}^{-1}(\alpha)}\right)}{1-\alpha} \leq \mathbb{E}((Y_{h})^{2})^{\frac{1}{2}} \frac{\mathbb{E} \left(\mathbf{1}^2_{Y_{h} \geq F_{Y_{h}^{-1}(\alpha)} }\right)^{\frac{1}{2}}}{1- \alpha}= ||Y_{h}||^{2} \sqrt{1-\alpha} \underset{h \to 0}{\longrightarrow} 0.$$
\end{proof}
\subsection{Proof of Proposition \ref{superasympto} : asymptotic behavior of the plug-in estimator of the superquantile}
\subsection{Mathematical tools}
We first give some technical or classical results that we will use in the forthcoming proofs.
\subsubsection{Ordered statistics and Beta function}
\label{repres}
Let us recall some results on ordered statistics (see \cite{order}).
First of all, let $(U_{i})_{i=1 \dots n}$ be an independant sample from the uniform distribution on $[0, 1]$. Then,
\begin{equation}
\begin{aligned}
\label{consistanceuniforme}
U_{(n)} \underset{a.s}{\longrightarrow}1.
\end{aligned}
\end{equation}
and
\begin{equation}
\begin{aligned}
\label{TCLuniforme}
n \left(1-U_{(n)} \right) \underset{\mathcal{L}}{\longrightarrow}W
\end{aligned}
\end{equation}
where $W$ has an exponential distribution of parameter 1.
Let now $(Y_{i})_{i=1 \dots n+1}$ be an independent sample from the standard exponential distribution. It's well known that
\begin{equation}
\label{representation}
U_{(i)}:= \displaystyle\sum_{j=1}^{i}Y_{j}\left(\displaystyle\sum_{j=1}^{n+1}Y_{j}\right)^{-1}
\end{equation}
has the same distribution as the $i^{th}$ ordered statistics of an i.i.d sample of size $n$ uniformly distributed on $[0, 1]$, that is Beta distribution of parameters $i$ and $n-i+1$ denoted $\mathcal{B}(i, n-i+1)$.
It is also known that when $(X_i)_{1\leq i \leq n}$ is a sample of cumulative distribution $F_X$, this equality in law holds
\begin{equation}
\label{loi}
X_{(i)} \stackrel{\mathcal{L}}{=} F^{-1}_X(U_{(i)}).
\end{equation}
Recall that $\mathcal{B}(a, b)$ distribution has the following density
$$f_{\mathcal{B}(a, b)}(x) = \frac{x^{a-1}(1-x)^{b-1}}{B(\alpha, \beta)}\mathbf{1}_{x \in [0, 1]}, \, \, a, b>0 $$
where
\begin{equation}
B(a, b)= \int_0^1 t^{a-1}(1-t)^{b-1} dt = \frac{\Gamma(a) \Gamma(b)}{\Gamma(a+b)}.
\label{defbeta}
\end{equation}
A classical property of the Beta function is
\begin{equation}
\forall (x, y) \in \mathbb{R}^+, \, \, B(x+1, y) = \frac{x}{x+y} B(x, y).
\label{betasym}
\end{equation}
Generalizing the definition of the factorials, we set for $n \in \mathbb{N}^*$
$$\left(n- \frac{1}{2}\right)! := \left(n- \frac{1}{2}\right) \left(n- 1- \frac{1}{2}\right) \dots \left(\frac{1}{2}\right),$$
we have for $i \in \mathbb{N}^*, n\geq i+2$
\begin{equation}
B\left(i, n-i- \frac{5}{2} +1\right) = \frac{(i-1)! \left(n-i-2- \frac{1}{2}\right)!}{\left(n-2- \frac{1}{2}\right)!},
\label{beta}
\end{equation}
\begin{equation}
\left(n- \frac{1}{2} \right) ! = \frac{(2n)!}{(2^n)^2 n!}.
\label{factoriel}
\end{equation}
Indeed, Equation \eqref{beta} comes directly from the Definition \eqref{defbeta} and to see Equation \eqref{factoriel}, we fix $k= \frac{1}{2}$ and notice that
$$2^n(n-k)! = (2n-1)(2n-3) \dots 3 \times 1 =
\frac{(2n)!}{2n(2n-2) \dots 6 \times 4 \times 2}
= \frac{(2n)!}{2^n n!} .$$
Moreover, the cumulative distribution function of the Beta distribution is the regularized incomplete Beta function $I_x(a, b)$. This function satisfies, when $a$ and $b$ are positive integers
\begin{equation*}
\begin{aligned}
I_x(a, b) = \mathbb{P} \left( \mathcal{B}(a+b-1,x) \geq a \right)
\end{aligned}
\end{equation*}
and then, Bernstein inequality for Bernoulli distribution (see for example Theorem 8.2 of \cite{livreaurelien}) gives
\begin{equation}
\label{devincomplete}
\begin{aligned}
I_x(\alpha, \beta) \leq \exp \left( - \frac{3}{8} (a+b-1) x\right)
\end{aligned}
\end{equation}
as soon as $\frac{a}{n} \geq 2 x$.
\subsubsection{Technical lemma}
\begin{lemma}
\label{lemmetechnic}
Let $\delta >1$ and $\beta \in ]0, 1[$.
Then $n^{-1} \displaystyle\sum_{i= \lfloor n \beta \rfloor}^{n-1} \left(1 - \frac{i}{n+1}\right)^{-\delta}= \mathcal{O}\left(\sqrt{n}\right)$ if and only if $\delta \leq \frac{3}{2}.$
\end{lemma}
\begin{proof}
Let $\delta>1$.
We have to characterize the $\delta$ for which
$n^{-\frac{3}{2}}\displaystyle\sum_{i= \lfloor n \beta \rfloor}^{n-1} \left(1 - \frac{i}{n+1}\right)^{-\delta}$ is bounded when $n$ goes to infinity. Let us make the index change $j:=n+1-i$. Those sums become
$$n^{-\frac{3}{2}}\displaystyle\sum_{j=2 }^{n+1- \lfloor n \beta \rfloor} \left(1-\frac{n+1-j}{n+1} \right)^{-\delta} = \frac{n^{\frac{-3}{2}}}{(n+1)^{-\delta}}\displaystyle\sum_{j=2 }^{n+1- \lfloor n \beta \rfloor} \frac{1}{j^{\delta}} \underset{n \to + \infty}{\sim} n^{\delta - \frac{3}{2}} \left(\zeta(\delta)-1 \right),$$
where $\zeta$ denote the Zeta function. Then, if $\delta > \frac{3}{2}$, $\zeta(\delta)$ is finite and the behaviour of the sum is the same as the one of $n^{-\frac{3}{2} + \delta}$ which diverges to $+ \infty$. On the contrary, if $1 < \delta \leq \frac{3}{2}$, $\zeta(\delta)$ is still finite but $n^{-\frac{3}{2} + \delta}$ is bounded and so does the sum.
\end{proof}
\subsubsection{A corollary of Lindenberg-Feller theorem}
To prove the asymptotic normality, we use a central limit theorem which is a corollary of the Lindeberg-Feller theorem (see lemma 1 in \cite{uniform}).
\begin{proposition}
\label{corlinde}
Let $(Y_{1}, \dots, Y_{n})$ be an independent sample of exponential variables of parameter 1 and $(\alpha_{j, n})_{j \leq n, \, n\geq 2}$ be a triangle array of real numbers.
If $Q_{n}= n^{-1} \displaystyle\sum_{j=1}^{n} \alpha_{j, n} (Y_{j}-1)$ and $\sigma_{n}^{2}=\frac{1}{n} \displaystyle\sum_{j=1}^{n} \alpha_{j, n}^{2}$, then
$$ \frac{\sqrt{n}Q_{n}}{\sigma_{n}} \Longrightarrow \mathcal{N}(0, 1)$$
if and only if $\max_{1 \leq j \leq n} | \alpha_{j, n} | = o(n^{\frac{1}{2}} \sigma_{n})$.
If furthermore $\sigma_{n}$ converges in probability to $\sigma$ then by Slutsky's lemma
$$ \sqrt{n}Q_{n} \Longrightarrow \mathcal{N}(0, \sigma^{2}).$$
\end{proposition}
\subsection{Proof of i) of Theorem \ref{superasympto} : consistency of the plug-in estimator (\ref{estimateurSQ})}
\begin{proof}
We aim to show consistency of the estimator (\ref{estimateurSQ}).
Let us first notice that
$$Q_{\alpha} = \frac{\mathbb{E} \left(X \mathbf{1}_{X \geq F_X^{-1}(x)(\alpha)} \right)}{1-\alpha}
= \frac{\int_{\mathbb{R}}x \mathbf{1}_{X \geq F_X^{-1}(\alpha)} f_X(x) dx}{1- \alpha} = \frac{\int_{\alpha}^1 F_X^{-1}(y) dy}{1-\alpha}. $$
Thus, we need to show that
$$\frac{1}{n} \sum_{i = \lfloor n \alpha \rfloor }^n X_{(i)} - \int_{\alpha}^1 F_X^{-1}(y) dy \underset{n \to + \infty}{\longrightarrow} 0 \, \, a.s.$$
In the sequel we omit the index $X$ in $F_X^{-1}$ because there is no ambiguity. Let us introduce the two following quantities and show their convergence to 0 in probability. $$A_n = \frac{1}{n} \sum_{i = \lfloor n \alpha \rfloor }^n X_{(i)} - \frac{1}{n} \sum_{i = \lfloor n \alpha \rfloor }^n F^{-1}\left(\frac{i}{n+1}\right), $$ and $$B_n = \frac{1}{n} \sum_{i = \lfloor n \alpha \rfloor }^n F^{-1}\left(\frac{i}{n+1}\right) - \int_{\alpha}^1 F^{-1}(y) dy.$$
Let us first deal with $A_n$.
We know by (\ref{loi}) that $X_{(i)} \sim F_X^{-1}\left( U_{(i)} \right)$ where $U_{(i)}$ is distributed like the $i^{th}$ ordered statistic of a uniform sample. Thus, defining $U_{(i)}$ with distribution $\mathcal{B}(i, n+1-i)$, it holds that,
$$A_n=\frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^n\left( F^{-1}\left(U_{(i)}\right) - F^{-1}\left(\frac{i}{n+1} \right) \right).$$
We now need to split the sum in two parts. First, let us deal with the last term in the sum (which gives actually the biggest contribution). By the mean value theorem, there exists $w_n \in \left[ U_{(n)}, \frac{n}{n+1} \right]$ (where we use non-oriented interval) such that
$$\frac{1}{n} \left( F^{-1}(U_{(n)}) - F^{-1}\left( \frac{n}{n+1} \right) \right) = \frac{1}{n} \left( U_{(n)} - \frac{n}{n+1} \right)l(w_n).$$
Since (\ref{consistanceuniforme}) holds, and thanks to assumption \textbf{H3}, there exists a constant $C_1$ such that for $n$ big enough
\begin{equation}
\begin{aligned}
\label{pareil}
\left| \frac{1}{n} \left( F^{-1}(U_{(n)}) - F^{-1}\left( \frac{n}{n+1} \right) \right) \right|& \leq \frac{C_1}{n} \frac{1}{(1-w_n)^{2-\epsilon_l}} \left|U_{(n)} - \frac{n}{n+1} \right| \\
& \leq \frac{1}{n} \frac{C_1}{\left(1- \frac{n}{n+1} \right)^{2-\epsilon_l}} \left|U_{(n)} - \frac{n}{n+1} \right| + \frac{1}{n} \frac{C_1}{\left( 1- U_{(n)} \right)^{2-\epsilon_l}} \left|U_{(n)} - \frac{n}{n+1} \right|\\
& \leq \frac{C_1}{n} (n+1)^{2-\epsilon_l}|U_{(n)} -1 | + \frac{C_1}{n} (n+1)^{2- \epsilon} \left| \frac{n}{n+1}-1 \right| \\
& + \frac{C_1}{n} \frac{1}{\left( 1- U_{(n)} \right)^{2 - \epsilon_l}} \left| U_{(n)} -1 \right| + \frac{C_1}{n} \frac{1}{\left( 1- U_{(n)} \right)^{2 - \epsilon_l}} \left| \frac{n}{n+1} -1 \right| \\
& \leq C_1W_n\frac{(n+1)^{2-\epsilon_l}}{n^2} + C_1\frac{(n+1)^{2-\epsilon_l}}{n(n+1)} + \frac{C_1}{W_n^{1-\epsilon_l} n^{\epsilon_l}} + \frac{C_1}{W_n^{2-\epsilon_l}} \frac{n^{1-\epsilon_l}}{n+1} \\
\end{aligned}
\end{equation}
where $W_n=n\left(1- U_{(n)} \right)$. Thanks to (\ref{TCLuniforme}) and the Slutsky lemma, we have shown the convergence in probability to 0 of this term. Terms for $i=n-1$ and $i=n-2$ can be treated exactlty in the same way.
\vspace{0.3cm}
Let us now deal with the remaining sum (for i from $\lfloor n \alpha \rfloor$ to $n-3$) that we still denote $A_n$. By the mean value theorem, we will upper-bound $A_n$ by a quantity depending on $l$. Since we know the behaviour of $l$ on every compact set and in the neighborhood of 1, we begin by showing that when $n$ becomes big enough, $U_{(\lfloor n \alpha \rfloor)}$ is far away from $0$. Let $\frac{\alpha}{2}\geq \epsilon> 0$ be a positive real number. We have,
$$A_n= A_n \mathbf{1}_{U_{( \lfloor n \alpha \rfloor ) }< \epsilon} + A_n \mathbf{1}_{U_{( \lfloor n \alpha \rfloor ) }\geq \epsilon}.$$
Then, for $\eta'>0$, thanks to (\ref{devincomplete}) of recallings and because for $n$ big enough $\frac{\lfloor n \alpha \rfloor}{n} \geq \frac{\alpha}{2} \geq \epsilon$, we get
\begin{equation}
\label{zero}
\begin{aligned}
\P( A_n \mathbf{1}_{U_{( \lfloor n \alpha \rfloor )} < \epsilon} > \eta') \leq \P(U_{( \lfloor n \alpha \rfloor )} < \epsilon) \leq \exp \left( - \frac{3n \epsilon}{8} \right). \\
\end{aligned}
\end{equation}
Then, it is enough to show the convergence to 0 in probability of $A_n':=A_n \mathbf{1}_{U_{( \lfloor n \alpha \rfloor )} \geq \epsilon}$. Let us show its converges to 0 in $L^1$.
By the mean value theorem, we know that there exists $w_i \in ]U_{(i)}, \frac{i}{n+1}[$ (we do not know if $U_{(i)}$ is smaller or bigger than $\frac{i}{n+1}$ but in the sequel we still denote the segment bewteen $U_{(i)}$ and $\frac{i}{n+1}$ in this way) such that
\begin{equation*}
\begin{aligned}
\mathbb{E} ( |A_n'|) \leq \frac{1}{n} \displaystyle\sum_{i= \lfloor n-3 \alpha \rfloor}^n \mathbb{E} \left[ \Big|U_{(i)} - \frac{i}{n+1} \Big| l(w_i) \mathbf{1}_{U_{(\lfloor n \alpha \rfloor)} \geq \epsilon} \right] .\\
\end{aligned}
\end{equation*}
But, for all $i \in \{ \lfloor n \alpha \rfloor, \dots, n-3\}$, we get
$U_{(i)} \geq U_{(\lfloor n \alpha \rfloor)} \geq \epsilon$. Moreover, for $n$ big enough, we get, $\frac{i}{n+1} \geq \frac{\lfloor n \alpha \rfloor}{n+1} \geq \frac{\alpha}{2} \geq \epsilon$. Then,
$$\forall i \in \{ \lfloor n \alpha \rfloor, \dots, n-3\}, \, \, w_i \in [\epsilon, 1 [.$$
Then, according to \textbf{H3} and Remark \ref{appelle} (here since we do not deal with biggest terms in the sum we only need a weaker assumption than \textbf{H3}), we get
\begin{itemize}
\item for any arbitrary $\eta >0$, there exists $\xi_{\eta} >0$ such that
$$\forall t \in ]1 - \xi_{\eta}, 1[, \, \, l(t) \leq \frac{\eta}{(1-t)^2}$$
\item because $l$ is continuous, there exists a constant $C$ such that
$$\forall t \in [\epsilon, 1- \xi_{\eta}], \, \, l(t) \leq C.$$
\end{itemize}
Let us then look at the sum on two differents events.
\begin{equation*}
\begin{aligned}
\mathbb{E}(|A_n'|) & \leq \frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \mathbb{E} \left[ \Big|U_{(i)} - \frac{i}{n+1} \Big| l(w_i) \mathbf{1}_{U_{(\lfloor n \alpha \rfloor)} \geq \epsilon} \mathbf{1}_{w_i >1-\xi_{\eta}}\right] \\
& + \frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \mathbb{E} \left[ \Big|U_{(i)} - \frac{i}{n+1} \Big| l(w_i) \mathbf{1}_{U_{(\lfloor n \alpha \rfloor)} \geq \epsilon} \mathbf{1}_{w_i \leq 1-\xi_{\eta}}\right] \\
& := T_n^1+T_n^2\\
\end{aligned}
\end{equation*}
But,
\begin{equation*}
\begin{aligned}
T_n^{1} & \leq \frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \mathbb{E} \left[ \Big|U_{(i)} - \frac{i}{n+1} \Big| l(w_i) \mathbf{1}_{\epsilon \leq w_i \leq 1-\xi_{\eta}}\right] \\
& \leq \frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \mathbb{E} \left[ \Big|U_{(i)} - \frac{i}{n+1} \Big| C \right]\\
&\leq \frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} C \sqrt{\mbox{Var}(U_{(i)})} \\
& \leq \frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} C \sqrt{\frac{i(n-i+1)}{(n+1)^2(n+2)}} \\
&\sim \frac{1}{\sqrt{n}} \frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} C \sqrt{\frac{i}{n}\left(1-\frac{i}{n}\right)} = \mathcal{O} \left( \frac{1}{\sqrt{n}} \right).\\
\end{aligned}
\end{equation*}
thanks to convergence of Rieman's sum of the continuous function $x \mapsto \sqrt{x(1-x)}$ on $[\alpha, 1]$. Then, this terms goes to 0. Let us conclude by dealing with the last term.
\begin{equation}
\label{eqmax}
\begin{aligned}
T_n^{2} & \leq \frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \mathbb{E} \left[ \Big|U_{(i)} - \frac{i}{n+1} \Big| l(w_i) \mathbf{1}_{w_i > 1-\xi_{\eta}}\right] \\
&\leq \frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \mathbb{E} \left[ \Big|U_{(i)} - \frac{i}{n+1} \Big| \frac{\eta}{(1-w_i)^2} \right]\\
& \leq \frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \mathbb{E} \left[ \Big|U_{(i)} - \frac{i}{n+1} \Big| \max \left( \frac{\eta}{(1-U_{(i)})^2},\frac{\eta}{(1-\frac{i}{n+1})^2}\right) \right].\\
\end{aligned}
\end{equation}
But, by the Cauchy-Schwartz inequality,
\begin{equation*}
\begin{aligned}
\frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \mathbb{E} \left[ \big|U_{(i)} - \frac{i}{n+1} \Big| \frac{\eta}{(1-U_{(i)})^2} \right]& \leq \frac{\eta}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \sqrt{\mbox{Var}(U_{(i)}) \mathbb{E} \left( \frac{1}{(1-U_{(i)})^4}\right)} \\
\end{aligned}
\end{equation*}
Since $i < n-2$, all the terms of the sum can be expressed using Beta functions and then the expectation is finite (and this is for this reason that we ca not include biggest terms $i=n$, $i=n-1$ and $i=n-2$ in this reasonning). Indeed,
\begin{small}
\begin{equation*}
\begin{aligned}
\frac{\eta}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \sqrt{\mbox{Var}(U_{(i)}) \mathbb{E} \left( \frac{1}{(1-U_{(i)})^4}\right) } & \leq \frac{\eta}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \sqrt{\frac{i(n-i+1)}{(n+1)^2(n+2)}} \sqrt{\frac{B(i, n+1-i) }{\int_0^1 x^{i-1}(1-x)^{n+1-i-4} dx}} \\
& = \frac{\eta}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \sqrt{\frac{i(n-i+1)}{(n+1)^2(n+2)}} \sqrt{\frac{B(i, n+1-i-3)}{B(i, n+1-i)}} \\
& = \frac{\eta}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \sqrt{\frac{n(n-1)(n-2)}{(n+1)^2(n+2)}\frac{i(n-i+1)}{(n-i)(n-1-i)(n-2-i)}} \\
\end{aligned}
\end{equation*}
\end{small}
where we used recallings on Beta function. The final term has the same behaviour when $n$ goes to $+ \infty$ that
$$ \eta \frac{1}{n^{\frac{3}{2}}} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \frac{1}{\left( 1- \frac{i}{n} \right)^{\frac{3}{2}}}:= \eta \times V_n$$
Since Lemma \ref{lemmetechnic} implies that $V_n$ is bounded independently of $\eta$ and because $\eta$ is arbitrary small, we have shown the converge to 0. Then the term in $U_{(i)}$ on the maximum of Equation (\ref{eqmax}) converges to 0. The second term can be computed in the same way
\begin{equation*}
\begin{aligned}
\frac{\eta}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \sqrt{\mbox{Var}(U_{(i)}) \mathbb{E} \left( \frac{1}{(1-\frac{i}{n+1})^4}\right) }\sim \frac{1}{\sqrt{n}}\frac{\eta}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-3} \frac{1}{\left( 1- \frac{i}{n+1} \right)^{\frac{3}{2}}}. \\
\end{aligned}
\end{equation*}
which also converges to 0 thanks to Lemma \ref{lemmetechnic}. Finally, $A_n'$ converges to 0 in $L^1$ and so in probability. So is $A_n$.
\vspace{0.3cm}
Let us now study the term $$B_n= \frac{1}{n} \displaystyle\sum_{i = \lfloor n \alpha \rfloor }^n F^{-1}\left(\frac{i}{n+1}\right) - \int_{\alpha}^1 F^{-1}(y) dy.$$ We will show that this term converges to 0 thanks to a generalized Riemann sum convergence due to the monotonicity of $F^{-1}$.
\begin{remark}
\label{help}
To begin with, for $\epsilon >0$ it is easy to show that
$$B_n^{\epsilon}= \frac{1}{n} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{\lfloor n (1 - \epsilon) \rfloor} F^{-1}\left( \frac{i}{n+1} \right) - \int_{\alpha}^{1- \epsilon} F^{-1}(y) dy$$
converges to 0. Indeed, it is the convergence of the Riemann sum for the continuous function $F^{-1}$.
\end{remark}
Let us fix $\epsilon >0$. According to the previous remark, we split the forthcoming sum in two parts.
$$B_n= \frac{1}{n+1} \sum_{i = \lfloor n \alpha \rfloor }^{\lfloor n (1- \epsilon) \rfloor} F^{-1}\left(\frac{i}{n+1}\right) + \frac{1}{n+1} \sum_{\lfloor n (1- \epsilon) \rfloor +1}^{n-1} F^{-1}\left(\frac{i}{n+1}\right) := S_n^1 + S_n^2. $$
Since the quantile function is non-decreasing on $[\alpha, 1]$, we have :
\begin{alignat*}{1}
& \int_{\frac{\lfloor n \alpha \rfloor -1}{n+1}}^{\frac{\lfloor n (1- \epsilon) \rfloor}{n+1}} F^{-1}(t) dt + \int_{\frac{\lfloor n (1- \epsilon) \rfloor}{n+1}}^\frac{n-1}{n+1} F^{-1}(t) dt:=C_n^1+C_n^2 \\
& \leq \frac{1}{n+1} \sum_{i = \lfloor n \alpha \rfloor }^{\lfloor n (1- \epsilon) \rfloor} F^{-1}\left(\frac{i}{n+1}\right) + \frac{1}{n+1} \sum_{\lfloor n (1- \epsilon) \rfloor +1}^{n-1} F^{-1}\left(\frac{i}{n+1}\right) :=S_n^1+S_n^2\\
& \leq \int_{\frac{\lfloor n \alpha \rfloor }{n+1}}^{\frac{\lfloor n (1- \epsilon) \rfloor}{n+1}} F^{-1}(t) dt + \int_{\frac{\lfloor n (1- \epsilon) \rfloor}{n+1}}^\frac{n}{n+1} F^{-1}(t) dt:=D_n^1+D_n^2. \\
\end{alignat*}
Then, we have :
$$(C_n^1-S_n^1) + (C_n^2 - D_n^2) \leq (S_n^2 - D_n^2) \leq (D_n^1 - S_n^1) $$
If we show that $C_n^2-D_n^2$ converge to 0, we can conclude using comparison theorem, beacause the convergence of $D_n^1 - S_n^1$ and $C_n^1 - S_n^1$ to 0 is true thanks to the Remark \ref{help}. Let us then show this convergence
As in the neighborhood of 1, $l(t)=o\left((1-t)^{-2}\right)$ (Remark \ref{appelle}), we also have $F^{-1}(t) = o \left( (1-t)^{-1} \right). $
Then, for $\epsilon >0$, there exist $N$ such that for $n \geq N$ :
$$C_n^2-D_n^2 = - \int_{\frac{n-1}{n+1}}^{\frac{n}{n+1}} F^{-1}(t) dt
\leq \epsilon \int_{\frac{n-1}{n+1}}^{\frac{n}{n+1}} \frac{1}{1-t} dt
= \epsilon \ln \left(2\right). $$
Finally, $S_n^2- D_n^2$ converges to 0 a.s. So that, the same holds for $B_n$.
We have shown that $A_n+B_n$ converge to 0 in probability. So under our hypothesis, the superquantile is consistent in probability.
\vspace{0.3cm}
\begin{remark}
\label{remarkordresup}
Using the same arguments, we can show that under stronger hypothesis on the quantile function $F^{-1}(t)=o \left(\frac{1}{(1-t)^{\frac{1}{2}} }\right)$ (that is the case in ii) of Proposition \ref{superasympto}), we have
$$- \int_{\frac{n-1}{n+1}}^{\frac{n}{n+1}} F^{-1}(t) dt \leq \epsilon \int_{\frac{n-1}{n+1}}^{\frac{n}{n+1}} \frac{1}{(1-t)^{\frac{1}{2}}} dt = \epsilon -2(1- \sqrt{2}) \frac{1}{\sqrt{n}}. $$
Then
$$ \sqrt{n} \left( \frac{1}{n} \sum_{i = \lfloor n \alpha \rfloor }^n F^{-1}(\frac{i}{n+1}) - \int_{\alpha}^1 F^{-1}(y) dy \right) \underset{n\to+\infty}{\longrightarrow} 0.$$
We will use this result in the next part.
\end{remark}
\end{proof}
\subsubsection{Proof of ii) of Proposition \ref{superasympto} : asymptotic normality of the plug-in estimator (\ref{estimateurSQ})}
Let us prove the asymptotic normality of the estimator of the superquantile. To begin with, we can make some technical remarks.
\begin{remark}
\label{primitive}
The assumption on $L$ implies that there exists $\epsilon_l>0$ and $\epsilon_{F^{-1}}>0$ such that $ l(t)=O\left((1-t)^{- \frac{3}{2}+ \epsilon_l}\right)$, and $F^{-1}(t)=O\left((1-t)^{- \frac{1}{2}+ \epsilon_{F^{-1}}}\right).$ It also implies that in the neighborhood of 1, $L(t)=o\left((1-t)^{-\frac{5}{2}}\right)$.
\end{remark}
\begin{proof}
The proof stands in three steps. First we reformulate and simplify the problem and apply the Taylor Lagrange formula. Then, we show that the second order term converges to 0 in probability. In the third step, we identify the limit of the first order term.
\vspace{0.3cm}
\textbf{Step 1 : Taylor-Lagrange formula}
\vspace{0.3cm}
Let us first omit $\alpha^{-1}$.
We have to study the convergence in distribution of
$$ \sqrt{n} \left( \frac{1}{n} \displaystyle\sum_{i=\lfloor n \alpha \rfloor }^{n} X_{(i)} - \int_{\alpha}^{1} F^{-1}(y) dy \right).$$
We have already noticed (Remarks \ref{remarkordresup} and \ref{primitive}) that
$$\sqrt{n} \left[ \frac{1}{n} \displaystyle\sum_{i=\lfloor n \alpha \rfloor }^{n} F^{-1}\left(\frac{i}{n+1}\right) - \int_{\alpha}^{1} F^{-1}(y) dy \right] \underset{n \to + \infty}\longrightarrow 0 .$$
Thus, Slutsky's lemma, allows us to study only the convergence in law of
$$\sqrt{n} \left[ \frac{1}{n} \displaystyle\sum_{i=\lfloor n \alpha \rfloor }^{n} X_{(i)} - \frac{1}{n} \displaystyle\sum_{i=\lfloor n \alpha \rfloor }^{n} F^{-1}\left(\frac{i}{n+1}\right) \right]. $$
The quantile function $F^{-1}$ is two times differentiable so that we may apply the first order Taylor-Lagrange formula. Using the same argument as in the proof of i), we introduce $U_{(i)}$ a random variable distributed as a $\mathcal{B}(i, n+1-i)$. Considering an equality in law we then have
\begin{alignat*}{1}
\sqrt{n}\left(\frac{1}{n} \displaystyle\sum_{i=\lfloor n \alpha \rfloor +1}^{n} \left[ X_{(i)}-F^{-1}\left(\frac{i}{n+1}\right) \right] \right) & \stackrel{\mathcal{L}}{=} \sqrt{n} \left[\frac{1}{n} \displaystyle\sum_{i=\lfloor n \alpha \rfloor +1}^{n} \left( U_{(i)} - \frac{i}{n+1} \right) \frac{1}{f\Big(F^{-1}(\frac{i}{n+1})\Big)}\right] \\
& + \frac{1}{\sqrt{n}} \displaystyle\sum_{i=\lfloor n \alpha \rfloor +1}^{n} \left[ \int_{\frac{i}{n+1}}^{U_{(i)}} \frac{f'(F^{-1}(t))}{ \left[f(F^{-1}(t)) \right]^{3}} \left(U_{(i)} - t \right) dt \right]. \\
\end{alignat*}
Let us call $\sqrt{n} Q_{n}$ the first-order term and $R_{n}$ the second-order one.
\vspace{0.3cm}
\textbf{Step 2 : The second-order term converges to 0 in probability}
\vspace{0.3cm}
Let us show that $R_{n}$ converges to 0 in probability. We will use the same decomposition as for $A_n$. First, we deal with the last term : $i=n$. Still using (\ref{consistanceuniforme}) and \textbf{H3} we have the existence of a constant $C_2$ such that for $n$ big enough,
\begin{small}
\begin{equation*}
\begin{aligned}
\left|\frac{1}{\sqrt{n}} \int_{\frac{n}{n+1}}^{U_{(n)}} L(t) \left( U_{(n)} -t \right) dt \right| & \leq \frac{1}{\sqrt{n}} \max \left( \frac{C_2}{\left( 1- U_{(n)} \right)^{\frac{5}{2}- \epsilon_L}}, \frac{C_2}{\left( 1- \frac{n}{n+1} \right)^{\frac{5}{2}- \epsilon_L}}\right) \frac{\left(U_{(n)} - \frac{n}{n+1}\right)^2}{2}\\
& = \frac{1}{\sqrt{n}} \frac{\left(U_{(n)} - \frac{n}{n+1} \right)^2}{2} \frac{C_2}{\left(1-U_{(n)}\right)^{\frac{5}{2} - \epsilon_L}} + \frac{1}{\sqrt{n}} \frac{\left(U_{(n)} - \frac{n}{n+1} \right)^2}{2} \frac{C_2}{\left(1-\frac{n}{n+1}\right)^{\frac{5}{2} - \epsilon_L}}\\
\end{aligned}
\end{equation*}
\end{small}
which converges to 0 in probability exactly as in (\ref{pareil}) thanks to (\ref{TCLuniforme}) and Slutsky lemma. This is the same idea for the term $i=n-1$.
\vspace{0.3cm}
Let us now deal with the remaining sum (for $i$ from $ \lfloor n \alpha \rfloor$ to $n-2$) that we still denote $R_n$. We use same kind of reasonning that for $A_n$. First, we still have for $\frac{\alpha}{2}\geq \epsilon >0$,
\begin{equation*}
\begin{aligned}
R_n= R_n\mathbf{1}_{U_{( \lfloor n \alpha \rfloor)} < \epsilon} + R_n\mathbf{1}_{U_{( \lfloor n \alpha \rfloor)} \geq \epsilon}
\end{aligned}
\end{equation*}
The first term converges in probability to 0 using the same argument as in (\ref{zero}). Let us then deal with the second term that we denote $R_n'$ and show its convergence in $L^1$. Since \textbf{H4} gives also informations on a neighborhood of 1 and on every compact set, we will use the same kind of argument as before (and so Remark \ref{appelle}). For $\eta>0$, there exists, thanks to \textbf{H4}, a real number $\xi_{\eta}$ such that
\begin{itemize}
\item $\forall t \in ]1- \xi_{\eta}, 1[, |L(t)|\leq \frac{\eta}{(1-t)^{ \frac{5}{2}}}$.
\item On $[\epsilon, 1- \xi_{\eta}]$ which is a compact set, the function $L$ is bounded by a constant $C_3$.
\end{itemize}
Moreover, since $U_{( \lfloor n \alpha \rfloor)} \geq \epsilon$, we have already seen that
$$ ]U_{(i)}, \frac{i}{n+1}[ \subset [\epsilon, 1[, \, \, (i= \lfloor n \alpha \rfloor, \lfloor n \alpha \rfloor +1, \, \, \dots \, \, n).$$
Finally, we get, by denoting $m_i=\min\left( U_{(i)}, \frac{i}{n+1} \right)$ and $M_i=\max\left( U_{(i)}, \frac{i}{n+1} \right)$, $(i= \lfloor n \alpha \rfloor, \lfloor n \alpha \rfloor +1, \, \, \dots \, \, n)$,
\begin{small}
\begin{equation*}
\begin{aligned}
\mathbb{E}(|R_n'|) & \leq \frac{1}{\sqrt{n}} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-2} \mathbb{E} \left( \Big| \int_{U_{(i)}}^{\frac{i}{n+1}} \left( U_{(i)} -t \right) L(t)dt \Big| \right)\\
& \leq \frac{1}{\sqrt{n}} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-2} \mathbb{E} \left( \int_{m_i}^{1- \xi_{\eta}} \Big| U_{(i)} -t \Big| |L(t)| dt \right) + \frac{1}{\sqrt{n}} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-2} \mathbb{E} \left( \int_{1- \xi_{\eta}}^{M_i} \Big| U_{(i)} -t \Big| |L(t)| dt \right) \\
& \leq \frac{1}{\sqrt{n}} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-2} C_3 \mathbb{E} \left( \frac{\left( U_{(i)}-\frac{i}{n+1} \right)^2}{2} \right) + \frac{1}{\sqrt{n}} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-2} \mathbb{E} \left(\max \left( \frac{\eta}{\left( 1 - U_{(i)} \right)^{\frac{5}{2}}}, \frac{\eta}{\left( 1 - \frac{i}{n+1} \right)^{\frac{5}{2}}} \right) \frac{\left( U_{(i)} - \frac{i}{n+1}\right)^2}{2} \right) \\
& \leq \frac{1}{\sqrt{n}} \frac{C_3}{2(n+2)} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-2} \frac{i}{n+1} \left( 1 - \frac{i}{n+1} + \frac{1}{n+1} \right) :=S_n^1 \\
&+ \frac{1}{\sqrt{n}} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^n \mathbb{E} \left(\frac{\eta}{\left( 1 - U_{(i)} \right)^{\frac{5}{2}}} \frac{\left( U_{(i)} - \frac{i}{n+1}\right)^2}{2} \right) :=S_n^2 \\
& + \frac{1}{\sqrt{n}} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-2} \mathbb{E} \left(\frac{\eta}{\left( 1 - \frac{i}{n+1}\right)^{\frac{5}{2}} } \frac{\left( U_{(i)} - \frac{i}{n+1}\right)^2}{2} \right):=S_n^3.\\
\end{aligned}
\end{equation*}
\end{small}
$S_n^1$ converges to 0 as the Riemann sum of the continuous function $x \mapsto x(1-x)$ multiplied by $n^{- \frac{1}{2}}$. We have also
\begin{equation*}
\begin{aligned}
S_n^3 &= \frac{1}{\sqrt{n}} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-2} \frac{\eta}{\left( 1 - \frac{i}{n+1}\right)^{\frac{5}{2}}} \frac{i(n+1-i)}{(n+2)(n+1)^2} \\
\end{aligned}
\end{equation*}
which has the same behaviour when $n$ goes to $+\infty$ that
$$ \eta \frac{1}{\sqrt{n^{\frac{3}{2}}}} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-2} \frac{1}{\left(1- \frac{i}{n}\right)^{\frac{3}{2}}} := \eta V_n$$
which goes to 0 because thanks to Lemma \ref{lemmetechnic}, $V_n$ is bounded independently of $\eta$ and because $\eta$ is arbitrary small. Finally, to study the converges to 0 of $S_n^2$, we have to compute
\begin{alignat*}{1}
\mathbb{E} \left(\frac{\left(U_{(i)} - \frac{i}{n+1}\right)^2}{\left(U_{(i)} - 1\right)^\frac{5}{2}} \right) & = \frac{1}{B(i, n+1-i)} \int_0^1 x^{i-1}(1-x)^{n-i-\frac{5}{2}} \left(x- \frac{i}{n+1} \right)^2 dx \\
& = \frac{1}{B(i, n+1-i)} \Big( \int_0^1 x^{i+1}(1-x)^{n-i-\frac{5}{2}} dx -2 \frac{i}{n+1} \int_0^1 x^{i}(1-x)^{n-i-\frac{5}{2}} dx \\
& + \left(\frac{i}{n+1} \right)^2 \int_0^1 x^{i-1}(1-x)^{n-i-\frac{5}{2}} \Big). \\
\end{alignat*}
Let us call this last quantity $I_n^i$. For $i < n-1$,
\begin{alignat*}{1}
I_n^i & = \frac{1}{B(i, n+1-i)} \Big[ B\left(i+2, n-i-\frac{5}{2} +1\right) - 2 \frac{i}{n+1} B \left(i+1, n-i-\frac{5}{2} +1 \right) \\
& + \left(\frac{i}{n+1}\right)^2 B \left(i, n-i-\frac{5}{2} +1 \right) \Big]. \\
\end{alignat*}
So that using \eqref{betasym} we obtain
\begin{alignat*}{1}
I_n^i & = \frac{B \left(i, n-i-\frac{5}{2}+1\right)}{B(i, n+1-i)} \left[ \frac{i(i+1)}{\left(n-\frac{5}{2} +2\right)\left(n- \frac{5}{2} +1\right)} \right. \\
& - \left.2 \frac{i^2}{\left(n-\frac{5}{2} +1\right)(n+1)} \frac{i(i+1)}{\left(n-\frac{5}{2} +2\right)\left(n- \frac{5}{2} +1\right)}
+ \left(\frac{i}{n+1}\right)^2 \frac{i(i+1)}{\left(n-\frac{5}{2} +2\right)\left(n- \frac{5}{2} +1\right)} \right]. \\
\end{alignat*}
Let $E_n^i$ be such that $I_n^i= \frac{B(i, n-i-\frac{5}{2}+1)}{B(i, n+1-i)} E_n^i$. Expanding $E_i^n$ gives when $n$ goes to infinity
$$E_i^n \sim \frac{1}{n} \frac{i}{n+1} \left( 1- \frac{i}{n+1} \right).$$
Let us study the term $\frac{B\left(i, n-i-\frac{5}{2} +1\right)}{B(i, n+1-i)}$. Using $\eqref{defbeta}$ and $\eqref{factoriel}$, we obtain
\begin{alignat*}{1}
\frac{B\left(i, n-i-\frac{5}{2} +1\right)}{B(i, n+1-i)}
& = \frac{n!}{\left(n-2- \frac{1}{2}\right)!} \frac{\left(n-i-2- \frac{1}{2}\right)!}{(n-i)!} \\
& =\frac{n(n-1)}{(n-i-1)(n-i)} \frac{\left(2(n-i-2)\right)! ((n-2)!)^2 2^{2i}}{((n-i-2)!)^2 (2(n-2))!} \\
\end{alignat*}
Since each $i$ can be written as $i= \lfloor n \beta \rfloor$ with $\beta<1$, $n-i$ goes to infinity when $n$ goes to infinity and we can apply the Stirling formula:
\begin{alignat*}{1}
\frac{(2(n-i-2))! }{((n-i-2)!)^2 } & \underset{n\to+\infty}{\sim } \frac{\sqrt{2(n-i-2) 2 \pi} \left( \frac{2(n-i-2)}{e}\right)^{2(n-i-2)}}{2 \pi (n-i-2) \left(\frac{n-i-2}{e}\right)^{n-i-2}} \underset{n\to+\infty}{\sim} \frac{2^{2(n-i-2)}}{\sqrt{\pi(n-i-2)}}.
\end{alignat*}
Likewise,
\begin{alignat*}{1}
\frac{(2(n-2))! }{((n-2)!)^2 } & \underset{n\to+\infty}{\sim} \frac{2^{2(n-2)}}{\sqrt{\pi(n-2)}}.
\end{alignat*}
Then, when $n$ goes to infinity
$$ \frac{B\left(i, n-i-\frac{5}{2} +1\right)}{B(i, n+1-i)}\underset{n\to+\infty}{\sim} \frac{1}{(1- \frac{i}{n+1})^{\frac{5}{2}}}.$$
Hence, we obtain
$$ I_i^n =\frac{B\left(i, n-i-\frac{5}{2} +1\right)}{B(i, n+1-i)} E_i^n \underset{n\to+\infty}{\sim} \frac{1}{n} \frac{\frac{i}{n+1}}{(1- \frac{i}{n+1})^\frac{3}{2}}. $$
Finally,
$$I_i^n \leq \frac{2}{n} \frac{\frac{i}{n+1}}{(1- \frac{i}{n+1})^\frac{3}{2}}.$$
and
\begin{alignat*}{1}
\frac{\eta}{ \sqrt{n}} \displaystyle\sum_{i= \lfloor n \alpha \rfloor}^{n-2} \mathbb{E} \left( \left|L(U_{(i)}) \right| \frac{\left(U_{(i)} - \frac{i}{n+1}\right)^2}{2} \right)
& \leq \frac{\eta}{2 \sqrt{n}} \displaystyle\sum_{i=\lfloor n \alpha \rfloor}^{n-2} \mathbb{E} \left( \frac{(U_{(i)} - \frac{i}{n+1})^2}{\left(U_{(i)} - 1\right)^\frac{5}{2}} \right) \\
& \leq \frac{\eta}{ \sqrt{n}} \frac{1}{n} \displaystyle\sum_{i=\lfloor n \alpha \rfloor}^{n-2} \frac{\frac{i}{n+1}}{\left(1- \frac{i}{n+1}\right)^\frac{3}{2}} \\ & \sim \eta \frac{1}{n^{\frac{3}{2}}} \displaystyle\sum_{i=\lfloor n \alpha \rfloor}^{n-2} \frac{1}{\left(1- \frac{i}{n} \right)^{\frac{3}{2}}}.\\
\end{alignat*}
Finally, $R_n'$ converges in $L^1$ to 0 and so in probability. Hence, $R_n$ converges to 0 in probability.
\vspace{0.3cm}
\textbf{Step 3 : Identification of the limit}
\vspace{0.3cm}
Our goal is to find the limit of $\sqrt{n}Q_{n}$. Let us reorganize the expression of $Q_{n}$ to have a more classical form (sum of independent random variables) and to allow the use of the Proposition \ref{corlinde}.
Denoting by $$ \bar{Y}= \frac{\displaystyle\sum_{j=1}^{n+1}Y_{j}}{n+1},$$ we have thanks to (\ref{representation})
\begin{small}
\begin{equation*}
\begin{aligned}
Q_{n} &= \frac{1}{n}\displaystyle\sum_{i=\lfloor n \alpha \rfloor +1}^{n} \left( \frac{\displaystyle\sum_{j=1}^{i}Y_{j}}{\displaystyle\sum_{k=1}^{n+1}Y_{k}} - \frac{i}{n+1} \right) l\left(\frac{i}{n+1}\right) \\
& = \frac{1}{n}\displaystyle\sum_{j=1}^{n} \left[ \left( \frac{Y_{j}}{\displaystyle\sum_{k=1}^{n+1}Y_{k}} - \frac{1}{n+1} \right) \displaystyle\sum_{i=\sup(\lfloor n \alpha \rfloor +1, j)}^n l\left(\frac{i}{n+1}\right) \right] \\
&= \frac{n+1}{\displaystyle\sum_{k=1}^{n+1}Y_{k}} \frac{1}{n(n+1)} \left[ \sum_{j=1}^{\lfloor n \alpha \rfloor +1} \left((Y_{j} - \bar{Y}) \displaystyle\sum_{i= \lfloor n \alpha \rfloor +1}^{n} l\left(\frac{i}{n+1}\right) \right) + \displaystyle\sum_{j= \lfloor n \alpha \rfloor +2}^{n} \left( (Y_{j}- \bar{Y}) \displaystyle\sum_{i=j}^{n} l\left(\frac{i}{n+1}\right) \right) \right]. \\
\end{aligned}
\end{equation*}
\end{small}
where we have permuted the two sums. The law of large numbers gives that $\bar{Y}$ converges a.s to 1 when $n$ goes to infinity. Then, thanks to Slutsky's lemma, we only need to study
$$ \frac{1}{n(n+1)} \left[ \sum_{j=1}^{\lfloor n \alpha \rfloor +1} \left((Y_{j} - \bar{Y}) \displaystyle\sum_{i= \lfloor n \alpha \rfloor +1}^{n} l\left(\frac{i}{n+1}\right) \right) + \displaystyle\sum_{j= \lfloor n \alpha \rfloor +2}^{n} \left( (Y_{j}- \bar{Y}) \displaystyle\sum_{i=j}^{n} l\left(\frac{i}{n+1}\right) \right) \right].$$
We set $\forall j \leq n$, $G^{n}_{j}:= \displaystyle\sum_{i=j}^{n} l\left(\frac{i}{n+1}\right),$ $G^{n}_{n+1}:=0$, $H^{n}: = \displaystyle\sum_{j= \lfloor n \alpha \rfloor +2}^{n} G_{j}$. Then
\begin{small}
\begin{equation*}
\begin{aligned}
Q_{n} & = \frac{1}{n(n+1)} \left[ \displaystyle\sum_{j=1}^{\lfloor n \alpha \rfloor +1} \left( \frac{(n - \lfloor n \alpha \rfloor )G^{n}_{ \lfloor n \alpha \rfloor +1}- H^{n}}{n+1} \right) Y_{j}
+ \displaystyle\sum_{\lfloor n \alpha \rfloor +2}^{n+1} \left(G^{n}_{j} - \frac{H^{n}}{n+1} + G^{n}_{\lfloor n \alpha \rfloor +1} \frac{-1 - \lfloor n \alpha \rfloor}{n+1} \right) Y_{j} \right] \\
& =\frac{1}{n(n+1)} \Bigg[ \displaystyle\sum_{j=1}^{\lfloor n \alpha \rfloor+1} \left( \frac{(n - \lfloor n \alpha \rfloor)G^{n}_{ \lfloor n \alpha \rfloor +1}- H^{n}}{n+1} \right) (Y_{j}-1) \\
&+ \displaystyle\sum_{\lfloor n \alpha \rfloor +2}^{n+1} \left(G^{n}_{j} - \frac{H^{n}}{n+1} + G^{n}_{\lfloor n \alpha \rfloor +1} \frac{-1 - \lfloor n \alpha \rfloor}{n+1} \right) (Y_{j}-1) \Bigg], \\
\end{aligned}
\end{equation*}
\end{small}
because
\begin{small}
\begin{alignat*}{1}
& \frac{1}{n(n+1)} \left[ \displaystyle\sum_{j=1}^{\lfloor n \alpha \rfloor +1} \left( \frac{n - \lfloor n \alpha \rfloor)G^{n}_{ \lfloor n \alpha \rfloor +1}- H^{n}}{n+1} \right)+ \displaystyle\sum_{\lfloor n \alpha \rfloor +2}^{n+1} \left(G^{n}_{j} - \frac{H^{n}}{n+1} + G^{n}_{\lfloor n \alpha \rfloor +1} \frac{-1 - \lfloor n \alpha \rfloor}{n+1} \right) \right] \\
& =\frac{1}{n(n+1)} \Bigg[ \left( \frac{G^{n}_{\lfloor n \alpha \rfloor} (n- \lfloor n \alpha \rfloor) -H^{n} }{n+1}\right)(\lfloor n \alpha \rfloor +1) \\
&+ \left( \frac{G^{n}_{\lfloor n \alpha \rfloor} (-1 - \lfloor n \alpha \rfloor)}{n+1}(n - \lfloor n \alpha \rfloor ) - \frac{H^{n}}{n+1}(n- \lfloor n \alpha \rfloor) + H^{n}\right) \Bigg]\\
& = 0.
\end{alignat*}
\end{small}
Finally, we obtain
\begin{small}
\begin{equation*}
\begin{aligned}
Q_{n} & =
\frac{1}{n(n+1)} \Bigg[ \displaystyle\sum_{j=1}^{\lfloor n \alpha \rfloor +1} \left( \frac{(n - \lfloor n \alpha \rfloor)G^{n}_{ \lfloor n \alpha \rfloor +1}- H^{n}}{n+1} \right) (Y_{j}-1) + \displaystyle\sum_{\lfloor n \alpha \rfloor +2}^{n+1} \left(G^{n}_{j} - \frac{H^{n}}{n+1} + G^{n}_{\lfloor n \alpha \rfloor +1} \frac{-1 - \lfloor n \alpha \rfloor}{n+1} \right) (Y_{j}-1) \Bigg] \\
& = \frac{1}{n+1} \displaystyle\sum_{j=1}^{n+1} \alpha_{j, n} (Y_{j}-1),\\
\end{aligned}
\end{equation*}
\end{small}
where
$$\alpha_{j, n}= \left( \frac{(n - \lfloor n \alpha \rfloor)G^{n}_{ \lfloor n \alpha \rfloor +1}- H^{n}}{n(n+1)} \right), \, \, \forall j \leq \lfloor n \alpha \rfloor +1$$
and
$$\alpha_{j, n}= \left( \frac{G^{n}_{j}(n+1) - H^{n} - G^{n}_{\lfloor n \alpha \rfloor +1} (1 + \lfloor n \alpha \rfloor)}{n(n+1)} \right) , \, \, \forall j \geq \lfloor n \alpha \rfloor + 2.$$
Let us check the assumptions of the Proposition \ref{corlinde}. To begin with, let us show that $\sigma_{n}^{2}$ converges. We have
\begin{alignat*}{1}
\sigma_{n}^{2} & =\frac{1}{n+1} \displaystyle\sum_{j=1}^{n+1} \alpha_{j, n}^{2} \\
&= \frac{\lfloor n \alpha \rfloor +1}{n+1} \left[\frac{(n - \lfloor n \alpha \rfloor)G^{n}_{ \lfloor n \alpha \rfloor +1}- H^{n}}{n(n+1)}\right]^{2} \\
& + \frac{1}{n+1} \displaystyle\sum_{j=\lfloor n \alpha \rfloor +2}^{n+1}\left( \frac{G^{n}_{j}(n+1)}{n(n+1)} \right)^{2} +2\frac{1}{n+1} \displaystyle\sum_{j=\lfloor n \alpha \rfloor +2}^{n+1} \frac{G^{n}_{j}(n+1)( - H^{n} - G^{n}_{\lfloor n \alpha \rfloor +1} (1 + \lfloor n \alpha \rfloor))}{n^{2}(n+1)^{2}} \\
& + \frac{1}{n+1} \displaystyle\sum_{j=\lfloor n \alpha \rfloor +2}^{n+1}\left(\frac{- H^{n} - G^{n}_{\lfloor n \alpha \rfloor +1} (1 + \lfloor n \alpha \rfloor)}{n(n+1)} \right)^{2}. \\
\end{alignat*}
Let us work with the two terms which depend on $G^{n}_{j}$. The first term can be expanded as
\begin{small}
\begin{alignat*}{1}
\frac{1}{n+1} \displaystyle\sum_{j=\lfloor n \alpha \rfloor +2}^{n+1}\left( \frac{G^{n}_{j}(n+1)}{n(n+1)} \right)^{2} & = \frac{1}{n^{2}(n+1)} \displaystyle\sum_{j=\lfloor n \alpha \rfloor +2}^{n+1}\Big( G^{n}_{j} \Big)^{2} \\
& = \frac{1}{n^{2}(n+1)} \displaystyle\sum_{j=\lfloor n \alpha \rfloor +2}^{n+1} \left( \displaystyle\sum_{i=j}^{n} l\left(\frac{i}{n+1}\right)\right)^{2} \\
& = \frac{1}{n(n+1)} \displaystyle\sum_{i_{1}=\lfloor n \alpha \rfloor +2}^{n+1} \displaystyle\sum_{i_{2}=\lfloor n \alpha \rfloor +2}^{n+1} \left(\frac{-1 + (i_{1} \wedge i_{2}) - \lfloor n \alpha \rfloor}{n+1}\right)l\left(\frac{i_{1}}{n+1}\right)l\left(\frac{i_{2}}{n+1}\right). \\
\end{alignat*}
\end{small}
The second term may be rewritten as
\begin{small}
$$ 2\frac{1}{n+1} \displaystyle\sum_{j=\lfloor n \alpha \rfloor +2}^{n+1} \frac{G^{n}_{j}(n+1)( - H^{n} - G^{n}_{\lfloor n \alpha \rfloor +1} (1 + \lfloor n \alpha \rfloor))}{n^{2}(n+1)^{2}}
= -2\frac{(H^{n})^{2}}{n^{2}(n+1)^{2}} - 2 \frac{(1 + \lfloor n \alpha \rfloor )}{(n+1)^{2}n^{2}} H^{n} G^{n}_{ \lfloor n \alpha \rfloor +1}.$$
\end{small}
Finally,
\begin{alignat*}{1}
\sigma_{n}^{2}& = \frac{\lfloor n \alpha \rfloor +1}{n+1} \left[\frac{(n - \lfloor n \alpha \rfloor)G^{n}_{ \lfloor n \alpha \rfloor +1}- H^{n}}{n(n+1)}\right]^{2} \\
& + \frac{1}{n(n+1)} \displaystyle\sum_{i_{1}=\lfloor n \alpha \rfloor +2}^{n+1} \displaystyle\sum_{i_{2}=\lfloor n \alpha \rfloor +2}^{n+1} \left(\frac{-1}{n+1}
+ \frac{\min(i_{1}, i_{2})}{n+1} - \frac{\lfloor n \alpha \rfloor}{n+1} \right) l(\frac{i_{1}}{n+1})l(\frac{i_{2}}{n+1}) \\
& -2\frac{(H^{n})^{2}}{n^{2}(n+1)^{2}} - 2 \frac{(1 + \lfloor n \alpha \rfloor )}{(n+1)^{3}n^{2}} H^{n} G^{n}_{ \lfloor n \alpha \rfloor +1} + \frac{n - \lfloor n \alpha \rfloor -1 }{n+1} \left(\frac{ H^{n} + G^{n}_{\lfloor n \alpha \rfloor +1} (1 + \lfloor n \alpha \rfloor)}{n(n+1)} \right)^{2}. \\
\end{alignat*}
Let us first notice that, if we denote $$K^{n}= \displaystyle\sum_{i_{1}=\lfloor n \alpha \rfloor +2}^{n+1} \displaystyle\sum_{i_{2}=\lfloor n \alpha \rfloor +2}^{n+1} \frac{\min(i_{1}, i_{2})}{n+1} l\left(\frac{i_{1}}{n+1}\right)l\left(\frac{i_{2}}{n+1}\right) $$ and $$T^{n} = \displaystyle\sum_{i =\lfloor n \alpha \rfloor }^{n} \frac{i}{n} l\left( \frac{i}{n+1}\right)$$ then
$$ H^{n} = nT^{n} - (\lfloor n \alpha \rfloor +1) G^{n}_{ \lfloor n \alpha \rfloor +1}.$$
So that
\begin{equation*}
\begin{aligned}
\sigma_{n}^{2} & \underset{n\to+\infty}{\sim} \alpha\frac{(G^{n}_{ \lfloor n \alpha \rfloor +1}-T^{n})^{2}}{n^{2}} + \frac{K^{n} - \alpha (G^{n}_{ \lfloor n \alpha \rfloor +1})^{2}}{n^{2}} - \frac{-2(T^{n} - \alpha G^{n}_{ \lfloor n \alpha \rfloor +1})^{2}}{n^{2}}\\
& -2 \frac{\alpha(G^{n_{ \lfloor n \alpha \rfloor +1}}T^{n}- \alpha (G^{n}_{ \lfloor n \alpha \rfloor +1})^{2})}{n^{2}} + \frac{(1- \alpha)(T^{n})^{2}}{n^{2}} \\
& \underset{n\to+\infty}{\sim} \frac{K^{n}-( T^{n})^{2}}{n^{2}}. \\
\end{aligned}
\end{equation*}
Let us show that this last quantity converges to $\sigma^2 = \int_{\alpha}^1 \int_{\alpha}^1 \frac{\min(x, y) -xy}{f(F^{-1}(x)) f(F^{-1}(y))} < \infty$. Indeed it is a generalized Rieman sum. First, we show that the function
$$g : (x, y) \mapsto \frac{\min(x, y) -xy}{f(F^{-1}(x)) f(F^{-1}(y))}$$ is integrable on $]\alpha, 1[ \times ]\alpha, 1[$. Indeed, around 1,
$$g(x, y) = O \left(\frac{\min(x, y) -xy}{(1-x)^{\frac{3}{2}- \epsilon_l}(1-y)^{\frac{3}{2}- \epsilon_l}} \right)$$
which is integrable on this domain because for $\beta$ close to 1
$$\int_{\alpha}^{\beta} \int_{\alpha}^{\beta} \frac{\min(x, y) -xy}{(1-x)^{\frac{3}{2}- \epsilon_l}(1-y)^{\frac{3}{2}- \epsilon_l}} dxdy \sim C(\alpha) \beta(1- \beta)^{2 \epsilon_l}.$$
and $\epsilon_l>0$ (here again we see that we really need this $\epsilon_L>0$).
As we have already seen, the results on Riemann's sum in dimension 2, give by the continuity of the function $(x, y) \mapsto \frac{\min(x, y) -xy}{f(F^{-1}(x)) f(F^{-1}(y))}$ that for all $\alpha <\beta <1$ :
$$
\sigma_{n, \beta}^2 := \frac{1}{n^2}\sum_{i_1=\lfloor n \alpha \rfloor}^{\lfloor n \beta \rfloor} \sum_{i_1=\lfloor n \alpha \rfloor}^{\lfloor n \beta \rfloor} \frac{\frac{\min(i_1, i_2)}{n} - \frac{i_1i_2}{n^2} }{f\left(F^{-1}\left(\frac{i_1}{n+1}\right)\right) f\left(F^{-1}\left(\frac{i_2}{n+1}\right)\right)}
\longrightarrow \int_{\alpha}^{\beta} \int_{\alpha}^{\beta} \frac{\min(x, y) -xy}{f(F^{-1}(x)) f(F^{-1}(y))} dxdy.
$$
We have to study the remaining part of the sum to conclude. Let us fix $\beta$ close to 1 and deal with
$$r_{n, \beta}^2 := \frac{1}{n^2}\sum_{i_1=\lfloor n \beta \rfloor}^{n} \sum_{i_1=\lfloor n \beta \rfloor}^{n} \frac{\frac{\min(i_1, i_2)}{n} - \frac{i_1i_2}{n^2} }{f \left(F^{-1}\left(\frac{i_1}{n+1}\right)\right) f\left(F^{-1}\left(\frac{i_2}{n+1}\right)\right)}.$$
First of all, let us notice that
$$r_{n, \beta}^2= \int_{\beta}^1 \int_{\beta}^1 g\left(\frac{\lfloor n x \rfloor}{n},\frac{\lfloor n y \rfloor}{n} \right) dx dy. $$
We want to show that $$\lim\limits_{n \to + \infty} r_{n, \beta}^2= \int_{\beta}^1\int_{\beta}^1 g(x, y) dx dy.$$
Let us first show that, using Lebesgue theorem, we can permute the limit in $n$ and the double integrale, in this way
$$\lim\limits_{n \to + \infty} r_{n, \beta}^2 = \int_{\beta}^1\int_{\beta}^1 \lim\limits_{n \to + \infty}g\left(\frac{\lfloor n x \rfloor}{n},\frac{\lfloor n y \rfloor}{n} \right).$$
\begin{itemize}
\item[1)] Let $(x, y)$ be fixed in $[\beta, 1[ \times [\beta, 1[$ and $n$. Then $$g\left(\frac{\lfloor n x \rfloor}{n},\frac{\lfloor n y \rfloor}{n} \right) \longrightarrow g(x, y)$$ by continuity. And $g$ is integrable on $[\beta, 1[ \times [\beta, 1[$ as we saw before.
\item[2)] Let $(x, y)$ be fixed in $[\beta, 1[ \times [\beta, 1[$ and $n$. Let us denote $x_n = \frac{\lfloor n x \rfloor}{n}$ and $y_n=\frac{\lfloor n y \rfloor}{n}$. By hypothesis
$$g\left(\frac{\lfloor n x \rfloor}{n},\frac{\lfloor n y \rfloor}{n} \right) \leq C \frac{\min(x_n, y_n) -x_ny_n}{(1-x_n)^{\frac{3}{2}- \epsilon_l}(1-y_n)^{\frac{3}{2}- \epsilon_l}}.$$
By separating the two cases and using monotony we obtain that
$$g\left(\frac{\lfloor n x \rfloor}{n},\frac{\lfloor n y \rfloor}{n} \right) \leq C h(x, y)$$
where $$h:(x, y) \mapsto \frac{\min(x, y)}{\left(1- \min(x, y)\right)^{\frac{3}{2}- \epsilon_l}\left(1- \max(x, y)^{\frac{3}{2}- \epsilon_l-1}\right)}$$
is integrable on $[\beta, 1[ \times [\beta, 1[$.
\end{itemize}
Then, the Lebesgue theorem allows us to permute integration and limit so that, we have shown that $\sigma_n^2 \longrightarrow \sigma^2$and the first assumption of Proposition \ref{corlinde} holds.
\vspace{0.3cm}
Let us now deal with the second assumption of Proposotion \ref{corlinde} about the maximum of the $\alpha_{i, n}$.
For $j \leq \lfloor n \alpha \rfloor +1$, we have
\begin{alignat*}{1}
\alpha_{j, n} = & \frac{(n - \lfloor n \alpha \rfloor +1)G^{n}_{ \lfloor n \alpha \rfloor +1} - H^{n}}{n(n+1)}. \\
\end{alignat*}
Using the previous computations, for $n$ large enough we have
$$\frac{ (\alpha_{j, n})^{2} }{n \sigma^{2}_{n}} \underset{n \to + \infty}{\sim} \frac{(G^{n}_{ \lfloor n \alpha \rfloor +1} - \frac{T^{n}}{n})^{2}}{K_{n} - \frac{T_{n}^{2}}{n^{2}}} \frac{1}{n}, $$
But the convergence $$ \frac{\left(K^{n} - \left(\frac{T^{n}}{n}\right)\right)^{2}}{n^{4} } \underset{n \to + \infty}{\longrightarrow} \int_{\alpha}^{1} \int_{\alpha}^{1} (\min(x, y) - xy) l(x) l(y) dxdy $$ implies the convergence $$\frac{(G^{n}_{ \lfloor n \alpha \rfloor +1} - \frac{T^{n}}{n})^{2}}{n^{4}} \underset{n \to + \infty}{\longrightarrow} \int_{\alpha}^{1} (1-x) l(x) dx.$$ Indeed
$$\int_{\alpha}^{1} \int_{\alpha}^{1} (\min(x, y) -xy)l(x) l(y) dx dy = \int_{\alpha}^{1} \int_{\alpha}^{1} (y(1-x)) l(x) l(y) dx dy + \int_{\alpha}^{1} x l(x) \int_{x}^{1} (1-y)l(y) dy dx.$$
So that,
$$\frac{ (\alpha_{j, n})^{2} }{n \sigma^{2}_{n}} \underset{n \to + \infty}{\sim} \frac{C}{n} \underset{n \to + \infty}{\longrightarrow} 0$$
when $n$ goes to infinity. If $j \geq \lfloor n \alpha \rfloor +2 $ the same property holds as
$$\alpha_{j, n}= \frac{(n+1)G_{j}^{n}-H^{n} - G^{n}_{\lfloor n \alpha \rfloor}( \lfloor n \alpha \rfloor +1) }{n(n+1)} \underset{n\to+\infty}{\sim} \frac{(n+1) G_{j^{n}} - T^{n}}{n^{2}}. $$
Hence, we may apply Proposition \ref{corlinde} and conclude that
$$\sqrt{n}Q_{n} \Longrightarrow \mathcal{N}(O, \sigma^{2}).$$
where $\sigma^{2} = \int_{\alpha}^{1} \int_{\alpha}^{1} (\min(x, y) -xy) l(x) l(y) dxdy$. Finally, just multiply by $(1-\alpha)^{-1}$ to get the final result.
\vspace{0.3cm}
\textbf{Step 4 : Conclusion}
The Slutsky lemma allows to conclude using the results of steps 1 and 3.
\end{proof}
\section{Conclusion}
The superquantile was introduced because the usual quantile was not subadditive and does not give enough information on what was happening in the tail-distribution. This quantity is interesting because it satisfies the axioms of a coherent measure of risk. In this paper, we have introduced a new coherent measure of risk with the help of the Bregman divergence associated to a strictly convex function $\gamma$. They are rich tools because of the diversity of the functions $\gamma$ that can be chosen according to the problem we study. We shown throught different examples that a judicious choice for the function $\gamma$ can make the Bregman superquantile a more interesting measure of risk than the classical superquantile (Bregman superquantile can be finite even in infinite mean model, it is continous in more cases, more \textit{robust} etc...). Moreover, we have introduced a Monte Carlo estimator of the Bregman superquantile which is statistically powerful thanks to the strictly convex (and so settling) function $\gamma$.
The theoretical properties obtained in this paper are confirmed on several numerical test cases.
More precisely, geometrical and harmonic superquantiles are more robust than the classical superquantile.
This robustness is particularly important in in finance and risk assessment studies.
For instance, in risk assessment, when dealing with real data, geometrical and harmonic statistics have been proved to be more relevant than classical statistics.
For example \cite{chahan04} prove the usefulness of the geometrical mean and variance for the analysis of air quality measurements.
As an illustration, we have applied the geometrical and harmonic superquantiles on real data coming from a radiological impact code used in the nuclear industry.
Further studies will try to apply these criteria in probabilistic assessment of physical components reliability using numerical simulation codes \cite{derdev08}.
However, Monte Carlo estimators are no longer applicable in this context and efficient estimators have to be developed.
Ideas involving response surface technique should be developed (see for example \cite{cangar08} for quantile estimation and \cite{becgin12} for rare event probability estimation).
Finally those Bregman superquantiles are interesting because they can be linked with several previous works in economy. There is for example a strong link between capital allocations and Bregman superquantile. It would be interesting to apply our results on this economic theory in a future work.
\newpage
\begin{center}
\textbf{ACKNOWLEDGEMENT}
\end{center}
We warmly thank two anonymous referees for their careful reading that helped us to improve the paper.
\bibliographystyle{plain}
|
\section{Introduction}
Cosmology without the mechanism of inflation (see e.g.\ \cite{Mukhanov:book})
seems inconceivable. Not only does inflation solve many conceptual
problems of the old hot big-bang theory, it is also in excellent
agreement with experimental data of ever increasing precision
\cite{Komatsu:2010fb, Hinshaw:2012aka, Ade:2013zuv,Ade:2013uln}. In
fact, it is hard to devise a mechanism different from inflation that
could solve all cosmological obstacles and, at the same time, does not
contradict any observation.
Despite all its success, however, the nature and the origin of
inflation remain unexplained so far. Most inflationary models are
based on a scalar field $\varphi$ ---the inflaton. Its
identification with a scalar field in particle-physics models is still
debated. For instance, in the model of non-minimal Higgs inflation,
$\varphi$ has been identified with the observed Standard Model Higgs
boson
\cite{Bezrukov:2007ep,Barvinsky:2008ia,Bezrukov:2008ej,DeSimone:2008ei,Barvinsky:2009fy,Bezrukov:2009db,Bezrukov:2010jz,Barvinsky:2009ii,
Allison:2013uaa}.\footnote{See also variants of Higgs inflation such
as the `new Higgs inflation' model, where a minimally coupled
scalar field with a non-canonical kinetic term coupled to the
Einstein tensor was considered
\cite{Germani:2010gm,Germani:2010ux,Germani:2014hqa}.} However, even
in scenarios where $\varphi$ is embedded in a realistic theory, the
question of why inflation has started in the first place remains
mostly unresolved.
Inflation presupposes an already pre-existing classical background on
which tiny primordial quantum fluctuations can propagate
\cite{Starobinsky1979,Mukhanov:1981xt,Hawking:1982cz,Guth:1982ec,Starobinsky:1982ee,Bardeen:1983qw}.
During inflation, these quantum fluctuations experience an effective
quantum-to-classical transition; see
\cite{PS96,Kiefer:2006je,Kiefer:2008ku} and the references therein.
At the most fundamental level, a classical background does not exist. The reason is that,
given the fundamental quantum character of all matter interactions, it
is expected that gravity (and therefore spacetime) has to be quantized as well.
There are many different approaches to quantum gravity
\cite{Ori09,Kiefer:2012boa,Kiefer:2013jqa}.
In the traditional canonical approach, when restricting to cosmology,
we obtain the Wheeler--DeWitt equation that governs the quantum dynamics of the universe.
This is a differential equation which has to be accompanied by a proper boundary condition.
In the absence of a fully developed theory of quantum gravity,
cosmological considerations offer at least a heuristic guideline for
natural choices of such boundary conditions.
The hope is that this will ultimately lead to observable quantum
cosmological effects
\cite{Calcagni:2012vw,KK12a,KK12b,Bini:2013fea,Sasha14}. There also exists a
path-integral analogue to the canonical approach, which will be used here.
Two of the most influential proposals for boundary conditions
are the so called
no-boundary \cite{Hawking:1982,Hartle:1983ai} and tunneling
\cite{Vilenkin:1982de,Vilenkin:1983xq,Linde:1983mx, Vilenkin:1984wp,
Rubakov:1984ki, Zeldovich:1984vk} conditions; see, for example,
\cite{Kiefer:2012boa} for a review.
The underlying picture behind the tunneling condition is that our
universe as a whole was created by a `quantum tunneling from
nothing'. (As we shall briefly discuss below, however, this picture can
at best be seen as a metaphor.)
The tunneling condition seems to be preferred over the no-boundary
condition, in the sense that it can lead to a successful
post-nucleating phase of inflation; see \cite{Barvinsky:2009jd} and
references therein.
It is, however, not self-evident that an inflationary model and the
tunneling process can always be combined into one consistent
scenario. Typically, the tunneling proposal is believed to give rise
to a sustainable inflationary phase because it predicts a conditional
probability of the field values peaked at large values of the potential. This accords with
chaotic inflation where, in its simplest incarnation, the potential is
a monomial of $\varphi$.\footnote{On the other hand, the no-boundary
proposal can accommodate inflation after introducing a re-weighting
of the probability \cite{HHH1,HHH2,HHH3,Hartle:2010dq}.}
The purpose of this paper is to derive a compatibility condition testing whether the origin of inflationary models favored by the 2013
\textsc{Planck} release \cite{Ade:2013uln, Martin:2013gra,
Tsujikawa:2013ila} can consistently be explained by a quantum
tunneling of the universe. The requirement of consistency then might
lead to restrictions for the parameters of the underlying inflationary
models and therefore to testable quantum cosmological predictions. For completeness, a separate confrontation with \textsc{Bicep2} results \cite{Ade:2014xna} is also carried out, regardless of the ongoing debate on their ultimate validity \cite{MoSe}.
The paper is structured as follows. In section \ref{II}, we introduce
the formalism of the Euclidean instanton and construct the quantum
cosmological tunneling distribution.
In section \ref{III}, we consider the model of natural inflation and
derive a consistency condition for the tunneling scenario.
We conclude in section \ref{IV} by summarizing our results and comment
on a similar analysis for different models of inflation as well as on
a more ambitious quantum analysis.
\section{Quantum origin of the cosmos}\label{II}
\subsection{Effective action and de Sitter instanton}
The quantum tunneling can be described in terms of instantons ---solutions to the Euclidean equations of motion.
The effective action $\Gamma$ is defined as
\begin{align}
e^{-\Gamma}:=\int[{\cal D}g]\,e^{-S_{\rm{eff}}[g]},\,
\end{align}
with the matter effective action $S_{\rm{eff}}$ \emph{defined} by the `quantum average' over matter fields $\Phi(x)$
\begin{align}
e^{-S_{\rm{eff}}[g]}:=\int\,[{\cal D}\Phi]\,e^{-S[g,\,\Phi]}\,.\label{EffAct1}
\end{align}
In practice, it is usually impossible to calculate the full effective action $\Gamma$ exactly
and one has to resort to a loop expansion. Here, we have split the functional integrals into two parts,
distinguishing between the geometrical and the matter part. In what follows, we consider the theory described
by the action $S$ as a quantum field theory in curved spacetime and neglect graviton loops, that is,
we only consider (\ref{EffAct1}) in a classical background described by the metric $g_{\mu\nu}$.
Following \cite{Barvinsky:2009jd}, we consider the matter effective action
\begin{align}
S_{\rm{eff}}[g]=\int\text{d}^4x\,\sqrt{g}\,\frac{M_{\rm{P}}^2}{2}\left[2\,\Lambda_{\rm{eff}}-R(g)+\dots\right]\,.\label{EHAction1}
\end{align}
Here, $\Lambda_{\rm{eff}}$ is the effective cosmological constant, $M_{\rm{P}}=m_{\rm{P}}/\sqrt{8\,\pi}\approx2.43\times 10^{18}$ GeV is the reduced Planck mass (in units where $\hbar=c=1$) and $R(g)$ is the Ricci scalar constructed from the Euclidean metric field $g_{\mu\nu}(x)$.
The ellipsis stands for higher-order curvature and gradient terms that we do not take into account.
In the cosmological context of slow-roll inflation driven by a real scalar field $\varphi$, the vacuum energy density during inflation is dominated by the nearly constant potential $V(\varphi)$.
This leads to the identification $M_{\rm{P}}^2\,\Lambda_{\rm{eff}}\approx V_{\rm{eff}}(\varphi)$ and justifies the omission of gradient terms in (\ref{EHAction1}).
Once we have calculated the effective action (\ref{EHAction1}),
we can specialize to a fixed closed Friedmann--Lema\^itre--Robertson--Walker (FLRW) background with line element
\begin{align}
\text{d}s^2=N^2(\tau)\,\text{d}\tau^2+a^2(\tau)\,\text{d}^2\Omega^{(3)}.\label{LineElFRW}
\end{align}
Here, $a(\tau)$ and $N(\tau)$ are the Euclidean scale factor and lapse function, while $\text{d}^2\Omega^{(3)}$ is the volume element of the three-dimensional sphere.
In the background metric (\ref{LineElFRW}), the effective action (\ref{EHAction1}) reduces to
\begin{align}
S_{\rm{eff}}[a,\,N]=12\,\pi^2\,M_{\rm{P}}^2\,\int\text{d}\tau\,N\,\left[-\frac{1}{N^2}\left(\frac{\text{d}a}{\text{d}\tau}\right)^2a-a+H_{\rm{eff}}^2\,a^3\right]\,,\label{EffActDS}
\end{align}
where we have identified the effective cosmological constant $\Lambda_{\rm{eff}}\equiv3\,H_{\rm{eff}}^2$
with the effective Hubble parameter $H_{\rm{eff}}$.
The instanton is a solution of the Euclidean Friedmann equations, which are obtained by varying (\ref{EffActDS}) with respect to $N$,
\begin{align}
\frac{1}{N^2} \left(\frac{\text{d}a}{\text{d}\tau}\right)^2=1-H_{\rm{eff}}^2\,a^2\,.\label{EEOM1}
\end{align}
This equation has one turning point $a_{+}:=a(\tau_{+}):=1/H_{\rm{eff}}$ so that the real solution interpolates
between $a_{-}:=a(\tau_{-}):=a(0)=0$ and $a_{+}$. Depending on the sign of $N$,
the gauge choice of the Lagrange multiplier $N$ describes two disjoint equivalence classes of instantons.
It is sufficient to consider the representative values $N_{\pm}:=\pm 1$.
The explicit solution to the differential equation (\ref{EEOM1})
then reads
\begin{align}
a(\tau)=\frac{1}{H_{\rm{eff}}}\text{sin}\left(H_{\rm{eff}}\,\tau\right)\,,\label{ESola1}
\end{align}
where we have fixed the integration constant by the condition
\begin{align}
\frac{\text{d}\,a}{\text{d}\tau}\Big|_{a=0}=1\,,
\end{align}
which is provided by the constraint equation (\ref{EEOM1}). We have also chosen
the geometrical meaningful positive root of (\ref{EEOM1}) to obtain (\ref{ESola1}).
The turning point $a_{+}$, corresponding to the equator of the Euclidean half sphere where $a(\tau)$ is maximized, determines the `moment of nucleation' $\tau_{+}:=\pi/(2\,H_{\rm{eff}})$.
The tunneling process is then described by attaching the Euclidean half sphere to the inflationary Lorentzian regime at $\tau_{+}$.
At the boundary, we analytically continue (\ref{ESola1}) to Lorentzian signature $\tau\to \rmi\,t$:
\begin{align}
a_{\rm{L}}(t)=a\left(\frac{\pi}{2\,H_{\rm{eff}}}+\rmi t\right)
=\frac{1}{H_{\rm{eff}}}\,\text{cosh}\left(H_{\rm{eff}}\,t\right)\,.
\end{align}
The instanton is obtained by inserting (\ref{ESola1}) into (\ref{EffActDS}) and integrating from $\tau_{-}$ to $\tau=\tau_{+}$:
\begin{align}
S_{\rm{eff}}^{\text{on-shell}}[a,N_{\pm}]=\mp 8\pi^2\,\frac{M_{\rm{P}}^2}{H_{\rm{eff}}^2}\,.\label{EffFRw}
\end{align}
Neglecting graviton loops, we obtain the tunneling instanton $\Gamma^{\text{on-shell}}$ for the choice $N_-=-1$ \cite{Barvinsky:2009jd},
\begin{align}
\Gamma^{\text{on-shell}}(\varphi):= S_{\rm{eff}}[a,-1]=24\,\pi^2\,\frac{M_{\rm{P}}^4}{V_{\rm{eff}}(\varphi)}\,,\label{EInst}
\end{align}
where we have again used the identification
$H_{\rm{eff}}^2\equiv\Lambda_{\rm{eff}}/3\equiv V_{\rm{eff}}/(3\,M_{\rm{P}}^2)$ in the last step (note that
$V_{\rm{eff}}>0$).
It should be mentioned that the analogy with the quantum mechanical
tunneling is of at most heuristic value.
In fact, the presented derivation of the instanton simply corresponds
to a solution of the Euclidean equations of motion subject to some
specially chosen boundary condition. Here, `tunneling' is then
simply \emph{defined} by this choice. In the ordinary quantum
mechanical tunneling problem,
say the example of the spontaneous decay of an $\alpha$-particle,
tunneling is described by a wave function that contains only outgoing
modes.
But in this case, there is always a fixed reference phase
$\propto\text{exp}(-{\rm i}\,\omega\,t)$ with respect to which one can define
outgoing and incoming modes unambiguously \cite{Zeh88,Zeh07}.
The sign in front of the frequency $\omega$ and the external time
parameter $t$ in the exponential is fixed by the sign of the time
derivative in the Schr\"odinger equation.
If the wave function corresponds to a plane wave
$\propto\text{exp}(-{\rm i}\,\omega t\mp\,{\rm i}\,k\,x)$, a relative
minus sign with respect to the sign of $t$
corresponds to outgoing modes $k$. In contrast, in the context of the
quantum tunneling of the universe as a whole, there is no such simple
notion of ingoing and outgoing modes, as there is no notion of an
external time parameter anymore at the fundamental level \cite{Kiefer:2012boa}.
In the absence of any reference phase, the definition of incoming and
outgoing becomes meaningless. The only notion of time one can
introduce at the fundamental level is that of `internal time'. In this
case, the role of time can be played by one or more configuration-space degrees of
freedom. In the context of cosmological minisuperspace, the scale
factor has a preferred role as internal time parameter, in the sense
that its associated kinetic term comes with a relative minus sign
compared to the matter degrees of freedom. This is a consequence of
the indefinite nature of the minisuperspace DeWitt metric. Time as an
external parameter can only be recovered at a semi-classical level
\cite{Kiefer:2012boa,Kiefer:2013jqa}. All
this does not invalidate the construction presented here but simply
serves to clarify our notion of `tunneling' and emphasizes the difference with respect to the ordinary quantum mechanical tunneling
problem.
\subsection{Tunneling distribution function and initial conditions for inflation}\label{sec2.2}
The interpretation of the wave function of the universe is largely an open problem \cite{Kiefer:2012boa}.
One heuristic approach is to interpret peaks in the (absolute square of the) wave function as a prediction; see \cite{Hartle:1987}.
Recently, this idea was applied to the model of non-minimal Higgs inflation \cite{Barvinsky:2009jd, Barvinsky:2012zz} and we will follow a similar idea in the present paper with the purpose of presenting a general construction that can serve as a tool to derive predictions from quantum cosmology.
In the semi-classical approximation to quantum cosmology, the no-boundary proposal does usually not give a wave function which is peaked at a field value high enough for inflation \cite{Barvinsky:1990ga}. Such a peak may arise in models of eternal inflation using the landscape picture \cite{Hartle:2010dq} but we will not discuss them here.
For this reason, we will only address the tunneling proposal. From it, using (\ref{EInst}),
the probability distribution in the semi-classical limit is found to be
\begin{align}
{\cal T}(\varphi):=e^{-\Gamma^{\text{on-shell}}(\varphi)}=\text{exp}\left[-\frac{24\,\pi^2\,M_{\rm{P}}^4}{V_{\rm{eff}}(\varphi)}\right]\,.\label{ProbDistr1}
\end{align}
A peak corresponds to a maximum of (\ref{ProbDistr1}).
Finding this peak is equivalent to finding the maxima of the potential $V_{\rm{\max}}:=V_{\rm{eff}}(\varphi_{\rm{\max}})$.
This leads to the simple conditions
\begin{align}
\frac{\text{d}\,V_{\rm{eff}}(\varphi)}{\text{d}\varphi}\Big|_{\varphi=\varphi_{\rm{max}}}=0,\qquad \frac{\text{d}^2\,V_{\rm{eff}}(\varphi)}{\text{d}\varphi^2}\Big|_{\varphi=\varphi_{\rm{max}}}<0\,.\label{MaxCond}
\end{align}
The peak $\varphi_{\rm{max}}$ in (\ref{ProbDistr1}) corresponds to the value of $\varphi$ that selects the most probable value of $\Lambda_{\rm{eff}}=V_{\rm{eff}}(\varphi_{\rm{max}})/M_{\rm{P}}^2$ for which the universe starts after tunneling.
In this way, the quantum scale of inflation was obtained in \cite{Barvinsky:1995gi,Barvinsky:1996ce, Barvinsky:1998qh, Barvinsky:1998rn,Barvinsky:1999qn}.
A high value of $V_{\rm{max}}$ is necessary to start an inflationary evolution after tunneling.
Therefore, $\varphi_{\rm{max}}$ can be interpreted as setting the initial conditions for inflation.
In the inflationary slow-roll regime, $\varphi\approx\text{const}$ and the energy density is completely dominated by $V_{\rm{max}}$.
The peak value $\varphi_{\rm{max}}$ allows one to determine the energy scale of inflation by
\begin{align}
E_{\rm{infl}}^{\rm{QC}}:=V_{\rm{max}}^{1/4}\,\label{EInfQC}.
\end{align}
One should bear in mind that
$E_{\rm{infl}}^{\rm{QC}}\neq\varphi_{\rm{max}}$ in general, as
$\varphi$ is only a coordinate in the field configuration space and as
such does not have any direct physical meaning (although both
$E_{\rm{infl}}^{\rm{QC}}$ and $\varphi_{\rm{max}}$ have the same
physical dimension of an energy).
Only the (effective) potential itself can serve as a meaningful observable.
Classical inflationary models predict an energy scale
\begin{align}
E_{\rm{infl}}^{\rm{model}}:={}& V_{*}^{1/4}\,,\label{EInfModel}
\end{align}
where $V_*:=V(\varphi_*)$ and $\varphi_{*}$ denotes the field value evaluated at the moment $k_{*}=H_{*}\,a_{*}$
when the pivot mode $k_*$ (to be chosen according to the observational window of the experiment) first crosses
the Hubble scale.
Inflationary models allowing for a quantum cosmological origin in the sense discussed here must therefore satisfy the approximate consistency condition
\begin{align}
E_{\rm{infl}}^{\rm{QC}}\approx E_{\rm{infl}}^{\rm{model}}\,,\label{ConsistencyCond1}
\end{align}
that is, the energy scale of the inflationary model must be of the
same order as the prediction from quantum cosmology. In principle,
this is an exact relation and one could derive a very precise
prediction of quantum cosmology. However, since in most situations
only a truncated loop expansion of $\Gamma$ (and therefore of ${\cal
T}$) is available, one cannot expect this condition to
be satisfied exactly at the perturbative level. It is well known that
radiative corrections can change the shape of the effective potential
and, in particular, its extrema which determine the peak position.
If the amplitude of the tensor power spectrum at the Hubble-scale crossing, $A_{\rm{t}\,*}$,
is known, one can introduce a third scale, the energy scale of inflation inferred
from this amplitude.
Using, e.g., the relations (152) and (216) from \cite{BaumannTASI}, one
gets the following expression for the inferred ``observed'' energy scale of inflation:
\begin{align}
E_{\rm{infl}}^{\rm{obs}}=M_{\rm{P}}\,\left(\frac{3}{2}\,\pi^2\,A_{\rm{t}\,*}\right)^{1/4}
\approx M_{\rm{P}}\,\left(\frac{3}{2}\,\pi^2\,A_{\rm{s}\,*}\,r_{*}\right)^{1/4}.
\end{align}
In the first step, we have used the well-known expression for tensor modes $A_{\rm{t}}\propto H^2\propto E_{\rm{infl}}^4$ at horizon crossing.
To first order in the slow-roll approximation, the scale $E_{\rm{infl}}^{\rm{obs}}$ can thus be expressed in terms of the tensor-to-scalar ratio $r:=
A_{\rm{t}}/A_{\rm{s}}$ and the amplitude of the scalar perturbations
$A_{\rm{s}}$, which is fixed by the measured temperature anisotropies
of the cosmic microwave background. For the pivot scale $k_{*}=
0.002\,\, \text{Mpc}^{-1}$, the best fit of the \textsc{Planck}+WP
data by the $\Lambda$CDM model yields the following $1\sigma$
experimental bound on $A_{\rm{s}\,*}$ \cite{Ade:2013uln}:
\begin{align}
\text{ln}\left(10^{10}\,A_{\rm{s}\,*}\right)=3.089^{+0.024}_{-0.027}\,.\label{AsCMB}
\end{align}
Until recently, observations gave only an upper bound on $r$. If, however,
the recent announcement by the \textsc{Bicep2} experiment
of the discovery of primordial gravitational waves \cite{Ade:2014xna}
is confirmed, this will yield a model-independent determination of
the energy scale of inflation.
Assuming that this is the case, one
has $r=0.20^{+0.07}_{-0.05}$ for the primordial graviton contribution,
or $r=0.16^{+0.06}_{-0.05}$ if currently best available dust models
are taken into account. Taking the central values
$A_{\rm{s}\,*}=2.2\times10^{-9}$ and $r=0.16$, this leads to an energy
scale
\begin{align}
E_{\rm{infl}}^{\rm{obs}}\approx 2.06\times 10^{16} \,\text{GeV}\,,
\end{align}
which can be roughly taken as upper bound for the inflationary energy if the \textsc{Bicep2} constraint is ignored.
We can thus make contact with experiments via the extended consistency condition
\begin{align}
E_{\rm{infl}}^{\rm{QC}}\approx E_{\rm{infl}}^{\rm{model}}\approx E_{\rm{infl}}^{\rm{obs}}\,.\label{ConsistencyCond2}
\end{align}
Since probabilistic arguments in the context of cosmology involve
difficult conceptual questions, we shortly summarize the underlying
assumptions allowing for a consistent application and interpretation
of the tunneling scenario presented here.
\begin{enumerate}
\item \emph{A classical background must have emerged}. In the underlying
full quantum theory, there is not yet any notion of `background'
or `classical', but only that of a pure quantum state,
corresponding to the wave function of the universe.
To understand the semi-classical limit, two steps must be performed
\cite{Kiefer:2012boa}. First, one must employ a Born--Oppenheimer type
of approximation scheme to find wave functions with a semi-classical
behavior. Second, one must invoke the process of decoherence
\cite{deco} to understand the degree of classical behavior.
The semi-classical wave function can be understood as
one branch of the full wave function in the Everett
interpretation of quantum mechanics. It is this semi-classical
branch of the full wave function which is used to construct the
probability distribution (\ref{ProbDistr1}). The emergence of a
quasi-classical background via decoherence is induced by the
division of the configuration space into
system and environment; in concrete models, the system consists of
global degrees of freedom such as the scale factor and the inflaton,
and the environment consists of small density fluctuations and small
gravitational waves \cite{Zeh86,CK87,BKKM97}.
The inevitable interaction with the
environment then leads to the entanglement of the system with the
environment. The process of decoherence describes the effective
influence of the environment on the system by integrating out the
overwhelmingly many inaccessible environmental degrees of freedom and
leads to a suppression of quantum correlations in the reduced density
matrix for the system. The exponential suppression of the non-diagonal
elements of the reduced density matrix then corresponds to an
effective classicalization of the system. Once the classical behavior
of the `background' is understood, one can address the
quantum-to-classical transition of inhomogeneous degrees of freedom
\cite{Kiefer:2006je,Kiefer:2008ku}.
\item \emph{The universe `nucleates' into a homogeneous and isotropic
universe}. The concordance model of cosmology is based on the
cosmological principle which implies homogeneity and isotropy around
any point in space, when averaged over scales larger than around
$100$ Mpc. This assumption is supported {\it a posteriori} by empirical
evidence from observations of the large-scale structure within the
observable patch of our universe.
\item \emph{Right after the tunneling process, the universe starts a
phase of accelerated expansion} (inflation).
This assumption is also supported by observational evidence and {\it a
posteriori} justifies the use of the de Sitter instanton.
\item \emph{The probability distribution should possess a sharp peak}.
Without such a peak, there would be no clear selection mechanism for
the most probable value of $\varphi$. If no peak is present and no other criterion is found, one must
refer to the anthropic principle as the only selection principle.
\item \emph{We have to assume some kind of `principle of
mediocrity'} \cite{Vilenkin:2011dd} in order to attribute
predictive power to the result for $E_{\rm{inf}}^{\rm QC}$. This simply means that
in order to interpret a deviation from the peak of ${\cal T}$ in a
probabilistic sense, we have to assume that in the multiverse
context our universe is not very special. Otherwise, a deviation of
the measured $E_{\rm{inf}}^{\rm{obs}}$ from the calculated
$E_{\rm{inf}}^{\rm QC}$, determined by the peak of the tunneling
distribution, would have no predictive power at all, for it might
perfectly be that we simply live in a very improbable branch of the
universe located at the far end of the tail of the probability
distribution, without any contradiction and without being able to
draw any conclusion from it. In contrast, if we assume instead that
our semi-classical branch of the universe is indeed for some reason
mediocre with respect to all other branches, a strong discrepancy
between the measured $E_{\rm{inf}}^{\rm{obs}}$ and the calculated
$E_{\rm{inf}}^{\rm QC}$ could indicate a falsification of the underlying
inflationary model used to calculate $E_{\rm{inf}}^{\rm QC}$. This
assumption is rather speculative and, of course, tightly related to the inevitable problem of
having only one sample universe.
\end{enumerate}
\section{Natural inflation}\label{III}
In what follows, we will focus on a tree-level analysis for the model of natural inflation \cite{Freese:1990rb}.
There are several other models favored by \textsc{Planck} data.
Among them, we mention Starobinsky's $R+R^2$ model \cite{Starobinsky:1980te}, inflation with a strong non-minimal coupling \cite{Fakir:1990eg,Spokoiny:1984bd,Salopek:1988qh,Ketov:2012jt,Kallosh:2013tua} (including non-minimal Higgs-inflation \cite{Bezrukov:2007ep,Barvinsky:2008ia,Bezrukov:2008ej,Barvinsky:2009fy,DeSimone:2008ei,Barvinsky:2009ii,Bezrukov:2010jz,Bezrukov:2009db}) and effective string-inspired models (see, e.g., \cite{Dvali:1998pa,Cicoli:2008gp} and related work).
These models predict a tiny tensor-to-scalar ratio which would be in agreement with the upper bound on $r$ derived by \textsc{Planck}, but
in the light of the \textsc{Bicep2} data they are under some pressure. In contrast, the natural-inflation scenario fits the \textsc{Bicep2} data easily \cite{CKT}.
Moreover, while natural inflation already admits a quantum cosmological analysis at the tree level,
the same is not true for the remaining models listed above. Although the procedure of our tunneling analysis
is applicable in general also for these models, they all share the common feature that their tree-level potentials
become nearly flat for high energies and thus do not feature a strict maximum. Hence, there is no sharp peak
in (\ref{ProbDistr1}). Note, however, that radiative corrections will in general change the structure of the effective potential
such that a tunneling analysis may become possible.
This was, for example, the case in \cite{Barvinsky:2009jd} where the renormalization-group
flow of the Higgs potential due to loop contributions of heavy Standard Model particles leads to the formation
of an additional minimum for high energies, thereby creating a maximum in between the two minima.
We will comment on different models and radiative corrections in section \ref{IV}.
The potential for natural inflation reads \cite{Freese:1990rb}
\begin{equation}
V=\Lambda^4\,\left[1+\text{cos}\left(\varphi/f\right)\right]\,.\label{NatInfPot}
\end{equation}
Here, $\varphi$ is interpreted as a pseudo Nambu--Goldstone boson taking values
on a circle with radius $f$ and angle $\varphi/f\in[0,\,2\,\pi)$. the constants $\Lambda$ and $f$
have dimension of mass and determine the height and the slope of the potential; in the model of natural inflation, one expects $f=O(M_{\rm{P}})$ and $\Lambda\approx M_{\rm{GUT}}\sim 10^{16}$ GeV, the grand-unification scale.
\subsection{Tunneling analysis}
The extrema of (\ref{NatInfPot}) are obtained by the condition
\begin{align}
\frac{\text{d} V}{\text{d}\varphi}\Big|_{\varphi=\varphi_{\rm{ext}}}=-\frac{\Lambda^4}{f}\,\text{sin}\left(\varphi_{\rm{ext}}/f\right)=0\,,
\end{align}
leading to $\varphi_{\rm{ext}}=n\,\pi\,f,\; n\in\mathbb{Z}$. If $\varphi_{\rm{ext}}$ is a maximum,
\begin{align}
\frac{\text{d}^2 V}{\text{d}\varphi^2}\Big|_{\varphi=\varphi_{\rm{ext}}}=
-\frac{\Lambda^4}{f^2}\,\text{cos}\left(n\pi\right)<0\,;\label{SecDervPotNaturalTree}
\end{align}
peak values correspond to even $n$, i.e., due to periodicity,
\begin{align}
\varphi_{\rm{max}}:=2\,\pi nf\,.
\end{align}
The potential at $\varphi_{\rm{max}}$ has the value
\begin{align}
V_{\rm{max}}=2\,\Lambda^4\,.\label{VatMax}
\end{align}
The predictability of the tunneling distribution ${\cal T}(\varphi)$ defined in (\ref{ProbDistr1})
is determined by the sharpness of the peak $\varphi_{\rm{max}}$. The sharpness is here defined as
\begin{align}
{\cal S}:=\frac{(\Delta\varphi)^2}{\left(E_{\rm{infl}}^{\rm{QC}}\right)^2}
\equiv \frac{(\Delta\varphi)^2}{\sqrt{V_{\rm max}}}\,.
\end{align}
Here, the variance $(\Delta\varphi)^2$ measures the width of the peak, and $E_{\rm{infl}}^{\rm{QC}}$ defines the height of the peak.
In view of (\ref{NatInfPot}), the distribution (\ref{ProbDistr1}) is clearly symmetric around $\varphi_{\rm{max}}$.
In order to get a rough estimate for the width $\Delta\varphi$,
we can fit ${\cal T}$ to a normal distribution around the peak $\varphi_{\max}$.
Therefore, we take $\varphi_{\rm{max}}$ as the mean and expand
$\Gamma^{\text{on-shell}}$ from (\ref{EInst}) around $\varphi_{\rm{\max}}$ to second order:
\begin{eqnarray}
\Gamma^{\text{on-shell}}(\varphi)&=&\Gamma^{\text{on-shell}}(\varphi_{\rm{\max}})+\frac{1}{2}
\Gamma^{\text{on-shell}''}(\varphi_{\rm{\max}})(\varphi-\varphi_{\rm{\max}})^2\nonumber\\
&\equiv&
\Gamma^{\text{on-shell}}(\varphi_{\rm{\max}})+\frac{1}{2}\frac{(\varphi-\varphi_{\rm{\max}})^2}{(\Delta\varphi)^2}.
\end{eqnarray}
This leads to the identification
\begin{align}
(\Delta\varphi)^2:= \left.\frac{1}{\Gamma^{\text{on-shell}''}}\right|_{\varphi=\varphi_{\rm{max}}}=\frac{1}{6\,\pi^2}\,\frac{f^2\,\Lambda^4}{M_{\rm{P}}^4}\,,\label{varianceGaussFit}
\end{align}
where primes denote derivatives with respect to $\varphi$. The sharpness of the peak ${\cal S}$ is then estimated as
\begin{align}
{\cal S}=\frac{(\Delta\varphi)^2}{\left(E_{\rm{infl}}^{\rm{QC}}\right)^2}
\approx\frac{1}{6\,\pi^2}\,\frac{f^2\,\Lambda^2}{M_{\rm{P}}^4}\sim\frac{\Lambda^2}{M_{\rm{P}}^2}\sim 10^{-4}\,,
\end{align}
where we have used $f\sim M_{\rm{P}}$ and, using (\ref{EInfQC}) and (\ref{VatMax}),
$E^{\rm{QC}}_{\rm{inf}}\sim\Lambda$.
\subsection{Slow-roll analysis}
The cosmological parameters in the inflationary slow-roll analysis are completely determined by the potential and its derivatives.
The first two slow-roll parameters are given by
\begin{align}\label{srpa}
\epsilon_{\rm{v}}:={}&\frac{M_{\rm{P}}^2}{2}\,\left(\frac{V'}{V}\right)^2=\frac{M_{\rm{P}}^2}{2\, f^2}\,\text{tan}^2\left[\varphi/(2\, f)\right],\qquad \eta_{\rm{v}}:={}M_{\rm{P}}^2\,\left(\frac{V''}{V}\right)=-\frac{M_{\rm{P}}^2\, \text{cos}(\varphi/f)}{f^2\, \left[1 + \text{cos}(\varphi/f)\right]}\,,
\end{align}
with $\epsilon_{\rm{v}}\ll1$ and $|\eta_{\rm{v}}|\ll 1$ during inflation.
The scalar spectral index and the tensor-to-scalar ratio read
\begin{align}
n_{\rm{s}}={}&1+2\,\eta_{\rm{v}}-6\,\epsilon_{\rm{v}}=-\frac{M_{\rm{P}}^2}{f^2}\,\frac{3-\text{cos}(\varphi/f)}{1+\text{cos}(\varphi/f)}\,,\label{SpectralIndex}\\
r={}&16\,\epsilon_{\rm{v}}=\frac{8\,M_{\rm{P}}^2}{ f^2}\,\text{tan}^2\left[\varphi/(2\, f)\right]\,.\label{TensorToScalarRatio}
\end{align}
All cosmological observables have to be evaluated at $\varphi_{*}$, the field value that corresponds to the moment where the pivot mode $k_{*}$ first crosses the Hubble scale.
The number of e-folds $N_*$, which is a measure of how long inflation lasted, connects the end of inflation $\varphi_{\rm{end}}$ with the value $\varphi_{*}$:
\begin{align}
N_{*}=\int_{\varphi_{*}}^{\varphi_{\rm end}}\frac{\text{d}\varphi}{M_{P}^2}\,\frac{V}{V'}=\frac{2\,f^2}{M_{\rm{P}}^2}\,\ln\left[\frac{\text{sin}\left(\frac{\varphi_{\rm{end}}}{2\,f}\right)}{\text{sin}\left(\frac{\varphi_{*}}{2\,f}\right)}\right]\,.\label{EFoldsNatural}
\end{align}
The value $\varphi_{\rm{end}}$ that determines the upper integration bound in (\ref{EFoldsNatural}) is defined by the breakdown of the slow-roll approximation at $\epsilon_{\rm{v}}(\varphi_{\rm{end}}):=1$,
\begin{align}
\varphi_{\rm{end}}=2\, f\, \text{arctan} (\sqrt{2}\, f/M_{\rm{P}})\,.\label{PhiEnd}
\end{align}
Inserting (\ref{PhiEnd}) in (\ref{EFoldsNatural}), solving for $\varphi_{*}$ and parametrizing $f$
in units of $M_{\rm{P}}$, we find
\begin{align}
\varphi_{*}={}&2 M_{\rm{P}}\,\alpha\, \text{arcsin}\left(\frac{\alpha\,e^{-N_{*}/2 \alpha^2}}{\sqrt{
1/2 + \alpha^2}}\right),\label{phiStar}
\end{align}
where $\alpha:=f/M_{\rm{P}}$.
Evaluating the potential (\ref{NatInfPot}) at $\varphi_{*}$ yields
\begin{align}
V(\varphi_{*})=2\,\Lambda^4\,\left[1-\delta_{V}(\alpha,N_{*})\right]\,,
\end{align}
where we have defined
\begin{align}
\delta_{V}(N_{*},\alpha):=\frac{2\,e^{-N_{*}/\alpha^2}\,\alpha^2}{1+2\,\alpha^2}\,.
\end{align}
The consistency condition (\ref{ConsistencyCond1}) implies $V_{\rm{max}}=V(\varphi_{*})$, or $ \delta_{V}=0$. Since $N_{*}$ should be in the range $50\lesssim N_{*}\lesssim 60$, this requires $\alpha=0$ for (\ref{ConsistencyCond1}) to be satisfied exactly.
However, by inspection of (\ref{NatInfPot}), $\alpha=0$ is not allowed.
Moreover, for $N_{*}=60$, \textsc{Planck} constraints on $(n_{\rm{s}},\,r)$ \cite{Ade:2013zuv,Ade:2013uln} imply the following bound on $\alpha$ \cite{Tsujikawa:2013ila}:
\begin{align}
\label{alphaPlanck}
\alpha>4.6\quad(95\%\;\,\text{CL})\,.
\end{align}
For fixed $N_{*}$,
the function $\delta_{V}(N_{*},\,\alpha)$ varies between zero and one.
As can be seen in figure \ref{fig1}, $\delta_{V}(N_{*},\,\alpha)$ first grows rapidly from zero at $\alpha=0$
to $0.9$ at around $\alpha=20$ and then slowly asymptotes to 1 for $\alpha\to\infty$.
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.8]{DeltaVofAlphaNaturalInflation
\caption{\label{fig1} The function $\delta_V(N_{*},\alpha)$ as a function of $\alpha$ for values of $N_{*}\in[50,60]$. The upper line corresponds to $N_{*}=50$, the lower line to $N_{*}=60$. The inset shows the region with $\alpha$ in the $68\%$ CL range $5.1< \alpha < 7.9$ (see (\ref{PlanckDataRange})).
}
\end{center}
\end{figure}
\noindent As already mentioned in the introduction, the consistency condition will lead in general to an exact constraint.
But since we only have considered the tree-level approximation to obtain $E^{\rm{QC}}_{\rm{inf}}$ and also made the
slow-roll approximation to obtain $E_{\rm{inf}}^{\rm{model}}$, this relation cannot be expected to hold exactly.
Nevertheless, the quantum cosmological analysis leads to an approximate consistency requirement that excludes certain values of $\alpha$ for a given $N_*$.
\noindent Since $\delta_{V}$ enters as the difference $1-\delta_{V}$ in $V(\varphi_{*})$,
it will only lead to significant changes in $E_{\rm{inf}}^{\rm{model}}$ when $\delta_V\approx1$.
For example, a $\delta_{V}\approx 0.9999$ will lead to
$V(\varphi_{*})=2\,\Lambda^4\left(1-\delta_{V}\right)=2\,\Lambda^4\,10^{-4}$
and will affect the energy scale of inflation by one order of magnitude,
$E_{\rm{inf}}^{\rm{model}}=V^{1/4}(\varphi_{*})=2^{1/4}\times 10^{-1}\,\Lambda$.
Then, even the approximate consistency condition
$E^{\rm{QC}}_{\rm{inf}}\approx E_{\rm{inf}}^{\rm{model}}$ would no longer hold.
This case would correspond to a value of $\alpha\approx 710$ for $N_{*}=50$ and $\alpha\approx 780$ for $N_{*}=60$
and is depicted in figure \ref{fig2}. In order for the consistency condition to hold at least approximately,
we have to impose in this model the constraint
\begin{equation}
\alpha\ll 700.
\end{equation}
This is, of course, compatible with the \textsc{Planck} constraint (\ref{alphaPlanck}).
\begin{figure}[ht!]
\begin{center}
\includegraphics[scale=0.9]{VofAlphaNaturalInflation
\caption{\label{fig2} A zoomed-in region of the function $V(N_{*},\alpha)/(2\,\Lambda^4)=1-\delta_V$ as a function of $\alpha$ for values of $N_{*}\in[50,60]$. The upper purple line corresponds to $N_{*}=60$, the lower purple line to $N_{*}=50$. The lower area, colored in light red (in black-and-white printing: light gray), corresponds to the region where $E_{\rm{inf}}^{\rm{model}}<10^{-1}\,E_{\rm{inf}}^{\rm QC}$.
}
\end{center}
\end{figure}
\begin{figure}[ht!]
\begin{center}
\includegraphics[scale=0.9]{rNaturalInflationBicep2TwoSigma
\caption{\label{fig3} The tensor-to-scalar ratio $r_{*}$ as a function of $\alpha$ for values of $N_{*}\in[50,60]$.
The $68$ \% CL and $95$ \% CL shaded regions
experimentally excluded by \textsc{Bicep2} are bounded from above by,
respectively, the orange (upper) and red (lower) horizonal line. The upper bounds on $r_{*}$ are not shown,
as they do not constrain $\alpha$. The upper curve corresponds to $N_{*}=50$, the lower one to $N_{*}=60$.
For $N_{*}\gtrsim 52$, $r_{*}$ falls in the forbidden $68$ \% CL region for all values of $\alpha$.
The inset shows the intersection of the $N_{*}\in[50,60]$ band with the $95$ \% CL bound.}
\end{center}
\end{figure}
Although a quantum cosmological bound on $\alpha$ derived in this way is rather arbitrary and clearly depends
on the aimed precision for the approximate quantum cosmological consistency condition to hold, it is obvious
that this bound is not as restrictive as the constraints on $\alpha$ coming from the comparison of the
inflationary model itself with observational data, which require $\alpha\sim O(10)$.
As described in section \ref{sec2.2}, the observed value $E_{\rm{inf}}^{\rm{obs}}$ is very close to the quantum
cosmological predicted value of $E_{\rm{inf}}^{\rm{QC}}\approx2^{1/4}\,\Lambda$ with $\Lambda\approx10^{16}$ GeV
around the GUT scale. However, the observational constraints on the spectral index and the tensor-to-scalar ratio by \textsc{Planck} yield a much sharper condition on $\alpha$ than the restrictions from quantum cosmology.
For $N_{*}=60$, \textsc{Planck} data \cite{Ade:2013zuv,Ade:2013uln} constrain $\alpha$ to lie in the interval \cite{Tsujikawa:2013ila}
\begin{align}
5.1<\alpha<7.9\quad(68\%\,\text{CL})\,.\label{PlanckDataRange}
\end{align}
As shown in figure \ref{fig1}, in this range $\delta_V\approx 0.1\div 0.5$ is small enough to respect the condition (\ref{ConsistencyCond2}), at least within the order of magnitude of the envisaged accuracy.
For the `classical' natural inflation model, a quick estimate for the constraint analogous
to (\ref{PlanckDataRange}) can be obtained by looking at the intersection points in
the $(n_{\rm{s}\,*},\alpha)$ and $(r_*,\alpha)$ planes between the experimental $68$ \% CL and $95$ \% CL
bounds on $n_{\rm{s}\,*}$ and $r_{*}$ and the corresponding model-dependent analytic expressions
for $n_{\rm{s}\,*}$ and $r_{*}$. The expressions for $n_{\rm{s}}$ and $r$ given in
(\ref{SpectralIndex})--(\ref{TensorToScalarRatio}) and evaluated at $\varphi_{*}$ take a particularly simple form when expressed in terms of $\delta_V$ and $\alpha$:
\begin{align}
n_{\rm{s}\,*}={}1+\frac{1}{\alpha^2}\,\frac{\delta_V(N_{*},\,\alpha)+1}{\delta_V(N_{*},\alpha)-1}\,,\qquad r_{*}={}\frac{8}{\alpha^2}\,\frac{\,\delta_V(N_{*},\,\alpha)}{1-\,\delta_V(N_{*},\,\alpha)}\,.
\end{align}
Here we show the results of this procedure by confronting the analytically obtained $r_{*}$ with the estimated bounds on $r_{*}$ from the combined \textsc{Planck}+\textsc{WP}+highL+\textsc{Bicep2} 1$\sigma$ and 2$\sigma$ contours of the tensor-to-scalar ratio for fixed central value $n_{s\,*}=0.96$ \cite{Ade:2014xna}.
The 1$\sigma$ contour leads to the bounds $r_{*}=0.20^{+0.07}_{-0.05}$ for the tensor-to-scalar ratio. Comparison with figure \ref{fig3} leads to the lower bound $\alpha\gtrsim 22$ for $N_{*}=50$.
The 2$\sigma$ contour of $r_{*}$ for the central value $n_{s\,*}=0.96$ roughly yields the bounds $r_{*}=0.20^{+0.1}_{-0.1}$ \cite{Ade:2014xna}, constraining $\alpha\gtrsim 8$ for $N_{*}=50$ and $\alpha\gtrsim 11$ for $N_{*}=60$, as can be seen in the inset of figure \ref{fig3}. A more elaborate likelihood analysis refines these rough estimates \cite{CKT}:
at the 1$\sigma$ level, $\alpha\gtrsim 9$ for $N_{*}=50$, while at the 2$\sigma$ level $\alpha\gtrsim 7\div 8$ for $N_*=50$ and $\alpha\gtrsim 9$ for $N_*=60$.
Thus, by comparing the classical inflationary predictions with observational data we have obtained a constraint on $\alpha$ of the order of magnitude $\alpha\sim O(10)$ far below the threshold $\alpha\approx 700$ at which a conflict with the quantum cosmological compatibility constraint would arise. We can therefore conclude that, to a good approximation, the consistency condition is satisfied for all experimentally allowed values of $\alpha$ according to both \textsc{Planck} and \textsc{Bicep2} data.
\section{Conclusions}\label{IV}
The purpose of our paper is to present a general method to discussing quantum cosmological consistency conditions
for inflation and to present a concrete example in detail. We have focused on the tunneling condition for the wave function
of the universe because it allows one to implement these consistency conditions in a straightforward
manner. In principle, however, the method can also be used to study other conditions such as the no-boundary condition,
although this condition does not lead to the prediction of inflation in the usual situations.
A central concept in our analysis is the use of the effective action. Because this action can
in general not be evaluated exactly,
it is necessary to perform a loop expansion. The example we have discussed in detail here is natural inflation.
There, the restriction to the tree-level approximation seems sufficient because at tree level
the potential (\ref{NatInfPot}) features a strict maximum necessary
for the tunneling analysis, while this is not the case for the inflationary models mentioned at the beginning of
section \ref{III}. For these models, therefore, one has to go at least to the one-loop level. This will be
the topic of future investigations.
All inflationary single-field models favored by recent \textsc{Planck} data can be collectively covered by the class of scalar-tensor theories with the action
\begin{align}
S=\int\text{d}^4x\,\sqrt{|g|}\,\left[U(\varphi)\,R-\frac{G(\varphi)}{2}\,(\nabla\varphi)^2-V(\varphi)\right]\label{ActionScalarTensor1}\,,
\end{align}
where $U(\varphi)$, $G(\varphi)$ and $V(\varphi)$ are arbitrary functions of the inflaton field $\varphi$.
Quantum corrections usually modify the shape of the effective potential and the location of its extrema. Thus, even for the inflationary models where a tunneling analysis was not applicable at the tree level, already at the one-loop level the changing structure of the effective potential may lead to a strict maximum such that a tunneling analysis becomes possible.
The one-loop divergences for the action (\ref{ActionScalarTensor1}), necessary for the renormalization of (\ref{ActionScalarTensor1}), can be extracted from \cite{Steinwachs:2011zs} where a more general action with a $O(N)$-symmetric scalar multiplet was considered.\footnote{A similar analysis has been performed in \cite{Shapiro:1995yc} for a single scalar field. For the divergences that can be absorbed in the functional couplings $U$, $G$ and $V$, the results of \cite{Shapiro:1995yc} coincide with those derived in \cite{Steinwachs:2011zs}.}
However, two important points have to be taken into account for such a quantum analysis.
First, if the inflaton field is coupled to additional matter, matter loop contributions usually lead to a significant modification of the effective potential. This fact was crucial in the renormalization-group improved investigation of the tunneling scenario for non-minimal Higgs inflation \cite{Barvinsky:2009jd}.
Second, the analysis of $f(R)$ theories and models with a non-minimal coupling to gravity requires extra care.
The tunneling formalism presented here has been developed for a minimally coupled scalar field.
By a conformal transformation of the metric field and a subsequent redefinition of the scalar field, the action (\ref{ActionScalarTensor1}) can be brought to the so-called Einstein-frame parametrization, which, in the absence of matter, formally resembles the situation of a scalar field minimally coupled to gravity.
It is well known that $f(R)$ theories with $f_{,RR}\neq0$ also admit an on-shell reformulation as scalar-tensor theories of the type (\ref{ActionScalarTensor1}) with $G(\varphi)=0$.
Therefore, they can ultimately be cast as well in the Einstein-frame parametrization.
While field reparametrizations lead to equivalent descriptions at the tree level, quantum divergences induce a frame dependence of the off-shell effective action \cite{Kamenshchik:2014waa}. In \cite{Vilkovisky:1984st}, the origin of this parametrization dependence was traced back to the non-covariant definition of the off-shell effective action on configuration space and in \cite{Steinwachs:2013tr,Kamenshchik:2014waa} this idea was applied to the cosmological context. In particular, when applied to the debate `Jordan frame vs. Einstein frame', it was pointed out that within a non-covariant formalism quantum corrections will naturally induce a frame dependence when the conformal transformation of the metric field as well as the transformation of the scalar field are viewed as field reparametrizations in configuration space. The tunneling consistency condition may thus serve not only as a tool to distinguish between competing models of inflation, but also to select a preferred parametrization \emph{in the absence} of a covariant formulation.
In our tree-level analysis of natural inflation, we have derived a consistency condition which restricted the parameter $\alpha=f/M_{\rm{P}}$ for a given number of e-folds $N_{*}$.~Our result ensures consistency with the quantum cosmological tunneling origin. For the tree-level analysis of the natural inflation model,
the restriction of $\alpha$ from the quantum cosmological consistency condition
is much weaker than the observational constraints on the inflationary parameters itself.
We have found that natural inflation with a quantum tunneling origin
is consistent with the 2013 \textsc{Planck} release as well as with a large tensor-to-scalar ratio as found by \textsc{Bicep2}. Making use of the general formalism presented here for the natural inflation model, the analysis can easily be extended to all kind of inflationary models, including their modification by quantum corrections. Such investigations will shed further light on the relation between a fundamental theory of quantum cosmology and cosmological observations.
\acknowledgments{The work of G. C. is under a Ram\'on y Cajal contract. C. K. thanks the Max Planck Institute for Gravitational Physics (Albert Einstein Institute) in Potsdam, Germany, for kind hospitality while part of this work was done. C. S. is grateful to A. Yu. Kamenshchik for fruitful discussions and valuable comments. }
|
\section{Introduction}
Let $V$ be a vector space
over a field ${\mathbb F}$ and denote by ${\mathrm{PG}}\,(V,{\mathbb F})$ the usual projective geometry given by the lattice of subspaces of $V$.
If ${\mathbb F}$ is the
finite field with $q$ elements ${\mathbb F}_q$ and $\dim V=n$, then we
shall write, as customary, ${\mathrm{PG}}\,(n-1,q):={\mathrm{PG}}\,(V,{\mathbb F}_q)$.
Recall that if ${\mathbb K}$ is a subfield of ${\mathbb F}$ and $[{\mathbb F}:{\mathbb K}]=t$
then $V$ is also endowed with
the structure of a vector space $\hat{V}$ of dimension $rt$ over ${\mathbb K}$.
We shall denote by ${\mathrm{PG}}\,(V,{\mathbb K})$,
the projective geometry given by the lattice of the subspaces
of $V$ with $V$ is regarded as a vector space over ${\mathbb K}$.
As each point of ${\mathrm{PG}}\,(V,{\mathbb F})$ corresponds to
a $(t-1)$--dimensional projective
subspace of ${\mathrm{PG}}\,(V,{\mathbb K})$, it is possible to represent the
projective space ${\mathrm{PG}}\,(V,{\mathbb F})$ as a subvariety $\mathcal V_{rt}$
of the Grassmann manifold $\mathcal G_{rt,t}$ of the
$t$--dimensional vector subspaces of $V$;
see \cite{V}.
A \emph{linear set} of ${\mathrm{PG}}\,(V,{\mathbb F})$ is a set of points defined by an
additive subgroup of $V$. More in detail, let ${\mathbb K}\leq{\mathbb F}$, as above,
and suppose $W$ to be a vector space of dimension $m+1$ over
${\mathbb K}$. Then, the \emph{${\mathbb K}$--linear set} $\Lambda$
of ${\mathrm{PG}}\,(V,{\mathbb F})$ defined by $W$ consists
of all points of ${\mathrm{PG}}\,(V,{\mathbb F})$ of the form
\[ \Lambda=\{ \langle X\otimes{\mathbb F} \rangle | X\in W \}. \]
Linear sets have been widely used to investigate
several different aspects of finite geometry, the two most remarkable
being blocking sets and finite semifields.
Following the approach pioneered by Schubert in
\cite{schubert}, it can be seen how the representation of subspaces
on the Grassmann manifold $\mathcal G$ might provides an important tool for the study of their
behaviour and their intersections.
In the present paper, we are interested in the representation of a
${\mathbb K}$--linear set $\Lambda$ on $\mathcal G$ and in determining the space of linear
equations defining it as linear section of $\mathcal V_{rt}$.
Throughout this paper, when discussing Grassmannians we shall use
vector dimension for the spaces under consideration,
whereas we shall consider projective dimension
when discussing projective spaces.
As for algebraic varieties $V$ defined over a field ${\mathbb F}$,
we shall always mean by \emph{dimension}
always mean the dimension of the variety
$\overline{\mathcal{V}}$, regarded over the algebraic closure
$\overline{{\mathbb F}}$ of
${\mathbb F}$, defined by the same equations as $\mathcal V$.
\section{Grassmannians and Schubert varieties}
Fix an $n$ dimensional vector space $V=V_n({\mathbb F})$ over ${\mathbb F}$ and
write $G(n,k)$, $k<n$, for the set of all the $k$--subspaces of $V$.
It is well known that $G(n,k)$ is endowed with the structure
of a partial linear space and it can be embedded via the
Pl\"{u}cker map
\[ \varepsilon_{k}:\begin{cases}
G(n,k) \rightarrow \bigwedge^k V \\
W=\langle v_1,v_2,\ldots,v_k \rangle \mapsto
v_1 \wedge v_2 \wedge \cdots \wedge v_k
\end{cases}\]
in
the projective
space ${\mathrm{PG}}\,(\bigwedge^k V,{\mathbb F})$; here $\dim_{{\mathbb F}}\bigwedge^k V = \binom{n}{k}$.
The image of $\varepsilon_{k}$, say $\mathcal G_{nk}$, is an algebraic
variety of ${\mathrm{PG}}\,(\bigwedge^k V,{\mathbb F})$ whose
points correspond exactly to the totally decomposable
$1$--dimensional subspaces of $\bigwedge^k V$.
We now recall some basic properties of alternating multilinear
forms. Let $U$ be a vector
space defined on ${\mathbb F}$ and let $V^k:=\displaystyle\underbrace{V \times
V \times \cdots \times V}_{k \text{ times}}$. A $k$--linear map $f:
V^k \longrightarrow U$ is \emph{alternating} if $f(v_1,v_2,\ldots,v_k)=0$ when $v_i=v_j$ for some $i \neq j$.
This implies that
$\forall i,j \in \{1,2,\ldots,k\}$,
$f(v_1,\ldots,v_i,\ldots,v_j,\ldots,v_k)=
-f(v_1,\ldots,v_j,\ldots,v_i,\ldots,v_k)$.
\begin{theorem}[Universal property of the $k^{th}$ exterior power of a
vector space, {\cite[Theorem 14.23]{ALA}}]
\label{ut}
A map $f: V^k \longrightarrow U$ is alternating
$k$--linear if, and only if, there is a
linear map $\overline{f}:
\bigwedge^k V \longrightarrow U$ with $\overline{f}(v_1 \wedge v_2
\wedge \cdots \wedge v_k) = f(v_1,v_2,\ldots,v_k)$.
The map $\overline{f}$ is uniquely determined.
\end{theorem}
\begin{corollary}
The ${\mathbb F}$--vector space
\[ {\mathrm{Alt}}\,^k(V,U):=\{f: V^k \longrightarrow U| f \text{ is $k$--linear and alternating}\} \]
is isomorphic to the ${\mathbb F}$--vector space
$\mathrm{Hom}(\bigwedge^kV,U)$.
\end{corollary}
In particular, let
$(\bigwedge^k V)'$ be the dual of $\bigwedge^k V$. Then,
$(\bigwedge^k V)'\simeq {\mathrm{Alt}}\,^k(V,{\mathbb F})$.
Furthermore,
we also have
$(\bigwedge^k V)' \simeq \bigwedge^{n-k}V$.
Actually, $(\bigwedge^k V)'$
is spanned by linear maps of type acting on the pure vectors of
$\bigwedge^k V$ as
\[
v_1 \wedge v_2 \wedge \cdots \wedge v_k \mapsto v_1 \wedge v_2 \wedge \cdots \wedge v_k \wedge w_{k+1} \wedge \cdots \wedge w_n,
\]
and extended by linearity; see \cite[Chapter 5]{Gr} for more details.
Here
$(w_{k+1},w_{k+2},\ldots,w_n)\in V^{n-k}$ is a fixed $(n-k)$--ple.
Let $F=A_1< A_2<\cdots< A_k$ be a proper flag consisting of $k$
subspaces of $V$.
The\- \emph{Schubert variety}\- $\Omega(F)=\Omega(A_1,A_2,\ldots,A_k)$
induced by $F$ is the subvariety of $\mathcal G_{nk}$ corresponding to
all $W\in G(n,k)$ such that $\dim W\cap A_i\geq i$ for all $i=1,\ldots,k$.
It is well known, see \cite[Corollary 5]{Kleiman_Laksov} and
\cite[Chapter XIV]{HP}, that a
Schubert variety is actually a linear section of the Grassmannian.
Furthermore, as the general linear group is flag--transitive,
all Schubert varieties defined by flags of the same kind,
i.e. with the same list of dimensions $a_i=\dim A_i$, turn out to be
projectively equivalent.
In the present work we shall be mostly concerned with Schubert varieties
of a very specific form, namely those for which $a_1=h\leq n-k$ and
$a_i=n-k+i$ for any $i=2,\ldots,k$.
Under these assumptions, $\Omega(A_1):=\Omega(A_1,A_2,\ldots,A_k)$
depends only on $A_1$ and corresponds to
the set of all $k$--subspaces with non--trivial intersection with
$A_1$.
Indeed, using once more \cite[\S 2, Corollary 5]{Kleiman_Laksov},
we see that $\Omega(A_1)$ is
the complete intersection of $\mathcal G_{nk}$ with a linear subspace of
codimension $\binom{n-h}{k}$, meaning that the
subspace of the dual of $\bigwedge^k V$ of
the elements vanishing on $\Omega(A_1)$ has dimension $\binom{n-h}{k}$.
Using Theorem \ref{ut} we can provide
a description of the space of the linear maps vanishing on $\Omega(A_1)$.
For any $k$--linear map $f:V^k\to U$,
define the \emph{kernel} of $f$ as
\[ \ker f=\{w
\in V | f(w,v_2,\ldots,v_k)=0, \forall v_i \in V \}. \]
It is straightforward to see that $\ker f$ is a subspace of $V$;
when $f$ is alternating and non--null,
the dimension of $\ker f$ is trivially bounded from
above, as recalled by the following proposition.
\begin{proposition}
\label{p2}
The kernel of a non--null $k$--linear alternating
map $f$ of an $n$--dimensional vector space $V$ has dimension at most $n-k$.
\end{proposition}
\begin{proof}
By Theorem \ref{ut}, $f$ can be
regarded as a linear functional
$\overline{f}:\wedge^k V\to{\mathbb F}$
where
\[ f(v_1,\ldots,v_k)=\overline{f}(v_1\wedge v_2 \ldots\wedge v_k). \]
Let $E=\langle v_1,\ldots,v_k\rangle$ and observe that
$f(E):=f(v_1,\ldots,v_k)=0$ when $\dim E<k$ or $\dim E\cap\ker f>0$
In particular, if $\dim\ker f>n-k$ we always have
$\dim E\cap\ker f>0$ for $\dim E \geq k$; this gives $f\equiv 0$.
\end{proof}
\begin{proposition}\label{kernel}
The subspace of $(\bigwedge^k V)'$
consisting of the linear forms vanishing on
$\Omega(A_1)$ is isomorphic to the
subspace of the $k$--linear alternating maps whose kernel contains $A_1$.
In particular, if $h=\dim A_1\leq n-k$, then there exists a basis for this
subspace consisting
of maps whose kernel contains $A_1$ and has dimension $n-k$.
\end{proposition}
\begin{proof}
Let $f: \bigwedge^k V \to {\mathbb F}$ be a linear function vanishing
on $\Omega(A_1)$. In particular, $f$ vanishes on all subspaces $E$ with
$\dim E\cap A_1>0$. Thus, by the definition of kernel, $A_1\leq\ker f$.
If $h>n-k$, then by Proposition \ref{p2}
the only function vanishing on $\Omega(A_1)$
is $f\equiv 0$ and there is nothing to prove.
Let now $h \leq n-k$. By $(\bigwedge^k V)' \simeq \bigwedge^{n-k} V$, let us
consider the linear maps:
\[
v_1 \wedge v_2 \wedge \cdots \wedge v_k \mapsto v_1 \wedge v_2 \wedge \cdots \wedge v_k \wedge w_{k+1} \wedge \cdots \wedge w_n \]
where $\{w_{k+1},w_{k+2},\ldots,w_n \}$ is a set of $n-k$ linearly
independent vectors such that $A_1 \leq \langle
w_{k+1},w_{k+2},\ldots,w_n \rangle$. The kernel of such a map is the
subspace $\langle w_{k+1},w_{k+2},\ldots,w_n \rangle$.
It is well known that the dimension
of the Pl\"{u}cker embedding of the $(n-k)$--subspaces containing a
fixed $h$--dimensional subspace is $\binom{n-h}{k}$.
As, by \cite[\S 2, Corollary 5]{Kleiman_Laksov}
this is also the dimension of the space of the linear functions vanishing on $\Omega(A_1)$, we have the
aforementioned linear maps can be used to also determine a basis for it.
\end{proof}
\section{Desarguesian spreads and linear sets}
A $(t-1)$--\emph{spread} $\mathcal S$ of ${\mathrm{PG}}\,(V,{\mathbb F})$ is a partition of the
point-set of ${\mathrm{PG}}\,(V,{\mathbb F})$
in subspaces of fixed projective dimension $t-1$. It is well known, see
\cite{Segre}, that spreads exist if and only if $t|n$. Henceforth, let
$n=rt$ and denote by $V_1$ a ${\mathbb F}$--vector space such that
$\dim_{{\mathbb F}}V_1=n+1$ and $V < V_1$. Under these
assumptions we can embed ${\mathrm{PG}}\,(V,{\mathbb F})$ as a
hyperplane in ${\mathrm{PG}}\,(V_1,{\mathbb F})$. Consider
the point--line geometry
$A(\mathcal S)$ whose points
are the points of ${\mathrm{PG}}\,(V_1,{\mathbb F})$ not contained in ${\mathrm{PG}}\,(V,{\mathbb F})$ and whose
lines of are the subspaces of ${\mathrm{PG}}\,(V_1,{\mathbb F})$
intersecting ${\mathrm{PG}}\,(V,{\mathbb F})$ in exactly one spread element.
We say that
$\mathcal S$ is a Desarguesian spread if $A(\mathcal S)$ is a Desarguesian
affine space.
Here we shall focus on spaces defined over finite fields.
We recall that, up to projective equivalence,
Desarguesian spreads are unique and their
automorphism group contains a copy of ${\mathrm{PGL}}\,(r,q^t)$.
There are basically two main ways
to represent a Desarguesian spread.
Let $V:=V(r,q^t)$ be the standard $r$--dimensional vector space over
${\mathbb F}_{q^t}$ and write ${\mathrm{PG}}\,(r-1,q^t)={\mathrm{PG}}\,(V,q^t)$.
When we regard $V$ as an ${\mathbb F}_q$--vector space,
$\dim_{{\mathbb F}_q}V(r,q^t)=rt$; hence,
${\mathrm{PG}}\,(V,q)$ corresponds to ${\mathrm{PG}}\,(rt-1,q)$;
furthermore, a point $\langle (x_0,x_1,\ldots,x_{r-1}) \rangle$ of
${\mathrm{PG}}\,(r-1,q^t)$ corresponds to the $(t-1)$--dimensional subspace of
${\mathrm{PG}}\,(rt-1,q)$ given by $\{\lambda (x_0,x_1,\ldots,x_{r-1}), \lambda
\in {\mathbb F}_{q^t}\}$. This is the so called the \emph{
${\mathbb F}_q$--linear representation} of
$\langle (x_0,x_1,\ldots,x_{r-1}) \rangle$. The set $\mathcal S$,
consisting of the $(t-1)$--dimensional subspaces of ${\mathrm{PG}}\,(rt-1,q)$ that are the linear representation of a point of ${\mathrm{PG}}\,(r-1,q^t)$, is
a partition of the point set of ${\mathrm{PG}}\,(rt-1,q)$ and it is the
${\mathbb F}_q$--linear representation of ${\mathrm{PG}}\,(r-1,q^t)$.
\begin{theorem}[\cite{BarlCof}]
The ${\mathbb F}_q$--linear representation of ${\mathrm{PG}}\,(r-1,q^t)$ is a Desarguesian
spread of ${\mathrm{PG}}\,(rt-1,q)$ and conversely.
\end{theorem}
Throughout this paper we shall extensively
use the following result: if $\sigma$
is a ${\mathbb F}_q$--linear collineation of ${\mathrm{PG}}\,(n-1,q^t)$ of order $t$, then the
subset ${\mathrm{Fix}\,}(\sigma)$ of all elements of ${\mathrm{PG}}\,(n-1,q^t)$ point--wise
fixed by $\sigma$ is a subgeometry isomorphic to ${\mathrm{PG}}\,(n-1,q)$.
This is a straightforward consequence of the fact that there is just one
conjugacy class of ${\mathbb F}_q$--linear collineations of order $t$ in
$P\Gamma L(n,q^t)$, namely that of $\mu: X\to X^q$.
In particular,
all subgeometries ${\mathrm{PG}}\,(n-1,q)$ are projectively equivalent to the
set of fixed points of the map
$(x_0,x_1,\ldots,x_{n-1}) \mapsto (x_0^q,x_1^q,\ldots,x_{n-1}^q)$.
\begin{lemma}\label{subspace}\cite[Lemma 1]{LunardonNS}
Let $\Sigma \simeq {\mathrm{PG}}\,(n-1,q)$ be a subgeometry of ${\mathrm{PG}}\,(n-1,q^t)$ and
let $\sigma$ be the ${\mathbb F}_q$--linear collineation of order $t$ such
that $\Sigma={\mathrm{Fix}\,}(\sigma)$. Then a subspace $\Pi$ of ${\mathrm{PG}}\,(n-1,q^t)$ is
fixed set--wise by $\sigma$ if and only if
$\Pi \cap \Sigma$ has the
same projective dimension as $\Pi$.
\end{lemma}
Take now $V$ to be a $rt$--dimensional projective space over
${\mathbb F}_{q^t}$ and let $U_i$ be the subspace of $V$ defined by the equations
$x_j=0$, $\forall j \notin \{ir+1,ir+2,\ldots,(i+1)r\}$.
Then, clearly, $V=U_0\oplus U_1 \oplus \cdots \oplus U_{t-1}$.
For any $(x_0,\ldots,x_{rt-1}) \in V$,
write $\mathbf{x}^{(i)}=x_{ir},\ldots, x_{(i+1)r-1}$, and consider
the ${\mathbb F}_q$--linear collineation of ${\mathrm{PG}}\,(rt-1,q^t)$ of order $t$ given
by
\[ \sigma:
(\mathbf{x}^{(0)},\mathbf{x}^{(1)},\ldots,\mathbf{x}^{(t-1)}) \mapsto
(\mathbf{x}^{(t-1)q},\mathbf{x}^{(0)q},\ldots,\mathbf{x}^{(t-2)q}). \]
As seen above,
the set ${\mathrm{Fix}\,}\sigma$ is a subgeometry
${\mathrm{PG}}\,(rt-1,q^t)$ isomorphic to ${\mathrm{PG}}\,(rt-1,q)$: in the remainder
of this section we shall denote such subgeometry just as
${\mathrm{PG}}\,(rt-1,q)$.
In particular, we see that ${\mathrm{Fix}\,}\sigma={\mathrm{PG}}\,(rt-1,q)$ consists of points of the form $\{(\mathbf{x},\mathbf{x}^{q},\ldots,\mathbf{x}^{q^{t-1}}),\mathbf{x}=x_0,x_1,\ldots,x_{r-1}; x_i \in {\mathbb F}_{q^t}\}$.
Observe that we have
$\sigma(U_i)=U_{{i+1}\pmod t}$ and the semilinear collineation $\sigma$
acts cyclically on the $U_i$;
furthermore, for any $u\in U_0$, $u\neq0$,
we have $u^{\sigma^i}\in U_i$ and the set
$\{u^{\sigma^i} : i=1,\ldots, t\}$ is linearly independent.
In particular, the subspace
$\Pi^*_u=\langle
\mathbf{u},\mathbf{u}^{\sigma},\ldots,\mathbf{u}^{\sigma^{t-1}} \rangle$
has projective dimension $t-1$.
The set
$\mathcal S^*=\{\Pi^*_u, u \in U_0\}$
consists of $(t-1)$--spaces and it
is a ${\mathbb F}_q$--rational normal $t$--fold scroll of ${\mathrm{PG}}\,(rt-1,q^t)$ over
${\mathrm{PG}}\,(r-1,q^t)={\mathrm{PG}}\,(U_0,q^t)$.
Any subspace $\Pi^*_u$ is fixed set-wise by $\sigma$;
hence, by Lemma \ref{subspace}$, \Pi_u:=\Pi_u^* \cap \Sigma$ has the same
projective dimension $t-1$. The collection of $(t-1)$--subspaces
$\mathcal S=\{\Pi_u | u \in U_0 \}$ is a spread of ${\mathrm{PG}}\,(rt-1,q)$, see \cite{Segre},
also called the \emph{Segre spread} of ${\mathrm{PG}}\,(rt-1,q)$.
\begin{theorem}[\cite{Desarguesian_spread}]
The Segre spread of ${\mathrm{PG}}\,(rt-1,q)$, obtained as the intersection with
${\mathrm{PG}}\,(rt-1,q)$ with a ${\mathbb F}_q$--rational normal $t$--fold scroll
of ${\mathrm{PG}}\,(rt-1,q^t)$
over ${\mathrm{PG}}\,(r-1,q^t)$, is a Desarguesian spread.
\end{theorem}
The correspondence between linear representations and Segre
spreads is given as follows:
\[
\langle u \rangle_{{\mathbb F}_q} \in {\mathrm{PG}}\,(U_0,q^t)\simeq{\mathrm{PG}}\,(r-1,q^t) \mapsto \langle
u,u^{\sigma},\ldots,u^{\sigma^{t-1}} \rangle \cap {\mathrm{PG}}\,(rt-1,q).
\]
Throughout this paper, we shall silently
identify the two aforementioned representations
of Desarguesian spreads.
In particular, a spread element will be regarded indifferently as a
$(t-1)$--subspace of ${\mathrm{PG}}\,(rt-1,q)$ of type
\[\{(\lambda u,\lambda^q
u^q,\ldots,\lambda^{q^{t-1}} u^{q^{t-1}}), \lambda \in {\mathbb F}_{q^t}\} \]
and as its projection $\langle u \rangle \in {\mathrm{PG}}\,(U_0,q^t)$.
Fix now a Desarguesian $(t-1)$--spread $\mathcal S$ of ${\mathrm{PG}}\,(rt-1,q)$ and fix
also a subspace $\Pi$ of ${\mathrm{PG}}\,(rt-1,q)$ of projective dimension $m$.
The set $\Lambda$ of all elements of $\mathcal S$ with non--empty intersection
with $\Pi$ is a \emph{linear set} of rank $m+1$. In other words,
$\Lambda$ may be regarded as the set of all points of
${\mathrm{PG}}\,(r-1,q^t)$ whose coordinates are \emph{defined} by a vector space
$W$ over ${\mathbb F}_q$ of dimension $m+1$.
Linear sets are used for several remarkable constructions in finite
geometry;
see \cite{linear_set} for a survey.
In order to avoid the trivial case $\Lambda=\mathcal S$, we shall assume
$m+1\leq tr-t$. When $m+1=rt-t$ we shall
say that the linear set has \emph{maximum rank}.
Furthermore, as we are interested in \emph{proper} linear sets of
${\mathrm{PG}}\,(r-1,q^t)$, that is linear sets which are not contained in any
hyperplane of ${\mathrm{PG}}\,(r-1,q^t)$, we have
$\langle \Lambda \rangle = {\mathrm{PG}}\,(r-1,q^t) $; hence, $\Lambda$
must contain a frame of ${\mathrm{PG}}\,(r-1,q^t)$ and $m+1 \geq r$.
Throughout this paper
a linear set will always be understood
to have rank $m+1$ with $r \leq m+1 \leq rt-t$.
We point out that,
when regarded point sets of ${\mathrm{PG}}\,(r-1,q^t)$, linear sets provide a
generalization of the notion of subgeometry over ${\mathbb F}_q$.
This is shown by the following theorem.
\begin{theorem}[\cite{lunardon_polverino_04}\label{linearset}]
Take
$r\leq m+1 \leq t(r-1)$ and
let $\Lambda$ be the projection in ${\mathrm{PG}}\,(m,q^t)$
of\- a\- subgeometry $\Theta\cong {\mathrm{PG}}\,(m,q)$
onto a ${\mathrm{PG}}\,(r-1,q^t)$. Then, $\Lambda$ is a ${\mathbb F}_q$--linear set of
${\mathrm{PG}}\,(r-1,q^t)$ of rank $m+1$.
Conversely, when $\Lambda$ is a linear set of
${\mathrm{PG}}\,(r-1,q^t)$
of rank $m+1$, then either $\Lambda$ is a canonical subgeometry of
${\mathrm{PG}}\,(r-1,q^t)$ or there exists a subspace $\Omega \cong {\mathrm{PG}}\,(m-r,q^t)$
\- of\- ${\mathrm{PG}}\,(m,q^t)$\- disjoint from ${\mathrm{PG}}\,(r-1,q^t)$ and a subgeometry
$\Theta\simeq{\mathrm{PG}}\,(m,q)$ disjoint from $\Omega$ such that $\Lambda$ is the
projection of $\Theta$ from $\Omega$ on ${\mathrm{PG}}\,(r-1,q^t)$.
\end{theorem}
In particular,
when $m+1=r$, we have $\Lambda \cong {\mathrm{PG}}\,(r-1,q)$ and this is
the unique linear set of rank $r$, up to
projective equivalence. When $m+1 > r$, there are several
non--equivalent linear set of any given rank; they
do not even have the same number of points. As $r$ and $t$
grow, the number of non--equivalent linear sets also grows,
so any attempt for classification is
hopeless.
We end this section by showing that a linear set, when considered as a
subset of a Desarguesian spread, is a projection of a family of
maximal subspaces of a suitable Segre variety. We are aware that the
same result appears in the manuscript \cite{LVdV2014}, but we here present
a different and shorter proof which might be of independent interest.
The embedding:
$$
{\mathrm{PG}}\,(V_1,{\mathbb F})\times{\mathrm{PG}}\,(V_2,{\mathbb F})\times \cdots \times {\mathrm{PG}}\,(V_t,{\mathbb F})\rightarrow {\mathrm{PG}}\,(V_1\otimes V_2 \otimes \cdots \otimes V_t,{\mathbb F})\\
$$
$$(v_1,v_2,\ldots,v_t)\mapsto v_1\otimes v_2 \otimes \cdots \otimes v_t
$$
is the so called \emph{Segre embedding} of
${\mathrm{PG}}\,(V_1,{\mathbb F})\times{\mathrm{PG}}\,(V_2,{\mathbb F})\times \cdots \times {\mathrm{PG}}\,(V_t,{\mathbb F})$ in
${\mathrm{PG}}\,(V_1\otimes V_2\otimes\cdots\otimes V_t,{\mathbb F})$. Its image, comprising
the simple tensors
of ${\mathrm{PG}}\,(V_1\otimes V_2 \otimes \cdots \otimes V_t,{\mathbb F})$,
is an algebraic variety: the
\emph{Segre variety}.
Suppose $t=2$ and $\dim V_i=n_i$ for $i=1,2$. Then,
the Segre variety of ${\mathrm{PG}}\,(n_1n_2-1,{\mathbb F})$, say $\Sigma_{n_1n_2}$,
contains two families of maximal subspaces:
$\{\Pi_w, w \in V_1\}$, with $\Pi_w$ the $n_2$--dimensional vector space
$\{w \otimes v, v\in V_2\}$, and $\{\Pi_u, u \in V_2\}$, with $\Pi_u$
the $n_1$--dimensional vector space $\{v \otimes u, v\in V_1\}$. For an
introduction to the study of this
topic see, for instance, \cite[Chapter 25]{Thas}.
A $(t-1)$--regulus of rank $r-1$ of ${\mathrm{PG}}\,(rt-1,q)$ is a collection of
$(t-1)$--dimensional projective subspaces of type $\langle
P,P^{\gamma_1},\ldots,P^{\gamma_{t-1}} \rangle$, where $P \in \Gamma$,
$P^{\gamma_i} \in \Gamma_i$ with $\Gamma,\Gamma_1,\ldots,\Gamma_{t-1}$ being
$(r-1)$--dimensional subspaces spanning ${\mathrm{PG}}\,(rt-1,q)$ and the
collineations $\gamma_i$ defined such that $\gamma_i : \Gamma \rightarrow
\Gamma_i$, $i=1,2,\ldots,t-1$; see \cite{desarg}.
Let now $\Sigma_{rt} \subset {\mathrm{PG}}\,(rt-1,q)$ be
the Segre variety of ${\mathrm{PG}}\,(r-1,q) \times {\mathrm{PG}}\,(t-1,q)$. We
recall the following result.
\begin{theorem}[\cite{desarg}]
\label{t7}
Any $(t-1)$--regulus of rank $r-1$ of a ${\mathrm{PG}}\,(rt-1,q)$ is the system of
maximal subspaces of dimension $t-1$ of the Segre variety $\Sigma_{rt}$ and
conversely.
\end{theorem}
Using theorems \ref{linearset} and \ref{t7}
we can formulate the following geometric description.
\begin{theorem}\label{segre}
By field reduction, either the points of a ${\mathbb F}_q$--linear set $\Lambda$ of
${\mathrm{PG}}\,(r-1,q^t)$ correspond to the system of
maximal subspaces of dimension $t-1$ of the
Segre variety $\Sigma_{rt}$
or there exists a subspace $\Theta
\cong {\mathrm{PG}}\,((m-r+1)t-1,q)$ of ${\mathrm{PG}}\,((m+1)t-1,q)$, disjoint from
\-${\mathrm{PG}}\,(rt-1,q)$ and a Segre variety $\Sigma_{m+1,t}$
also disjoint from $\Theta$ such that the field reduction of the
points of $\Lambda$
corresponds to the projection of the $(t-1)$--maximal
subspaces of $\Sigma_{m+1,t}$ from $\Theta$
on ${\mathrm{PG}}\,(rt-1,q)$.
\end{theorem}
\begin{proof}
Write $m+1=r$; then, $\Lambda\cong {\mathrm{PG}}\,(r-1,q)$. As all the Desarguesian
subgeometries of the same dimension
are projectively equivalent, we can suppose without loss of generality
$\Lambda=\{\langle(x_0,x_1,\ldots,x_{r-1})\rangle_{{\mathbb F}_{q^t}} | x_i \in {\mathbb F}_q\}$.
For any point $\mathbf{x}:=(x_0,x_1,\ldots,x_{r-1})$, the corresponding spread
element in ${\mathrm{PG}}\,(rt-1,q)$
is $\{(\lambda
\mathbf{x},\lambda^q \mathbf{x},\ldots,\lambda^{q^{t-1}}
\mathbf{x}), \lambda \in {\mathbb F}_{q^t}\}$. Let
$(1,\xi_1,\ldots,\xi_{t-1})$ be a basis for ${\mathbb F}_{q^t}$ regarded
as ${\mathbb F}_q$-vector space and
\[ \gamma_i: (\mathbf{x}^{(0)},\mathbf{x}^{(1)},\ldots,\mathbf{x}^{(t-1)})
\mapsto
(\xi_i \mathbf{x}^{(0)},\xi_i^{q}\mathbf{x}^{(1)},\ldots,\xi_i^{q^{t-1}}\mathbf{x}^{(t-1)}).
\]
Observe that
the collineations $\gamma_i$ of ${\mathrm{PG}}\,(rt-1,q^t)$ all fix ${\mathrm{PG}}\,(rt-1,q)$
set-wise; thus, for all $i=1,\ldots,t-1$ they act also as
collineations of ${\mathrm{PG}}\,(rt-1,q)$.
Let $P$ be the point $(\mathbf{x},\mathbf{x},\ldots,\mathbf{x})$, then
\[ \{(\lambda \mathbf{x},\lambda^q \mathbf{x},\ldots,\lambda^{q^{t-1}}
\mathbf{x}), \lambda \in {\mathbb F}_{q^t}\}=\langle P,P^{\gamma_1},\ldots,P^{\gamma_{t-1}}
\rangle_{{\mathbb F}_q}; \]
so, by \cite{desarg}, the linear representation of
a subgeometry is the system of
maximal subspaces of dimension $t-1$ of the Segre variety $\Sigma_{rt}$
If $m+1 > r$, then, by Theorem \ref{linearset} and by the well--known fact
that the subspace spanned by any two elements of a Desarguesian
spread is partitioned by spread elements, we have the statement.
\end{proof}
As a system of maximal subspaces of a Segre variety is always
a partition of the point-set of the variety, when we regard a linear set
$\Lambda$ of
${\mathrm{PG}}\,(r-1,q^t)$ as a set of points of ${\mathrm{PG}}\,(rt-1,q)$, rather than as a
particular collection of $(t-1)$--subspaces, we see
that $\Lambda$ is
either a Segre variety $\Sigma_{rt}$ or, for $m+1>r$ the projection of a Segre
variety $\Sigma_{m+1,t}$ on a ${\mathrm{PG}}\,(rt-1,q)$. We point out that
Segre varieties and their
projection share several combinatorial and geometric properties;
see, for example, \cite{Thas_VMald}.
\section{Representation of linear sets on the Grassmannian}
The image under the Pl\"{u}cker embedding of a
Desarguesian spread $\mathcal S$ of ${\mathrm{PG}}\,(rt-1,q)$
determines the algebraic variety $\mathcal V_{rt}$; this variety
actually lies in a subgeometry ${\mathrm{PG}}\,(r^t-1,q)$;
see
\cite{Segre,LunardonNS,V}.
We briefly recall a few essential properties of
$\mathcal V_{rt}$.
Let $V:=V(rt,q^t)$ and let $\varepsilon_t:G(rt,t)\to{\mathrm{PG}}\,(\bigwedge^t V,q^t)$ be
the usual Pl\"{u}cker embedding of
the $(t-1)$--projective subspaces of ${\mathrm{PG}}\,(rt-1,q^t)$
in ${\mathrm{PG}}\,(\bigwedge^t V,q^t)$. Denote by $\mathcal G^*_{rt,t}$
the image of such embedding.
Recall that the subgeometry ${\mathrm{PG}}\,(rt-1,q)$ is the set of fixed points of
$ \sigma:
(\mathbf{x}^{(0)},\mathbf{x}^{(1)},\ldots,\mathbf{x}^{(t-1)}) \mapsto
(\mathbf{x}^{(t-1)q},\mathbf{x}^{(0)q},\ldots,\mathbf{x}^{(t-2)q})$.
As
${\mathrm{PG}}\,(\binom{rt}{t}-1,q^t)={\mathrm{PG}}\,(\bigwedge^t V,q^t)$ is spanned by its
totally decomposable vectors, that is its tensors of
rank $1$, we can define a collineation
$\sigma^*$ of ${\mathrm{PG}}\,(\bigwedge^t V,q^t)$ as
\[
\sigma^*: v_0 \wedge v_1 \wedge \cdots \wedge v_{t-1} \mapsto
v_0^{\sigma} \wedge v_1^{\sigma} \wedge \cdots \wedge v_{t-1}^{\sigma}.
\]
The collineation $\sigma^*$ turns out to be a ${\mathbb F}_q$--linear collineation
of order $t$ of
${\mathrm{PG}}\,(\binom{rt}{t}-1,q^t)$; hence,
the set of its fixed points is a subgeometry ${\mathrm{PG}}\,(\binom{rt}{t}-1,q)$.
By Lemma \ref{subspace}, a subspace of ${\mathrm{PG}}\,(rt-1,q^t)$ meets
${\mathrm{PG}}\,(rt-1,q)$ in a subspace of the
same dimension if, and only if, it is fixed set-wise by $\sigma$. Clearly, any subspace of ${\mathrm{PG}}\,(rt-1,q)$ is contained
in exactly one subspace of ${\mathrm{PG}}\,(rt-1,q^t)$ of the same
dimension. Thus, the Grassmannian of the
$(t-1)$--subspaces of ${\mathrm{PG}}\,(rt-1,q)$, say $\mathcal G_{rt,t}$, can be
obtained as the intersection $\mathcal G_{rt,t}=\mathcal G^*_{rt,t}(V)\cap{\mathrm{Fix}\,}(\sigma^*)$.
Recall now the decomposition
$V=U_0 \oplus U_1 \oplus \cdots \oplus U_{t-1}$ and let
$V^{\otimes t}:=\displaystyle \underbrace{V \otimes V \otimes \cdots
\otimes V}_{t\text{ times}}$. Denote by $I$ be the two-sided ideal of the
tensor algebra $\mathcal{T}(V)=\displaystyle\sum_{i=0}^{\infty}
V^{\otimes i}$ generated by $\{v \otimes v, v \in V\}$. As
$\bigwedge^t V=\frac{V^{\otimes t}}{V^{\otimes t} \cap I}$ and
$u_0\otimes u_1\otimes\cdots\otimes u_{t-1}\not\in I$ when
$u_i\in U_i$ and $u_i\neq 0$,
we can identify (with a slight abuse of notation) the element
$u_0\otimes u_1
\otimes \cdots \otimes u_{t-1}$ with
$u_0\wedge u_1 \wedge \cdots \wedge u_{t-1}$.
In particular,
we shall regard
$U_0 \otimes U_1 \otimes \cdots \otimes U_{t-1}$,
as a subspace of $\bigwedge^t V$,
write $\varepsilon_t(\Pi^*_{u})=u\otimes
u^{\sigma}\otimes \cdots \otimes u^{\sigma^{t-1}}$ and regard
$\mathcal V_{rt}$ as a subvariety of $\mathcal G_{rt,t}$.
Let now $\Sigma$
be the Segre variety of ${\mathrm{PG}}\,(r^t-1,q^t)$ consisting of the simple tensors
of $U_0\otimes U_1 \otimes \cdots \otimes U_{t-1}$,
and denote by $\sigma^{\dagger}$ the ${\mathbb F}_q$--linear collineation induced
by $\sigma$ on ${\mathrm{PG}}\,(U_0\otimes U_1 \otimes \cdots \otimes U_{t-1},q^t)$;
in particular,
$\sigma^{\dagger}(u_0\otimes u_1 \otimes\cdots \otimes u_{t-1})= u_{t-1}^q\otimes
u_0^{q} \otimes\cdots \otimes u_{t-2}^q $ and
$\mathcal V_{rt}=\Sigma \cap{\mathrm{Fix}\,}(\sigma^{\dagger})$.
Actually, $\mathcal V_{rt}$ is also as the image of the map
\[ \alpha:(x_0,\ldots,x_{r-1})\in{\mathrm{PG}}\,(r-1,q^t)\mapsto
(\prod_{i=0}^{t-1} x_{f(i)}^{q^i})_{f \in \mathfrak F}\in {\mathrm{PG}}\,(r^t-1,q) \subset {\mathrm{PG}}\,(r^t-1,q) \]
where $\mathfrak F=\{ f: \{0,\ldots, t-1\}\to\{0,\ldots,r-1\} \}$.
Here, $\alpha$ is the map that makes the following diagram commute:
$$\begin{array}{ccc}
& \alpha & \\
\mathrm{PG}(r-1,q^t) & \longrightarrow & \mathrm{PG}(r^t-1,q) \\
& & \\
\text{field reduction} \searrow & & \nearrow \varepsilon_t \\
& \mathrm{PG}(rt-1,q) & \\
& \mathcal S=\text{ Desarguesian Spread} & \\
\end{array}
$$
Let now $\Sigma_{rt}$ be the Segre embedding
of ${\mathrm{PG}}\,(r-1,{\mathbb F}_q) \times {\mathrm{PG}}\,(t-1,{\mathbb F}_q)$.
It is well known that the Pl\"{u}cker embedding of
a family of maximal subspaces of dimension $t-1$ of $\Sigma_{rt}$
is a Veronese variety of dimension $r-1$ and degree $t$; see, for
instance, \cite[Exercise 9.23]{H}.
By Theorem \ref{segre}, the field reduction of a
subgeometry ${\mathrm{PG}}\,(r-1,q)$ of ${\mathrm{PG}}\,(r-1,q^t)$ consists of the family of
maximal subspaces of dimension $t-1$ of $\Sigma_{rt}$. Up to
isomorphism, we can indeed assume ${\mathrm{PG}}\,(r-1,q)=\{(x_0,x_1,\ldots,x_{r-1}), x_i
\in{\mathbb F}_q\}$. The image under $\alpha$ of such a set is, clearly, a
Veronese variety of dimension $r-1$ and degree $t$, the complete
intersection of $\mathcal V_{rt}$ with a subspace of dimension
$\binom{r-1+t}{t}-1$.
As a consequence of Theorem
\ref{segre}, the image of a linear set of rank $m+1$ on
$\mathcal V_{rt}$ is the projection of a Veronese variety of dimension $m$
and degree $t$. Hence, the dimension of such a variety is at most $m$.
\begin{lemma}\label{min_m}
A minimal subspace $\Pi$ defining a linear set $\Lambda$ of
${\mathrm{PG}}\,(rt-1,q)$ is spanned by points
$\{P_0,P_1,\ldots P_{m}\}$ such that $\forall i=0,1,\ldots,m$
the spread element containing $P_i$ intersects $\Pi$ only in $P_i$.
\end{lemma}
\begin{proof}
Let $\Pi$ be a minimal defining subspace for $\Lambda$ and
suppose that every spread element intersects $\Pi$ in at least a
line. Consider a hyperplane $\Pi'$ of $\Pi$.
As $\Pi'$ meets each spread element with non--empty intersection
with $\Pi$, we have that $\Pi'$ and $\Pi$
determine the same linear set and $\Pi'<\Pi$ --- a contradiction.
Thus, we can assume that $\Pi$ contains
at least a point $P$ such that the spread element through $P$
intersects $\Pi$ only in $P$. According to the terminology of
\cite{linear_set}, $P$ is a point of the linear set
of \emph{weight $1$}.
Suppose now that $\Pi$ is not spanned by its points of
weight $1$. Then, there is
a hyperplane $\Pi'$ in $\Pi$ containing
all of these points. A spread element either intersects $\Pi$ in only
one point $P$, hence $P \in \Pi'$, or it intersects $\Pi$ at least a
line;
thus it must intersect
also $\Pi'$. It follows $\Pi'$ and $\Pi$ determine the same linear set
and $\Pi'<\Pi$, contradicting the minimality of $\Pi$ again.
\end{proof}
From now on, when we say that a linear set $\Lambda$ has rank $m+1$, we suppose
that $m$ is the minimum possible; in particular
the defining subspace of $\Lambda$
is taken to be of the type of Lemma \ref{min_m}.
\begin{proposition}
The image of a linear set $\Lambda$ of ${\mathrm{PG}}\,(rt-1,q)$ of rank $m+1$ on
the Grassmannian, hence on $\mathcal V_{rt}$, is an algebraic variety of
dimension $m$, the projection of a Veronese variety of dimension $m$
and degree $t$.
\end{proposition}
\begin{proof}
By Theorem \ref{segre} and the above remarks,
the image of $\Lambda$, say $\mathcal V$, is the
projection of a Veronese variety of dimension $m$.
Thus, its dimension
is at most $m$. Let $\Pi=\langle P_0,P_1,\ldots,P_m\rangle$ be a
subspace determining $\Lambda$ and suppose that
each $P_i$ is of weight $1$.
Write $\Pi_i=\langle P_0,\ldots,P_i\rangle$ and let
$\Lambda_i$ be the linear set determined by $\Pi_i$, with
corresponding image $\mathcal V_i$. Then we
have $\mathcal V_0 \subsetneq \mathcal V_1 \subsetneq \cdots \subsetneq \mathcal V_{m-1}
\subsetneq \mathcal V$. Hence, the dimension of $\mathcal V$ is $m$.
\end{proof}
It has been shown in \cite{V},
that the image of a linear set of a ${\mathrm{PG}}\,(1,q^t)$
is a linear section of $\mathcal V_{2t}$. We can now generalize this result.
\begin{theorem}
\label{t5}
The image of a linear set $\Lambda$
of rank $m+1$ is the intersection of $\mathcal V_{rt}$
with a linear subspace of codimension
at most $\binom{rt-m-1}{t}$.
In particular, this image is the intersection of the images of
$\binom{rt-m-1}{t}$ linear sets of maximum rank.
\end{theorem}
\begin{proof}
Let $\Pi={\mathrm{PG}}\,(W,q)$ be a defining subspace of ${\mathrm{PG}}\,(rt-1,q)$ for
$\Lambda$.
Write $\Omega=\Omega(W)$
for the Schubert variety that is the Pl\"{u}cker
embedding of the $t$-subspaces with non--trivial intersection with
$W$. Then, the image of the linear set on $\mathcal V_{rt}$ is $\Omega
\cap \mathcal V_{rt}$ and $\Omega$ is the complete intersection of the
Grassmannian with a subspace of codimension $\binom{rt-m-1}{t}$.
The statement now follows from Proposition \ref{kernel}.
\end{proof}
We now want to provide some insight on the space of all linear
equations vanishing on $\mathcal V_{rt} \cap \Omega$.
Obviously, any subspace ${\mathrm{PG}}\,(m,q)$ of ${\mathrm{PG}}\,(rt-1,q) \subset {\mathrm{PG}}\,(rt-1,q^t)$
is determined by $n=rt-1-m$ independent ${\mathbb F}_q$--linear equations. These can
always be
taken of the form
\begin{equation}
\label{eqns}
\mathrm{Tr}\,(\displaystyle\sum_{i=0}^{r-1}a_{ji}x_i)=0,\qquad
j=1,2,\ldots,n,
\end{equation}
where $\mathrm{Tr}\,:{\mathbb F}_{q^t}\to {\mathbb F}_q$ is
the usual trace function.
A spread element has non--empty
intersection with the ${\mathrm{PG}}\,(m,q)$ given by the equations in
\eqref{eqns} if, and only if, there exists a non--zero
$\lambda \in {\mathbb F}_{q^t}$ such that
\[
\mathrm{Tr}\,((\displaystyle\sum_{i=0}^{r-1}a_{ji}x_i)\lambda)=0 \qquad
j=1,2,\ldots,n.
\]
In other words, this is the same as to require that
the $(rt-m-1)\times t$ matrix
$$
M=\begin{pmatrix}
\displaystyle\sum_{i=0}^{r-1}a_{1i}x_i & (\displaystyle\sum_{i=0}^{r-1}a_{1i}x_i)^q & \cdots & (\displaystyle\sum_{i=0}^{r-1}a_{1i}x_i)^{q^{t-1}} \\
\displaystyle\sum_{i=0}^{r-1}a_{2i}x_i & (\displaystyle\sum_{i=0}^{r-1}a_{2i}x_i)^q & \cdots & (\displaystyle\sum_{i=0}^{r-1}a_{2i}x_i)^{q^{t-1}} \\
\cdots & \cdots & \cdots & \cdots \\
\displaystyle\sum_{i=0}^{r-1}a_{ni}x_i & (\displaystyle\sum_{i=0}^{r-1}a_{ni}x_i)^q & \cdots & (\displaystyle\sum_{i=0}^{r-1}a_{ni}x_i)^{q^{t-1}}
\end{pmatrix}
$$
cannot have full rank; thus, each of its minors of order $t$
must be singular.
This condition corresponds to
a set of $\binom{rt-m-1}{t}$ equations,
each of them determining a hyperplane section of $\mathcal V_{rt}$.
We remark that, as we expect from Proposition \ref{kernel},
every set of $t$ equations in \rif{eqns} determines a $(rt-t-1)$--dimensional
subspace containing ${\mathrm{PG}}\,(m,q)$, hence a linear set of maximum
rank containing the given one.
Clearly, not all of the equations obtained above are always
linearly independent of $\mathcal V_{rt}$.
For instance, if there were a minor $M_0$ of order
$t-1$ in $M$ which is non--singular for any
choice of $x_i\neq 0$, then $rt-m-t$ equations would
suffice.
The rest of this paper is devoted to investigate
the dimension the space of the linear equations
vanishing on the image of a linear set on $\mathcal V_{rt}$.
As we have already remarked, for any fixed rank $m+1 >r$,
there are many non--equivalent linear sets; here we propose
an unifying approach for linear sets of the same rank.
Let $\Pi={\mathrm{PG}}\,(W,q)$ be a $m$--subspace defining a linear set of ${\mathrm{PG}}\,(rt-1,q)$, ${\mathrm{PG}}\,(W^*,q^t)$ be the $m$--dimensional projective
subspace of ${\mathrm{PG}}\,(rt-1,q^t)$ such that
${\mathrm{PG}}\,(W^*,q^t) \cap {\mathrm{PG}}\,(rt-1,q)=\Pi$, and
$\Omega^*=\Omega(W^*) \subset \mathcal G_{rt\,t}^*$ be
the Schubert variety of the $t$--subspaces with non--trivial
intersection with $W^*$.
Let also $\Sigma$ be the Segre variety of the simple tensors
of $U_0\otimes U_1 \otimes \cdots \otimes U_{t-1}$; recall that we
can identify $\Sigma$ with the set of
$\{u_0 \wedge u_1 \wedge \cdots \wedge u_{t-1}, u_i \in U_i\}$ in
$\bigwedge V^t$.
The lifting $\sigma^*$ of the ${\mathbb F}_q$--linear collineation $\sigma$ to
${\mathrm{PG}}\,(\binom{rt}{t}-1,q^t)$ acts as $\sigma^*(v_1 \wedge v_2 \wedge
\cdots \wedge v_t)=v_1^{\sigma} \wedge v_2^{\sigma} \wedge \cdots
\wedge v_t^{\sigma}$.
As $\sigma$ permutes the $U_i$'s,
$\sigma^*$ fixes $\Sigma$ set-wise.
Since $W^*$ is also fixed set-wise by $\sigma$, see Lemma \ref{subspace}, we
see
that $\Omega^*$ is set-wise fixed by $\sigma^*$.
Lemma \ref{subspace} guarantees
$\dim_{{\mathbb F}_{q}}
\Omega \cap \mathcal V_{rt}=\dim_{{\mathbb F}_{q^t}} \Omega^* \cap \Sigma$, hence we shall
determine
$\dim_{{\mathbb F}_{q^t}} \Omega^* \cap \Sigma$.
As there exists an embedding
$\phi:U\otimes U^{\sigma} \otimes \cdots \otimes U^{\sigma^{t-1}}\to\bigwedge^t V$,
there is also a canonical projection
$\phi':(\bigwedge^t V)'\rightarrow (U\otimes U^{\sigma} \otimes \cdots \otimes
U^{\sigma^{t-1}})'$, where $(\bigwedge^t V)'$ and
$(U\otimes U^{\sigma} \otimes \cdots \otimes
U^{\sigma^{t-1}})'$ are the duals of respectively $\bigwedge^t V$ and
$U\otimes U^{\sigma} \otimes \cdots \otimes
U^{\sigma^{t-1}}$.
Let $\mathcal{F}$ be the subspace of $(\bigwedge^t V)'$ consisting
of the linear
functions vanishing on $\Omega^*$, and let $\phi'_/$ be the restriction of
$\phi'$ to $\mathcal{F}$. We are interested in the dimension of the
image of $\phi'_/$. The nucleus of $\phi_/'$ consists of the $t$--linear
alternating forms $f$ such that $\ker f$ contains $W^*$ and
$f(u_0,u_1,\ldots,u_{t-1})=0$ for all $u_i \in U_i$. Such a space is
isomorphic to the space of the $t$--linear forms $\overline{f}$
defined on a subspace $W^{\natural}$ complement of $W^*$ in $V$, with
$\overline{f}(\overline{u_1},\overline{u_2},\ldots,\overline{u_t})=0$
for all $\overline{u_i} \in \overline{U_i}$, where $\overline{U_i}$ is the
projection of $U_i$ on $W^{\natural}$ from $W^*$.
Observe that $\dim
\overline{U_i}=\dim \langle U_i,W^*\rangle\cap W^{\natural}=\dim\langle U_i,W^*\rangle+\dim W^{\natural}-\dim\langle U_i,W^*,W^{\natural}\rangle= \dim U_i-\dim (U_i\cap W^*)$, so $W^{*\sigma}=W^*$ and
$U_{i+1}=U_i^{\sigma}$ imply that $\dim \overline{U_i}=\dim
\overline{U_0}$, for all $i=1,\ldots,t-1$.
\begin{proposition}\label{projection2}
We have $\dim U_i \cap W^*= h>0$ if, and only if,
the linear set $\Lambda$ contains a
${\mathbb F}_{q^t}$--projective subspace of dimension $h-1$. If the linear
set is proper, that is it spans\- ${\mathrm{PG}}\,(r-1,q^t)$\- but\- it is not
${\mathrm{PG}}\,(r-1,q^t)$, this can occur only for $r\geq 3$.
Furthermore, $h \leq \frac{m+1-r}{t-1}$ in general and
$h = m+1-r$ if
$t=2$.
\end{proposition}
\begin{proof}
A proper linear set $\Lambda$, when considered as a subset of ${\mathrm{PG}}\,(r-1,q^t)$,
spans the whole projective space; hence, the projection of $\Pi=PG(W,q)$ on
${\mathrm{PG}}\,(U_0,q^t)={\mathrm{PG}}\,(r-1,q^t)$ necessarily
spans ${\mathrm{PG}}\,(U_0,q^t)$. It follows
that the projection of ${\mathrm{PG}}\,(W^*,q^t)$ also spans ${\mathrm{PG}}\,(U_0,q^t)$.
For $t=2$,
this implies that
$\dim U_1 \cap W^*=m+1-r$ and $m+1-r>0$ can occur only if $r \geq 3$,
since $r \leq m+1 \leq t(r-1)$.
Suppose now $t>2$ and
let $Z=U_i\cap W^*$; then,
$\langle Z^{\sigma^i}, i=0,\ldots,t-1\rangle \subseteq W^*$.
For any $P\in{\mathrm{PG}}\,(Z,q^t)$,
the projective $(t-1)$--space
$\langle P,P^{\sigma},\ldots,P^{\sigma^{t-1}}\rangle \cap {\mathrm{PG}}\,(rt-1,q)$
is a spread element completely contained in ${\mathrm{PG}}\,(W,q)$. In particular,
${\mathrm{PG}}\,(W,q)$
contains a subspace of dimension $ht-1$ completely partitioned
by spread elements. Thus there exists
a projective subspace ${\mathrm{PG}}\,(h-1,q^t)$ completely contained in
the linear set $\Lambda$. Write $m+1=ht+k$ and let $W^*_1$ be a subspace of
dimension $k$ disjoint from $\langle
Z^{\sigma^i},i=0,\ldots,t-1\rangle \subseteq W^*$. Then $\Lambda$ is a
cone with vertex a ${\mathrm{PG}}\,(h-1,q^t)$ and base
$\Lambda_1$, with $\Lambda_1$ the
linear set induced by $W_1:=W^*_1 \cap {\mathrm{PG}}\,(rt-1,q)$.
In order to have a proper linear set,
we need $\dim \langle \Lambda_1 \rangle =r-h$ and $r-h >0$, so
$k \geq r-h$; hence, $ht \leq m+1-r+h$. Since $m+1 \leq rt-t$,
we have $h \leq \frac{m+1-r}{t-1}$. We can have $h >0$
only if $m+1 \geq t-1+r$, but we also have $m+1 \leq rt-t$, hence we get $rt-t \geq t-1+r$ and so $r \geq 3$.
\end{proof}
\begin{theorem}
Let $c:=\dim \overline{U_i}$. The map $\phi_/$ is injective
if, and only if, $m+1 > rt-t-c$. This is always the case for $t=2$,
$(r,t)=(2,3)$ and for $t \geq 3$ with $m+1 > tr-t-1-\frac{2}{t-2}$.
\end{theorem}
\begin{proof}
The kernel of $\phi_/$ is
the space of
the alternating $t$--linear forms defined on the vector space $W^{\natural}$ of
dimension $rt-m-1$ and such that $f(u_0,u_1,\ldots,u_{t-1})=0$
$\forall u_i \in \overline{U_i}$ or, equivalently,
the space of of the linear forms defined on
$\bigwedge ^t W^{\natural}$ vanishing on
all the points that are the Pl\"{u}cker embedding of
a $t$--space with non-trivial
intersection with each $\overline{U_i}$.
For $t+c > rt-m-1$, every
$t$--subspace intersects every $\overline{U_i}$ non--trivially. This implies $f\equiv 0$ and $\phi$ is injective.
By Proposition \ref{projection2}, $\frac{rt-m-1}{t-1}
\leq c \leq r$ and $c=2r-m-1$ for $t=2$. Hence, when $t=2$, the
condition $m+1 > rt-t-c=2r-2-2r+m+1$ is always fulfilled. Suppose
now $t \geq 3$. By Proposition \ref{projection2},
we have $rt-t-c \leq rt-t-\frac{rt-m-1}{t-1}$; hence,
$m+1 > rt-t-\frac{rt-m-1}{t-1}$ implies $m+1 > rt-t-c$. Thus,
$m+1 > rt-t-\frac{rt-m-1}{t-1}$ if, and only if,
$m+1 > rt-t-1-\frac{2}{t-2}$. When $t=3$, this is equivalent to $m+1 > 3r-6$,
a condition
which is obviously always fulfilled
for $r=2$.
If $t+c \leq rt-m-1$, then the image via the Pl\"{u}cker embedding of the
$t$--spaces with non--trivial intersection with a $\overline{U_i}$ is a
Schubert variety cut on the Grassmannian by
a linear subspace of codimension
$\binom{rt-m-1-c}{t}$; hence, the dimension of the kernel of the map
$\phi_/$ is at least $\binom{rt-m-1-c}{t} \geq 1$.
\end{proof}
\begin{corollary}
Let ${\mathrm{PG}}\,(W,q) \subset {\mathrm{PG}}\,(rt-1,q)$ be the $m$--dimensional
subspace
defining a linear set $\Lambda$ and ${\mathrm{PG}}\,(W^*,q^t)$ be the unique subspace
of ${\mathrm{PG}}\,(rt-1,q^t)$ such that ${\mathrm{PG}}\,(W^*,q^t)\cap {\mathrm{PG}}\,(rt-1,q)={\mathrm{PG}}\,(W,q)$.
Take $W^{\natural}$ such that $V(rt,q^t)=W^*\oplus W^{\natural}$ and
let also $\overline{U_i}$ be the
projection of $U_i$ on $W^{\natural}$. Write $c=\dim\overline{U_i}$.
Then, the image of $\Lambda$ is
the complete intersection of $\mathcal V_{rt}$ with a linear subspace of
codimension $\binom{rt-m-1}{t}$ if, and only if, $m+1 > rt-t-c$. This is
always the case for $t=2$, $(r,t)=(2,3)$ and for $t \geq 3$ and $m+1 >
tr-t-1-\frac{2}{t-2}$. If $m+1 \leq rt-t-c$, then the image of
$\Lambda$ is the complete intersection of $\mathcal V_{rt}$ with a linear
subspace of codimension $\dim \langle u_0 \wedge u_1 \wedge \ldots
\wedge u_{t-1}, u_i \in \overline{U_i}\rangle <\binom{rt-m-1}{t}$.
\end{corollary}
We can provide a complete description for the case $t=3$.
\begin{theorem}
\label{t=3}
Let $t=3$, $r>2$ and $m+1 \leq 3r-3-c$.
Then, the codimension of
$\langle u_0 \wedge u_1 \wedge u_{2}, u_i \in \overline{U_i}\rangle$
in $\bigwedge ^{t}W^{\natural}$ is
$3\binom{3r-m-1-c}{3}$.
\end{theorem}
\begin{proof}
As the projection of ${\mathrm{PG}}\,(W^*,q^t)$ on ${\mathrm{PG}}\,(U_0,q^t)$ spans
${\mathrm{PG}}\,(U_0,q^t)$ we have $\dim \langle U_i,U_j \rangle \cap
W^*=m+1-r$; hence, $\dim \langle \overline{U_i} \overline{U_j} \rangle= 2r-m-1+r=3r-m-1=\dim
W^{\natural}$. Thus, $\langle \overline{U_i} \overline{U_j} \rangle=W^{\natural}$ for any $i\neq j$.
Let $\Omega_i$ be the Schubert variety of the $t$--subspaces with
non--trivial intersection with $\overline{U_i}$ and let $\mathcal F_i$ the space
of the linear functions defined on $\bigwedge ^{t}W^{\natural}$ vanishing on
$\Omega_i$. By a slight abuse of notation, identify the elements of
$\mathcal F_i$ with the corresponding trilinear alternating maps
defined on $W^{\natural} \times W^{\natural} \times W^{\natural}$; the kernel of any element of
$\mathcal F_i$ contains $\overline{U_i}$. Suppose $f_i+f_j=0$ with $f_i \in
\mathcal F_i$, $f_j \in \mathcal F_j$, $i \neq j$. Then, the kernel
of $f_i$ contains $\langle \overline{U_i},\overline{U_j} \rangle=W^{\natural}$, so
$f_i=f_j=0$. Suppose now $f_0+f_1+f_2=0$, with $f_i \in \mathcal F_i
\setminus\{0\}$ and $i=0,1,2$. For every $u_2 \in \overline{U_2}$,
$f_0(\cdot,\cdot,u_2)$ is a bilinear map vanishing on $\langle
\overline{U_0},\overline{U_1} \rangle=W^{\natural}$; hence, it
is identically $0$ and the kernel of $f_0$
would contain $\langle\overline{U_0},\overline{U_2}\rangle=W^{\natural}$. This would
imply $f_0=0$, a contradiction. Hence $\dim \langle \mathcal F_1,\mathcal F_2,\mathcal F_3\rangle=3\dim\mathcal F_i$
\end{proof}
\begin{corollary}
Let $t=3$ and $r>2$. Suppose ${\mathrm{PG}}\,(W,q) \subset {\mathrm{PG}}\,(3r-1,q)$ to be the
$m$--subspace defining the linear set $\Lambda$. Let also
${\mathrm{PG}}\,(W^*,q^3)$ be the
unique subspace of ${\mathrm{PG}}\,(3r-1,q^3)$ such that ${\mathrm{PG}}\,(W^*,q^3)\cap
{\mathrm{PG}}\,(3r-1,q)={\mathrm{PG}}\,(W,q)$ and take $W^{\natural}$ such that $V(3r,q^3)=W^*\oplus W^{\natural}$.
Denote by
$\overline{U_i}$ the projection of $U_i$ on $W^{\natural}$ and write $c=\dim\overline{U_i}$.
Assume also $m+1 \leq 3r-3-c$.
Then, the image of $\Lambda$ is the complete intersection of
$\mathcal V_{r,3}$ with a linear subspace of codimension
$\binom{3r-m-1}{3}-3\binom{3r-m-1-c}{3}$.
\end{corollary}
When $t>3$ and $m+1 \leq 3r-3-\dim\overline{U_i}$, it is not possible, in
general, to provide a formula for the codimension of the image of a
linear set on $\mathcal V_{rt}$ depending only on $m$, as shown by the
following example.
In ${\mathrm{PG}}\,(5,q^4)$, take the linear set $\Lambda_1$ of rank $9$ given by
$\{(x,x^q,y,y^q,y^{q^2},z), x,y \in {\mathbb F}_{q^4}, z \in {\mathbb F}_q\}$. The
subspace $W_1$ of ${\mathrm{PG}}\,(23,q)$ defining $\Lambda_1$ is
\[ \{(x,x^q,y,y^q,y^{q^2},z,x^q,x^{q^2},y^q,y^{q^2},y^{q^3},z,x^{q^2},x^{q^3},y^{q^2},y^{q^3},y,z,x^{q^3},x,y^{q^3},y,y^q,z),
x,y \in {\mathbb F}_{q^4}, z \in {\mathbb F}_q\}; \]
hence, the subspace $W_1^*$ of
rank $9$ of ${\mathrm{PG}}\,(23,q^4)$ containing $W_1$ is
\[ \{(x_1,x_2,x_5,x_6,x_7, x_9, x_2,x_3,x_6,x_7,x_8,x_9,x_3,x_4,x_7,x_8,x_5,x_9,x_4,x_1,x_8,x_5,x_6,x_9),
x_i \in {\mathbb F}_{q^4}\}. \]
A complement is
\[ W_1^{\natural}=\{(0,0,0,0,0,0,y_1,0,y_2,y_3,0,y_4,y_5,0,y_6,y_7,y_8,y_9,y_{10},y_{11},y_{12},y_{13},y_{14},y_{15}), y_i \in {\mathbb F}_{q^4}\}. \]
Let $\overline{U_i}$ be the projection of
$U_i$ on $W_1^{\natural}$. By a straightforward calculation, we get $c=\dim
\overline{U_i}=6$, $\dim \overline{U_0}\cap \overline{U_1}=\overline{U_0}\cap
\overline{U_3}=1$ and $\overline{U_0}\cap \overline{U_2}=0$. Then, the
number of equations defining the image of $\Lambda_1$ on $\mathcal V_{6,4}$
is
$\binom{rt-m-1}{t}-4\binom{rt-m-1-c}{t}+4\binom{rt-m-1-2c+1}{t}=865$.
Consider now the following linear set $\Lambda_2$ of the same rank:
$\{(x,y,y^q,z,z^q,z^{q^2}),x\in {\mathbb F}_{q^2}, y \in {\mathbb F}_{q^4}|\mathrm{Tr}\,(y)=0,z
\in {\mathbb F}_{q^4}\}$, where $\mathrm{Tr}\,:{\mathbb F}_{q^4}\rightarrow {\mathbb F}_q$
is the trace function.
In ${\mathrm{PG}}\,(23,q)$, we have
\[ \{(x,y,y^q,z,z^q,z^{q^2},x^q,y^q,y^{q^2},z^q,z^{q^2},z^{q^3},x,y^{q^2},-y-y^q-y^{q^2},z^{q^2},z^{q^3},z,x^q,-y-y^q-y^{q^2},y,z^{q^3},z,z^q)\}; \]
hence in ${\mathrm{PG}}\,(23,q^4)$ we get
$W_2^*=\{(x_1,x_3,x_4,x_6,x_7,x_8,x_2,x_4,x_5,x_7,x-8,x_9,x_1,x_5,-x_3-x_4-x_5,x_8,
x_9,x_6,x_2,-x_3-x_4-x_5,x_3,x_9,x_6,x_7),x_i\in {\mathbb F}_{q^4}\}$.
A complement is
\[ W_2^{\natural}=\{(0,0,0,0,0,0,0,y_1,0,y_2,y_3,0,y_4,y_5,y_6,y_7,y_8,y_9,y_{10},y_{11},y_{12},y_{13},y_{14},y_{15}),
y_i \in {\mathbb F}_{q^4}\}. \]
We see that $c=\dim \overline{U_i}=6$, $\dim
\overline{U_0}\cap \overline{U_1}=\overline{U_0}\cap \overline{U_3}=0$
and $\overline{U_0}\cap \overline{U_2}=1$. Thus, the number of equation
defining the image of $\Lambda_2$ on $\mathcal V_{6,4}$ is
$\binom{rt-m-1}{t}-4\binom{rt-m-1-c}{t}+2\binom{rt-m-1-2c+1}{t}=863\neq865$.
{\bf Remark}. Even if it is not possible to provide a formula for the
codimension of the image of a linear set on $\mathcal V_{rt}$ depending only
on $m$ for $t>3$ and $m+1 \leq 3r-3-\dim\overline{U_i}$, the above arguments show
a possible way to actually determine its value on a case--by--case basis,
as this codimension is, in general, the same as $\dim \langle u_0
\wedge u_1 \wedge \ldots \wedge u_{t-1}, u_i \in
\overline{U_i}\rangle$.
|
\subsection{Discretization}
The provisional solution is found by
discretizing~\eqref{e:picardFull} and~\eqref{e:inextensMethod1} as
\begin{align*}
{\mathbf{x}}_{j}^{n+1} = {\mathbf{x}}_{j}^{n} + \Delta t_{n}
\left({\mathbf{v}}_{\infty}({\mathbf{x}}^{n}_{j}) +
\sum_{k=1}^{M} \SS({\mathbf{x}}^{n}_{j},{\mathbf{x}}^{n}_{k})(
-{\mathcal{B}}({\mathbf{x}}^{n}_{k})({\mathbf{x}}^{n+1}_{k}) +
{\mathcal{T}}({\mathbf{x}}^{n}_{k})(\sigma^{n+1}_{k}))\right),
\end{align*}
and
\begin{align*}
\mathrm{Div}({\mathbf{x}}^{n}_{j})\left(\sum_{k=1}^{M} \SS({\mathbf{x}}^{n}_{j},{\mathbf{x}}^{n}_{k})(
-{\mathcal{B}}({\mathbf{x}}^{n}_{k})({\mathbf{x}}^{n+1}_{k}) +
{\mathcal{T}}({\mathbf{x}}^{n}_{k})(\sigma^{n+1}_{k}))\right)=0.
\end{align*}
The SDC updates~\eqref{e:sdcUpdateFull}
and~\eqref{e:inextensUpdateMethod1} are discretized as
\begin{align*}
{\mathbf{e}}^{n+1}_{{\mathbf{x}}_{j}} = {\mathbf{e}}^{n}_{{\mathbf{x}}_{j}} +
{\mathbf{r}}_{j}^{n+1} - {\mathbf{r}}_{j}^{n}
+\Delta t_{n} \left(\sum_{k=1}^{M}
\SS(\txx{j}^{n+1},\txx{k}^{n+1})\left(
-{\mathcal{B}}(\txx{k}^{n+1})({\mathbf{e}}^{n+1}_{{\mathbf{x}}_{k}}) +
{\mathcal{T}}(\txx{k}^{n+1})(\delta^{n+1}_{\sigma_{k}})
\right)\right),
\end{align*}
and
\begin{align*}
\mathrm{Div}(\txx{j}^{n+1})&\left(\sum_{k=1}^{M}
\SS(\txx{j}^{n+1},\txx{k}^{n+1})\left(
-{\mathcal{B}}(\txx{k}^{n+1})({\mathbf{e}}^{n+1}_{{\mathbf{x}}_{k}}) +
{\mathcal{T}}(\txx{k}^{n+1})(\delta^{n+1}_{\sigma_{k}})
\right)\right) \\
&=-\mathrm{Div}(\txx{j}^{n+1})\left({\mathbf{v}}_{\infty}(\txx{j}^{n+1})
+\sum_{k=1}^{M} \SS(\txx{j}^{n+1},\txx{k}^{n+1})\left(
-{\mathcal{B}}(\txx{k}^{n+1})(\txx{k}^{n+1}) +
{\mathcal{T}}(\txx{k}^{n+1})(\tsigma{k}^{n+1})\right)\right).
\end{align*}
With this discretization of~\eqref{e:inextensUpdateMethod1}, we see that
if ${\mathbf{e}}^{n}_{{\mathbf{x}}_{k}}=0$ and $\delta^{n}_{\sigma_{k}}=0$, then
\begin{align*}
\mathrm{Div}(\txx{j}^{n+1})\left({\mathbf{v}}_{\infty}(\txx{j}^{n+1})
+\sum_{k=1}^{M} \SS(\txx{j}^{n+1},\txx{k}^{n+1})\left(
-{\mathcal{B}}(\txx{k}^{n+1})(\txx{k}^{n+1}) +
{\mathcal{T}}(\txx{k}^{n+1})(\tsigma{k}^{n+1})\right)\right) = 0,
\end{align*}
which means that
\begin{align*}
\mathrm{Div}(\txx{j}^{n})\left( \sum_{k=1}^{M}
{\mathbf{v}}(\txx{j}^{n};\txx{k}^{n})\right) = 0.
\end{align*}
In other words, by using the
discretization~\eqref{e:numericSDCImplicit}, the fixed point of the SDC
iteration exactly satisfies the inextensibility condition.
\subsection{Spectral Deferred Correction}
In its original development~\cite{dut:gre:rok2000}, SDC iteratively
constructed a high-order solution of the IVP
\begin{align}
\frac{d{\mathbf{x}}}{dt} = f({\mathbf{x}},t), \quad t \in [0,T],
\quad {\mathbf{x}}(0) = {\mathbf{x}}_{0}.
\label{e:ivp}
\end{align}
While classical deferred correction
methods discretize the time derivative in~\eqref{e:ivp}, SDC uses a
Picard integral to avoid unstable numerical differentiation.
Equation~\eqref{e:ivp} is reformulated as
\begin{align}
{\mathbf{x}}(t) = {\mathbf{x}}_{0} + \int_{0}^{t} f({\mathbf{x}},\tau) d\tau, \quad t \in [0,T].
\label{e:picardIVP}
\end{align}
Any time stepping scheme can be used to generate a provisional solution
$\txx{}(t)$. Given this provisional solution $\txx{}(t)$
of~\eqref{e:picardIVP}, the residual is defined as
\begin{align}
{\mathbf{r}}(t;\txx{}) = {\mathbf{x}}_{0} - \txx{}(t) + \int_{0}^{t} f(\txx{},\tau)
d\tau, \quad t \in [0,T].
\label{e:picardResidual}
\end{align}
The integral in~\eqref{e:picardResidual} is approximated by a
$p^{th}$-order accurate quadrature rule. The error ${\mathbf{e}} = {\mathbf{x}} -
\txx{}$ satisfies
\begin{align}
{\mathbf{e}}(t) = {\mathbf{r}}(t;\txx{}) + \int_{0}^{t}
(f(\txx{} + {\mathbf{e}},\tau) - f(\txx,\tau)) d\tau, \quad
t \in [0,T],
\label{e:picardCorrect}
\end{align}
and an approximation $\tilde{{\mathbf{e}}}$ of ${\mathbf{e}}$ is computed, generally
using the same numerical method used to compute $\txx{}$. Finally, the
provisional solution is updated to $\txx{} + \tilde{{\mathbf{e}}}$ and the
procedure is repeated as many times as desired. The process of forming
the residual ${\mathbf{r}}$, approximating the error ${\mathbf{e}}$, and updating the
provisional solution is referred to as an SDC correction or iteration.
The accuracy of SDC depends on the discretization of~\eqref{e:picardIVP}
and~\eqref{e:picardCorrect} and the accuracy of the quadrature rule
required to evaluate ${\mathbf{r}}(t;\txx{})$. An abstract error analysis for
applying deferred correction methods to the operator equation $F(y)=0$
has been preformed in~\cite{boh:ste1984,lin1980,ske1982,ste1973}.
In~\cite{han:str2011}, this abstract framework was applied
to~\eqref{e:ivp} by defining $F({\mathbf{x}}) = (-{\mathbf{x}}'(t) + f({\mathbf{x}},t), -{\mathbf{x}}(0) +
{\mathbf{x}}_{0})$, and the main result is state in Theorem 4.2. Here we only
summarize the result to avoid introducing additional notation.
\begin{theorem}
\label{thm:convergence}
Let $f \in C^{\infty}$ have bounded derivatives, and ${\mathbf{x}}$ be the unique
solution of~\eqref{e:ivp}. Suppose that the time integrator $\phi$ is
stable. That is, there exists some $S > 0$ which only depends on $f$
and $T$, such that for all ${\mathbf{y}}, {\mathbf{z}} \in {\mathbb{R}}^{m}$,
\begin{align*}
\|{\mathbf{y}} - {\mathbf{z}}\| \leq S\|\phi_{m}({\mathbf{y}}) - \phi_{m}({\mathbf{z}})\|.
\end{align*}
Moreover, suppose that $\phi_{m}$ is of order $k$ meaning that
$\phi_{m}({\mathbf{x}}) = \mathcal{O}(\Delta t^{k})$, where $\Delta t = T/m$.
Suppose that a numerical solution $\txx{}$ of ${\mathbf{x}}$ satisfies
\begin{align*}
\|\txx{}(T) - {\mathbf{x}}(T)\| \leq C \Delta t^{\ell},
\end{align*}
where $C$ depends only on the derivatives of $f$ and on $T$, but not on
$m$. Assuming the residual ${\mathbf{r}}$ is computed exactly, if $\tee{}$ is
formed with the order $p$ time integrator $\phi$ to approximate ${\mathbf{e}}$,
then
\begin{align*}
\|\tee{}(T) + \txx{}(T) - {\mathbf{x}}(T)\| \leq C \Delta t^{\ell+k}.
\end{align*}
However, since ${\mathbf{r}}$ is approximated with a $p^{th}$-order quadrature,
the asymptotic error is
\begin{align*}
\|\tee{}(T) + \txx{}(T) - {\mathbf{x}}(T)\| = \mathcal{O}(\Delta t^{\min(\ell+k,p)}).
\end{align*}
\end{theorem}
Theorem~\ref{thm:convergence} tells us that by estimating the error
${\mathbf{e}}$ with a first-order method, which is the only order we consider in
the SDC framework, the order of accuracy is increased by one, with the
constraint that this convergence is limited by the accuracy of the
quadrature rule for approximating~\eqref{e:picardResidual}. However,
the theorem states nothing about the stability of SDC.
In~\cite{dut:gre:rok2000}, the authors consider three discretizations
of~\eqref{e:picardIVP} and~\eqref{e:picardCorrect}: fully explicit,
fully implicit, and linear combinations of the two. We do not consider
explicit methods since the governing equations are stiff. The other two
methods could be used for vesicle suspensions, but both would require
solving non-linear equations.
We have successfully used a variant of IMEX methods for vesicle
suspensions~\cite{qua:bir2013b,rah:vee:bir2010,vee:gue:zor:bir2009}, and
we couple these methods with SDC in this work. IMEX
methods~\cite{asc:ruu:wet1995} are a family of time integrators that
treat some terms (generally, linear) implicitly and other terms
explicitly. IMEX methods for additive splittings $\dot{{\mathbf{x}}}(t) =
F_{\mathrm{EXPLICIT}}({\mathbf{x}},t) + F_{\mathrm{IMPLICIT}}({\mathbf{x}},t) =
F_{E}({\mathbf{x}},t) + F_{I}({\mathbf{x}},t)$ of~\eqref{e:ivp} were first applied to SDC
by Minion in~\cite{min2003}. We summarize the numerical results
from~\cite{min2003} since their behaviour resembles results that we will
observe. First, the Van der Pol oscillator is considered in a
non-stiff, mildly stiff, and stiff regime, and the number of SDC
iterations ranges from two to six. In the non-stiff regime, the error
behaves according to Theorem~\ref{thm:convergence}. In the mildly stiff
case, the correct asymptotic result is observed, but not until $\Delta
t$ is much smaller. In the stiff regime, the convergence behaviour
differs considerably from the formal error. This behaviour is
attributed to order reduction which is further analyzed. The author
proceeds to claim that ``the correct asymptotic convergence rates would
be observed given sufficiently small $\Delta t$; however, this is not
the relevant issue in most applications". Two other examples considered
are a system of differential equations resulting from a spatial
discretization of an advection-diffusion equation, and Burgers equation.
Convergence rates up to fifth-order are achieved. As before, as the
system becomes stiffer, smaller time steps are required before the
expected convergence behaviour is observed.
\subsection{SDC for Vesicle Suspensions}
Let $\{\gamma_{j}\}_{j=1}^{M}$ be a collection of vesicles
parameterized by ${\mathbf{x}}_{j}$ and with tension $\sigma_{j}$
(Figure~\ref{f:schematics}). We use the integro-differential equation
from~\cite{vee:gue:zor:bir2009} to balance the bending and tension
forces of the vesicle with the stress jump of the fluid across the
vesicle membrane, and an algebraic constraint to enforce the
inextensibility condition. We start by defining the velocity of
vesicle $j$ due to the hydrodynamic forces from vesicle $k$
\begin{align*}
{\mathbf{v}}({\mathbf{x}}_{j};{\mathbf{x}}_{k}) = {\mathbf{v}}_{\infty}({\mathbf{x}}_{j})\delta_{j,k} + \SS({\mathbf{x}}_{j},{\mathbf{x}}_{k})
(-{\mathcal{B}}({\mathbf{x}}_{k})({\mathbf{x}}_{k}) + {\mathcal{T}}({\mathbf{x}}_{k})(\sigma_{k})),
\end{align*}
where $\delta_{j,k}$ is the Kronecker delta function,
\begin{align*}
&s = \|{\mathbf{x}}'\|, \quad \quad \rho = \|{\mathbf{x}} - {\mathbf{y}}\|, \\
&\SS({\mathbf{x}}_{j},{\mathbf{x}}_{k})({\mathbf{f}}) = \frac{1}{4\pi}\int_{\gamma_{k}} \left(
-\log \rho + \frac{({\mathbf{x}} - {\mathbf{y}}) \otimes ({\mathbf{x}} - {\mathbf{y}})}{\rho^{2}} \right)
{\mathbf{f}}({\mathbf{y}}) ds_{{\mathbf{y}}}, \quad {\mathbf{x}} \in \gamma_{j}, \\
&{\mathcal{B}}({\mathbf{x}})({\mathbf{f}}) = \frac{d^{4}{\mathbf{f}}}{ds^{4}}, \\
&{\mathcal{T}}({\mathbf{x}})(\sigma) =
\frac{d}{ds}\left(\sigma\frac{d{\mathbf{x}}}{ds}\right),
\end{align*}
and ${\mathbf{v}}_{\infty}$ is the background velocity (unconfined flows) or the
velocity due to solid walls (confined flows). In the case of confined
flows, we use the double-layer potential of an unknown density function
${\boldsymbol\eta}$ defined on the boundary of the solid walls. The extra equation
comes from a non-slip boundary condition on the solid walls and the
details are presented in~\cite{rah:vee:bir2010}. We point out that
${\mathbf{v}}$ is not symmetric, meaning that ${\mathbf{v}}({\mathbf{x}}_{j};{\mathbf{x}}_{k}) \neq
{\mathbf{v}}({\mathbf{x}}_{k};{\mathbf{x}}_{j})$ for all $j \neq k$, and $\SS$, ${\mathcal{B}}$, ${\mathcal{T}}$ are
all linear in their second argument.
The notation we are using is chosen so that terms such as
${\mathcal{B}}({\mathbf{x}}^{N})({\mathbf{x}}^{N+1})$, which approximates
${\mathcal{B}}({\mathbf{x}}^{N+1})({\mathbf{x}}^{N+1})$, are understood to mean
\begin{align*}
{\mathcal{B}}({\mathbf{x}}^{N})({\mathbf{x}}^{N+1}) = \frac{d^{4}}{ds^{4}} {\mathbf{x}}^{N+1},
\quad s = \|{\mathbf{x}}'^{N}\|,
\end{align*}
where ${\mathbf{x}}'$ is the derivative of ${\mathbf{x}}$ with respect to its
parameterization variable. The tension $\sigma_{j}$ acts as a Lagrange
multiplier to satisfy the inextensibility constraint
\begin{align}
\mathrm{Div}({\mathbf{x}}_{j})\left(\sum_{k=1}^{M}{\mathbf{v}}({\mathbf{x}}_{j};{\mathbf{x}}_{k})\right) = 0,
\label{e:inextens}
\end{align}
where
\begin{align*}
\mathrm{Div}({\mathbf{x}})({\mathbf{f}}) = \frac{d{\mathbf{x}}}{ds} \cdot \frac{d{\mathbf{f}}}{ds}, \quad s =
\|{\mathbf{x}}'\|,
\end{align*}
which is also linear in its second argument.
Equation~\eqref{e:inextens} can be eliminated using the Schur complement
of the tension to write $\sigma_{j}$ in terms of the positions
${\mathbf{x}}_{k}$, $k=1,\ldots,M$. Then, ${\mathbf{v}}({\mathbf{x}}_{j};{\mathbf{x}}_{k})$ can be written
entirely in terms of ${\mathbf{x}}_{j}$ and ${\mathbf{x}}_{k}$, and the no-slip boundary
condition of vesicle $j$ gives
\begin{align}
\frac{d{\mathbf{x}}_{j}}{dt} = \sum_{k=1}^{M}{\mathbf{v}}({\mathbf{x}}_{j};{\mathbf{x}}_{k}), \quad
j=1,\ldots,M.
\label{e:diffEqn}
\end{align}
The formulation~\eqref{e:diffEqn} is easiest to present our numerical
methods, but we in fact do not eliminate the tension in our
implementation. The resulting changes to the SDC formulation are
presented in Appendix~\ref{a:appendix1}.
\begin{figure}[htps]
\begin{center}
\begin{tabular}{cc}
\ifTikz
\scalebox{0.9}{\input{unboundedGeom.tikz}} &
\scalebox{0.9}{\input{boundedGeom.tikz}}
\fi
\end{tabular}
\end{center}
\mcaption{Two typical vesicle suspensions. Left: $M$ vesicles are
submerged in an unbounded shear flow. Right: $13$ vesicles are in a
bounded domain. In the right configuration, the flow is driven by
Dirichlet boundary conditions on $\Gamma$ (The internal cylinders are
rotating.)}{f:schematics}
\end{figure}
Following the SDC method we reformulate~\eqref{e:diffEqn} as
\begin{align}
{\mathbf{x}}_{j}(t) = {\mathbf{x}}_{j}(0) + \int_{0}^{t}
\sum_{k=1}^{M}{\mathbf{v}}({\mathbf{x}}_{j};{\mathbf{x}}_{k})d\tau,
\quad t \in [0,T].
\label{e:picardEqn}
\end{align}
We form a provisional solution $\txx{}$ at a set of quadrature nodes in
$[0,\Delta t]$ using a method that we describe in the next section
(Section~\ref{s:vesiclePicard}), and then evaluate
\begin{align}
{\mathbf{r}}_{j}(t;\txx{}) = {\mathbf{x}}_{j}(0) - \txx{j}(t) + \int_{0}^{t}
\sum_{k=1}^{M}{\mathbf{v}}(\txx{j};\txx{k})d\tau,
\quad t \in [0,T],
\label{e:picardRes}
\end{align}
with a quadrature rule that we also define in the next section
(Section~\ref{s:vesicleQuadrature}). Then, the error $\exx{j} = {\mathbf{x}}_{j}
- \txx{j}$ satisfies
\begin{align*}
\exx{j}(t) = {\mathbf{x}}_{j}(0) - \txx{j}(t) + \int_{0}^{t}
\sum_{k=1}^{M}{\mathbf{v}}(\txx{j} + \exx{j};\txx{k} + \exx{k})
d\tau, \quad t \in [0,T],
\end{align*}
which we write using the residual ${\mathbf{r}}$ as
\begin{align}
\exx{j}(t) = {\mathbf{r}}_{j}(t;\txx{}) + \int_{0}^{t}
\sum_{k=1}^{M}({\mathbf{v}}(\txx{j}+\exx{j};\txx{k} + \exx{k}) -
{\mathbf{v}}(\txx{j};\txx{k})) d\tau,
\quad t \in [0,T].
\label{e:sdcUpdate}
\end{align}
We define the new provisional solution as $\txx{} +
\tilde{{\mathbf{e}}}_{{\mathbf{x}}}$, where $\tilde{{\mathbf{e}}}_{{\mathbf{x}}}$ is an approximate
solution of~\eqref{e:sdcUpdate}. Again, this procedure of computing the
residual~\eqref{e:picardRes}, numerically solving~\eqref{e:sdcUpdate}
for the error, and updating the provisional solution is what we call an
SDC iteration. Assuming that we take $n_{{\mathrm{sdc}}}$ first-order SDC
iterations and $p$ is the quadrature error for computing the residual
${\mathbf{r}}$, from Theorem~\ref{thm:convergence}, we expect that the asymptotic
rate of convergence is $\min(n_{{\mathrm{sdc}}},p)$.
\subsection{Spatial Discretization}
Following our previous
work~\cite{qua:bir2013b,rah:vee:bir2010,vee:gue:zor:bir2009}, we use a
Lagrangian formulation by marking each vesicle with $N$ tracker
points. Since vesicles are modeled as smooth closed curves, all the
derivatives are computed spectrally using the fast Fourier transform
(FFT). Near-singular integrals are handled using the near-singular
integration scheme outlined in~\cite{qua:bir2013b}. Finally, we use
Alpert's high-order Gauss-trapezoid quadrature rule~\cite{alp1999} with
accuracy $\mathcal{O}(h^{8} \log h)$ to evaluate the single-layer potential,
and the trapezoid rule for the double-layer potential which has
spectral accuracy since its kernel is smooth.
\subsection{Temporal Discretization}
A fully explicit discretization of~\eqref{e:diffEqn} results in a stiff
system for multiple reasons. A stiffness analysis
in~\cite{vee:gue:zor:bir2009} reveals that the leading sources of
stiffness, and the corresponding time step restrictions, are:
\begin{itemize}
\item $\SS({\mathbf{x}}_{j},{\mathbf{x}}_{j}){\mathcal{B}}({\mathbf{x}}_{j})({\mathbf{x}}_{j})$ (self-hydrodynamic
bending force) -- $\Delta t \sim \Delta s^{3}$;
\item $\SS({\mathbf{x}}_{j},{\mathbf{x}}_{j}){\mathcal{T}}({\mathbf{x}}_{j})({\mathbf{x}}_{j})$ (self-hydrodynamic
tension force) -- $\Delta t \sim \Delta s$;
\item $\mathrm{Div}({\mathbf{x}}_{j})({\mathbf{x}}_{j})$ (self-inextensibility force) -- $\Delta
t \sim \Delta s$;
\item $\SS({\mathbf{x}}_{j},{\mathbf{x}}_{k})({\mathcal{B}}({\mathbf{x}}_{k})({\mathbf{x}}_{k}) +
{\mathcal{T}}({\mathbf{x}}_{k})(\sigma_{k}))$, $j \neq k$ (inter-vesicle hydrodynamic
forces) -- depends on the inter-vesicle distance.
\end{itemize}
The leading sources of stiffness result from the intra-vesicle
interactions; but, for concentrated suspensions, inter-vesicle
interactions become significant and introduce stiffness.
To address these multiple sources of stiffness, we use a variant of
IMEX~\cite{asc:ruu:wet1995} time integrators. IMEX methods were
developed for the problem $\dot{{\mathbf{x}}}(t) = F_{E}({\mathbf{x}}) + F_{I}({\mathbf{x}})$,
where $F_{E}$ is non-stiff and is treated explicitly, and $F_{I}$ is
stiff and is treated implicitly. For problems with this additive
splitting, the family of time integrators is
\begin{align*}
\frac{\beta {\mathbf{x}}^{n+1} - {\mathbf{x}}^{0}}{\Delta t} =
F_{E}({\mathbf{x}}^{e}) + F_{I}({\mathbf{x}}^{n+1}),
\end{align*}
where ${\mathbf{x}}^{0}$ and ${\mathbf{x}}^{e}$ are linear combinations of previous time
steps and $\beta > 0$. Unfortunately,~\eqref{e:diffEqn} does not have
an additive splitting between stiff and non-stiff terms. However, we
have observed first- and second-order convergence~\cite{qua:bir2013b}
for the time integrator
\begin{align*}
\frac{\beta {\mathbf{x}}^{n+1}_{j} - {\mathbf{x}}^{0}_{j}}{\Delta t} =
\sum_{k=1}^{M} {\mathbf{v}}({\mathbf{x}}_{j}^{e};{\mathbf{x}}_{k}^{n+1}), \quad j=1,\dots,M,
\end{align*}
where
\begin{align*}
{\mathbf{v}}({\mathbf{x}}_{j}^{e};{\mathbf{x}}_{k}^{n+1}) = \SS({\mathbf{x}}^{e}_{j},{\mathbf{x}}^{e}_{k})
(-{\mathcal{B}}({\mathbf{x}}^{e}_{k})({\mathbf{x}}^{n+1}_{k}) +
{\mathcal{T}}({\mathbf{x}}^{e}_{k})(\sigma^{n+1}_{k})).
\end{align*}
Note that in this formulation, it is the operators involved in ${\mathbf{v}}$,
such as the bending ${\mathcal{B}}({\mathbf{x}}_{k})$ and the single-layer potential
$\SS({\mathbf{x}}_{j},{\mathbf{x}}_{k})$, that are discretized explicitly at ${\mathbf{x}}^{e}$.
In line with our previous work, we use the first-order method given by
$\beta = 1$, ${\mathbf{x}}^{0}={\mathbf{x}}^{n}$, and ${\mathbf{x}}^{e}={\mathbf{x}}^{n}$, and the
second-order backward difference formula (BDF) given by $\beta = 3/2$,
${\mathbf{x}}^{0} = 2{\mathbf{x}}^{n} - 1/2{\mathbf{x}}^{n-1}$, and ${\mathbf{x}}^{e} = 2{\mathbf{x}}^{n} -
{\mathbf{x}}^{n-1}$.
\subsection{Quadrature Formula}
\label{s:vesicleQuadrature}
SDC requires a quadrature formula to approximate~\eqref{e:picardRes},
the residual of the Picard integral. The quadrature rule effects the
accuracy of SDC (Theorem~\ref{thm:convergence}). However, the
stability of SDC is also effected by the quadrature rule, and this has
been investigated in depth in~\cite{bou:lay:min2003}. We do not use
equi-spaced nodes since they are vulnerable to the Runge phenomenon.
In order to achieve the highest possible accuracy, we could use
Gaussian nodes; however, in order to avoid extrapolation, we would like
both endpoints to be quadrature nodes. Therefore, we use Gauss-Lobatto
points, $0 = t_{1} < \cdots < t_{p}=\Delta t$, since they include both
endpoints, have maximal order $2p-3$, and have successfully been used
by other groups~\cite{bou:lay:min2003,bou:min2010,min2003}.
Alternatively, we could use Radau quadrature formula which include the
left endpoint, but not the right.
For computational efficiency, the accuracy of the quadrature formula
should not exceed the expected rate of convergence. This is especially
important in three dimensions, which is not considered in this work,
since multiple variables must be stored at each of the quadrature nodes
for future SDC iterations.
\subsection{Picard Integral Discretization}
\label{s:vesiclePicard}
Equations~\eqref{e:picardEqn} and~\eqref{e:sdcUpdate} have similar
structure, and we take advantage of this structure in their
discretizations. For adaptivity, we want a scheme that easily allows
for variable time step sizes. For this reason, we use only first-order
methods for~\eqref{e:picardEqn} and~\eqref{e:sdcUpdate}. When desired,
we use SDC iterations to increase the accuracy.
The first-order provisional solution is found by
discretizing~\eqref{e:picardEqn} as
\begin{align}
{\mathbf{x}}_{j}^{n+1} = {\mathbf{x}}_{j}^{n} + \Delta t_{n}
\sum_{k=1}^{M} {\mathbf{v}}({\mathbf{x}}_{j}^{n};{\mathbf{x}}_{k}^{n+1}), \quad
n=0,\ldots,p-1,
\label{e:numericProvisionalImplicit}
\end{align}
where $\Delta t_{n} = t_{n+1} - t_{n}$. This is exactly the first-order
time integrator we introduced in~\cite{qua:bir2013b}. As we march in
time with~\eqref{e:numericProvisionalImplicit}, we save the variables
required for near-singular integration for future SDC iterations. Then,
we evaluate the residual~\eqref{e:picardRes} using the Gauss-Lobatto
quadrature rule.
In line with~\cite{min2003}, which considers semi-implicit SDC methods,
we would like to discretize~\eqref{e:sdcUpdate} as
\begin{align*}
{\mathbf{e}}^{n+1}_{{\mathbf{x}}_{j}} = {\mathbf{e}}^{n}_{{\mathbf{x}}_{j}} +
{\mathbf{r}}_{j}^{n+1} - {\mathbf{r}}_{j}^{n} +
\Delta t_{n}\sum_{k=1}^{M}(
{\mathbf{v}}(\txx{j}^{n}+{\mathbf{e}}^{n}_{{\mathbf{x}}_{j}};\txx{k}^{n+1}+{\mathbf{e}}^{n+1}_{{\mathbf{x}}_{k}}) -
{\mathbf{v}}(\txx{j}^{n},\txx{k}^{n+1})).
\end{align*}
The issue with this formulation is that it requires additional storage
and computations to find the velocity due to the vesicle parameterized
by $\txx{k}^{n} + {\mathbf{e}}^{n}_{{\mathbf{x}}_{k}}$. Again, in three dimensions
this restriction is even more prohibitive. We have experimented with
other discretizations of~\eqref{e:sdcUpdate}. The simplest such one is
\begin{align*}
{\mathbf{e}}^{n+1}_{{\mathbf{x}}_{j}} = {\mathbf{e}}^{n}_{{\mathbf{x}}_{j}} +
{\mathbf{r}}_{j}^{n+1} - {\mathbf{r}}_{j}^{n}.
\end{align*}
This discretization is consistent with the governing
equations\footnote{If ${\mathbf{e}}^{n}_{{\mathbf{x}}_{j}}$ converges to 0, then
${\mathbf{r}}_{j}^{n+1}={\mathbf{r}}_{j}^{n}$, and by~\eqref{e:picardRes}, we have
solved~\eqref{e:picardEqn} up to quadrature error.}, but,
experimentally, SDC converges only if a very small time step is used.
To allow for larger time steps, we can include the implicit term in the
discretization
\begin{align*}
{\mathbf{e}}^{n+1}_{{\mathbf{x}}_{j}} = {\mathbf{e}}^{n}_{{\mathbf{x}}_{j}} +
{\mathbf{r}}_{j}^{n+1} - {\mathbf{r}}_{j}^{n} +
\Delta t_{n} \sum_{k=1}^{M}
({\mathbf{v}}(\txx{j}^{n};\txx{k}^{n+1}+{\mathbf{e}}_{{\mathbf{x}}_{k}}^{n+1}) -
{\mathbf{v}}(\txx{j}^{n};\txx{k}^{n+1})),
\end{align*}
where we use a slight abuse of notation by defining
\begin{align*}
{\mathbf{v}}(\txx{j}^{n};\txx{k}^{n+1} + {\mathbf{e}}_{{\mathbf{x}}_{k}}^{n+1}) =
\SS(\txx{j}^{n},\txx{k}^{n})
(-{\mathcal{B}}(\txx{k}^{n})(\txx{k}^{n+1}+{\mathbf{e}}_{{\mathbf{x}}_{k}}^{n+1}) +
{\mathcal{T}}(\txx{k}^{n})(\sigma^{n+1}_{k})).
\end{align*}
(Note how none of the operators depend on the error ${\mathbf{e}}$.)
Appendix~ref{a:appendix1} presents this same discretization without the
abuse of notation. While this discretization allows larger time steps,
it is incompatible with the inextensibility constraint (see
Appendix~\ref{a:appendix1}). A discretization that is is compatible
with the inextensibility constraint is
\begin{align}
{\mathbf{e}}^{n+1}_{{\mathbf{x}}_{j}} = {\mathbf{e}}^{n}_{{\mathbf{x}}_{j}} +
{\mathbf{r}}_{j}^{n+1} - {\mathbf{r}}_{j}^{n} +
\Delta t_{n} \sum_{k=1}^{M}
({\mathbf{v}}(\txx{j}^{n+1};\txx{k}^{n+1}+{\mathbf{e}}_{{\mathbf{x}}_{k}}^{n+1}) -
{\mathbf{v}}(\txx{j}^{n+1};\txx{k}^{n+1})),
\label{e:numericSDCImplicit}
\end{align}
where we are using the same abuse of notation. Note
that~\eqref{e:numericSDCImplicit} only requires evaluating the velocity
field due to the vesicle configuration given by $\txx{k}^{n+1}$. Since
these velocity fields are required to form residual ${\mathbf{r}}$, no additional
velocity fields need to be formed. However, there is a loss in
accuracy, but we have observed experimentally that if only
one or two SDC corrections are used, it does not reduce the rate of
convergence.
To summarize, the main steps for using SDC to solve~\eqref{e:picardEqn}
are
\begin{enumerate}
\item Find a first-order provisional solution $\tilde{{\mathbf{x}}}$
using~\eqref{e:numericProvisionalImplicit}.
\item \label{step:residual} Compute the residual ${\mathbf{r}}$ by
approximating the integral in~\eqref{e:picardRes} with the
Gauss-Lobatto quadrature rule.
\item Use~\eqref{e:numericSDCImplicit} to approximate the error
$\tee{}$.
\item Define the new provisional solution to be $\txx{} + \tee{}$.
\item Go to step~\ref{step:residual}.
\end{enumerate}
\subsection{Preconditioning}
\label{s:preco}
Equations~\eqref{e:numericProvisionalImplicit}
and~\eqref{e:numericSDCImplicit} are ill-conditioned and require a large
number of GMRES iterations~\cite{vee:gue:zor:bir2009}. To reduce the
cost of the linear solves, we used a block-diagonal preconditioner
in~\cite{qua:bir2013b} which is formed and factorized in matrix form at
each time step. Using this preconditioner, the number of preconditioned
GMRES iterations depends only on the magnitude of the inter-vesicle
interactions, which in turn is a function of the proximity of the
vesicles. For further savings, we freeze and factorize the
preconditioner at the first Gauss-Lobatto point, and this preconditioner
is used for all the subsequent Gauss-Lobatto points and SDC iterates.
By freezing the preconditioner, we significantly reduce the number of
GMRES iterations, and we only require one matrix factorization per time
step, which is the number of factorizations required when we
precondition our time integrators introduced in~\cite{qua:bir2013b}.
\subsection{Complexity Estimates}
Here we summarize the cost of the most expensive algorithms required in
our formulation.
\begin{itemize}
\item \emph{Matrix-vector multiplication:} For unbounded flows, if $M$
vesicles are each discretized with $N$ points, the bending and tension
calculations require $\mathcal{O}(MN\log N)$ operations using the FFT, and the
single-layer potential requires $\mathcal{O}(MN)$ operations using the FMM.
If the solid wall is discretized with $N_{{\mathrm{wall}}}$ points, using the FMM,
the matrix-vector multiplication requires $\mathcal{O}(MN\log N + N_{{\mathrm{wall}}})$
operations.
\item \emph{Computing the residual:}
Given a provisional solution $\tilde{{\mathbf{x}}}$, computing the residual
${\mathbf{r}}$ is equivalent to $p$ matrix-vector multiplications. Therefore,
computing the residual requires $\mathcal{O}(p(MN\log N + N_{{\mathrm{wall}}}))$
operations.
\item \emph{Forming the provisional solution and SDC corrections:}
Equations~\eqref{e:numericProvisionalImplicit}
and~\eqref{e:numericSDCImplicit} require solving the same linear system
(only the right-hand sides are different), and our preconditioner
results in a mesh-independent number of GMRES iterations. Therefore,
if $n_{{\mathrm{gmres}}}$ total iterations are required to find the provisional
solution, then $n_{{\mathrm{sdc}}}$ SDC iterations requires
$\mathcal{O}(n_{{\mathrm{gmres}}}p(n_{{\mathrm{sdc}}}+1)(MN\log N + N_{{\mathrm{wall}}}))$ operations.
\item \emph{Forming the preconditioner:} The preconditioner is computed
and stored in matrix form and requires $\mathcal{O}(MN^{2}\log N)$ operations
per time step by using Fourier differentiation and the FFT. The
preconditioner must be factorized which requires $\mathcal{O}(MN^{3})$
operations. This is computed only once per time step and is reused at
all the additional Gauss-Lobatto quadrature points. In two-dimensions,
this cost is acceptable since the number of unknowns on each vesicle is
relatively small. However, in three-dimensions, this cost is
unacceptable and different preconditioners will need to be constructed.
\end{itemize}
\subsection{Relaxation}
We consider a single vesicle, initialized as a three-to-one ellipse, in
a Stokes fluid with no background velocity. We discretize the vesicle
with $N=96$ points and the time horizon is $T=2$ which is large enough
that the vesicle comes within 10\% of its steady state solution. We use
$p=5$ Gauss-Lobatto quadrature points so that the quadrature's order of
accuracy (seven in this case) is a few orders larger than the order of
the time integrator (up to five). For this example, there are no calls
to the FMM because the self-interactions are evaluated directly. We
report the errors and CPU timings in
Tables~\ref{t:noSDCrelaxation}--\ref{t:SDC34relaxation}. We observe
that:
\begin{itemize}
\item Each SDC iteration significantly improves the accuracy of the
solution. However, we are unable to achieve third- and higher-order
results. We expect that these convergence rates will be achieved for
smaller values of $\Delta t$, but, at these required values for $\Delta
t$, other sources of error, such as the GMRES tolerance or machine
precision, will dominate. This behaviour is observed by Minion
in~\cite{min2003}.
\item Comparing the left entires of Tables~\ref{t:noSDCrelaxation}
and~\ref{t:SDC12relaxation}, we see that the CPU time increases by more
than four-fold. This is a result of SDC having to form the solution at
the intermediate Gauss-Lobatto points.
\item Comparing the two second-order solvers (BDF and $n_{{\mathrm{sdc}}}=1$), we
can not conclusively pick the faster method. However, simulations using
SDC corrections are compatible with adaptive time stepping while with
BDF, they are not.
\end{itemize}
\begin{table}[htps]
\begin{center}
\begin{tabular}{c|ccc|ccc}
& \multicolumn{3}{c|}{$\boldnsdc{0}$} &
\multicolumn{3}{c}{{\bf BDF}} \\
$m$ & $e_{A}$ & $e_{L}$ & CPU & $e_{A}$ & $e_{L}$ & CPU \\
\hline
$125$ & $3.44$E$-6$ & $3.40$E$-5$ & $1.0$
& $6.13$E$-8$ & $1.56$E$-6$ & $1.1$ \\
$250$ & $1.74$E$-6$ & $1.74$E$-5$ & $2.1$
& $1.40$E$-8$ & $3.57$E$-7$ & $2.1$ \\
$500$ & $8.76$E$-7$ & $8.83$E$-6$ & $4.2$
& $3.10$E$-9$ & $7.84$E$-8$ & $4.2$ \\
$1000$ & $4.40$E$-7$ & $4.44$E$-6$ & $8.4$
& $2.41$E$-10$ & $1.75$E$-8$ & $8.3$
\end{tabular}
\mcaption{The errors in area and length and the CPU time for a single
vesicle in a {\bf relaxation} flow with a {\bf constant} time step size
using $\boldnsdc{0}$ (left) and {\bf BDF} (right). The CPU times for
Tables~\ref{t:noSDCrelaxation}--\ref{t:SDC34relaxation} are relative to
the cheapest simulation ($m=125$ and $n_{{\mathrm{sdc}}}=0$) which took
approximately 38 seconds. Both methods converge with their expected
rate of convergence.}{t:noSDCrelaxation}
\begin{tabular}{c|ccc|ccc}
& \multicolumn{3}{c|}{$\boldnsdc{1}$} &
\multicolumn{3}{c}{$\boldnsdc{2}$} \\
$m$ & $e_{A}$ & $e_{L}$ & CPU & $e_{A}$ & $e_{L}$ & CPU \\
\hline
$125$ & $2.75$E$-8$ & $1.30$E$-8$ & $4.6$
& $3.99$E$-10$ & $5.29$E$-11$ & $7.2$ \\
$250$ & $7.28$E$-9$ & $2.58$E$-9$ & $9.0$
& $1.98$E$-11$ & $6.54$E$-12$ & $16$ \\
$500$ & $1.88$E$-9$ & $4.45$E$-10$ & $17$
& $1.04$E$-11$ & $6.37$E$-13$ & $29$ \\
$1000$ & $4.78$E$-10$ & $6.88$E$-11$ & $35$
& $5.37$E$-12$ & $4.77$E$-14$ & $57$
\end{tabular}
\mcaption{The errors in area and length and the CPU time for a single
vesicle in a {\bf relaxation} flow with a {\bf constant} time step size
using $\boldnsdc{1}$ (left) and $\boldnsdc{2}$ (right). We achieve
second-order convergence with one SDC correction, but third-order
convergence is only observed for the error in length with two SDC
corrections. However, the error in area has plateaued which indicates
that the GMRES tolerance has limited the error in
area.}{t:SDC12relaxation}
\begin{tabular}{c|ccc|ccc}
& \multicolumn{3}{c|}{$\boldnsdc{3}$} &
\multicolumn{3}{c}{$\boldnsdc{4}$} \\
$m$ & $e_{A}$ & $e_{L}$ & CPU & $e_{A}$ & $e_{L}$ & CPU \\
\hline
$125$ & $9.77$E$-11$ & $3.94$E$-13$ & $10$
& $1.10$E$-10$ & $1.07$E$-14$ & $13$ \\
$250$ & $2.03$E$-11$ & $3.13$E$-14$ & $20$
& $2.74$E$-11$ & $3.11$E$-15$ & $26$ \\
$500$ & $4.30$E$-12$ & $2.22$E$-15$ & $41$
& $6.60$E$-12$ & $1.11$E$-16$ & $49$ \\
$1000$ & $7.11$E$-13$ & $1.33$E$-16$ & $80$
& $1.33$E$-12$ & $1.44$E$-15$ & $101$
\end{tabular}
\mcaption{The errors in area and length and the CPU time for a single
vesicle in a {\bf relaxation} flow with a {\bf constant} time step size
using $\boldnsdc{3}$ (left) and $\boldnsdc{4}$ (right). The error in
length achieves fourth-order convergence with three SDC corrections, but
again, the error in area has plateaued. The error in area continues to
plateau with four SDC corrections, and the error in length has reached
machine precision.}{t:SDC34relaxation}
\end{center}
\end{table}
\subsection{Extensional}
We consider two vesicles placed symmetrically around the origin with
the background velocity ${\mathbf{v}}_{\infty} = (-x,y)$
(Figure~\ref{f:extensionalGeom}). We discretize both vesicles with
$N=96$ points and the time horizon is $T=24$ which is long enough that
the distance between the vesicles at the time horizon is
$0.4\sqrt{\Delta s}$, where $\Delta s$ is the arclength spacing. We
use $p=4$ Gauss-Lobatto quadrature points so that ${\mathbf{r}}$ has fifth-order
accuracy which is at least two orders more accurate than all the
reported time integrators. We report results using zero, one, and two
SDC corrections, and BDF in Tables~\ref{t:noSDCextensional}
and~\ref{t:SDC12extensional}. Again, for some of the results, the
expected convergence rates are not observed. Other groups (see, for
example,~\cite{min2003,wei2013}) have observed that for stiff systems,
very small time steps must be taken before the asymptotic convergence
rates are achieved. As before, we expect that other sources of error
will dominate once the temporal asymptotic regime is achieved.
\begin{figure}[htps]
\begin{center}
\begin{tabular}{ccccc}
\ifTikz
\input{extensionalSnapsTime1.tikz} &
\input{extensionalSnapsTime2.tikz} &
\input{extensionalSnapsTime3.tikz} &
\input{extensionalSnapsTime4.tikz} &
\input{extensionalSnapsTime5.tikz}
\fi
\end{tabular}
\end{center}
\mcaption{Two vesicles discretized with $N=96$ points. The vesicles are
initially placed symmetrically around the origin and the background
velocity is ${\mathbf{v}}_{\infty} = (-x,y)$.}{f:extensionalGeom}
\end{figure}
As the vesicles approach one another, their shape is nearly static and
their velocities decrease. Therefore, we test our adaptive time
stepping strategy using zero, one, and two SDC corrections. We expect
that larger time steps can be taken as the vesicles come closer
together. The errors in area and length, the number of accepted and
rejected time steps, the number of FMM calls, and the CPU times are
reported in Table~\ref{t:adaptiveFirstOrderExtensional} ($n_{{\mathrm{sdc}}}=0$),
Table~\ref{t:adaptiveSecondOrderExtensional} ($n_{{\mathrm{sdc}}}=1$), and
Table~\ref{t:adaptiveThirdOrderExtensional} ($n_{{\mathrm{sdc}}}=2$). The
adaptive time stepping strategy does a good job of attaining the desired
tolerance while not having too many rejected time steps. In
Figure~\ref{f:extensionalSummary}, we plot the time step size, the
location of the rejected time steps, and the errors in area and length
for a constant time step size and for an adaptive time step size. While
the constant time step size commits a negligible amount of error shortly
after the initial condition, too much error is accumulated at the
beginning of the simulation. This indicates that a smaller time step
should be taken near the start of the simulation, and then it can be
increased later in the simulation. The left plot of
Figure~\ref{f:extensionalSummary} exactly demonstrates this behaviour.
Comparing
Tables~\ref{t:noSDCextensional}--\ref{t:adaptiveThirdOrderExtensional},
we observe the following behaviors:
\begin{itemize}
\item Unsurprisingly, the errors resulting from the first-order time
integrator are much larger than those resulting from higher-order
time integrators. These larger errors have a very adverse effect
when using adaptive time stepping since very small time steps must be
taken to maintain the requested local truncation error.
\item As we saw in the {\em relaxation} example, the two second-order
methods (BDF and $n_{{\mathrm{sdc}}}=1$) achieve similar errors with respect to
CPU time. Again, only the integrator $n_{{\mathrm{sdc}}}=1$ is compatible with
adaptive time stepping, so we no longer report results using BDF.
\item Considering adaptive time steps, to achieve a four digits of
accuracy, our second-order adaptive time integrator is 11 times
faster than our first-order adaptive time integrator. Furthermore,
if seven digits of accuracy is requested, our third-order adaptive
time integrator is 23\% faster.
\item In general, when smaller tolerances are requested, additional
SDC iterations, rather than smaller time step sizes, should be used
to increase the accuracy. For instance, to achieve seven digits of
accuracy, using two rather than one SDC correction results in a 25\%
savings in CPU time. This behaviour was also observed by
Minion~\cite{min2003} for the Van der Pol equation: ``higher-order
methods are again more efficient when higher precision is required."
\end{itemize}
\begin{table}[htps]
\begin{center}
\begin{tabular}{c|cccc|cccc}
& \multicolumn{4}{c|}{$\boldnsdc{0}$} &
\multicolumn{4}{c}{{\bf BDF}} \\
$m$ & $e_{A}$ & $e_{L}$ & \# fmm & CPU &
$e_{A}$ & $e_{L}$ & \# fmm & CPU \\
\hline
$300$ & $2.46$E$-4$ & $1.27$E$-3$ & $3.99$E$3$ & $1.0$
& $5.76$E$-5$ & $5.67$E$-4$ & $3.88$E$3$ & $0.9$ \\
$600$ & $1.24$E$-4$ & $6.64$E$-4$ & $7.53$E$3$ & $1.8$
& $1.55$E$-5$ & $2.00$E$-4$ & $7.34$E$3$ & $1.8$ \\
$1200$ & $6.24$E$-5$ & $3.46$E$-4$ & $1.44$E$4$ & $3.6$
& $4.12$E$-6$ & $6.54$E$-5$ & $1.40$E$4$ & $3.6$ \\
$2400$ & $3.15$E$-5$ & $1.79$E$-4$ & $2.75$E$4$ & $6.8$
& $1.06$E$-6$ & $1.92$E$-5$ & $2.71$E$4$ & $6.7$
\end{tabular}
\mcaption{The errors in area and length and the CPU time for two
vesicles in an {\bf extensional} flow with a {\bf constant} time step
size using $\boldnsdc{0}$ (left) and ${\bf BDF}$ (right). The CPU
times for
Tables~\ref{t:noSDCextensional}--\ref{t:adaptiveThirdOrderExtensional}
are relative to the cheapest simulation ($m=300$ and $n_{{\mathrm{sdc}}}=0$)
which took approximately 613 seconds. We achieve the desired
first-order results, but second-order results are not yet achieved.
With additional time steps, second-order convergence is achieved (see
Table $7$ in~\cite{qua:bir2013b}).}{t:noSDCextensional}
\begin{tabular}{c|cccc|cccc}
& \multicolumn{4}{c|}{$\boldnsdc{1}$} &
\multicolumn{4}{c}{$\boldnsdc{2}$} \\
$m$ & $e_{A}$ & $e_{L}$ & \# fmm & CPU &
$e_{A}$ & $e_{L}$ & \# fmm & CPU \\
\hline
$300$ & $1.37$E$-6$ & $3.08$E$-6$ & $2.38$E$4$ & $4.6$
& $7.39$E$-7$ & $7.16$E$-8$ & $3.61$E$4$ & $7.1$ \\
$600$ & $1.03$E$-6$ & $1.32$E$-6$ & $4.59$E$4$ & $8.9$
& $1.56$E$-7$ & $8.78$E$-9$ & $6.95$E$4$ & $14$ \\
$1200$ & $4.51$E$-7$ & $5.24$E$-7$ & $8.91$E$4$ & $18$
& $3.26$E$-8$ & $2.23$E$-9$ & $1.35$E$5$ & $26$ \\
$2400$ & $1.60$E$-7$ & $1.76$E$-7$ & $1.75$E$5$ & $34$
& $8.39$E$-9$ & $7.25$E$-10$ & $2.65$E$5$ & $52$ \\
$4800$ & $4.95$E$-8$ & $4.92$E$-8$ & $3.46$E$5$ & $67$
& $2.90$E$-9$ & $1.94$E$-10$ & $5.28$E$5$ & $104$
\end{tabular}
\mcaption{The errors in area and length and the CPU time for two
vesicles in an {\bf extensional} flow with a {\bf constant} time step
size using $\boldnsdc{1}$ (left) and $\boldnsdc{2}$ (right). While
each SDC correction reduces the error, the desired asymptotic rates of
convergence are not achieved. As the ratios of successive errors are
approaching the expected values, we expect that we have not taken
enough time steps to observe the asymptotic rate of convergence. Also,
we observe that one SDC correction and BDF have comparable errors with
respect to computational work.}{t:SDC12extensional}
\begin{tabular}{ccccccc}
Tolerance & $e_{A}$ & $e_{L}$ & Accepts & Rejects & \# fmm & CPU \\
\hline
$1$E$-2$ & $2.10$E$-4$ & $1.34$E$-3$
& $76$ & $28$ & $1.16$E$3$ & $0.3$ \\
$1$E$-3$ & $3.55$E$-5$ & $2.22$E$-4$
& $510$ & $30$ & $4.69$E$3$ & $1.4$ \\
$1$E$-4$ & $6.73$E$-6$ & $4.11$E$-5$
& $4803$ & $34$ & $3.98$E$4$ & $12$
\end{tabular}
\mcaption{The errors in area and length, the CPU time, and the number
of accepted and rejected time steps for two vesicles in an {\bf
extensional} flow with an {\bf adaptive} time step size using
$\boldnsdc{0}$. For the larger tolerances, the desired tolerance is
achieved in an acceptable amount of CPU time. However, first-order
methods require too small of time steps to achieve four digits of
accuracy. With these smaller tolerances, higher-order methods should
be used.}{t:adaptiveFirstOrderExtensional}
\begin{tabular}{ccccccc}
Tolerance & $e_{A}$ & $e_{L}$ & Accepts & Rejects & \# fmm & CPU \\
\hline $1$E$-2$ & $4.24$E$-3$ & $2.93$E$-4$ & $26$ & $18$ & $5.12$E$3$
& $0.9$ \\ $1$E$-3$ & $2.90$E$-4$ & $8.42$E$-5$ & $28$ & $16$ &
$4.32$E$3$ & $0.8$ \\ $1$E$-4$ & $1.01$E$-6$ & $7.10$E$-6$ & $45$ &
$24$ & $5.90$E$3$ & $1.1$ \\ $1$E$-5$ & $7.08$E$-6$ & $8.97$E$-7$
& $97$ & $16$ & $8.14$E$3$ & $1.5$ \\
$1$E$-6$ & $8.79$E$-7$ & $4.40$E$-7$
& $289$ & $21$ & $2.13$E$4$ & $4.1$ \\
$1$E$-7$ & $9.18$E$-8$ & $3.04$E$-9$
& $891$ & $24$ & $6.07$E$4$ & $12$
\end{tabular}
\mcaption{The errors in area and length, the CPU time, and the number of
accepted and rejected time steps for two vesicles in an {\bf
extensional} flow with an {\bf adaptive} time step size and
$\boldnsdc{1}$. This second-order method is able to achieve much
smaller tolerances than $n_{{\mathrm{sdc}}}=0$. However, when seven digits
of accuracy is requested, the number of required time steps becomes
unacceptable. In this case, a third-order method should be
used.}{t:adaptiveSecondOrderExtensional}
\begin{tabular}{ccccccc}
Tolerance & $e_{A}$ & $e_{L}$ & Accepts & Rejects & \# fmm & CPU \\
\hline
$1$E$-5$ & $3.07$E$-6$ & $1.53$E$-6$
& $59$ & $23$ & $1.06$E$4$ & $2.0$ \\
$1$E$-6$ & $6.48$E$-7$ & $1.11$E$-7$
& $143$ & $33$ & $2.06$E$4$ & $4.0$ \\
$1$E$-7$ & $8.90$E$-8$ & $1.68$E$-9$
& $430$ & $22$ & $4.75$E$4$ & $9.2$
\end{tabular}
\mcaption{The errors in area and length, the CPU time, and the number of
accepted and rejected time steps for two vesicles in an {\bf
extensional} flow with an {\bf adaptive} time step size using
$\boldnsdc{2}$. We see that if smaller tolerances are desired, it is
advantageous to use additional SDC corrections to allow for larger time
steps.}{t:adaptiveThirdOrderExtensional}
\end{center}
\end{table}
\begin{figure}[htps]
\begin{tabular}{cc}
\ifTikz
\input{extensionalAdaptiveDT.tikz} &
\input{extensionalAdaptiveErrors.tikz}
\fi
\end{tabular}
\mcaption{Results for the {\bf extensional} flow using $\boldnsdc{1}$.
Left: The time step size using 2400 {\bf constant} time step sizes
(dashed) and an {\bf adaptive} time step size (solid). The open
circles indicate the 24 times when the time step size is rejected. As
expected, the time step size increases as the vesicles come closer to a
steady state. Right: The errors in area and length using {\bf
constant} (dashed) and {\bf adaptive} (solid) time steps, and the
desired error (black) of the adaptive time step. When using adaptive
time stepping, the error in area nearly achieves the desired error
indicating that we are almost selecting the optimal time step size.
However, when using a constant time step size, a large amount of error
is committed at the start of the simulation, and then very little error
is committed for the duration of the simulation. The result is a
nearly three-fold increase in the CPU time. The CPU time is further
reduced by using adaptive time stepping with
$n_{{\mathrm{sdc}}}=2$.}{f:extensionalSummary}
\end{figure}
\subsection{Stenosis}
We consider a single vesicle discretized with $N=128$ points in a
constricted tube discretized with $N_{\mathrm{wall}}=256$ points
(Figure~\ref{f:stenosisGeom}). At this resolution, our FMM
implementation of the double-layer potential is slower than a direct
evaluation. Therefore, the FMM is only used for the single-layer
potentials. The time horizon is $T=15$ which is sufficiently long that
the vesicle passes through the constriction. We again use $p=4$
Gauss-Lobatto quadrature points. We check the rates of convergence for
a varying number of SDC corrections in Tables~\ref{t:SDC01stenosis}
and~\ref{t:SDC23stenosis}. We see that first- and second-order
convergence is achieved in Table~\ref{t:SDC01stenosis}. While the
error continues to decrease with each SDC iteration, as before,
additional orders of convergence are not achieved for the presented
values of $\Delta t$.
\begin{figure}[htps]
\centering
\begin{tabular}{ccc}
\ifTikz
\input{stenosisSnapsTime1.tikz} &
\input{stenosisSnapsTime2.tikz} &
\input{stenosisSnapsTime3.tikz} \\
\input{stenosisSnapsTime4.tikz} &
\input{stenosisSnapsTime5.tikz} &
\input{stenosisSnapsTime6.tikz} \\
\input{stenosisSnapsTime7.tikz} &
\input{stenosisSnapsTime8.tikz} &
\input{stenosisSnapsTime9.tikz}
\fi
\end{tabular}
\mcaption{A single vesicle discretized with $N=128$ points passing
through a constricted tube discretized with $N_{{\mathrm{wall}}}=256$ points. The
boundary condition at the intake and outtake has a parabolic-profile and
on the rest of the solid wall is zero.}{f:stenosisGeom}
\end{figure}
In Figure~\ref{f:stenosisErrors}, we plot the errors in area and length
using $n_{{\mathrm{sdc}}}=1$ and a constant time step size. Unsurprisingly, the
errors increase when the vesicle passes through the constriction. In
this interval, a smaller time step should be taken. In
Tables~\ref{t:adaptiveFirstOrderStenosis}--\ref{t:adaptiveThirdOrderStenosis},
we report results for adaptive time stepping with different tolerances
and different numbers of SDC corrections. In
Figure~\ref{f:stenosisSummary}, we plot the errors in area and length
using $n_{{\mathrm{sdc}}}=1$ with and without adaptive time stepping. We also
plot the time step size and the location of the rejected time steps. As
expected, a much smaller time step size is taken as the vesicle passes
through the constriction.
We observe similar behavior as we observed for the {\em extensional}
example. In particular,
\begin{itemize}
\item To achieve three digits of accuracy with $n_{{\mathrm{sdc}}}=0$, a fixed
time step requires more than 6,000 time steps, and with adaptive time
steps, over 30,000 time steps are required. This sharp increase is
due to a very a small time step that must be taken to keep the local
truncation error below the required threshold as the vesicle passes
through the constriction (for this example, 65\% of the time steps
are smaller than $10^{-4}$). Using the second-order integrator
$n_{{\mathrm{sdc}}}=1$ with adaptive time stepping, only $172 \times 3$
adaptive time steps are required\footnote{Recall that $p-1=3$
additional time steps are required because of the intermediate
Gauss-Lobatto quadrature points.} are required to achieve 3 digits of
accuracy. The resulting speedup is a factor of greater than 22.
\item Comparing the two adaptive time stepping integrators
$n_{{\mathrm{sdc}}}=1$ and $n_{{\mathrm{sdc}}}=2$, six digits of accuracy can be computed
with 16\% less CPU time by using $n_{{\mathrm{sdc}}}=2$. Again, this indicates
that if smaller tolerances are desired, additional SDC corrections
should be used to increase the accuracy of each time step.
\end{itemize}
\begin{table}[htps]
\begin{center}
\begin{tabular}{c|cccc|cccc}
& \multicolumn{4}{c|}{$\boldnsdc{0}$} &
\multicolumn{4}{c}{$\boldnsdc{1}$} \\
$m$ & $e_{A}$ & $e_{L}$ & \# fmm & CPU &
$e_{A}$ & $e_{L}$ & \# fmm & CPU \\
\hline
$750$ & $1.16$E$-2$ & $4.24$E$-2$ & $1.90$E$4$ & $1.0$
& $6.05$E$-5$ & $2.57$E$-5$ & $1.26$E$5$ & $6.9$ \\
$1500$ & $5.96$E$-3$ & $2.12$E$-2$ & $3.79$E$4$ & $2.2$
& $1.50$E$-5$ & $4.48$E$-6$ & $2.50$E$5$ & $13$ \\
$3000$ & $3.03$E$-3$ & $1.06$E$-2$ & $7.55$E$4$ & $4.2$
& $3.73$E$-6$ & $7.00$E$-7$ & $4.96$E$5$ & $26$ \\
$6000$ & $1.53$E$-3$ & $5.28$E$-3$ & $1.51$E$5$ & $8.7$
& $9.30$E$-7$ & $1.03$E$-7$ & $9.83$E$5$ & $53$
\end{tabular}
\mcaption{The errors in area and length and the CPU time for a single
vesicle in a constricted tube ({\bf stenosis}) with a {\bf constant}
time step size using $\boldnsdc{0}$ (left) and $\boldnsdc{1}$ (right).
The CPU times for
Tables~\ref{t:SDC01stenosis}--\ref{t:adaptiveThirdOrderStenosis} are
relative to the cheapest simulation ($m=750$ and $n_{{\mathrm{sdc}}}=0$) which
took approximately $2.30$E$3$ seconds. We achieve the expected first-
and second-order convergence rates.}{t:SDC01stenosis}
\begin{tabular}{c|cccc|cccc}
& \multicolumn{4}{c|}{$\boldnsdc{2}$} &
\multicolumn{4}{c}{$\boldnsdc{3}$} \\
$m$ & $e_{A}$ & $e_{L}$ & \# fmm & CPU &
$e_{A}$ & $e_{L}$ & \# fmm & CPU \\
\hline
$750$ & $2.74$E$-5$ & $1.08$E$-6$ & $1.94$E$5$ & $11$
& $3.51$E$-7$ & $4.38$E$-8$ & $2.62$E$5$ & $13$ \\
$1500$ & $7.16$E$-6$ & $7.36$E$-8$ & $3.84$E$5$ & $21$
& $5.97$E$-8$ & $1.30$E$-9$ & $5.19$E$5$ & $27$ \\
$3000$ & $1.83$E$-6$ & $5.31$E$-9$ & $7.60$E$5$ & $39$
& $9.30$E$-9$ & $4.60$E$-11$ & $1.03$E$6$ & $56$ \\
$6000$ & $4.63$E$-7$ & $4.53$E$-10$ & $1.51$E$6$ & $81$
& $1.35$E$-9$ & $2.02$E$-12$ & $2.03$E$6$ & $108$
\end{tabular}
\mcaption{The errors in area and length and the CPU time for a single
vesicle in a constricted tube ({\bf stenosis}) with a {\bf constant}
time step size using $\boldnsdc{2}$ (left) and $\boldnsdc{3}$ (right).
We do not obtain the expected convergence rates, but we do see that each
SDC correction does result in an increase in the accuracy. We could
attempt to reach the asymptotic convergence rates, but this most likely
would require temporal resolutions where other sources of error will
dominate.}{t:SDC23stenosis}
\begin{tabular}{ccccccc}
Tolerance & $e_{A}$ & $e_{L}$ & Accepts & Rejects & \# fmm & CPU \\
\hline
$1$E$-2$ & $2.57$E$-3$ & $7.87$E$-3$
& $3397$ & $21$ & $1.08$E$5$ & $5.7$ \\
$1$E$-3$ & $2.95$E$-4$ & $9.44$E$-4$
& $33554$ & $12$ & $1.06$E$6$ & $59$ \\
\end{tabular}
\mcaption{The errors in area and length, the CPU time, and the number
of accepted and rejected time steps for a single vesicle in a
constricted tube ({\bf stenosis}) with an {\bf adaptive} time step size
using $\boldnsdc{0}$. Since the time integrator is first-order, small
time steps have to be taken. The result is that the CPU time is too
large at the reported tolerances.}{t:adaptiveFirstOrderStenosis}
\begin{tabular}{ccccccc}
Tolerance & $e_{A}$ & $e_{L}$ & Accepts & Rejects & \# fmm & CPU \\
\hline
$1$E$-2$ & $2.60$E$-3$ & $8.03$E$-4$
& $68$ & $30$ & $2.06$E$4$ & $1.2$ \\
$1$E$-3$ & $5.55$E$-4$ & $1.11$E$-4$
& $172$ & $68$ & $4.80$E$4$ & $2.6$ \\
$1$E$-4$ & $6.82$E$-5$ & $1.18$E$-5$
& $503$ & $63$ & $1.11$E$5$ & $5.9$ \\
$1$E$-5$ & $7.20$E$-6$ & $9.08$E$-7$
& $1559$ & $54$ & $3.15$E$5$ & $17$ \\
$1$E$-6$ & $7.35$E$-7$ & $5.60$E$-8$
& $4904$ & $40$ & $9.61$E$5$ & $51$ \\
\end{tabular}
\mcaption{The errors in area and length, the CPU time, and the number
of accepted and rejected time steps for a single vesicle in a
constricted tube ({\bf stenosis}) with an {\bf adaptive} time step size
using $\boldnsdc{1}$. This second-order method is able to take much
larger time steps than the first-order
method.}{t:adaptiveSecondOrderStenosis}
\begin{tabular}{ccccccc}
Tolerance & $e_{A}$ & $e_{L}$ & Accepts & Rejects & \# fmm & CPU \\
\hline
$1$E$-5$ & $7.19$E$-6$ & $4.63$E$-8$
& $828$ & $136$ & $2.96$E$5$ & $16$ \\
$1$E$-6$ & $7.54$E$-7$ & $2.49$E$-9$
& $2558$ & $105$ & $8.21$E$5$ & $43$ \\
\end{tabular}
\mcaption{The errors in area and length, the CPU time, and the number
of accepted and rejected time steps for a single vesicle in a
constricted tube ({\bf stenosis}) with an {\bf adaptive} time step size
using $\boldnsdc{2}$. We see that if small tolerances are desired,
additional SDC iterations should be used to achieve a higher-order time
integrator.}{t:adaptiveThirdOrderStenosis}
\end{center}
\end{table}
\begin{figure}[htps]
\centering
\begin{tabular}{cccc}
\ifTikz
\input{stenosisErrorsSDC750.tikz} &
\input{stenosisErrorsSDC1500.tikz} &
\input{stenosisErrorsSDC3000.tikz} &
\input{stenosisErrorsSDC6000.tikz}
\fi
\end{tabular}
\mcaption{The errors in area and length for a single vesicle in a
constricted tube ({\bf stenosis}) with a {\bf constant} time step size
using $\boldnsdc{1}$ (right side of Table~\ref{t:SDC01stenosis}).
Notice that the error increases sharply as the vesicle passes through
the constriction for all the time step sizes. This indicates that a
smaller time step size should be taken as the vesicle passes through
the constriction.}{f:stenosisErrors}
\end{figure}
\begin{figure}[htps]
\begin{tabular}{cc}
\ifTikz
\input{stenosisAdaptiveDT.tikz} &
\input{stenosisAdaptiveErrors.tikz}
\fi
\end{tabular}
\mcaption{Results for the {\bf stenosis} flow using $\boldnsdc{1}$.
Left: The time step size using 6000 {\bf constant} time step sizes
(dashed) and an {\bf adaptive} time step size (solid). The open circles
indicate the 40 times when the time step size is rejected. Notice that
the time step size decreases as the vesicle passes through the
constriction. Right: The errors in area and length using {\bf constant}
(dashed) and {\bf adaptive} (solid) time steps, and the desired error
(black) of the adaptive time step. When using adaptive time stepping,
the error in area nearly achieves the desired error everywhere except
when the vesicle passes through the constriction, where the error in
area actually drops. However, when using a constant time step size,
very little error is committed at the start of the simulation, and then
the majority of the error is committed as the vesicle passes through the
constriction. Even though the CPU savings are negligible (it is about
4\%), the adaptive time stepping method did not require any trial and
error to find an appropriate time step size. The user simply specifies
that they desire six digits accuracy and it is achieved. In addition,
by using an additional SDC correction, the CPU savings is increased to
19\%.}{f:stenosisSummary}
\end{figure}
\subsection{Couette}
\label{s:couette}
We consider eight vesicles discretized with $N=192$ points in a couette
apparatus where each wall is discretized with $N_{\mathrm{wall}}=512$
points (Figure~\ref{f:couetteGeom}). At this resolution, our FMM
implementation of the double-layer potential is faster than a direct
evaluation. Therefore, we use the FMM to evaluate all the single- and
double-layer potentials. We use $p=4$ Gauss-Lobatto quadrature points
and take the long time horizon $T=50$. At the time horizon, the inner
boundary has made five complete rotations with a constant velocity, and
the outer boundary is stationary. The inner and outer boundaries are
aligned so that there is a narrow region that the vesicles must pass
through. As a vesicle passes through this region, a smaller time step
should be taken. We check the convergence rates with a constant time
step size using zero and one SDC correction in
Tables~\ref{t:SDC0couette} and~\ref{t:SDC1couette}. For both time
integrators, the expected first- and second-order convergence rates are
achieved.
\begin{figure}[htps]
\centering
\begin{tabular}{cccccc}
\ifTikz
\input{couetteSnapsTime1.tikz} &
\input{couetteSnapsTime2.tikz} &
\input{couetteSnapsTime3.tikz} &
\input{couetteSnapsTime4.tikz} &
\input{couetteSnapsTime5.tikz} &
\input{couetteSnapsTime6.tikz} \\
\input{couetteSnapsTime7.tikz} &
\input{couetteSnapsTime8.tikz} &
\input{couetteSnapsTime9.tikz} &
\input{couetteSnapsTime10.tikz} &
\input{couetteSnapsTime11.tikz} &
\input{couetteSnapsTime12.tikz}
\fi
\end{tabular}
\mcaption{Eight vesicles discretized with $N=192$ points in a couette
apparatus whose solid walls are each discretized with $N_{{\mathrm{wall}}}=512$
points. The outer boundary is stationary and the inner boundary has
constant angular velocity and has completed five full rotations at
$T=50$. A single vesicle is colored in blue to help with visualization.
The interactions between the vesicles and the solid walls complicate and
simplify throughout the simulation and the simulation benefits from
adaptive time stepping.}{f:couetteGeom}
\end{figure}
In Figure~\ref{f:couetteErrors}, we plot the errors in area and length
using $n_{{\mathrm{sdc}}}=1$ with a constant time step size. If $500$ time steps
are taken, the error in length has a sharp jump caused by the blue
vesicle in Figure~\ref{f:couetteGeom}. By reducing the time step size,
the errors decrease at the expected rate, but there is still a jump in
the errors near $t=2$. In Table~\ref{t:adaptiveSecondOrderCouette}, we
report results for $n_{{\mathrm{sdc}}}=1$ with adaptive time step sizes and two
different tolerances. In Figure~\ref{f:couetteSummary}, we plot the
errors in area and length with and without adaptive time stepping. We
also plot the time step size and the locations of rejected time steps.
We see that a small time step is taken initially as the vesicles smooth,
and then much larger time steps can be taken once the vesicles are
smooth and separated from one another and the solid walls.
While the desired error is achieved, the final error is much smaller
than the tolerance when compared to the {\em extensional} and {\em
stenosis} examples. This is a result of having more vesicles since the
local truncation error is estimated with the maximum change in error,
where the maximum is taken over all the vesicles. In contrast to the
previous two examples, different vesicles commit the largest error at
different times. The result is that the global error of our adaptive
time stepping method is much less than the desired tolerance. A
possible solution to this problem is to adjust the allowable truncation
error at each time step. In particular, if we again assume that the
time horizon is $T=1$ and the desired tolerance is $\epsilon$, then
condition~\eqref{e:errorBounds} is replaced with
\begin{align*}
|A(t + \Delta t) - A(t)| \leq A(t) \frac{\Delta t}{1-t}
\left(\epsilon - \frac{|A(t) - A(0)|}{A(0)} \right),
\end{align*}
and a similar condition for the length. Using this new strategy with
the time integrator from Table~\ref{t:adaptiveSecondOrderCouette}, with
a tolerance of $1$E$-1$, the final error is $5.14$E$-2$, the number of
accepted time steps is reduced to 222, but 84 time steps are rejected.
Using a tolerance of $1$E$-2$, the final error is $9.62$E$-3$ with 583
accepted time steps, and the number of rejected time steps is only
slightly increased to 74. While this strategy does offer a speedup, we
expect that multiscale time integrators, where each vesicle's time step
size can differ from the others', will overall result in a more robust
solver.
We have also run the same simulation with a coarser discretization. We
leave all the parameters unchanged with the exception of $N=92$ points
per vesicle and $N_{{\mathrm{wall}}}=192$ points per solid wall. Using adaptive
time stepping with one SDC correction, a tolerance of $1$E$-1$ and
$5$E$-2$ are attainable. However, with a tolerance of $1$E$-2$, the
error condition~\eqref{e:errorBounds} can not be satisfied at the
initial condition for all $\Delta t$. Therefore, the spatial error is
too large for condition~\eqref{e:errorBounds} to be satisfied. We also
ran this resolution at a later time when the vesicle shapes are
smoothed. There, the desired local truncation error is attainable and
the simulation can successfully complete. This indicates that spatial
adaptivity is required, and this is a current extension we are
developing.
Our conclusions from this example are:
\begin{itemize}
\item Considering the time integrator $n_{{\mathrm{sdc}}}=1$, to achieve one
digit of accuracy, adaptive time stepping is at least 33\% faster.
However, to achieve two digits accuracy, with our current limitations
(fixed spatial resolution, and each vesicle having the same time step
size), adaptive time stepping does not conclusively offer a speedup.
However, unlike the simulations with fixed time step size, no trial and
error procedure to choose a time step size that results in the desired
global error is required.
\item With multiple vesicles, it is harder to achieve the desired
global error. By adjusting the allowable local truncation error
throughout the simulation, the final error is much closer to the
tolerance. However, each vesicle must still have the same time step
size. We plan to develop integrators that use different time step
sizes for each vesicle.
\item Without spatial adaptivity, excessively large resolutions are
required for the entirety of the simulation.
\end{itemize}
\begin{table}[htps]
\begin{center}
\begin{tabular}{c|cccc}
\multicolumn{4}{c}{$\boldnsdc{0}$} \\
$m$ & $e_{A}$ & $e_{L}$ & \# fmm & CPU \\
\hline
2000 & $1.24$E$-1$ & $3.50$E$-1$ & $2.73$E$5$ & $1.0$ \\
4000 & $2.04$E$-2$ & $1.21$E$-1$ & $5.28$E$5$ & $2.1$ \\
8000 & $5.99$E$-3$ & $5.79$E$-2$ & $1.05$E$6$ & $3.8$
\end{tabular}
\mcaption{The errors in area and length and the CPU time for eight
vesicles in a {\bf couette} apparatus with a {\bf constant} time step
size using $\boldnsdc{0}$. The CPU for
Tables~\ref{t:SDC0couette}--\ref{t:adaptiveSecondOrderCouette} are
relative to the cheapest simulation ($m=2000$ and $n_{{\mathrm{sdc}}}=0$) which
took approximately $1.44$E$5$ seconds. We achieve the expected
first-order convergence.}{t:SDC0couette}
\begin{tabular}{c|cccc}
& \multicolumn{4}{c}{$\boldnsdc{1}$} \\
$m$ & $e_{A}$ & $e_{L}$ & \# fmm & CPU \\
\hline
$500$ & $7.14$E$-3$ & $5.39$E$-2$ & $4.27$E$5$ & $1.5$ \\
$1000$ & $2.79$E$-3$ & $4.68$E$-5$ & $8.28$E$5$ & $2.9$ \\
$2000$ & $7.04$E$-4$ & $9.34$E$-6$ & $1.69$E$6$ & $6.0$
\end{tabular}
\mcaption{The errors in area and length and the CPU time for eight
vesicles in a {\bf couette} apparatus with a {\bf constant} time step
size using $\boldnsdc{1}$. At the reported resolutions, we are just
beginning to achieve second-order convergence.}{t:SDC1couette}
\begin{tabular}{ccccccc}
Tolerance & $e_{A}$ & $e_{L}$ & Accepts & Rejects & \# fmm & CPU \\
\hline
$1$E$-1$ & $2.65$E$-2$ & $2.07$E$-2$
& $256$ & $68$ & $2.90$E$5$ & $1.0$ \\
$1$E$-2$ & $4.21$E$-3$ & $3.86$E$-4$
& $780$ & $65$ & $6.97$E$5$ & $2.5$
\end{tabular}
\mcaption{The errors in area and length, the CPU time, and the number of
accepted and rejected time steps for eight vesicles in a {\bf couette}
apparatus with an {\bf adaptive} time step size using
$\boldnsdc{1}$.}{t:adaptiveSecondOrderCouette}
\end{center}
\end{table}
\begin{figure}[htps]
\centering
\begin{tabular}{ccc}
\ifTikz
\input{couetteErrorsSDC500.tikz} &
\input{couetteErrorsSDC1000.tikz} &
\input{couetteErrorsSDC2000.tikz}
\fi
\end{tabular}
\mcaption{The errors in area and length for eight vesicles in a {\bf
couette} apparatus with a {\bf constant} time step size using
$\boldnsdc{1}$. For all the time step sizes, there is a sharp increase
in the errors around $t=2$ which is committed by the vesicle colored in
blue in Figure~\ref{f:couetteGeom}. Adaptive time stepping
automatically resolves these sharp jumps and avoids the need to use
trial and error to find an appropriate time step size.}{f:couetteErrors}
\end{figure}
\begin{figure}[htps]
\begin{tabular}{cc}
\ifTikz
\input{couetteAdaptiveDT.tikz} &
\input{couetteAdaptiveErrors.tikz}
\fi
\end{tabular}
\mcaption{Results for the {\bf couette} apparatus using
$\boldnsdc{1}$. Left: The time step size using a {\bf constant} time
step (dashed) and an {\bf adaptive} time step size (solid). The open
circles indicate the 65 times when the time step size is rejected. We
see that the time step size is smallest when the error is large (near
$t=2$) and is largest when the vesicles and walls are well-separated
(near $t=15$ and $t=30$). Right: The errors in area and length using
{\bf constant} (dashed) and {\bf adaptive} (solid) time steps, and the
desired error (black) of the adaptive time step. Since each vesicle
must use the same time step size, the actual error does not closely
follow the desired error.}{f:couetteSummary}
\end{figure}
\section{Introduction\label{s:intro}}
\input introduction.tex
\section{Formulation\label{s:formulation}}
\input formulation.tex
\section{Numerical Scheme\label{s:numerics}}
\input numerics.tex
\section{Adaptive Time Stepping\label{s:adaptive}}
\input adaptive.tex
\section{Results\label{s:results}}
\input results.tex
\section{Conclusions\label{s:conclusions}}
\input conclusions.tex
\begin{appendices}
\section{Complete SDC Formulation \label{a:appendix1}}
\input appendixA.tex
\end{appendices}
\bibliographystyle{plainnat}
\biboptions{sort&compress}
|
\@startsection {section}{1}{\z@}{10pt plus 2pt minus 2pt{\@startsection {section}{1}{\z@}{10pt plus 2pt minus 2pt}
{8pt plus 2pt minus 2pt}{\centering\normalsize\sc
\edef\@svsec{\Roman{section}.\ }}}
\def\Roman{section}{\Roman{section}}
\def\@startsection {subsection}{2}{\z@}{10pt plus 2pt minus 2pt{\@startsection {subsection}{2}{\z@}{10pt plus 2pt minus 2pt}
{6pt plus 2pt minus 2pt}{\normalsize\sl
\edef\@svsec{\Alph{subsection}.\ }}}
\def\Alph{subsection}{\Alph{subsection}}
\long\def\@makecaption#1#2{
\vskip6pt\begin{center} #1 #2 \end{center}\par\vskip 1pt}
\def\fnum@figure{\raggedright{\footnotesize Fig. \thefigure }.%
\footnotesize}
\def\fnum@table{\footnotesize TABLE \Roman{table}\\\footnotesize\sc}
\def\Roman{table}{\Roman{table}}
\topsep 0.47\baselineskip
\partopsep \z@
\parsep \z@
\itemsep \z@
\itemindent -1em
\leftmargin 2em
\leftmargini 2em
\leftmarginii 1em
\leftmarginiii 1.5em
\leftmarginiv 1.5em
\leftmarginv 1.0em
\leftmarginvi 1.0em
\labelsep 0.5em
\labelwidth \z@
\def\@listi{\leftmargin\leftmargini \topsep 2pt plus 1pt minus 1pt}
\let\@listI\@listi
\def\@listii{\leftmargin\leftmarginii\labelwidth\leftmarginii%
\advance\labelwidth-\labelsep \topsep 2pt}
\def\@listiii{\leftmargin\leftmarginiii\labelwidth\leftmarginiii%
\advance\labelwidth-\labelsep \topsep 2pt}
\def\@listiv{\leftmargin\leftmarginiv\labelwidth\leftmarginiv%
\advance\labelwidth-\labelsep \topsep 2pt}
\def\@listv{\leftmargin\leftmarginv\labelwidth\leftmarginv%
\advance\labelwidth-\labelsep \topsep 2pt}
\def\@listvi{\leftmargin\leftmarginvi\labelwidth\leftmarginvi%
\advance\labelwidth-\labelsep \topsep 2pt}
\def0.97{0.97}
\makeatother
\begin{document}
\title{\Large\bf Decomposition of Diagonal Hermitian Quantum Gates Using Multiple-Controlled Pauli \emph{Z} Gates
\date{}
\author{\normalsize
\begin{tabular}[]{c c}
\hspace{0 pt}\large Mahboobeh Houshmand & \hspace{8 pt} \large Morteza Saheb Zamani \\\\
\hspace{0 pt} \large Mehdi Sedighi & \hspace{8 pt} \large Mona Arabzadeh\\\\
\end{tabular}
\\
\normalsize \textbf{\emph{Department of Computer Engineering and Information Technology}}\\
\normalsize \textbf{\emph{Amirkabir University of Technology}}\\
\normalsize \textbf{\emph{Tehran, Iran}}\\\\
\normalsize \emph{\{houshmand,szamani,msedighi,m.arabzadeh\}@aut.ac.ir}\\
}
\maketitle
\thispagestyle{empty}
\newcounter {TCounter}
\newcounter {LCounter}
\newcounter {PCounter}
\newcounter {CCounter}
\newcounter {DCounter}
\newcounter {ECounter}
\newtheorem{theorem}[TCounter]{\textbf{Theorem}}
\newtheorem{lemma}[LCounter]{\textbf{Lemma}}
\newtheorem{proposition}[PCounter]{\textbf{Proposition}}
\newtheorem{corollary}[CCounter]{\textbf{Corollary}}
\newtheorem{definition}[DCounter]{\textbf{Definition}}
\newtheorem{example}[ECounter]{\textbf{Example}}
\newenvironment{proofs}[1][\textbf{Proof}]{\begin{trivlist}
\item[\hskip \labelsep {\bfseries #1}]}{\end{trivlist}}
\newenvironment{remark}[1][\textbf{Remark}]{\begin{trivlist}Alg:1
\item[\hskip \labelsep {\bfseries #1}]}{\end{trivlist}}
\newcommand{\qed}{\nobreak \ifvmode \relax \else
\ifdim\lastskip<1.5em \hskip-\lastskip
\hskip1.5em plus0em minus0.5em \fi \nobreak
\vrule height0.75em width0.5em depth0.25em\fi}
{\small\bf Abstract---
Quantum logic decomposition refers to decomposing a given quantum gate to a set of physically implementable gates. An approach has been presented to decompose arbitrary diagonal quantum gates to a set of multiplexed-rotation gates around $z$ axis.
In this paper, a special class of diagonal quantum gates, namely diagonal Hermitian quantum gates, is considered and a new perspective to the decomposition problem with respect to decomposing these gates is presented. It is first shown that these gates can be decomposed to a set that solely consists of multiple-controlled $Z$ gates. Then a binary representation for the diagonal Hermitian gates is introduced. It is shown that the binary representations of multiple-controlled $Z$ gates form a basis for the vector space that is produced by the binary representations of all diagonal Hermitian quantum gates. Moreover, the problem of decomposing a given diagonal Hermitian gate is mapped to the problem of writing its binary representation in the specific basis mentioned above.
Moreover, CZ gate is suggested to be the two-qubit gate in the decomposition library, instead of previously used CNOT gate.
Experimental results show that the proposed approach can lead to circuits with lower costs in comparison with the previous ones.}
\\
\\
{\small\bf Keywords---
Diagonal Hermitian quantum gates, Optimization, Decomposition
}
\@startsection {section}{1}{\z@}{10pt plus 2pt minus 2pt {Introduction}
Quantum logic decomposition refers to decomposing a given quantum gate to a set of physically implementable gates by quantum technologies. This set of gates typically consists of CNOT and single-qubit gates, called ``basic gate" library~\cite{Barenco95} or CNOT and single-qubit rotation gates, called ``elementary gate" library~\cite{bullock2004}. Many studies have been performed on the decomposition of quantum gates. Barenco et al.~\cite{Barenco95} showed that the number of CNOT gates required to implement an arbitrary quantum gate on $n$ qubits was $O(n^3 4^n)$. In~\cite{George2001}, it was shown that applying the QR matrix decomposition in linear algebra to decompose quantum gates could lead to the same result.
The highest known lower bound on the number of CNOT gates to decompose an $n$-qubit quantum gate was reported as $\left\lceil {\frac{1}{4}(4^n - 3n - 1)} \right\rceil$ in~\cite{shende-2004}. The improved method based on QR decomposition~\cite{vartiainen-2004} and Cosine-Sine Decomposition (CSD)~\cite{Mottonen-2004,Shende06,saeedi2010block} achieve the order of $O(4^n)$ CNOT gates.
To further reduce the number of elementary gates, the decomposition problem of more specific quantum operators, such as two-qubit gates~\cite{vidal-2004-69,Shende-1-2004,vatan-2004-69} and controlled unitary gates~\cite{song-2003-3} has been considered. Among the specific quantum gates are the diagonal quantum operators. The decomposition problem for these gates has been addressed in~\cite{Hogg1999,Bullock2003,bullock2004,Shende06}.
In~\cite{Bullock2003} it has been shown that any two-qubit diagonal matrix can be implemented by at most five elementary gates up to a global phase. The authors of~\cite{bullock2004} prove that the lowest number of required elementary gates for decomposition of arbitrary $n$-qubit diagonal quantum gates is $2^n-1$ and present asymptotically optimal circuits for these gates which require at most $2^{n+1}-3$ elementary gates. Using the definition of quantum multiplexer gates in~\cite{Shende06}, the diagonal gates of~\cite{bullock2004} can be decomposed to a set of multiplexed-rotation gates around $z$ axis.
In this paper, a new perspective to the decomposition problem with respect to decomposing $n$-qubit diagonal Hermitian gates is presented. At first, it is shown that such gates can be decomposed to a set that solely consists of $\mathrm{C^kZ}$ gates, $0\leq k \leq n-1$. Then a binary representation for diagonal Hermitian quantum gates is introduced and the binary representations of the set of $\mathrm{C^kZ}$ gates are mapped to a basis for vector space that is produced by these binary numbers. After that, the problem of decomposing a diagonal Hermitian quantum gate is mapped to the problem of writing these binary numbers in that specific basis.
Although CNOT is the common two-qubit gate in existing decomposition libraries, in some technologies such as ion-trap, which is one of the most promising candidates for realization of scalable quantum computers~\cite{nakahara}, CZ gate has been directly implemented~\cite{Nielsen10} and CNOT is realized by using a middle CZ gate and two rotation gates around the $y$ axis that apply on the target qubit. Therefore, a new library that contains CZ gate as a two-qubit gate can be considered. $\mathrm{C^kZ}$ gates for $k>1$ can in turn be decomposed to CZ and single-qubit gates. CZ gate is also useful in creating a parallel structure for quantum circuits using one-way quantum computation model~\cite{PhysRevLett.86,par}, as the input quantum circuits to this approach are assumed to contain CZ gates. Single-qubit gates in ion-trap technology should be constructed from rotation gates around $y$ and $x$ axes~\cite{Nielsen10}.
Diagonal gates appear in some decomposition methods such as QR \cite{vartiainen-2004} and CSD \cite{Bergholm:2004} and in the case that these diagonal gates are Hermitian, the proposed approach can be applied.
The paper is organized as follows. In the next section, some preliminaries are presented.
Section~\ref{sec:proposed} explains the proposed approach. Experimental results are presented in Section~\ref{sec:results} and finally, Section~\ref{sec:conc} concludes the paper.
\@startsection {section}{1}{\z@}{10pt plus 2pt minus 2pt{Preliminaries}
In this section, preliminaries about quantum states and the quantum gates used in this paper are introduced.
\@startsection {subsection}{2}{\z@}{10pt plus 2pt minus 2pt{Quantum States and Quantum Gates}
\label{sec:qubits}
Quantum bits or \emph{qubits} are quantum analogues of classical bits. A qubit is a unit vector in a two-dimensional Hilbert space, $\mathcal{H}_2$, for which an orthonormal basis, denoted by $\{$$\left\vert 0\right\rangle$,
$\left\vert 1\right\rangle$$\}$, has been fixed. Unlike classical bits, qubits can be in a superposition of $\left\vert 0\right\rangle$ and $\left\vert 1\right\rangle$ like $\alpha\left\vert 0\right\rangle+\beta\left\vert 1\right\rangle$ where $\alpha$ and $\beta$ are complex numbers such that
$|\alpha|^2 + |\beta|^2 = 1$. If such a superposition is measured with respect to the basis $\{$$\left\vert 0\right\rangle$, $\left\vert 1\right\rangle$$\}$, then the classic outcome of 0 is observed with the probability of $|\alpha|^2$ and the classical result of 1 is observed with the probability of $|\beta|^2$. If 0 is obtained, the state of the system after measurement will collapse to $\left\vert 0\right\rangle$ and if 1 is obtained, it will be $\left\vert 1\right\rangle$.
Every $n$-qubit quantum gate is a linear transformation represented by a unitary matrix defined on an $n$-qubit Hilbert
space. A matrix $U$ is \emph{unitary} if $UU^{\dagger} = I$, where $U^{\dagger}$ is the conjugate transpose of the matrix $U$.
Some useful single-qubit gates are the elements of the Pauli set:
\[\sigma_0=
I=%
\begin{bmatrix}
1 & 0\\
0 & 1
\end{bmatrix},
\sigma_1=
X=%
\begin{bmatrix}
0 & 1\\
1 & 0
\end{bmatrix},
\sigma_2=
Y=%
\begin{bmatrix}
0 & -i\\
i & 0
\end{bmatrix},
\sigma_3=
Z=%
\begin{bmatrix}
1 & 0\\
0 & -1
\end{bmatrix}.\]
Another class of useful unitary gates on a single qubit are rotation operators around $x$, $y$ and $z$ axis with the angle $\alpha$ in the Bloch sphere, as shown below:
\[
R_x (\alpha )=%
\begin{bmatrix}
{{\rm cos}\frac{\alpha }{{\rm 2}}} & { - i{\rm sin}\frac{\alpha }{{\rm 2}}}\\
{- i{\rm sin}\frac{\alpha }{{\rm 2}}} & {{\rm cos}\frac{\alpha }{{\rm 2}}}
\end{bmatrix},
R_y (\alpha )=%
\begin{bmatrix}
{{\rm cos}\frac{\alpha }{{\rm 2}}} & { - {\rm sin}\frac{\alpha }{{\rm 2}}}\\
{{\rm sin}\frac{\alpha }{{\rm 2}}} & {{\rm cos}\frac{\alpha }{{\rm 2}}}
\end{bmatrix},
R_z (\alpha )=%
\begin{bmatrix}
{e^{ - i\frac{\alpha }{2}} } & 0\\
0 & {e^{ i\frac{\alpha }{2}} }
\end{bmatrix}.\]
Hadamard, $H$, and \emph{T} are two other known single-qubit gates where:
\[
H=%
\frac{1}{\sqrt{2}}\begin{bmatrix}
1 & 1\\
1 & -1
\end{bmatrix},
T=
\begin{bmatrix}
e^{i\frac{\pi }{8}} & 0\\
0 & e^{-i\frac{\pi }{8}}
\end{bmatrix}
.
\]
If $U$ is a gate that operates on a single qubit, then controlled-$U$ is a gate that operates on two qubits, i.e., control and target qubits, and \emph{U} is applied to the target qubit if the control qubit is $\left\vert 1\right\rangle$ and leaves it unchanged otherwise. For example, controlled-$Z$ (CZ) and controlled-NOT (CNOT) gates perform the $Z$ and $X$ operators respectively on the target qubit if the control qubit is $\left\vert 1\right\rangle$. Otherwise, the target qubit remains unchanged. $\mathrm{C^{k}U}$ gates have $k$ control qubits and one target. When all $k$ control qubits are in the state $|1\rangle$, gate $U$ is applied on the target qubit and no action is taken otherwise. $\mathrm{C^2NOT}$ gate is called Toffoli gate.
A quantum circuit consists of quantum gates interconnected by quantum wires carrying qubits with time flowing from left or right. The unitary matrix of the quantum circuit is evaluated by either dot product or tensor product of the unitary matrices of those quantum gates. The net effect of the gates which are applied to the same subset of qubits in series is computed by the dot product which is the same as the known matrix multiplication. The adjacent gates which act on independent subsets of qubits can be applied in parallel and their overall net effect is computed by their tensor product as defined as follows. Let $A$ be an $m\times n$ matrix and let B be a $p\times q $ matrix. Then $A \otimes B$ is an $(mp)\times(nq)$ matrix called the tensor product (Kronecker product) of $A$ and $B$ as defined below:
\[
A \otimes B = \left[ \begin{array}{l}
a_{11} B,a_{12} B,...,a_{1n} B \\
a_{21} B,a_{22} B,...,a_{2n} B \\
\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,... \\
a_{m1} B,a_{m2} B,...,a_{mn} B \\
\end{array} \right]
\]
where $a_{ij}$ shows the element in the $i^{th}$ column and the $j^{th}$ row of matrix $A$.
\@startsection {subsection}{2}{\z@}{10pt plus 2pt minus 2pt{Hermitian Quantum Gates}
\label{sec:ckz}
A matrix $\mathbb{H}$ is called Hermitian~\cite{Nielsen10} or self-adjoint if $\mathbb{H}^\dag=\mathbb{H}$. The Pauli matrices are some examples of Hermitian matrices. All eigenvalues of a Hermitian matrix are real numbers. On the other hand, the eigenvalues of a unitary matrix have modulus equal to 1 and hence the eigenvalues of a Hermitian unitary matrix are either +1 or -1.
An $n$-qubit diagonal unitary matrix can be represented as $\sum\limits_{i = 0}^{2^n - 1} {\lambda_i|i\rangle \langle i|}$ where $\lambda_i^,s$ are the eigenvalues of the matrix and $|\lambda_i|=1$.
Since the diagonal elements of a diagonal matrix are its eigenvalues, the diagonal elements of a diagonal Hermitian quantum gate are either +1 or -1.
\@startsection {subsection}{2}{\z@}{10pt plus 2pt minus 2pt{The Properties of $\mathrm{C^{k}Z}$ Gates}
\label{sec:hermiti}
Since $\mathrm{C^kZ}$ gates are used in the proposed decompostion method, some of their properties are discussed in this section.
The matrix representation of the CZ gate is as follows:
\[
\text{CZ}=%
\begin{bmatrix}
1 & 0& 0&0\\
0 & 1&0&0\\
0&0&1&0\\
0&0&0&-1
\end{bmatrix}.\]
Figure~\ref{fig:cz} shows the circuit representation of CZ gate.
\begin{figure}
\centering
\[
\Qcircuit @C=1.0em @R=1.4em{
& \ctrl{1}&\qw\\
& \ctrl{0}&\qw
}
\]
\caption{The circuit representation of CZ gate.}
\label{fig:cz}
\end{figure}
In this paper, $Z$ gate is also denoted as $\mathrm{C^{0}Z}$ gate. The $\mathrm{C^{k}Z}$ gates for $k\ge 0$ are symmetric with respect to exchanging qubits. They negate the input qubits when all of them are in the state $|1\rangle$ and otherwise leave them unchanged. These gates are diagonal and self-inverse and they all commute with each other. Consider a circuit $\mathcal{C}$ that solely consists of these gates. It is worth mentioning that if any of these gates may occur at most once in $\mathcal{C}$, as these gates commute and if one gate appears more than once in $\mathcal{C}$, they can be moved next to each other and be canceled out. The $\mathrm{C^{k}Z}$ gates for $k>1$ can in turn be decomposed to CZ and single qubit gates. As an example, the decomposition of $\mathrm{C^2Z}$ gate to CZ and single-qubit gates by the approach of~\cite{Shende09} is shown in Figure~\ref{fig:C2Z}.
Using the equation $HXH=Z$, a $\mathrm{C^kZ}$ gate can be easily replaced by a $\mathrm{C^kNOT}$ gate in the middle and two Hadamard gates which act on target qubit, as shown in Figure~\ref{fig:czcnot}.
Similarly, $\mathrm{C^kNOT}$ gates can be replaced by $\mathrm{C^kZ}$ gates at the cost of inserting two Hadamard gates.
\begin{figure*}
\scriptsize
\centering
\[
\Qcircuit @C=1.0em @R=0.7em {
&\ctrl{1} & \qw & & & \qw & \qw &\qw &\ctrl{2} &\qw &\qw &\qw &\ctrl{2} &\qw &\ctrl{1} &\gate{T} &\ctrl{1} &\qw &\qw \\
&\ctrl{1} & \qw &=& & \qw & \ctrl{1} & \qw & \qw &\qw &\ctrl{1} &\qw &\qw &\gate{TH} & \ctrl{0} & \gate{HT^\dag H}&\ctrl{0} & \gate{H}&\qw\\
&\ctrl{0} & \qw & & &\gate{H} &\ctrl{0}&\gate{HT^\dag H} &\ctrl{-2}&\gate{HTH}&\ctrl{-1}&\gate{HT^\dag H}&\ctrl{-2}&\gate{HT}&\qw &\qw &\qw &\qw &\qw
}
\]
\caption{The decomposition of $\mathrm{C^2Z}$ gate to CZ and single-qubit gates~\cite{Shende09}.}
\label{fig:C2Z}
\end{figure*}
\begin{figure}
\centering
\[
\Qcircuit @C=0.8em @R=0.0000001em @!R{
&k&&&/\qw &\ctrl{2} &\qw & \qw &&&k&& &/\qw &\ctrl{2} &\qw &\qw \\
& &&& & & & & &&= && &&& & & & & \\
& &&&\qw &\ctrl{0} &\qw &\qw & && & & &\gate{H} &\targ &\gate{H} &\qw
}
\]
\caption{Circuit equivalence of $\mathrm{C^kZ}$ gates and $\mathrm{C^kNOT}$ gates.}
\label{fig:czcnot}
\end{figure}
\@startsection {section}{1}{\z@}{10pt plus 2pt minus 2pt{Proposed approach}
\label{sec:proposed}
In this section, some definitions and notations are introduced before proceeding to the main part of the paper.
\@startsection {subsection}{2}{\z@}{10pt plus 2pt minus 2pt{Definitions and Notations}
\label{sec:def}
Consider $\mathbb{D}_n=\{d_n, n \in N\}$ where each $d_n$ is a diagonal Hermitian matrix on $n$ qubits.
For a given $d_n$, its $i^{th}$ diagonal element is denoted as $\lambda_i$ for $0\leq i \leq 2^{n}-1$. Based on the definition of $d_n$, each element of it is either +1 or -1. Without the loss of generality, it can be assumed that the first element of a given $d_n$, i.e., $\lambda_0=1$. If $\lambda_0=-1$, one can substitute $d_n$ with $-d_n$.
This new matrix is equivalent to the old one up to a global phase. As a result, $2^{n}-1$ independent diagonal elements of $d_n$, i.e., $\lambda_1,...,\lambda_{2^n-1}$, are left.
The following two functions will be used later in the proposed approach.
\begin{align*}
b=\text{Binary}(d_n)
\,,d_n=\text{DiagMat}(b)
\end{align*}
The first function, takes a matrix ($d_n$) and returns a corresponding $(2^n-1)$-bit binary number $b$. To this end, the elements with the +1 and -1 values of $d_n$ are first replaced by 0 and 1 respectively and then the sequence of the diagonal elements of this matrix (excluding $\lambda_0$) are written in $b$ from right to left. For example, $b=(1, 0, 0)=\text{Binary(CZ)}$. We call $b$ the binary representation of $d_n$. The second function, $d_n=\text{DiagMat}(b)$ is the reverse of the previous one which takes a binary number $b$ and returns the corresponding matrix. For example, CZ=DiagMat(1, 0, 0).
When a number of $d_n$ gates are cascaded, a matrix is produced whose $i^{th}$ diagonal element is obtained by the multiplication of all of the $i^{th}$ diagonal elements of these gates. It can be readily verified that the multiplication operation on the set $\{ - 1,1\}$ is equivalent to XOR operation on the set $\{ 1,0\}$
where algebric value -1 corresponds to binary value 1 and algebraic value 1 corresponds to binary value 0.
Consider $z=\mathrm{XOR}_{1\leq i \leq k}(x^i)$ a function which takes $k$ $n$-bit binary numbers, $x^i$, $1\leq i \leq k$ and returns a binary number $z$ where $z_j$, i.e., the $j^{th}$ bit of $z$, $0\leq j\leq n-1$ is computed by xoring all of the $j^{th}$ bits of $x^i$.
Therefore, in order to compute the multiplication of $d_n^i$ gates, $1\leq i \leq k$, the following equation is used:
\begin{equation}
\prod\nolimits_{i = 1}^k {d_n^i } = \text{DiagMat(XOR}_{1 \le i \le k} (\text{Binary}(d_n^i)))
\label{eq:eq1}
\end{equation}
The number of different $\mathrm{C^kZ}$ gates for $0 \leq k \leq n-1$ which act on $n$ qubits is equal to $2^{n}-1$.
To prove this, consider that there are $\binom{n}{1}$ possible $\mathrm{C^0Z}$ gates, $\binom{n}{2}$ possible $\mathrm{C^1Z}$ gates, $\binom{n}{3}$ possible $\mathrm{C^{2}Z}$ gates, ..., and $\binom{n}{n}$ possible $\mathrm{C^{n-1}Z}$ gates in the same way. On the other hand, the number of $k$-combinations for all $k$ equals to the number of subsets of a set with $n$ members:
\[\binom{n}{0}+\binom{n}{1}+\binom{n}{2}+...+\binom{n}{n}=2^{n}\]
Therefore, the number of the different $\mathrm{C^kZ}$ gates is equal to:
\begin{equation}
\binom{n}{1}+\binom{n}{2}+...+\binom{n}{n}=2^{n}-1.\nonumber
\label{eq:num}
\end{equation}
In the rest of the paper, to specify different $\mathrm{C^kZ}$ gates which act on $n$ qubits, they are indexed.
To this end, first the qubits in an $n$-qubit circuit are numbered from 0 to $n$-1 from top to bottom and for each $\mathrm{C^kZ}$ gate a corresponding $n$-bit binary string, denoted by $\mathbf{b}$ can be produced. The $j^{th}$ bit of $\mathbf{b}$ from right, $0\leq j \leq n-1$ is 1, if $\mathrm{C^kZ}$ is applied on the $j^{th}$ qubit of the circuit and is 0, otherwise. Then, the $\mathrm{C^kZ}$ gate is shown by $i$, the decimal value of $\mathbf{b}$, which ranges from 1 to $2^n-1$, as $\mathbf{CZ}_i$. For example, in a two-qubit circuit, the binary strings of the gates $Z\otimes I$, $I\otimes Z$, and CZ are produced as $01$, $10$ and $11$ and they are denoted as $\mathbf{CZ}_1$, $\mathbf{CZ}_2$ and $\mathbf{CZ}_3$, respectively.
\@startsection {subsection}{2}{\z@}{10pt plus 2pt minus 2pt{\label{sec:decom}Decomposition of $\mathbb{D}_n$ Gates}
In this section, we propose an approach to decompose a given $d_n$ matrix to $\mathbf{CZ}_i$ gates for $1\leq i\leq 2^n-1$.
\begin{lemma}
\label{lemma1}
The set of $\mathrm{{C^{k}Z}}$ gates for $k\geq0$ are independent, i.e., none of them can be decomposed to a combination of the other $\mathrm{{C^{k}Z}}$ gates.
\end{lemma}
\begin{proofs}
The proof is done by contradiction. Suppose a gate $\mathrm{C^{k}Z}$ can be decomposed to a set of gates that may consist of $\mathrm{C^{0}Z, C^1Z, ..., C^{k-1}Z}$ gates. This set of gates produces a circuit which is called $\mathcal{C}$. Now, consider $\mathrm{C^{m}Z}$ a gate which has the smallest number of control qubits (i.e., $m$) among all these gates.
Let the qubits on which $\mathrm{C^{m}Z}$ gate acts be $|1\rangle$ and the remaining input qubits be $|0\rangle$. As other gates in $\mathcal{C}$ have more control lines than $\mathrm{C^{m}Z}$ gate and at least one of them is in the state $|0\rangle$, none of them activates. As $\mathrm{C^{m}Z}$ gate acts, the output of $\mathcal{C}$ is the negative of its inputs. On the other hand, $\mathrm{C^{k}Z}$ gate has no effect on its input because at least one of its control lines is in the $|0\rangle$ state. Therefore, there is at least one input state for which the outputs of $\mathcal{C}$ and $\mathrm{C^{k}Z}$ gate differ which is a contradiction. Hence, $\mathrm{C^{k}Z}$ gates for $k\geq0$ are independent.
\end{proofs}
\begin{theorem}
Every $d_n$ gate can be decomposed to a combination of $\mathbf{CZ}_i$ gates for $1 \leq i \leq 2^n-1$.
\end{theorem}
\begin{proofs}
The Binary function as introduced in Section~\ref{sec:def} returns a $(2^n-1)$-bit binary number. On the other hand, there are $2^n-1$ independent $\mathbf{CZ}_i$ gates, as shown in Lemma~\ref{lemma1}. Therefore, the binary representations of these $2^n-1$ gates are also independent and form a basis for the vector space produced by $(2^n-1)$-bit binary numbers when their combination is performed by the XOR operator. Therefore, each diagonal Hermitian quantum gate can be written as a combination of $\mathbf{CZ}_i$ gates where their multiplication is performed using Equation~$\left(\ref{eq:eq1}\right)$. These gates may have a coefficient of 1 if they exist and a coefficient of 0 if they do not exist in the decomposed circuit. In other words, the following equation should be used:
\begin{equation}
\text{Binary}(d_n)=
\text{XOR}_{1 \le i \le 2^n-1 }(a_i.\text{Binary}(\textbf{CZ}_i))
\label{eq:eqset}
\end{equation}
where ${a_i}^,s$ are binary variables and dot represents the logical AND operator which is applied on $a_i$ and each bit of the binary representation of $\textbf{CZ}_i$. It is worth mentioning that as we had assumed that the first elements of the diagonal Hermitian matrices are always +1 and the first element of all $\mathbf{CZ}_i$ gates are +1, the equation corresponding to them always satisfies.
By solving Equation~(\ref{eq:eqset}), $a_i^,s$ can be found.
\end{proofs}
Example 1 clarifies the discussion.\\
\textbf{Example 1.} Suppose a two-qubit diagonal Hermitian matrix $A=|0\rangle \langle 0|+|1\rangle \langle 1|-|2\rangle \langle 2|+|3\rangle \langle 3|$ is given. The function Binary(A) returns (0,1,0). In order to apply the proposed approach, the gates $\mathbf{CZ}_1$, $\mathbf{CZ}_2$ and $\mathbf{CZ}_3$ are considered. (1, 1, 0)=Binary($\mathbf{CZ}_1$), (1, 0, 1)=Binary($\mathbf{CZ}_2$) and (1, 0, 0)=Binary($\mathbf{CZ}_3$). We have the following set of equations:
\begin{equation}
a_{1}.(1, 1, 0)\oplus a_{2}.(1, 0, 1) \oplus a_{3}.(1, 0, 0)=(0, 1, 0)\nonumber
\end{equation}
where $a_i^,s$ are binary variables.
Therefore, the following equation set must be solved:
\begin{equation}
\left\{{\begin{array}{*{20}c}
{a_1.(1)\oplus a_2.(1) \oplus a_3.(1) = 0}\\
{a_1.(1)\oplus a_2.(0) \oplus a_3.(0) = 1}\\
{a_1.(0)\oplus a_2.(1) \oplus a_3.(0) = 0}\\
\end{array}}\right.
\label{eq:set}
\end{equation}
By solving Equation~(\ref{eq:set}), $a_1$, $a_2$ and $a_3$ are found as 1, 0 and 1, respectively. This solution set implies that the gates $\mathbf{CZ}_1$ and $\mathbf{CZ}_3$ exist and $\mathbf{CZ}_2$ does not exist in the synthesized circuit, respectively. The circuit that implements the $A$ gate is shown in Figure~\ref{fig:fig2}.
\begin{figure}
\centering
\[
\Qcircuit @C=1.0em @R=.7em {
&\ctrl{1} &\gate{Z}&\qw \\
& \ctrl{-1} &\qw & \qw
}
\]
\caption{The circuit that implements matrix $A=|0\rangle \langle 0|-|1\rangle \langle 1|-|2\rangle \langle 2|+|3\rangle \langle 3|$.}
\label{fig:fig2}
\end{figure}
\@startsection {subsection}{2}{\z@}{10pt plus 2pt minus 2pt{Cost Analysis}
\label{sec:cost}
To decompose a given $d_n$ gate, the proposed approach in the worst case produces a circuit that consists of all possible $2^n-1$ $\mathbf{CZ}_i$ gates (as explained in Section~\ref{sec:def}). Without using any ancilla qubits, $\mathrm{C^{n-1}Z}$ gates can be decomposed to $O(n^2)$ elementary gates based on the results of~\cite[Corollary 7.6]{Barenco95}. Therefore, the cost complexity of the proposed approach in the worst case is $O(n^22^n)$ in terms of elementary gates.
On the other hand, as mentioned in Section~\ref{sec:ckz}, the $\mathrm{C^kZ}$ gates can be replaced by $\mathrm{C^kNOT}$ gates at the cost of inserting two Hadamard gates~\cite{Shende09} and there are many studies on the efficient decomposition of $\mathrm{C^kNOT}$ gates in the literature. For example, in~\cite{Barenco95} it is shown that in a circuit on $n$ qubits, using one ancilla, a $\mathrm{C^{n-2}NOT}$ gate, $n\ge7$ can be decomposed to $8(n-2)$ Toffoli gates and using $m-2$ ancilla qubits, $m \in \{3,4,..,\lceil n/2 \rceil\}$, $n\ge5$, a $\mathrm{C^mNOT}$ gate can be decomposed to $4(m-2)$ Toffoli gates. These studies can be applied to decrease the worst case cost complexity of the proposed approach.
\@startsection {section}{1}{\z@}{10pt plus 2pt minus 2pt{Experimental Results} \label{sec:results}
As there are $2^n-1$ independent diagonal elements in a $\mathbb{D}_n$ matrix, each of which can take two distinct values, the number of all possible $\mathbb{D}_n$ gates is $2^{2^n-1}$. For example, the numbers of all possible $\mathbb{D}_2$, $\mathbb{D}_3$ and $\mathbb{D}_4$ gates are 8, 128 and 32678, respectively.
The proposed approach and the approach of~\cite{bullock2004} for decomposing $\mathbb{D}_i$ gates for $i=2,3,4$ were implemented in MATLAB and MAPLE, respectively on a workstation with 4GB RAM and Core i5 2.40GHz CPU.
Table~\ref{table} compares the results of the proposed approach and the approach of~\cite{bullock2004} to decompose some members of $\mathbb{D}_i$ for $i=2, 3, 4$.
In this table, each matrix is denoted by a number equal to the decimal value of its binary representation. For example, $(1,0,0)$=CZ is denoted by 4. The second column of the table contains the indices of the produced basis state in the proposed approach (i.e., $\mathbf{CZ}_i$ gates), as obtained by the procedure explained in Section~\ref{sec:proposed}. The angles of the produced $R_z(\alpha)$ gates for the approach of~\cite{bullock2004}, according to the circuit structure of Figure~\ref{fig:previous}, are determined in the third column.
The numbers of required CZ and single-qubit gates in terms of rotation gates around $y$ and $x$ axes for each circuit are reported and compared for these two approaches. Possible optimizations are performed on the circuits produced by the approach of~\cite{bullock2004}. The optimizations of the CNOT gates for this approach are performed by cancelling CNOT gates whenever possible using the transformation rules of these gates~\cite{iwama2002transformation}. The utilized transformation rules are cancelling two adjacent identical CNOT gates and changing the order of independent CNOT gates. Moreover, by using the equation of $H^2=I$, whenever two adjacent Hadamard gates appear in the circuit, they cancel and therefore the number of single-qubit gates can be decreased.
In this paper, the library of CZ and rotation gates around $y$ and $x$ axes is considered. To decompose the circuits based on the previously known elementary library which includes CNOT gate, CZ gates can be readily replaced by a middle CNOT gate and two Hadaramd gates which act on the target qubit (as shown in Figure~\ref{fig:czcnot}). After this replacement, the circuits should be further optimized by cancelling two adjacent Hadamard gates.
\begin{figure}
\centering
\scriptsize
\[
\Qcircuit @C=0.7em @R=0.5em @!R {
&\qw \gategroup{7}{10}{8}{14}{.9em}{--} & \qw &\qw &\qw &\qw &\qw &\qw &\ctrl{3} &\qw & \qw &\qw &\qw &\qw &\qw &\qw &\ctrl{3} &\qw\\
&\qw & \qw &\qw &\ctrl{2} &\qw &\qw &\qw &\qw &\qw & \qw &\qw &\ctrl{2} &\qw &\qw &\qw &\qw &\qw\\
&\qw &\ctrl{1} &\qw &\qw &\qw &\ctrl{1} &\qw &\qw &\qw &\ctrl{1} &\qw &\qw &\qw &\ctrl{1} &\qw &\qw &\qw\\
&\gate{R_z(o)H} &\ctrl{0} &\gate{H R_z(n)H} &\ctrl{0} &\gate{H R_z(m) H} &\ctrl{0} &\gate{H R_z(l) H} &\ctrl{0} &\gate{HR_z(k)H} &\ctrl{0} &\gate{H R_z(j)H} &\ctrl{0} &\gate{H R_z(i) H} &\ctrl{0} &\gate{H R_z(h) H} &\ctrl{0} &\qw\\
& & & & & & & & & & & & & & & & & &\\
& & & & & & & & & & & & & & & & & &\\
&\qw \gategroup{6}{1}{9}{15}{2.3em}{..} & \qw &\qw &\ctrl{2} &\qw &\qw &\qw &\ctrl{2} &\qw &\ctrl{1} &\qw &\ctrl{1} &\gate{R_z(a)} &\qw & & & &\\
&\qw & \ctrl{1} &\qw &\qw &\qw &\ctrl{1} &\qw &\qw &\gate{R_z(c)H} & \ctrl{0} &\gate{H R_z(b)H} &\ctrl{0} &\gate{H} &\qw & & & &\\
&\gate{R_z(g)H} &\ctrl{0} &\gate{H R_z(f)H} &\ctrl{0} &\gate{H R_z(e) H} &\ctrl{0} &\gate{H R_z(d) H} &\ctrl{0} &\gate{H} &\qw &\qw &\qw &\qw &\qw & & & &\\
&\gate{H} &\qw &\qw &\qw &\qw &\qw &\qw &\qw &\qw &\qw &\qw &\qw &\qw &\qw & & & &
}
\]
\caption{The general circuit structure to implement four-qubit diagonal gates by~\cite{bullock2004} using CZ and single-qubit gates, up to a global phase. The dashed and the dotted parts of the figure can be assumed as general circuit structures for two and three-qubit diagonal gates, respectively. The circuit of the bottom is the continuation of the top circuit.}
\label{fig:previous}
\end{figure}
The experimental results on all possible $\mathbb{D}_2$, $\mathbb{D}_3$ and $\mathbb{D}_4$ gates show that the proposed approach improves the cost of CZ and single-qubit gates of the method of~\cite{bullock2004} by $37.4\%$, $34.3\%$, $10.9\%$, $31\%$, $4.6\%$ and $24.7\%$, respectively on average. As can be seen, when the number of qubits increases, the average improvement decreases. However, as Table~\ref{table} shows, with the increase of the number of qubits, there are always some cases where the proposed approach leads to significantly better results.
\begin{center}
\scriptsize
\begin{longtable}{|l||c|l|c|c|l|c|c|c|c|}%
\caption{The comparison of obtained circuits by the proposed approach and the method of [Bullock and Markov 2004] to decompose some members of $\mathbb{D}_2$, $\mathbb{D}_3$ and $\mathbb{D}_4$ gates. $a,b,...,0$ are the angles of $R_z(\alpha)$ gates of Figure~\ref{fig:previous}. The single qubit gates are of types $R_x$ and $R_y$.}\label{table}\\
\hline\multirow{2}{*}{$n$} &\multirow{2}{*}{Matrix} &\multicolumn{3}{c|}{The Proposed Approach} &\multicolumn{3}{c|}{\cite{bullock2004}} &\multicolumn{2}{c|}{Imp. (\%)} \\
& &Basis &\#CZ &\#1qu &$a,b,c$ &\#CZ &\#1qu &\#CZ &\#1qu\\
\hline\hline\endfirsthead
\multicolumn{10}{c}%
{{\bfseries \tablename\ \Roman{table}{} -- continued from previous page}} \\
\hline\multirow{2}{*}{$n$} &\multirow{2}{*}{Matrix} &\multicolumn{3}{c|}{The Proposed Approach} &\multicolumn{3}{c|}{\cite{bullock2004}} &\multicolumn{2}{c|}{Imp. (\%)} \\
& &Basis &\#CZ &\#1qu &$a,b,c,d,e,f,g,h,i,j,k,l,m,n,o$ &\#CZ &\#1qu &\#CZ &\#1qu\\
\hline\hline\endhead
\hline\hline\multicolumn{10}{|c|}{Continued on next page} \\
\hline\endfoot
\hline\endlastfoot
2 &1 &2,3 &1 &3 &$\frac{-\pi}{2}, \frac{\pi}{2}, \frac{\pi}{2}$ &2 &12 &50 &75\\
&2 &1,3 &1 &3 &$\frac{\pi}{2}, \frac{\pi}{2}, \frac{-\pi}{2}$ &2 &12 &50 &75\\
&3 &1,2 &0 &6 &$0, \pi, 0$ &2 &9 &100 &33.3\\
&4 &3 &1 &0 &$\frac{\pi}{2}, \frac{-\pi}{2}, \frac{\pi}{2}$ &2 &12 &50 &100\\
&7 &1,2,3 &1 &6 &$\frac{\pi}{2}, \frac{\pi}{2}, \frac{\pi}{2}$ &2 &12 &50 &50\\\hline\hline
\hline\hline
\multirow{2}{*}{$n$} &\multirow{2}{*}{Matrix} &\multicolumn{3}{c|}{The Proposed Approach} &\multicolumn{3}{c|}{\cite{bullock2004}} &\multicolumn{2}{c|}{Imp. (\%)} \\
& &Basis &\#CZ &\#1qu &$a,b,c,d,e,f,g$ &\#CZ &\#1qu &\#CZ &\#1qu\\\hline\hline
3 &15 &1,2,4,6 &1 &9 &$\frac{-\pi}{2}, \frac{\pi}{2}, 0, \frac{\pi}{2}, \frac{\pi}{2}, 0, 0 $ &6 &24 &83.3 &62.5 \\
&18 &2,3,5,6 &3 &3 &$0, \frac{\pi}{2}, 0, \frac{-\pi}{2}, 0, \frac{\pi}{2}, 0 $ &6 &21 &50 &85.7 \\
&20 &5,6 &2 &0 &$0, \frac{\pi}{2}, 0, 0, \frac{-\pi}{2}, 0, \frac{\pi}{2} $ &6 &18 &66.6 &100 \\
&27 &1,2,4,5 &1 &9 &$0, \frac{\pi}{2}, \frac{-\pi}{2}, 0, \frac{\pi}{2}, \frac{\pi}{2}, 0 $ &6 &21 &83.3 &57.1 \\
&45 &1,4 &0 &6 &$0, 0, 0, \pi, 0, 0, 0 $ &2 &9 &100 &33.3 \\
&51 &2,4 &0 &6 &$0, 0, 0, 0, 0, \pi, 0 $ &2 &9 &100 &33.3 \\
&54 &2,5 &1 &3 &$0, \frac{\pi}{2}, \frac{\pi}{2}, 0, \frac{-\pi}{2}, \frac{\pi}{2}, 0 $ &6 &21 &83.3 &85.7 \\
&65 &4,5,6 &2 &3 &$0, \frac{-\pi}{2}, 0, 0, \frac{\pi}{2}, 0, \frac{\pi}{2} $ &6 &18 &66.6 &83.3 \\
&99 &2,4,5 &1 &6 &$0, \frac{-\pi}{2}, \frac{\pi}{2}, 0, \frac{\pi}{2}, \frac{\pi}{2}, 0 $ &6 &21 &83.3 &71.4 \\
&113 &3,4,6 &2 &3 &$\frac{\pi}{2}, \frac{-\pi}{2}, 0, 0, 0, \frac{\pi}{2}, \frac{\pi}{2} $ &4 &21 &50 &85.7 \\
\hline\hline
\multirow{2}{*}{$n$} &\multirow{2}{*}{Matrix} &\multicolumn{3}{c|}{The Proposed Approach} &\multicolumn{3}{c|}{\cite{bullock2004}} &\multicolumn{2}{c|}{Imp. (\%)} \\
& &Basis &\#CZ &\#1qu &$a,b,c,d,e,f,g,h,i,j,k,l,m,n,o$ &\#CZ &\#1qu &\#CZ &\#1qu\\\hline\hline
4 &4680 &3,6,9,12 &4 &0 &$ 0, 0, 0, \frac{\pi}{2}, 0, 0, 0, 0, 0, \frac{-\pi}{2}, 0, \frac{\pi}{2}, 0, 0, 0 $&8 &15 &50 &100 \\
&10376 &11,12 &7 &27 &$ 0, 0, 0, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, 0, \frac{-\pi}{4}, \frac{\pi}{4}, 0, 0, \frac{-\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{2} $&10 &33 &30 &18.1 \\
&14602 &1,5,8,10 &2 &6 &$ \frac{\pi}{4}, \frac{-\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4} $&14 &54 &85.7 &88.8 \\
&21760 &1,9 &1 &3 &$ \frac{\pi}{2}, 0, 0, 0, 0, 0, 0, \frac{\pi}{2}, 0, 0, 0, 0, 0, 0, \frac{-\pi}{2} $&2 &9 &50 &66.6 \\
&23280 &2,9 &1 &3 &$ 0, \frac{\pi}{2}, \frac{\pi}{2}, 0, 0, 0, 0, 0, 0, 0, \frac{-\pi}{2}, \frac{\pi}{2}, 0, 0, 0 $&6 &18 &83.3 &83.3 \\
&24428 &1,4,5,10 &2 &6 &$ \frac{\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4} $&14 &54 &85.7 &88.8 \\
&27030 &1,2,4,8 &0 &12 &$ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \pi, 0, 0, 0 $&6 &6 &100 &-100 \\
&38460 &2,4,9 &1 &6 &$ 0, 0, 0, 0, \frac{\pi}{2}, \frac{\pi}{2}, 0, 0, 0, \frac{-\pi}{2}, 0, 0, \frac{\pi}{2}, 0, 0 $&10 &24 &90 &75 \\
&40044 &3,4,10 &2 &3 &$ 0, 0, 0, 0, 0, \frac{\pi}{2}, \frac{\pi}{2}, 0, \frac{\pi}{2}, \frac{-\pi}{2}, 0, 0, 0, 0, 0 $&8 &40 &75 &92.5 \\
&43520 &9 &1 &0 &$ \frac{\pi}{2}, 0, 0, 0, 0, 0, 0, \frac{-\pi}{2}, 0, 0, 0, 0, 0, 0, \frac{\pi}{2} $&2 &9 &50 &100 \\
&49258 &9,6,8 &2 &3 &$ \frac{-\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4} $&14 &54 &85.7 &94.4 \\
&51884 &4,5,6,9,10 &4 &3 &$ 0, 0, 0, 0, \frac{\pi}{2}, 0, \frac{\pi}{2}, 0, 0, 0, \frac{-\pi}{2}, 0, 0, 0, \frac{\pi}{2} $&8 &15 &50 &80 \\
&63120 &2,5,6,9,10 &4 &3 &$ \frac{\pi}{2}, 0, \frac{\pi}{2}, 0, 0, 0, 0, 0, \frac{-\pi}{2}, 0, 0, 0, \frac{\pi}{2}, 0, 0 $&6 &18 &33.3 &83.3 \\
&63916 &1,4,6,9,10 &3 &6 &$ \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{-\pi}{4}, \frac{-\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4} $&14 &54 &78.5 &88.8 \\
&64598 &2,4,6,8,9 &2 &9 &$ \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{\pi}{4}, \frac{-\pi}{4} $&14 &54 &85.7 &83.3 \\
\end{longtable}
\end{center}
\@startsection {section}{1}{\z@}{10pt plus 2pt minus 2pt{Conclusions}
\label{sec:conc}
In this paper, we presented a new approach to decompose $n$-qubit diagonal Hermitian gates, denoted by $\mathbb{D}_n$. First, a function was defined that takes the diagonal elements of a specific $n$-qubit diagonal Hermitian quantum gate, $d_n$, and returns a binary representation for it. Then we showed that $\mathbb{D}_n$ gates can be decomposed to a circuit that is solely constructed of $\mathrm{C^kZ}$ gates for $0\leq k\leq n-1$. Moreover, it was proved that the binary representations of this set of gates, form a basis for the vector space that is produced by the binary representations of $\mathbb{D}_n$ gates, denoted by $b$, where their combination is performed using logical XOR gates. After that, it was shown that the decomposition problem of $\mathbb{D}_n$ gates can be mapped to writing $b$ in that basis where the coefficients of 1 and 0 imply that the corresponding basis state exists and does not exists in the produced circuit, respectively. Experimental results on all possible two, three and four-qubit diagonal Hermitian gates show that the proposed approach improves the cost of CZ and single-qubit gates of best previous method for synthesizing arbitrary diagonal gates by $37.4\%$, $34.3\%$, $10.9\%$, $31\%$, $4.6\%$ and $24.7\%$, respectively on average. As a future work, applying some post processing after the proposed decomposition to optimize the produced circuits is being considered. Moreover, we are seeking more general unitary matrices to which the proposed perspective can be applied.
\footnotesize
|
\section{\label{I}Introduction}
Taylor's seminal paper on the swimming of microscopic organisms (Taylor 1951) generated a wide range of studies of swimming at low Reynolds number (Shapere $\&$ Wilczek 1989a,b; Felderhof $\&$ Jones 1994a,b; Lauga $\&$ Powers 2009; Garstecki $\&$ Cieplak 2009; Swan et al. 2011). Taylor considered the swimming of a planar sheet with sinusoidal transverse undulation. Soon after, Lighthill (1952) published a similar calculation for a squirming sphere. The theory is based on Stokes' equations for flow velocity and pressure of a viscous incompressible fluid (Happel $\&$ Brenner 1973). A no-slip boundary condition applies at the interface of body and fluid. The swimming speed and the rate of dissipation are calculated in perturbation theory to second order in the surface displacements.
Taylor's calculation is easily extended to the swimming of a planar slab. The independent distortions of the upper and lower surface of the slab allows a wider class of motions. Taylor's model is recovered in the limit of a thin slab with parallel motion of the two surfaces.
Taylor's calculation was extended by Reynolds (1965), Tuck (1968), Blake (1971), and Childress (1981) to include squirming motion of the planar sheet. We show in the following that their calculation of the swimming velocity was incorrect. For the more general motion the second order flow velocity differs on the two sides of the sheet, and this was not taken into account appropriately. In Sec. 3 of this article we consider the sheet with a superposition of undulating and squirming distortions in detail. For pure undulation and pure squirming there is no problem, and the sheet has a well-defined swimming velocity. However, for the superposition of the two modes of flow the second order velocity shows a discontinuity at the undisplaced sheet due to interference of flows with different symmetry on the two sides. The discontinuity was discussed by Reynolds (1965), but he interpreted the situation as a pumping mechanism.
We resolve the difficulty by considering a finite system, in particular a circular disk. It then becomes evident that the difference in surface velocity on either side corresponds to a torque exerted on the fluid. The torque must be canceled by adding a Stokes flow with no-slip boundary condition on the disk but with rotation of the fluid at infinity (Felderhof $\&$ Jones 1994a). Hence one infers a net rotational swimming velocity in the laboratory frame.
Subsequently we study the mean translational and rotational swimming of a circular disk in more detail. Since both sides of the disk can have independent surface distortions, a wide range of flow situations is possible. For the optimal stroke the efficiency, defined from the ratio of speed and power, is a factor $\sqrt{2}$ larger than for Taylor's undulating sheet.
For the optimal stroke the mean rotational swimming velocity vanishes. For the general stroke the combined translational and rotational swimming leads to a circular orbit. At any point the disk is perpendicular to the plane of the circle and tangential to the circle. Consequently, the possibility of slightly different strokes on the two sides of its body allows a swimming flat microorganism (Opalina or Paramecium) to change direction.
\section{\label{2}Taylor's swimming sheet}
We consider a planar slab immersed in a viscous
incompressible fluid of shear viscosity $\eta$. The slab at rest is bounded by two planes at distance $2d$. We choose Cartesian coordinates $x,y,z$ such that the upper plane is at $y=d$ and the lower at $y=-d$. The rest shape of the slab is denoted as $S_0$. We shall consider a prescribed time-dependent shape $S(t)$ leading to swimming motion in the $x$ direction. The distortions are decomposed into $S_+(t)$ for the upper surface and $S_-(t)$ for the lower surface. The fluid is set in motion by the time-dependent distortions of the
slab. At low Reynolds
number and on a slow time scale the flow velocity
$\vc{v}(\vc{r},t)$ and the pressure $p(\vc{r},t)$ in the rest frame satisfy the
Stokes equations
\begin{equation}
\label{2.1}\eta\nabla^2\vc{v}-\nabla p=0,\qquad\nabla\cdot\vc{v}=0.
\end{equation}
The surface displacement
$\vc{\xi}(\vc{s},t)$ is defined as the vector distance
\begin{equation}
\label{2.2}\vc{\xi}=\vc{s}'-\vc{s}
\end{equation}
of a point $\vc{s}'$ on the displaced surface $S(t)$ from the
point $\vc{s}$ on the slab $S_0$. We decompose $\vc{s}$ into $\vc{s}_+=(x,d,z)$ for the upper surface and $\vc{s}_-=(x,-d,z)$ for the lower surface. The fluid
velocity $\vc{v}(\vc{r},t)$ is required to satisfy
\begin{equation}
\label{2.3}\vc{v}(\vc{s}+\vc{\xi}(\vc{s},t))=\frac{\partial\vc{\xi}(\vc{s},t)}{\partial t}.
\end{equation}
This amounts to a no-slip boundary condition. The displacement is decomposed into $\vc{\xi}_\pm(t)$ for the upper and lower surface. We consider displacements which do not depend on $z$ and take the form $\vc{\xi}_\pm=(\xi_{\pm x}(x,t),\xi_{\pm y}(x,t),0)$. As a consequence the flow velocity $\vc{v}$ and pressure $p$ do not depend on $z$, and the problem is effectively two-dimensional.
First we recall Taylor's well-known calculation of the swimming of a thin planar sheet due to a running wave of surface undulations (Taylor 1951). The sheet can be regarded as the limiting case of a slab with $d\rightarrow 0$ and $S_+(t)$ and $S_-(t)$ moving in parallel so that $\vc{\xi}_+(t)=\vc{\xi}_-(t)$.
Taylor considered transverse oscillations of the sheet with surface distortions described by the running wave
\begin{equation}
\label{2.4}\vc{\xi}(x,t)=A\sin(kx-\omega t)\vc{e}_y,
\end{equation}
with amplitude $A$, positive wavenumber $k$, and positive frequency $\omega$, so that the wave propagates in the positive $x$ direction. He derived the flow pattern from a stream function $\psi(x,y,t)$, defined such that the components of velocity are
\begin{equation}
\label{2.5}v_x=u=\frac{\partial\psi}{\partial y},\qquad v_y=v=-\frac{\partial\psi}{\partial x}.
\end{equation}
(We use a different sign convention.) Stokes' equations Eq. (1) are satisfied provided that the stream function satisfies the biharmonic equation
\begin{equation}
\label{2.6}\nabla^4\psi=0.
\end{equation}
To terms linear in $A$ the stream function is found to be given by
\begin{equation}
\label{2.7}\psi_1=A\frac{\omega}{k}(1+k|y|)e^{-k|y|}\sin(kx-\omega t).
\end{equation}
The first order velocity components are
\begin{equation}
\label{2.8}u_1=-A\omega kye^{-k|y|}\sin(kx-\omega t),\qquad v_1=-A\omega (1+k|y|)e^{-k|y|}\cos(kx-\omega t).
\end{equation}
The corresponding pressure disturbance is given by
\begin{equation}
\label{2.9}p_1=\mp2A\eta\omega ke^{-k|y|}\cos(kx-\omega t),
\end{equation}
where the upper (lower) sign corresponds to the upper (lower) half-space. The velocity components satisfy the symmetry relations
\begin{equation}
\label{2.10}u_1(x,y,t)=-u_1(x,-y,t),\qquad
v_1(x,y,t)=v_1(x,-y,t).
\end{equation}
We call this mirror-inverse symmetry. It corresponds to reflection of the vector in the plane $y=0$ and inversion.
Taylor showed that to second order in the amplitude $A$ the sheet swims in the negative $x$ direction with speed
\begin{equation}
\label{2.11}|U_2|=\frac{1}{2}\omega kA^2.
\end{equation}
Alternatively this result can be derived from the observation that the second order time-averaged flow $\overline{\vc{v}}_2,\overline{p}_2$ must satisfy the Stokes equations Eq. (1) with the boundary condition (Felderhof $\&$ Jones 1994a)
\begin{equation}
\label{2.12}\overline{\vc{v}}_2(\vc{s})=-\overline{(\vc{\xi}\cdot\nabla)\vc{v}_1}\big|_{y=0},
\end{equation}
where the overline indicates averaging over a period $T=2\pi/\omega$. The expression on the right-hand side follows from the boundary condition Eq. (3). It takes the value $|U_2|$, so that Eq. (1) is solved by
\begin{equation}
\label{2.13}\overline{\vc{v}}_2(\vc{r})=|U_2|\vc{e}_x,\qquad\overline{p}_2=0.
\end{equation}
This implies that the swimming velocity of the sheet in the laboratory frame is $\vc{U}_2=-|U_2|\vc{e}_x$.
The required power equals the rate of dissipation of energy in the fluid. Taylor calculated this from the work done per unit area against the stress $\vc{\sigma}=\eta(\nabla\vc{v}+\widetilde{\nabla\vc{v}})-p\vc{I}$. The mean rate of dissipation per unit area is to second order (Felderhof $\&$ Jones 1994a)
\begin{equation}
\label{2.14}\overline{D}_2=-2\overline{\vc{v}_1\cdot\vc{\sigma}_1\cdot\vc{e}_y}\big|_{y=0}.
\end{equation}
The factor 2 arises from the fact that both sides of the sheet must be considered. It turns out that the viscous stress does not contribute, so that Eq. (14) reduces to
\begin{equation}
\label{2.15}\overline{D}_2=2\overline{\frac{\partial\xi_{y}}{\partial t}p_1}\big|_{y=0}.
\end{equation}
Substituting from Eqs. (4) and (9) one finds Taylor's result
\begin{equation}
\label{2.16}\overline{D}_2=2\eta\omega^2 kA^2.
\end{equation}
We define the dimensionless efficiency as
\begin{equation}
\label{2.17}E_2=4\eta\omega\frac{|U_2|}{\overline{D}_2}.
\end{equation}
In the present case this equals unity. We have shown elsewhere (Felderhof 2009) how the efficiency is affected by walls at $y=\pm L$.
Taylor's calculation can be generalized easily to a slab of thickness $2d$ with surface displacements $\vc{\xi}_{\pm}(x,t)$ of the upper and lower surface at $y=\pm d$, both given by Eq. (4). The flow pattern in the upper half-space is simply shifted by $d$ in the positive $y$ direction, and in the lower half-space by $d$ in the negative $y$ direction. In the above expressions $y$ must be replaced by $y\mp d$. The speed and the power are the same as for the infinitesimally thin sheet.
\section{\label{III}Undulating and squirming sheet}
Taylor's calculation for a thin planar sheet has been generalized to include squirming distortions, besides the transverse motions. Tuck (1968) called such motions longitudinal and calculated the corresponding swimming velocity, including corrections due to inertia. The motion is also included in the derivations of Reynolds (1965), Blake (1971), and Childress (1981). We begin the discussion by considering the squirming motion by itself.
We consider longitudinal oscillations of the sheet with surface distortions described by the running wave
\begin{equation}
\label{3.1}\vc{\xi}(x,t)=B\sin(kx-\omega t+\alpha)\vc{e}_x,
\end{equation}
with a phase shift $\alpha$ with respect to Eq. (4). We assume that $B$ is positive and consider values of $\alpha$ in the range $-\pi\leq\alpha\leq\pi$. The corresponding stream function is given by
\begin{equation}
\label{3.2}\psi_1=-B\omega ye^{-k|y|}\cos(kx-\omega t+\alpha).
\end{equation}
The first order velocity components are
\begin{equation}
\label{3.3}u_1=B\omega(-1+k|y|e^{-k|y|})\cos(kx-\omega t+\alpha),\qquad v_1=-B\omega kye^{-k|y|}\sin(kx-\omega t+\alpha).
\end{equation}
The corresponding pressure disturbance is given by
\begin{equation}
\label{3.4}p_1=-2B\eta\omega ke^{-k|y|}\sin(kx-\omega t+\alpha).
\end{equation}
The velocity components satisfy the symmetry relations
\begin{equation}
\label{3.5}u_1(x,y,t)=u_1(x,-y,t),\qquad
v_1(x,y,t)=-v_1(x,-y,t).
\end{equation}
We call this mirror symmetry.
For the linear combination of the displacements in Eqs. (4) and (18) we find for the velocity to be used in the second order boundary condition as in Eq. (12)
\begin{eqnarray}
\label{3.6}\vc{v}_2(\vc{s}_\pm,t)&=&-(\vc{\xi}\cdot\nabla)\vc{v}_1\big|_{y=0\pm}\nonumber\\
&=&\omega k\big[A^2 \sin^2\varphi-B^2\sin^2(\varphi+\alpha)\mp2AB\sin\varphi\cos(\varphi+\alpha)\big]\vc{e}_x,
\end{eqnarray}
where we have abbreviated
\begin{equation}
\label{3.7}\varphi=kx-\omega t.
\end{equation}
Thus in general the second order fluid velocity tends to different values on either side of the sheet.
Averaging over a period we find
\begin{equation}
\label{3.8}\overline{\vc{v}}_2(\vc{s}_\pm)=\frac{1}{2}\omega k\big[A^2-B^2\pm 2AB\sin\alpha\big]\vc{e}_x,
\end{equation}
with different values on both sides of the sheet if $\alpha>0$ and both $A$ and $B$ differ from zero. Blake (1971) and Childress (1981) explicitly consider only the upper side of the sheet, and arrive at the result for the swimming velocity of the sheet $\overline{\vc{U}}_{2BC}=\overline{U}_{2BC}\vc{e}_x$ with
\begin{equation}
\label{3.9}\overline{U}_{2BC}=\frac{1}{2}\omega k\big[-A^2+B^2-2AB\sin\alpha\big],
\end{equation}
(Blake considered the case $\alpha=\pi/2$ in his Eq. (40)). This would allow one to use the phase angle $\alpha$ to optimize the speed for fixed $A$ and $B$. However, the result is wrong, since in the derivation only the upper sign in Eq. (25) was considered.
The rate of dissipation turns out to be the same on both sides of the sheet. Calculating as in Eq. (14) one finds for the mean rate of dissipation per unit area
\begin{equation}
\label{3.10}\overline{D}_2=2\eta\omega^2 k(A^2+B^2),
\end{equation}
independent of the phase angle $\alpha$.
The solution of the second order flow problem for the infinite sheet is not obvious. In order to understand the physical situation we must regard the sheet as an idealization of the problem for a finite system.
\section{\label{IV}Circular disk}
Specifically we consider a circular disk of radius $a$ in the horizontal $x,z$ plane and of height $2d$ in the $y$ direction. The origin is taken at the center of the disk. We assume that the disk is thin with $d$ much smaller than $a$. In the theory of elasticity this would be called a plate (Green $\&$ Zerna 1968), and it would be surrounded by vacuum, rather than by a viscous incompressible fluid. We assume that the upper and lower surface perpendicular to the $y$ axis are distorted in phase with equal displacements on both sides of the disk,
\begin{equation}
\label{4.1}\vc{\xi}(\vc{s}_+,t)=\vc{\xi}(\vc{s}_-,t)=A\sin(kx-\omega t)\vc{e}_y+B\sin(kx-\omega t+\alpha)\vc{e}_x.
\end{equation}
The fluid moves on account of the no-slip boundary condition.
We consider distortions such that the wavelength $\lambda=2\pi/k$ is much larger than the thickness $2d$ and much smaller than the horizontal dimensions. The boundary conditions
\begin{equation}
\label{4.2}\vc{v}_2(\vc{r},t)\big|_{y=\pm d+}=\vc{v}_2(\vc{s}_\pm,t),
\end{equation}
with right-hand side given by Eq. (23), must be applied within the upper and lower circular areas $S_\pm$ parallel to the $xz$ plane.
Averaging over a period $T=2\pi/\omega$ this becomes
\begin{equation}
\label{4.3}\overline{\vc{v}}_2(\vc{s}_\pm)=g_\pm\vc{e}_x,\qquad \vc{s}_\pm\in S_\pm,
\end{equation}
with constants $g_\pm$ given by the factor in Eq. (25). The symmetric part is identified as minus the translational swimming velocity
\begin{equation}
\label{4.4}\overline{U}_2=-\frac{1}{2}(g_++g_-)=\frac{1}{2}\omega k(-A^2+B^2).
\end{equation}
The antisymmetric part $\frac{1}{2}(g_+-g_-)$ gives rise to a rotational swimming velocity. The difference in fluid velocity on both sides of the disk corresponds to a uniform force dipole layer. At large distances the flow in the external fluid tends to the flow generated by a point torque at the origin. From the force dipole layer the torque acting on the fluid is calculated as
\begin{equation}
\label{4.5}\vc{N}=-\pi\eta(g_+-g_-)a^2\vc{e}_z=-2\pi\eta\omega ka^2AB\sin\alpha\;\vc{e}_z.
\end{equation}
Note that for a uniform force dipole layer only the area covered is relevant, so that a similar calculation would work for an elliptical disk, for example.
At large distances the flow tends to
\begin{equation}
\label{4.6}\hat{\vc{v}}_2(\vc{r})\approx\frac{1}{8\pi\eta}\vc{N}\times\frac{\vc{r}}{r^3},
\qquad\hat{p}_2(\vc{r})\approx 0,\qquad\mathrm{as}\;r\rightarrow\infty.
\end{equation}
Outside the disk the flow velocity $\hat{\vc{v}}_2(\vc{r})$ satisfies Laplace's equation and the pressure $\hat{p}_2$ vanishes. The flow velocity is identical with the vector potential
of a magnet in the shape of the disk with magnetization in the plane of the disk.
In order to find the complete second order flow pattern at large distances we must add the Stokes flow which vanishes at the surface of the disk (Felderhof $\&$ Jones 1994a), and which cancels the term shown in Eq. (33). This flow has the asymptotic behavior
\begin{equation}
\label{4.7}\vc{v}^{\mathrm{St}}_2(\vc{r})\approx -\overline{\vc{\Omega}}_2\times\vc{r}-\frac{1}{8\pi\eta}\vc{N}\times\frac{\vc{r}}{r^3},
\qquad p^{\mathrm{St}}_2(\vc{r})\approx 0,\qquad\mathrm{as}\;r\rightarrow\infty,
\end{equation}
where $\overline{\vc{\Omega}}_2$ is related to $\vc{N}$ by (Happel $\&$ Brenner 1973)
\begin{equation}
\label{4.8}\vc{N}=-\frac{32}{3}\pi\eta a^3\overline{\vc{\Omega}}_2.
\end{equation}
The complete mean second order flow tends to
\begin{equation}
\label{4.9}\overline{\vc{v}}_2(\vc{r})\approx-\overline{\vc{\Omega}}_2\times\vc{r}-\overline{U}_2\vc{e}_x,
\qquad\overline{p}_2(\vc{r})\approx 0,\qquad\mathrm{as}\;r\rightarrow\infty.
\end{equation}
This identifies $\overline{\vc{\Omega}}_2$ as the mean rotational swimming velocity of the disk in the laboratory frame. From Eqs. (32) and (35) we find
\begin{equation}
\label{4.10}\overline{\vc{\Omega}}_2=\frac{3}{16a}\;\omega k AB\sin\alpha\;\vc{e}_z.
\end{equation}
This shows that on time average the disk rotates steadily about a diameter perpendicular to the direction of translational swimming. Since the translational swimming velocity is always parallel to the plane of the disk this implies that on a slow time scale the center of the disk runs through a circular orbit in the laboratory frame. The circle has radius $|U_2|/|\Omega_2|$ and is traversed in time $2\pi/|\Omega_2|$. At any point the disk is orthogonal and tangential to the circle.
\section{\label{V}Twofold strokes}
In this section we consider more complicated modes of motion in which the disk has different distortions on the two sides. We can idealize the disk as an infinite slab with surface displacements of the upper and lower surface given by
\begin{eqnarray}
\label{5.1}\vc{\xi}(\vc{s}_+,t)&=&A\sin(kx-\omega t)\vc{e}_y+B\sin(kx-\omega t+\alpha)\vc{e}_x,\nonumber\\
\vc{\xi}(\vc{s}_-,t)&=&C\sin(kx-\omega t+\beta)\vc{e}_y+D\sin(kx-\omega t+\beta+\gamma)\vc{e}_x.
\end{eqnarray}
The no-slip boundary condition Eq. (3) is imposed on both sides. The flow in the upper half-space is the same as before, and in the lower half-space it is modified only by amplitudes and phase shifts. Hence the mean second order flow velocity at the two surfaces becomes
\begin{eqnarray}
\label{5.2}\overline{\vc{v}}_2(\vc{s}_+)&=&\frac{1}{2}\omega k\big[A^2-B^2+2AB\sin\alpha\big]\vc{e}_x,\nonumber\\
\overline{\vc{v}}_2(\vc{s}_-)&=&\frac{1}{2}\omega k\big[C^2-D^2-2CD\sin\gamma\big]\vc{e}_x.
\end{eqnarray}
Therefore the mean translational swimming velocity of the disk is
\begin{equation}
\label{5.3}\overline{\vc{U}}_2=\frac{1}{4}\omega k\big[-A^2+B^2-C^2+D^2-2AB\sin\alpha+2CD\sin{\gamma}\big]\;\vc{e}_x,
\end{equation}
and the mean rotational swimming velocity is
\begin{equation}
\label{5.4}\overline{\vc{\Omega}}_2=\frac{3}{64a}\;\omega k\big[A^2-B^2-C^2+D^2+2AB\sin\alpha+2CD\sin{\gamma}\big]\;\vc{e}_z.
\end{equation}
The mean rate of dissipation per unit area is given by
\begin{equation}
\label{5.5}\overline{D}_2=\eta\omega^2 k(A^2+B^2+C^2+D^2),
\end{equation}
in generalization of Eq. (27).
A case of particular interest corresponds to $C=-A,\;D=-B,\;\beta=0,\;\gamma=\alpha$. This is a squeezing mode in which the lower surface is distorted in the opposite direction to the upper one. Clearly the two swimming velocities are the same as for the stretching mode studied earlier. For the symmetric squeezing mode corresponding to $C=-A,\;D=B,\;\beta=0,\;\gamma=\alpha$ the translational swimming velocity is given by Eq. (26), corresponding to Blake's Eq. (36), and the rotational swimming velocity vanishes. Blake did not distinguish the slab with symmetric squeezing mode, shown in his Fig. 1, from the two-dimensional waving infinitesimally thin sheet considered by Taylor and Tuck.
The tip of the displacement vector at any point of the surface, as given by Eq. (38), in general runs through an elliptical orbit as a function of time. The semi-axes, the ellipticity, and the tilt angle can be calculated as Stokes parameters (Bohren $\&$ Huffman 1983). The parameters in Eq. (38) can be varied to optimize the swimming speed for given power. This corresponds to maximization of the efficiency $E_2$, defined in Eq. (17). In particular for $C=A,\;D=B,\gamma=-\alpha$, or for $C=-A,\;D=-B,\gamma=-\alpha$, the mean swimming velocity becomes equal to the expression of Blake and Childress given by Eq. (26). For such strokes the efficiency $E_2$ is maximal for $A=\pm(1+\sqrt{2}) B,\;\alpha=\pi/2$, and equal to $\sqrt{2}$, substantially larger than unity, the value for the Taylor mode. For the optimum stroke the rotational swimming velocity vanishes.
\section{\label{VI}Discussion}
In the above we have studied the swimming of a circular disk caused by synchronized running distortion waves on the upper and lower surface of the disk. For asymmetric situations the disk performs both translational and rotational swimming.
The swimming of a slab generalizes the swimming of a sheet studied originally by Taylor (1951). His calculation has been extended to swimming with account of fluid inertia (Reynolds 1965, Tuck 1968), and to swimming in confined geometry (Katz 1974). Such studies could also be performed for a slab. The results can be applied to the swimming of a circular disk.
The calculations can be extended to the case of an elliptical disk with distortion waves on either side running in the direction of the long or short semi-axis. The mean translational swimming velocity remains the same as Eq. (40). The coefficient of the mean rotational velocity in Eq. (41) is modified and can be calculated from the known expressions for the rotational friction coefficient of an ellipsoid (Kim $\&$ Karrila 1991).
In the same article Taylor studied also the swimming of two sheets in parallel. Elsewhere we have studied the efficiency of swimming in this situation (Felderhof 2012). Similarly one could consider the swimming of two parallel circular disks. The calculations performed here suggest that such studies would be feasible and interesting.
\newpage
|
\section{Introduction}
Automorphic forms and their Hecke eigenvalues are of tremendous importance in number theory. These eigenvalues carry a lot of interesting arithmetic information, such as the number of points on elliptic curves or traces of Frobenius in Galois representations. One of the most successful methods for computing automorphic forms for~$\GL_2$ over number fields uses the Jacquet-Langlands correspondence.
This result transfers the problem to a quaternion algebra, in which it is often easier to solve. This approach has its roots in the theory of Brandt matrices and has been successfully used by Demb\'el\'e-Donnelly
and Greenberg-Voight~\cite{vd}
to compute Hecke eigenvalues of Hilbert modular forms.
In both methods, a crucial step is to test whether an ideal is principal and to produce a generator in this case: this is the \emph{principal ideal problem} that we are considering in this paper.
\par The principal ideal problem naturally splits into two cases: definite and indefinite algebras. In the definite case, Demb\'el\'e and Donnelly described an algorithm and Kirschmer and Voight proved that this algorithm runs in polynomial time when the base field is \emph{fixed}, so we focus on the remaining indefinite case. In that case, testing whether an ideal is principal reduces to the same problem over the base field by Eichler's theorem (Theorem~\ref{thmeichlerclass}), but finding a generator is difficult. Kirschmer and Voight~\cite{kv} provide an algorithm that improves on naive enumeration\footnote{Trying every linear combination of a basis until we find a generator.}, without analysing its complexity.
\par In this paper, we present a probabilistic algorithm using a factor base and an auxiliary data structure to solve the principal ideal problem. Our algorithm is inspired by Buchmann's algorithm~\cite{buchmann} for computing the class group of a number field. However, it is not easy to adapt this technique to quaternion algebras. Indeed, the set of right ideals of an order does not form a group under multiplication.
In fact, for most pairs of ideals, multiplication is not well-defined. We are able to salvage the factor base technique in the case of indefinite quaternion algebras by algorithmically realizing the strong approximation property (Theorem~\ref{thmstrongapprox}).
The main point is that if every ideal were two-sided, Buchmann's method would work unchanged. Our algorithm is divided in two parts. Because the algebra is indefinite, every ideal is equivalent to an ``almost two-sided'' ideal: a local algorithm (Algorithm~\ref{preduce}) makes this equivalence effective. The global algorithm (Algorithm~\ref{greduce}) uses a factor base: by linear algebra it cancels out the valuations of the \emph{norm} of the ideal and then corrects the ideal locally at every prime to make it two-sided.
We implemented our algorithm as in Magma. It performs well in practice, compared to the built-in Magma function implementing Kirschmer and Voight's algorithm.
\par The paper is organized as follows. We first recall basic properties of quaternion algebras, Eichler's theorems and Bruhat-Tits trees in Section~\ref{secrappels}. We then proceed to algorithms in Section~\ref{secalgos}. In Section~\ref{secred}, we define local and global reduction structures and Algorithm~\ref{isprincipal}, solving the principal ideal problem. In Section~\ref{secstruct}, Algorithm~\ref{gbuild} constructs the needed local and global reduction structures:
the first one uses units constructed from commutative suborders, and the second one is inspired by Buchmann's algorithm. In Section~\ref{seccr}, we introduce a compact representation for quaternions to prevent coefficient explosion in the previous algorithms. Section~\ref{seccomplexity} provides a complexity analysis of our algorithms: assuming suitable heuristics, we prove a subexponential running time. Section~\ref{secex} presents examples
\section{Background on quaternion algebras and Bruhat-Tits trees}\label{secrappels}
When~$G$ is a group and~$S\subset G$ is a subset, we write~$\gp{S}$ for the subgroup generated by~$S$. When the group~$G$ acts on a set~$X$, we say that~$S$ acts transitively on~$X$ if~$\gp{S}$ does.
\subsection{Quaternion algebras}
Good references for this section and the next one are~\cite{kv},~\cite{mfv} and~\cite{voightbook}. Let~$\fld$ be a number field with ring of integers~$\Z_\fld$ and discriminant~$\disc_\fld$. We write~$N:\fld\to\Q$ for the norm.
Let~$\prm$ be a prime of~$\Z_\fld$. We write~$\fld_\prm$ for the~$\prm$-adic completion of~$\fld$, we let $v_\prm$ be the $\prm$-adic valuation and we write~$\fldfin{\prm}$ for the residue field~$\Z_\fld/\prm$. When~$S$ is a set of primes of~$\Z_\fld$, we write~$\Z_{\fld,S}$ the ring of~$S$-integers in~$\fld$.
\par Let~$\alg$ be a quaternion algebra over~$\fld$ with reduced norm~$\nrd$. Let~$v$ be a place of~$\fld$. The place~$v$ is \emph{split} or~\emph{ramified} according to whether~$\alg\otimes_{\fld}\fld_v\cong\mat_2(\fld_v)$ or not. The \emph{reduced discriminant~$\delta_\alg$} of~$\alg$ is the product of the ramified primes and its \emph{absolute discriminant} is the integer~$\Delta_\alg=\disc_\fld^4N(\delta_\alg)^2$.
Let~$\order$ be a maximal order in~$\alg$. We write~$\order^1$ for the group~$\{x\in\order\ |\ \nrd(x)=1\}$. A \emph{lattice}~$I\subset\alg$ is a finitely generated~$\Z_\fld$-submodule such that~$\fld I=\alg$. The \emph{right order}~$\order_r(I)$ of~$I$ is the set~$\{x\in\alg\ |\ Ix\subset I\}$ and the \emph{left order~$\order_l(I)$} is defined analogously. A \emph{right $\order$-ideal} is a lattice~$I$ such that~$\order_r(I)=\order$. The ideal~$I$ is \emph{integral} if~$I\subset\order$ and~$I$ is \emph{two-sided} if~$\order_l(I)=\order_r(I)$. The inverse~$I^{-1}$ of~$I$ is~$\{x\in\alg\ |\ xI\subset\order\}$. If~$I,J$ are lattices such that~$\order_r(I)=\order_l(J)$, we define their product~$IJ$ to be the lattice generated by the set~$\{xy\ :\ x\in I,\ y\in J\}$. If~$I$ is an~$\order$-ideal we have~$II^{-1}=\order_l(I)$ and~$I^{-1}I=\order_r(I)$. The \emph{reduced norm~$\nrd(I)$} of an $\order$-ideal~$I$ is the~$\Z_\fld$-module generated by the reduced norms of elements in~$I$. The reduced norm of ideals is
multiplicative.
For a right~$\order$-ideal~$I$ we define~$\na(I)=N(\nrd(I))$ and for an element~$x\in\alg^\times$ we set~$\na(x)=\na(x\order)$.
Let~$\prm$ be a prime of~$\Z_\fld$. There exists a unique two-sided~$\order$-ideal~$\Prm$ such that every two-sided $\order$-ideal having reduced norm a power of~$\prm$ is a power of~$\Prm$. We have~$\Prm=\prm\order$ if~$\prm$ splits in~$\alg$ and~$\Prm^2=\prm\order$ if~$\prm$ ramifies in~$\alg$: such an ideal~$\Prm$ is called a prime of~$\order$, and every two-sided $\order$-ideal is a product of primes of~$\order$.
The set of right~$\order$-ideals is equipped with an action of the group of two-sided~$\order$-ideals by multiplication on the right and an action of the group~$\alg^\times$ by multiplication on the left. Two right~$\order$-ideals~$I,J$ are \emph{equivalent} if there exists~$x\in\alg^\times$ such that~$xI=J$, that is if they lie in the same orbit modulo~$\alg^\times$. The set of equivalence classes of right~$\order$-ideals is written~$\Cl(\order)$. An ideal is~\emph{principal} if it is equivalent to the unit ideal~$\order$. If~$S$ is a set of primes of~$\Z_\fld$, the~\emph{$S$-order} associated with~$\order$ is the ring~$\order_S=\Z_{\fld,S}\order$ and the group of~\emph{$S$-units} (relative to~$\order$) in~$\alg$ is~$\order_S^\times$.
\subsection{Eichler's condition and theorems}
A quaternion algebra~$\alg$ satisfies the \emph{Eichler condition} or is~\emph{indefinite} if there exists an infinite place of the base field~$\fld$ at which~$\alg$ is split. Indefinite algebras satisfy the following properties.
\begin{theorem}[(Consequence of strong approximation)]\label{thmstrongapprox}
Let~$\order$ be a maximal order in a quaternion algebra~$\alg$ over a number field~$\fld$, satisfying the Eichler condition. Let~$\prm$ be a prime of~$\Z_\fld$ that splits in~$\alg$ and~$k$ a positive integer. Then the map
\[\order^1 \longrightarrow \SL_2(\Z_\fld/\prm^k)\]
is surjective.
\end{theorem}
\begin{theorem}[(Integral version of Eichler's norm theorem)]\label{thmeichlernorm}
Let~$\order$ be a maximal order in a quaternion algebra~$\alg$ over a number field~$\fld$ satisfying the Eichler condition. Let~$S$ be a finite set of primes of~$\Z_\fld$. Let~$\Z_{\fld,S,\alg}^\times$ be the set of~$S$-units that are positive at every real place of~$\fld$ that ramifies in~$\alg$. Then the reduced norm
\[\nrd : \order_S^\times\longrightarrow \Z_{\fld,S,\alg}^\times\]
is surjective.
\end{theorem}
\begin{theorem}[(Eichler)]\label{thmeichlerclass}
Let~$\order$ be a maximal order in a quaternion algebra~$\alg$ over a number field~$\fld$ satisfying the Eichler condition. Let~$\clalg$ be the ray class group with modulus the product of the real places of~$\fld$ that ramify in~$\alg$. Then the reduced norm induces a bijection
\[\Cl(\order)\xrightarrow{\sim}\clalg\text{.}\]
\end{theorem}
In other words, two right~$\order$-ideals are equivalent if and only if the classes of their norm in~$\clalg$ are equal. Note that since~$\Cl(\order)$ is not a group, this map is only a bijection of \emph{sets}.
\subsection{The Bruhat-Tits tree}\label{secbtt}
The standard reference for this section is~\cite{serretrees}. Let~$\fld$ be a field with a discrete valuation~$v$. Let~$R$ be its valuation ring,~$\pi$ a uniformizer and~$\fldfin{}=R/\pi R$ the residue field. An \emph{$R$-lattice in~$\fld^2$} is an~$R$-submodule of rank~$2$ in~$\fld^2$. We define the \emph{Bruhat-Tits tree~$\tree$}, which we write~$\tree_\prm$ when~$\fld$ is the $\prm$-adic completion of a number field. The set of vertices of~$\tree$ is the set of homothety classes of $R$-lattices in~$\fld^2$. Let~$L,L'$ be two such $R$-lattices. There exists an ordered~$R$-basis~$(e_1,e_2)$ of~$L$ and integers~$a,b$ such that~$(\pi^a e_1,\pi^b e_2)$ is an $R$-basis of~$L'$. The integer~$|a-b|$ depends only on the homothety classes of~$L,L'$ and is called their \emph{distance}.
By definition, there is an edge in the tree~$\tree$ between every pair of vertices at distance~$1$. The graph~$\tree$ is an infinite tree. If~$P,Q$ are two vertices, the unique path of minimum length between~$P$ and~$Q$ is called the \emph{segment}~$PQ$ and the distance~$d(P,Q)$ equals the length of the segment~$PQ$. The set of vertices at distance~$1$ from a given vertex is in natural bijection with~$\Proj^1(\fldfin{})$.
The group~$\GL_2(\fld)$ acts on the tree and preserves the distance, and this action factors through~$\PGL_2(\fld)$ and is transitive on the set of vertices. The stabilizer of the vertex~$P_0$ corresponding to the $R$-lattice~$R^2$ is~$\fld^\times\GL_2(R)$ and the stabilizer of any vertex is a conjugate of this group. The group~$\SL_2(R)$ acts transitively on the set of vertices at a fixed distance from~$P_0$. For every~$g\in\mat_2(R)\setminus\pi\mat_2(R)$, the Smith normal form shows that~$d(g\act P_0,P_0)=v(\det(g))$. The tree is illustrated in Figure~\ref{figtree} where we label some vertices~$P$ with a matrix~$g$ such that~$P=g\act P_0$.
\begin{figure}[thb]
\begin{center}
\begin{tikzpicture}[btt,
level/.style={
level distance/.expanded=\ifnum#1>1 \tikzleveldistance/1.5\else\tikzleveldistance\fi,
nodes/.expanded={\ifodd#1 fill=black\else fill=white\fi}
}]
\draw (-0.3,-0.2) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}1&0\\0&1\end{smallmatrix}\right)$};
\draw (-0.3,-0.8) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}2&0\\0&1\end{smallmatrix}\right)$};
\draw (-0.6,-1.22) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}2&0\\1&2\end{smallmatrix}\right)$};
\draw (0.11,-1.53) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}4&0\\0&1\end{smallmatrix}\right)$};
\draw (0.11,-1.93) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}8&0\\0&1\end{smallmatrix}\right)$};
\draw (0.8,-1.22) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}4&0\\1&2\end{smallmatrix}\right)$};
\draw (-0.5,0.6) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}1&0\\0&2\end{smallmatrix}\right)$};
\draw (-0.68,1.1) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}1&0\\0&4\end{smallmatrix}\right)$};
\draw (-1.28,0.1) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}1&0\\2&4\end{smallmatrix}\right)$};
\draw (0.55,0.6) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}1&0\\1&2\end{smallmatrix}\right)$};
\draw (1.28,0.1) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}1&0\\1&4\end{smallmatrix}\right)$};
\draw (0.7,1.1) node[draw=none,scale=0.9] {$\left(\begin{smallmatrix}1&0\\3&4\end{smallmatrix}\right)$};
\levelu
\end{tikzpicture}
\end{center}
\caption{The Bruhat-Tits tree for~$\fld=\Q_2$}
\label{figtree}
\end{figure}
\begin{theorem}\label{thmbtt}
Let~$P,Q$ be two vertices of the tree~$\tree$ with~$d(P,Q)=1$. Then the action of the group~$G=\SL_2(\fld)$ on the vertices of~$\tree$ has exactly two orbits~$G\act P$ and~$G\act Q$
\end{theorem}
The connection between the Bruhat-Tits tree and ideals is the following: a right $\mat_2(R)$-ideal is always principal, generated by an element of~$\GL_2(\fld)$. Such an ideal is two-sided if and only if it is generated by an element of~$\fld^\times\GL_2(R)$. So there is a~$\GL_2(\fld)$-equivariant bijection between set of the vertices of the Bruhat-Tits tree and the quotient of the set of right~$\mat_2(R)$-ideals modulo the action of the group of two-sided $\mat_2(R)$-ideals.
\section{Algorithms}\label{secalgos}
We want to adapt the classical subexponential algorithms for computing the class group of a number field due to Hafner and McCurley~\cite{hmc} in the quadratic case and Buchmann~\cite{buchmann} in the general case to indefinite quaternion algebras by using a factor base: a fixed finite set of primes of~$\Z_\fld$. To simplify the notations, we set~$\Delta=\Delta_\alg$.
\begin{definition}\label{deffb}
The \emph{factor base} for~$\alg$ is a finite set~$\fb$ of primes of~$\Z_\fld$ that generates the group~$\clalg$.
\end{definition}
We say that a fractional ideal~$\idl$ of~$\fld$ is~\emph{$\fb$-smooth} or simply \emph{smooth} if it is a product of the primes in~$\fb$. Let~$I$ be a right~$\order$-ideal~$I$. When~$I$ is integral, we say that~$I$ is \emph{smooth} if its reduced norm is. When~$I$ is arbitrary, it is \emph{smooth} if it can be written~$I=J\idl$ with~$\idl$ a smooth fractional ideal of~$\fld$ and $J$ an integral smooth right $\order$-ideal. Equivalently, the ideal~$I$ is smooth if and only if~$I_\prm=\order_\prm$ for all~$\prm \notin \fb$. An element~$x\in A^\times$ is \emph{smooth} if the ideal~$x\order$ is smooth, or equivalently if~$x\in\order_S^\times$ with~$S=\fb$.
\medskip
We equip~$\mat_2(\R)$ and~$\mat_2(\C)$ with the usual positive definite quadratic form~$Q$ given by the sum of the squares of the absolute values of the coefficients, and we equip the Hamiltonian quaternion algebra~$\Hamil$ with the positive definite quadratic form~$Q=\nrd$. For each infinite place of~$\fld$ represented by a complex embedding~$\sigma$, we fix an isomorphism~$\sigma': \alg\otimes_{\fld}\fld_{\sigma}\cong M$ extending~$\sigma$, where~$M$ is one of~$\mat_2(\R)$, $\mat_2(\C)$ or~$\Hamil$. This defines a positive definite quadratic form~$T_2 : \alg\otimes_{\Q}\R \to \R$ by setting~$T_2(x) = \sum_{\sigma}[\fld_{\sigma}:\R]\cdot Q(\sigma'(x))$ for all~$x\in\alg\otimes_{\Q}\R$, giving covolume~$\Delta^{1/2}$ to the lattice~$\order$.
We represent a lattice in~$\alg$ by a $\Z_\fld$-pseudobasis (see~\cite{kv}). When~$L$ is a lattice in~$\alg$, we can enumerate its elements by increasing value of~$T_2$ with the Kannan--Fincke--Pohst algorithm~\cite{fp,kannan}. We represent this enumeration with a routine~\texttt{NextElement} that outputs a new element of~$L$ every time we call~\texttt{NextElement}($L$), ordering them by increasing value of~$T_2$.
\subsection{The reduction algorithms}\label{secred}
In this section, we describe the reduction structures and the corresponding reduction algorithms.
We start with the local reduction, which is an effective version of the fact that every integral right~$\order$-ideal of norm~$\prm^2$ is equivalent to the two-sided ideal~$\prm\order$ (Theorem~\ref{thmeichlerclass}). We perform this reduction by making algorithmic the reduction theory of~$\SL_2(\fld_\prm)$ on the Bruhat-Tits tree~$\tree_\prm$ (Section~\ref{secbtt}). The point is that this reduction needs only a small number of units: this leads to the definition of the $\prm$-reduction structure
\begin{definition}
Let~$\prm$ be a prime that splits in~$\alg$, let~$\order_0=\order$ and let~$P_0$ be the fixed point of~$\order_0^\times$ in the Bruhat-Tits tree~$\tree_\prm$. A \emph{$\prm$-reduction structure} is given by the following data:
\begin{enumerate}
\item the left order~$\order_1$ of an integral right $\order$-ideal of norm~$\prm$, and the fixed point~$P_1$ of~$\order_{1}^\times$ in~$\tree_\prm$;
\item for each~$b\in\{0,1\}$ and for each~$P\in\tree_\prm$ at distance~$1$ from~$P_b$, an element~$g\in\order_b^\times$ such that~$g\act P = P_{1-b}$.
\end{enumerate}
\end{definition}
Such a structure exists by strong approximation (Theorem~\ref{thmstrongapprox}). Note that if~$I$ is an integral right~$\order$-ideal of norm~$\prm$ such that~$\order_1=\order_l(I)$, we have~$I_\prm = x\order_\prm$ for some~$x\in\alg^\times$, and~$P_1 = x\act P_0$. We represent the points at distance~$1$ from~$P_b$ by elements of~$\Proj^1(\fldfin{\prm})$ and we compute the action on these points via explicit splitting maps~$\iota_{b}: \order_b/\prm\order_b \to \mat_2(\fldfin{\prm})$.
\medskip
This structure provides everything we need to perform reduction in the Bruhat-Tits tree. The following algorithm corresponds to the standard reduction procedure (Theorem~\ref{thmbtt}), which is illustrated in Figure~\ref{figred}. The idea is to use successive ``rotations'' (elements in~$\SL_2(\fld_\prm)$ having a fixed point in the tree) around the adjacent vertices~$P_0$ and $P_1$ to send an arbitrary vertex to one of the vertices~$P_b$: every rotation around a vertex decreases the distance to the other one.
\medskip
To realize this procedure, we need to perform the following subtask: given a right $\order$-ideal~$I$, find~$x\in I$ such that~$I_{\prm} = x\order_{\prm}$. A simple idea is to let~$e=v_\prm(\nrd(J))$ and to draw elements~$x\in J/\prm\order$ uniformly at random until~$v_\prm(\nrd(x))=e$. To obtain a deterministic algorithm, we can adapt Euclid's algorithm in the matrix ring~$\mat_2(\ring)$ with~$\ring = \Z_\fld/\prm^{e+1}$. This is done in~\cite{euclmat}, except that the base ring~$\ring$ is assumed to be a domain. We adapt the argument to our case. First note that we have a well-defined~$\prm$-adic valuation~$v=v_\prm$ in the ring~$\ring$. Let~$a,b\in\ring$. We have~$v(ab)\le v(a)+v(b)$ whenever~$ab\neq 0$, and~$a\mid b$ if and only if~$v(b)\ge v(a)$. If~$a\neq 0$, there is a Euclidean division taking the following simple form: if~$a\mid b$ then~$b = a\cdot (b/a) + 0$, and otherwise~$b = a\cdot 0 + b$. In every case we have written~$b = aq+r$
with~$r=0$ or~$v(r)<v(a)$.
Adapting this in the matrix ring leads to the following Euclidean division algorithm, where for convenience we write~$w = v\circ\det$. The idea is to work with~$A$ in Smith normal form, and if~$A$ is a diagonal matrix, dividing by~$A$ is almost the same as dividing by the diagonal coefficients. The difference is that we have to ensure that~$\det(R)\neq 0$ unless~$R=0$.
\begin{subalgorithm}[\texttt{DivideMatrix}]\label{divmat}
\begin{algorithmic}[1]
\INPUT two matrices~$A,B\in\mat_2(\ring)$ with~$\det A \neq 0$, where~$\ring = \Z_F/\prm^i$.
\OUTPUT two matrices~$Q,R\in\mat_2(\ring)$ such that~$B = AQ+R$, and~($R=0$ or~$w(R)<w(A)$).
\STATE let~$A'=UAV$ be the Smith form of~$A$ with~$U,V\in\SL_2(\ring)$ and~$A = (\begin{smallmatrix}a & 0\\ 0 & b\end{smallmatrix})$
\STATE $B'\sto UB$
\STATE let~$B''=B'W$ be the Hermite form of~$B'$ with~$W\in\SL_2(\ring)$ and~$B'' = (\begin{smallmatrix}c & 0\\ e & f\end{smallmatrix})$
\IF{$a\mid c$}
\IF{$b\mid f$}
\IF{$b\mid e$}
\STATE $Q \sto \bigl(\begin{smallmatrix}c/a & 0\\ e/b & f/b\end{smallmatrix}\bigr)$, $R \sto \bigl(\begin{smallmatrix}0 & 0\\ 0 & 0\end{smallmatrix}\bigr)$\label{stepcompletediv}
\ELSE
\STATE $Q \sto \bigl(\begin{smallmatrix}c/a & 1\\ 0 & f/b\end{smallmatrix}\bigr)$, $R \sto \bigl(\begin{smallmatrix}0 & -a\\ e & 0\end{smallmatrix}\bigr)$
\ENDIF
\ELSE
\STATE $Q \sto \bigl(\begin{smallmatrix}c/a-1 & 0\\ 0 & 0\end{smallmatrix}\bigr)$, $R \sto \bigl(\begin{smallmatrix}a & 0\\ e & f\end{smallmatrix}\bigr)$
\ENDIF
\ELSE
\IF{$b\mid f$}
\STATE $Q \sto \bigl(\begin{smallmatrix}0 & 0\\ 0 & f/b-1\end{smallmatrix}\bigr)$, $R \sto \bigl(\begin{smallmatrix}c & 0\\ e & b\end{smallmatrix}\bigr)$
\ELSE
\STATE $Q \sto \bigl(\begin{smallmatrix}0 & 0\\ 0 & 0\end{smallmatrix}\bigr)$, $R \sto \bigl(\begin{smallmatrix}c & 0\\ e & b\end{smallmatrix}\bigr)$
\ENDIF
\ENDIF
\RETURN $VQW^{-1}$, $U^{-1}RW^{-1}$
\end{algorithmic}
\end{subalgorithm}
\begin{proposition}
Subalgorithm~\ref{divmat} is correct.
\end{proposition}
\begin{proof}
By case-by-case analysis, we have~$B'' = A'Q+R$, and either~$R=0$ (Step~\ref{stepcompletediv}) or~$\det(R)\neq 0$ and~$w(R)<w(A')$. Let~$Q' =VQW^{-1}$ and~$R'=U^{-1}RW^{-1}$ be the matrices returned by the algorithm. We have~$B = U^{-1}B' = U^{-1}B''W^{-1} = U^{-1}(A'Q+R)W^{-1} = U^{-1}A'V^{-1}Q'+R' = AQ'+R'$. Since~$U,V$ and~$W$ have determinant~$1$, we have~$R=0$ if and only if~$R'=0$, and~$w(R')=w(R)<w(A')=w(A)$, proving the correctness of the algorithm.
\end{proof}
\begin{subalgorithm}[\texttt{GCDMatrix}]\label{gcdmat}
\begin{algorithmic}[1]
\INPUT two matrices~$A,B\in\mat_2(\ring)$ with~$\det A \neq 0$, where~$\ring = \Z_F/\prm^i$.
\OUTPUT a matrix~$D$ such that~$A\mat_2(\ring)+B\mat_2(\ring) = D\mat_2(\ring)$.
\STATE $Q,R\sto $ \texttt{DivideMatrix}($A$, $B$)
\IF{$R=0$}
\RETURN $A$
\ELSE
\RETURN \texttt{GCDMatrix}($R$, $A$)\label{steprec}
\ENDIF
\end{algorithmic}
\end{subalgorithm}
\begin{proposition}
Subalgorithm~\ref{gcdmat} is correct.
\end{proposition}
\begin{proof}
In Step~\ref{steprec} we have~$\det(R)\neq 0$ by the properties of Subalgorithm~\ref{divmat}, so the recursive call to~\texttt{GCDMatrix} is valid. The rest of the proof is the same as with the usual Euclidean algorithm.
\end{proof}
\begin{subalgorithm}[\texttt{LocalGenerator}]\label{locgene}
\begin{algorithmic}[1]
\INPUT an integral right~$\order$-ideal~$I$ and a prime~$\prm$, for some maximal order~$\order$.
\OUTPUT an element~$x\in I$ such that~$I_\prm = x\order_\prm$.
\STATE $b_1,\dots,b_n \sto $ an $LLL$-reduced~$\Z$-basis of~$I$
\STATE $e\sto v_\prm(\nrd(b_1))$
\STATE $\ring\sto \Z_\fld/\prm^{e+1}$
\STATE $B_1,\dots,B_n\sto $ images of~$b_1,\dots,b_n$ in~$\mat_2(\ring)$
\STATE $D\sto B_1$\label{stepinit}
\FOR{$i=1$ to~$n$}
\STATE $D\sto $ \texttt{GCDMatrix}($D$, $B_i$)
\ENDFOR
\STATE let~$\mu_1,\dots,\mu_n\in\Z$ be such that~$\sum\mu_i B_i=D$\label{stepsolve}
\RETURN $\sum \mu_i b_i$
\end{algorithmic}
\end{subalgorithm}
\begin{proposition}
Subalgorithm~\ref{locgene} is correct.
\end{proposition}
\begin{proof}
By definition of the norm of an ideal, we have~$v_\prm(\nrd(I))\le v_\prm(\nrd(b_1)) = e$. Because of the choice of~$\ring$, at Step~\ref{stepinit} we have~$\det(D)\neq 0$, so the calls to \texttt{GCDMatrix} are valid. By the properties of Subalgorithm~\ref{gcdmat}, at Step~\ref{stepsolve} we have~$D\mat_2(\ring) = B_1\mat_2(\ring) + \dots + B_1\mat_2(\ring)$ so the integers~$\mu_1,\dots,\mu_n$ exist. Now let~$x = \sum \mu_i b_i\in I$ be the output of the algorithm, so that~$v_\prm(\nrd(x))\le e$, hence~$\prm^{e+1}\order_\prm\subset x\order_\prm$. Let~$y\in I_\prm$. By reduction modulo~$\prm^{e+1}$ there exists~$a\in\order_\prm$ and~$b\in\prm^{e+1}\order$ such that~$y = xa+b\in x\order_\prm$, proving the result.
\end{proof}
Now we can present the local reduction algorithm.
\begin{subalgorithm}[\texttt{PReduce}]\label{preduce}
\begin{algorithmic}[1]
\INPUT an integral right $\order$-ideal $I$, a prime $\prm$ and a $\prm$-reduction structure.
\OUTPUT an integer~$r$, an element~$c\in \alg^\times$ and an integral~$\order$-ideal~$J$ such that~$cI = J\prm^r$ and~$v_\prm(\nrd(J))\in\{0,1\}$.
\STATE $r\sto$ largest integer such that~$I\subset \order\prm^r$, $J\sto I\prm^{-r}$\label{stepts}
\STATE $k\sto v_\prm(\nrd(J))
\STATE $c \sto 1$, $b\sto 0$
\STATE $x \sto$ \texttt{LocalGenerator}($I$, $\prm$)
\STATE $Q\sto x\act P_0$
\REPEAT
\STATE $P\sto $ point at distance~$1$ from~$P_b$ in the segment~$P_bQ$
\STATE $(c,J,Q)\sto g\act (c,J,Q)$ where~$g\in\order_b^\times$ is such that~$g\act P=P_{1-b}$\label{stepmul}
\LINEIF{$b=1$}{$J\sto J\prm^{-1}$, $r\sto r+1$, $k\sto k-2$}\label{stepdivide}
\STATE $b\sto 1-b$
\UNTIL{$k<2$}
\RETURN $J, c, r$
\end{algorithmic}
\end{subalgorithm}
In Step~\ref{stepts}, we have~$2r = v_\prm(\nrd(\tsidl))$ where~$\tsidl$ is the two-sided $\order$-ideal generated by~$I$. We can compute~$\tsidl$ as follows: if~$w_1,\dots,w_n$ is a~$\Z$-basis of~$\order$ and~$b_1,\dots,b_n$ is a~$\Z$-basis of~$I$, then~$\tsidl = \sum_{i,j}\Z w_ib_j$.
\medskip
\begin{figure}[thb]
\begin{center}
\begin{tikzpicture}[btt,
level/.style={
level distance/.expanded=\ifnum#1>1 \tikzleveldistance/1.5\else\tikzleveldistance\fi,
nodes/.expanded={\ifodd#1 fill=black\else fill=white\fi}
}]
\begin{scope}[xshift=-\scsh,yshift=\scsh]
\draw (0,0) node[draw=none,above,minimum size=20pt]{$P_0$};
\draw (0,-1) node[draw=none,below,minimum size=20pt]{$P_1$};
\draw (180:2) node[draw=none]{$Q$};
\draw (210:1.15) node[draw=none]{$g$};
\draw[->, >=stealth] (165:1) arc (165:255:1);
\draw (-67:2.15) node[draw=none]{$gQ$};
\draw (-1.9,-1.7) node[draw=none]{$k=4$};
\draw (-1.9,-1.95) node[draw=none]{$b=0$};
\path[rotate=30]
node[scale=\markscale]{
child[dotted] {
node[solid,scale=\markscale]{
child[solid]{
node{}
\levelt
}
child[dotted]{
node[solid]{}
child[dotted]{
node[solid]{}
child[solid]{
node[solid]{}
\levelc
}
child[dotted]{
node[solid,scale=\markscale]{
\levelc
}
}
child[solid]{
node[solid]{}
\levelq
}
}
}
child {
node{}
\leveld
}
child[dashed] {
node[solid]{}
child[solid] {
node[solid] {}
\levelt
}
child[dashed] {
node[solid]{}
child[solid]{
node[solid] {}
\levelq
}
child[dashed]{
node[solid] {}
child[densely dashed]{
node[solid, scale=\markscale]{
\levelc
}
child[solid]{
node{}
\levelc
}
}
}
};
\end{scope}
\begin{scope}[xshift=\scsh,yshift=\scsh
\draw (0,0) node[draw=none,above,minimum size=20pt]{$P_0$};
\draw (0,-1) node[draw=none,below,minimum size=20pt]{$P_1$};
\draw (-68:2.15) node[draw=none]{$Q$};
\draw (39:1.6) node[draw=none]{$gQ$};
\draw (-33:1) node[draw=none]{$g$};
\draw[->, >=stealth] (0.6,-1.2) arc (-35:75:0.7);
\draw (-1.9,-1.7) node[draw=none]{$k=4$};
\draw (-1.9,-1.95) node[draw=none]{$b=1$};
\path[rotate=30]
node[scale=\markscale]{
child[dotted] {
node[solid,scale=\markscale]{
child[solid]{
node[solid]{}
\levelt
}
child[dashed]{
node[solid]{}
child[dashed]{
node[solid]{}
child[solid]{
node[solid]{}
\levelc
}
child[densely dashed]{
node[solid, scale=\markscale]{
\levelc
}
}
child[solid]{
node[solid]{}
\levelq
}
}
}
child[dotted] {
node[solid]{}
child[solid]{
node{}
\levelt
}
child[dotted]{
node[solid, scale=\markscale]{
\levelt
}
}
child {
node{}
\leveld
};
\end{scope}
\begin{scope}[xshift=-\scsh,yshift=-\scsh
\draw (0,0) node[draw=none,above,minimum size=20pt]{$P_0$};
\draw (0,-1) node[draw=none,below,minimum size=20pt]{$P_1$};
\draw (39:1.6) node[draw=none]{$Q$};
\draw (-115:1.35) node[draw=none]{$gQ$};
\draw[<-, >=stealth] (-75:1) arc (-75:15:1);
\draw (-30:1.15) node[draw=none]{$g$};
\draw (-1.9,-1.7) node[draw=none]{$k=2$};
\draw (-1.9,-1.95) node[draw=none]{$b=0$};
\path[rotate=30]
node[scale=\markscale]{
child[dotted] {
node[solid,scale=\markscale]{
child[dotted]{
node[solid, scale=\markscale]{
\levelt
}
child[solid]{
node{}
\levelt
}
}
child[dashed]{
node[solid]{}
child[solid]{
node[solid]{}
\levelt
}
child[dashed]{
node[solid, scale=\markscale]{
\levelt
}
}
child {
node{}
\leveld
};
\end{scope}
\begin{scope}[xshift=\scsh,yshift=-\scsh
\draw (0,0) node[draw=none,above,minimum size=20pt]{$P_0$};
\draw (0,-1) node[draw=none,below,minimum size=20pt]{$P_1$};
\draw (-112:1.85) node[draw=none]{$Q$};
\draw (215:1) node[draw=none]{$g$};
\draw[->, >=stealth] (-0.6,-1.2) arc (215:105:0.7);
\draw (-1.9,-1.7) node[draw=none]{$k=2$};
\draw (-1.9,-1.95) node[draw=none]{$b=1$};
\path[rotate=30]
node[scale=\markscale]{
child[dotted] {
node[solid,scale=\markscale]{
child[dashed]{
node[solid, scale=\markscale]{
\levelt
}
child[solid]{
node{}
\levelt
}
}
child {
node{}
\leveld
}
child {
node{}
\leveld
};
\end{scope}
\end{tikzpicture}
\end{center}
\caption{\texttt{PReduce} (Subalgorithm~\ref{preduce})}
\label{figred}
\end{figure}
\begin{proposition}\label{proppreduce}
Subalgorithm~\ref{preduce} is correct.
\end{proposition}
\begin{proof}
Since~$k$ decreases by~$2$ every two iterations and is positive by the loop condition, the algorithm terminates. We now prove that the output is correct.
\par First, the distance~$d(P_b,Q)$ decreases by~$1$ during each execution of the loop: before Step~\ref{stepmul}, we have~$d(P_{1-b},gQ)=d(gP,gQ)=d(P,Q)=d(P_b,Q)-1$ since~$P$ is at distance~$1$ from~$P_b$ on the segment~$P_bQ$. We claim that before or after a complete execution of the loop, we have~$d(P_0,Q)=k$ (see Figure~\ref{figred}). The claim is true before the first iteration: we have~$x\in\order\setminus\prm\order$ so~$d(P_0,Q)=d(P_0,xP_0)=v_\prm(\nrd(x))=v_\prm(\nrd(J))=k$. We only need to prove that the equality~$d(P_0,Q)=k$ is preserved when~$b=1$. In that case, $P_1$ is in the segment~$P_0Q$ because of the previous iteration, so that~$d(P_0,Q)=1+d(P_1,Q)=k$.
\par We now prove that before or after a complete execution of the loop, the $\order$-ideal~$J$ is integral and~$v_\prm(\nrd(J))=k$. This property clearly holds before the first iteration. Before Step~\ref{stepdivide}, after two iterations~$b=0$ and~$b=1$, $J$ and~$Q$ have been multiplied by an element~$h$ such that~$v_\prm(\nrd(h))=0$ and~$d(P_0,hQ)=d(P_0,Q)-2$, so~$hJ$ is divisible by~$\prm$. Step~\ref{stepdivide} hence preserves integrality and updates~$k$ according to the valuation of~$\nrd(J)$.
\par We now prove the proposition. The element~$c$ is a product of elements of~$\order_0^\times$ and~$\order_1^\times$ so~$c$ is a~$\prm$-unit with~$\nrd(c)\in\Z_\fld^\times$. We have just proved that~$J$ is integral, and by Step~\ref{stepdivide} the value of~$r$ is such that~$cI = J\prm^r$. Because of the loop condition, after the algorithm terminates we have~$v_\prm(\nrd(J))=k\in\{0,1\}$.
\end{proof}
We now explain how to perform global reduction. We use linear algebra to control the valuations of a smooth ideal and then perform local reduction at every prime to get an ``almost two-sided ideal''. The first step is similar to its commutative analogue: we need sufficiently many ``relations'' (smooth elements in~$\alg^\times$) so that the quotient of the factor base by the norms of the relations is the ray class group~$\clalg$. This leads to the definition of a G-reduction structure.
\begin{definition}\label{defgred}
A \emph{G-reduction structure} is given by the following data:
\begin{enumerate}
\item a~$\prm$-reduction structure for each~$\prm\in\fb$ that splits in~$\alg$;
\item\label{defgred2} a finite set of elements~$X\subset \order\cap \alg^\times$ and a map~$\phi:\clalg \rightarrow \gp{\fb}$ that is a lift of an isomorphism~$\clalg \xrightarrow{\sim} \gp{\fb}/\gp{\nrd(X)}$ and such that~$\phi(1)=\Z_\fld$.
\end{enumerate}
\end{definition}
The following algorithm performs global reduction. In order to avoid explosion of the size of the ideal in the local reduction, we extract the two-sided part, allowing us to reduce all exponents modulo~$2$. The remaining part stays small and gets $\prm$-reduced, while the two-sided part is only multiplied by powers of primes.
\begin{subalgorithm}[\texttt{GReduce}]\label{greduce}
\begin{algorithmic}[1]
\INPUT a smooth integral right $\order$-ideal~$I$ and a G-reduction structure.
\OUTPUT an integral ideal $J$, an element~$c\in \alg^\times$ and a two-sided ideal $\tsidl$ such that~$cI=J\tsidl$ and~$I$ is principal if and only if~$J=\order$ and~$\tsidl=\order$.
\STATE $\idl\sto\nrd(I)$
\STATE $\idlb\sto\phi(\idl)$ where~$\idl$ is seen as an element of~$\clalg$
\STATE let~$e\in\Z^X$ be such that~$\nrd(y)\idl = \idlb$ where~$y = \prod_{x\in X}x^{e_x}$.
\STATE $c\sto \prod_{x\in X}x^{e_x \mathrm{mod}\, 2}$, $f\sto \prod_{x\in X}\nrd(x)^{\floor{e_x/2}}$ \COMMENT{Extract two-sided part}
\STATE $J\sto cI$\label{stepgmul}
\STATE $\tsidl\sto$ two-sided ideal generated by~$J$
\STATE $J\sto J\tsidl^{-1}$ \COMMENT{Extract two-sided part}\label{stepgts}
\STATE $\tsidl\sto f\tsidl$
\FOR{$\prm\in\fb$ dividing~$\nrd(J)$ and splitting in~$\alg$}\label{stepgloop}
\STATE $J,c',r \sto$ \texttt{PReduce}($J, \prm$)
\STATE $c\sto c'c$, $\tsidl\sto\prm^r\tsidl$
\ENDFOR
\RETURN $J, cf, \tsidl$
\end{algorithmic}
\end{subalgorithm}
\begin{proposition}\label{propgreduce}
Subalgorithm~\ref{greduce} is correct.
\end{proposition}
\begin{proof}
Since Step~\ref{stepgts} and \texttt{PReduce} preserve integrality (Proposition~\ref{proppreduce}), the output~$J$ is integral. The relation~$cI=J\tsidl$ is clear by tracking the multiplications. If the output is such that~$J=\order$ and~$\tsidl=\order$, then~$I=c^{-1}\order$ is principal. Conversely, if~$I$ is principal, then~$\idl=\nrd(I)$ is trivial in the class group~$\clalg$ so~$\idlb=\phi(\mathrm{cl}(\idl))=\Z_F$. After Step~\ref{stepgmul}, we have~$\nrd(J)=f^{-2}\Z_F$. After Step~\ref{stepgts}, we have multiplied~$J$ by a two-sided ideal, so~$v_\prm(\nrd(J))$ is even for all primes~$\prm$ splitting in~$\alg$. Since~$J$ is not divisible by a two-sided ideal, $\nrd(J)$ is not divisible by primes that ramify in~$\alg$. We obtain~$J=\order$ at the end of the loop by the properties of~\texttt{PReduce} so~$\nrd(\tsidl)=\Z_F$. Since~$\tsidl$ is two-sided, it is entirely determined by its norm so~$\tsidl=\order$.
\end{proof}
Finally, we reduce the general case to the smooth case by the noncommutative analogue of standard randomizing techniques. We generate a random smooth $\order$-ideal by the following procedure, to which we refer as \texttt{RandomLeftIdeal}($\order$). For each~$\prm\in\fb$, pick a nonnegative integer~$k$. Let~$\iota:\order\to\mat_2(\Z_F/\prm^k)$ be a splitting map. Let~$M\in\mat_2(\Z_F/\prm^k)$ be a random upper-triangular matrix with zero determinant and compute~$R_\prm = \order\iota^{-1}(M)+\prm^k\order$. Finally, return~$\bigcap_\prm R_\prm$. Choose the exponents~$k$ such that~$\na(R)\approx\Delta$.
\par It not clear at the moment what the best distribution for the exponents is.
A simpler idea would be to use random products~$\prod_\prm \prm^{k_\prm}$. In our experience, this leads to poorly randomized ideals. This is clear in the case~$\fld=\Q$: the randomized ideals are simply integer multiples of~$\order$.
\begin{lmsalgorithm}[\texttt{IsPrincipal}] \label{isprincipal}
\begin{algorithmic}[1]
\INPUT an integral right $\order$-ideal~$I$ and a G-reduction structure.
\OUTPUT an integral ideal $J$, an element~$c\in \alg^\times$ and a two-sided ideal $\tsidl$ such that~$cfI=J\tsidl$ and~$I$ is principal if and only if~$J=\order$ and~$\tsidl=\order$.
\STATE $R\sto$ \texttt{RandomLeftIdeal}($\order$)
\STATE $x\sto$ \texttt{NextElement}($I^{-1}\cap R$
\LINEIF{$xI$ is not smooth}{ \algorithmicreturn \textbf{ FAIL }}\label{stepsmooth}
\STATE $J, c, \tsidl \sto$ \texttt{GReduce}($xI$)
\RETURN $J, cx, \tsidl$
\end{algorithmic}
\end{lmsalgorithm}
By Proposition~\ref{propgreduce}, if Algorithm~\ref{isprincipal} does not return FAIL, its output is correct. In practice, we repeat Algorithm~\ref{isprincipal} until it returns the result.
\subsection{Building the reduction structures}\label{secstruct}
Now we explain how to build the previous reduction structures. The local reduction structure needs units in~$\order$. In general it is difficult to compute the whole unit group~$\order^\times$: for instance in the Fuchsian case, the minimal number of generators is at least~$\Delta^{3/8+o(1)}$ (this follows from the theory of signatures of Fuchsian groups~\cite[Section 4.3]{katok} and a volume formula~\cite[Theorem 11.1.1]{macreid}), which makes it hopeless to find a subexponential method. However, we can find some units in~$\order$ by considering commutative suborders and computing generators of their unit group with Buchmann's algorithm. Heuristically, these units are sufficiently random for our purpose. This strategy is implemented by the following algorithms.
\begin{subalgorithm}[\texttt{P1Search}] \label{p1search}
\begin{algorithmic}[1]
\INPUT a maximal order~$\orderb$ and a prime~$\prm$.
\OUTPUT a set of elements~$X\subset\orderb^\times$ acting transitively on~$\Proj^1(\fldfin{\prm})$.
\STATE $X\sto\emptyset$
\REPEAT
\STATE $x\sto$ \texttt{NextElement}($\orderb$)\label{stepnexteltp1s}
\STATE $\fldext\sto \fld(x)$
\IF{$\fldext/\fld$ has positive relative unit rank}\label{stepposrk}
\STATE $\ring\sto\Z_\fldext\cap\orderb$
\STATE\label{stepunits} $X\sto X\,\cup$ a set of generators of $\ring^\times$
\ENDIF
\UNTIL{$X$ acts transitively on~$\Proj^1(\fldfin{\prm})$}
\RETURN $X$
\end{algorithmic}
\end{subalgorithm}
In Step~\ref{stepunits}, we can compute the unit group~$\ring^\times$ with the algorithms of Kl\"uners and Pauli~\cite{rings}. Note that we actually do not need the full group~$\ring^\times$: a subgroup of finite index is sufficient.
\begin{proposition}\label{propp1search}
Subalgorithm~\ref{p1search} is correct.
\end{proposition}
\begin{proof}
By strong approximation (Theorem~\ref{thmstrongapprox}), the group~$\orderb^\times$ acts transitively on~$\Proj^1(\fldfin{\prm})$. This group is finitely generated, so after finitely many iterations we will have enumerated a set of generators and the algorithm will terminate. By the loop condition, the output is correct.
\end{proof}
\begin{subalgorithm}[\texttt{PBuild}] \label{pbuild}
\begin{algorithmic}[1]
\INPUT a maximal order~$\order$ and a prime~$\prm$.
\OUTPUT a $\prm$-reduction structure.
\STATE $I\sto$ an integral right $\order$-ideal of norm~$\prm$
\STATE $\order_1\sto\order_l(I)$
\STATE $X\sto$ \texttt{P1Search}($\order, \prm$)
\STATE from~$X$, for each~$P$ at distance~$1$ from~$P_0$ compute an element~$g\in\order^\times$ such that~$g\act P=P_1
\STATE $X\sto$ \texttt{P1Search}($\order_1, \prm$)\label{stepo1p1search}
\STATE from~$X$, for each~$P$ at distance~$1$ from~$P_1$ compute an element~$g\in\order_1^\times$ such that~$g\act P=P_0
\RETURN the $\prm$-reduction structure
\end{algorithmic}
\end{subalgorithm}
\begin{remark
Let~$g,g'\in\order^\times$ be such that~$g\act P_1 = g'\act P_1$ and let~$h=g^{-1}g'$. Then~$h\act P_1=P_1$ so~$h\in\order^\times\cap\order_1^\times$. This allows us to construct many elements in~$\order_1^\times$ before Step~\ref{stepo1p1search}. If we have sufficiently many such units, which often happens in practice, they will act transitively on the points~$P\neq P_0$ at distance~$1$ from~$P_1$. In this case, in Step~\ref{stepo1p1search} we will only need to find one element~$g\in\order_1^\times$ such that~$g\act P_0\neq P_0$.
\end{remark}
We build the global reduction structure in a way similar to the commutative case: we look for small relations in smooth ideals. In addition, we get a good starting point thanks to the inclusion~$\Z_\fld\subset\order$: the units~$\Z_{\fld,\fb}^\times$ provide all the relations up to a~$2$-elementary Abelian group.
\begin{lmsalgorithm}[\texttt{GBuild}] \label{gbuild}
\begin{algorithmic}[1]
\INPUT a maximal order~$\order$ and a factor base~$\fb$.
\OUTPUT a G-reduction structure.
\FOR{$\prm\in\fb$ that splits in~$\alg$}
\STATE \texttt{PBuild}($\order, \prm$)\label{steppbuild}
\ENDFOR
\STATE $X\sto$ integral generators of~$\Z_{\fld,\fb}^\times$\label{stepsunits}
\FOR{$\prm\in\fb$}
\STATE $I\sto $ integral $\order$-ideal of norm~$\prm$\label{stepidnormp}
\REPEAT\label{loop1}
\STATE $x\sto$ \texttt{NextElement}($I$)\label{stepnextgb1}
\UNTIL{$x$ is smooth}
\STATE $X\sto X\cup\{x\}$
\ENDFOR
\WHILE{$\gp{\fb}/\gp{\nrd(X)}\not\cong\clalg$}\label{loop2}
\STATE $x\sto$ \texttt{NextElement}($\order$)\label{stepnextgb2}
\LINEIF{$x$ is smooth}{$X\sto X\cup\{x\}$}
\ENDWHILE
\RETURN the G-reduction structure
\end{algorithmic}
\end{lmsalgorithm}
\begin{remark}
The various calls to~\texttt{PBuild} in Step~\ref{steppbuild} are not completely independent: we can keep the elements in~$\order^\times$ from one call for other ones.
\end{remark}
\begin{proposition}\label{propgbuild}
Algorithm~\ref{gbuild} is correct.
\end{proposition}
\begin{proof}
Let~$I$ be the ideal in Step~\ref{stepidnormp}. There exists a smooth ideal~$J$ equivalent to~$I$, let~$x\in\alg^\times$ be such that~$I=xJ$. Then~$x\order = IJ^{-1}$, so~$\nrd(J)x\in I\bar{J}\subset I$ is smooth. It will be enumerated at some point, so the loop starting at Step~\ref{loop1} terminates. Since~$\fb$ generates the class group~$\clalg$, by Eichler's theorem (Theorem~\ref{thmeichlerclass}) we have~$\gp{\fb}/\nrd(\alg^\times)\cong\clalg$, so there exists a finite set of~$\fb$-smooth elements~$X\subset\order$ such that~$\gp{\fb}/\gp{\nrd(X)}\cong\clalg$. We will enumerate this set at some point, so the loop starting at Step~\ref{loop2} terminates. So Algorithm~\ref{gbuild} terminates, and by Proposition~\ref{propp1search} and Step~\ref{loop2} it returns a correct G-reduction structure.
\end{proof}
\begin{remark}
We have restricted to maximal orders to simplify the exposition, but this restriction can be weakened as follows. Let~$\order$ be an arbitrary order, and let~$S$ be the set of primes~$\prm$ of~$\Z_\fld$ such that~$\order$ is not $\prm$-maximal. The set $S$ is finite, so we can choose a factor basis disjoint from~$S$. Then our algorithms work unchanged for right $\order$-ideals except one point: Theorem~\ref{thmeichlerclass} characterizing principal ideals might no longer hold. If we restrict to Eichler orders, that is intersections of two maximal orders, Theorem~\ref{thmeichlerclass} still holds. Otherwise we need to find the suitable class group and change Definition~\ref{defgred} \ref{defgred2} accordingly.
\end{remark}
\subsection{Compact representations}\label{seccr}
In the previous algorithms, the cost of elementary operations is important. Representing units as linear combinations of a basis of~$\order$ could be catastrophic: the classical example of real quadratic fields suggests that fundamental units in commutative orders, such as those produced by Subalgorithm~\ref{p1search}, can have exponential size. This problem is classically circumvented by representing units in \emph{compact form}: a product of small~$S$-units with possibly large exponents. The problem is reduced to computing efficiently with those compact representations.
A natural notion of compact representation in~$\order$ would be to take ordered products of $S$-units in~$\order$ but we do not know how to compute efficiently with such a general representation. Instead we use a more restricted notion:
we group the units belonging to a common commutative suborder, in which we can compute efficiently. This leads to the following definition.
\begin{definition}
A \emph{compact representation} in~$\order$ is:
\begin{enumerate}
\item an element~$x\in\order$, or
\item\label{typetwo} a product~$y = \prod_{i=1}^{r}y_i^{e_i}$ where the exponents are signed integers, the elements~$y_i$ all lie in a single ring~$\ring\subset \order$ containing~$\Z_\fld$ and such that~$y\in\ring^\times$, together with a~$\Z$-basis of the integral closure of~$\ring$ and a factorization of its conductor, or
\item an ordered product of compact representations.
\end{enumerate}
A product~$y$ as in~\ref{typetwo} will be called a \emph{representation of type~\ref{typetwo}}.
\end{definition}
We describe the algorithms for representations of type~\ref{typetwo}, and they naturally extend by multiplicativity to arbitrary compact representations. We first explain the algorithm for local evaluation of compact representations. Since the product represents a unit, we can replace it with a product of local units, avoiding loss of precision despite the large exponents.
\begin{subalgorithm}[\texttt{EvalCR}] \label{evalcr}
\begin{algorithmic}[1]
\INPUT a representation~$y = \prod_{i=1}^{r}y_i^{e_i}\in\ring$ of type~\ref{typetwo}, an ideal~$\idl\subset\Z_\fld$.
\OUTPUT an element~$z\in\order$ such that~$z=y\pmod{\idl\order}$.
\STATE $\fldext\sto$ field of fractions of~$\ring$
\STATE $\cnd\sto$ conductor of~$\ring$ inside~$\Z_\fldext$
\STATE $\prod_{j=1}^{k}\Prm_j^{w_j}\sto$ factorization of~$\idl\cnd\Z_\fldext$
\FOR{$j=1$ to $k$}
\STATE $\phi \sto$ reduction map onto~$\Z_\fldext/\Prm_j^{w_j}$
\STATE $\pi\sto$ uniformizer in~$\Prm_j$, $v\sto v_{\Prm_j}$
\STATE $z_j\sto \prod_{i=1}^{r}\phi(y_i\pi^{-v(y_i)})^{e_i}$
\ENDFOR
\STATE $z\sto$ element in~$\Z_\fldext$ such that~$z=z_j\pmod{\Prm_j^{w_j}}$
\RETURN $z$
\end{algorithmic}
\end{subalgorithm}
\begin{proposition}\label{propevalcr}
Subalgorithm~\ref{evalcr} is correct.
\end{proposition}
\begin{proof}
First, we claim that for all~$j\le k$, we have~$z_j=y\pmod{\Prm_j^{w_j}}$. Since~$\nrd(y)\in\Z_\fld^\times$, we have~$\sum_{i=1}^r v(y_i)e_i=0$ so~$\prod_{i=1}^{r}(y_i\pi^{-v(y_i)})^{e_i} = \prod_{i=1}^{r}\pi^{-v(y_i)e_i}\prod_{i=1}^{r}y_i^{e_i}=y$. Since~$y_i\pi^{-v(y_i)}$ is integral at~$\Prm_j$, we can apply~$\phi$ to it and the claim follows. This implies that the output~$z$ of the algorithm satisfies~$z=y\pmod{\idl\cnd\Z_\fldext}$. In particular, we have~$z=y\pmod{\cnd\Z_\fldext}$ so~$z-y\in\ring$. Since~$y\in\ring$ we get~$z\in\ring\subset\order$. The relation~$\idl\cnd\Z_\fldext\subset\idl\ring\subset\idl\order$ shows that~$z=y\pmod{\idl\order}$.
\end{proof}
We can now explain how to multiply an ideal by a compact representation. To know an ideal, it suffices to know it up to large enough precision: we reduce the problem to local evaluation.
\begin{subalgorithm}[\texttt{MulCR}] \label{mulcr}
\begin{algorithmic}[1]
\INPUT a representation~$y = \prod_{i=1}^{r}y_i^{e_i}\in\ring$ of type~\ref{typetwo}, an integral right~$\order$-ideal~$I$.
\OUTPUT the ideal~$yI$.
\STATE $\idl\sto \nrd(I)$
\STATE $z\sto$ \texttt{EvalCR}($y, \idl$)
\RETURN $zI+\idl\order$
\end{algorithmic}
\end{subalgorithm}
\begin{proposition}\label{propmulcr}
Subalgorithm~\ref{mulcr} is correct.
\end{proposition}
\begin{proof}
By Proposition~\ref{propevalcr}, we have~$z=y\pmod{\idl\order}$ so~$(z-y)I\subset(z-y)\order\subset\idl\order$, which gives~$zI+\idl\order=yI+\idl\order$. Since~$y\in\order^\times$, we have~$\nrd(yI)=\nrd(I)=\idl$, so~$\idl\order\subset yI$ and finally~$yI+\idl\order=yI$. Therefore the output of the algorithm is correct.
\end{proof}
\section{Complexity analysis}\label{seccomplexity}
We perform a complete complexity analysis of our algorithms, assuming suitable heuristics. To simplify the notations, we set~$L(x)=\exp(\sqrt{\log x\log\log x})$. In every complexity estimate, the degree of the base field~$\fld$ is fixed. When we mention a complexity of the form~$\secompl$, we always implicitly mean~$\secompl$ times a polynomial in the size of the input. We fix a parameter~$\alpha>0$. We will analyse our algorithm using the general paradigm that with a factor base of subexponential size, elements have a subexponential probability of being smooth. However, recall from Definition~\ref{deffb} that the factor base~$\fb$ is assumed to generate the ray class group~$\clalg$, so we need the following heuristic.
\begin{heuristic}\label{heurfb}
There exists a constant~$c=c_\alpha$ such that for every quaternion algebra~$\alg$ with absolute discriminant~$\Delta$, the set of primes having norm less than~$c\cdot L(\Delta)^{\alpha}$ generate the class group~$\clalg$.
\end{heuristic}
This heuristic is a theorem under the Generalized Riemann Hypothesis~\cite{bach}. By Minkowski's bound, Heuristic~\ref{heurfb} is also true unconditionnally with the restriction that~$\log\Delta\gg_{\alpha} (\log |\disc_\fld|)^2$.
From now on, we assume Heuristic~\ref{heurfb} and we assume that the factor base~$\fb$ is the set of primes having norm less than~$c\cdot L(\Delta)^{\alpha}$. Note that this implies the bound~$\#\fb \le L(\Delta)^{\alpha+o(1)}$.
\medskip
We start by analysing the complexity of elementary operations: Subalgorithm~\ref{locgene} and the algorithms of Section~\ref{seccr}.
\begin{lemma}\label{lemcompllocgene}
Subalgorithm~\ref{locgene} terminates in time polynomial in the size of the input.
\end{lemma}
\begin{proof}
Subalgorithm~\ref{divmat} (\texttt{DivideMatrix}) is made of a constant number of elementary operations, so it is polynomial. In Subalgorithm~\ref{gcdmat} (\texttt{GCDMatrix}) with~$\ring = \Z_\fld/\prm^i$, since the valuation of the determinant decreases at every recursive call, there are at most~$i$ such calls. When Subalgorithm~\ref{locgene} (\texttt{LocalGenerator}) calls Subalgorithm~\ref{gcdmat}, we have~$i=e+1=v_{\prm}(\nrd(b_1))+1 = O(\log \na(I))$ by lattice reduction. So the algorithm is polynomial in the size of the input.
\end{proof}
\begin{lemma}\label{lemcomplcr}
Given the factorization of~$\idl$, Subalgorithm~\ref{evalcr} terminates in time polynomial in the size of the input. Given the factorization of~$\nrd(I)$, Subalgorithm~\ref{evalcr} terminates in time polynomial in the size of the input.
\end{lemma}
\begin{proof}
In Subalgorithm~\ref{evalcr}, the number of iterations of the loop is polynomial in the size of the input. The only operations that could possibly not be polynomial in the size of the input are the computation of~$\Z_\fldext$ and the factorization of~$\idl\cnd\Z_L$, but~$\Z_\fldext$ and the factorization of~$\cnd$ are contained in the compact representation, and the factorization of~$\idl$ is assumed to be given. So the algorithm is polynomial in the size of its input. Since the HNF of the output can be computed in polynomial time~\cite{hnf}, Subalgorithm~\ref{evalcr} is also polynomial.
\end{proof}
The restriction on the factorization is not a problem in our application: every ideal on which we call these algorithms is smooth.
\medskip
Since our algorithms use their commutative counterparts, we have to make assumptions on the algorithms used to compute commutative unit groups.
\begin{heuristic}\label{heurfastunits}
There is an explicit algorithm that, given a number field~$\fldb$ with discriminant~$\disc_\fldb$, an order~$\ring$ in~$\fldb$ and a bound~$b = L(\disc_\fldb)^{O(1)}$, computes a set~$U$ of integral generators for the~$S$-unit group of~$\Z_\fldb$, where~$S$ is the set of primes of norm less than~$b$, and generators for the unit group~$\ring^\times$ expressed as products of elements in~$U$, in expected time~$L(\disc_\fldb)^{O(1)}$.
\end{heuristic}
This is a strong hypothesis. However, under the Generalized Riemann Hypothesis it is a theorem for quadratic number fields~\cite{hmc,voll} and experience has shown that it is not an unreasonable assumption\footnote{The PARI developers experimented extensively with this algorithm in the past twenty years, as implemented in the PARI/GP function \texttt{bnfinit}, while building and checking tables of number fields of small degree [Ref: \url{http://pari.math.u-bordeaux1.fr/pub/pari/packages/nftables/}], as well as with number fields of much larger degree. E.g. the class group and unit group of $\fld = \Q[t]/(t^{90}-t^2-1)$, $d_\fld > 10^{175}$ can be computed in a few hours (Karim Belabas, personal communication).}.
\medskip
Since the reduction algorithms depend on the structures that are given as input, we analyse the algorithms building the reduction structures before the reduction algorithms. We start with Subalgorithm~\ref{p1search}, for which we need some heuristic assumptions.
\begin{heuristics}\label{heurunits} In Subalgorithm~\ref{p1search}, we assume the following.
\begin{enumerate}[(i)]
\item\label{heurunits1} If~$\fld$ is totally real, a positive proportion of~$x$ satisfies the condition of Step~\ref{stepposrk} that~$\fld(x)/\fld$ has positive relative unit rank.
\item\label{heurunits2} The images in~$\PGL_2(\fldfin{\prm})$ of the units produced at Step~\ref{stepunits} are uniformly distributed in the image of~$\order^\times$ in~$\PGL_2(\fldfin{\prm})$.
\end{enumerate}
\end{heuristics}
\begin{proposition}\label{propcomplp1search}
Assuming Heuristics~\ref{heurfastunits} and~\ref{heurunits}, the expected running time of Subalgorithm~\ref{p1search} is at most~$\secompl$.
\end{proposition}
\begin{proof}
We first prove that the expected number of iterations of the loop is~$O(1)$. If $\fld$ is not totally real, the relative unit rank condition is always satisfied, so by Heuristic~\ref{heurunits}~(\ref{heurunits1}) a positive proportion of the iterations of the loop produce a unit.
By strong approximation, the image of~$\order^\times$ in~$\PGL_2(\fldfin{\prm})$ contains~$\PSL_2(\fldfin{\prm})$, and the index is at most~$2$. By Heuristic~\ref{heurunits}~(\ref{heurunits2}), with probability at least~$1/2$ the image of the unit produced at Step~\ref{stepunits} is in~$\PSL_2(\fldfin{\prm})$, and the corresponding images are equidistributed in~$\PSL_2(\fldfin{\prm})$. By~\cite{lubo}, the probability that two random elements of~$\PSL_2(\fldfin{\prm})$ generate this group is bounded below by a constant.
Therefore, after an expected number of iterations~$O(1)$, the image of~$X$ generates~$\PSL_2(\fldfin{\prm})$ and hence acts transitively on~$\Proj^1(\fldfin{\prm})$.
\par Each computation of a unit group in Step~\ref{stepunits} takes expected time~$\secompl$ by Heuristic~\ref{heurfastunits} since the discriminant of~$\Z_\fld[x]$ is~$\Delta^{O(1)}$. The units are stored in compact representation and we use Subalgorithm~\ref{evalcr} (\texttt{EvalCR}) to compute the action on~$\Proj^1(\fldfin{\prm})$, so by Lemma~\ref{lemcomplcr} this takes total time~$\secompl$.
\end{proof}
\begin{heuristics}\label{heursmoothdist} Let~$\Z_{\fld,\fb,\alg}^\times$ be the group of~$\fb$-units in~$\Z_\fld$ that are positive at every real place ramified in~$\alg$. In Algorithm~\ref{gbuild}, we assume the following.
\begin{enumerate}[(i)]
\item\label{heursmoothdist1} There exists a constant~$\beta>0$ such that the elements~$x$ produced in Steps~\ref{stepnextgb1} and~\ref{stepnextgb2} are smooth with probability at least~$L(\Delta)^{-\beta+o(1)}$.
\item\label{heursmoothdist2} The norms of the smooth elements produced in Step~\ref{stepnextgb2} are equidistributed in~$\Z_{\fld,\fb,\alg}^\times / \Z_{\fld,\fb}^{\times 2}$.
\end{enumerate}
\end{heuristics}
By comparison with the case of integers~\cite[Equation (1.16) and Section 1.3]{granville}, $\beta=1/(2\alpha)$ could be a reasonable value.
\begin{theorem}\label{thmcomplgbuild}
Assume Heuristics~\ref{heurfastunits},~\ref{heurunits}, \ref{heurfb} and~\ref{heursmoothdist}. Then, given a maximal order~$\order$ in an indefinite quaternion algebra~$\alg$, Algorithm~\ref{gbuild} (\texttt{GBuild}) terminates in expected time~$\secompl$.
\end{theorem}
\begin{proof}
There are~$2\cdot \#\fb = \secompl$ calls to Subalgorithm~\ref{p1search} (\texttt{P1Search}). By Proposition~\ref{propcomplp1search}, these calls take total time~$\secompl$. The computation of the group~$\Z_{\fld,\fb}^\times$ takes time~$L(\disc_\fld)^{O(1)}=\secompl$ by Heuristic~\ref{heurfastunits}. By Heuristic~\ref{heursmoothdist}~(\ref{heursmoothdist1}), the expected number of iterations in the loop starting at Step~\ref{loop1} is at most~$\secompl$ for each~$\prm\in\fb$, so the total expected number of iterations of this loop is at most~$\#\fb\cdot \secompl=\secompl$.
\par We study the loop starting at Step~\ref{loop2}. After Step~\ref{stepsunits}, we have~$\gp{\fb}/\gp{\nrd(X)}=\gp{\fb}/\Z_{\fld,\fb}^{\times 2}$. Let~$C$ be the group~$\Z_{\fld,\fb,\alg}^\times / \Z_{\fld,\fb}^{\times 2}$. There is an exact sequence
\[ 1 \longrightarrow C \longrightarrow \gp{\fb}/\Z_{\fld,\fb}^{\times 2} \longrightarrow \clalg \longrightarrow 1\text{,}\]
so that the loop terminates if and only if the image of~$\nrd(X)$ generates the group~$C$. We have~$\# C = 2^{\#\fb + O(1)}$, so by Heuristic~\ref{heursmoothdist}~(\ref{heursmoothdist2}) this happens after we find an expected number of~$\#\fb+O(1)$ smooth elements. By Heuristic~\ref{heursmoothdist}~(\ref{heursmoothdist1}), the total expected number of iterations of this loop is at most~$\#\fb\cdot \secompl=\secompl$. Checking the loop condition~$\gp{\fb}/\gp{\nrd(X)}\not\cong\clalg$ amounts to linear algebra, so it also takes time~$\secompl$. This proves the theorem.
\end{proof}
\begin{theorem}
Assume Heuristics~\ref{heurfastunits},~\ref{heurunits}, \ref{heurfb} and~\ref{heursmoothdist}. Then, given the G-reduction structure computed by Algorithm~\ref{gbuild} and a smooth integral right $\order$-ideal~$I$, Algorithm~\ref{greduce} (\texttt{GReduce}) terminates in expected time~$\secompl$.
\end{theorem}
\begin{proof}
First, since by Theorem~\ref{thmcomplgbuild}, Algorithm~\ref{gbuild} terminates in expected time~$\secompl$ so in particular the expected size of its output is also at most~$\secompl$. In Subalgorithm~\ref{greduce} (\texttt{GReduce}), the first part is linear algebra so it takes time~$\secompl$. By Lemma~\ref{lemcomplcr}, all the elementary operations can be performed in time polynomial in the size of their input.
We analyse the calls to Subalgorithm~\ref{preduce} (\texttt{PReduce}). In this subalgorithm, the variable~$k$ decreases by~$2$ every two iterations and the initial value of~$k$ is bounded by~$v_\prm(\nrd(J))$, so the algorithm terminates after at most~$v_\prm(\nrd(J))$ iterations by the loop condition. So the total number of iterations in the calls to Subalgorithm~\ref{preduce} is bounded by~$\sum_{\prm\in\fb}v_\prm(\nrd(J))\le \log_2\na(J) \le \log_2(N_X\cdot \na(I))$ by Step~\ref{stepgmul} of Subalgorithm~\ref{greduce}, where~$N_X = \prod_{x\in X}\na(x)$. But~$\log\na(I)$ is polynomial in the size of the input and~$\log N_X$ is polynomial in the size of the G-reduction structure, which is~$\secompl$. This proves the theorem.
\end{proof}
Finally, for a general integral right $\order$-ideal, repeated attempts with Algorithm~\ref{isprincipal} (\texttt{IsPrincipal}) takes total expected time~$\secompl$ if we assume the following heuristic.
\begin{heuristic}\label{heursmoothen}
There exists a constant~$\gamma>0$ such that in Step~\ref{stepsmooth} of Algorithm~\ref{isprincipal}, the element~$x$ is smooth with probability at least~$L(\Delta)^{-\gamma+o(1)}$.
\end{heuristic}
Again, by comparison with the case of integers~\cite{granville}, $\gamma=1/(\sqrt{2}\alpha)$ could be a reasonable value.
\section{Examples}\label{secex}
We have implemented the above algorithms in the computer algebra system Magma~\cite{magma}. In this section, we demonstrate how our algorithms work and perform in practice. Every computation was performed on a $2.5$~GHz Intel Xeon E5420 processor with Magma v2.20-5 from the PLAFRIM experimental testbed.
\paragraph{Example 1.}Let~$\alg$ be the quaternion algebra over~$\Q$ generated by two elements~$i,j$ such that~$i^2=3$, $j^2=-1$ and~$ij=-ji$. The algebra~$\alg$ is ramified at~$2$ and~$3$ and unramified at every other place: $\alg$ is indefinite and our method applies. Let~$\order$ be the maximal order~$\Z+\Z i+\Z j+\Z\omega$ where~$\omega=(1+i+j+ij)/2$. We construct a reduction structure with Algorithm~\ref{gbuild} (\texttt{GBuild}) and factor base~$\fb=\{2,3,5,7,11,13,17\}$. Let~$I=19\order+a\order$ where~$a=-3-4i+j$, so that~$\nrd(I)=19\Z$. We use Algorithm~\ref{isprincipal} (\texttt{IsPrincipal}) to compute a generator of~$I$. It finds an elemen
~$x=(7+i-9j-3\omega)/19\in I^{-1}$ such that~$\nrd(xI)=7\Z$, so that~$xI$ is smooth. The linear algebra phase in Subalgorithm~\ref{greduce} (\texttt{GReduce}) compute
~$c=-1-2i-j+\omega$ having norm~$-7$ and~$f=1/7$. We obtain~$J=49\order+b\order$ with~$b=-17-8i+j$ and~$\tsidl=7^{-1}\order$ before the local reduction. We reduce the ideal~$J$ at~$7$.
In Subalgorithm~\ref{preduce} (\texttt{PReduce}), at the first iteration we have~$P=P_1$ so~$c=1$. In the second iteration we hav
~$c=(-9-5i-7j-3\omega)/7$: the element~$c$ has norm~$1$ and~$r=1$. After multiplying every element, we obtain the outpu
~$c=7/19\cdot(8+4i+3j-11\omega), f=1/7$ an
~$x=(cf)^{-1}=3+4i-3j-11\omega$ has norm~$-19$: $x$ is a generator of the ideal~$I$.
\paragraph{Example 2.}Let~$\fld$ be the complex cubic field of discriminant~$-23$, which is generated by an element~$t$ such that~$t^3-t+1=0$. Let~$\alg$ be the quaternion algebra over~$\fld$ generated by two elements~$i,j$ such that~$i^2=2t^2+t-3$,~$j^2=-5$ and~$ij=-ji$. The algebra~$\alg$ is ramified at the real place of~$\fld$ and the discriminant~$\delta_\alg$ has norm~$5$. All the maximal orders in~$\alg$ are conjugate and we compute one of them with Magma. Algorithm~\ref{gbuild} (\texttt{GBuild}) constructs the reduction structure in~$4$ seconds. We then compute the~$22$ primes of~$\fld$ coprime to~$\delta_\alg$ and having norm less than~$100$. For every such prime~$\prm$, we construct a random integral right $\order$-ideal~$I$ with norm~$\prm$. Since~$\clalg$ is trivial, they are all principal. We apply Algorithm~\ref{isprincipal} (\texttt{IsPrincipal}) to compute a generator of each of these ideals.
This computation takes~$0.3$ seconds per ideal on average with a maximum of~$0.9$ seconds. As a comparison, we compute generators for the same ideals with the function provided by Magma. This computation takes~$4$~hours per ideal on average with a maximum of~$69$h, and~$5$ of the~$23$ ideals take less that~$0.1$~seconds.
\paragraph{Example 3.}When the base field is totally real and the algebra is ramified at every real place except one, there is an algorithm of Voight~\cite{vfuchsian} for computing the unit group of an order. In~\cite[p.~25]{vd}, Demb\'el\'e and Voight mention but do not describe an unpublished algorithm using this computation to speedup ideal principalization. This algorithm is provided in Magma\footnote{\texttt{IsPrincipal(<Any> I, <GrpPSL2> Gamma) -> BoolElt, AlgQuatElt}} and improves on the algorithm of~\cite{kv}.
Let~$\fld$ be the real cubic field of discriminant~$3132=2^2\cdot 3^3\cdot 29$, which is generated by an element~$t$ such that~$t^3-15t+6=0$. Let~$\alg$ be the quaternion algebra over~$\fld$ generated by two elements~$i,j$ such that~$i^2=-1$,~$j^2=(141t^2+57t-2092)/2$ and~$ij=-ji$. The algebra~$\alg$ is ramified at two of the three real places of~$\fld$ and no finite place.
We compute a maximal order in~$\alg$ and then construct the reduction structure in~$14$ seconds.
We produce a random integral ideal of norm~$\prm$ for each prime~$\prm$ having norm less than~$100$ and compute a generator for each of them with our algorithm. The computation takes~$1.5$~seconds per ideal on average with a maximum of~$4.6$~seconds. With Magma we compute the unit group~$\order^\times$ in~$8$~minutes and then compute generators for the same ideals with the units-assisted algorithm~\cite[p.~25]{vd} provided by Magma. The computation takes~$1$ hour per ideal on average with a maximum of~$17$h, and~$10$ of the~$23$ ideals take less that~$0.5$~seconds. Magma tends to return smaller generators than our algorithm. Magma is fast whenever there exists a small generator and our algorithm is faster when this is not the case.
\begin{figure}[tbh]
\centering
\includegraphics*[height=9.5cm,keepaspectratio=true]{moy.eps}
\caption{Running time of the algorithms}\label{timings}
\end{figure}
\paragraph{Example 4.}In order to understand the practical behaviour of the algorithms, we conduct the following experiment. We draw algebras~$\alg$ and ideals~$I$ at random\footnote[2]{Let~$x$ be uniformly distributed in~$[0,70]$. This value controls the size of the discriminant. Let~$k$ be~$1$ or~$2$, each with probability~$1/2$. This is the number of prime factors of the discriminant of the algebra. Let~$t$ be uniformly distributed in~$[0,1]$. This value controls the part of the size of the discriminant coming from the base field or from the algebra. Let~$d$ be the smallest fundamental discriminant larger that~$\exp(tx/4)$, and let~$\fld=\Q(\sqrt{d})$.
Let~$\prm_1$ be the prime of~$\Z_\fld$ with smallest norm larger than~$\exp((1-t)x/2k)$, and if~$k=2$ let~$\prm_2$ be the prime with smallest norm larger than~$N(\prm_1)$. Let~$\alg$ be the quaternion algebra ramified exactly at~$\prm_i$ for~$i\le k$ and at~$k\,\mathrm{mod}\, 2$ real places of~$\fld$, and let~$\order$ be a maximal order in~$\alg$. Let~$\Delta = d^4N(\delta_\alg)^2$ be the absolute discriminant of~$\alg$. Let~$y$ be uniformly distributed in~$[0,1]$, and let~$\prm$ be the prime of~$\Z_\fld$ of smallest norm larger than~$y\Delta^{1/2}$, coprime to~$\delta_\alg$ and such that the class of~$\prm$ in~$\clalg$ is trivial.
Finally, let~$I$ be a random integral right $\order$-ideal of norm~$\prm$.}. In every random test case, we compute our reduction structure with Algorithm~\ref{gbuild} (\texttt{GBuild}), and we compute a generator of the ideal~$I$ with Algorithm~\ref{isprincipal} (\texttt{IsPrincipal}). We also compute a generator of~$I$ with the function provided by Magma. In every case, we interrupt any algorithm that takes more than~$1000$ seconds to terminate.
The result of~$15\,000$ such test cases is plotted in Figure~\ref{timings}: the discriminant~$\Delta$ and the time are both in logarithmic scale, and each plot~$(D,T)$ is such that~$T$ is the average of the running time of the algorithm over the discriminants~$\Delta\in[D/10,10D]$. We do not plot the running time when more than~$50\%$ of the executions were interrupted, since the corresponding value is no longer meaningful.
\begin{acknowledgements}
I would like to thank Karim Belabas and Andreas Enge for helpful discussions and careful reading of early versions of this paper. I also want to thank an anonymous referee many comments and corrections and for suggesting a deterministic algorithm for computing a local generator and Pierre Lezowski for explaining to me Euclidean algorithms over matrix rings.
\end{acknowledgements}
|
\section{Introduction}
$C_{60}$ molecules rotate. Even in crystalline fullerite \cite{Yannoni,Neumann}, or in ``peapods''inside single walled carbon nanotubes \cite{Zou}, or in molecular layers on top of crystals
\cite{Sanchez},
they present quasi-free rotation even down to low temperatures. The orientational disorder of the rotating $C_{60}$ molecules made the crystalline structural determination of the solid state be completed only about five years after their discovery. This was achieved by smartly breaking the spherical symmetry of $C_{60}$ by the addition of a functional unit yielding as result a crystal with orientational order \cite{HAWKINS,Fagan}.
Solid fullerene, or fullerite $C_{60}$, is a very unusual solid in the sense that it is made of rotationally mobile molecules. Below 249K it has an orientationally ordered phase reminiscent of a ferromagnet or a nematic liquid crystal. Near $T_c=250-260K$ it presents a first order transition to the orientationally disordered phase similar to a paramagnet or the liquid crystalline isotropic phase. At room temperature, the face-centered cubic (fcc) structure of fullerite has, in each cubic unit cell, four molecules rotating nearly freely at frequencies about \cite{JOHNSON} $10^{11}$ Hz. Much new physics comes from this rotational degree of freedom. The rapid rotation has important consequences on the nanotribological properties of fullerite single crystal surfaces \cite{Liang}, on the interaction between the $C_{60}$ molecules \cite{Yan} and on the electronic transport in the crystal \cite{Katz}.
\begin{figure}[hpt
\centering
\includegraphics[width=6cm,height=8cm]{fig1.eps}
\caption{Rotating Fulerene. The angular velocity $\vec{\Omega}$ is in the $z$ direction.}\label{fulereno}
\end{figure}
Even though rotation is so important to the determination of the physical characteristics of $C_{60}$, very few works \cite{shen,shen2} have been published on its influence on electronic properties of the molecule. An obvious question is how rotation affects the electronic energy spectrum of $C_{60}$. This question was addressed by Shen and Zhang \cite{shen} who computed, by aproximative methods, an energy shift proportional to the rotational angular velocity of the molecule.
In this article, using the continuum approach for the electronic structure of $C_{60}$, we confirm the result of Shen and Zhang and show that rotation also introduces a Zeeman-like spliting of the energy levels. As a consequence, the zero modes predicted \cite{vozmediano,vozmedianoprl,osipov} for static $C_{60}$, no longer appear. Nevertheless, in the zero rotational angular velocity limit, the zero modes, as well as the stactic $C_{60}$ energy spectrum are recovered. We use a Dirac equation continuum model based in the tight-binding approximation, first presented in \cite{vozmedianoprl,vozmediano}, which is limited to electronic states close to the Fermi level. In this model the lattice structure disappears and the pentagons in $C_{60}$ are simulated by localized fictitious gauge fluxes.
The organization of the paper is as follows. In Section II we describe the continuum model for rotating fullerene, obtaining the Dirac operator for the problem. The solution of the eigenvalue problem is presented in Section III, where we find an exact analytical solution for both, the eigenfunction and the energy spectrum. Finally, in Section IV we present our conclusions.
\section{The continuum model for rotating fullerene}
\subsection{Graphene: from tight-binding to field theory}
Our starting point is the fact that, in the long wavelenght limit, the electronic spectrum of graphene can be obtained from an effective Dirac equation in (2+1) dimensions. This was found using the tight binding \cite{wallace} and effective-mass \cite{DiVincenzo} approximation. From the tight-binding Hamiltonian for electrons in graphene, considering only nearest-neighbour hopping, the energy bands are given by \cite{RevModPhys.81.109}
\begin{eqnarray}
E_\pm=\pm &t&\Biggl(3+2\cos(\sqrt{3}k_ya) \nonumber \\
&+&4\left. \cos(\frac{\sqrt{3}}{2}k_ya)\cos(\frac{3}{2}k_xa)\right)^{1/2},
\label{equation}
\end{eqnarray}
where $t$ is the nearest-neighbour hopping energy and $a$ is the carbon-carbon distance. Expanding this relation around the Dirac points $\vec{K}$ in the Brillouin zone, {\it i.e.}, writing $\vec{k}=\vec{K}+\vec{q}$, with $\mid\vec{q}\mid\ll\mid\vec{K}\mid$, one obtains
\begin{equation}
E_\pm(\vec{q})=\pm v_F\mid\vec{q}\mid+O[(q/K)^2].
\label{dispersion}
\end{equation}
This linear dispersion relation justifies the effective theory for the low-energy electrons of graphene as two-dimensional massless Dirac fermions. The electronic states are represented by the two-component wave function $\psi=(\psi_A, \psi_B)^T$, called \textit{pseudospin}, which represents the two graphene sublattices $A$ and $B$. The wave function $\psi$ obeys the massless Dirac equation in two dimensions
\begin{equation}
-i \hbar v_F\sigma^\mu\partial_\mu\psi(r)=E\psi(r),
\label{dirac1}
\end{equation}
where $\sigma^\mu$ are the Pauli matrices ($\mu=1, 2$) and $v_F$ is the Fermi velocity.
Besides the pseudospin, there is a second spin-like entity due to the fact that there are two independent wave vectors, $\vec{K}_+=\vec{K}$ and $\vec{K}_-=-\vec{K}$, which give the same dispersion relation (\ref{dispersion}). This degree of freedom is called $K$-spin. The tensor product between the two ``spin'' spaces yields a four dimensional space expanded by the kets $\vert\vec{K}_\pm A\rangle$ and $\vert\vec{K}_\pm B\rangle$. Therefore, each wave function $\psi_A$ and $\psi_B$ will have two subcomponents related with the two Dirac points: $\psi_A=(\psi_{A}^{K_+}, \psi_{A}^{K_-})^T$ and $\psi_B=(\psi_{B}^{K_+}, \psi_{B}^{K_-})^T$. The Pauli matrices which act on the $K$ part of the wave function will be denoted by $\tau^\mu$. The $K$-spin plays an important role in one of the fictitious gauge fields flux related to the existence of pentagons in the fullerene molecule. This will be seen in the next subsections where the Dirac operator (\ref{dirac1}) will be modified in order to incorporate the spherical geometry, the disclinations (pentagons) and rotation of the fullerene molecule.
\subsection{Disclinations: from graphene to fullerene}
\begin{figure}[hpt
\centering
\includegraphics[width=5.5cm,height=5cm]{fig2.eps}
\caption{Cutting a $\pi/3$ angular section of the graphene sheet. Connecting the dangling bonds on the sides OA and OB one gets a cone with a pentagon on the vertex. Grey and black circles represent sublattices A and B, respectively.}\label{corte}
\end{figure}
To transform a graphene sheet into a $C_{60}$ fullerene one needs to introduce curvature into the hexagon-tiled plane. If we cut a $\pi/3$ wedge out of the plane and join the dangling bonds at the opposing edges (see Fig. 2), a cone is created with a pentagon at its tip. The pentagons are then defects, called disclinations, in the otherwise perfect hexagonal tiling of the graphene sheet. In the continuum limit, each disclination carries a $\delta$-function curvature singularity \cite{vickers}
\begin{equation}
R_{12}^{\,\,\,\,\,\,\,12}=\lambda \delta (x)\delta(y), \label{R12}
\end{equation}
where $\lambda$ is the angle characterising the removed wedge which, in the case of the pentagon, is $\frac{\pi}{3}$.
To get a spherical shape, 12 pentagons, symmetrically arranged are needed. The result is $C_{60}$ which, in the continuum limit, has 12 curvature singularities which integrate out to give the total curvature of the sphere. These curvature singularities act as flux tubes giving rise to a Aharonov-Bohm-like phase\cite{carvalho,osipov1992}. Therefore, the ``elastic" geometric fluxes incorporate the topological origin of the defects and should be included in the Dirac equation (\ref{dirac1}).
Disclinations are characterized by the Frank vector $\Theta_i$. In the geometric theory of defects \cite{Katanaev} the curvature associated to the disclination is the surface density of Frank vectors, such that
\begin{equation}
\Omega^ {ij}=\int\int dx^{\mu} \wedge dx^{\nu} R_{\mu\nu}^{\,\,\,\,\,\,ij}, \label{Omega}
\end{equation}
with
\begin{equation}
\Theta_i = \epsilon_{ijk}\Omega^ {jk}. \label{Thetai}
\end{equation}
The curvature tensor, in terms of the $SO(3)$ connection $\omega_{\nu j}^{\,\,\,\,\,\,i}$, is
\begin{equation}
R_{\mu\nu j}^{\,\,\,\,\,\,\,\,\,i}=\partial_{\mu}\omega_{\nu j}^{\,\,\,\,\,\,i} - \omega_{\mu j}^{\,\,\,\,\,\,k}\omega_{\nu k}^{\,\,\,\,\,\,i} - (\mu \leftrightarrow \nu). \label{Rtensor}
\end{equation}
Since we have a two-dimensional system, the rotations are restricted by the subgroup $SO(2)$, which is Abelian. This implies that the quadratic terms in (\ref{Rtensor}) vanish and the curvature tensor can be seen as the curl of the $SO(2)$ connection. This way, the Frank vector is then given by
\begin{equation}
\Omega^ {ij}= \oint dx^{\mu} \omega_{\mu}^{\,\,\,\,\,ij}. \label{omegaij}
\end{equation}
Using (\ref{R12}), (\ref{Omega}), (\ref{Thetai}) and $\lambda=\frac{\pi}{3}$ one gets
\begin{equation}
\Theta^3 = \frac{\pi}{3}.
\end{equation}
Equivalently, in the gauge theory of disclinations \cite{fieldW} the Frank vector is obtained from the flux of an Abelian gauge field or, the circuit integral of its vector potential $W_{\mu}$,
\begin{equation}
\Theta^3 = \oint dx^\mu W_\mu = \frac{\pi}{3}, \label{flux1}
\end{equation}
analogous to (\ref{omegaij}), which gives a geometric interpretation to the gauge field $W_{\mu}$. From here on we proceed like \cite{osipov} including the Abelian gauge field $W_{\mu}$ in the Dirac equation via minimal coupling.
A second and third gauge fields are needed to fix the jump in the wave function on the cone as one goes around its tip. As one can see in Fig. \ref{corte}, the alternating black and gray sublattices of graphene will have a discontinuity after the wedge is removed and the dangling bonds are joined. Gray atoms will be connected to gray atoms at the junction of the bonds. The pseudospin and K-spin parts of the wave function acquire phases when passing the gray-gray junction as the cone tip is circulated. In order to make the discontinuity disappear in the continuous model these phases need to be removed. This is done with the help of fluxes through the apex of the cone of fictitious non-Abelian gauge fields $\omega_{\mu}$ and $a_{\mu}$, designed \cite{Lammert} to compensate those phases. The fluxes produced by these fields are
\begin{equation}
\oint dx^\mu \omega_\mu = -\frac{\pi}{6}\sigma^3 \label{flux2}
\end{equation}
and
\begin{equation}
\oint dx^\mu a_\mu = \frac{\pi}{2}\tau^2 , \label{flux3}
\end{equation}
where the integrals should be computed along closed curves around the vertex $O$ in Fig. \ref{corte}. Notice that $\sigma^3$ is the usual Pauli matrix acting on the pseudospin and $\tau^2$ is also the Pauli matrix, this time acting on the $K$-spin. The fluxes (\ref{flux1}), (\ref{flux2}) and (\ref{flux3}) modify the wave function as the Aharonov-Bohm flux $\oint dx^{\mu} A_{\mu}$ does, as generators of infinitesimal rotations in the respective Hilbert spaces where they act.
The $\omega_{\mu}$ field is purely geometrical. In fact, it is the spin connection, which is part of the the covariant derivative in curvilinear coordinates. So it is naturally included in the Dirac equation on the sphere and therefore it does not contribute to the total flux. Each of the twelve disclinations of $C_{60}$ contributes with a flux of each kind, (\ref{flux1}) and (\ref{flux3}). The problem of incorporating these discreet fluxes into the Dirac equation may be approximated by replacing them altogether by their average over the sphere. That is, effectively substituting them with the flux of a ('t Hooft-Polyakov) monopole at the center of the sphere \cite{osipov,vozmedianoprl,vozmediano}. The electronic structure of $C_{60}$ near the Fermi level may then be obtained by solving Dirac equation on the sphere including the gauge field of the monopole.
Like the electric charge, which is $\frac{1}{4\pi}$ times the flux of electric field through a closed surface enclosing it, the monopole charge can be obtained from the expressions (\ref{flux1}) and (\ref{flux3}), respectively:
\begin{equation}
g_W=\frac{1}{4 \pi}\cdot12 \cdot \frac{\pi}{3}=1
\end{equation}
for the $W_{\mu}$ field and
\begin{equation}
g_a=\frac{1}{4\pi} \cdot12 \cdot \frac{\pi}{2}=\frac{3}{2}
\end{equation}
for the $a_{\mu}$ field.
In spherical coordinates, the Abelian gauge field $W_{\mu}$ is given by \cite{osipov}
\begin{equation}
W_\theta = 0 , W_\varphi = g_W \cos\theta \label{W}
\end{equation}
and the non-Abelian gauge field $a_{\mu}$ by \cite{osipov,vozmedianoprl,vozmediano}
\begin{equation}
a_\theta = 0 , \\ \\ \\ a_\varphi = g_a \cos\theta \tau^2 . \label{a}
\end{equation}
The free particle Dirac operator on the sphere is \cite{abrikosov}
\begin{equation}
-i\hbar c \left[ \sigma_x \left( \partial_{\theta} + \frac{\cot \theta}{2} \right) + \frac{i\sigma_y}{\sin \theta} \partial_{\varphi} \right].
\end{equation}
Therefore, the equivalent of Eq. (\ref{dirac1}) for fullerene is \cite{osipov}
\begin{eqnarray}
-i \hbar v_F \Bigl[ &\sigma_x &\left( \partial_{\theta} + \frac{\cot \theta}{2} \right) \nonumber \\
&+&\left. \frac{\sigma_y}{\sin \theta} \left(\partial_{\varphi}-ia_{\varphi}-iW_{\varphi}\right) \right] \psi =E\psi .
\label{equation2}
\end{eqnarray}
\subsection{Rotation}
Since we want to study the effects of rotation on the electronic spectrum of $C_{60}$ we need to solve Dirac equation in the rotating reference frame attached to the molecule. In order to obtain some insight into the not so intuitive motion in a non-inertial frame, we start this subsection by reviewing the classical physics of a free particle in a rotating medium.
The velocity transformation between the static laboratory frame (primed) and the rotating frame (unprimed) is
\begin{equation}
\vec{v}\,' =\vec{v} + \left( \vec{\Omega} \times \vec{r} \right).
\end{equation}
That is, the speed of the particle as measured in the laboratory is the sum of its speed in the rotating frame plus the speed $\vec{\Omega} \times \vec{r}$ of the motion done with the rotating frame. We consider that the origin of both frames coincide with the center of the molecule. This leads to the following Lagrangian for the free particle in the rotating frame \cite{landau2}
\begin{equation}
L=\frac{mv^2}{2}+ m\vec{v}\cdot (\vec{\Omega}\times\vec{r})+ \frac{m}{2}(\vec{\Omega}\times\vec{r})^2 . \label{lagr}
\end{equation}
This gives the canonical momentum
\begin{equation}
\vec{p}=\frac{\partial L}{\partial \vec{v}}=m\vec{v}+m(\vec{\Omega}\times\vec{r}) .\label{cmomentum}
\end{equation}
Using (\ref{lagr}) and (\ref{cmomentum}) one obtains the classical Hamiltonian for the free particle in the rotating frame
\begin{equation}
H=\vec{p} \cdot \vec{v} - L = \frac{mv^2}{2}-\frac{1}{2}m(\vec{\Omega}\times\vec{r})^2 ,
\end{equation}
or, in terms of $\vec{p}$,
\begin{equation}
H=\frac{p^2}{2m}-\vec{\Omega}\cdot\vec{L} , \label{hamil2}
\end{equation}
where $\vec{L}=\vec{r}\times\vec{p}$ is the orbital angular momentum of the particle. Equation (\ref{hamil2}) expresses also the form of the Schr\"odinger Hamiltonian for the free particle in the rotating frame, as seen in reference \cite{ni} which also gives the Dirac Hamiltonian
\begin{equation}
H = \beta mc^2+ c \vec{\alpha} \cdot \vec{p} - \vec{\Omega}\cdot (\vec{L}+\vec{S}) , \label{diracH}
\end{equation}
where $\vec{S}$ is the real spin operator and $\vec{\alpha} $ and $\beta$ are the Dirac matrices. This is essentially the free particle Dirac Hamiltonian plus a coupling between rotation and total angular momemtum, like in (\ref{hamil2}).
Considering the rotation axis to be in the $z$ direction we then have, from (\ref{equation2}) and (\ref{diracH}), that for the rotating fullerene
\begin{eqnarray}
& - & i \hbar v_F \left[ \sigma_x \left( \partial_{\theta} + \frac{\cot \theta}{2} \right) \right. \nonumber \\
& + & \left. \frac{i\sigma_y}{\sin \theta} \left(\partial_{\varphi}-ia_{\varphi}-iW_{\varphi}\right)
\right] \psi - \Omega J_z \psi =E\psi ,
\label{equation}
\end{eqnarray}
where
\begin{equation}
J_z = -i\hbar \partial_{\varphi} + S_z \label{jz}
\end{equation}
is the $z$-component of the total angular momentum. The innocent looking expression (\ref{jz}) is not as obvious as it seems. The orbital angular momentum is obtained from the mechanical moment $\vec{\pi}$ as $\vec{L}=\vec{r}\times \vec{\pi}$. For a particle of charge $q$ moving in the presence of an ordinary magnetic field, $\vec{\pi}=\vec{p}-\frac{q}{c}\vec{A}$, where $\vec{A}$ is the vector potential associated to the magnetic field. For a particle moving in the presence of a magnetic monopole field of charge $g$, the operator $\vec{L}=\vec{r}\times \vec{\pi}$ is no longer the angular momentum since it does not obey the usual commutation relations of the angular momentum, which must be respected in order to guarantee its conservation. It is then necessary to fix this by adding to it the angular momentum stored in the magnetic field of the monopole \cite{jackson}, $-g\hbar \hat{r}$, so as to fix this problem. In the same way, non-Abelian monopoles contribute to the mechanical momentum via minimal coupling and also store angular momentum in their fields \cite{goddard}. Taking this into consideration, we have \cite{vozmediano,osipov}
\begin{eqnarray}
J_z = & - & i\hbar \left(\partial_{\varphi}-\frac{1}{4}\left[ \sigma_x , \sigma_y \right] cos\theta -ia_{\varphi}-iW_{\varphi}\right)\nonumber \\
& + & \frac{\hbar}{2} \sigma_z \cos \theta + \hbar g_a \tau^2 \cos\theta +\hbar g_W \cos\theta + S_z
\end{eqnarray}
Taking into consideration (\ref{W}) and (\ref{a}) and the fact that $\left[ \sigma_x , \sigma_y \right] =2i\sigma_z $, one gets (\ref{jz}).
\section{Electronic Structure}
In equation (\ref{equation}) the only operator acting in the $K$ part of the wave function is $\tau^2$ which appears only in the gauge field $a_{\varphi}$. So, the equation is already diagonalized for this operator and we can replace it with its eigenvalue $k=\pm 1$, which correspond, respectively, to the $K$-spin states $\vec{K}_+$ and $\vec{K}_-$. In the same way the real spin operator $S_z$ appears only in the total angular momentum $J_z$ and is replaced with the eigenvalue $s_z= \frac{\hbar}{2}s$, where $s=+1(-1)$ for spin up(down). The result is the two-component Dirac equation
\begin{eqnarray}
& - & i \hbar v_F\sigma_{x}\left(\partial_{\theta} + \frac{\cot\theta}{2}\right) \psi - \frac{i \hbar v_F\sigma_{y}}{\sin\theta}(\partial_{\varphi} - i A \cos\theta) \psi \nonumber \\
& + & i \hbar \Omega \partial_\varphi \psi - \frac{s \hbar \Omega}{2} \psi = E \psi
\label{dirac2}
\end{eqnarray}
where $A=(a^k_\varphi+W_\varphi)/\cos\theta=\frac{3}{2}k+1$ with $a^k_{\varphi}=\frac{3}{2}k\cos\theta$ being the $K$-spin eigenvalue of $a_{\varphi}$. Writing
\begin{equation}
\left(
\begin{array}{c}
\psi_A \\
\psi_B \\
\end{array}
\right) = \displaystyle\sum_{j}\frac{e^{i \ell \varphi}}{\sqrt{2\pi}}
\left(
\begin{array}{c}
u_{\ell} (\theta) \\
v_{\ell} (\theta) \\
\end{array}
\right), \ell =0, \pm1, \pm2, ...
\end{equation}
we will have two coupled equations for $u_{\ell}$ and $v_{\ell}$ given by
\begin{eqnarray}
-i \hbar v_F\left(\partial_\theta + \left[\frac{1}{2}-A\right]\cot \theta+\frac{\ell}{\sin \theta}\right)v_{\ell}= \nonumber \\
\left(E+\ell \hbar \Omega +\frac{s \hbar \Omega}{2}\right)u_{\ell}
\label{acoplada}
\end{eqnarray}
and
\begin{eqnarray}
-i \hbar v_F\left(\partial_\theta + \left[\frac{1}{2}+A\right] \cot \theta - \frac{\ell}{\sin \theta}\right) u_{\ell} = \nonumber \\
\left(E+\ell \hbar \Omega + \frac{s \hbar \Omega}{2} \right)v_{\ell}
\label{acoplada1}
\end{eqnarray}
which can be easily decoupled, becoming second order differential equations.
The uncoupled equation for $u_{\ell}$, after making $x=\cos\theta$, is
\begin{eqnarray}
\left[\partial_x (1-x^2) \partial_x - \frac{(\ell -Ax)^2 + \frac{1}{4} + A - \ell x}{1-x^2}\right] u_{\ell}\nonumber \\
= \left[- \frac{1}{v_F^2\hbar^2}\left(E+\ell \hbar \Omega +\frac{s \hbar \Omega}{2}\right)^2+\frac{1}{4}\right]u_{\ell}.
\label{eq}
\end{eqnarray}
The asymptotic behavior of this equation suggests the following $ansatz$
\begin{equation}
u_j = (1-x)^\mu (1+x)^\nu \overline{u}_j (x),
\label{asymp}
\end{equation}
where
\begin{equation}
\mu = \frac{1}{2}\left| \ell - A - \frac{1}{2} \right| , \nu = \frac{1}{2}\left| \ell + A + \frac{1}{2} \right| .
\end{equation}
Inserting (\ref{asymp}) in (\ref{eq}) results in the following equation for $\overline{u}_j (x)$
\begin{eqnarray}
(1-x^2)\partial^{2}_{x} \overline{u}_{\ell} + [2(\nu - \mu) - 2(\mu + \nu +1)x]\partial_x \overline{u}_{\ell} \nonumber \\ + \left(- 2 \mu \nu - \mu -\nu - \frac{1}{2}(\ell^2 - A^2 + \frac{1}{4} - A) \right.\nonumber \\
\left. + \frac{1}{v_F^2\hbar^2}\left(E+\ell \hbar \Omega +\frac{s \hbar \Omega}{2}\right)^2-\frac{1}{4}\right)\overline{u}_{\ell} = 0,
\label{hiper}
\end{eqnarray}
which can be rewritten as
\begin{equation}
z(1-z)\partial^{2}_{x}\overline{u}_{\ell} + [c-(a+b+1)z]\partial_x \overline{u}_{\ell} - ab\overline{u}_{\ell} =0,
\label{hiper1}
\end{equation}
where
\begin{eqnarray}
a&=&\frac{1}{2}+\mu +\nu -\frac{1}{2}\left(4\mu^2 (4+8\nu)\mu +4\nu^2 +4\nu +1+4\alpha \right)^{\frac{1}{2}}, \nonumber \\
b&=&\frac{1}{2}+\mu +\nu +\frac{1}{2}\left(4\mu^2 (4+8\nu)\mu +4\nu^2 +4\nu +1+4\alpha \right)^{\frac{1}{2}}, \nonumber \\
c&=&2\nu +1 ,\\
z&=&\frac{1}{2}+\frac{1}{2}x ,\nonumber
\end{eqnarray}
and
\begin{eqnarray}
\alpha &=&- 2 \mu \nu - \mu -\nu - \frac{1}{2}(\ell^2 - A^2 + \frac{1}{4} - A) \nonumber \\
&&+ \frac{1}{v_F^2\hbar^2}\left(E+\ell \hbar \Omega +\frac{s \hbar \Omega}{2}\right)^2-\frac{1}{4} .
\end{eqnarray}
Equation (\ref{hiper1}) is clearly the hypergeometric equation whose solution is
\begin{eqnarray}
\overline{u}_j&=&C_1\;\; {}_2F_1(a,b;c;z) \nonumber \\
&&+C_2z^{1-c}{}_2F_1(a+1-c,b+1-c;2-c;z),
\end{eqnarray}
where $C_1$ and $C_2$ are normalization constants and ${}_2F_1$ is the hypergeometric function.
Physics requires that the wave function be square integrable. So, the hypergeometric series has to be convergent. This happens for arbitrary $a$, $b$ and $c$ when $-1<z<1$, and for $c>a+b$ when $z=\pm1$. As $-1\leq x\leq 1$, we have that $0\leq z\leq 1$. Therefore, in order to have a convergent series, the condition $c>a+b$ has to be satisfied.
We have that
\begin{equation}
a+b=1+2\mu + 2\nu > c
\end{equation}
because $\mu$ is a positive constant. So, we have a divergent solution.
This problem is solved when we choose $a=-n$, where $n$ is an integer, which guarantees that we have a polynomial solution of degree $n$. From this condition, we obtain the energy spectrum as
\begin{eqnarray}
E^{(u)}_{n,\ell ,s}=\hbar v_F\biggl[\biggl(n&+&\mu +\nu +\frac{1}{2}\biggr)^2 \nonumber \\
&-&A-A^2\biggr]^{1/2}-\left(\ell +\frac{s}{2}\right)\hbar \Omega \; .
\label{spectrum1}
\end{eqnarray}
When $\Omega =0$ in (\ref{spectrum1}), one gets the energy spectrum obtained in \cite{osipov} for static fullerene.
Rotation adds two terms to the energy spectrum. The first one, the energy shift $-\ell \hbar \Omega$, was already obtained in \cite{shen} from the Aharonov-Carmi phase\cite{arahonov.carmi, arahonov.carmi1}. The second is the spin-rotation coupling, analogous to the Zeeman coupling between magnetic field and spin.
Analogous to what as was done to $u_{\ell}$, one can uncouple Eqs.(\ref{acoplada}) and (\ref{acoplada1}) for $v_{\ell}$ and use the $ansatz$
\begin{equation}
v_{\ell} = (1-x)^\gamma (1+x)^\delta \overline{v}_{\ell} (x),
\end{equation}
where
\begin{equation}
\gamma = \frac{1}{2}\left| \ell - A + \frac{1}{2} \right| , \delta = \frac{1}{2}\left| \ell + A - \frac{1}{2} \right| \; ,
\end{equation}
to get another energy spectrum
\begin{eqnarray}
E^{(v)}_{n,\ell ,s}=\hbar v_F\biggl[\biggl(n&+&\gamma +\delta +\frac{1}{2}\biggr)^2 \nonumber \\
&-&A-A^2\biggr]^{1/2}-\left(\ell + \frac{s}{2}\right)\hbar \Omega .
\label{spectrum2}
\end{eqnarray}
The spectra (\ref{spectrum1}) and (\ref{spectrum2}) have to be the same, {\it i.e.}, the condition $E^{(u)}_{n,\ell}=E^{(v)}_{n,\ell}$ must be satisfied. There are two possible cases where this happens. The first one is when $\mu+\nu=\gamma+\delta$. In this case the possible values of $\ell$ are $|\ell|\geq ||A|+1/2|$. The energy spectrum becomes
\begin{eqnarray}
E_{n,\ell,s}=\hbar v_F\biggl[\biggl(n&+&|\ell| +\frac{1}{2}\biggr)^2 \nonumber \\
&-& A-A^2\biggr]^{1/2}-\left(\ell +\frac{s}{2}\right)\hbar \Omega \label{spec}
\end{eqnarray}
and the eigenfunctions are
\begin{eqnarray}
u_{\ell}&=&(1-x)^\mu (1+x)^\nu [ C_1\;\; {}_2F_1(-n,b;c;z) \nonumber \\
&+&C_2z^{1-c}{}_2F_1(-n+1-c,b+1-c;2-c;z) ] \nonumber
\end{eqnarray}
and
\begin{eqnarray}
v_{\ell}=&&(1-x)^\gamma (1+x)^\delta [ C_1\;\; {}_2F_1(-n,b';c';z) \nonumber \\
&+&C_2z^{1-c'}{}_2F_1(-n+1-c',b'\nonumber \\
&+&1-c';2-c';z) ] ,
\end{eqnarray}
where $c'$ and $b'$ have the same form as $c$ and $b$ with the exchange of $\mu$ by $\gamma$ and $\nu$ by $\delta$.
The second case is when
\begin{equation}
E_{0,\ell , s}=-\left(\ell +\frac{s}{2}\right)\hbar \Omega, \label{nonzeromode}
\end{equation}
which makes the right hand side of equations (\ref{acoplada}) and (\ref{acoplada1}) vanish. In this case, the possible values for $\ell$ and $n$ are determined by $|\ell|\leq ||A|-1/2|$ and $n=0$. Without rotation, this is a zero mode. Therefore, rotation not only shifts the zero mode to $-\ell \hbar \Omega$ but also splits it into two states. The eigenfunctions for this case are given by
\begin{eqnarray}
u_{\ell}=0, \;\;\;\;\;v_{\ell}=C (1-x)^\gamma (1+x)^\delta .
\label{zero}
\end{eqnarray}
Since $A=\frac{3}{2}k+1=-1/2, 5/2$ and we must have $|\ell|\leq ||A|-1/2|$ then, when $A=-1/2$, $\ell =0$, and when $A=5/2$, $\ell =0,\pm 1, \pm 2$. Therefore the degeneracy of the states given by (\ref{nonzeromode}) would be $6$ but the Zeeman splitting reduces it to $3$.
The overall spectrum obtained in this continuum approach is given by Eqs. (\ref{spec}) and (\ref{nonzeromode}). Both equations indicate that the contribution of rotation is a shift of $-\left(\ell +s_z\right)\hbar \Omega$ to the static molecule spectrum. We observe that the integer part, $\ell \hbar \Omega$, of the shift has been predicted
in reference \cite{shen} using an approximation based on the Aharonov-Carmi effect\cite{arahonov.carmi1,arahonov.carmi}. The non-integer part of the shift, which includes a splitting of levels analogous to Zeeman's, is related to the geometric phase acquired by the electrons in rotating $C_{60}$ studied in \cite{shen2}. Both effects derive from the spin-rotation coupling.
\section{Concluding Remarks}
In this paper, we report the results of the electronic structure of rotating fullerene obtained using the continuum model based in an effective Dirac equation. The model includes fictitious non-Abelian gauge field fluxes which simulate the pentagons of the $C_{60}$ molecule. There are two main differences between the energy spectrum here obtained and the static ones from previous studies. The first one is a shift that comes from the coupling between the orbital angular momentum and rotation and which removes the zero modes. The second one is a Zeeman-like splitting of the levels due to the coupling between the real spin and rotation. This suggests the exciting possibility of observing Electron Spin Resonance without a magnetic field by tuning the rotation speed with a variation of temperature. The occurrence of rapidly rotating molecules in solid fullerene is an experimental reality and therefore we expect that our work may contribute to a better understanding of the electronic properties of those fascinating solids. Moreover, a rotating system is of course accelerated and therefore, non-inertial. Non-inertial systems are undistinguishable from systems in the presence of a gravitational field, according to the Equivalence Principle of General Relativity. So, what in fact we are obtaining in the present work is a gravitational effect in quantum mechanics which may be observed with conventional experimental techniques. This problem becomes even more interesting if one considers an accelerated rotation and also the presence of a magnetic field, which we expect to address in forthcoming publications.
{\bf Acknowledgements}: This work was partially supported by CNPq, CNPQ-MICINN binational, INCTFCx, CAPES-WEISSMANN binational and CAPES-NANOBIOTEC. We are also grateful to Maria Vozmediano for the hospitality and illuminating discussions during the stay of JRFL and FM at Instituto de Ciencias de los Materiales de Madrid, where this work began.
|
\section{Two lexicographic orders for binary words
\begin{figure
{\jjbls
\begin{verbatim}
0: [ . . . . . ] { } [ . . . . . ] { }
1: [ 1 . . . . ] { 0 } [ . . . . 1 ] { 4 }
2: [ 1 1 . . . ] { 0, 1 } [ . . . 1 . ] { 3 }
3: [ 1 1 1 . . ] { 0, 1, 2 } [ . . . 1 1 ] { 3, 4 }
4: [ 1 1 1 1 . ] { 0, 1, 2, 3 } [ . . 1 . . ] { 2 }
5: [ 1 1 1 1 1 ] { 0, 1, 2, 3, 4 } [ . . 1 . 1 ] { 2, 4 }
6: [ 1 1 1 . 1 ] { 0, 1, 2, 4 } [ . . 1 1 . ] { 2, 3 }
7: [ 1 1 . 1 . ] { 0, 1, 3 } [ . . 1 1 1 ] { 2, 3, 4 }
8: [ 1 1 . 1 1 ] { 0, 1, 3, 4 } [ . 1 . . . ] { 1 }
9: [ 1 1 . . 1 ] { 0, 1, 4 } [ . 1 . . 1 ] { 1, 4 }
10: [ 1 . 1 . . ] { 0, 2 } [ . 1 . 1 . ] { 1, 3 }
11: [ 1 . 1 1 . ] { 0, 2, 3 } [ . 1 . 1 1 ] { 1, 3, 4 }
12: [ 1 . 1 1 1 ] { 0, 2, 3, 4 } [ . 1 1 . . ] { 1, 2 }
13: [ 1 . 1 . 1 ] { 0, 2, 4 } [ . 1 1 . 1 ] { 1, 2, 4 }
14: [ 1 . . 1 . ] { 0, 3 } [ . 1 1 1 . ] { 1, 2, 3 }
15: [ 1 . . 1 1 ] { 0, 3, 4 } [ . 1 1 1 1 ] { 1, 2, 3, 4 }
16: [ 1 . . . 1 ] { 0, 4 } [ 1 . . . . ] { 0 }
17: [ . 1 . . . ] { 1 } [ 1 . . . 1 ] { 0, 4 }
18: [ . 1 1 . . ] { 1, 2 } [ 1 . . 1 . ] { 0, 3 }
19: [ . 1 1 1 . ] { 1, 2, 3 } [ 1 . . 1 1 ] { 0, 3, 4 }
20: [ . 1 1 1 1 ] { 1, 2, 3, 4 } [ 1 . 1 . . ] { 0, 2 }
21: [ . 1 1 . 1 ] { 1, 2, 4 } [ 1 . 1 . 1 ] { 0, 2, 4 }
22: [ . 1 . 1 . ] { 1, 3 } [ 1 . 1 1 . ] { 0, 2, 3 }
23: [ . 1 . 1 1 ] { 1, 3, 4 } [ 1 . 1 1 1 ] { 0, 2, 3, 4 }
24: [ . 1 . . 1 ] { 1, 4 } [ 1 1 . . . ] { 0, 1 }
25: [ . . 1 . . ] { 2 } [ 1 1 . . 1 ] { 0, 1, 4 }
26: [ . . 1 1 . ] { 2, 3 } [ 1 1 . 1 . ] { 0, 1, 3 }
27: [ . . 1 1 1 ] { 2, 3, 4 } [ 1 1 . 1 1 ] { 0, 1, 3, 4 }
28: [ . . 1 . 1 ] { 2, 4 } [ 1 1 1 . . ] { 0, 1, 2 }
29: [ . . . 1 . ] { 3 } [ 1 1 1 . 1 ] { 0, 1, 2, 4 }
30: [ . . . 1 1 ] { 3, 4 } [ 1 1 1 1 . ] { 0, 1, 2, 3 }
31: [ . . . . 1 ] { 4 } [ 1 1 1 1 1 ] { 0, 1, 2, 3, 4 }
\end{verbatim}
\caption{\label{fig:subsetlex}
Two lexicographic orders for the subsets of a 5-element set:
ordering using the sets as lists of elements (left)
and ordering using the characteristic words (right).
Dots are used to denote zeros in the characteristic words.}
\end{figure}
Two simple ways to represent subsets of a set $\{0,1,2,\ldots,n-1\}$ are
as lists of elements and as the characteristic word, the binary word with
ones at the positions corresponding to the elements in the subset.
Sorting all subsets by list of characteristic words gives the right columns of
figure \ref{fig:subsetlex}, sorting by the list of elements gives the left
columns.
We will concern ourselves with the generalization of the ordering by the lists
which we call \jjterm{subset-lex} order.
The algorithms for computing the successor or predecessor
of a (nonempty) subset $S$ in subset-lex order
are well-known\footnote{For example,
an algorithm for the $k$-subsets of an $n$-set is given
in \cite[Algorithm LEXSUB, p.18]{nw-combalg},
see section \ref{sect:ksubset} for an implementation.}.
In the following let $z$ be the last, and $y$ the next to last element in $S$.
We call an element minimal or maximal if it is respectively
the smallest or greatest element in the superset
\begin{alg}[Next-SL2]\label{alg:next-sl2}
Compute the successor of the subset $S$.
\end{alg}
\begin{enumerate}
\item If there is just one element, and it is maximal, stop.
\item If $z$ is not maximal, append $z+1$.
\item Otherwise remove $z$ and $y$, then append $y+1$.
\end{enumerate}
\begin{alg}[Prev-SL2
Compute the predecessor of the subset $S$.
\end{alg}
\begin{enumerate}
\item If there is just one element, and it is minimal, stop.
\item If $z-1 \in S$, remove $z$.
\item Otherwise remove $z$, append $z-1$ and the maximal element.
\end{enumerate}
C++ implementations of all algorithms discussed here are given in the FXT
library \cite{fxt}.
Here and elsewhere we give slightly simplified versions of the actual code,
like omitting modifiers such as \texttt{public}, \texttt{protected}, and
\texttt{private}. The type \texttt{ulong} is short for \texttt{unsigned long}.
{\codesize
\begin{listing}{1}
// FILE: src/comb/subset-lex.h
class subset_lex
// Nonempty subsets of the set {0,1,2,...,n-1} in subset-lex order.
// Representation as list of parts.
// Loopless generation.
{
ulong n_; // number of elements in set, should have n>=1
ulong k_; // index of last element in subset
ulong n1_; // == n - 1 for n >=1, and == 0 for n==0
ulong *x_; // x[0...k-1]: subset of {0,1,2,...,n-1}
subset_lex(ulong n)
// Should have n>=1,
// for n==0 one set with the element zero is generated.
{
n_ = n;
n1_ = (n_ ? n_ - 1 : 0 );
x_ = new ulong[n_ + (n_==0)];
first();
}
\end{listing}
}%
We will always use the method names
\texttt{first()} and \texttt{last()}
for the first and last element in
the respective list of combinatorial objects.
{\codesize
\begin{listing}{1}
ulong first()
{
k_ = 0;
x_[0] = 0;
return k_ + 1;
}
ulong last()
{
k_ = 0;
x_[0] = n1_;
return k_ + 1;
}
\end{listing}
}%
The methods for computing the successor and predecessor
will always be called \texttt{next()} and \texttt{prev()}
respectively.
{\codesize
\begin{listing}{1}
ulong next()
// Return number of elements in subset.
// Return zero if current is last.
{
if ( x_[k_] == n1_ ) // last element is max?
{
if ( k_==0 ) return 0;
--k_; // remove last element
x_[k_] += 1; // increment last element
}
else // add next element from set:
{
++k_;
x_[k_] = x_[k_-1] + 1;
}
return k_ + 1;
}
ulong prev()
// Return number of elements in subset.
// Return zero if current is first.
{
if ( k_ == 0 ) // only one element ?
{
if ( x_[0]==0 ) return 0;
x_[0] -= 1; // decrement last (and only) element
++k_;
x_[k_] = n1_; // append maximal element
}
else
{
if ( x_[k_] == x_[k_-1] + 1 ) --k_; // remove last element
else
{
x_[k_] -= 1; // decrement last element
++k_;
x_[k_] = n1_; // append maximal element
}
}
return k_ + 1;
}
\end{listing}
}%
A program demonstrating usage of the class is
{\codesize
\begin{listing}{1}
// FILE: demo/comb/subset-lex-demo.cc
ulong n = 6;
subset_lex S(n);
do
{
// visit subset
}
while ( S.next() );
\end{listing}
}%
\section{Subset-lex order for multisets}
\begin{figure
\vspace*{-5mm
{\jjbls
\begin{verbatim}
0: [ . . . . ] { }
1: [ 1 . . . ] { 0 }
2: [ 1 1 . . ] { 0, 1 }
3: [ 1 2 . . ] { 0, 1, 1 }
4: [ 1 2 1 . ] { 0, 1, 1, 2 }
5: [ 1 2 2 . ] { 0, 1, 1, 2, 2 }
6: [ 1 2 2 1 ] { 0, 1, 1, 2, 2, 3 }
7: [ 1 2 2 2 ] { 0, 1, 1, 2, 2, 3, 3 }
8: [ 1 2 2 3 ] { 0, 1, 1, 2, 2, 3, 3, 3 }
9: [ 1 2 1 1 ] { 0, 1, 1, 2, 3 }
10: [ 1 2 1 2 ] { 0, 1, 1, 2, 3, 3 }
11: [ 1 2 1 3 ] { 0, 1, 1, 2, 3, 3, 3 }
12: [ 1 2 . 1 ] { 0, 1, 1, 3 }
13: [ 1 2 . 2 ] { 0, 1, 1, 3, 3 }
14: [ 1 2 . 3 ] { 0, 1, 1, 3, 3, 3 }
15: [ 1 1 1 . ] { 0, 1, 2 }
16: [ 1 1 2 . ] { 0, 1, 2, 2 }
17: [ 1 1 2 1 ] { 0, 1, 2, 2, 3 }
18: [ 1 1 2 2 ] { 0, 1, 2, 2, 3, 3 }
19: [ 1 1 2 3 ] { 0, 1, 2, 2, 3, 3, 3 }
20: [ 1 1 1 1 ] { 0, 1, 2, 3 }
21: [ 1 1 1 2 ] { 0, 1, 2, 3, 3 }
22: [ 1 1 1 3 ] { 0, 1, 2, 3, 3, 3 }
23: [ 1 1 . 1 ] { 0, 1, 3 }
24: [ 1 1 . 2 ] { 0, 1, 3, 3 }
25: [ 1 1 . 3 ] { 0, 1, 3, 3, 3 }
26: [ 1 . 1 . ] { 0, 2 }
27: [ 1 . 2 . ] { 0, 2, 2 }
28: [ 1 . 2 1 ] { 0, 2, 2, 3 }
29: [ 1 . 2 2 ] { 0, 2, 2, 3, 3 }
30: [ 1 . 2 3 ] { 0, 2, 2, 3, 3, 3 }
31: [ 1 . 1 1 ] { 0, 2, 3 }
32: [ 1 . 1 2 ] { 0, 2, 3, 3 }
33: [ 1 . 1 3 ] { 0, 2, 3, 3, 3 }
34: [ 1 . . 1 ] { 0, 3 }
35: [ 1 . . 2 ] { 0, 3, 3 }
36: [ 1 . . 3 ] { 0, 3, 3, 3 }
37: [ . 1 . . ] { 1 }
38: [ . 2 . . ] { 1, 1 }
39: [ . 2 1 . ] { 1, 1, 2 }
40: [ . 2 2 . ] { 1, 1, 2, 2 }
41: [ . 2 2 1 ] { 1, 1, 2, 2, 3 }
42: [ . 2 2 2 ] { 1, 1, 2, 2, 3, 3 }
43: [ . 2 2 3 ] { 1, 1, 2, 2, 3, 3, 3 }
44: [ . 2 1 1 ] { 1, 1, 2, 3 }
45: [ . 2 1 2 ] { 1, 1, 2, 3, 3 }
46: [ . 2 1 3 ] { 1, 1, 2, 3, 3, 3 }
47: [ . 2 . 1 ] { 1, 1, 3 }
48: [ . 2 . 2 ] { 1, 1, 3, 3 }
49: [ . 2 . 3 ] { 1, 1, 3, 3, 3 }
50: [ . 1 1 . ] { 1, 2 }
51: [ . 1 2 . ] { 1, 2, 2 }
52: [ . 1 2 1 ] { 1, 2, 2, 3 }
53: [ . 1 2 2 ] { 1, 2, 2, 3, 3 }
54: [ . 1 2 3 ] { 1, 2, 2, 3, 3, 3 }
55: [ . 1 1 1 ] { 1, 2, 3 }
56: [ . 1 1 2 ] { 1, 2, 3, 3 }
57: [ . 1 1 3 ] { 1, 2, 3, 3, 3 }
58: [ . 1 . 1 ] { 1, 3 }
59: [ . 1 . 2 ] { 1, 3, 3 }
60: [ . 1 . 3 ] { 1, 3, 3, 3 }
61: [ . . 1 . ] { 2 }
62: [ . . 2 . ] { 2, 2 }
63: [ . . 2 1 ] { 2, 2, 3 }
64: [ . . 2 2 ] { 2, 2, 3, 3 }
65: [ . . 2 3 ] { 2, 2, 3, 3, 3 }
66: [ . . 1 1 ] { 2, 3 }
67: [ . . 1 2 ] { 2, 3, 3 }
68: [ . . 1 3 ] { 2, 3, 3, 3 }
69: [ . . . 1 ] { 3 }
70: [ . . . 2 ] { 3, 3 }
71: [ . . . 3 ] { 3, 3, 3 }
\end{verbatim}
\caption{\label{fig:msubsetlex}
Subsets of the multiset $\{0^1, 1^2, 2^2, 3^3 \}$ in subset-lex order.
Dots are used to denote zeros in the (generalized) characteristic words.
}
\end{figure}
For the internal representation of the subsets of a multiset, we could use lists
of pairs $(e, m)$ where $e$ is the element and $m$ its multiplicity. This
choice would lead to loopless algorithms as will become clear in a moment. The
disadvantage of this representation, however, is that the generalizations we
will find would have a slightly more complicated form.
We will instead use generalized characteristic words,
where nonzero entries can be at most the multiplicity
of the element in question.
For a multiset $\{0^{m(0)},\, 1^{m(1)},\, \ldots,\, k^{m(k)}\}$
these are the mixed radix numbers with radix vector
$[m(0)+1,\, m(1)+1,\, \ldots,\, m(k)+1]$.
The subsets of the set $\{0^1,\, 1^2,\, 2^2,\, 3^3 \}$ in subset-lex order
are shown in figure \ref{fig:msubsetlex}.
The algorithms for computing the successor or predecessor
of a (nonempty) subset $S$ of a multiset in subset-lex are now given.
In the following let $z$ be the last element in $S$.
\begin{samepage}
\begin{alg}[Next-SL]\label{alg:next-sl}
Compute the successor of the subset $S$.
\end{alg}
\begin{enumerate}
\item If the multiplicity of $z$ is not maximal, increase it; return.
\item If $z$ is not maximal, append $z+1$ with multiplicity 1; return.
\item If $z$ is the only element, stop.
\item Remove $z$ and decrease the multiplicity of next to last nonzero element $y$,
then append $y+1$ with multiplicity 1.
\end{enumerate}
\end{samepage}
The description omits the case of the empty set,
the implementation takes care of this case by intially pointing
to the leftmost digit of the characteristic word, which is zero.
With the first call to the routine it is changed to 1.
\begin{alg}[Prev-SL]\label{alg:prev-sl}
Compute the predecessor of the subset $S$.
\end{alg}
\begin{enumerate}
\item If the set is empty, stop.
\item Decrease the multiplicity of $z$.
\item If the new multiplicity is nonzero, return.
\item Increase the multiplicity of $z-1$ and append the maximal element.
\end{enumerate}
Clearly both algorithms are loopless for the representation using a list of pairs.
With the representation as characteristic words algorithm \ref{alg:next-sl}
involves a loop in the last step, a scan for the next to last nonzero element.
This could be prevented by maintaining an additional list of positions where the
characteristic word is nonzero.
Algorithm \ref{alg:prev-sl} does stay loopless, the position of
the maximal element does not need to be sought.
Our implementations will always use a variable
holding the position of the last nonzero position
in the characteristic word.
This is called the \eqq{current track} in the implementations.
We will not make the equivalents of algorithm \ref{alg:next-sl} loopless, as
the maintenance of the additional list would render algorithm \ref{alg:prev-sl}
slower (at least as long as mixed calls to both shall be allowed),
but see section \ref{sect:loopless-next} for one such example.
A step in either direction has either one or three positions in the
characteristic word changed. The number of transitions with one change
are more frequent with higher multiplicities. The worst case occurs
if all multiplicities are one (corresponding to binary words), when
(about) half of the steps involve only one transition.
The C++ implementation uses the sentinel technique
to reduce the number of conditional branches.
{\codesize
\begin{listing}{1}
// FILE: src/comb/mixedradix-subset-lex.h
class mixedradix_subset_lex
{
ulong n_; // number of digits (n kinds of elements in multiset)
ulong tr_; // aux: current track
ulong *a_; // digits of mixed radix number
// (multiplicities in subset).
ulong *m1_; // nines (radix minus one) for each digit
// (multiplicity of kind k in superset).
mixedradix_subset_lex(ulong n, ulong mm, const ulong *m=0)
{
n_ = n;
a_ = new ulong[n_+2]; // two sentinels, one left, one right
a_[0] = 1; a_[n_+1] = 1;
++a_; // nota bene
m1_ = new ulong[n_+2];
m1_[0] = 0; m1_[n_+1] = 0; // sentinel with n==0
++m1_; // nota bene
mixedradix_init(n_, mm, m, m1_); // set up m1_[], omitted
first();
}
\end{listing}
}%
The omitted routine \texttt{mixedradix\_init()}
sets all elements of the array \texttt{m1[]} to \texttt{mm}
(fixed radix case),
unless the pointer \texttt{m} is nonzero, then the
elements behind \texttt{m} are copied into \texttt{m1[]}
(mixed radix case).
{\codesize
\begin{listing}{1}
void first()
{
for (ulong k=0; k<n_; ++k) a_[k] = 0;
tr_ = 0; // start by looking at leftmost (zero) digit
}
void last()
{
for (ulong k=0; k<n_; ++k) a_[k] = 0;
ulong n1 = ( n_ ? n_ - 1 : 0 );
a_[n1] = m1_[n1];
tr_ = n1;
}
\end{listing}
}%
Note the \texttt{while}-loop in the \texttt{next()} method.
{\codesize
\begin{listing}{1}
bool next()
{
ulong j = tr_;
if ( a_[j] < m1_[j] ) // easy case 1: increment
{
a_[j] += 1;
return true;
}
// here a_[j] == m1_[j]
if ( j+1 < n_ ) // easy case 2: append (move track to the right)
{
++j;
a_[j] = 1;
tr_ = j;
return true;
}
a_[j] = 0;
// find first nonzero digit to the left:
--j;
while ( a_[j] == 0 ) { --j; } // may read sentinel a_[-1]
if ( (long)j < 0 ) return false; // current is last
a_[j] -= 1; // decrement digit to the left
++j;
a_[j] = 1;
tr_ = j;
return true;
}
\end{listing}
}%
The method \texttt{prev()} is indeed loopless:
{\codesize
\begin{listing}{1}
bool prev()
{
ulong j = tr_;
if ( a_[j] > 1 ) // easy case 1: decrement
{
a_[j] -= 1;
return true;
}
else
{
if ( tr_ == 0 )
{
if ( a_[0] == 0 ) return false; // current is first
a_[0] = 0; // now word is first (all zero)
return true;
}
a_[j] = 0;
--j; // now looking at next track to the left
if ( a_[j] == m1_[j] ) // easy case 2: move track to left
{
tr_ = j; // move track one left
}
else
{
a_[j] += 1; // increment digit to the left
j = n_ - 1;
a_[j] = m1_[j]; // set rightmost digit = nine
tr_ = j; // move to rightmost track
}
return true;
}
}
\end{listing}
}%
The performance of the generator is quite satisfactory.
A system based on an AMD \texttt{Phenom(tm) II X4 945} processor
clocked at 3.0 GHz and the GCC C++ compiler version 4.9.0 \cite{GCC}
were used for measuring.
The updates using \texttt{next()} cost about
12 cycles for base 2 (the worst case) and
9 cycles for base 16.
Using \texttt{prev()} takes about
8.5 cycles with base 2 and
4.2 cycles with base 16,
the latter figure corresponding to a rate of 780 million subsets per second.
These and all following such figures are average values, obtained by measuring
the total time for the generation of all words of a certain size and dividing
by the number of words.
\section{Loopless computation of the successor for non-adjacent forms}\label{sect:loopless-next}
\begin{figure
{\jjbls
\begin{verbatim}
colex Gray code subset-lex iset
0: [ . . . . ] [ 4 . 2 . ] [ . . . . ] { }
1: [ 1 . . . ] [ 3 . 2 . ] [ 1 . . . ] { 0 }
2: [ 2 . . . ] [ 2 . 2 . ] [ 2 . . . ] { 0 }
3: [ 3 . . . ] [ 1 . 2 . ] [ 3 . . . ] { 0 }
4: [ 4 . . . ] [ . . 2 . ] [ 4 . . . ] { 0 }
5: [ . 1 . . ] [ . . 1 . ] [ 4 . 1 . ] { 0, 2 }
6: [ . 2 . . ] [ 1 . 1 . ] [ 4 . 2 . ] { 0, 2 }
7: [ . 3 . . ] [ 2 . 1 . ] [ 4 . . 1 ] { 0, 3 }
8: [ . . 1 . ] [ 3 . 1 . ] [ 3 . 1 . ] { 0, 2 }
9: [ 1 . 1 . ] [ 4 . 1 . ] [ 3 . 2 . ] { 0, 2 }
10: [ 2 . 1 . ] [ 4 . . . ] [ 3 . . 1 ] { 0, 3 }
11: [ 3 . 1 . ] [ 3 . . . ] [ 2 . 1 . ] { 0, 2 }
12: [ 4 . 1 . ] [ 2 . . . ] [ 2 . 2 . ] { 0, 2 }
13: [ . . 2 . ] [ 1 . . . ] [ 2 . . 1 ] { 0, 3 }
14: [ 1 . 2 . ] [ . . . . ] [ 1 . 1 . ] { 0, 2 }
15: [ 2 . 2 . ] [ . 1 . . ] [ 1 . 2 . ] { 0, 2 }
16: [ 3 . 2 . ] [ . 2 . . ] [ 1 . . 1 ] { 0, 3 }
17: [ 4 . 2 . ] [ . 3 . . ] [ . 1 . . ] { 1 }
18: [ . . . 1 ] [ . 3 . 1 ] [ . 2 . . ] { 1 }
19: [ 1 . . 1 ] [ . 2 . 1 ] [ . 3 . . ] { 1 }
20: [ 2 . . 1 ] [ . 1 . 1 ] [ . 3 . 1 ] { 1, 3 }
21: [ 3 . . 1 ] [ . . . 1 ] [ . 2 . 1 ] { 1, 3 }
22: [ 4 . . 1 ] [ 1 . . 1 ] [ . 1 . 1 ] { 1, 3 }
23: [ . 1 . 1 ] [ 2 . . 1 ] [ . . 1 . ] { 2 }
24: [ . 2 . 1 ] [ 3 . . 1 ] [ . . 2 . ] { 2 }
25: [ . 3 . 1 ] [ 4 . . 1 ] [ . . . 1 ] { 3 }
\end{verbatim}
\caption{\label{fig:naf}
Mixed radix numbers of length 4 in falling factorial base that are non-adjacent forms (NAF),
in co-lexicographic order, minimal-change order (Gray code),
and subset-lex order (with the sets of positions of nonzero digits).
}
\end{figure}
\begin{figure
{\jjbls
\begin{verbatim}
0: [ 4 . 2 . ] (0, 4) (1, 3) (2) [ 4 3 2 1 0 ]
1: [ 3 . 2 . ] (0, 3) (1, 4) (2) [ 3 4 2 0 1 ]
2: [ 2 . 2 . ] (0, 2) (1, 4) (3) [ 2 4 0 3 1 ]
3: [ 1 . 2 . ] (0, 1) (2, 4) (3) [ 1 0 4 3 2 ]
4: [ . . 2 . ] (0) (1) (2, 4) (3) [ 0 1 4 3 2 ]
5: [ . . 1 . ] (0) (1) (2, 3) (4) [ 0 1 3 2 4 ]
6: [ 1 . 1 . ] (0, 1) (2, 3) (4) [ 1 0 3 2 4 ]
7: [ 2 . 1 . ] (0, 2) (1, 3) (4) [ 2 3 0 1 4 ]
8: [ 3 . 1 . ] (0, 3) (1, 2) (4) [ 3 2 1 0 4 ]
9: [ 4 . 1 . ] (0, 4) (1, 2) (3) [ 4 2 1 3 0 ]
10: [ 4 . . . ] (0, 4) (1) (2) (3) [ 4 1 2 3 0 ]
11: [ 3 . . . ] (0, 3) (1) (2) (4) [ 3 1 2 0 4 ]
12: [ 2 . . . ] (0, 2) (1) (3) (4) [ 2 1 0 3 4 ]
13: [ 1 . . . ] (0, 1) (2) (3) (4) [ 1 0 2 3 4 ]
14: [ . . . . ] (0) (1) (2) (3) (4) [ 0 1 2 3 4 ]
15: [ . 1 . . ] (0) (1, 2) (3) (4) [ 0 2 1 3 4 ]
16: [ . 2 . . ] (0) (1, 3) (2) (4) [ 0 3 2 1 4 ]
17: [ . 3 . . ] (0) (1, 4) (2) (3) [ 0 4 2 3 1 ]
18: [ . 3 . 1 ] (0) (1, 4) (2, 3) [ 0 4 3 2 1 ]
19: [ . 2 . 1 ] (0) (1, 3) (2, 4) [ 0 3 4 1 2 ]
20: [ . 1 . 1 ] (0) (1, 2) (3, 4) [ 0 2 1 4 3 ]
21: [ . . . 1 ] (0) (1) (2) (3, 4) [ 0 1 2 4 3 ]
22: [ 1 . . 1 ] (0, 1) (2) (3, 4) [ 1 0 2 4 3 ]
23: [ 2 . . 1 ] (0, 2) (1) (3, 4) [ 2 1 0 4 3 ]
24: [ 3 . . 1 ] (0, 3) (1) (2, 4) [ 3 1 4 0 2 ]
25: [ 4 . . 1 ] (0, 4) (1) (2, 3) [ 4 1 3 2 0 ]
\end{verbatim}
\caption{\label{fig:naf-invol}
Gray code for the length-4 non-adjacent forms in falling factorial base (left),
together with the corresponding involutions in cycle form (middle) and
in array form (right).
}
\end{figure}
The method of using a list of nonzero positions in the words to obtain loopless
methods for both \texttt{next()} and \texttt{prev()} has been implemented for
words where no two adjacent digits are nonzero (non-adjacent forms, NAF).
In the following algorithm for the computation of the successor
let $a_i$ be the digits of the length-$n$ NAF where $0\leq{}i<n$.
\begin{alg}[Next-NAF-SL
Compute the successor of a non-adjacent form in subset-lex order.
\end{alg}
\begin{enumerate}
\item Let $j$ be the position of the last nonzero digit.
\item If $a_j$ is not maximal, increment it and return.
\item If $j+2 < n$, set $a_{j+2}=1$ (append new digit) and return.
\item Set $a_j=0$ (set last nonzero digit to zero).
\item If $j+1<n$, set $a_{j+1}=1$ (move digit right) and return.
\item If there is just one nonzero digit, stop.
\item Let $k$ be the position of the nearest nonzero digit left of $j$.
\item Set $a_k=a_k-1$ (decrement digit to the left).
\item If $a_k=0$, set $a_{k+1}=1$ (move digit right) and return.
\item Otherwise, $a_{k+2}=1$ (append digit two positions to the right).
\end{enumerate}
In the following implementation the array \texttt{iset[]}
holds the positions of the nonzero digits.
We only read the last or second last element,
the adjustments of the length (variable \texttt{ni})
have been left out in the description above.
The implementation handles all $n\geq{}0$ correctly,
for the all-zero word \texttt{iset[]} is of length one
and its only element points to the leftmost digit.
{\codesize
\begin{listing}{1}
// FILE: src/comb/mixedradix-naf-subset-lex.h
class mixedradix_naf_subset_lex
{
ulong *iset_; // Set of positions of nonzero digits
ulong *a_; // digits
ulong *m1_; // nines (radix minus one) for each digit
ulong ni_; // number of elements in iset[]
ulong n_; // number of digits
void first()
{
for (ulong k=0; k<n_; ++k) a_[k] = 0;
// iset[] initially with one element zero:
iset_[0] = 0;
ni_ = 1;
if ( n_==0 ) // make things work for n == 0
{
m1_[0] = 0;
a_[0] = 0;
}
}
\end{listing}
}%
The computation of the successor is
{\codesize
\begin{listing}{1}
bool next()
{
ulong j = iset_[ni_-1];
const ulong aj = a_[j] + 1;
if ( aj <= m1_[j] ) // can increment last digit
{
a_[j] = aj;
return true;
}
if ( j + 2 < n_ ) // can append new digit
{
iset_[ni_] = j + 2;
a_[j+2] = 1; // assume all m1[] are nonzero
++ni_;
return true;
}
a_[j] = 0; // set last nonzero digit to zero
if ( j + 1 < n_ ) // can move last digit to the right
{
a_[j+1] = 1;
iset_[ni_-1] = j + 1;
return true;
}
if ( ni_ == 1 ) return false; // current is last
// Now we look to the left:
const ulong k = iset_[ni_-2]; // nearest nonzero digit to the left
const ulong ak = a_[k] - 1; // decrement digit to the left
a_[k] = ak;
if ( ak == 0 ) // move digit one to the right
{
a_[k+1] = 1;
iset_[ni_-2] = k + 1;
--ni_;
return true;
}
else // append digit two positions to the right
{
a_[k+2] = 1;
iset_[ni_-1] = k + 2;
return true;
}
}
\end{listing}
}%
The update via \texttt{next()} takes about 7 cycles.
The updates for co-lexicographic order and the Gray code
respectively take about 15 and 21 cycles.
We note that the falling factorial NAFs of length $n$ are one-to-one
with involutions (self-inverse permutations) of $n+1$ elements.
Process the NAF from left to right; for $a_i=0$ let the
next unused element of the permutation be a fixed point
(and mark it as used); for $a_i\neq{}0$ put the next unused
element in a cycle with the $a_i$th unused element (and mark both as used).
A (simplistic) implementation of this method
is given in the program
\fileref{demo/comb/perm-involution-naf-demo.cc}.
\begin{figure
{\jjbls
\begin{verbatim}
colex Gray code subset-lex
0: ........ .1..1..1 1....... = { 0 }
1: .......1 .1..1... 1.1..... = { 0, 2 }
2: ......1. .1..1.1. 1.1.1... = { 0, 2, 4 }
3: .....1.. .1....1. 1.1.1.1. = { 0, 2, 4, 6 }
4: .....1.1 .1...... 1.1.1..1 = { 0, 2, 4, 7 }
5: ....1... .1.....1 1.1..1.. = { 0, 2, 5 }
6: ....1..1 .1...1.1 1.1..1.1 = { 0, 2, 5, 7 }
7: ....1.1. .1...1.. 1.1...1. = { 0, 2, 6 }
8: ...1.... .1.1.1.. 1.1....1 = { 0, 2, 7 }
9: ...1...1 .1.1.1.1 1..1.... = { 0, 3 }
10: ...1..1. .1.1...1 1..1.1.. = { 0, 3, 5 }
11: ...1.1.. .1.1.... 1..1.1.1 = { 0, 3, 5, 7 }
12: ...1.1.1 .1.1..1. 1..1..1. = { 0, 3, 6 }
13: ..1..... ...1..1. 1..1...1 = { 0, 3, 7 }
14: ..1....1 ...1.... 1...1... = { 0, 4 }
15: ..1...1. ...1...1 1...1.1. = { 0, 4, 6 }
16: ..1..1.. ...1.1.1 1...1..1 = { 0, 4, 7 }
17: ..1..1.1 ...1.1.. 1....1.. = { 0, 5 }
18: ..1.1... .....1.. 1....1.1 = { 0, 5, 7 }
19: ..1.1..1 .....1.1 1.....1. = { 0, 6 }
20: ..1.1.1. .......1 1......1 = { 0, 7 }
21: .1...... ........ .1...... = { 1 }
22: .1.....1 ......1. .1.1.... = { 1, 3 }
23: .1....1. ....1.1. .1.1.1.. = { 1, 3, 5 }
24: .1...1.. ....1... .1.1.1.1 = { 1, 3, 5, 7 }
25: .1...1.1 ....1..1 .1.1..1. = { 1, 3, 6 }
26: .1..1... ..1.1..1 .1.1...1 = { 1, 3, 7 }
27: .1..1..1 ..1.1... .1..1... = { 1, 4 }
28: .1..1.1. ..1.1.1. .1..1.1. = { 1, 4, 6 }
29: .1.1.... ..1...1. .1..1..1 = { 1, 4, 7 }
30: .1.1...1 ..1..... .1...1.. = { 1, 5 }
31: .1.1..1. ..1....1 .1...1.1 = { 1, 5, 7 }
32: .1.1.1.. ..1..1.1 .1....1. = { 1, 6 }
33: .1.1.1.1 ..1..1.. .1.....1 = { 1, 7 }
34: 1....... 1.1..1.. ..1..... = { 2 }
35: 1......1 1.1..1.1 ..1.1... = { 2, 4 }
36: 1.....1. 1.1....1 ..1.1.1. = { 2, 4, 6 }
37: 1....1.. 1.1..... ..1.1..1 = { 2, 4, 7 }
38: 1....1.1 1.1...1. ..1..1.. = { 2, 5 }
39: 1...1... 1.1.1.1. ..1..1.1 = { 2, 5, 7 }
40: 1...1..1 1.1.1... ..1...1. = { 2, 6 }
41: 1...1.1. 1.1.1..1 ..1....1 = { 2, 7 }
42: 1..1.... 1...1..1 ...1.... = { 3 }
43: 1..1...1 1...1... ...1.1.. = { 3, 5 }
44: 1..1..1. 1...1.1. ...1.1.1 = { 3, 5, 7 }
45: 1..1.1.. 1.....1. ...1..1. = { 3, 6 }
46: 1..1.1.1 1....... ...1...1 = { 3, 7 }
47: 1.1..... 1......1 ....1... = { 4 }
48: 1.1....1 1....1.1 ....1.1. = { 4, 6 }
49: 1.1...1. 1....1.. ....1..1 = { 4, 7 }
50: 1.1..1.. 1..1.1.. .....1.. = { 5 }
51: 1.1..1.1 1..1.1.1 .....1.1 = { 5, 7 }
52: 1.1.1... 1..1...1 ......1. = { 6 }
53: 1.1.1..1 1..1.... .......1 = { 7 }
54: 1.1.1.1. 1..1..1.
\end{verbatim}
\caption{\label{fig:bitfibrep}
Binary non-adjacent forms of length 8
in co-lexicographic order, minimal-change order (Gray code),
and subset-lex order (with the corresponding sets without
consecutive elements).
}
\end{figure}
The binary NAFs of length 8 are shown in figure \ref{fig:bitfibrep}.
Algorithms for the generation of these NAFs as binary words
for both lexicographic order and the Gray code are given
in \cite[p.75-77]{fxtbook}.
Here we give the implementations for computing successor
and predecessor for the nonempty NAFs in subset-lex order.
The routine for the successor needs to
start with the word that has a single set bit at
the highest position of the desired word length.
{\codesize
\begin{listing}{1}
// FILE: src/bits/fibrep-subset-lexrev.h
ulong next_subset_lexrev_fib(ulong x)
{
ulong x0 = x & -x; // lowest bit
ulong xs = x0 >> 2;
if ( xs != 0 ) // easy case: set bit right of lowest bit
{
x |= xs;
return x;
}
else // lowest bit at index 0 or 1
{
if ( x0 == 2 ) // at index 1
{
x -= 1;
return x;
}
x ^= x0; // clear lowest bit
x0 = x & -x; // new lowest bit ...
x0 >>= 1; x -= x0; // ... is moved one to the right
return x;
}
}
\end{listing}
}%
The all-zero word is returned as successor
of the word whose value is 1.
The routine for the predecessor can
be started with the all-zero word.
{\codesize
\begin{listing}{1}
ulong prev_subset_lexrev_fib(ulong x)
{
ulong x0 = x & -x; // lowest bit
if ( x & (x0<<2) ) // easy case: next higher bit is set
{
x ^= x0; // clear lowest bit
return x;
}
else
{
x += x0; // move lowest bit to the left and
x |= (x0!=1); // set rightmost bit unless blocked by next bit
return x;
}
}
\end{listing}
}%
\section{Compositions}
\begin{figure
{\jjbls
\begin{verbatim}
0: ...... [ 1 1 1 1 1 1 1 ]
1: .....1 [ 1 1 1 1 1 2 ]
2: ....1. [ 1 1 1 1 2 1 ]
3: ....11 [ 1 1 1 1 3 ]
4: ...1.. [ 1 1 1 2 1 1 ]
5: ...1.1 [ 1 1 1 2 2 ]
6: ...11. [ 1 1 1 3 1 ]
7: ...111 [ 1 1 1 4 ]
8: ..1... [ 1 1 2 1 1 1 ]
9: ..1..1 [ 1 1 2 1 2 ]
10: ..1.1. [ 1 1 2 2 1 ]
11: ..1.11 [ 1 1 2 3 ]
12: ..11.. [ 1 1 3 1 1 ]
13: ..11.1 [ 1 1 3 2 ]
14: ..111. [ 1 1 4 1 ]
15: ..1111 [ 1 1 5 ]
16: .1.... [ 1 2 1 1 1 1 ]
17: .1...1 [ 1 2 1 1 2 ]
18: .1..1. [ 1 2 1 2 1 ]
19: .1..11 [ 1 2 1 3 ]
20: .1.1.. [ 1 2 2 1 1 ]
21: .1.1.1 [ 1 2 2 2 ]
22: .1.11. [ 1 2 3 1 ]
23: .1.111 [ 1 2 4 ]
24: .11... [ 1 3 1 1 1 ]
25: .11..1 [ 1 3 1 2 ]
26: .11.1. [ 1 3 2 1 ]
27: .11.11 [ 1 3 3 ]
28: .111.. [ 1 4 1 1 ]
29: .111.1 [ 1 4 2 ]
30: .1111. [ 1 5 1 ]
31: .11111 [ 1 6 ]
32: 1..... [ 2 1 1 1 1 1 ]
33: 1....1 [ 2 1 1 1 2 ]
34: 1...1. [ 2 1 1 2 1 ]
35: 1...11 [ 2 1 1 3 ]
36: 1..1.. [ 2 1 2 1 1 ]
37: 1..1.1 [ 2 1 2 2 ]
38: 1..11. [ 2 1 3 1 ]
39: 1..111 [ 2 1 4 ]
40: 1.1... [ 2 2 1 1 1 ]
41: 1.1..1 [ 2 2 1 2 ]
42: 1.1.1. [ 2 2 2 1 ]
43: 1.1.11 [ 2 2 3 ]
44: 1.11.. [ 2 3 1 1 ]
45: 1.11.1 [ 2 3 2 ]
46: 1.111. [ 2 4 1 ]
47: 1.1111 [ 2 5 ]
48: 11.... [ 3 1 1 1 1 ]
49: 11...1 [ 3 1 1 2 ]
50: 11..1. [ 3 1 2 1 ]
51: 11..11 [ 3 1 3 ]
52: 11.1.. [ 3 2 1 1 ]
53: 11.1.1 [ 3 2 2 ]
54: 11.11. [ 3 3 1 ]
55: 11.111 [ 3 4 ]
56: 111... [ 4 1 1 1 ]
57: 111..1 [ 4 1 2 ]
58: 111.1. [ 4 2 1 ]
59: 111.11 [ 4 3 ]
60: 1111.. [ 5 1 1 ]
61: 1111.1 [ 5 2 ]
62: 11111. [ 6 1 ]
63: 111111 [ 7 ]
\end{verbatim}
}%
\caption{\label{fig:comp-nz}
The compositions of 7 together with their run-length
encodings as binary words (where dots denote zeros),
lexicographic order.
A succession of $k$ ones, followed by a zero, in the run-length encoding
stands for a part $k+1$, one trailing zero is implied.
}
\end{figure}
One of the most simple algorithms in combinatorial generation
may be the computation of the successor of a composition
represented as a list of parts for the lexicographic ordering.
In the following let $z$ be the last element of the composition.
\begin{samepage}
\begin{alg}[Next-Comp
Compute the successor of a composition in lexicographic order.
\end{alg}
\begin{enumerate}
\item If there is just one part, stop (this is the last composition).
\item Add $1$ to the second last part (and remove $z$) and
append $z-1$ ones at the end.
\end{enumerate}
\end{samepage}
\begin{samepage}
\begin{alg}[Prev-Comp
Compute the predecessor of a composition in lexicographic order.
\end{alg}
\begin{enumerate}
\item If the number of parts is maximal (composition into all ones), stop.
\item If $z>1$, replace $z$ by $z-1,\,1$ (move one unit right); return
\item Otherwise, replace the tail $y,\, 1^q$ ($y$ followed by $q$ ones) by $y-1, q+1$.
\end{enumerate}
\end{samepage}
Figure \ref{fig:comp-nz} shows the compositions of 7
in lexicographic order. The corresponding run-length encodings
appear in lexicographic order as well.
The algorithm for the successor can be made loopless
when care is taken that only ones are left beyond
the end of the current composition.
{\codesize
\begin{listing}{1}
// FILE: src/comb/composition-nz.h
class composition_nz
// Compositions of n into positive parts, lexicographic order.
{
ulong *a_; // composition: a[1] + a[2] + ... + a[m] = n
ulong n_; // composition of n
ulong m_; // current composition is into m parts
ulong next()
// Return number of parts of generated composition.
// Return zero if the current is the last composition.
{
if ( m_ <= 1 ) return 0; // current is last
// [*, Y, Z] --> [*, Y+1, 1, 1, 1, ..., 1] (Z-1 trailing ones)
a_[m_-1] += 1;
const ulong z = a_[m_];
a_[m_] = 1;
// all parts a[m+1], a[m+2], ..., a[n] are already ==1
m_ += z - 2;
return m_;
}
\end{listing}
}%
\begin{figure
{\jjbls
\begin{verbatim}
0: 111111 [ 7 ] ...... [ 7 ]
1: 11111. [ 6 1 ] 1..... [ 1 6 ]
2: 1111.. [ 5 1 1 ] 11.... [ 1 1 5 ]
3: 1111.1 [ 5 2 ] 111... [ 1 1 1 4 ]
4: 111..1 [ 4 1 2 ] 1111.. [ 1 1 1 1 3 ]
5: 111... [ 4 1 1 1 ] 11111. [ 1 1 1 1 1 2 ]
6: 111.1. [ 4 2 1 ] 111111 [ 1 1 1 1 1 1 1 ]
7: 111.11 [ 4 3 ] 1111.1 [ 1 1 1 1 2 1 ]
8: 11..11 [ 3 1 3 ] 111.1. [ 1 1 1 2 2 ]
9: 11..1. [ 3 1 2 1 ] 111.11 [ 1 1 1 2 1 1 ]
10: 11.... [ 3 1 1 1 1 ] 111..1 [ 1 1 1 3 1 ]
11: 11...1 [ 3 1 1 2 ] 11.1.. [ 1 1 2 3 ]
12: 11.1.1 [ 3 2 2 ] 11.11. [ 1 1 2 1 2 ]
13: 11.1.. [ 3 2 1 1 ] 11.111 [ 1 1 2 1 1 1 ]
14: 11.11. [ 3 3 1 ] 11.1.1 [ 1 1 2 2 1 ]
15: 11.111 [ 3 4 ] 11..1. [ 1 1 3 2 ]
16: 1..111 [ 2 1 4 ] 11..11 [ 1 1 3 1 1 ]
17: 1..11. [ 2 1 3 1 ] 11...1 [ 1 1 4 1 ]
18: 1..1.. [ 2 1 2 1 1 ] 1.1... [ 1 2 4 ]
19: 1..1.1 [ 2 1 2 2 ] 1.11.. [ 1 2 1 3 ]
20: 1....1 [ 2 1 1 1 2 ] 1.111. [ 1 2 1 1 2 ]
21: 1..... [ 2 1 1 1 1 1 ] 1.1111 [ 1 2 1 1 1 1 ]
22: 1...1. [ 2 1 1 2 1 ] 1.11.1 [ 1 2 1 2 1 ]
23: 1...11 [ 2 1 1 3 ] 1.1.1. [ 1 2 2 2 ]
24: 1.1.11 [ 2 2 3 ] 1.1.11 [ 1 2 2 1 1 ]
25: 1.1.1. [ 2 2 2 1 ] 1.1..1 [ 1 2 3 1 ]
26: 1.1... [ 2 2 1 1 1 ] 1..1.. [ 1 3 3 ]
27: 1.1..1 [ 2 2 1 2 ] 1..11. [ 1 3 1 2 ]
28: 1.11.1 [ 2 3 2 ] 1..111 [ 1 3 1 1 1 ]
29: 1.11.. [ 2 3 1 1 ] 1..1.1 [ 1 3 2 1 ]
30: 1.111. [ 2 4 1 ] 1...1. [ 1 4 2 ]
31: 1.1111 [ 2 5 ] 1...11 [ 1 4 1 1 ]
32: ..1111 [ 1 1 5 ] 1....1 [ 1 5 1 ]
33: ..111. [ 1 1 4 1 ] .1.... [ 2 5 ]
34: ..11.. [ 1 1 3 1 1 ] .11... [ 2 1 4 ]
35: ..11.1 [ 1 1 3 2 ] .111.. [ 2 1 1 3 ]
36: ..1..1 [ 1 1 2 1 2 ] .1111. [ 2 1 1 1 2 ]
37: ..1... [ 1 1 2 1 1 1 ] .11111 [ 2 1 1 1 1 1 ]
38: ..1.1. [ 1 1 2 2 1 ] .111.1 [ 2 1 1 2 1 ]
39: ..1.11 [ 1 1 2 3 ] .11.1. [ 2 1 2 2 ]
40: ....11 [ 1 1 1 1 3 ] .11.11 [ 2 1 2 1 1 ]
41: ....1. [ 1 1 1 1 2 1 ] .11..1 [ 2 1 3 1 ]
42: ...... [ 1 1 1 1 1 1 1 ] .1.1.. [ 2 2 3 ]
43: .....1 [ 1 1 1 1 1 2 ] .1.11. [ 2 2 1 2 ]
44: ...1.1 [ 1 1 1 2 2 ] .1.111 [ 2 2 1 1 1 ]
45: ...1.. [ 1 1 1 2 1 1 ] .1.1.1 [ 2 2 2 1 ]
46: ...11. [ 1 1 1 3 1 ] .1..1. [ 2 3 2 ]
47: ...111 [ 1 1 1 4 ] .1..11 [ 2 3 1 1 ]
48: .1.111 [ 1 2 4 ] .1...1 [ 2 4 1 ]
49: .1.11. [ 1 2 3 1 ] ..1... [ 3 4 ]
50: .1.1.. [ 1 2 2 1 1 ] ..11.. [ 3 1 3 ]
51: .1.1.1 [ 1 2 2 2 ] ..111. [ 3 1 1 2 ]
52: .1...1 [ 1 2 1 1 2 ] ..1111 [ 3 1 1 1 1 ]
53: .1.... [ 1 2 1 1 1 1 ] ..11.1 [ 3 1 2 1 ]
54: .1..1. [ 1 2 1 2 1 ] ..1.1. [ 3 2 2 ]
55: .1..11 [ 1 2 1 3 ] ..1.11 [ 3 2 1 1 ]
56: .11.11 [ 1 3 3 ] ..1..1 [ 3 3 1 ]
57: .11.1. [ 1 3 2 1 ] ...1.. [ 4 3 ]
58: .11... [ 1 3 1 1 1 ] ...11. [ 4 1 2 ]
59: .11..1 [ 1 3 1 2 ] ...111 [ 4 1 1 1 ]
60: .111.1 [ 1 4 2 ] ...1.1 [ 4 2 1 ]
61: .111.. [ 1 4 1 1 ] ....1. [ 5 2 ]
62: .1111. [ 1 5 1 ] ....11 [ 5 1 1 ]
63: .11111 [ 1 6 ] .....1 [ 6 1 ]
\end{verbatim}
}%
\caption{\label{fig:comp-nz-rl-and-sl}
The compositions of 7 together with their run-length
encodings as binary words (where dots denote zeros),
in an order corresponding to the complemented binary Gray code
(left) and in subset-lex order (right).
A succession of $k$ ones, followed by a zero, in the run-length encoding
stands for a part $k+1$, one trailing zero is implied (left).
The roles of ones and zeros are reversed for the subset-lex order (right).
}
\end{figure}
Figure \ref{fig:comp-nz-rl-and-sl} shows two orderings for the compositions.
The ordering corresponding to the (complemented) binary Gray code is shown in
the left columns. We will call this order \jjterm{RL-order}, as the
compositions correspond to the run-lengths of the binary words in lexicographic
order.
For comparison with the new algorithm we give the (loopless) algorithms.
In the following let $m$ be the number of parts in the composition
and $x,y,z$ the last three parts.
\begin{samepage}
\begin{alg}[Next-Comp-RL
Compute the successor of a composition in RL-order.
\end{alg}
\begin{enumerate}
\item If $m$ is odd:
if $z\geq{}2$, replace $z$ by $z-1, 1$ and return;
otherwise ($z=1$) replace $y,1$ by $y+1$ and return.
\item If $m$ is even:
if $y\geq{}2$, replace $y,z$ by $y-1, 1, z$ and return;
otherwise ($y=1$) replace $x,1,z$ by $x+1,z$ and return.
\end{enumerate}
\end{samepage}
The next algorithm is obtained from the previous simply
by swapping \eqq{even} and \eqq{odd} in the description.
\begin{samepage}
\begin{alg}[Prev-Comp-RL
Compute the predecessor of a composition in RL-order.
\end{alg}
\begin{enumerate}
\item If $m$ is even:
if $z\geq{}2$, replace $z$ by $z-1, 1$ and return;
otherwise ($z=1$) replace $y,1$ by $y+1$ and return.
\item If $m$ is odd:
if $y\geq{}2$, replace $y,z$ by $y-1, 1, z$ and return;
otherwise ($y=1$) replace $x,1,z$ by $x+1,z$ and return.
\end{enumerate}
\end{samepage}
With each transition, at most three parts (at the end of
the composition) are changed.
The number of parts changes by $1$
with each step\footnote{See
\fileref{src/comb/composition-nz-rl.h} for an implementation.}.
An alternative algorithm for this ordering is given
in \cite[ex.12, sect.7.2.1.1, p.308]{DEK4}, see
also \cite{misra}.
Now we give the algorithms for subset-lex order
(the first is essentially given in \cite{mansour-ll-comp}).
\begin{alg}[Next-Comp-SL
Compute the successor of a composition in subset-lex order.
\end{alg}
\begin{enumerate}
\item If $z=1$ and there are at most two parts, stop.
\item If $z\geq{}2$, replace $z$ by $1, z-1$ (move all but one unit to the right); return.
\item Otherwise ($z=1$) add $1$ to the third last part and
remove $z$ (move one unit two places left: $x,y,1$ is replaced by $x+1, y$).
\end{enumerate}
For the next algorithm let $y,\,z$ be the two last elements.
\begin{samepage}
\begin{alg}[Prev-Comp-SL
Compute the predecessor of a composition in subset-lex order.
\end{alg}
\begin{enumerate}
\item If there is just one part, stop.
\item If $y=1$, replace $y,\,z$ by $z+1$ (add $z$ to the left); return.
\item Otherwise ($y\geq{}2$) replace $y,\,z$ by $y-1,\,z,\,1$ (move one unit two places right).
\end{enumerate}
\end{samepage}
At most two parts are changed by either method and
these span at most those three parts at the end of the composition.
Again, the number of parts changes by $1$ with each step.
Note that also the initializations are loopless
for the preceding two orderings.
We give the crucial parts of the implementation.
{\codesize
\begin{listing}{1}
// FILE: src/comb/composition-nz-subset-lex.h
class composition_nz_subset_lex
{
ulong *a_; // composition: a[1] + a[2] + ... + a[m] = n
ulong n_; // composition of n
ulong m_; // current composition is into m parts
void first()
{
a_[0] = 0;
a_[1] = n_;
m_ = ( n_ ? 1 : 0 );
}
void last()
{
if ( n_ >= 2 )
{
a_[1] = n_ - 1;
a_[2] = 1;
for (ulong j=2; j<=n_; ++j) a_[j] = 1;
m_ = 2;
}
else
{
a_[1] = n_;
m_ = n_;
}
}
\end{listing}
}%
For the methods \texttt{next()} and \texttt{prev()}
we give one of the two implementations found in the file.
Both methods return the number of parts in the generated
composition and return zero if there are no more
compositions.
{\codesize
\begin{listing}{1}
ulong next()
{
const ulong z = a_[m_];
if ( z<=1 ) // move one unit two places left
{ // [*, X, Y, 1] --> [*, X+1, Y]
if ( m_ <= 2 ) return 0; // current is last
m_ -= 1;
a_[m_-1] += 1;
return m_;
}
else // move all but one unit right
{ // [*, Y, Z] --> [*, Y, 1, Z-1]
a_[m_] = 1;
m_ += 1;
a_[m_] = z - 1;
return m_;
}
}
\end{listing}
}%
{\codesize
\begin{listing}{1}
ulong prev()
{
if ( m_ <= 1 ) return 0; // current is first
const ulong y = a_[m_-1];
if ( y==1 ) // add Z to left place
{ // [*, 1, Z] --> [*, Z+1]
const ulong z = a_[m_];
a_[m_-1] = z + 1;
a_[m_] = 1;
m_ -= 1;
return m_;
}
else // move one unit two places right
{ // [*, Y, Z] --> [*, Y-1, Z, 1]
a_[m_-1] = y - 1;
m_ += 1;
// a[m] == 1 already
return m_;
}
}
\end{listing}
}%
The method \texttt{next()} takes about
5 cycles for lexicographic order,
9 cycles for RL-order,
and 6 cycles for subset-lex order.
The method \texttt{prev()} takes about
10.5 cycles for lexicographic order
and is identical in performance
to \texttt{next()} for the other orders.
Note that the last parts in the successive compositions in
lexicographic order give the (one-based)
ruler function (sequence \jjseqref{A001511} in \cite{oeis}), see
\fileref{src/comb/ruler-func1.h} for the trivial implementation, compare to
\fileref{src/comb/ruler-func.h} (giving sequence \jjseqref{A007814} in \cite{oeis})
which uses the techniques from \cite{ehrlichll} and \cite{bitnergray}
as described in \cite[Algorithm L, p.290]{DEK4}.
An loopless implementation for the generation of the compositions into odd parts in
subset-lex order is given in \fileref{src/comb/composition-nz-odd-subset-lex.h}.
\subsubsection*{Ranking and unranking}
An unranking algorithm for the subset-lex order
is obtained by observing (see figure \ref{fig:comp-nz-rl-and-sl})
that the composition into one part has rank 0
and otherwise the first part and the remaining parts
are easily determined.
\begin{alg}[Unrank-Comp-SL]\label{alg:unrank-comp-sl}
Recursive routine $U( r,\, n,\, C[\,],\, m )$ for the computation
of the composition $C[\,]$ of $n$ with rank $r$ in subset-lex order.
The auxiliary variable $m\geq{}0$ is the index
of the part about to be written in $C[\,]$.
The initial call is $F(r,\, n,\, C[\,],\, 0)$.
\end{alg}
\begin{enumerate}
\item If $r=0$, set $C[m]=n$ and return $m+1$ (the number of parts).
\item Set $f=0$ (the tentative first part).
\item\label{alg:loop}
Set $f=f+1$ and $t=2^{n-f-1}$.
\item If $r<t$ (can use part $f$),
set $C[m]=f$ and return $F( r,\, n-f,\, C[\,],\, m+1 )$.
\item Set $r=r-t$ and go to step \ref{alg:loop}.
\end{enumerate}
We give an iterative implementation.
{\codesize
\begin{listing}{1}
// FILE: src/comb/composition-nz-rank.cc
ulong composition_nz_subset_lex_unrank(ulong r, ulong *x, ulong n)
{
if ( n==0 ) return 0;
ulong m = 0;
while ( true )
{
if ( r==0 ) // composition into one part
{
x[m++] = n;
return m;
}
r -= 1;
ulong t = 1UL << (n-1);
for (ulong f=1; f<n; ++f) // find first part f >= 1
{
t >>= 1; // == 2**(n-f-1)
// == number of compositions of n with first part f
if ( r < t ) // first part is f
{
x[m++] = f;
n -= f;
break;
}
r -= t;
}
}
}
\end{listing}
}%
The following algorithm for computing the rank
is modeled as inverse of the method above.
\begin{alg}[Rank-Comp-SL]\label{alg:rank-comp-sl}
Computation of the rank $r$ of a composition $C[\,]$ of $n$ in subset-lex order.
\end{alg}
\begin{enumerate}
\item Set $r=0$ and $e=0$ (position of part under consideration).
\item\label{alg:loop2}
Set $f=C[e]$ (part under consideration).
\item If $f=n$, return $r$.
\item Set $r=r+1$, $t=2^{n-1}$, and $n=n-f$.
\item While $f>1$, set $t=t/2$, $r=r+t$, and $f=f-1$.
\item Set $e=e+1$ and go to step \ref{alg:loop2}
\end{enumerate}
An implementation is
{\codesize
\begin{listing}{1}
ulong composition_nz_subset_lex_rank(const ulong *x, ulong m, ulong n)
{
ulong r = 0; // rank
ulong e = 0; // position of first part
while ( e < m )
{
ulong f = x[e];
if ( f==n ) return r;
r += 1;
ulong t = 1UL << (n-1);
n -= f;
while ( f > 1 )
{
t >>= 1;
r += t;
f -= 1;
}
e += 1;
}
return r; // return r==0 for the empty composition
}
\end{listing}
}%
Conversion functions between binary words in lexicographic order and subset-lex
order are shown in \cite[pp.71-72]{fxtbook}.
{\codesize
\begin{listing}{1}
// FILE: src/bits/bitlex.h
ulong negidx2lexrev(ulong k)
{
ulong z = 0;
ulong h = highest_one(k);
while ( k )
{
while ( 0==(h&k) ) h >>= 1;
z ^= h;
++k;
k &= h - 1;
}
return z;
}
ulong lexrev2negidx(ulong x)
{
if ( 0==x ) return 0;
ulong h = x & -x; // lowest one
ulong r = (h-1);
while ( x^=h )
{
r += (h-1);
h = x & -x; // next higher one
}
r += h; // highest bit
return r;
}
\end{listing}
}%
Based on these, alternative ranking and unranking functions
can be given\footnote{See
\fileref{src/comb/composition-nz-rank.cc}
where the corresponding routines for all three orders shown here are implemented.}.
Ranking and unranking methods for the compositions
into odd parts can be obtained by replacing
$2^{n-f-1}$ (number of compositions of $n$ with first part $f$)
in the algorithms \ref{alg:unrank-comp-sl} and \ref{alg:rank-comp-sl}
by the Fibonacci numbers $F_{n-f-1}$
(number of compositions of $n$ into odd parts with first part $f$),
where $n\geq{}1$, $f<n$, and $f$ odd
(see sequence \jjseqref{A242086} in \cite{oeis}).
\section{Partitions as weakly increasing lists of parts}
We now give algorithms for the computation
of the successor for partitions represented
as weakly increasing lists of parts.
The generation of all partitions in this representation
is the subject of \cite{kelleher-thesis}.
\subsection{All partitions}
\begin{figure
{\jjbls
\begin{verbatim}
1: [ 1 1 1 1 1 1 1 1 1 1 1 ] [ 11 ]
2: [ 1 1 1 1 1 1 1 1 1 2 ] [ 1 10 ]
3: [ 1 1 1 1 1 1 1 1 3 ] [ 1 1 9 ]
4: [ 1 1 1 1 1 1 1 2 2 ] [ 1 1 1 8 ]
5: [ 1 1 1 1 1 1 1 4 ] [ 1 1 1 1 7 ]
6: [ 1 1 1 1 1 1 2 3 ] [ 1 1 1 1 1 6 ]
7: [ 1 1 1 1 1 1 5 ] [ 1 1 1 1 1 1 5 ]
8: [ 1 1 1 1 1 2 2 2 ] [ 1 1 1 1 1 1 1 4 ]
9: [ 1 1 1 1 1 2 4 ] [ 1 1 1 1 1 1 1 1 3 ]
10: [ 1 1 1 1 1 3 3 ] [ 1 1 1 1 1 1 1 1 1 2 ]
11: [ 1 1 1 1 1 6 ] [ 1 1 1 1 1 1 1 1 1 1 1 ]
12: [ 1 1 1 1 2 2 3 ] [ 1 1 1 1 1 1 1 2 2 ]
13: [ 1 1 1 1 2 5 ] [ 1 1 1 1 1 1 2 3 ]
14: [ 1 1 1 1 3 4 ] [ 1 1 1 1 1 2 4 ]
15: [ 1 1 1 1 7 ] [ 1 1 1 1 1 2 2 2 ]
16: [ 1 1 1 2 2 2 2 ] [ 1 1 1 1 1 3 3 ]
17: [ 1 1 1 2 2 4 ] [ 1 1 1 1 2 5 ]
18: [ 1 1 1 2 3 3 ] [ 1 1 1 1 2 2 3 ]
19: [ 1 1 1 2 6 ] [ 1 1 1 1 3 4 ]
20: [ 1 1 1 3 5 ] [ 1 1 1 2 6 ]
21: [ 1 1 1 4 4 ] [ 1 1 1 2 2 4 ]
22: [ 1 1 1 8 ] [ 1 1 1 2 2 2 2 ]
23: [ 1 1 2 2 2 3 ] [ 1 1 1 2 3 3 ]
24: [ 1 1 2 2 5 ] [ 1 1 1 3 5 ]
25: [ 1 1 2 3 4 ] [ 1 1 1 4 4 ]
26: [ 1 1 2 7 ] [ 1 1 2 7 ]
27: [ 1 1 3 3 3 ] [ 1 1 2 2 5 ]
28: [ 1 1 3 6 ] [ 1 1 2 2 2 3 ]
29: [ 1 1 4 5 ] [ 1 1 2 3 4 ]
30: [ 1 1 9 ] [ 1 1 3 6 ]
31: [ 1 2 2 2 2 2 ] [ 1 1 3 3 3 ]
32: [ 1 2 2 2 4 ] [ 1 1 4 5 ]
33: [ 1 2 2 3 3 ] [ 1 2 8 ]
34: [ 1 2 2 6 ] [ 1 2 2 6 ]
35: [ 1 2 3 5 ] [ 1 2 2 2 4 ]
36: [ 1 2 4 4 ] [ 1 2 2 2 2 2 ]
37: [ 1 2 8 ] [ 1 2 2 3 3 ]
38: [ 1 3 3 4 ] [ 1 2 3 5 ]
39: [ 1 3 7 ] [ 1 2 4 4 ]
40: [ 1 4 6 ] [ 1 3 7 ]
41: [ 1 5 5 ] [ 1 3 3 4 ]
42: [ 1 10 ] [ 1 4 6 ]
43: [ 2 2 2 2 3 ] [ 1 5 5 ]
44: [ 2 2 2 5 ] [ 2 9 ]
45: [ 2 2 3 4 ] [ 2 2 7 ]
46: [ 2 2 7 ] [ 2 2 2 5 ]
47: [ 2 3 3 3 ] [ 2 2 2 2 3 ]
48: [ 2 3 6 ] [ 2 2 3 4 ]
49: [ 2 4 5 ] [ 2 3 6 ]
50: [ 2 9 ] [ 2 3 3 3 ]
51: [ 3 3 5 ] [ 2 4 5 ]
52: [ 3 4 4 ] [ 3 8 ]
53: [ 3 8 ] [ 3 3 5 ]
54: [ 4 7 ] [ 3 4 4 ]
55: [ 5 6 ] [ 4 7 ]
56: [ 11 ] [ 5 6 ]
\end{verbatim}
\caption{\label{fig:partition-asc}
The partitions of 11 as weakly increasing lists of parts,
lexicographic order (left) and subset-lex order (right).
}
\end{figure}
Figure \ref{fig:partition-asc}
shows the partitions of 11 as weakly increasing lists of parts,
in lexicographic order (left) and in subset-lex order (right).
We first give a description of the computation of
the successor in lexicographic order.
The three last parts of the partition are denoted by $x$, $y$, and $z$.
\begin{samepage}
\begin{alg}[Next-Part-Asc
Compute the successor of a partition in lexicographic order.
\end{alg}
\begin{enumerate}
\item If there is just one part, stop.
\item If $z-1 < y+1$, replace $y,z$ by $y+z$; return.
\item Otherwise, change $y$ to $y+1$ and append
parts $y+1$ as long as there are at least $y+1$ units left.
\item Add the remaining units to the last part.
\end{enumerate}
\end{samepage}
The second step of the algorithm can be merged
into the other steps, as it is a special case of them.
{\codesize
\begin{listing}{1}
// FILE: src/comb/partition-asc.h
class partition_asc
{
ulong *a_; // partition: a[1] + a[2] + ... + a[m] = n
ulong n_; // integer partitions of n
ulong m_; // current partition has m parts
\end{listing}
}%
The implementation is correct for all $n\geq{}0$,
for $n=0$ the empty list of parts is generated.
{\codesize
\begin{listing}{1}
ulong next()
// Return number of parts of generated partition.
// Return zero if the current partition is the last.
{
if ( m_ <= 1 ) return 0; // current is last
ulong z1 = a_[m_] - 1; // take one unit from last part
m_ -= 1;
const ulong y1 = a_[m_] + 1; // add one unit to previous part
while ( y1 <= z1 ) // can put part Y+1
{
a_[m_] = y1;
z1 -= y1;
m_ += 1;
}
a_[m_] = y1 + z1; // add remaining units to last part
return m_;
}
\end{listing}
}%
The computation of the successor in subset-lex order is loopless.
\begin{samepage}
\begin{alg}[Next-Part-Asc-SL
Compute the successor of a partition in subset-lex order.
A sentinel zero shall precede list of elements.
\end{alg}
\begin{enumerate}
\item If $z\geq{}2\,y$, replace $y,z$ by $y,y,z-y$ (extend to the right); return.
\item If $z-1\geq{}y+1$, replace $y,z$ by $y+1,z-1$ (add one unit to the left); return.
\item If $z=1$ (the all-ones partition) do the following.
Stop if the number of parts is $\leq{}3$,
otherwise replace the tail $[1,1,1,1]$ by $[2,2]$ and return.
\item If the number of parts is 2, stop.
\item Replace $x, y, z$ by $x+1, y+z-1$ (add one unit to second left, add rest to end) and return.
\end{enumerate}
\end{samepage}
Note that with each update all but the last two parts
are the same as in the predecessor.
This can be an advantage over the lexicographic order for
computations where partial results for prefixes can be reused.
{\codesize
\begin{listing}{1}
// FILE: src/comb/partition-asc-subset-lex.h
class partition_asc_subset_lex
{
ulong *a_; // partition: a[1] + a[2] + ... + a[m] = n
ulong n_; // partition of n
ulong m_; // current partition has m parts
explicit partition_asc_subset_lex(ulong n)
{
n_ = n;
a_ = new ulong[n_+1+(n_==0)];
// sentinel a[0] set in first()
first();
}
void first()
{
a_[0] = 1; // sentinel: read (once) by test for right-extension
a_[1] = n_ + (n_==0); // use partitions of n=1 for n=0 internally
m_ = 1;
}
\end{listing}
}%
The implementation is again correct for all $n\geq{}0$.
{\codesize
\begin{listing}{1}
ulong next()
// Loopless algorithm.
{
ulong y = a_[m_-1]; // may read sentinel a[0]
ulong z = a_[m_];
if ( z >= 2*y ) // extend to the right:
{ // [*, Y, Z] --> [*, Y, Y, Z-Y]
a_[m_] = y;
a_[m_+1] = z - y; // >= y
++m_;
return m_;
}
z -= 1; y += 1;
if ( z >= y ) // add one unit to the left:
{ // [*, Y, Z] --> [*, Y+1, Z-1]
a_[m_-1] = y;
a_[m_] = z;
return m_;
}
if ( z == 0 ) // all-ones partition
{
if ( n_ <= 3 ) return 0; // current is last
// [1, ..., 1, 1, 1, 1] --> [1, ..., 2, 2]
m_ -= 2;
a_[m_] = 2; a_[m_-1] = 2;
return m_;
}
// add one unit to second left, add rest to end:
// [*, X, Y, Z] --> [*, X+1, Y+Z-1]
a_[m_-2] += 1;
a_[m_-1] += z;
m_ -= 1;
m_ -= (m_ == 1); // last if partition is into one part
return m_;
}
\end{listing}
}%
The updates via \texttt{next()}
take about 15 cycles for lexicographic order
and about 13 cycles for subset-lex order.
\subsection{Partitions into odd and into distinct parts}
\begin{figure
{\jjbls
\begin{verbatim}
0: [ 19 ] [ 19 ]
1: [ 1 18 ] [ 1 1 17 ]
2: [ 1 2 16 ] [ 1 1 1 1 15 ]
3: [ 1 2 3 13 ] [ 1 1 1 1 1 1 13 ]
4: [ 1 2 3 4 9 ] [ 1 1 1 1 1 1 1 1 11 ]
5: [ 1 2 3 5 8 ] [ 1 1 1 1 1 1 1 1 1 1 9 ]
6: [ 1 2 3 6 7 ] [ 1 1 1 1 1 1 1 1 1 1 1 1 7 ]
7: [ 1 2 4 12 ] [ 1 1 1 1 1 1 1 1 1 1 1 1 1 1 5 ]
8: [ 1 2 4 5 7 ] [ 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 3 ]
9: [ 1 2 5 11 ] [ 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ]
10: [ 1 2 6 10 ] [ 1 1 1 1 1 1 1 1 1 1 1 1 1 3 3 ]
11: [ 1 2 7 9 ] [ 1 1 1 1 1 1 1 1 1 1 1 3 5 ]
12: [ 1 3 15 ] [ 1 1 1 1 1 1 1 1 1 1 3 3 3 ]
13: [ 1 3 4 11 ] [ 1 1 1 1 1 1 1 1 1 3 7 ]
14: [ 1 3 4 5 6 ] [ 1 1 1 1 1 1 1 1 1 5 5 ]
15: [ 1 3 5 10 ] [ 1 1 1 1 1 1 1 1 3 3 5 ]
16: [ 1 3 6 9 ] [ 1 1 1 1 1 1 1 3 9 ]
17: [ 1 3 7 8 ] [ 1 1 1 1 1 1 1 3 3 3 3 ]
18: [ 1 4 14 ] [ 1 1 1 1 1 1 1 5 7 ]
19: [ 1 4 5 9 ] [ 1 1 1 1 1 1 3 3 7 ]
20: [ 1 4 6 8 ] [ 1 1 1 1 1 1 3 5 5 ]
21: [ 1 5 13 ] [ 1 1 1 1 1 3 11 ]
22: [ 1 5 6 7 ] [ 1 1 1 1 1 3 3 3 5 ]
23: [ 1 6 12 ] [ 1 1 1 1 1 5 9 ]
24: [ 1 7 11 ] [ 1 1 1 1 1 7 7 ]
25: [ 1 8 10 ] [ 1 1 1 1 3 3 9 ]
26: [ 2 17 ] [ 1 1 1 1 3 3 3 3 3 ]
27: [ 2 3 14 ] [ 1 1 1 1 3 5 7 ]
28: [ 2 3 4 10 ] [ 1 1 1 1 5 5 5 ]
29: [ 2 3 5 9 ] [ 1 1 1 3 13 ]
30: [ 2 3 6 8 ] [ 1 1 1 3 3 3 7 ]
31: [ 2 4 13 ] [ 1 1 1 3 3 5 5 ]
32: [ 2 4 5 8 ] [ 1 1 1 5 11 ]
33: [ 2 4 6 7 ] [ 1 1 1 7 9 ]
34: [ 2 5 12 ] [ 1 1 3 3 11 ]
35: [ 2 6 11 ] [ 1 1 3 3 3 3 5 ]
36: [ 2 7 10 ] [ 1 1 3 5 9 ]
37: [ 2 8 9 ] [ 1 1 3 7 7 ]
38: [ 3 16 ] [ 1 1 5 5 7 ]
39: [ 3 4 12 ] [ 1 3 15 ]
40: [ 3 4 5 7 ] [ 1 3 3 3 9 ]
41: [ 3 5 11 ] [ 1 3 3 3 3 3 3 ]
42: [ 3 6 10 ] [ 1 3 3 5 7 ]
43: [ 3 7 9 ] [ 1 3 5 5 5 ]
44: [ 4 15 ] [ 1 5 13 ]
45: [ 4 5 10 ] [ 1 7 11 ]
46: [ 4 6 9 ] [ 1 9 9 ]
47: [ 4 7 8 ] [ 3 3 13 ]
48: [ 5 14 ] [ 3 3 3 3 7 ]
49: [ 5 6 8 ] [ 3 3 3 5 5 ]
50: [ 6 13 ] [ 3 5 11 ]
51: [ 7 12 ] [ 3 7 9 ]
52: [ 8 11 ] [ 5 5 9 ]
53: [ 9 10 ] [ 5 7 7 ]
\end{verbatim}
\caption{\label{fig:partition-asc-odd-dist}
The partitions of 19 into distinct parts (left) and odd parts (right) in subset-lex order.}
\end{figure}
It is not very difficult to adapt the algorithms
to specialized partitions.
We give the algorithms for partitions into distinct and odd parts
by their implementations.
Computation of the successor for partitions into distinct parts
in lexicographic order:
{\codesize
\begin{listing}{1}
// FILE: src/comb/partition-dist-asc.h
ulong next()
{
if ( m_ <= 1 ) return 0; // current is last
ulong s = a_[m_] + a_[m_-1];
ulong k = a_[m_-1] + 1;
m_ -= 1;
// split s into k, k+1, k+2, ..., y, z where z >= y + 1:
while ( s >= ( k + (k+1) ) )
{
a_[m_] = k;
s -= k;
k += 1;
m_ += 1;
}
a_[m_] = s;
return m_;
}
\end{listing}
}%
Computation of the successor for partitions into odd parts
in lexicographic order:
{\codesize
\begin{listing}{1}
// FILE: src/comb/partition-odd-asc.h
ulong next()
{
const ulong z = a_[m_]; // can read sentinel a[0] if n==0
const ulong y = a_[m_-1]; // can read sentinel a[0] (a[-1] for n==0)
ulong s = y + z; // sum of parts we scan over
ulong k; // min value of next term
if ( z >= y+4 ) // add last 2 terms
{
if ( m_ == 1 ) return 0; // current is last
k = y + 2;
a_[m_-1] = k;
s -= k;
}
else // add last 3 terms
{
if ( m_ <= 2 ) return 0; // current is last
const ulong x = a_[m_-2];
s += x;
k = x + 2;
m_ -= 2;
}
const ulong k2 = k + k;
while ( s >= k2 + k )
{
a_[m_] = k; s -= k; m_ += 1;
a_[m_] = k; s -= k; m_ += 1;
}
a_[m_] = s;
return m_;
}
\end{listing}
}%
All routines in this section return the number of parts in the generated
partition and return zero if there are no more partitions.
Computation of the successor for partitions into distinct parts
in subset-lex order:
{\codesize
\begin{listing}{1}
// FILE: src/comb/partition-dist-asc-subset-lex.h
ulong next()
// Loopless algorithm.
{
ulong y = a_[m_-1]; // may read sentinel a[0]
ulong z = a_[m_];
if ( z >= 2*y + 3 ) // can extend to the right
{ // [*, Y, Z] --> [*, Y, Y+1, Z-1]
y += 1;
a_[m_] = y;
a_[m_+1] = z - y; // >= y
++m_;
return m_;
}
else // add to the left
{
z -= 1; y += 1;
if ( z > y ) // add one unit to the left
{ // [*, Y, Z] --> [*, Y+1, Z-1]
if ( m_<=1 ) return 0; // current is last
a_[m_-1] = y;
a_[m_] = z;
return m_;
}
else // add to one unit second left
// and combine last with second last
{ // [*, X, Y, Z] --> [*, X+1, Y+Z]
if ( m_<=2 ) return 0; // current is last
a_[m_-2] += 1;
a_[m_-1] += z;
--m_;
return m_;
}
}
}
\end{listing}
}%
Computation of the successor for partitions into odd parts
in subset-lex order:
{\codesize
\begin{listing}{1}
// FILE: src/comb/partitions-odd-asc-subset-lex.h
ulong next()
// Loopless algorithm.
{
ulong y = a_[m_-1]; // may read sentinel a[0]
ulong z = a_[m_];
if ( z >= 3*y ) // can extend to the right
{ // [*, Y, Z] --> [*, Y, Y, Y, Z-2*Y]
a_[m_] = y;
a_[m_+1] = y;
a_[m_+2] = z - 2 * y;
m_ += 2;
return m_;
}
if ( m_ >= n_ ) // all-ones partition
{
if ( n_ <= 5 ) return 0; // current is last
// [1, ..., 1, 1, 1, 1, 1, 1] --> [1, ..., 3, 3]
m_ -= 4;
a_[m_] = 3;
a_[m_-1] = 3;
return m_;
}
ulong z2 = z - 2;
ulong y2 = y + 2;
if ( z2 >= y2 ) // add 2 units to the left
{ // [*, Y, Z] --> [*, Y+2, Z-2]
a_[m_-1] = y2;
a_[m_] = z2;
return m_;
}
if ( m_==2 ) // current is last (happens only for n even)
return 0;
// here m >= 3; add to second or third left:
ulong x2 = a_[m_-2] + 2;
ulong s = z + y - 2;
if ( x2 <= z2 )
// add 2 units to third left, repeat part, put rest at end
{ // [*, X, Y, Z] --> [*, X+2, X+2, Y+Z-2 -X-2]
a_[m_-2] = x2;
a_[m_-1] = x2;
a_[m_] = s - x2;
return m_;
}
if ( m_==3 ) // current is last (happens only for n odd)
return 0;
// add 2 units to third left, combine rest into second left
// [*, W, X, Y, Z] --> [*, W+2, X+Y+Z-2]
a_[m_-3] += 2;
a_[m_-2] += s;
m_ -= 2;
return m_;
}
\end{listing}
}%
The updates via \texttt{next()} of partitions into distinct parts take about 11
cycles for lexicographic order and about 9 cycles for subset-lex order. The
respective figures for the partitions into odd parts are 13 and 13.5 cycles.
Algorithms for ranking and unranking are obtained by replacing $2^{n-f-1}$ (the
number of compositions of $n$ with first part $f$) in algorithms
\ref{alg:unrank-comp-sl} and \ref{alg:rank-comp-sl} by the expression for the
number of partitions of $n$ (of the desired kind) with first part $f$.
\section{Subset-lex order for restricted growth strings (RGS)}
By modifying the algorithms for the successor and predecessor (\ref{alg:next-sl}
and \ref{alg:prev-sl}) for subset-lex order to adhere to certain conditions,
one obtains the equivalents for RGS.
These can often be given without any difficulty.
We give two examples,
RGS for set partitions and
RGS for $k$-ary Dyck words.
\subsection{RGS for set partitions}
\begin{figure
{\jjbls
\begin{verbatim}
lexicographic subset-lex
1: [ . . . . . ] {1 2 3 4 5} [ . . . . . ] {1 2 3 4 5}
2: [ . . . . 1 ] {1 2 3 4} {5} [ . 1 . . . ] {1 3 4 5} {2}
3: [ . . . 1 . ] {1 2 3 5} {4} [ . 1 1 . . ] {1 4 5} {2 3}
4: [ . . . 1 1 ] {1 2 3} {4 5} [ . 1 2 . . ] {1 4 5} {2} {3}
5: [ . . . 1 2 ] {1 2 3} {4} {5} [ . 1 2 1 . ] {1 5} {2 4} {3}
6: [ . . 1 . . ] {1 2 4 5} {3} [ . 1 2 2 . ] {1 5} {2} {3 4}
7: [ . . 1 . 1 ] {1 2 4} {3 5} [ . 1 2 3 . ] {1 5} {2} {3} {4}
8: [ . . 1 . 2 ] {1 2 4} {3} {5} [ . 1 2 3 1 ] {1} {2 5} {3} {4}
9: [ . . 1 1 . ] {1 2 5} {3 4} [ . 1 2 3 2 ] {1} {2} {3 5} {4}
10: [ . . 1 1 1 ] {1 2} {3 4 5} [ . 1 2 3 3 ] {1} {2} {3} {4 5}
11: [ . . 1 1 2 ] {1 2} {3 4} {5} [ . 1 2 3 4 ] {1} {2} {3} {4} {5}
12: [ . . 1 2 . ] {1 2 5} {3} {4} [ . 1 2 2 1 ] {1} {2 5} {3 4}
13: [ . . 1 2 1 ] {1 2} {3 5} {4} [ . 1 2 2 2 ] {1} {2} {3 4 5}
14: [ . . 1 2 2 ] {1 2} {3} {4 5} [ . 1 2 2 3 ] {1} {2} {3 4} {5}
15: [ . . 1 2 3 ] {1 2} {3} {4} {5} [ . 1 2 1 1 ] {1} {2 4 5} {3}
16: [ . 1 . . . ] {1 3 4 5} {2} [ . 1 2 1 2 ] {1} {2 4} {3 5}
17: [ . 1 . . 1 ] {1 3 4} {2 5} [ . 1 2 1 3 ] {1} {2 4} {3} {5}
18: [ . 1 . . 2 ] {1 3 4} {2} {5} [ . 1 2 . 1 ] {1 4} {2 5} {3}
19: [ . 1 . 1 . ] {1 3 5} {2 4} [ . 1 2 . 2 ] {1 4} {2} {3 5}
20: [ . 1 . 1 1 ] {1 3} {2 4 5} [ . 1 2 . 3 ] {1 4} {2} {3} {5}
21: [ . 1 . 1 2 ] {1 3} {2 4} {5} [ . 1 1 1 . ] {1 5} {2 3 4}
22: [ . 1 . 2 . ] {1 3 5} {2} {4} [ . 1 1 2 . ] {1 5} {2 3} {4}
23: [ . 1 . 2 1 ] {1 3} {2 5} {4} [ . 1 1 2 1 ] {1} {2 3 5} {4}
24: [ . 1 . 2 2 ] {1 3} {2} {4 5} [ . 1 1 2 2 ] {1} {2 3} {4 5}
25: [ . 1 . 2 3 ] {1 3} {2} {4} {5} [ . 1 1 2 3 ] {1} {2 3} {4} {5}
26: [ . 1 1 . . ] {1 4 5} {2 3} [ . 1 1 1 1 ] {1} {2 3 4 5}
27: [ . 1 1 . 1 ] {1 4} {2 3 5} [ . 1 1 1 2 ] {1} {2 3 4} {5}
28: [ . 1 1 . 2 ] {1 4} {2 3} {5} [ . 1 1 . 1 ] {1 4} {2 3 5}
29: [ . 1 1 1 . ] {1 5} {2 3 4} [ . 1 1 . 2 ] {1 4} {2 3} {5}
30: [ . 1 1 1 1 ] {1} {2 3 4 5} [ . 1 . 1 . ] {1 3 5} {2 4}
31: [ . 1 1 1 2 ] {1} {2 3 4} {5} [ . 1 . 2 . ] {1 3 5} {2} {4}
32: [ . 1 1 2 . ] {1 5} {2 3} {4} [ . 1 . 2 1 ] {1 3} {2 5} {4}
33: [ . 1 1 2 1 ] {1} {2 3 5} {4} [ . 1 . 2 2 ] {1 3} {2} {4 5}
34: [ . 1 1 2 2 ] {1} {2 3} {4 5} [ . 1 . 2 3 ] {1 3} {2} {4} {5}
35: [ . 1 1 2 3 ] {1} {2 3} {4} {5} [ . 1 . 1 1 ] {1 3} {2 4 5}
36: [ . 1 2 . . ] {1 4 5} {2} {3} [ . 1 . 1 2 ] {1 3} {2 4} {5}
37: [ . 1 2 . 1 ] {1 4} {2 5} {3} [ . 1 . . 1 ] {1 3 4} {2 5}
38: [ . 1 2 . 2 ] {1 4} {2} {3 5} [ . 1 . . 2 ] {1 3 4} {2} {5}
39: [ . 1 2 . 3 ] {1 4} {2} {3} {5} [ . . 1 . . ] {1 2 4 5} {3}
40: [ . 1 2 1 . ] {1 5} {2 4} {3} [ . . 1 1 . ] {1 2 5} {3 4}
41: [ . 1 2 1 1 ] {1} {2 4 5} {3} [ . . 1 2 . ] {1 2 5} {3} {4}
42: [ . 1 2 1 2 ] {1} {2 4} {3 5} [ . . 1 2 1 ] {1 2} {3 5} {4}
43: [ . 1 2 1 3 ] {1} {2 4} {3} {5} [ . . 1 2 2 ] {1 2} {3} {4 5}
44: [ . 1 2 2 . ] {1 5} {2} {3 4} [ . . 1 2 3 ] {1 2} {3} {4} {5}
45: [ . 1 2 2 1 ] {1} {2 5} {3 4} [ . . 1 1 1 ] {1 2} {3 4 5}
46: [ . 1 2 2 2 ] {1} {2} {3 4 5} [ . . 1 1 2 ] {1 2} {3 4} {5}
47: [ . 1 2 2 3 ] {1} {2} {3 4} {5} [ . . 1 . 1 ] {1 2 4} {3 5}
48: [ . 1 2 3 . ] {1 5} {2} {3} {4} [ . . 1 . 2 ] {1 2 4} {3} {5}
49: [ . 1 2 3 1 ] {1} {2 5} {3} {4} [ . . . 1 . ] {1 2 3 5} {4}
50: [ . 1 2 3 2 ] {1} {2} {3 5} {4} [ . . . 1 1 ] {1 2 3} {4 5}
51: [ . 1 2 3 3 ] {1} {2} {3} {4 5} [ . . . 1 2 ] {1 2 3} {4} {5}
52: [ . 1 2 3 4 ] {1} {2} {3} {4} {5} [ . . . . 1 ] {1 2 3 4} {5}
\end{verbatim}
\caption{\label{fig:setpart-rgs}
Restricted growth strings and corresponding set partitions in
lexicographic and subset-lex order.
}
\end{figure}
We consider the restricted growth strings (RGS)
$a_0,\,a_1,\,\ldots,\,a_{n-1}$ such that
$a_0=0$ and $a_k \leq{} 1 + \max(a_0,\, a_1,\, \ldots,\, a_{k-1})$.
These RGS are counted by the Bell numbers, see sequence \jjseqref{A000110} in \cite{oeis}.
Figure \ref{fig:setpart-rgs} shows the all such RGS of length 5
in lexicographic and subset-lex order.
The set partitions are obtained by putting
all $i$ such that $a_i=k$ into the same set.
In the following algorithm we assume that
$m_k=1 + \max(a_0,\, a_1,\, \ldots,\, a_{k-1})$
and that there is a sentinel $a_{-1}=1$.
\begin{alg}[Next-Setpart-RGS
Compute the successor in subset-lex order.
\end{alg}
\begin{enumerate}
\item Let $j$ be the index of the last nonzero digit ($j=1$ for the all-zero RGS).
\item If $a_j < m_j$, set $a_j=a_j+1$ and return.
\item If $j+1 < n$, set $a_{j+1}=1$ and return.
\item Set $a_j=0$.
\item Set $j$ to the position of the next nonzero digit to the left.
If $j=-1$, stop.
\item Set $a_j=a_j-1$ and $a_{j+1}=1$.
\end{enumerate}
The implementation has to take care of updating
the array $m[]$ which has an additional (write-only)
element at its end.
{\codesize
\begin{listing}{1}
// FILE: src/comb/setpart-rgs-subset-lex.h
class setpart_rgs_subset_lex
{
ulong *a_; // digits of the RGS
ulong *m_; // maximum + 1 in prefix
ulong tr_; // current track
ulong n_; // number of digits in RGS
\end{listing}
}%
The computation of the successor is correct for all $n\geq{}0$,
we omit the constructor, which sets $m_0=1$ for $n=0$ to
cover this case.
{\codesize
\begin{listing}{1}
bool next()
{
ulong j = tr_;
if ( a_[j] < m_[j] ) // easy case 1: can increment track
{
if ( n_ <= 1 ) return false; // handle n <= 1 correctly
a_[j] += 1;
return true;
}
const ulong j1 = j + 1;
if ( j1 < n_ ) // easy case 2: can attach
{
m_[j1] = m_[j] + 1;
a_[j1] = +1;
tr_ = j1;
return true;
}
a_[j] = 0;
m_[j] = m_[j-1];
// Find nonzero track to the left:
do { --j; } while ( a_[j] == 0 ); // can read sentinel
if ( (long)j < 0 ) return false; // current is last
if ( a_[j] == m_[j] ) m_[j+1] = m_[j];
a_[j] -= 1;
++j;
a_[j] = 1;
tr_ = j;
return true;
}
\end{listing}
}%
An update takes about 9.5 cycles for subset-lex order
and 12.5 cycles for lexicographic order.
\subsection{RGS for $k$-ary Dyck words}
\begin{figure
{\jjbls
\begin{verbatim}
1: [ . . . . ] 1..1..1..1.. [ . . . . ] 1..1..1..1..
2: [ . . . 1 ] 1..1..1.1... [ . 1 . . ] 1.1...1..1..
3: [ . . . 2 ] 1..1..11.... [ . 2 . . ] 11....1..1..
4: [ . . 1 . ] 1..1.1...1.. [ . 2 1 . ] 11...1...1..
5: [ . . 1 1 ] 1..1.1..1... [ . 2 2 . ] 11..1....1..
6: [ . . 1 2 ] 1..1.1.1.... [ . 2 3 . ] 11.1.....1..
7: [ . . 1 3 ] 1..1.11..... [ . 2 4 . ] 111......1..
8: [ . . 2 . ] 1..11....1.. [ . 2 4 1 ] 111.....1...
9: [ . . 2 1 ] 1..11...1... [ . 2 4 2 ] 111....1....
10: [ . . 2 2 ] 1..11..1.... [ . 2 4 3 ] 111...1.....
11: [ . . 2 3 ] 1..11.1..... [ . 2 4 4 ] 111..1......
12: [ . . 2 4 ] 1..111...... [ . 2 4 5 ] 111.1.......
13: [ . 1 . . ] 1.1...1..1.. [ . 2 4 6 ] 1111........
14: [ . 1 . 1 ] 1.1...1.1... [ . 2 3 1 ] 11.1....1...
15: [ . 1 . 2 ] 1.1...11.... [ . 2 3 2 ] 11.1...1....
16: [ . 1 1 . ] 1.1..1...1.. [ . 2 3 3 ] 11.1..1.....
17: [ . 1 1 1 ] 1.1..1..1... [ . 2 3 4 ] 11.1.1......
18: [ . 1 1 2 ] 1.1..1.1.... [ . 2 3 5 ] 11.11.......
19: [ . 1 1 3 ] 1.1..11..... [ . 2 2 1 ] 11..1...1...
20: [ . 1 2 . ] 1.1.1....1.. [ . 2 2 2 ] 11..1..1....
21: [ . 1 2 1 ] 1.1.1...1... [ . 2 2 3 ] 11..1.1.....
22: [ . 1 2 2 ] 1.1.1..1.... [ . 2 2 4 ] 11..11......
23: [ . 1 2 3 ] 1.1.1.1..... [ . 2 1 1 ] 11...1..1...
24: [ . 1 2 4 ] 1.1.11...... [ . 2 1 2 ] 11...1.1....
25: [ . 1 3 . ] 1.11.....1.. [ . 2 1 3 ] 11...11.....
26: [ . 1 3 1 ] 1.11....1... [ . 2 . 1 ] 11....1.1...
27: [ . 1 3 2 ] 1.11...1.... [ . 2 . 2 ] 11....11....
28: [ . 1 3 3 ] 1.11..1..... [ . 1 1 . ] 1.1..1...1..
29: [ . 1 3 4 ] 1.11.1...... [ . 1 2 . ] 1.1.1....1..
30: [ . 1 3 5 ] 1.111....... [ . 1 3 . ] 1.11.....1..
31: [ . 2 . . ] 11....1..1.. [ . 1 3 1 ] 1.11....1...
32: [ . 2 . 1 ] 11....1.1... [ . 1 3 2 ] 1.11...1....
33: [ . 2 . 2 ] 11....11.... [ . 1 3 3 ] 1.11..1.....
34: [ . 2 1 . ] 11...1...1.. [ . 1 3 4 ] 1.11.1......
35: [ . 2 1 1 ] 11...1..1... [ . 1 3 5 ] 1.111.......
36: [ . 2 1 2 ] 11...1.1.... [ . 1 2 1 ] 1.1.1...1...
37: [ . 2 1 3 ] 11...11..... [ . 1 2 2 ] 1.1.1..1....
38: [ . 2 2 . ] 11..1....1.. [ . 1 2 3 ] 1.1.1.1.....
39: [ . 2 2 1 ] 11..1...1... [ . 1 2 4 ] 1.1.11......
40: [ . 2 2 2 ] 11..1..1.... [ . 1 1 1 ] 1.1..1..1...
41: [ . 2 2 3 ] 11..1.1..... [ . 1 1 2 ] 1.1..1.1....
42: [ . 2 2 4 ] 11..11...... [ . 1 1 3 ] 1.1..11.....
43: [ . 2 3 . ] 11.1.....1.. [ . 1 . 1 ] 1.1...1.1...
44: [ . 2 3 1 ] 11.1....1... [ . 1 . 2 ] 1.1...11....
45: [ . 2 3 2 ] 11.1...1.... [ . . 1 . ] 1..1.1...1..
46: [ . 2 3 3 ] 11.1..1..... [ . . 2 . ] 1..11....1..
47: [ . 2 3 4 ] 11.1.1...... [ . . 2 1 ] 1..11...1...
48: [ . 2 3 5 ] 11.11....... [ . . 2 2 ] 1..11..1....
49: [ . 2 4 . ] 111......1.. [ . . 2 3 ] 1..11.1.....
50: [ . 2 4 1 ] 111.....1... [ . . 2 4 ] 1..111......
51: [ . 2 4 2 ] 111....1.... [ . . 1 1 ] 1..1.1..1...
52: [ . 2 4 3 ] 111...1..... [ . . 1 2 ] 1..1.1.1....
53: [ . 2 4 4 ] 111..1...... [ . . 1 3 ] 1..1.11.....
54: [ . 2 4 5 ] 111.1....... [ . . . 1 ] 1..1..1.1...
55: [ . 2 4 6 ] 1111........ [ . . . 2 ] 1..1..11....
\end{verbatim}
\caption{\label{fig:dyck-words}
The 55 RGS for the 3-ary Dyck words of length 12, in
lexicographic order (left) and subset-lex order (right).
}
\end{figure}
Figure \ref{fig:dyck-words} shows
the 55 RGS for the 3-ary Dyck words of length 12, in
both lexicographic and subset-lex order.
The $j$th value in each RGS contains the distance of the position $j$th one in
the Dyck word from its maximal value $k\cdot{}j$
The sequences of numbers of $k$-ary Dyck words are entries
\jjseqref{A000108} ($k=2$, Catalan numbers),
\jjseqref{A001764} ($k=3$),
\jjseqref{A002293} ($k=4$), and
\jjseqref{A002294} ($k=5$) in \cite{oeis}.
It shall suffice to give the implementation for the (loopless) computation of
the predecessor in subset-lex order.
A sentinel \texttt{a[-1]=+1} is used.
{\codesize
\begin{listing}{1}
// FILE: src/comb/dyck-rgs-subset-lex.h
class dyck_rgs_subset_lex
{
ulong *a_; // digits of the RGS: a_[k] <= as[k-1] + 1
ulong tr_; // current track
ulong n_; // number of digits in RGS
ulong i_; // k-ary Dyck words: i = k - 1
void last()
{
for (ulong k=0; k<n_; ++k) a_[k] = 0;
tr_ = n_ - 1;
// make things work for n <= 1:
if ( n_==0 )
{
tr_ = 0;
a_[0] = 1;
}
if ( n_>=2 ) a_[tr_] = i_;
}
\end{listing}
}%
All cases $n\geq{}0$ are handled correctly.
{\codesize
\begin{listing}{1}
bool prev()
// Loopless algorithm.
{
if ( n_<=1 ) return false; // just one RGS
ulong j = tr_;
if ( a_[j] > 1 ) // can decrement track
{
a_[j] -= 1;
return true;
}
const ulong aj = a_[j]; // zero or one
a_[j] = 0;
--j;
if ( a_[j] == a_[j-1] + i_ ) // move track to the left
{
--tr_;
return true;
}
if ( j==0 ) // current or next is last
{
if ( aj == 0 ) return false;
return true;
}
a_[j] += 1; // increment left digit
tr_ = n_ - 1; // move to right end
a_[tr_] = a_[tr_-1] + i_; // set to max value
return true;
}
\end{listing}
}%
One update takes about $10$ cycles for both
lexicographic and subset-lex order.
For the generation of other restricted growth strings in subset-lex order,
see the following files.
{\jjbls
\begin{verbatim}
// Ascent sequences, see OEIS sequence A022493:
src/comb/ascent-rgs-subset-lex.h
src/comb/ascent-rgs.h // lexicographic order
// RGS for Catalan objects, see OEIS sequence A000108:
src/comb/catalan-rgs-subset-lex.h (loopless prev())
src/comb/catalan-rgs.h // lexicographic order
// Standard Young tableaux, represented as ballot sequences,
// see OEIS sequence A000085:
src/comb/young-tab-rgs-subset-lex.h
src/comb/young-tab-rgs.h // lexicographic order
\end{verbatim}
}%
\section{A variant of the subset-lex order}
\begin{figure
\vspace*{-10mm
{\jjbls
\begin{verbatim}
0: [ . . . . ] { }
1: [ . . . 1 ] { 3 }
2: [ . . . 2 ] { 3, 3 }
3: [ . . . 3 ] { 3, 3, 3 }
4: [ . . 1 3 ] { 3, 3, 3, 2 }
5: [ . . 2 3 ] { 3, 3, 3, 2, 2 }
6: [ . 1 2 3 ] { 3, 3, 3, 2, 2, 1 }
7: [ . 2 2 3 ] { 3, 3, 3, 2, 2, 1, 1 }
8: [ 1 2 2 3 ] { 3, 3, 3, 2, 2, 1, 1, 0 }
9: [ 1 1 2 3 ] { 3, 3, 3, 2, 2, 1, 0 }
10: [ 1 . 2 3 ] { 3, 3, 3, 2, 2, 0 }
11: [ . 1 1 3 ] { 3, 3, 3, 2, 1 }
12: [ . 2 1 3 ] { 3, 3, 3, 2, 1, 1 }
13: [ 1 2 1 3 ] { 3, 3, 3, 2, 1, 1, 0 }
14: [ 1 1 1 3 ] { 3, 3, 3, 2, 1, 0 }
15: [ 1 . 1 3 ] { 3, 3, 3, 2, 0 }
16: [ . 1 . 3 ] { 3, 3, 3, 1 }
17: [ . 2 . 3 ] { 3, 3, 3, 1, 1 }
18: [ 1 2 . 3 ] { 3, 3, 3, 1, 1, 0 }
19: [ 1 1 . 3 ] { 3, 3, 3, 1, 0 }
20: [ 1 . . 3 ] { 3, 3, 3, 0 }
21: [ . . 1 2 ] { 3, 3, 2 }
22: [ . . 2 2 ] { 3, 3, 2, 2 }
23: [ . 1 2 2 ] { 3, 3, 2, 2, 1 }
24: [ . 2 2 2 ] { 3, 3, 2, 2, 1, 1 }
25: [ 1 2 2 2 ] { 3, 3, 2, 2, 1, 1, 0 }
26: [ 1 1 2 2 ] { 3, 3, 2, 2, 1, 0 }
27: [ 1 . 2 2 ] { 3, 3, 2, 2, 0 }
28: [ . 1 1 2 ] { 3, 3, 2, 1 }
29: [ . 2 1 2 ] { 3, 3, 2, 1, 1 }
30: [ 1 2 1 2 ] { 3, 3, 2, 1, 1, 0 }
31: [ 1 1 1 2 ] { 3, 3, 2, 1, 0 }
32: [ 1 . 1 2 ] { 3, 3, 2, 0 }
33: [ . 1 . 2 ] { 3, 3, 1 }
34: [ . 2 . 2 ] { 3, 3, 1, 1 }
35: [ 1 2 . 2 ] { 3, 3, 1, 1, 0 }
36: [ 1 1 . 2 ] { 3, 3, 1, 0 }
37: [ 1 . . 2 ] { 3, 3, 0 }
38: [ . . 1 1 ] { 3, 2 }
39: [ . . 2 1 ] { 3, 2, 2 }
40: [ . 1 2 1 ] { 3, 2, 2, 1 }
41: [ . 2 2 1 ] { 3, 2, 2, 1, 1 }
42: [ 1 2 2 1 ] { 3, 2, 2, 1, 1, 0 }
43: [ 1 1 2 1 ] { 3, 2, 2, 1, 0 }
44: [ 1 . 2 1 ] { 3, 2, 2, 0 }
45: [ . 1 1 1 ] { 3, 2, 1 }
46: [ . 2 1 1 ] { 3, 2, 1, 1 }
47: [ 1 2 1 1 ] { 3, 2, 1, 1, 0 }
48: [ 1 1 1 1 ] { 3, 2, 1, 0 }
49: [ 1 . 1 1 ] { 3, 2, 0 }
50: [ . 1 . 1 ] { 3, 1 }
51: [ . 2 . 1 ] { 3, 1, 1 }
52: [ 1 2 . 1 ] { 3, 1, 1, 0 }
53: [ 1 1 . 1 ] { 3, 1, 0 }
54: [ 1 . . 1 ] { 3, 0 }
55: [ . . 1 . ] { 2 }
56: [ . . 2 . ] { 2, 2 }
57: [ . 1 2 . ] { 2, 2, 1 }
58: [ . 2 2 . ] { 2, 2, 1, 1 }
59: [ 1 2 2 . ] { 2, 2, 1, 1, 0 }
60: [ 1 1 2 . ] { 2, 2, 1, 0 }
61: [ 1 . 2 . ] { 2, 2, 0 }
62: [ . 1 1 . ] { 2, 1 }
63: [ . 2 1 . ] { 2, 1, 1 }
64: [ 1 2 1 . ] { 2, 1, 1, 0 }
65: [ 1 1 1 . ] { 2, 1, 0 }
66: [ 1 . 1 . ] { 2, 0 }
67: [ . 1 . . ] { 1 }
68: [ . 2 . . ] { 1, 1 }
69: [ 1 2 . . ] { 1, 1, 0 }
70: [ 1 1 . . ] { 1, 0 }
71: [ 1 . . . ] { 0 }
\end{verbatim}
\caption{\label{fig:subset-lexrev}
Subsets of the set $\{0^1, 1^2, 2^2, 3^3 \}$ in subset-lexrev order.
Dots denote zeros in the (generalized) characteristic words.
Note the sets are printed starting with the largest element.
}
\end{figure}
An order obtained by processing the positions in the characteristic
words as for subset-lex order but with reversed priorities is
shown in figure \ref{fig:subset-lexrev}.
We will call this ordering \jjterm{subset-lexrev}.
The algorithms for computing successor and predecessor
are easily obtained from those for the subset-lex order,
we just give the implementation for \texttt{prev()},
which is (again) loopless.
{\codesize
\begin{listing}{1}
// FILE: src/comb/mixedradix-subset-lexrev.h
bool prev()
{
ulong j = tr_;
if ( a_[j] > 1 ) // easy case: just decrement
{
a_[j] -= 1;
return true;
}
a_[j] = 0;
++j; // now looking at next track to the right
if ( j >= n_ ) // was on rightmost track (last two steps)
{
bool q = ( a_[j] != 0 );
a_[j] = 0;
return q;
}
if ( a_[j] == m1_[j] ) // semi-easy case: move track to left
{
tr_ = j; // move track one right
return true;
}
else
{
a_[j] += 1; // increment digit to the right
j = 0;
a_[j] = m1_[j]; // set leftmost digit = nine
tr_ = j; // move to leftmost track
return true;
}
}
\end{listing}
}%
\begin{figure
{\jjbls
\begin{verbatim}
lex colex subset-lexrev subset-lex
1: [ . . . . . ] [ . . . . . ] [ . . . . . ] [ . . . . . ]
2: [ . . . . 1 ] [ . . . . 1 ] [ . . . . 1 ] [ . 1 2 3 3 ]
3: [ . . . . 2 ] [ . . . 1 1 ] [ . . . . 2 ] [ . 1 2 3 4 ]
4: [ . . . . 3 ] [ . . 1 1 1 ] [ . . . . 3 ] [ . 1 2 2 2 ]
5: [ . . . . 4 ] [ . 1 1 1 1 ] [ . . . . 4 ] [ . 1 2 2 3 ]
6: [ . . . 1 1 ] [ . . . . 2 ] [ . . . 1 4 ] [ . 1 2 2 4 ]
7: [ . . . 1 2 ] [ . . . 1 2 ] [ . . . 2 4 ] [ . 1 1 3 3 ]
8: [ . . . 1 3 ] [ . . 1 1 2 ] [ . . . 3 4 ] [ . 1 1 3 4 ]
9: [ . . . 1 4 ] [ . 1 1 1 2 ] [ . . 1 3 4 ] [ . 1 1 2 2 ]
10: [ . . . 2 2 ] [ . . . 2 2 ] [ . . 2 3 4 ] [ . 1 1 2 3 ]
11: [ . . . 2 3 ] [ . . 1 2 2 ] [ . 1 2 3 4 ] [ . 1 1 2 4 ]
12: [ . . . 2 4 ] [ . 1 1 2 2 ] [ . 1 1 3 4 ] [ . 1 1 1 1 ]
13: [ . . . 3 3 ] [ . . 2 2 2 ] [ . . 1 2 4 ] [ . 1 1 1 2 ]
14: [ . . . 3 4 ] [ . 1 2 2 2 ] [ . . 2 2 4 ] [ . 1 1 1 3 ]
15: [ . . 1 1 1 ] [ . . . . 3 ] [ . 1 2 2 4 ] [ . 1 1 1 4 ]
16: [ . . 1 1 2 ] [ . . . 1 3 ] [ . 1 1 2 4 ] [ . . 2 3 3 ]
17: [ . . 1 1 3 ] [ . . 1 1 3 ] [ . . 1 1 4 ] [ . . 2 3 4 ]
18: [ . . 1 1 4 ] [ . 1 1 1 3 ] [ . 1 1 1 4 ] [ . . 2 2 2 ]
19: [ . . 1 2 2 ] [ . . . 2 3 ] [ . . . 1 3 ] [ . . 2 2 3 ]
20: [ . . 1 2 3 ] [ . . 1 2 3 ] [ . . . 2 3 ] [ . . 2 2 4 ]
21: [ . . 1 2 4 ] [ . 1 1 2 3 ] [ . . . 3 3 ] [ . . 1 3 3 ]
22: [ . . 1 3 3 ] [ . . 2 2 3 ] [ . . 1 3 3 ] [ . . 1 3 4 ]
23: [ . . 1 3 4 ] [ . 1 2 2 3 ] [ . . 2 3 3 ] [ . . 1 2 2 ]
24: [ . . 2 2 2 ] [ . . . 3 3 ] [ . 1 2 3 3 ] [ . . 1 2 3 ]
25: [ . . 2 2 3 ] [ . . 1 3 3 ] [ . 1 1 3 3 ] [ . . 1 2 4 ]
26: [ . . 2 2 4 ] [ . 1 1 3 3 ] [ . . 1 2 3 ] [ . . 1 1 1 ]
27: [ . . 2 3 3 ] [ . . 2 3 3 ] [ . . 2 2 3 ] [ . . 1 1 2 ]
28: [ . . 2 3 4 ] [ . 1 2 3 3 ] [ . 1 2 2 3 ] [ . . 1 1 3 ]
29: [ . 1 1 1 1 ] [ . . . . 4 ] [ . 1 1 2 3 ] [ . . 1 1 4 ]
30: [ . 1 1 1 2 ] [ . . . 1 4 ] [ . . 1 1 3 ] [ . . . 3 3 ]
31: [ . 1 1 1 3 ] [ . . 1 1 4 ] [ . 1 1 1 3 ] [ . . . 3 4 ]
32: [ . 1 1 1 4 ] [ . 1 1 1 4 ] [ . . . 1 2 ] [ . . . 2 2 ]
33: [ . 1 1 2 2 ] [ . . . 2 4 ] [ . . . 2 2 ] [ . . . 2 3 ]
34: [ . 1 1 2 3 ] [ . . 1 2 4 ] [ . . 1 2 2 ] [ . . . 2 4 ]
35: [ . 1 1 2 4 ] [ . 1 1 2 4 ] [ . . 2 2 2 ] [ . . . 1 1 ]
36: [ . 1 1 3 3 ] [ . . 2 2 4 ] [ . 1 2 2 2 ] [ . . . 1 2 ]
37: [ . 1 1 3 4 ] [ . 1 2 2 4 ] [ . 1 1 2 2 ] [ . . . 1 3 ]
38: [ . 1 2 2 2 ] [ . . . 3 4 ] [ . . 1 1 2 ] [ . . . 1 4 ]
39: [ . 1 2 2 3 ] [ . . 1 3 4 ] [ . 1 1 1 2 ] [ . . . . 1 ]
40: [ . 1 2 2 4 ] [ . 1 1 3 4 ] [ . . . 1 1 ] [ . . . . 2 ]
41: [ . 1 2 3 3 ] [ . . 2 3 4 ] [ . . 1 1 1 ] [ . . . . 3 ]
42: [ . 1 2 3 4 ] [ . 1 2 3 4 ] [ . 1 1 1 1 ] [ . . . . 4 ]
\end{verbatim}
\caption{\label{fig:catalan-step-rgs}
RGS corresponding to certain lattice paths (see text)
in lexicographic, co-lexicographic, subset-lexrev, and subset-lex order.
}
\end{figure}
One use of the subset-lexrev order is the development of algorithms for the
generation of orderings for objects where the subset-lex order exhibits no
simple structure.
As an example we consider the restricted growth strings corresponding
to lattice paths from $(0,0)$ to $(n,n)$ with steps
$(+1,0)$ and $(+1,+1)$ that do not go below the diagonal.
These RGS are words $a_0,\,a_1, \ldots,\, a_{n-1}$
such that $a_0=0$, $a_j\leq{}j$, and $a_{j+1}\geq{}a_j$
(the final $a_n=n$ is omitted).
These RGS are counted by the Catalan numbers,
see sequence \jjseqref{A000108} in \cite{oeis}.
Figure \ref{fig:catalan-step-rgs} shows the all such RGS of length 5
in lexicographic, co-lexicographic, subset-lexrev, and subset-lex
order. The listing in subset-lex order does not seem to have
any apparently useful features, while the listing in subset-lexrev
order suggests the following method of generation.
We now assume a sentinel $a_n=+\infty$ at the end of the word.
\begin{alg}[Next-Catalan-Step-RGS
Compute the successor in subset-lexrev order.
\end{alg}
\begin{enumerate}
\item Let $a_t$ be the first nonzero digit.
\item If $a_t < a_{t+1}$ and $a_t<t$, increment $a_t$ and return.
\item If $t\geq{}2$, set $a_{t-1}=1$ and return.
\item Remove all leading ones (setting them to zero),
let $j$ be the position of the first digit $a_j\neq{}1$.
\item If the word is all-zero (that is, $j=n$), stop.
\item Set $a_j=a_j-1$ and set $a_{j-1}=1$.
\end{enumerate}
The implementation correctly handles all cases $n\geq{}0$.
{\codesize
\begin{listing}{1}
// FILE: src/comb/catalan-step-rgs-subset-lexrev.h
class catalan_step_rgs_subset_lexrev
{
ulong *a_; // RGS
ulong n2_; // aux: min(n,2).
ulong tr_; // aux: track we are looking at
ulong n_; // length of RGS
\end{listing}
}%
The routine for the successor returns the
position of the rightmost change.
{\codesize
\begin{listing}{1}
ulong next()
{
const ulong a0 = a_[tr_];
if ( a0 < a_[tr_+1] ) // may read sentinel
{
if ( a0 < tr_ ) // can increment
{
a_[tr_] = a0 + 1;
return tr_;
}
}
if ( tr_ != 1 ) // can move left and increment (from 0 to 1)
{
--tr_;
a_[tr_] = 1;
return tr_;
}
// remove ones:
ulong j = tr_;
do { a_[j] = 0; } while ( a_[++j] == 1 );
if ( j==n2_ ) return 0; // current was last
// decrement first value != 1:
ulong aj = a_[j] - 1;
a_[j] = aj;
// move left and restore the 1:
tr_ = j - 1;
a_[tr_] = 1;
return j; // rightmost change at j==tr+1
}
\end{listing}
}%
An update takes about 10.5 cycles,
whereas the updates for lexicographic and co-lexicographic
order respectively take 9 and 8 cycles.
\begin{figure
{\jjbls
\begin{verbatim}
0: [ 1 1 1 1 1 1 1 1 1 1 ] [ 1 1 1 1 1 1 1 1 1 1 ]
1: [ 1 1 1 1 1 1 1 1 2 . ] [ 2 1 1 1 1 1 1 1 1 . ]
2: [ 1 1 1 1 1 1 1 3 . . ] [ 2 2 1 1 1 1 1 1 . . ]
3: [ 1 1 1 1 1 1 2 2 . . ] [ 3 1 1 1 1 1 1 1 . . ]
4: [ 1 1 1 1 1 1 4 . . . ] [ 2 2 2 1 1 1 1 . . . ]
5: [ 1 1 1 1 1 2 3 . . . ] [ 3 2 1 1 1 1 1 . . . ]
6: [ 1 1 1 1 2 2 2 . . . ] [ 4 1 1 1 1 1 1 . . . ]
7: [ 1 1 1 1 1 5 . . . . ] [ 2 2 2 2 1 1 . . . . ]
8: [ 1 1 1 1 2 4 . . . . ] [ 3 2 2 1 1 1 . . . . ]
9: [ 1 1 1 1 3 3 . . . . ] [ 3 3 1 1 1 1 . . . . ]
10: [ 1 1 1 2 2 3 . . . . ] [ 4 2 1 1 1 1 . . . . ]
11: [ 1 1 2 2 2 2 . . . . ] [ 5 1 1 1 1 1 . . . . ]
12: [ 1 1 1 1 6 . . . . . ] [ 2 2 2 2 2 . . . . . ]
13: [ 1 1 1 2 5 . . . . . ] [ 3 2 2 2 1 . . . . . ]
14: [ 1 1 1 3 4 . . . . . ] [ 3 3 2 1 1 . . . . . ]
15: [ 1 1 2 2 4 . . . . . ] [ 4 2 2 1 1 . . . . . ]
16: [ 1 1 2 3 3 . . . . . ] [ 4 3 1 1 1 . . . . . ]
17: [ 1 2 2 2 3 . . . . . ] [ 5 2 1 1 1 . . . . . ]
18: [ 2 2 2 2 2 . . . . . ] [ 6 1 1 1 1 . . . . . ]
19: [ 1 1 1 7 . . . . . . ] [ 3 3 2 2 . . . . . . ]
20: [ 1 1 2 6 . . . . . . ] [ 4 2 2 2 . . . . . . ]
21: [ 1 1 3 5 . . . . . . ] [ 3 3 3 1 . . . . . . ]
22: [ 1 2 2 5 . . . . . . ] [ 4 3 2 1 . . . . . . ]
23: [ 1 1 4 4 . . . . . . ] [ 5 2 2 1 . . . . . . ]
24: [ 1 2 3 4 . . . . . . ] [ 4 4 1 1 . . . . . . ]
25: [ 2 2 2 4 . . . . . . ] [ 5 3 1 1 . . . . . . ]
26: [ 1 3 3 3 . . . . . . ] [ 6 2 1 1 . . . . . . ]
27: [ 2 2 3 3 . . . . . . ] [ 7 1 1 1 . . . . . . ]
28: [ 1 1 8 . . . . . . . ] [ 4 3 3 . . . . . . . ]
29: [ 1 2 7 . . . . . . . ] [ 4 4 2 . . . . . . . ]
30: [ 1 3 6 . . . . . . . ] [ 5 3 2 . . . . . . . ]
31: [ 2 2 6 . . . . . . . ] [ 6 2 2 . . . . . . . ]
32: [ 1 4 5 . . . . . . . ] [ 5 4 1 . . . . . . . ]
33: [ 2 3 5 . . . . . . . ] [ 6 3 1 . . . . . . . ]
34: [ 2 4 4 . . . . . . . ] [ 7 2 1 . . . . . . . ]
35: [ 3 3 4 . . . . . . . ] [ 8 1 1 . . . . . . . ]
36: [ 1 9 . . . . . . . . ] [ 5 5 . . . . . . . . ]
37: [ 2 8 . . . . . . . . ] [ 6 4 . . . . . . . . ]
38: [ 3 7 . . . . . . . . ] [ 7 3 . . . . . . . . ]
39: [ 4 6 . . . . . . . . ] [ 8 2 . . . . . . . . ]
40: [ 5 5 . . . . . . . . ] [ 9 1 . . . . . . . . ]
41: [10 . . . . . . . . . ] [10 . . . . . . . . . ]
\end{verbatim}
\caption{\label{fig:part-asc-subset-lexrev}
The partitions of 10 in subset-lexrev order,
as weakly increasing lists of parts (left) and
as weakly decreasing lists of parts (right).
}
\end{figure}
We mention that the subset-lexrev order coincides
for certain objects with well-known orderings.
For example, for partitions as lists of parts
as shown in figure \ref{fig:part-asc-subset-lexrev},
both orderings are by falling length of partitions as
major order,
the minor order is lexicographic in case of weakly increasing parts
and reverse co-lexicographic for weakly decreasing parts.
Similar observations can be made (for example) for compositions into a fixed
number of parts, for both subset-lex and subset-lexrev order.
\section{SL-Gray order for binary words}
\begin{figure
{\jjbls
\begin{verbatim}
[ . . . . ] [ . . . . ] [ . . . . ] [ . . . . ]
[ . . . 1 ] [ . . . 1 ] [ . . . 1 ] [ . . . 1 ]
[ . . 1 . ] [ . . 1 . ] [ . . 1 . ] v [ . . 1 1 ]
[ . . 1 1 ] [ . . 1 1 ] [ . . 1 1 ] ^ [ . . 1 . ]
[ . 1 . . ] [ . 1 . . ] v [ . 1 1 1 ] v [ . 1 1 . ]
[ . 1 . 1 ] [ . 1 . 1 ] [ . 1 1 . ] ^ [ . 1 1 1 ]
[ . 1 1 . ] [ . 1 1 . ] [ . 1 . 1 ] [ . 1 . 1 ]
[ . 1 1 1 ] [ . 1 1 1 ] ^ [ . 1 . . ] [ . 1 . . ]
[ 1 . . . ] v [ 1 1 1 1 ] v [ 1 1 . . ] v [ 1 1 . . ]
[ 1 . . 1 ] [ 1 1 1 . ] [ 1 1 . 1 ] ^ [ 1 1 . 1 ]
[ 1 . 1 . ] [ 1 1 . 1 ] [ 1 1 1 . ] v [ 1 1 1 1 ]
[ 1 . 1 1 ] [ 1 1 . . ] ^ [ 1 1 1 1 ] ^ [ 1 1 1 . ]
[ 1 1 . . ] [ 1 . 1 1 ] [ 1 . 1 1 ] v [ 1 . 1 . ]
[ 1 1 . 1 ] [ 1 . 1 . ] [ 1 . 1 . ] ^ [ 1 . 1 1 ]
[ 1 1 1 . ] [ 1 . . 1 ] [ 1 . . 1 ] [ 1 . . 1 ]
[ 1 1 1 1 ] ^ [ 1 . . . ] [ 1 . . . ] [ 1 . . . ]
[ . . . . . ] [ . . . . . ] [ . . . . . ] [ . . . . . ]
[ 1 . . . . ] [ 1 . . . . ] [ 1 . . . . ] [ 1 . . . . ]
[ 1 1 . . . ] [ 1 1 . . . ] [ 1 1 . . . ] [ 1 1 . . . ]
[ 1 1 1 . . ] [ 1 1 1 . . ] [ 1 1 1 . . ] [ 1 1 1 . . ]
[ 1 1 1 1 . ] [ 1 1 1 1 . ] [ 1 1 1 1 . ] [ 1 1 1 1 . ]
[ 1 1 1 1 1 ] [ 1 1 1 1 1 ] [ 1 1 1 1 1 ] [ 1 1 1 1 1 ]
[ 1 1 1 . 1 ] [ 1 1 1 . 1 ] [ 1 1 1 . 1 ] [ 1 1 1 . 1 ]
[ 1 1 . 1 . ] [ 1 1 . 1 . ] [ 1 1 . 1 . ] v [ 1 1 . . 1 ]
[ 1 1 . 1 1 ] [ 1 1 . 1 1 ] [ 1 1 . 1 1 ] [ 1 1 . 1 1 ]
[ 1 1 . . 1 ] [ 1 1 . . 1 ] [ 1 1 . . 1 ] ^ [ 1 1 . 1 . ]
[ 1 . 1 . . ] [ 1 . 1 . . ] v [ 1 . . . 1 ] v [ 1 . . 1 . ]
[ 1 . 1 1 . ] [ 1 . 1 1 . ] [ 1 . . 1 1 ] [ 1 . . 1 1 ]
[ 1 . 1 1 1 ] [ 1 . 1 1 1 ] [ 1 . . 1 . ] ^ [ 1 . . . 1 ]
[ 1 . 1 . 1 ] [ 1 . 1 . 1 ] [ 1 . 1 . 1 ] [ 1 . 1 . 1 ]
[ 1 . . 1 . ] [ 1 . . 1 . ] [ 1 . 1 1 1 ] [ 1 . 1 1 1 ]
[ 1 . . 1 1 ] [ 1 . . 1 1 ] [ 1 . 1 1 . ] [ 1 . 1 1 . ]
[ 1 . . . 1 ] [ 1 . . . 1 ] ^ [ 1 . 1 . . ] [ 1 . 1 . . ]
[ . 1 . . . ] v [ . . . . 1 ] v [ . . 1 . . ] [ . . 1 . . ]
[ . 1 1 . . ] [ . . . 1 1 ] [ . . 1 1 . ] [ . . 1 1 . ]
[ . 1 1 1 . ] [ . . . 1 . ] [ . . 1 1 1 ] [ . . 1 1 1 ]
[ . 1 1 1 1 ] [ . . 1 . 1 ] [ . . 1 . 1 ] [ . . 1 . 1 ]
[ . 1 1 . 1 ] [ . . 1 1 1 ] [ . . . 1 . ] v [ . . . . 1 ]
[ . 1 . 1 . ] [ . . 1 1 . ] [ . . . 1 1 ] [ . . . 1 1 ]
[ . 1 . 1 1 ] [ . . 1 . . ] ^ [ . . . . 1 ] ^ [ . . . 1 . ]
[ . 1 . . 1 ] [ . 1 . . 1 ] [ . 1 . . 1 ] v [ . 1 . 1 . ]
[ . . 1 . . ] [ . 1 . 1 1 ] [ . 1 . 1 1 ] [ . 1 . 1 1 ]
[ . . 1 1 . ] [ . 1 . 1 . ] [ . 1 . 1 . ] ^ [ . 1 . . 1 ]
[ . . 1 1 1 ] [ . 1 1 . 1 ] [ . 1 1 . 1 ] [ . 1 1 . 1 ]
[ . . 1 . 1 ] [ . 1 1 1 1 ] [ . 1 1 1 1 ] [ . 1 1 1 1 ]
[ . . . 1 . ] [ . 1 1 1 . ] [ . 1 1 1 . ] [ . 1 1 1 . ]
[ . . . 1 1 ] [ . 1 1 . . ] [ . 1 1 . . ] [ . 1 1 . . ]
[ . . . . 1 ] ^ [ . 1 . . . ] [ . 1 . . . ] [ . 1 . . . ]
\end{verbatim}
\caption{\label{fig:gray-rec}
Construction of Gray codes by reversing sublists:
for the binary reflected Gray code (top)
and for the binary SL-Gray order (bottom).
The symbols \texttt{v} and \texttt{\^{}} respectively mark
begin and end of the reversed sublists.
}
\end{figure}
A well-known construction for the binary reflected Gray code proceeds
by successively reversing ranges having identical prefixes that
end in a one, using increasingly longer prefixes,
see top of figure \ref{fig:gray-rec}.
The same construction, now using prefixes ending in a zero,
gives a Gray code for subset-lex order, see
bottom of figure \ref{fig:gray-rec}.
We will call this ordering the \jjterm{SL-Gray order}.
For the computation of the successor we keep
a variable $t$ for the current track and a variable
$d$ indicating the direction in which the track
will be moved if necessary.
Initially $t=0$ and $d=+1$.
\begin{samepage}
\begin{alg}[Next-SL-Gray
Compute the successor in SL-Gray order.
\end{alg}
\begin{enumerate}
\item If $d=+1$, do the following (try to append trailing ones):
\begin{enumerate}
\item If $a_t=0$, set $a_t=1$, $t=t+1$, and return.
\item Otherwise, set $d=-1$ (change direction),
$t=n-1$ (move to rightmost track),
$j=n-2$, $a_j=1-a_j$, and return.
\end{enumerate}
\item Otherwise ($d=-1$), do the following (try to remove trailing ones):
\begin{enumerate}
\item If $a_{t-1}=1$, set $a_t=0$, $t=t-1$, and return.
\item Otherwise, set $d=+1$ (change direction),
$j=t-2$, $a_j=1-a_j$, $t=t+1$ (move right), and return.
\end{enumerate}
\end{enumerate}
\end{samepage}
The algorithm is loopless.
In the implementation, the variables
\texttt{tr} and \texttt{dt} respectively
correspond to $t$ and $d$ in the algorithm.
Two sentinels $a_{-1}=a_n=+1$ are used.
{\codesize
\begin{listing}{1}
// FILE: src/comb/binary-sl-gray.h
class binary_sl_gray
{
ulong n_; // number of digits
ulong tr_; // aux: current track (0 <= tr <= n)
ulong dt_; // aux: direction in which track tries to move
ulong *a_; // digits
void first()
{
for (ulong k=0; k<n_; ++k) a_[k] = 0;
tr_ = 0;
dt_ = +1;
}
explicit binary_sl_gray(ulong n)
{
n_ = n;
a_ = new ulong[n_+2]; // sentinels at both ends
a_[n_+1] = +1; // != 0
a_[0] = +1;
++a_; // nota bene
first();
}
void first()
{
for (ulong k=0; k<n_; ++k) a_[k] = 0;
tr_ = 0;
dt_ = +1;
j_ = ( n_>=2 ? 1 : 0); // wrt. last word
dm_ = -1; // wrt. last word
}
\end{listing}
}%
The routines for successor and predecessor handle
all cases $n\geq{}0$ correctly.
{\codesize
\begin{listing}{1}
bool next()
{
if ( dt_ == +1 ) // try to append trailing ones
{
if ( a_[tr_] == 0 ) // can append // may read sentinel a[n]
{
a_[tr_] = 1;
++tr_;
}
else
{
dt_ = -1UL; // change direction
tr_ = n_ - 1;
ulong j_ = tr_ - 1;
// Is current last (only for n_ <= 1)?
if ( j_ > n_ ) return false;
a_[j_] = 1 - a_[j_];
}
}
else // dt_ == -1 // try to remove trailing ones
{
if ( a_[tr_-1] != 0 ) // can remove // tr - 1 >= 0
{
a_[tr_] = 0;
--tr_;
}
else
{
dt_ = +1; // change direction
ulong j_ = tr_ - 2;
if ( (long)j_ < 0 ) return false;
a_[j_] = 1 - a_[j_];
++tr_;
}
}
return true;
}
\end{listing}
}%
The routine to compute the predecessor is obtained
by (essentially) negating the direction the
track tries to move (variable \texttt{dt}):
{\codesize
\begin{listing}{1}
bool prev()
{
if ( dt_ != +1 ) // dt==-1 // try to append trailing ones
{
if ( a_[tr_+1] == 0 ) // can append // may read sentinel a[n]
{
a_[tr_+1] = 1;
++tr_;
}
else
{
dt_ = +1; // change direction
tr_ = n_ - 1;
ulong j_ = tr_ - 1;
a_[j_] = 1 - a_[j_];
++tr_;
}
}
else // dt_ == +1 // try to remove trailing ones
{
if ( tr_ == 0 )
{
if ( a_[0]==0 ) return false; // (only for n_ <= 1)
a_[0] = 0;
return true;
}
// tr - 1 >= -1 (can read low sentinel)
if ( a_[tr_-2] != 0 ) // can remove
{
a_[tr_-1] = 0;
--tr_;
}
else
{
dt_ = -1UL; // change direction
ulong j_ = tr_ - 3;
a_[j_] = 1 - a_[j_];
--tr_;
}
}
return true;
}
\end{listing}
}%
One update in either direction takes about 7.5 cycles.
An implementation for the generation of the
SL-Gray order in a binary word is
given in \fileref{src/bits/bit-sl-gray.h}.
The ranking algorithm is easily obtained by
observing that if the highest bit is not set
(and the word is nonzero),
the order for the remaining word is reflected,
as shown in figure \ref{fig:gray-rec}.
\begin{alg}[Binary-SL-Gray-Rank
Recursive routine $F(k,\,n)$ for the computation of the rank of
an $n$-bit word $k$ in binary SL-Gray order.
\end{alg}
\begin{enumerate}
\item If $k=0$, return 0.
\item Set $w=k$ and unset the highest bit in $w$.
\item If the highest bit of $k$ (at position $n-1$) is set,
return $1 + F(w,\, n-1)$.
\item Otherwise return $2^n - F(w,\, n-1)$.
\end{enumerate}
In the following implementation
the variable \texttt{ldn} must give
the number of bits in the SL-Gray code.
{\codesize
\begin{listing}{1}
// FILE: src/bits/bin-to-sl-gray.h
ulong sl_gray_to_bin(ulong k, ulong ldn)
{
if ( k==0 ) return 0;
ulong b = 1UL << (ldn-1); // mask for bit at end
ulong h = k & b; // bit at end
k ^= h;
ulong z = sl_gray_to_bin( k, ldn-1 ); // recursion
if ( h==0 ) return (b<<1) - z;
else return 1 + z;
}
\end{listing}
}%
A routine for the conversion of a binary
word into the corresponding word in SL-Gray order
(unranking algorithm) is
{\codesize
\begin{listing}{1}
ulong bin_to_sl_gray(ulong k, ulong ldn)
{
if ( ldn==0 ) return 0;
ulong b = 1UL << (ldn-1); // highest bit
ulong m = (b<<1) - 1; // mask for reversing direction
ulong z = b; // Gray code
k -= 1; // move all-zero word to begin
while ( b != 0 )
{
const ulong h = k & b; // bit under consideration
z ^= h; // with one, switch bit in Gray code
if ( !h ) k ^= m; // reverse direction with zero
k += 1; // SL-Gray
b >>= 1; // next lower bit
m >>= 1; // next smaller mask
}
return z;
}
\end{listing}
}%
\begin{figure
{\jjbls
\begin{verbatim}
0121030121014101210301210125210121030121014101210301210123
......1...1.1...1...1...1.1...1.....1...1.1...1...1...1.1...1...
.....1.1.1...1.1.1.1.1.1...1.1.1...1.1.1...1.1.1.1.1.1...1.1.1..
....1...1.....1.......1.....1...1.1...1.....1.......1.....1...1.
...1.......1.............1...............1.............1.......1
..1...............1.............................1...............
.1...............................1..............................
......1111..1111....1111..1111......1111..1111....1111..1111....
.....11..1111..11..11..1111..11....11..1111..11..11..1111..11...
....1111......11111111......1111..1111......11111111......1111..
...11111111..............1111111111111111..............11111111.
..1111111111111111..............................1111111111111111
.11111111111111111111111111111111...............................
\end{verbatim}
\caption{\label{fig:bin-sl-gray-delta}
The delta sequence (top) of the binary SL-Gray order,
starting after the initial slope
and indexing positions from the end of the words.
}
\end{figure}
The delta sequence (sequence of positions of changes),
starting after the initial slope and indexing positions from the end,
is shown in figure \ref{fig:bin-sl-gray-delta},
this is sequence \jjseqref{A217262} in \cite{oeis}.
It can be obtained from the ruler function (sequence \jjseqref{A007814})
by replacing $0$ by $w=01210$,
$1$ by $3$, $2$ by $141$, $3$ by $12521$, $4$ by $1236321$, \ldots,
$n$ by $123\ldots{}(n-1)(n+2)(n-1)\ldots{}321$:
{\jjbls
\begin{verbatim}
0 1 0 2 0 1 0 3 0 1 0 2 0 1 0 4 0 ... (ruler function)
w 3 w 141 w 3 w 12521 w 3 w 141 w 3 w 1236321 w ... =
01210 3 01210 141 01210 3 01210 12521 01210 3 01210 141 01210 3 01210 ...
\end{verbatim}
}%
It is easy to see that (for $n\geq{}4$) the
SL-Gray order for the $2^n$ words of length $n$
has $2^{n-2}-2$ successive transitions that are 3-close
(have a distance of 3), the rest are 1-close
(adjacent changes).
A Gray code where the maximal distance between successive
transitions is $k$ is called $k$-close.
In \cite[p.400]{fxtbook} it has been observed
that 1-close Gray codes exists for $n\leq{}6$
but not for $n=7$ (and it appears unlikely that
for any $n\geq{}8$ such a Gray code exists).
Does a 2-close Gray code exist?
We remark that the SL-Gray order is 2-close
if the radices for all digits are even and $\geq{}4$.
\section{SL-Gray order for mixed radix words}
\begin{figure
\vspace*{-8mm
{\jjbls
\begin{verbatim}
0: [ . . . . ] [ + + + + ] { }
1: [ 1 . . . ] [ + + + + ] { 0 }
2: [ 1 1 . . ] [ - + + + ] { 0, 1 }
3: [ 1 2 . . ] [ - + + + ] { 0, 1, 1 }
4: [ 1 2 1 . ] [ - - + + ] { 0, 1, 1, 2 }
5: [ 1 2 2 . ] [ - - + + ] { 0, 1, 1, 2, 2 }
6: [ 1 2 2 1 ] [ - - - + ] { 0, 1, 1, 2, 2, 3 }
7: [ 1 2 2 2 ] [ - - - + ] { 0, 1, 1, 2, 2, 3, 3 }
8: [ 1 2 2 3 ] [ - - - + ] { 0, 1, 1, 2, 2, 3, 3, 3 }
9: [ 1 2 1 3 ] [ - - - - ] { 0, 1, 1, 2, 3, 3, 3 }
10: [ 1 2 1 2 ] [ - - - - ] { 0, 1, 1, 2, 3, 3 }
11: [ 1 2 1 1 ] [ - - - - ] { 0, 1, 1, 2, 3 }
12: [ 1 2 . 1 ] [ - - - + ] { 0, 1, 1, 3 }
13: [ 1 2 . 2 ] [ - - - + ] { 0, 1, 1, 3, 3 }
14: [ 1 2 . 3 ] [ - - - + ] { 0, 1, 1, 3, 3, 3 }
15: [ 1 1 . 3 ] [ - - + - ] { 0, 1, 3, 3, 3 }
16: [ 1 1 . 2 ] [ - - + - ] { 0, 1, 3, 3 }
17: [ 1 1 . 1 ] [ - - + - ] { 0, 1, 3 }
18: [ 1 1 1 1 ] [ - - + + ] { 0, 1, 2, 3 }
19: [ 1 1 1 2 ] [ - - + + ] { 0, 1, 2, 3, 3 }
20: [ 1 1 1 3 ] [ - - + + ] { 0, 1, 2, 3, 3, 3 }
21: [ 1 1 2 3 ] [ - - + - ] { 0, 1, 2, 2, 3, 3, 3 }
22: [ 1 1 2 2 ] [ - - + - ] { 0, 1, 2, 2, 3, 3 }
23: [ 1 1 2 1 ] [ - - + - ] { 0, 1, 2, 2, 3 }
24: [ 1 1 2 . ] [ - - - + ] { 0, 1, 2, 2 }
25: [ 1 1 1 . ] [ - - - + ] { 0, 1, 2 }
26: [ 1 . 1 . ] [ - - + + ] { 0, 2 }
27: [ 1 . 2 . ] [ - - + + ] { 0, 2, 2 }
28: [ 1 . 2 1 ] [ - - - + ] { 0, 2, 2, 3 }
29: [ 1 . 2 2 ] [ - - - + ] { 0, 2, 2, 3, 3 }
30: [ 1 . 2 3 ] [ - - - + ] { 0, 2, 2, 3, 3, 3 }
31: [ 1 . 1 3 ] [ - - - - ] { 0, 2, 3, 3, 3 }
32: [ 1 . 1 2 ] [ - - - - ] { 0, 2, 3, 3 }
33: [ 1 . 1 1 ] [ - - - - ] { 0, 2, 3 }
34: [ 1 . . 1 ] [ - - - + ] { 0, 3 }
35: [ 1 . . 2 ] [ - - - + ] { 0, 3, 3 }
36: [ 1 . . 3 ] [ - - - + ] { 0, 3, 3, 3 }
37: [ . . . 3 ] [ - + + - ] { 3, 3, 3 }
38: [ . . . 2 ] [ - + + - ] { 3, 3 }
39: [ . . . 1 ] [ - + + - ] { 3 }
40: [ . . 1 1 ] [ - + + + ] { 2, 3 }
41: [ . . 1 2 ] [ - + + + ] { 2, 3, 3 }
42: [ . . 1 3 ] [ - + + + ] { 2, 3, 3, 3 }
43: [ . . 2 3 ] [ - + + - ] { 2, 2, 3, 3, 3 }
44: [ . . 2 2 ] [ - + + - ] { 2, 2, 3, 3 }
45: [ . . 2 1 ] [ - + + - ] { 2, 2, 3 }
46: [ . . 2 . ] [ - + - + ] { 2, 2 }
47: [ . . 1 . ] [ - + - + ] { 2 }
48: [ . 1 1 . ] [ - + + + ] { 1, 2 }
49: [ . 1 2 . ] [ - + + + ] { 1, 2, 2 }
50: [ . 1 2 1 ] [ - + - + ] { 1, 2, 2, 3 }
51: [ . 1 2 2 ] [ - + - + ] { 1, 2, 2, 3, 3 }
52: [ . 1 2 3 ] [ - + - + ] { 1, 2, 2, 3, 3, 3 }
53: [ . 1 1 3 ] [ - + - - ] { 1, 2, 3, 3, 3 }
54: [ . 1 1 2 ] [ - + - - ] { 1, 2, 3, 3 }
55: [ . 1 1 1 ] [ - + - - ] { 1, 2, 3 }
56: [ . 1 . 1 ] [ - + - + ] { 1, 3 }
57: [ . 1 . 2 ] [ - + - + ] { 1, 3, 3 }
58: [ . 1 . 3 ] [ - + - + ] { 1, 3, 3, 3 }
59: [ . 2 . 3 ] [ - + + - ] { 1, 1, 3, 3, 3 }
60: [ . 2 . 2 ] [ - + + - ] { 1, 1, 3, 3 }
61: [ . 2 . 1 ] [ - + + - ] { 1, 1, 3 }
62: [ . 2 1 1 ] [ - + + + ] { 1, 1, 2, 3 }
63: [ . 2 1 2 ] [ - + + + ] { 1, 1, 2, 3, 3 }
64: [ . 2 1 3 ] [ - + + + ] { 1, 1, 2, 3, 3, 3 }
65: [ . 2 2 3 ] [ - + + - ] { 1, 1, 2, 2, 3, 3, 3 }
66: [ . 2 2 2 ] [ - + + - ] { 1, 1, 2, 2, 3, 3 }
67: [ . 2 2 1 ] [ - + + - ] { 1, 1, 2, 2, 3 }
68: [ . 2 2 . ] [ - + - + ] { 1, 1, 2, 2 }
69: [ . 2 1 . ] [ - + - + ] { 1, 1, 2 }
70: [ . 2 . . ] [ - - + + ] { 1, 1 }
71: [ . 1 . . ] [ - - + + ] { 1 }
\end{verbatim}
\caption{\label{fig:mixedradix-sl-gray}
Mixed radix words (left) and corresponding subsets (right)
of the multiset $\{0^1, 1^2, 2^2, 3^3 \}$ in SL-Gray order,
and the array of directions used in the computation (middle).
}
\end{figure}
The generalization of the SL-Gray order to mixed radix words is obtained by
successive reversion of the sublists with identical prefix for all prefixes
ending in an even digit.
For the computation of the successor we keep
a variable $t$ for the current track and
variables $d_j=\pm{}1$ (an array) for the direction
the digit $a_j$ is currently moving in
(whether its is incremented or decremented).
We further use $m_j$ for the maximal value
the digit $a_j$ can have (the \eqq{nines}).
Initially all digits $a_k$ are zero,
all directions $d_k$ are $+1$, and $t=0$.
\begin{alg}[Next-SL-Gray
Compute the successor SL-Gray order.
\end{alg}
\begin{enumerate}
\item Set $j=t$ and $b = a_j + d_j$.
\item If $b\neq{}0$ and $b\leq{}m_j$, set $a_j=b$ and return (easy case).
\item Set $d_j=-d_j$ (change direction for digit $j$).
\item If $d_j = +1$ and $a_{j+1}=0$,
set $a_{j+1} = +1$, $t = j+1$ (move track right), and return.
\item If $d_j=-1$ and $a_{j-1} = m_{j-1}$,
set $a_j=0$, $d_{j-1} = -1$, $t=j-1$ (move track left), and return.
\item Find the position $p$ of the nearest digit $a_k$ to the left
such that $a_k+d_k$ is in the range $0,1,\ldots,m_k$ (a valid digit).
In the process, change $d_k$ for all $k$ where $p < k < j$.
\item Set $a_p = a_p + d_p$ (change digit, keep track).
\end{enumerate}
The details for termination have been omitted,
this is handled with the help of sentinels
in the implementation.
{\codesize
\begin{listing}{1}
// FILE: src/comb/mixedradix-sl-gray.h
class mixedradix_sl_gray
{
ulong n_; // number of digits
// (n kinds of elements in multiset, n>=1)
ulong tr_; // aux: current track
ulong *a_; // digits of mixed radix number
// (multiplicity of kind k in subset).
ulong *d_; // directions (either +1 or -1)
ulong *m1_; // nines (radix minus one) for each digit
// (multiplicity of kind k in set).
\end{listing}
}%
Sentinels are used in all arrays at index $-1$ and at index $n$
to handle termination.
{\codesize
\begin{listing}{1}
mixedradix_sl_gray(ulong n, ulong mm, const ulong *m=0)
// Must have n>=1
{
n_ = n;
a_ = new ulong[n_+2]; // all with sentinels at both ends
d_ = new ulong[n_+2];
m1_ = new ulong[n_+2];
// sentinels on the right:
a_[n_+1] = +1; // != 0
m1_[n_+1] = +1; // same as a_[n+1]
d_[n_+1] = +1; // positive
// sentinels on the left:
a_[0] = +2; // >= +2
m1_[0] = +2; // same as a_[0]
d_[0] = 0; // zero
++a_; ++d_; ++m1_; // nota bene
mixedradix_init(n_, mm, m, m1_);
first();
}
void first()
{
for (ulong k=0; k<n_; ++k) a_[k] = 0;
for (ulong k=0; k<n_; ++k) d_[k] = +1;
tr_ = 0;
}
\end{listing}
}%
The method for the successor handles all cases $n\geq{}0$ correctly.
The variable \texttt{a1} in the implementation corresponds to $b$ in the algorithm.
{\codesize
\begin{listing}{1}
bool next()
{
ulong j = tr_;
const ulong dj = d_[j];
const ulong a1 = a_[j] + dj; // a[j] +- 1
if ( (a1 != 0) && (a1 <= m1_[j]) ) // easy case
{
a_[j] = a1;
return true;
}
d_[j] = -dj; // change direction
if ( dj == +1 ) // so a_[j] == m1_[j] == nine
{
// Try to move track right with a[j] == nine:
const ulong j1 = j + 1;
if ( a_[j1] == 0 ) // can read high sentinel
{
a_[j1] = +1;
tr_ = j1;
return true;
}
}
else // here dj == -1, so a_[j] == 1
{
if ( (long)j <= 0 ) return false; // current is last
// Try to move track left with a[j] == 1:
const ulong j1 = j - 1;
if ( a_[j1] == m1_[j1] ) // can read low sentinel when n_ == 1
{
a_[j] = 0;
d_[j1] = -1UL;
tr_ = j1;
return true;
}
}
// find first changeable track to the left:
--j;
while ( a_[j] + d_[j] > m1_[j] ) // may read low sentinels
{
d_[j] = -d_[j]; // change direction
--j;
}
if ( (long)j < 0 ) return false; // current is last
// Change digit left but keep track:
a_[j] += d_[j];
return true;
}
\end{listing}
}%
\section{A Gray code for compositions}
\begin{figure
{\jjbls
\begin{verbatim}
0: 111111 [ 7 ]
1: 11111. [ 6 1 ]
2: 1111.. [ 5 1 1 ] 0: 111..... [ 4 1 1 1 1 1 ]
3: 1111.1 [ 5 2 ] 1: 11..1... [ 3 1 2 1 1 1 ]
4: 111.11 [ 4 3 ] 2: 11....1. [ 3 1 1 1 2 1 ]
5: 111.1. [ 4 2 1 ] 3: 11.....1 [ 3 1 1 1 1 2 ]
6: 111... [ 4 1 1 1 ] 4: 11...1.. [ 3 1 1 2 1 1 ]
7: 111..1 [ 4 1 2 ] 5: 11.1.... [ 3 2 1 1 1 1 ]
8: 11..11 [ 3 1 3 ] 6: 1.11.... [ 2 3 1 1 1 1 ]
9: 11..1. [ 3 1 2 1 ] 7: 1.1.1... [ 2 2 2 1 1 1 ]
10: 11.... [ 3 1 1 1 1 ] 8: 1.1...1. [ 2 2 1 1 2 1 ]
11: 11...1 [ 3 1 1 2 ] 9: 1.1....1 [ 2 2 1 1 1 2 ]
12: 11.1.. [ 3 2 1 1 ] 10: 1.1..1.. [ 2 2 1 2 1 1 ]
13: 11.1.1 [ 3 2 2 ] 11: 1...11.. [ 2 1 1 3 1 1 ]
14: 11.11. [ 3 3 1 ] 12: 1...1.1. [ 2 1 1 2 2 1 ]
15: 11.111 [ 3 4 ] 13: 1...1..1 [ 2 1 1 2 1 2 ]
16: 1.1111 [ 2 5 ] 14: 1.....11 [ 2 1 1 1 1 3 ]
17: 1.111. [ 2 4 1 ] 15: 1....1.1 [ 2 1 1 1 2 2 ]
18: 1.11.. [ 2 3 1 1 ] 16: 1....11. [ 2 1 1 1 3 1 ]
19: 1.11.1 [ 2 3 2 ] 17: 1..1..1. [ 2 1 2 1 2 1 ]
20: 1.1.11 [ 2 2 3 ] 18: 1..1...1 [ 2 1 2 1 1 2 ]
21: 1.1.1. [ 2 2 2 1 ] 19: 1..1.1.. [ 2 1 2 2 1 1 ]
22: 1.1... [ 2 2 1 1 1 ] 20: 1..11... [ 2 1 3 1 1 1 ]
23: 1.1..1 [ 2 2 1 2 ] 21: ..111... [ 1 1 4 1 1 1 ]
24: 1...11 [ 2 1 1 3 ] 22: ..11..1. [ 1 1 3 1 2 1 ]
25: 1...1. [ 2 1 1 2 1 ] 23: ..11...1 [ 1 1 3 1 1 2 ]
26: 1..... [ 2 1 1 1 1 1 ] 24: ..11.1.. [ 1 1 3 2 1 1 ]
27: 1....1 [ 2 1 1 1 2 ] 25: ..1.11.. [ 1 1 2 3 1 1 ]
28: 1..1.. [ 2 1 2 1 1 ] 26: ..1.1.1. [ 1 1 2 2 2 1 ]
29: 1..1.1 [ 2 1 2 2 ] 27: ..1.1..1 [ 1 1 2 2 1 2 ]
30: 1..11. [ 2 1 3 1 ] 28: ..1...11 [ 1 1 2 1 1 3 ]
31: 1..111 [ 2 1 4 ] 29: ..1..1.1 [ 1 1 2 1 2 2 ]
32: ..1111 [ 1 1 5 ] 30: ..1..11. [ 1 1 2 1 3 1 ]
33: ..111. [ 1 1 4 1 ] 31: ....111. [ 1 1 1 1 4 1 ]
34: ..11.. [ 1 1 3 1 1 ] 32: ....11.1 [ 1 1 1 1 3 2 ]
35: ..11.1 [ 1 1 3 2 ] 33: ....1.11 [ 1 1 1 1 2 3 ]
36: ..1.11 [ 1 1 2 3 ] 34: .....111 [ 1 1 1 1 1 4 ]
37: ..1.1. [ 1 1 2 2 1 ] 35: ...1..11 [ 1 1 1 2 1 3 ]
38: ..1... [ 1 1 2 1 1 1 ] 36: ...1.1.1 [ 1 1 1 2 2 2 ]
39: ..1..1 [ 1 1 2 1 2 ] 37: ...1.11. [ 1 1 1 2 3 1 ]
40: ....11 [ 1 1 1 1 3 ] 38: ...11.1. [ 1 1 1 3 2 1 ]
41: ....1. [ 1 1 1 1 2 1 ] 39: ...11..1 [ 1 1 1 3 1 2 ]
42: ...... [ 1 1 1 1 1 1 1 ] 40: ...111.. [ 1 1 1 4 1 1 ]
43: .....1 [ 1 1 1 1 1 2 ] 41: .1..11.. [ 1 2 1 3 1 1 ]
44: ...1.. [ 1 1 1 2 1 1 ] 42: .1..1.1. [ 1 2 1 2 2 1 ]
45: ...1.1 [ 1 1 1 2 2 ] 43: .1..1..1 [ 1 2 1 2 1 2 ]
46: ...11. [ 1 1 1 3 1 ] 44: .1....11 [ 1 2 1 1 1 3 ]
47: ...111 [ 1 1 1 4 ] 45: .1...1.1 [ 1 2 1 1 2 2 ]
48: .1..11 [ 1 2 1 3 ] 46: .1...11. [ 1 2 1 1 3 1 ]
49: .1..1. [ 1 2 1 2 1 ] 47: .1.1..1. [ 1 2 2 1 2 1 ]
50: .1.... [ 1 2 1 1 1 1 ] 48: .1.1...1 [ 1 2 2 1 1 2 ]
51: .1...1 [ 1 2 1 1 2 ] 49: .1.1.1.. [ 1 2 2 2 1 1 ]
52: .1.1.. [ 1 2 2 1 1 ] 50: .1.11... [ 1 2 3 1 1 1 ]
53: .1.1.1 [ 1 2 2 2 ] 51: .11.1... [ 1 3 2 1 1 1 ]
54: .1.11. [ 1 2 3 1 ] 52: .11...1. [ 1 3 1 1 2 1 ]
55: .1.111 [ 1 2 4 ] 53: .11....1 [ 1 3 1 1 1 2 ]
56: .11.11 [ 1 3 3 ] 54: .11..1.. [ 1 3 1 2 1 1 ]
57: .11.1. [ 1 3 2 1 ] 55: .111.... [ 1 4 1 1 1 1 ]
58: .11... [ 1 3 1 1 1 ]
59: .11..1 [ 1 3 1 2 ]
60: .111.. [ 1 4 1 1 ]
61: .111.1 [ 1 4 2 ]
62: .1111. [ 1 5 1 ]
63: .11111 [ 1 6 ]
\end{verbatim}
\caption{\label{fig:comp-gray}
The 64 compositions of 7 together with their binary encodings
as in figure \ref{fig:comp-nz} (left)
and the 56 compositions of 9 into exactly 6 parts (right).
}
\end{figure}
We now describe a Gray code for compositions where at most one unit is
moved in each step, all moves are 1-close or 2-close (these always
cross a part 1), and all moves are at the end of the current composition.
The compositions of 7 in this ordering are shown in
figure \ref{fig:comp-gray} (left), it can be obtained
from the lexicographic order by successively reversing the sublists
with identical prefixes that end in an odd part.
To keep matters simple for the iterative algorithm, we arrange the compositions
for $n$ odd such that the first part is only decreasing (starting with the
composition $[n]$) and otherwise such that the first part is increasing
(starting with the composition $[1,n-1]$). The compositions of 6 in this
ordering are obtained by dropping the first part 1 in all compositions of 7 in
the second half of the list, see figure \ref{fig:comp-gray}.
A recursive routine can be obtained by switching
between the recursive functions for lexicographic
and reversed lexicographic order whenever
an odd part has been written.
In the following a global array \texttt{a[]} is used.
{\codesize
\begin{listing}{1}
// FILE: src/comb/composition-nz-gray-rec-demo.cc
void F(ulong n, ulong m)
{
if ( n==0 ) { visit(m); return; }
for (ulong f=n; f!=0; --f) // first part decreasing
{
a[m] = f;
if ( 0 == (f & 1) ) F(n-f, m+1);
else B(n-f, m+1);
}
}
void B(ulong n, ulong m)
{
if ( n==0 ) { visit(m); return; }
for (ulong f=1; f<=n; ++f) // first part increasing
{
a[m] = f;
if ( 0 == (f & 1) ) B(n-f, m+1);
else F(n-f, m+1);
}
}
\end{listing}
}%
The initial call is \texttt{F(n, 0)} for $n$ odd
and \texttt{G(n,0)} for $n$ even.
For the iterative algorithm we use a function
$D(x)$ that shall return $+1$ if $x$ is odd and otherwise $-1$.
We will refer to the last three parts as respectively $x$, $y$, and $z$.
\begin{alg}[Next-Comp-Gray
Compute the successor of a composition.
\end{alg}
\begin{enumerate}
\item If $z=n-1$ and $n$ is odd, or $z=n$ and $n$ is even, stop.
\item If $z=1$, do the following:
\begin{enumerate}
\item If $D(y)=+1$, set $y=y+1$ and drop $z$; return.
\item Otherwise ($D(y)=-1$), set $y=y-1$ and append a part $1$; return.
\end{enumerate}
\item Otherwise ($z>1$) do the following:
\begin{enumerate}
\item If $z$ is odd, set $z=z-1$ and append a part $1$; return.
\item If $y>1$, set $y=y-D(y)$ and $z=z+D(y)$; return.
\item If $x>1$, set $x=x+D(x)$ and $z=z-D(x)$; return.
\item Otherwise ($x=1$) set $x=x+1$ and $z=z-1$.
\end{enumerate}
\end{enumerate}
The algorithm is loopless.
Note that the initialization is loopless as well.
The implementation handles all $n\geq$ correctly.
{\codesize
\begin{listing}{1}
// FILE: src/comb/composition-nz-gray2.h
class composition_nz_gray2
{
ulong *a_; // composition: a[1] + a[2] + ... + a[m] = n
ulong n_; // compositions of n
ulong m_; // current composition has m parts
ulong e_; // aux: detection of last composition
explicit composition_nz_gray2(ulong n)
{
n_ = n;
a_ = new ulong[n_+1+(n_==0)];
a_[0] = 0; // returned by last_part() when n==0
a_[1] = 0; // returned by first_part() when n==0
if ( n_ <= 1 ) e_ = n_;
else e_ = ( oddq(n_) ? n_ - 1 : n_ );
first();
}
void first()
{
if ( n_ <= 1 )
{
a_[1] = n_;
m_ = n_;
}
else
{
if ( oddq(n_) )
{
a_[1] = n_;
m_ = 1;
}
else
{
a_[1] = 1;
a_[2] = n_ - 1;
m_ = 2;
}
}
}
\end{listing}
}%
A few auxiliary routines are used.
{\codesize
\begin{listing}{1}
bool oddq(ulong x) const { return 0 != ( x & 1UL ); }
bool evenq(ulong x) const { return 0 == ( x & 1UL ); }
ulong par_to_dir_odd(ulong x) const
{
if ( oddq(x) ) return +1;
else return -1UL;
}
ulong par_to_dir_even(ulong x) const
{
if ( evenq(x) ) return +1;
else return -1UL;
}
\end{listing}
}%
We split the update into routines
for the cases $z==1$ and $z>1$ as
in the algorithm.
{\codesize
\begin{listing}{1}
ulong next_zeq1() // for Z == 1
{
const ulong y = a_[m_-1];
const ulong dy = par_to_dir_odd(y);
if ( dy == +1 )
{ // [*, Y, 1 ] --> [*, Y+1 ]
a_[m_-1] = y + 1;
m_ -= 1;
return m_;
}
else // dy == -1
{ // [*, Y, 1 ] --> [*, Y-1, 1, 1 ]
a_[m_-1] = y - 1;
m_ += 1;
a_[m_] = 1;
return m_;
}
}
ulong next_zgt1() // for Z > 1
{
const ulong z = a_[m_];
const ulong y = a_[m_-1];
if ( oddq(z) )
{ // [*, Z ] --> [*, Z-1, 1 ]
a_[m_] = z - 1;
m_ += 1;
a_[m_] = 1;
return m_;
}
if ( y != 1 ) // Y > 1
{ // [*, Y, Z ] --> [*, Y+-1, Z-+1 ]
const ulong dy = par_to_dir_even(y);
a_[m_-1] = y + dy;
a_[m_] = z - dy;
return m_;
}
else // Y == 1
{
const ulong x = a_[m_-2];
if ( x != 1 )
{ // [*, X, 1, Z ] --> [*, X+-1, 1, Z-+1 ]
const ulong dx = par_to_dir_odd(x);
a_[m_-2] = x + dx;
a_[m_] = z - dx;
return m_;
}
else // X == 1
{ // [*, X, 1, Z ] --> [*, X+1, 1, Z-1 ]
a_[m_-2] = x + 1;
a_[m_] = z - 1;
return m_;
}
}
}
\end{listing}
}%
The routine \texttt{next()}
returns number of parts in the new composition
and zero if there are no more compositions.
{\codesize
\begin{listing}{1}
ulong next()
{
ulong z = a_[m_];
if ( z == e_ ) return 0; // current is last
if ( z != 1 ) return next_zgt1();
else return next_zeq1();
}
\end{listing}
}%
One update takes about 11 cycles.
We remark that the sublists of compositions into a
fixed number of parts correspond to the combinations
in \jjterm{enup} order.
The list of binary words on the right of figure \ref{fig:comp-gray}
are those shown in \cite[fig.6.6-B (left), p.189]{fxtbook}.
\subsubsection*{Ranking and unranking}
Recursive routines for ranking and unranking
are obtained easily. The following implementations
are for the ordering with the first part decreasing.
{\codesize
\begin{listing}{1}
// FILE: src/comb/composition-nz-rank.cc
ulong
composition_nz_gray_rank(const ulong *x, ulong m, ulong n)
// Return rank r of composition x[], 0 <= r < 2**(n-1)
// where n is the sum of all parts.
{
if ( m <= 1 ) return 0;
ulong f = x[0]; // first part
ulong s = n - f; // remaining sum
ulong y = composition_nz_gray_rank(x+1, m-1, s);
if ( 0 == ( f & 1 ) ) // first part even
return ( 1UL << (s-1) ) + y;
else
return ( 1UL << s ) - 1 - y;
}
ulong
composition_nz_gray_unrank(ulong r, ulong *x, ulong n)
// Generate composition x[] of n with rank r.
// Return number of parts m of generated composition, 0 <= m <= n.
{
if ( r == 0 )
{
if ( n==0 ) return 0;
x[0] = n;
return 1;
}
ulong h = highest_one_idx(r);
ulong f = n - 1 - h; // first part
x[0] = f;
ulong b = 1UL << h; // highest one
r ^= b; // delete highest one
bool p = f & 1; // first part f odd ?
if ( p ) r = b - 1 - r; // change direction with odd f
return 1 + composition_nz_gray_unrank( r , x+1, n-f );
}
\end{listing}
}%
\section{Appendix: nonempty subsets with at most $k$ elements}\label{sect:ksubset}
\begin{figure
{\jjbls
\begin{verbatim}
1: 1..... { 0 } (continued)
2: 11.... { 0, 1 } 23: .1.11. { 1, 3, 4 }
3: 111... { 0, 1, 2 } 24: .1.1.1 { 1, 3, 5 }
4: 11.1.. { 0, 1, 3 } 25: .1..1. { 1, 4 }
5: 11..1. { 0, 1, 4 } 26: .1..11 { 1, 4, 5 }
6: 11...1 { 0, 1, 5 } 27: .1...1 { 1, 5 }
7: 1.1... { 0, 2 } 28: ..1... { 2 }
8: 1.11.. { 0, 2, 3 } 29: ..11.. { 2, 3 }
9: 1.1.1. { 0, 2, 4 } 30: ..111. { 2, 3, 4 }
10: 1.1..1 { 0, 2, 5 } 31: ..11.1 { 2, 3, 5 }
11: 1..1.. { 0, 3 } 32: ..1.1. { 2, 4 }
12: 1..11. { 0, 3, 4 } 33: ..1.11 { 2, 4, 5 }
13: 1..1.1 { 0, 3, 5 } 34: ..1..1 { 2, 5 }
14: 1...1. { 0, 4 } 35: ...1.. { 3 }
15: 1...11 { 0, 4, 5 } 36: ...11. { 3, 4 }
16: 1....1 { 0, 5 } 37: ...111 { 3, 4, 5 }
17: .1.... { 1 } 38: ...1.1 { 3, 5 }
18: .11... { 1, 2 } 39: ....1. { 4 }
19: .111.. { 1, 2, 3 } 40: ....11 { 4, 5 }
20: .11.1. { 1, 2, 4 } 41: .....1 { 5 }
21: .11..1 { 1, 2, 5 }
22: .1.1.. { 1, 3 }
\end{verbatim}
\caption{\label{fig:ksubset-lex}
Nonempty subsets of the set $\{0,1,2,\ldots,5\}$ with at most 3 elements.
}
\end{figure}
The modifications needed in algorithm \ref{alg:next-sl2}
for restricting the subsets to those with
a prescribed maximal number of elements
are quite small.
Without ado, we give the implementation.
{\codesize
\begin{listing}{1}
// FILE: src/comb/ksubset-lex.h
class ksubset_lex
{
ulong n_; // number of elements in set, should have n>=1
ulong j_; // number of elements in subset
ulong m_; // max number of elements in subsets
ulong *x_; // x[0...j-1]: subset of {0,1,2,...,n-1}
\end{listing}
}%
The computation of the successor is
{\codesize
\begin{listing}{1}
ulong next()
{
ulong j1 = j_ - 1;
ulong z1 = x_[j1] + 1;
if ( z1 < n_ ) // last element is not max
{
if ( j_ < m_ ) // append element
{
x_[j_] = x_[j1] + 1;
++j_;
return j_;
}
x_[j1] = z1; // increment last element
return j_;
}
else // last element is max
{
if ( j1 == 0 ) return 0; // current is last
--j_;
x_[j_-1] += 1;
return j_;
}
}
\end{listing}
}%
We omit the routine for the predecessor.
Updates take no more time than the updates for the unrestricted subsets.
\section*{Acknowledgment}
I am indebted to Edith Parzefall for
editing two earlier versions of this paper.
\input{bib.tex}
\end{document}
|
\section{Introduction}
The problem of an antiferromagnet on a triangular planar lattice has been
intensively studied
theoretically.~\cite{Kawamura_1985,Korshunov_1986,Anderson_1987,Plumer_1990,Chubukov_1991}
The ground state in the Heisenberg and XY models is a ``triangular'' planar
spin structure with three magnetic sublattices arranged 120$^\circ$ apart. The
orientation of the spin plane is not fixed in the exchange approximation in the
Heisenberg model. The applied static field does not remove the degeneracy of
the classical spin configurations. Therefore the usual small corrections such
as quantum and thermal fluctuations, and relativistic interactions in the
geometrically frustrated magnets play an important role in the formation of the
equilibrium state.~\cite{Korshunov_1986,Chubukov_1991,Rastelli_1996} The
magnetic phase diagrams of such 2D magnets strongly depend on the spin value of
magnetic ions.
CuCrO$_2$ is an example of quasi-two-dimensional antiferromagnet ($S=3/2$) with triangular lattice structure. Below $T_N\approx 24$~K CuCrO$_2$ exhibits spiral ordering to incommensurate spiral magnetic structure with a small deviation from regular 120$^\circ$ structure. The transition to the magnetically ordered state is accompanied by a small distortion of triangular lattice. We present a NMR study of low temperature magnetic structure of CuCrO$_2$ in the fields up to 15.5~T. These fields are small in comparison with exchange interactions within the triangular plane ($\mu_0H_{sat}\approx 280$~T). Thus, we can expect that in our experiments the exchange structure within individual plane will not be distorted significantly and the field evolution of NMR spectra in our experiments will be due to spin plane reorientation or change of interplane ordering. The microscopic properties of magnetic phases of this magnet are especially interesting because this material is multiferroic.~\cite{Seki_2008,Kimura_PRB_2008,Kimura_PRB_2009} The possibility to modify electric and magnetic domains with electric and magnetic fields make CuCrO$_2$ attractive for experimental study of the magnetoelectric coupling in this class of materials.
\section{Crystal and magnetic structure}
\begin{figure}[b!]
\includegraphics[width=.55\columnwidth,angle=0,clip]{fig1.eps}
\caption{(color online) Top: Crystal structure of CuCrO$_2$ projected on the $ab$-plane. The three layers, $\alpha\beta\gamma$, are the positions of Cr$^{3+}$ ions. Bottom: Reference angles $\psi$ and $\varphi$ as defined in the text; the gray bar corresponds to the projection of spin plane. The incommensurate wavevector $\vect{q}_{ic}$ is collinear with the base of the triangle (thick line).}
\label{fig1}
\end{figure}
The structure CuCrO$_2$ consists of magnetic Cr$^{3+}$ (3$d^3$, $S=3/2$), nonmagnetic Cu$^+$, and O$^{2-}$ triangular lattice planes (TLPs), which are stacked along $c$-axis in the sequence Cr-O-Cu-O-Cr (space group $R\bar{3}m$, $a=2.98$~\AA{}, $c =17.11$~\AA{} at room temperature~\cite{Poienar_2009}). The layer stacking sequences are $\alpha\gamma\beta$, $\beta\alpha\gamma$, and $\beta\beta\alpha\alpha\gamma\gamma$ for Cr, Cu and O ions, respectively. The crystal structure of CuCrO$_2$ projected on the $ab$-plane is shown on top portion of Fig.~1. The distances between the nearest planes denoted by different letters for copper and chromium ions and the pairs of planes for oxygen ions are $c/3$, whereas the distance between the nearest oxygen planes denoted by the same letters is $(1/3-0.22)c$ ~(Ref.~[\onlinecite{Poienar_2009}]). No structural phase transition has been reported at temperatures higher than N\'{e}el ordering temperature ($T > T_N\approx 24$~K). In the magnetically ordered state the triangular lattice is distorted, so that one side of the triangle becomes slightly smaller than two other sides: $\Delta a / a \simeq 10^{-4}$~(Ref.~[\onlinecite{Kimura_JPSJ_2009}]).
The magnetic structure of CuCrO$_2$ has been intensively investigated by neutron diffraction experiments.~\cite{Poienar_2009, Kadowaki_1990, Soda_2009, Soda_2010, Frontzek_2012} It was found that the magnetic ordering in CuCrO$_2$ occurs in two stages.~\cite{Frontzek_2012, Aktas_2013} At the higher transition temperature $T_{N1}=24.2$~K two dimensional (2D) ordered state within $ab$-planes sets in, whereas below $T_{N2}=23.6$~K three dimensional (3D) magnetic order with incommensurate propagation vector $\vect{q}_{ic}= (0.329, 0.329, 0)$ along the distorted side of TLPs~\cite{Kimura_JPSJ_2009} is established. The magnetic moments of Cr$^{3+}$ ions can be described by the expression
\begin{eqnarray}
\vect{M}_i=M_1\vect{e}_1\cos(\vect{q}_{ic}\cdot\vect{r}_i+\theta)+M_2\vect{e}_2\sin(\vect{q}_{ic}\cdot\vect{r}_i+\theta),
\label{eqn:spiral}
\end{eqnarray}
where $\vect{e}_1$ and $\vect{e}_2$ are two perpendicular unit vectors determining the spin plane orientation with the normal vector $\vect{n}=\vect{e}_1 \times \vect{e}_2$, $\vect{r}_i$ is the vector to the $i$-th magnetic ion and $\theta$ is an arbitrary phase. The spin plane orientation and the propagation vector of the magnetic structure are schematically shown in the bottom of Fig.~1. For zero magnetic field $\vect{e_1}$ is parallel to $[001]$ with $M_1 = 2.8(2)~\mu_B$, while $\vect{e_2}$ is parallel to $[1\bar{1}0]$ with $M_2 = 2.2(2)~\mu_B$ (Ref.~[\onlinecite{Frontzek_2012}]). The pitch angle between the neighboring Cr moments corresponding to the observed value of $\vect{q}_{ic}$ along the distorted side of TLP is equal to 118.5$^\circ$ which is very near to 120$^\circ$ expected for regular TLP structure.
Owing to the crystallographic symmetry in the ordered phase we can expect {\it six} magnetic domains at $T < T_N$. The propagation vector of each domain can be directed along one side of the triangle and can be positive or negative. As reported in Refs.~[\onlinecite{Soda_2010, Svistov_2013}], the distribution of the domains is strongly affected by the cooling history of the sample.
Inelastic neutron scattering data~\cite{Poienar_2010} has shown that CuCrO$_2$ can be considered as a quasi 2D magnet. The spiral magnetic structure is defined by the strong exchange interaction between the nearest Cr ions within the TLPs with exchange constant $J_{ab}=2.3$~meV. The inter-planar interaction is at least one order of magnitude weaker than the in-plane interaction.
Results of the magnetization, ESR and electric polarization experiments~\cite{Svistov_2013, Kimura_PRB_2009} has been discussed within the framework of the planar spiral spin structure at fields studied experimentally: $\mu_0H < 14$~T~$\ll \mu_0H_{sat}$ ($\mu_0H_{sat}\approx 280$~T). The orientation of the spin plane is defined by the biaxial crystal anisotropy. One {\it hard} axis for the normal vector $\vect{n}$ is parallel to the $c$ direction and the second axis is perpendicular to the direction of the distorted side of the triangle. The anisotropy along $c$ direction dominates with anisotropy constant approximately hundred times larger than that within $ab$-plane resulting from the distortions of the triangle structure. A magnetic phase transition was observed for the field applied perpendicular to one side of the triangle ($\vect{H}\parallel [1\bar{1}0]$) at $\mu_0H_c = 5.3$~T, which was consistently described~\cite{Soda_2010, Svistov_2013, Kimura_PRB_2009} by the reorientation of the spin plane from $(110)$ ($\vect{n}\perp\vect{H}$) to $(1\bar{1}0)$ $(\vect{n}\parallel\vect{H})$. This spin reorientation happens due to weak susceptibility anisotropy of the spin structure $\chi_{\parallel}\approx 1.05\chi_{\perp}$.
\section{Sample preparation and experimental details}
A single crystal of CuCrO$_2$ was grown by the flux method following~Ref.~[\onlinecite{Frontzek_2012}]. The crystal structure was confirmed by single crystal room-temperature x-ray spectroscopy. The magnetic susceptibility ($M(T)/H$) was measured at $\mu_0H=0.5$~T in the temperature range from 2 to 300~K using SQUID magnetometer. The obtained susceptibility curve $\chi_c$ was similar to the data from~Refs.~[\onlinecite{Okuda_2005, Kimura_PRB_2008}]. The N\'{e}el temperature $T_N\approx 24$~K and the Curie-Weiss temperature $\theta_{CW}=-204$~K obtained from the fitting of $M(T)$ at temperatures $150<T<300$~K are in agreement with the values given in~Ref.~[\onlinecite{Kimura_PRB_2008}].
NMR experiments were carried out using a home-built NMR spectrometer. Measurements were taken on a 17.5~T Cryomagnetics field-sweepable NMR magnet at the National High Magnetic Field Laboratory. $^{63,65}$Cu nuclei (nuclear spins $I=3/2$, gyromagnetic ratios $\gamma^{63}/2\pi=11.285$~MHz/T, $\gamma^{65}/2\pi=12.089$~MHz/T) were probed using pulsed NMR technique. In the figures that follow, the spectra shown by solid lines were obtained by summing fast Fourier transforms (FFT), while the spectra shown by circles were obtained by integrating the averaged spin-echo signals as the field was swept through the resonance line. NMR spin echoes were obtained using $1.5~\mu$s$~-~\tau_D~-~3~\mu$s ($\vect{H}\parallel\vect{c}$), $1.8~\mu$s$~-~\tau_D~-~3.6~\mu$s ($\vect{H}\perp\vect{c}$, $\mu_0H\sim 4.5$~T), $2.3~\mu$s$~-~\tau_D~-~4.6~\mu$s ($\vect{H}\perp\vect{c}$, $\mu_0H\sim 11.6$~T) pulse sequences, where the times between pulses $\tau_D$ were 15, 20, 15~$\mu$s, respectively. Measurements were carried out in the temperature range $4.2\leq T \leq 40$~K stabilized with a precision better than 0.03~K. The experimental setup allowed rotating the sample inside the excitation coil with respect to the static field $\vect{H}\perp\vect{c}$ during the experiment.
\section{Experimental results}
The $^{63,65}$Cu NMR spectra for the paramagnetic (Figs.~2a,3a) and ordered (Figs.~2b,3b) states for $\vect{H}\perp\vect{c}$ and $\vect{H}\parallel\vect{c}$ consist of two sets of lines, corresponding to $^{63}$Cu and $^{65}$Cu isotopes. Each set consists of three lines; one of them corresponds to the central line ($m_I = -1/2\leftrightarrow +1/2$) and two quadrupole satellites corresponding to ($\pm 3/2\leftrightarrow \pm 1/2$) transitions.
For the paramagnetic state the spectral shape was found to be independent of the magnetic field orientation within $ab$-plane. In contrast, the shapes of the spectra at temperatures below $T_N$ are strongly dependent on the field direction and cooling history. The spectra were studied under two cooling conditions: zero field cooling (ZFC) and field cooling (FC).
In the first set of the experiments the static field was oriented within $ab$-plane (Figs.~4-9). The static magnetic field during the FCs was directed parallel ($\vect{H}\parallel [110]$) or perpendicular ($\vect{H}\parallel [1\bar{1}0]$) to one side of the triangular structure. For all FC data, the sample was cooled in a field of 16.9~T from 40 to 4.2~K (or 5~K) with characteristic time $t_{cooling}\approx 40$~min, and then the measurements were performed. The sample was rotated about the $c$-axis. The direction of the external field given on the figures is measured with respect to the $[110]$ direction of the sample. Since the data obtained for the full rotation of the sample reveal a $180^\circ$ symmetry only data from $0^\circ$ to $150^\circ$ are shown for clarity.
\begin{figure}
\includegraphics[width=0.9\columnwidth,angle=0,clip]{fig2.eps}
\caption{(color online) $^{63,65}$Cu NMR spectra of the CuCrO$_2$ single crystal in the paramagnetic state (a) and in the ordered state (b) at the external magnetic field directed perpendicular to $c$ axis, $\vect{H}\parallel[110]$. The two sets of lines correspond to the signals from quadrupolar split $^{63}$Cu and $^{65}$Cu nuclei (see text). The peaks marked with crosses are spurious $^{63,65}$Cu and $^{27}$Al NMR signals from the probe.}
\label{fig2}
\end{figure}
\begin{figure}
\includegraphics[width=0.9\columnwidth,angle=0,clip]{fig3.eps}
\caption{(color online) Similar spectra as in Fig.~2 except with field applied parallel to $c$ axis, $\vect{H}\parallel[001]$.}
\label{fig3}
\end{figure}
The angular dependences within $ab$-plane were measured at the frequencies 55.3~MHz and 137~MHz. The lower frequency is chosen such that the central line of $^{63}$Cu NMR is situated at fields near 4.5~T, i.e. below reorientation field $\mu_0H_c=5.3$~T, whereas the higher frequency will place the resonance above $\mu_0H_c$.
The NMR spectra were measured at two temperatures $T_{high}=20$~K (Figs.~4,8) and $T_{low}=4.2$~K (Figs.~5,6,7) and $T_{low}=5$~K (Fig.~9). We chose these two temperature sets, since the domain walls in CuCrO$_2$ are mobile at high temperatures, whereas at low temperatures the walls are pinned.~\cite{Svistov_2013}
In the second set of the experiments the static field was oriented parallel to the $c$ direction. Representative ZFC spectra at 20~K measured at different fields are shown in Fig.~10.
\section{Discussion}
We shall discuss the results of the NMR experiments in the framework of planar spiral spin structure (Eq.~(\ref{eqn:spiral})) proposed from neutron diffraction experiments~\cite{Frontzek_2012} and carried out at $H=0$. Spin plane orientation of such structure is defined by weak relativistic interactions with the external field and crystal environment. Following~Ref.~[\onlinecite{Svistov_2013}], the anisotropic part of magnetic energy of CuCrO$_2$ can be written as:
\begin{eqnarray}
U=-\frac{\chi_{\parallel}-\chi_{\perp}}{2}(\vect{n}\vect{H})^2+\frac{A}{2}n_{z}^2+\frac{B}{2}n_{y}^2,
\label{eqn:energy}
\end{eqnarray}
where $A > B > 0$. For the arbitrary field direction within $ab$-plane the vector $\vect{n}$ will monotonously rotate from $\vect{n}\parallel[110]$ to $\vect{n}\parallel\vect{H}$. This can be defined by minimization of Eq.~(\ref{eqn:energy}), which can be rewritten as:
\begin{eqnarray}
U=-\frac{\Delta\chi}{2}H^2(\cos^2(\psi - \varphi)-\Bigr(\frac{H_{cy}}{H}\Bigl)^2\sin^2\varphi)+\frac{A}{2}n_{z}^2.
\label{eqn:energy_psi}
\end{eqnarray}
Here the angles $\psi$ and $\varphi$ define the directions of the vectors $\vect{H}$ and $\vect{n}$, respectively, as depicted in Fig.~1. Using value of reorientation field $\mu_0H_{cy}=5.3$~T we obtain the expected orientations of the spin planes relative to the field directions of our NMR experiments. For the field directed along ``thick'' side of the triangle within $ab$-plane (i.e. $\psi=0^\circ$) $\varphi=0^\circ$ at all fields. For the field directed along ``thin'' side of the triangle (i.e. $\psi=60^\circ$) the expected angles $\varphi$ for a given field are: $\varphi(\mu_0H=4.5$~T$)=22.15^\circ$ and $\varphi(\mu_0H=11.6$~T$)=54.3^\circ$. For the field direction perpendicular to ``thick'' side of triangle (i.e. $\psi=90^\circ$) below spin-flop $\varphi(\mu_0H < \mu_0H_{cy}=5.3$~T$)=0^\circ$ and above spin-flop $\varphi(\mu_0H > \mu_0H_{cy}=5.3$~T$)=90^\circ$. For the field direction perpendicular to ``thin'' side of triangle (i.e. $\psi=30^\circ$) the orientation of spin plane is defined by the angles: $\varphi(\mu_0H=4.5$~T$)=12.3^\circ$ and $\varphi(\mu_0H=11.6$~T$)=25.4^\circ$.
\begin{figure}
\includegraphics[width=1\columnwidth,angle=0,clip]{fig4.eps}
\caption{(color online) $^{63}$Cu NMR spectra ($m_I = -1/2\leftrightarrow +1/2$) measured at different angles between $\vect{H}$ applied within $ab$-plane and $[110]$ direction of the sample (red solid circles). ZFC to $T=20$~K, $\nu=55.3$~MHz. Black solid lines on the figure are calculated spectra in the model of the magnetic structure (Eq.~(\ref{eqn:spiral})) and orientations of spin planes (gray bars) shown in the bottom of the figure. $A$, $B$ and $C$ accord to three possible alignment of propagation vector of magnetic structure ($\vect{q}_{ic}$ is collinear to the triangle side marked thick). $\eta_A$, $\eta_B$, $\eta_C$ - the relative weights of the NMR signals from $A$, $B$, $C$ domains, which were used for the best coincidence with experiment.}
\label{fig4}
\end{figure}
\begin{figure}
\includegraphics[width=1\columnwidth,angle=0,clip]{fig5.eps}
\caption{(color online) Similar data to Fig.~4 except ZFC to $T=4.2$~K.}
\label{fig5}
\end{figure}
For $\vect{H}\parallel\vect{c}$ the field of spin reorientation transition $H_{cz}$ is expected much larger, than the fields in our experiments.~\cite{Svistov_2013} Therefore the spin plane orientation for $\vect{H}\parallel\vect{c}$ we expect the same as at $H=0$ ($\vect{n}\parallel [110]$).
Analyzing the NMR spectral shapes we found that they can be well described
within the model of spiral spin structure given by~Eq.~(\ref{eqn:spiral}) with
the incommensurate propagation vector directed along one side of the triangle
$q_{ic}=0.329$. Generally, the local magnetic field at Cu sites is the sum of
the long range dipole field $\vect{H}_{dip}$ and the transferred hyperfine
contact field produced by the nearest Cr$^{3+}$ moments. The $^{63,65}$Cu NMR
of CuCrO$_2$ in paramagnetic state was studied
in~Ref.~[\onlinecite{Verkhovskii_2011}], where it was shown that the effective
field at the copper site is proportional to the chromium moment with a
hyperfine field of 3.3~T/$\mu_B$. This value does not depend on the direction
of the chromium moment. The computed dipolar fields on the chromium nuclei in
paramagnetic phase are anisotropic and essentially smaller, than experimentally
observed, and equal to 0.17~T/$\mu_B$ and -0.08~T/$\mu_B$ for
$\vect{H}\parallel[001]$ and $\vect{H}\parallel[1\bar{1}0]$, respectively. From
these data we presumed that the effective field measured in paramagnetic
phase~\cite{Verkhovskii_2011} is mostly defined by the contact fields from six
nearest chromium magnetic ions. Although the contact field created by an
individual neighbor chromium moment (3.3/6~T/$\mu_B=0.55$~T/$\mu_B$) is much
larger than the dipole field, dipolar and contact hyperfine contributions prove
to be comparable in the ordered state. This is due to a strong compensation of
the contact fields from the six nearest chromium moments in the ordered state.
We took into account both contributions in our calculations.
The dipolar fields on the copper nuclei were computed by numerically summing
contributions from nearest neighbor Cr moments in the sphere of radius
$20$~\AA{}, further accounting for the moments farther away gives no noticeable effect.
The shape of the individual NMR line was taken Lorentzian for the fits.
The best fit with experiment $\vect{H}\perp\vect{c}$ was obtained with the value of Cr$^{3+}$ magnetic moments $M_1=M_2=0.91(5)~\mu_B$ at low fields $\mu_0H\sim 4.5$~T (Figs.~4,5,6,7) and $M_1=M_2=1.15(9)~\mu_B$ at high fields $\mu_0H\sim 11.6$~T (Figs.~8,9). Each individual linewidth is $\delta=20(5)$~mT for all fitted NMR spectra. This value is consistent with the linewidth measured in the paramagnetic phase.
Since the copper nuclei in CuCrO$_2$ are situated at a position of high symmetry, the NMR spectra from the magnetic domains with opposite directions of incommensurate vectors $+\vect{q}_{ic}$ and $-\vect{q}_{ic}$ are identical. In the analysis that follow we shall assign magnetic domains as letters $A$, $B$ and $C$, having in mind that each letter refers to two magnetic domains indistinguishable by NMR method.
The ZFC NMR spectra measured at frequency 55.3~MHz and temperatures 20~K and 4.2~K are shown on Figs.~4 and 5. For the field directed parallel to one side of the triangle ($\psi=0^\circ$, $60^\circ$, $120^\circ$) we expect that the resonance conditions of two domains ($B$, $C$) will be equivalent. A sketch of the expected spin plane orientations within $A$~($\psi=0^\circ$, $\varphi=0^\circ$), $B$~($\psi=60^\circ$, $\varphi=22.15^\circ$), $C$~($\psi=120^\circ$, $\varphi=157.85^\circ$) domains are shown at the bottom left of each figure. The dashed and dotted lines show the computed spectra for domains $A$ and $B+C$, respectively. The resulting spectrum has been obtained by summing the spectra from the three domains with relative weights $\eta_A$, $\eta_B$, $\eta_C$ and is shown as solid lines in the figures.
\begin{figure}
\includegraphics[width=1\columnwidth,angle=0,clip]{fig6.eps}
\caption{(color online) Similar data to Fig.~4 except FC $\vect{H}$(16.9~T)$\parallel[110]$ ($\psi=0^\circ$) to $T=4.2$~K. Background at some panels are due to overlapping with $^{65}$Cu NMR lines.}
\label{fig6}
\end{figure}
\begin{figure}
\includegraphics[width=1\columnwidth,angle=0,clip]{fig7.eps}
\caption{(color online) Similar data to Fig.~4 except FC $\vect{H}$(16.9~T)$\parallel[1\bar{1}0]$ ($\psi=30^\circ$) to $T=4.2$~K. Background at some panels are due to overlapping with $^{65}$Cu NMR lines.}
\label{fig7}
\end{figure}
If the relative sizes of the domains do not change during the rotation of the field we expect that the sum of the relative fractions of domain $A$ $\eta_A$ measured at three unique orientations $\psi=0^\circ$, $60^\circ$, $120^\circ$ i.e., $\eta_A(0^\circ)+\eta_A(60^\circ)+\eta_A(120^\circ)$, will be equal to unity. The same is true for the sum $\eta_C(30^\circ)+\eta_C(90^\circ)+\eta_C(150^\circ)$. However, the experimental values of the sums for $T=20$~K (Fig.~4) are $\eta_A(0^\circ)+\eta_A(60^\circ)+\eta_A(120^\circ)=1.36$ and $\eta_C(30^\circ)+\eta_C(90^\circ)+\eta_C(150^\circ)=0.65$. This deviation from unity shows that the domain sizes of the sample change with field rotation. These observations show that the size of the energetically favorable domains grows in expense of unfavorable domains. To our knowledge, this phenomenon is directly observed the first time through the NMR technique.
For measurements at $T=4.2$~K (Fig.~5) the sums are more close to unity: $\eta_A(0^\circ)+\eta_A(60^\circ)+\eta_A(120^\circ)=1.17$; $\eta_C(30^\circ)+\eta_C(90^\circ)+\eta_C(150^\circ)=0.88$. This implies that the mobility of domain walls increases with the temperature. We emphasize that the conversion of the domain structure implies changes not only in the spin plane orientation, but also in the direction of the wave vector $\vect{q}_{ic}$.
The field cooling of the sample enables us to prepare the sample with the energetically preferable domains. If the field during the cooling process was directed along one side of the triangle ($\psi=0^\circ$, Fig.~6) domain $A$ is preferable. In this case, field cooling the sample results in $\sim 85$~\% domain $A$ and $\sim 15$~\% $B$ and $C$. If the cooling field is directed perpendicular to one side of the triangle ($\psi=30^\circ$, Fig.~7) the domains $B$ and $C$ are preferable. In such a case, the NMR signal from unfavorable domain will be negligibly small and the two other domains will have nearly the same sizes. Interestingly, the relative part of the sample, where the direction of $\vect{q}_{ic}$ changes with field rotation at $T=4.2$~K has nearly the same intensity as for the ZFC procedure. The parameters defining the domain conversion for FC samples are $\eta_A(0^\circ)+\eta_A(60^\circ)+\eta_A(120^\circ)=1.12$ and $\eta_C(30^\circ)+\eta_C(90^\circ)+\eta_C(150^\circ)=0.91$ for the cooling field direction $\psi=0^\circ$ (Fig.~6). For the cooling field direction $\psi=30^\circ$ (Fig.~7) these parameters are 1.24 and 0.88, respectively. Thus, the domain conversion during the rotation of the static field $\mu_0H\approx4.5$~T at $T=4.2$~K takes place within 4-8~\% of the sample.
Such domain conversion is more definitely observed at higher fields $\mu_0H\approx11.6$~T $>\mu_0H_c=5.3$~T and higher temperature $T=20$~K (see Fig.~8). For the fields directed along one side of the triangular structure, the NMR spectra can be solely identified with a single energetically preferable domain with $\vect{q}_{ic}$ parallel to the applied field ($\psi=0^\circ$, $60^\circ$, $120^\circ$, spectra on the left panels of Fig.~8). This means, that at $T=20$~K, the field ($\sim11.6$~T) rotates not only the spin plane of magnetic structure, but also rebuilds the domain structure, so that only the energetically preferable domains are established in the sample, namely the domains with $\vect{q}_{ic}\parallel\vect{H}$. For the fields directed perpendicular to one side of the triangle ($\psi=30^\circ$, $90^\circ$, $150^\circ$, spectra on the right panels of Fig.~8) domains $A$ and $B$ are more energetically preferable. The ZFC spectra observed at high fields ($\sim11.6$~T) and low temperature (5~K) (Fig.~9) are qualitatively similar to ZFC spectra measured at low fields ($\sim4.5$~T). Only the sums defining the domain conversion of the sample, $\eta_A(0^\circ)+\eta_A(60^\circ)+\eta_A(120^\circ)=1.27$ and $\eta_C(30^\circ)+\eta_C(90^\circ)+\eta_C(150^\circ)=0.5$, are larger than those for the low field case (1.17 and 0.88, respectively).
\begin{figure}
\includegraphics[width=1\columnwidth,angle=0,clip]{fig8.eps}
\caption{(color online) Similar data to Fig.~4 except ZFC to $T=20$~K, $\nu=137$~MHz.}
\label{fig8}
\end{figure}
\begin{figure}
\includegraphics[width=1\columnwidth,angle=0,clip]{fig9.eps}
\caption{(color online) Similar data to Fig.~4 except ZFC to $T=5$~K, $\nu=137$~MHz.}
\label{fig9}
\end{figure}
The NMR spectra measured at field along the $c$-axis ({\it hard} axis for $\vect{n}$ vector of the structure) is 2.5 times broader, than those with fields aligned within the $ab$-plane, Fig.~10. The shape of the spectra depends on the field value. At low field range ($\mu_0H<\sim10$~T) the shape has two horns similar to those observed for fields oriented within $ab$-plane. For higher fields, an additional third peak appears in the middle of the spectra.
The low field spectra can be satisfactorily described by the model of incommensurate spiral spin structure similar to the structure proposed for zero field (Eq.~(\ref{eqn:spiral})). The best fit at $T=20$~K and $\mu_0H\approx9$~T is obtained with $M_1=M_2=2.1~\mu_B$, $\delta=20$~mT (Fig.~10, blue dotted line).
\begin{figure}
\includegraphics[width=0.95\columnwidth,angle=0,clip]{fig10.eps}
\caption{(color online) Representative $^{63,65}$Cu NMR spectra at different field (black solid lines) under ZFC condition, $T=20$~K. Shown are data taken with each field swept over a narrow range at fixed frequencies. The $x$-axis was adjusted so that each panel covers the same field range. The blue dotted line is calculated spectrum using (Eq.~(\ref{eqn:spiral})), red dot-dashed and green dashed lines are calculated spectra corresponding to incommensurate spiral magnetic structure with disorder in the $c$ direction and commensurate spiral magnetic structure. See text for details.}
\label{fig10}
\end{figure}
The change of the NMR spectral shape at higher fields indicates the field evolution of the magnetic structure. We suggest two possible models of the high field magnetic structure. The red dash-dotted line in Fig.~10 shows the result of computed NMR spectra using Eq.~(\ref{eqn:spiral}) with random phases $\theta$ for structures within different triangular planes with $M_1=M_2=2.2~\mu_B$, $\delta=20$~mT. Such model accounts for the incommensurate spiral magnetic structure within every $ab$-plane with disorder in the $c$ direction. In this case the experimental spectra can be described by the sum of the spectra from the parts of the sample with order and disorder along $c$ direction.
Another model for the three horn spectra is the commensurate magnetic structure with pitch angle close to $120^\circ$, in which one moment looks along $[00\bar{1}]$ direction, i.e. opposite to the applied field $\vect{H}$. Computed NMR spectrum for this case with $M_1=M_2=2.7~\mu_B$, $\delta=30$~mT is given in the Fig.~10, shown as green dashed line. A transition of the spiral structure from incommensurate to commensurate in large enough fields has been observed in other compounds with triangular structure.~\cite{Prozorova_1993, Prozorova_2004, Park_2007}
For both models the values of $M_1$, $M_2$ are closer to the maximum value expected for chromium magnetic moment $g\mu_BS$, compare to those values for perpendicular orientation. It is probable that the static field applied along $c$ direction suppresses the fluctuating part of the magnetic moments of Cr ions which could explain its large value when $\vect{H}\parallel\vect{c}$ compared to its value when $\vect{H}\perp\vect{c}$.
The observed ``third peak'' peculiarity possibly corresponds to a phase transition also seen by recent electric polarization experiments.~\cite{Zapf_2014}
\section{Conclusions}
$^{63,65}$Cu NMR spectra measured at $\vect{H}\perp[001]$ can be well described by the planar spiral magnetic structure with the oscillating components of the moments approximately 2.5 times smaller than obtained from neutron diffraction experiments.~\cite{Poienar_2009, Frontzek_2012} Rotation of the sample in a magnetic field results in the reorientation of the spin plane accompanied by the reorientation of the incommensurate wave vector of the structure. This wave vector follows the direction of magnetic field at high enough temperature and fields, whereas at low temperatures or low fields the propagation vector is defined by the cooling history of the sample. The results are consistent with previous results of electric polarization and ESR studies.~\cite{Svistov_2013, Kimura_PRB_2009}
The NMR study of the magnetic structure at $\vect{H}\parallel[001]$ shows that the low field magnetic structure is consistent with the structure proposed by neutron diffraction experiments. In fields higher than 10~T the magnetic structure is modified. The results can be described by the loss of long range ordering in the $c$ direction, or by the transition from incommensurate to commensurate structure. This observation opens interesting possibilities for future experimental investigations.
\acknowledgements We thank V.I. Marchenko and V.L. Matukhin for stimulating
discussions. Yu.A.S. is grateful to the Institute of International Education,
Fulbright Program, for financial support (Grant 68435029). H.D.Z. thanks for
the support from NSF-DMR through award DMR-1350002. This work was supported by
Russian Foundation for Basic Research, Program of Russian Scientific Schools
(Grant 13-02-00637). Work at the National High Magnetic Field Laboratory is
supported by the NSF Cooperative Agreement No. DMR-0654118, and by the State of
Florida.
|
\section{Introduction}
The deterministic quantum computation with one quantum bit (DQC1),
or the one clean qubit model,
proposed by Knill and Laflamme~\cite{KL} is a restricted model
of quantum computing. It was originally motivated by nuclear magnetic resonance (NMR)
quantum information processing, but now it has been extensively studied
in variety of contexts
to understand the border between quantum and classical
computing~\cite{Poulin,Poulin2,SS,Passante,JW,3mani,Datta,Datta2,Datta3,
MFF,Elham}.
As shown in Fig.~\ref{DQC1} (a), a DQC1 circuit consists of
\begin{itemize}
\item[1.]
the input state $|0\rangle\langle0|\otimes\frac{I^{\otimes n}}{2^n}$,
where $I\equiv|0\rangle\langle0|+|1\rangle\langle1|$
is the two-dimensional identity operator,
\item[2.]
a $poly(n)$-size quantum gate $U$,
\item[3.]
the computational basis measurement of the first qubit,
which gives the output bit $a\in\{0,1\}$.
\end{itemize}
The DQC1 model seems to be very weak, and in fact,
it does not support universal quantum computation
under reasonable assumptions~\cite{Ambainis}.
However, counter-intuitively, the DQC1 model
can efficiently solve some problems for which no efficient classical algorithm is
known, such as the spectral density estimation~\cite{KL},
testing integrability~\cite{Poulin},
calculation of fidelity decay~\cite{Poulin2},
approximation of the Jones and HOMFLY polynomials~\cite{SS,Passante,JW}
and an invariant of 3-manifolds~\cite{3mani}.
These results are surprising since if we replace the single pure input qubit
with the completely mixed qubit, then the output suddenly becomes trivially classically
simulatable. In other words, the single pure qubit hides some capacity
of unlocking the potential power of highly-mixed states
for the quantum speedup.
In fact, such an unexpected power of the DQC1 model was shown to
come from some non-classical correlations
between the mixed register and the pure qubit~\cite{Datta,Datta2,Datta3}.
In short, the DQC1 model is believed
to be a model of computation which is intermediate classical
and universal quantum computation.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.45\textwidth]{DQC1.eps}
\end{center}
\caption{(a) The DQC1 model. (b) The DQC1$_m$ model for $m=3$.
}
\label{DQC1}
\end{figure}
In Ref.~\cite{MFF},
a generalized version of
the DQC1 model, which is called the DQC1$_m$ model,
was introduced.
The DQC1$_m$ model is the same as the DQC1 model except
that not a single but $m$ output qubits are measured
in the computational basis
at the end of the computation (Fig.~\ref{DQC1} (b)).
In particular, the DQC1$_1$ model is equivalent to the DQC1 model.
It was shown in Ref.~\cite{MFF} that the output probability
distribution of the DQC1$_m$ model
for $m\ge3$
cannot be classically
efficiently sampled (within a multiplicative error) unless the polynomial hierarchy
collapses at the third level~\cite{MFF}.
The polynomial hierarchy~\cite{Toda} is a natural way of classifying the complexity
of problems (languages) beyond the usual NP (nondeterministic polynomial time,
which includes ``traveling salesman" and ``satisfiability" problems).
Since it is believed in computer science that
the polynomial hierarchy does not collapse,
it is unlikely that the DQC1$_m$ model for $m\ge3$ can be classically efficiently
sampled. This result is based on the ``postselection technique"
used for other restricted models of quantum computing,
such as the depth-four circuits model by Terhal and DiVincenzo~\cite{TD},
the instantaneous quantum polytime (IQP) model by Bremner, Jozsa, and Shepherd~\cite{BJS}, and
the Boson sampling model by Aaronson and Arkhipov~\cite{AA}.
Can we show an impossibility
of classical efficient simulation of the DQC1$_m$ model
for $m\le2$?
In particular, the case for $m=1$,
which corresponds to the original DQC1 model,
has been the long standing open problem.
In this paper, we show that if we can classically efficiently approximate
(exactly within a polynomial bit length
or in the fully polynomial randomized approximation scheme (FPRAS)
with at most a constant error)
the output probability distribution of the DQC1 model,
then $\mbox{BQP}\subseteq\mbox{BPP}$.
Although the belief of
$\mbox{BPP}\subsetneq\mbox{BQP}$
is relatively less solid than the belief of P$\neq$NP
or that the polynomial hierarchy does
not collapse, researchers in quantum computing
believe $\mbox{BPP}\subsetneq\mbox{BQP}$
(for example, there is an oracle $A$ relative to which
$\mbox{BPP}^A\neq\mbox{BQP}^A$~\cite{BB}).
Therefore our results suggest that
DQC1 model can unlikely be classically efficiently simulated
in these senses.
If we replace the single pure qubit $|0\rangle$ of the DQC1 model with
the completely-mixed qubit $I$,
the output probability distribution suddenly becomes easy to be classically
calculated.
Therefore, our results demonstrate
``a power of a single qubit" in the DQC1 model.
Furthermore, it is known that some quantum circuits,
which look more complicated than the DQC1 model,
such as the stabilizer circuits~\cite{GK} and matchgates~\cite{match},
can be classically efficiently simulated.
These facts therefore suggest that the superficially naive DQC1 model
indeed hides some ability of doing complicated quantum computing.
Note that definitions of classical simulatability considered in this paper
are based on the calculation, and therefore stronger than the sampling,
which was used in Refs.~\cite{BJS,MFF}.
It is still an open problem whether DQC1$_m$ model for $m\le2$
is hard to be classically simulated in the sense of sampling.
\section{First Result}
Let
$P(a):\{0,1\}\to[0,1]$
be the probability of obtaining the result $a\in\{0,1\}$ in the computational
basis measurement of the output qubit in the DQC1 model.
Let $P'(a)$ be an approximation of $P(a)$:
\begin{eqnarray*}
0\le P(a)-P'(a)\le\epsilon,
\end{eqnarray*}
where $0\le\epsilon\le\frac{1}{2^r}$
and $r$ is a polynomial of $n$ sufficiently larger than $n-1$.
If there exists a deterministic polynomial time
Turing machine that can calculate $P'(a)$,
then
$\mbox{BQP}\subseteq\mbox{BPP}$.
\section{Proof}
Let $L$ be a language in the class
BQP. This means that
there exists a constant (error tolerance) $0<\delta<\frac{1}{2}$,
a uniform family of unitary operators $\{V_w\}$
acting on $|0\rangle^{\otimes n}$, where $n=poly(|w|)$, and
a specified single qubit output register $o\in\{0,1\}$ for the
$L$-membership decision problem such that
\begin{itemize}
\item[1.]
if $w\in L$ then $Prob(o=0)\ge 1-\delta$ and
\item[2.]
if $w\notin L$ then $Prob(o=0)\le \delta$.
\end{itemize}
By using the majority voting technique, $\delta$ can be
exponentially small~\cite{Barak}.
For a unitary operator $V_w$,
we construct the DQC1 circuit of Fig.~\ref{DQC1circuit},
where the first $n$-qubit Toffoli gate is defined by
\begin{eqnarray*}
X\otimes |0\rangle\langle0|^{\otimes n}
+I\otimes (I^{\otimes n}-|0\rangle\langle0|^{\otimes n}),
\end{eqnarray*}
and $X\equiv|0\rangle\langle1|+|1\rangle\langle0|$
is the bit flip operator.
For this circuit, we obtain
\begin{eqnarray*}
P(a=0)=\frac{q}{2^{n-1}}
+\frac{1}{2}-\frac{1}{2^n},
\end{eqnarray*}
where
\begin{eqnarray*}
q\equiv\mbox{Tr}\Big[
\Big(|0\rangle\langle0|\otimes I^{\otimes (n-1)}\Big)
\Big(V_w|0\rangle\langle0|^{\otimes n}V_w^\dagger\Big)
\Big]
\end{eqnarray*}
is the probability of obtaining the positive result for
the BQP circuit.
By the assumption, there exists a deterministic polynomial time Turing machine
that can calculate $P'(a=0)$.
From the deterministic polynomial time Turing machine,
we can construct the probabilistic polynomial time Turing machine
that outputs $o\in\{0,1\}$
according to the probability
\begin{eqnarray*}
Prob(o=0)&=&
2^{n-1}\Big[P'(a=0)-\frac{1}{2}+\frac{1}{2^n}\Big]\\
Prob(o=1)&=&
1-2^{n-1}\Big[P'(a=0)-\frac{1}{2}+\frac{1}{2^n}\Big].
\end{eqnarray*}
Then, the probabilistic polynomial time Turing machine outputs
$o=0$ in the following way:
\begin{itemize}
\item[1.]
if $w\in L$ then
\begin{eqnarray*}
Prob(o=0)&=&
2^{n-1}\Big[P'(a=0)-\frac{1}{2}+\frac{1}{2^n}\Big]\\
&\ge&
2^{n-1}\Big[P(a=0)-\epsilon-\frac{1}{2}+\frac{1}{2^n}\Big]\\
&=&
2^{n-1}\Big[\frac{q}{2^{n-1}}-\epsilon\Big]\\
&=& q-\epsilon2^{n-1}\\
&\ge& (1-\delta)-\frac{1}{2^{r-(n-1)}},
\end{eqnarray*}
and
\item[2.]
if $w\notin L$ then
\begin{eqnarray*}
Prob(o=0)&=&
2^{n-1}\Big[P'(a=0)-\frac{1}{2}+\frac{1}{2^n}\Big]\\
&\le&
2^{n-1}\Big[P(a=0)-\frac{1}{2}+\frac{1}{2^n}\Big]\\
&=&
q\\
&\le&\delta,
\end{eqnarray*}
\end{itemize}
and therefore $\mbox{BQP}\subseteq\mbox{BPP}$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.25\textwidth]{DQC1circuit.eps}
\end{center}
\caption{
The DQC1 circuit created from the $n$-qubit Toffoli gate
and $V_w$.}
\label{DQC1circuit}
\end{figure}
\section{Second Result}
Let us define
\begin{eqnarray*}
Q(a)\equiv P(a)-\frac{1}{2}.
\end{eqnarray*}
The fully polynomial randomized approximation
scheme (FPRAS) means that we can obtain an approximation
$Q'(a)$ of $Q(a)$, which satisfies
\begin{eqnarray*}
Prob\Big(\Big|Q(a)-Q'(a)\Big|\le\epsilon Q(a)\Big)\ge 1-\eta
\end{eqnarray*}
for given $\epsilon>0$ and $0<\eta<1$,
within $poly(\epsilon^{-1},\ln\eta^{-1},n)$ time.
Our second result is that if a classical computer can calculate $Q'(a)$
in the FPRAS with $\epsilon<\frac{1}{2}$
and $\eta<\frac{1}{2}$, and within $poly(n)$ time,
then $\mbox{BQP}\subseteq\mbox{BPP}$.
Note that
$\eta=\frac{1}{4}$ is sufficient to reduce the failure probability
to an arbitrarily small value $\eta'$.
\section{Proof}
By the assumption, $Q'(a=0)$ can be calculated in the FPRAS.
Now we can construct the probabilistic polynomial time Turing machine
that outputs $o\in\{0,1\}$
according to the probability~\cite{largerthan1}
\begin{eqnarray*}
Prob(o=0)&=&
2^{n-1}\Big[Q'(a=0)+\frac{1}{2^n}\Big]\\
Prob(o=1)&=&
1-2^{n-1}\Big[Q'(a=0)+\frac{1}{2^n}\Big].
\end{eqnarray*}
Then, the probabilistic polynomial time Turing machine outputs
$o=0$ in the following way:
\begin{itemize}
\item[1.]
if $w\in L$ and $Q'(a=0)\ge Q(a=0)$ then
\begin{eqnarray*}
Prob(o=0)&=&
2^{n-1}\Big[Q'(a=0)+\frac{1}{2^n}\Big]\\
&\ge& 2^{n-1}\Big[Q(a=0)+\frac{1}{2^n}\Big]\\
&=&q\\
&\ge&1-\delta.
\end{eqnarray*}
\item[2.]
if $w\in L$ and $Q'(a=0)\le Q(a=0)$ then~\cite{affect}
\begin{eqnarray*}
Prob(o=0)&=&
2^{n-1}\Big[Q'(a=0)+\frac{1}{2^n}\Big]\\
&\ge&
2^{n-1}\Big[(1-\epsilon)Q(a=0)+\frac{1}{2^n}\Big]\\
&=&
(1-\epsilon)q+\frac{\epsilon}{2}\\
&\ge&
(1-\epsilon)(1-\delta)+\frac{\epsilon}{2}\\
&\ge&
(1-\epsilon)(1-\delta).
\end{eqnarray*}
\item[3.]
if $w\notin L$ and $Q'(a=0)\le Q(a=0)$ then
\begin{eqnarray*}
Prob(o=0)&=&
2^{n-1}\Big[Q'(a=0)+\frac{1}{2^n}\Big]\\
&\le&
2^{n-1}\Big[Q(a=0)+\frac{1}{2^n}\Big]\\
&=&q\\
&\le&\delta.
\end{eqnarray*}
\item[4.]
if $w\notin L$ and $Q'(a=0)\ge Q(a=0)$ then~\cite{affect}
\begin{eqnarray*}
Prob(o=0)&=&
2^{n-1}\Big[Q'(a=0)+\frac{1}{2^n}\Big]\\
&\le&
2^{n-1}\Big[(1+\epsilon)Q(a=0)+\frac{1}{2^n}\Big]\\
&=&(1+\epsilon)q-\frac{\epsilon}{2}\\
&\le&(1+\epsilon)\delta-\frac{\epsilon}{2}\\
&\le&(1+\epsilon)\delta.
\end{eqnarray*}
\end{itemize}
Therefore we conclude that $\mbox{BQP}\subseteq\mbox{BPP}$.
\section{Discussion}
In this paper, we have shown that the classical efficient approximation
(exactly within a polynomial bit length or in the FPRAS with at most
a constant error) of the output probability distribution
of the DQC1 model is impossible unless $\mbox{BQP}\subseteq\mbox{BPP}$.
Since it is believed that $\mbox{BPP}\subsetneq\mbox{BQP}$,
our results suggest that
it is unlikely that the DQC1 model can be simulated in these senses.
In Ref.~\cite{Datta3}, it was shown that the DQC1 model cannot be simulated
by using the tensor-network method, since the Schmidt rank increases exponentially.
Our first result can be considered as a generalization of their result:
whatever method is utilized, classical efficient simulation of the DQC1 model is impossible
(unless $\mbox{BQP}\subseteq\mbox{BPP}$).
As we have mentioned, our definitions of classical simulatability
are stronger than the sampling considered in Refs.~\cite{MFF,BJS}.
It will be a future study to generalize our results to the
sampling.
\acknowledgements
TM is supported by the Tenure Track System by MEXT Japan
and KAKENHI 26730003 by JSPS.
TK is supported by KAKENHI
26540002, 24106008, 24240001, 23246071
by JSPS.
|
\section{Introduction}
\setcounter{footnote}{0}
A large fraction of known Galactic black hole X-ray binaries
(BHXBs) are highly transient sources. They are thought
to be low-mass X-ray binaries, in which a stellar mass black
hole accretes gas from a low-mass donor star via Roche-lobe overflow.
They normally stay in a very faint state, but occasionally undergo
dramatic outbursts, changing their X-ray luminosities in several orders
of magnitude \citep[e.g.,][]{tan96} on a human timescale (typically
$\approx$10 days to several months for the whole period of an outburst).
Thus, they are excellent targets to study the physics
of black hole accretion over a wide range of mass accretion rates.
Along with the increase and decrease of the X-ray luminosity,
they exhibit several distinct ''states'' with different spectral
and timing properties. The most canonical ones among them are
so-called the ''high/soft state'' and the ''low/hard state'', which
appears in relatively high and low luminosities
($\sim 0.1 L_{\rm Edd}$\footnote{$L_{\rm Edd}$ is the Eddington luminosity:
$4 \pi G m_{\rm p} c M_{\rm BH}/ \sigma_{\rm T}$, where $G$, $m_{\rm p}$, $c$,
$M_{\rm BH}$, and $\sigma_{\rm T}$ represent the gravitational constant,
proton mass, speed of light, black hole mass, and Thomson scattering
cross-section, respectively.}, and $\lesssim 10^{-2} L_{\rm Edd}$,
respectively; see e.g., \citealt{mcc06, don07} for a review).
In the high/soft state, the X-ray flux is dominated by the thermal
emission from the optically-thick and geometrically-thin accretion
disk \citep[the standard disk;][]{sha73}, and the observed spectrum
is well reproduced with a multi-color disk model \citep{mit84}.
The inner disk radius stays constant during the high/soft state,
regardless of the significant change in X-ray luminosity
\citep[e.g.,][]{ebi93, kub04, ste10, shi11a}. This indicates that
the standard disk stably extends down to the
innermost stable circular orbit (ISCO). In the low/hard state,
BHXBs exhibit a hard, power-law shaped X-ray spectrum with a photon
index of $< 2$ and an exponential cutoff at $\approx$100 keV.
This spectral shape is generally described as unsaturated Compton
scattering of the soft X-ray photons from the accretion disk
by thermal electrons in hot inner flow or corona. Previous studies
suggested that the standard disk is truncated in the low/hard state
and eventually extends inward according to the increase of X-ray
luminosity \citep{tom09,shi11b,yam13,kol14}. However, in what
luminosity the standard disk reaches the ISCO is still controversial.
Further study is necessary to fully understand the inner disk structure
and its evolution in response to the change in mass accretion rate
in the low/hard state.
Fast time variability of BHXBs gives clues to reveal the structure
of the inner accretion disk and corona. In the low/hard state, the
power density spectra (PDSs) are roughly described by a strong
band-limited noise component, a flat-topped spectrum in the
$\nu P_\nu$ form between the low- and high-frequency breaks.
The low-frequency break moves to higher frequencies according
to the rapid rise of the X-ray flux. This suggests that
the break is associated with the truncation radius of the standard
disk \citep{ing12}. Extracting the energy spectra of the noise
component, \citet{axe05} found that the fast variability is produced
by Comptonized emission from the vicinity of the black hole.
Quasi-periodic oscillations (QPOs) are sometimes detected on
the band-limited noise. Previous studies reported that the observed
QPO frequencies are correlated to the spectral index and disk flux
\citep[e.g.,][]{sob00, tit04, sha07},
implying that the QPOs, as well as the band-limited noise, are
coupled with the inner disk structure.
H 1743$-$322 is a transient BHXB discovered by {\it Ariel-V} in 1977
\citep{kal77}. Since then, it has displayed more than 10 outbursts,
which were extensively observed by X-ray satellites.
This source exhibits outbursts recurrently with an interval
of $\approx$200 days \citep{shi12}. Some of them were ''failed''
outbursts, in which the high/soft state is missing
\citep[e.g.,][]{cap09, che10}.
High- and low-frequency QPOs were detected above 100 Hz and
below 10 Hz in many observations \citep[e.g.,][]{hom03, tom03, hom05}.
The X-ray spectra of H 1743$-$322 were also obtained many times in
various states, including wide-band data up to $\sim$100 keV taken with
{\it INTEGRAL} and {\it Suzaku} \citep{cap05, blu10}.
Ionized absorption lines, likely originated from disk winds, were
discovered in the high/soft state and the steep power-law
(or ''very high'') state in {\it Chandra} high-resolution spectra
\citep{mil06}. This suggests that the source has a high
inclination angle.
H 1743$-$322 was also observed in other wavelengths.
Optical and near-infrared counterparts were identified by
\citet{ste03} and \citet{bab03}, respectively. However, the
high Galactic extinction has hampered deeper monitoring
observations to measure the black hole mass dynamically.
Bright radio flares likely associated with relativistic jets
were detected in the early phase of the outbursts
\citep[e.g.,][]{rup03}. In the 2003 outburst, bipolar large-scale
X-ray jets were resolved with {\it Chandra} \citep{cor05}. Combining
the {\it Chandra} image with the radio data obtained with the {\it VLA}
during the same outburst, \citet{ste12} estimated the distance, the
inclination angle, and the spin parameter $a$ of the black hole
($a = cJ/GM_{\rm BH}^2$, where $J$ represents the angular momentum)
as $8.5 \pm 0.8$ kpc,
$75^\circ \pm 3^\circ$, and $0.2 \pm 0.3$, respectively.
They were derived by modeling the trajectories of
the bipolar jets with a symmetric kinematic model.
On 2012 September 24, the brightening of H 1743$-$322 was reported by
\citet{shi12} with {\it Monitor of All-sky X-ray Image} ({\it MAXI};
\citealt{mat09})/Gas Slit Camera (GSC; \citealt{mih11}). About 2 weeks
later, we triggered three sequential Target-of-Opportunity (ToO)
observations with {\it Suzaku} \citep{mit07}, when the X-ray flux
reached to the peak and started decreasing. Quasi-simultaneous
observations in the near-infrared ($J$, $H$, and $K_{\rm S}$) and
optical ($R$ and $i$) bands were also carried out with the
{\it Infrared Survey Facility} ({\it IRSF}) 1.4m telescope and the
{\it Kanata} 1.5m telescope at Higashi-Hiroshima Observatory,
respectively. The source was not detected in any bands, however.
In this paper, we report the results of spectral and timing studies
using the {\it Suzaku} datasets. Section 2 describes the detail of the
{\it Suzaku} observations and the data reduction. In Section 3, we
present the {\it Suzaku} light curves and power spectra. The spectral
analysis and the results are described in Section 4. The infrared and
optical observations are summarized in Section 5. Section 6 gives the
discussion of the {\it Suzaku} results and Section 7 is the summary and
conclusions. Errors represent the 90\% confidence range for a single
parameter of interest, unless otherwise specified.
Throughout the article, we refer to the table by \citet{and89} as
the Solar abundances. We assume the distance and inclination of
H 1743$-$322 as those determined by \citet{ste12}, $D=8.5$ kpc
and $i=75^\circ$.
\section{{\it Suzaku} Observations and Data Reduction}
\begin{deluxetable*}{lccc}
\tablecaption{Log of the {\it Suzaku} observations\label{tab_obslog_Suzaku}}
\tablewidth{0pt}
\tablehead{
\colhead{} & \colhead{Epoch-1} & \colhead{Epoch-2} & \colhead{Epoch-3}
}
\startdata
ObsID & 407005010 & 407005020 & 407005030 \\
start time (UT) & 2012 Oct. 4 18:46:29 & 2012 Oct. 10 15:30:34 & 2012 Oct. 12 09:43:11 \\
end time (UT) & 2012 Oct. 5 14:35:16 & 2012 Oct. 11 14:04:14 & 2012 Oct. 13. 08:59:13 \\
XIS window & 1/4 & 1/4 & 1/4 \\
burst option & 1.0 sec & 1.0 sec & 1.0 sec\\
net exposure & & & \\
XIS & 21 ksec & 21 ksec & 21 ksec \\
HXD & 42 ksec & 42 ksec & 43 ksec
\enddata
\end{deluxetable*}
\begin{figure}
\plotone{f1.eps}
\caption{Long-term light curves of H~1743$-$322 in
the 2--20 keV band obtained with
{\it MAXI}/GSC (top) and in the 15--50 keV band
from {\it Swift}/BAT (middle),
and the hardness ratio calculated from these two light curves (bottom). The
shadowed region indicate the {\it Suzaku} observations. MJD 56201 corresponds
to 2012 October 1.\label{fig_longtermLC}}
\end{figure}
{\it Suzaku} ToO observations of H 1743$-$322 were performed on 2012
October 4, 10, and 12, with a net exposure of $\approx$40 ksec in
each epoch. Figure~\ref{fig_longtermLC} plots the long-term X-ray
light curves of {\it MAXI}/GSC and {\it Swift}/Burst Alert Telescope
\citep[BAT;][]{bar05} and their hardness ratio, in which the {\it Suzaku}
epochs are marked with shadowed regions. The first observation
(hereafter ``Epoch-1'') is close to the peak of the BAT hard X-ray flux
in the 15--50 keV band and the second and third ones (``Epoch-2'' and
``Epoch-3'', respectively) are in the decaying phase, where the fluxes
are by $\approx$20\% lower than that of Epoch-1.
{\it Suzaku} carries three currently available X-ray
charge-coupled-devise (CCD) cameras named the X-ray Imaging Spectrometer
(XIS), which detect X-ray photons of 0.2--12 keV, and a hard X-ray
sensor called Hard X-ray Detector (HXD) composed of PIN diodes and GSO
scintillators covering the 10--70 keV and 40--600 keV bands,
respectively. The XIS consists of frontside-illuminated cameras
(FI-XISs; XIS-0 and XIS-3) and a backside-illuminated one (BI-XIS;
XIS-1), which is more sensitive to soft X-rays below $\approx$1.5 keV
than FI-XISs. In our observations, the 1/4 window mode and the 1.0 sec
burst option were employed for both FI- and BI-XISs to avoid pile-up
effects. A summary of the {\it Suzaku} observations is given in
Table~\ref{tab_obslog_Suzaku}.
We utilized HEAsoft 6.13 and {\it Suzaku} Calibration Database (CALDB)
released on 2013 March 5 for the reduction and analysis of our data. We
started with the ``cleaned'' events (whose process version is 2.8.16.34)
provided by the {\it Suzaku} team and followed the standard procedure
for the data reduction.
Source photons were extracted from a circular region with a radius of
110'' centering at the target position. Background events were collected
within a circle with a radius of 120'' taken in a blank-sky area outside
of the region for source extraction. The XIS response matrix and
ancillary response files were generated by {\tt xisrmfgen} and {\tt
xissimarfgen} \citep{ish07}, respectively. The non X-ray background
(NXB) was subtracted from the HXD data by using the modeled NXB files
produced by the HXD
team\footnote{http://www.astro.isas.ac.jp/suzaku/analysis/hxd/pinnxb/tuned/
for PIN, and http://www.astro.isas.ac.jp/suzaku/analysis/hxd/gsonxb/ for
GSO.}. In the spectral analysis, {\tt
ae\_hxd\_pinxinome11\_20110601.rsp} was used as the response file for
PIN, and {\tt ae\_hxd\_gsoxinom\_20100524.rsp} and {\tt
ae\_hxd\_gsoxinom\_crab\_20100526.arf} were as those for GSO.
Because H 1743$-$322 is located in a crowded region close to the center of
our Galaxy ($l,b = 357^\circ.26,-1^\circ.83$), the HXD spectra could be
contaminated with the Galactic diffuse X-ray background (GXB)
and nearby bright point sources. However, we find the
GXB contributes only $\approx$1.5\% of the total flux in the PIN bandpass,
and far smaller in that of GSO, which are negligible.
To estimate these values, we assumed the GXB spectrum at
$(l, b) = (28^\circ.5, 0^\circ.2)$ shown in \citet{kub10} and converted
the fluxes in the PIN and GSO energy ranges to those at the position
of H 1743$-$322 using the spatial profile of the GXB in the 16--70 keV band
obtained with {\it INTEGRAL}/IBIS \citep{kri07}. The Cosmic X-ray
background (CXB) is not subtracted from the HXD data, because its
contribution is even smaller than that of the GXB. We examined the
contamination of nearby sources using the {\it Swift}/BAT 70 month catalog
\citep{bau13}, considering the detector's transmission functions \citep{tak07}.
We also checked the variability of the sources through the {\it MAXI}/GSC and
{\it Swift}/BAT public light curves. The nearby sources were found to
contribute only $< 1\%$ to the observed PIN flux and $\approx$1\%
to that of GSO, both of which are ignorable.
\begin{figure*}
\plotone{f2.eps}
\caption{{\it Suzaku} XIS and HXD light curves of Epoch-1 (left), Epoch-2
(middle), and Epoch-3 (right). The shadowed regions are dipping periods.
\label{fig_suzakuLC}}
\end{figure*}
\section{X-ray Light Curves and Timing Properties}
Figure~\ref{fig_suzakuLC} presents the {\it Suzaku} light curves in the
three epochs. They sometimes display significant drops of the soft X-ray
flux in Epoch-2 and Epoch-3. These are likely ``absorption dips'', which
are thought to be caused by the obscuration of the central X-ray source
by the outer edge of the accretion disk hit by the accretion stream from
the companion star. As noticed in Figure~\ref{fig_lc_dip}, the
hardness ratio increases in these dipping periods. This indicates that
the dips are caused by photo-electric absorption, although the too
short dipping duration does not allow us to perform precise time-resolved
spectral analysis.
The presence of similar absorption dips in H 1743$-$322 was also
reported in previous observations \citep[e.g.,][]{hom05}. We define
the shadowed regions in Fig.~\ref{fig_suzakuLC} as the dipping periods,
in which the count rate in the 0.7--2 keV band is $\gtrsim$30\% smaller
than the averaged rate in the individual epochs, and exclude them in the
following analysis.
Figure~\ref{fig_psd} plots the normalized power density spectra
(PDSs) in the three epochs obtained with the PIN data.
The PDSs are dominated by a strong noise
components with a normalized power of $\approx 10^{-2}$ (rms/mean)$^2$
in the $\nu P_{\nu}$ form, consistent with those of typical BHXBs in the
low/hard state \citep[see e.g.,][]{mcc06}. A weak low frequency QPO is
also detected in each epoch. Fitting it with a Gaussian model, we
estimate the central frequency of the QPO as $0.206 \pm 0.003$ Hz in
Epoch-1, which become 1.4 times lower in Epoch-2 and Epoch-3 ($0.147 \pm
0.004$ Hz and $0.141 \pm 0.004$ Hz, respectively) as the {\it Swift/BAT}
flux is decreased by $\approx$ 20\%.
This QPO is detected in the XIS PDSs as well.
\begin{figure}
\epsscale{1.0}
\plotone{f3.eps}
\caption{{\it Suzaku} light curve and hardness ratio in Epoch-2 focused
on the dipping periods. (Upper) the XIS-0 light curve in the 0.7--2 keV band
binned in 64 sec. (Lower) The hardness ratio between the 2--10 keV and 0.7--2
keV bands. Similar hardening can also be seen in the dipping period of Epoch-3.
\label{fig_lc_dip}}
\end{figure}
\begin{figure}
\epsscale{1.0}
\plotone{f4.eps}
\caption{Normalized power density spectra (PDSs) in the three epochs created from
the PIN light curves with 0.1 sec bins. They are normalized in the way that their
integral gives the squared root mean squared fractional variability. White noise is
subtracted from all the PDSs.\label{fig_psd}}
\end{figure}
\section{Analysis of {\it Suzaku} Spectra}
We create the time-averaged spectra of the three epochs and analyze them
separately on XSPEC version 12.8.0. Considering the signal-to-noise
ratio and the reliability of the calibration, we use the data within
1--9 keV, 1--8 keV, and 14--70 keV for FI-XISs, BI-XIS, and PIN,
respectively. For the GSO data, we only consider the 50--130 keV (Epoch-1),
50--220 keV (Epoch-2), and 50--200 keV (Epoch-3) bands,
above which the signal
levels are overcome by the systematic errors in the simulated
background \citep[$\lesssim$1\% of the total count rate;][]{fuk09}. The
spectra and responses of FI-XISs (i.e., XIS-0 and XIS-3) are combined to
improve statistics. The XIS data of 1.7--1.9 keV are discarded to
avoid the uncertainties in the responses at energies around the
complex instrumental Si-K edge. To account for possible calibration
uncertainties, we add a 1\% systematic error to every spectral bin in the
XIS and HXD data, following \citet{mak08}. We find that the final
best-fit parameters are the same within their 90\% confidence ranges
even if we assign 0\%, 2\%, and 3\% systematic errors, and that the
confidence ranges themselves are increased only by $\lessapprox 10$\%
from the case of 0\% systematic error to that of 3\%.
We confirm that the conclusions are not affected.
The cross-normalization of the HXD with respect to FI-XISs was fixed at
1.16\footnote{http://www.astro.isas.ac.jp/suzaku/doc/suzakumemo/suzaku\\memo-2008-06.pdf},
while that of BI-XIS is left free. We examine the pileup effects in the
XIS data using the software {\tt aepileupcheckup.py}, which estimates
the pileup fraction at various radii from the center of the XIS point
spread function \citep{yam12}. We find that the pileup fraction is
negligibly small in all observations, less than 2\% even at the cores of
the PSFs. Figure~\ref{fig_spec} shows the time-averaged
spectra for the three epochs in the $\nu F_\nu$ form.
\begin{figure}
\plotone{f5.eps}
\caption{The time-averaged spectra in Epoch-1 (black, filled square), Epoch-2 (red, open circle),
and Epoch-3 (blue, cross). The spectral ratios of Epoch-2 and Epoch-3 with respect to Epoch-1
are plotted in the lower panel.\label{fig_spec}}
\end{figure}
\subsection{Dust-scattering Effects}
In observations of heavily absorbed sources (with $N_{\rm H} \gtrsim
10^{22}$ cm$^{-2}$), it must be borne in mind that dust scattering
in addition to photo-electric absorption can significantly affect the
observed soft X-ray spectra \citep{ued10,hor14}. Diffuse interstellar
(and circumbinary) dust can scatter-out the X-ray photons from the line
of sight and at the same time, scatter-in the photons emitted in
different directions. Consequently, we see a ''dust halo'' around a
point source. A dust halo was indeed resolved in GX 13$+$1
\citep{smi02}, which has almost the same column density ($N_{\rm H}
\approx 2 \times 10^{22}$ cm$^{-2}$) as our target. When we use
data that do not fully cover the dust halo,
those two scattering effects are not canceled out. In this
case, the observed spectra, particularly in the soft X-ray band, are
changed from what we would expect if there were no dust-scattering. This
must be taken into account for accurate spectral modeling. Even if all
the scattered photons are collected, the scattered-in and scattered-out
components no longer compensate each other in variable sources, because
of the difference in arrival time between the scattered photons and
those taking the straight path to the observer. In our case, the
typical delay time of the scattered photons in the extraction region for
the XIS is estimated to be $\sim 1.3$ day, by assuming a distance of 8.5
kpc. We confirm for all the {\it Suzaku} epochs that the flux
variability of H 1743$-$322 on that timescale is low enough to ignore
the variability of the scattered-in component, as in the case of
\citet{hor14}.
Following \citet{ued10} and \citet{hor14}, we utilize {\tt Dscat}, a
local multiplicative model implemented in XSPEC, to account for the
effects of dust scattering in the spectral analysis. {\tt Dscat}
estimates both scattering-in and -out components from the scattering
fraction (the fraction of the scattered-in photons included in the data)
and the hydrogen column density, by assuming the dust scattering cross
section for $R_{\rm V} = 3.1$ calculated by \citet{dra03}. We fix the
scattering fraction at 1.0 (i.e., all the scattered photons are
included) for the HXD, which has sufficiently large field of view to
cover the whole dust halo. For the XIS, we adopt the same value (0.65)
as \citet{hor14}, because we used the same region as theirs to extract
the source photons from the XIS data. In the following analysis, the
column density of the {\tt Dscat} model is always linked to that of the
interstellar photo-electric absorption, for which we employ {\tt phabs}
\citep{bal92}. We find that $\approx$30\%, $\approx$8\%, and
$\approx$3\% of the intrinsic flux from H1743$-$322 are reduced at 1
keV, 3 keV, and 5 keV by the effects of dust scattering,
respectively\footnote{Here $N_{\rm H} = 2 \times 10^{22}$ cm$^{-2}$ is
assumed.}.
\subsection{Modeling Time-averaged Spectra}
As noticed in Fig.~\ref{fig_spec}, the overall spectral shapes in the
three epochs are very similar to one another, although the Epoch-2 and
Epoch-3 spectra are slightly harder than that of Epoch-1. All of them
are roughly characterized by a hard power-law component with a photon
index of $\approx 1.6$, suggesting that the source is in the low/hard
state during the {\it Suzaku} observations. The hydrogen column density
is estimated to be $\approx$2.0 $\times 10^{22}$ cm$^{-2}$ for the
photo-electric absorption in cold interstellar medium. This value is
well within those obtained in previous studies of H 1743$-$322
\citep[1.6--2.3 $\times 10^{22}$ cm$^{-2}$; e.g.,][]{cap09,mil06}. A
high-energy rollover can be found in 50--100 keV, which is likely to
correspond to the electron temperature of the thermal corona. No
significant emission and absorption lines can be seen in the spectra.
We first fit the spectra with the {\tt nthcomp} model \citep{zdz96,
zyc99} with an interstellar absorption. The {\tt nthcomp} model
calculates a thermal Comptonization spectrum parametrized by a photon
index, an electron temperature of the Comptonizing corona ($kT_{\rm
e}$), and a seed-photon temperature. We could not constrain
$kT_{\rm e}$
from a joint fit with the XIS and HXD spectra, due to much
poorer statistics of the GSO data compared with those of the XIS that
dominate the total $\chi^2$ statistics.
We therefore determine $kT_{\rm e}$ only from
the GSO spectra\footnote{Here
the seed temperature is fixed at 0.1 keV to constrain $kT_{\rm e}$ so
that it does not affect the spectral shape in the GSO band.}, and fix
it at the best-fit value in fitting together with the XIS and PIN data.
From the joint fit of the
XIS$+$HXD spectra, we obtain $\chi^2/{\rm d.o.f.} = 1690/1158$
(Epoch-1), $1253/994$ (Epoch-2), and $1237/904$ (Epoch-3), which are far
from acceptable. As can be seen in the middle panel in
Fig.~\ref{fig_specfit}, significant residuals remain in the hard X-ray
band above 10 keV. A convex-shaped structure can be found, which is
likely originated from the reflection of Comptonized photons on the
disk.
We next apply a more sophisticated model to the spectra, considering the
general picture of the low/hard state. We use {\tt diskbb} model
\citep{mit84} as the direct emission from the standard disk and the {\tt
nthcomp} model as the Comptonization. The seed photon of the {\tt
nthcomp} component is assumed to be the multi-color disk emission, and
the seed temperature is tied to the inner disk temperature of the {\tt
diskbb} component. To account for the reflection component, the {\tt
ireflect} model \citep{mag95} is convolved to {\tt nthcomp}. This model
calculates a reflection spectrum using the ionization parameter ($\xi =
L_{\rm X}/(nR^2)$, where $L_{\rm X}$, $n$, and $R$ stand for the
incident X-ray luminosity, the electron number density and the distance
of reflector from the X-ray source, respectively), the temperature
($T_{\rm disk}$), inclination angle ($i$), and solid angle ($\Omega/2
\pi$) of the reflector. We fix $T_{\rm disk}$ at $30000$ K, which cannot
be constrained from the data.
The {\tt ireflect} model does not include any emission lines accompanied
by the reflection continuum. We thus combine a Gaussian component to
account for the iron K-$\alpha$ emission line, which is normally seen in
the low/hard state spectra most significantly. Considering numerical
studies \citep[e.g.,][]{mat91}, we link $\Omega/2 \pi$ of {\tt ireflect}
and the normalization of the Gaussian to keep the equivalent width with
respect to the reflection continuum at 1.0 keV.
We note that the equivalent width becomes more than an order of
magnitude smaller than what is expected in the numerical calculations,
when we unlinked the normalization of the Gaussian and the solid angle of
{\tt ireflect}.
The convolution model {\tt kdblur} \citep{lao91} is also incorporated to
model the relativistic blurring by the accretion disk around the
black hole. This model uses the index $\alpha$ describing the radial
dependence of the emissivity $\epsilon$ ($\propto r^{-\alpha}$), inner
and outer radii of the accretion disk, and the inclination angle. Here
we assume $\alpha = 3$ (i.e., a flat disk) and fix the inner radius at
$10 R_{\rm g}$ (where $R_{\rm g}$ is the gravitational radius, $GM_{\rm
BH}/c^2$), and the outer radius at $R_{\rm out} = 400$, which is the
maximum value of the {\tt kdblur} model. The integrated fitting model
is described as
{\tt phabs*Dscat*(diskbb+kdblur*(ireflect*nthcomp[re\\f]+gauss)+nthcomp[C])},
where {\tt nthcomp[ref]} and {\tt nthcomp[C]} represent the parts of
the Comptonized component that is and is not reflected on the disk,
respectively.
Because {\tt ireflect} and {\tt kdblur} are convolution models, we
extend the energy range to 0.01--1000 keV.
\begin{figure}
\plotone{f6a.eps}
\plotone{f6b.eps}
\plotone{f6c.eps}
\caption{The time-averaged spectra of Epoch-1 (top), Epoch-2 (middle),
and Epoch-3 (bottom) fitted with the {\tt nthcomp} model. The black
(below 9 keV), red (below 8 keV), green (14--70 keV), and blue
(above 50 keV) points correspond to the XIS-0$+$XIS-3, XIS-1, PIN, and
GSO data, respectively. The contribution of the scattered flux to the
XIS-1 data is shown in red dotted line. The data/model ratio for a single
{\tt nthcomp} model and the final best-fit model are plotted in the second
and third panels, respectively.\label{fig_specfit}}
\end{figure}
\begin{figure}
\plotone{f7a.eps}
\plotone{f7b.eps}
\plotone{f7c.eps}
\caption{The best-fit models in Epoch-1 (top), Epoch-2 (middle), and Epoch-3
(bottom) plotted in the $\nu F_\nu$ form.
\label{fig_specmodel}}
\end{figure}
The quality of fit is remarkably improved compared with that of the
single {\tt nthcomp} model, yielding $\chi^2/{\rm d.o.f.} = 1347/1155$
(Epoch-1), $984/991$ (Epoch-2), and $995/901$ (Epoch-3).
The spectra and the best-fit models in the individual epochs are plotted
in Figure~\ref{fig_specfit} and Figure~\ref{fig_specmodel},
respectively. The resultant parameters are given in
Table~\ref{tab_specfit}.
The inclusion of the relativistic blurring model to the reflection
component makes a significant statistical improvement ($\Delta \chi^2
=$ 70--90, where the degrees of freedom are not changed). When the
{\tt kdblur} component is excluded, we obtain $\chi^2/{\rm d.o.f.} =
1435/1155$, $1061/991$, and $1075/901$, for Epoch-1, -2, and -3,
respectively. Conversely, the addition of a narrow Gaussian
component at 6.4 keV with a 1-$\sigma$ width of 10 eV does not improve the
quality of fit.
Assuming an isotropic Comptonizated corona and the conservation of the total
number of photons from the accretion disk, the inner disk radius can be
derived from the formula in \citet{kub04};
\begin{eqnarray}
F^p_{\rm disk}+F^p_{\rm thc}2 \cos i &= 0.0165 \left[ \frac{r_{\rm in}^2\cos i}{(D/10\mbox{ kpc})^2}\right]
\left( \frac{T_{\rm in}}{1\mbox{ keV}} \right)^3 \nonumber \\
& \mbox{ photons } {\rm s}^{-1} \mbox{ }{\rm cm}^{-2}, \label{eq_Rin}
\end{eqnarray}
where $F^p_{\rm disk}$ and $F^p_{\rm thc}$ are the 0.01--100 keV photon
fluxes of the disk and thermal Comptonized components, respectively.
Extending the best-fit model to that energy range, we estimate
$F^p_{\rm disk} = 0.699$, $0.646$, and $0.489$ photons cm$^{-2}$ sec$^{-1}$
and $F^p_{\rm thc} = 1.27$, $1.06$, and $0.918$ photons cm$^{-2}$
sec$^{-1}$ in order from Epoch-1 to Epoch-3. The inner disk radii are
thus $r_{\rm in}=101^{+10}_{-9}$ $D_{8.5} (\cos i/\cos 75^\circ)^{-1/2}$ km,
$r_{\rm in}=119^{+24}_{-15}$ $D_{8.5} (\cos i/\cos 75^\circ)^{-1/2}$ km, and
$r_{\rm in}=92^{+19}_{-10}$ $D_{8.5} (\cos i/\cos 75^\circ)^{-1/2}$ km (where
$D_{8.5}$ represents the distance in units of 8.5 kpc). Here we only
include the 90\% confidence range of the inner disk temperature to
estimate the uncertainties of the radii. Multiplying a representative
correction factor (1.19) of boundary condition and spectral hardening
\citep{kub98}, the actual radii are calculated to be
$R_{\rm in} = 120^{+12}_{-11}$ $D_{8.5} (\cos i/\cos 75^\circ)^{-1/2}$ km,
$R_{\rm in} = 141^{+28}_{-17}$ $D_{8.5} (\cos i/\cos 75^\circ)^{-1/2}$ km, and
$R_{\rm in} = 110^{+23}_{-12}$ $D_{8.5} (\cos i/\cos 75^\circ)^{-1/2}$ km for
Epoch-1, -2, and -3, respectively.
The inner radius in the high/soft state is $R_{\rm in} = 64 \pm 2$
$D_{8.5} (\cos i/\cos 75^\circ)^{-1/2}$ km \citep{che10}, which
is estimated from the direct disk emission without Comptonization.
This value is taken from the first two rows of Table 2 in \citet{che10}
(which are originally from \citealt{mcc09} and \citealt{cap09}).
The fraction of the power-law component in the 2--20 keV flux is
less than 6\% for these two results and therefore $R_{\rm in}$
is changed by only a few \% by including the Comptonized disk
photons.
The inner radii obtained from the {\it Suzaku} data are 1.3--2.3 times
larger than the value in the high/soft state, suggesting that the
standard disk do not extend to the ISCO during the low/hard state
observations.
\begin{deluxetable*}{lcccc}
\tablecaption{Best-fit parameters of the {\tt nthcomp} model \label{tab_specfit}}
\tablewidth{0pt}
\tablehead{
\colhead{Component} & \colhead{Parameter} & \colhead{Epoch-1} & \colhead{Epoch-2}
& \colhead{Epoch-3}
}
\startdata
phabs & $N_{\rm H}$ ($10^{22}$ cm$^{-2}$) & $2.01^{+0.04}_{-0.06}$ & $1.99 \pm 0.04$
& $1.99^{+0.05}_{-0.06}$ \\
diskbb & $kT_{\rm in}$ (keV) & $0.28 \pm 0.02$ & $0.24^{+0.03}_{-0.02}$ & $0.27^{+0.02}_{-0.03}$ \\
& norm & $2.1^{+0.8}_{-1.0} \times 10^{3}$ & $3.1^{+2.7}_{-1.3} \times 10^{3}$
& $1.7^{+1.3}_{-0.8} \times 10^{3}$ \\
nthcomp & $\Gamma$ & $1.668^{+0.009}_{-0.005}$ & $1.633^{+0.009}_{-0.008}$ & $1.629 \pm 0.009$ \\
& $kT_{\rm e}$ (keV)\tablenotemark{a} & $< 35$
& $68^{+68}_{-20}$ & $61^{+144}_{-21}$ \\
& norm & $0.223^{+0.004}_{-0.006}$ & $0.174 \pm 0.003$ & $0.157 \pm 0.004$ \\
ireflect & $\Omega/2 \pi$ & $0.65^{+0.08}_{-0.12}$ & $0.56 \pm 0.07$ &
$0.55^{+0.08}_{-0.07}$ \\
& $\xi$ (erg cm sec$^{-1}$) & $< 6$ & $ < 31$ & $ < 68$ \\
& $T_{\rm disk}$ (K) & 30000 (fixed) & 30000 (fixed) & 30000 (fixed) \\
kdblur & $\alpha$\tablenotemark{b} & 3 (fixed) & 3 (fixed) & 3 (fixed) \\
& $R_{\rm in}$ ($R_{\rm g}$) & 10 (fixed) & 10 (fixed) & 10 (fixed) \\
& $R_{\rm out}$ ($R_{\rm out}$) & 400 (fixed) & 400 (fixed) & 400 (fixed) \\
& $i$ (deg) & 75 (fixed) & 75 (fixed) & 75 (fixed) \\
gauss\tablenotemark{c} & $E_{\rm cen}$ (keV) & 6.4 (fixed) & 6.4 (fixed) & 6.4 (fixed) \\
& $\sigma$ (keV) & 0.01 (fixed) & 0.01 (fixed) & 0.01 (fixed) \\
observed flux & 1--10 keV (ergs cm$^{-2}$ sec$^{-1}$) & $9.3 \times 10^{-10}$ & $7.5 \times 10^{-10}$
& $6.9 \times 10^{-10}$ \\
unabsorbed flux\tablenotemark{d} & 0.01--100 keV (ergs cm$^{-2}$ sec$^{-1}$) & $5.7 \times 10^{-9}$ &
$4.8 \times 10^{-9}$ & $4.4 \times 10^{-9}$ \\
& $\chi^2/{\rm d.o.f.}$ & $1347/1155$ & $984/991$ & $995/901$
\enddata
\tablenotetext{a}{The electron temperature is first estimated only from the GSO data
and it is fixed at the best-fit value in the joint fit with the XIS and HXD spectra.
Although the upper limit for Epoch-1 is not constrained, we adopted the
best-estimate value, $kT_{\rm e} = 57$ keV, in the simultaneous fit.}
\tablenotetext{b}{The emissivity index. The radial dependence of emissivity ($\epsilon$)
is expressed with $\epsilon \propto R^{-\alpha}$.}
\tablenotetext{c}{The normalization of the Gaussian is linked to the solid angle of
{\tt ireflect} component so that the equivalent width with respect to the reflection
continuum is always 1.0 keV.}
\tablenotetext{d}{The flux is corrected for dust-scattering.}
\end{deluxetable*}
It has been suggested that the Comptonization cloud in the
low/hard state has a complex structure. Using {\it Suzaku} broad-band
X-ray data, \citet{tak08, mak08, shi11b, yam13} found that the
time-averaged spectra can be described with two Comptonization
components that have different optical depths.
Moreover, \citet{yam13} detected the second Comptonization component by
analyzing the spectral variability. To test the ``double Compton'' model
in the H 1743$-$322 data, we add one more {\tt nthcomp} component and
its accompanying reflection component to our final model, and fit the
time-averaged spectra. We find, however, that the inclusion of the
second Comptonization component does not improve the quality of fit.
\subsection{Analysis of Short-term Spectral Variability}
We often find it difficult to accurately estimate the cool disk
component in the low/hard state by modeling time-averaged spectra,
because it is buried in the dominant Comptonized component and coupled
with the other structures such as the interstellar absorption and dust
scattering. In these cases, the analysis of the short-term spectral
variability can be a more powerful approach to separate the disk
component. The standard disk and Comptonization in the corona have
different properties of spectral variability on the $\sim$1-sec
timescale; the former generally shows significant variation, while
the latter is more stable \citep[e.g.,][]{chu01}.
Here we apply ''intensity-sorted spectroscopy'' described in \citet{mak08} and
\citet{yam13} to our {\it Suzaku} data. Using this technique, \citet{yam13}
successfully separated the cool disk component from the highly variable
Comptonization in Cyg X-1. We define the high- and low-intensity phases in
the same manner as theirs;
\begin{equation}
\{ t| C(t) > (1+f) \overline{C(t)_T} \}
\end{equation}
and
\begin{equation}
\{ t| C(t) < (1-f) \overline{C(t)_T} \},
\end{equation}
respectively, where $C(t)$ is the count rate at the time $t$ for an XIS
detector and $\overline{C(t)_T}$ is that averaged over $[t - T/2, t + T/2)$.
According to these criteria, we define the time intervals of high- and
low-intensity phases using XIS-1 light curves in the 1--10 keV band in
2.0 sec binning. We set $f = 0.2$ and $T = 64$ sec, so that we obtain
both the good photon statistics and a sufficiently high contrast of the
intensity between the two phases. The XIS-0$+$XIS-3, PIN, and GSO
spectra in these two phases are then extracted from the intervals
determined by XIS-1.
Thus, by using the independent dataset for the definition of the
time regions, we avoid selection biases merely caused by statistical
fluctuation in the photon counts in making the intensity-sorted spectra.
A summary of the intensity selection is given in Table~\ref{tab_isort_log},
and the XIS-1 light curves in low- and high-intensity phases are presented
in Figure~\ref{fig_lc_isort}.
The high and low-intensity spectra ($H(E)$ and $L(E)$, respectively) can be
separated to the constant component ($d(E)$) and variable components ($h(E)$
and $l(E)$) as
\begin{equation}
H(E) = \omega(E) (d(E)+h(E)) \label{eq_h}
\end{equation}
and
\begin{equation}
L(E) = \omega(E) (d(E)+l(E)), \label{eq_l}
\end{equation}
where $\omega(E)$ represents the photo-electric absorption (which can be
regarded as constant on the $\sim$1-sec timescale). Unlike \citet{yam13},
the constant component $d(E)$ in our data contains not only the
direct disk emission but also the scattered-in component,
whose short-term variability is smoothed out
due to the difference of the light traveling time of each scattered
photon. On the other hand, the observed variable component is composed
of the Comptonized photons without experiencing dust-scattering. For
simplicity, we assume the best-fit models obtained from the time-averaged
spectra as the scattered-in disk and Comptonization components, and subtract
them from both the high- and low-intensity spectra before calculating the
spectral ratio $H(E)/L(E)$.
Figure~\ref{fig_spec_ratio} plots the resultant spectral ratio in each
epoch. The overall shape is very similar to that of the Cyg X-1 spectra
in the low/hard state obtained in \citet{yam13}. A turnover below
$\sim 2$ keV can be seen, suggesting a significant suppression of
variability due to the contribution of constant disk component at low
energies. The spectral ratios of
XIS-0$+$XIS-3 without removing the scattered-in component are also
presented in pink for comparison. These demonstrate how significantly
dust scattering affect the results. We note that the turnover is intrinsic,
not generated by the dust scattering. The scattered-in component does not
have such a steep turnover, and the decline of variability is seen
in the spectral ratio including the scattered-in component as well.
Following the procedure of \citet{yam13}, we assume that the ratio
$H(E)/L(E)$ can be expressed with a single power-law ($\alpha E^{\beta}$)
in $E>2$ keV. We determine $\alpha$ and $\beta$ by fitting the data
in the 2--4 keV band (above which they may affected by the reflection
component). Using these parameters and Equation~\ref{eq_h} and \ref{eq_l},
we derive
\begin{equation}
\omega(E) d(E) = \frac{\alpha E^{\beta} L(E) - H(E)}{\alpha E^{\beta} - 1},
\end{equation}
where $d(E)$ is the direct disk spectra.
$d(E)$ is then fitted with {\tt phabs*Dscat*diskbb} to estimate the inner
temperature and the flux of the constant disk emission. Here the scattering
fraction of the {\tt Dscat} model is fixed at 0.0 (i.e., only the effect
of scattering-out is considered).
We assume $N_{\rm H} = 2.0 \times 10^{22}$ cm$^{-2}$, which is the averaged
value of those determined by the time-averaged spectra in the individual
epochs. However, the parameters are only very weakly constrained due to
the poor statistics in the
soft X-ray band around the turnover of the spectral ratios, although the values
are consistent with those estimated from the time-averaged spectra.
The fitting parameters are summarized in Table~\ref{tab_isort}.
\begin{figure}
\plotone{f8.eps}
\caption{The XIS-1 light curves of Epoch-1 (top), Epoch-2 (middle),
and Epoch-3 (bottom) in 1--10 keV for the high- (black) and
low-intensity (dark gray) phases. The bin size is 2.0 sec.\label{fig_lc_isort}}
\end{figure}
\begin{deluxetable}{lcccccc}
\tablecaption{Summary of the high- and low-intensity selection. \label{tab_isort_log}}
\tablewidth{0pt}
\tablehead{
\colhead{} & \multicolumn{2}{c}{Epoch-1}
& \multicolumn{2}{c}{Epoch-2} & \multicolumn{2}{c}{Epoch-3} \\
& \colhead{High} & \colhead{Low} & \colhead{High} & \colhead{Low} & \colhead{High} & \colhead{Low}
}
\startdata
frac. exp. (\%)\tablenotemark{a} & 24 & 27 & 24 & 28 & 25 & 29 \\
ave. rate (cnt sec$^{-1}$)\tablenotemark{b} &
27 & 18 & 22 & 15 & 20 & 14
\enddata
\tablenotetext{a}{The fraction of time that the source was in high-
and low-intensity phases with respect to the total exposure for each epoch.}
\tablenotetext{b}{Averaged XIS-0 count rates in 1--10 keV.}
\end{deluxetable}
\begin{figure}
\plotone{f9a.eps}
\plotone{f9b.eps}
\plotone{f9c.eps}
\caption{The ratios between the high-intensity and low-intensity spectra in
Epoch-1 (top), Epoch-2 (middle), and Epoch-3 (bottom). The ordinate axis in
each panel is logarithmically scaled.
The scattered-in components are already subtracted from the XIS spectra.
The same colors as Figure~\ref{fig_specfit} are used. The XIS-1 data,
which are used to determine the intervals of high- and low-intensity phases,
are not shown here, because they are affected by Poisson fluctuation of signal
counts. The XIS spectral ratio including the scattered-in component is also
plotted in pink (open circle).
\label{fig_spec_ratio}}
\end{figure}
\begin{deluxetable}{lccc}
\tablecaption{Results from the analysis of intensity-sorted spectra. \label{tab_isort}}
\tablewidth{0pt}
\tablehead{
\colhead{Parameter} & \colhead{Epoch-1} & \colhead{Epoch-2}
& \colhead{Epoch-3}
}
\startdata
\mbox{\boldmath $ h(E)/l(E)$}\tablenotemark{a} & & & \\
$\alpha$ & $2.1 \pm 0.1$ & $2.1 \pm 0.2$ & $2.0 \pm 0.1$ \\
$\beta$ & $-0.24 \pm 0.06$ & $-0.23 \pm 0.06$ & $-0.19 \pm 0.06$ \\
$\chi^2/{\rm d.o.f.}$ & $6.1/6$ & $8.1/6$ & $13.4/13$ \\ \hline
\mbox{\boldmath $\omega(E) d(E)$}\tablenotemark{b} & & & \\
$N_{\rm H}$ ($10^{22}$ cm$^{-2}$) & 2.0 (fixed) & 2.0 (fixed) & 2.0 (fixed) \\
scatf\tablenotemark{c} & 0.0 (fixed) & 0.0 (fixed) & 0.0 (fixed) \\
$T_{\rm in}$ & $0.22^{+0.40}_{-0.08}$ & $<0.33$ & $0.11^{+0.19}_{-0.06}$ \\
norm & $< 8 \times 10^{5}$ & $> 2 \times 10^3 $ & $< 1 \times 10^{12}$ \\
$\chi^2/{\rm d.o.f.}$ & $3.1/6$ & $4.2/3$ & $11.3/8$
\enddata
\tablenotetext{a}{$h(E)/l(E)$ is fitted with $\alpha E^{\beta}$ in the 2--4 keV band.}
\tablenotetext{b}{$\omega(E) d(E)$, calculated by using the best-fit values of
$\alpha$ and $\beta$, is fitted with {\tt phabs*Dscat*diskbb}.}
\tablenotetext{c}{The scattering fraction in the {\tt Dscat} model.}
\end{deluxetable}
\section{Near-infrared and optical observations and the results}
In addition to the X-ray observations with {\it Suzaku}, we performed
infrared photometric observations of H~1743$-$322 in the $J$ (1.25 $\mu$m),
$H$ (1.63 $\mu$m), and $K_{\rm s}$ (2.14 $\mu$m) bands, with the SIRIUS
camera \citep{nag03} on the 1.4m {\it IRSF} telescope
at South African Astronomical Observatory (SAAO)
during the 2012 outburst. Optical observations through $R_{\rm C}$ and
$I_{\rm C}$ filters were also carried out by using HOWPol
\citep[Hiroshima One-shot Wide-field Polarimeter;][]{kaw08} attached to
the 1.5m {\it Kanata} telescope at Higashi-Hiroshima Observatory. The
counterpart of H 1743$-$322 was not detected in any bands, however. The
typical seeing was $\approx$ 1''.8 for the {\it IRSF} $J$ band
and 2''.3 for the {\it Kanata} $R_{\rm C}$ band in full width at
half maximum.
In the analysis of the {\it IRSF} data, we combine all the frames taken
in one night and perform dark-subtraction, flat-fielding,
sky-subtraction, and combining the dithered images using the SIRIUS
pipeline software running on IRAF \citep[Image Reduction and Analysis
Facilities;][]{tod86} version 2.16. H 1743$-$322 is located in a
crowded region close to the Galactic center, and the near-infrared
counterpart in the {\it IRSF} images is buried in nearby bright
sources. We estimate the upper flux in each band from the counts on the
pixel at the position of H 1743$-$322. Count-flux conversion is made by
comparison with the stellar photometry of nearby sources with the 2MASS
magnitude \citep{skr06}. We reduce the {\it Kanata} data in a standard
way for the aperture photometry of optical CCD data. The upper
magnitudes in $R_{\rm C}$ and $I_{\rm C}$ band are determined by the
detection limits of a point source in the image, estimated from the
background level of source-free regions near H 1743$-$322.
Figure~\ref{fig_SED} presents the upper limits of the optical and
near-infrared fluxes on 2012 October 12, together with the
contemporaneous {\it Suzaku} spectrum (in Epoch-3).
The flux limits in the optical and near-infrared bands are corrected
for the Galactic extinction. Substituting the hydrogen column density
of H 1743$-$322 ($N_{\rm H} \approx 2 \times 10^{22}$ cm$^{-2}$) into the
$N_{\rm H}$ versus $A_{\rm V}$ relation
\citep[where $A_{\rm V}$ represents the extinction in the $V$ band;][]{pre95},
we derive $A_{\rm V} \approx 11$. We convert $A_{\rm V}$
to the extinction in the $R_{\rm C}$, $I_{\rm C}$, $J$, $H$, and
$K_{\rm S}$ bands using the extinction curve for $R_{\rm V} = 3.1$
given by \citet{car89}.
The upper limits of the extinction-corrected absolute
$R_{\rm C}$, $I_{\rm C}$, $J$, $H$, and $K_{\rm S}$ magnitudes
are estimated to be $-1.9$, $-0.5$, $-3.8$, $-5.2$, and $-5.8$,
respectively.
We find that if the companion is a main sequence star,
it should be a late-B or a less massive star \citep{wai92}.
The upper limits are consistent with the intrinsic disk flux including
the Compton-scattered photons (shown in red in Fig.~\ref{fig_SED}),
and with the possible contribution of the synchrotron emission from
the compact jet. If the positive relation between radio and X-ray fluxes
found by \citet{cor11} holds in the 2012 outburst, the {\it Suzaku} flux
in Epoch-3 ($5.2 \times 10^{-10}$ erg cm$^{-2}$
sec$^{-1}$in the 3--9 keV band) corresponds to 0.68 mJy at 8.5 GHz flux.
Assuming that the flat optically-thick synchrotron spectrum
extends up to the optical bands, we estimate the jet contribution to be
$\nu F_\nu \approx 4 \times 10^{-12}$ erg cm$^{-2}$ sec$^{-1}$ at
$5 \times 10^{14}$ Hz.
This is more than $\approx$10 times lower than the optical and near-infrared
upper flux limits.
\begin{figure}[tbp]
\plotone{f10.eps}
\caption{The spectral energy distribution of H 1743$-$322 on 2012 October 12.
The upper limits in the extinction-corrected
optical ($R_{\rm C}$ and $I_{\rm C}$) and near-infrared
($J$, $H$, and $K_{\rm S}$ bands) fluxes
are plotted in blue arrows. The black points
are the {\it Suzaku} spectra, corrected for neutral absorption and dust
scattering. Red solid line shows the intrinsic disk emission including the
Comptonized photons, where the outer disk is assumed to extend to infinity.
\label{fig_SED}}
\end{figure}
\section{Discussion}
\subsection{Failed Outburst in 2012 October}
We observed H 1743$-$322 with {\it Suzaku} at the flux peak and early
decaying phase during the outburst in 2012 September and October. The
time-averaged spectra in all the three epochs are approximated by
a power-law with a photon index of $\approx$1.6, which is
well within the typical values in the low/hard state
\citep[1.5--2.0; e.g.,][]{don07}. The PIN power spectra in
the 14--70 keV band show strong variability above $\approx$0.1 Hz with a
power of $1 \times 10^{-2}$ (rms$^2$/Hz$^2$), which most probably
corresponds to the flat part of the band-limited noise. These X-ray
spectral and timing properties support that the source was in the
low/hard state during the {\it Suzaku} observations.
As noticed in Fig.~\ref{fig_longtermLC}, the long-term MAXI data do
not show significant softening around the peak flux. This suggests
that the state transition from the low/hard state to the high/soft
state (hard-to-soft transition) did not occur in the late 2012 outburst.
This type of outburst is often classified as a ``failed outburst'', which
exhibits no hard-to-soft transition or an incomplete transition only
reaching the intermediate state.
A few failed outbursts have been reported in H 1743$-$322 so far
\citep{cap09,che10}. In addition to this source, there are few
other BHXBs that underwent failed outbursts, like XTE J1550$-$564
and MAXI J1836$-$164 \citep{stu05,fer12}.
The 1--500 keV unabsorbed luminosity in Epoch-1 (calculated
from the best-fit model) is estimated to be $6.1 \times 10^{37}$
$D_{8.5}^2$ erg sec$^{-1}$.
This is a few times lower than the luminosity at which H 1743$-$322
generally makes a hard-to-soft transition
\citep[e.g.,][]{mcc09,zho13}, suggesting that the mass
accretion rate was not enough to trigger a transition.
\subsection{Structure of Inner Disk and Corona}
The time-averaged {\it Suzaku} spectra are well described with thermal
emission from the standard disk, its Comptonized emission by a hot
corona, and the reflection component from the accretion
disk. The overall spectral shape is not largely changed in the three epochs,
while the 2--20 keV fluxes in Epoch-2 and Epoch-3 are
$\approx$20\% lower than that in Epoch-1. Although the inner disk
temperature and the photon index of the Comptonization are decreased by
10--30\% and 0.03 from Epoch-1 to the latter two epochs, the other
parameters remain the same within their 90\% confidence ranges. We
find that the double Comptonization model is not necessary to reproduce
the spectra. The same was suggested with the {\it Suzaku} spectrum of
MAXI J1305$-$704 in the low/hard state at $\sim 0.01 L_{\rm Edd}$
\citep{shi13}.
In the case of Cyg X-1, the softer Comptonization component is more clearly
seen in the brighter low/hard (or hard intermediate) states \citep{yam13}.
Thus, the luminosities of H 1743$-$322 in {\it Suzaku} observations
($\approx 0.04$ $D_{\rm 8.5}^2 M_{\rm 10}^{-1}$ $L_{\rm Edd}$ in the 1--500 keV
band) may not be sufficiently high to detect the softer Comptonization
component significantly.
As noticed in Fig.~\ref{fig_spec}, the GSO spectra suggest
the evidence of the high-energy rollover around 50--100 keV, likely
correspond to the electron temperature of the Comptonized corona.
From the spectral fit with the GSO data, the electron temperature
is estimated to be $kT_{\rm e} \approx 60$ keV.
We note, however, that the value may be somewhat overestimated
because the {\tt nthcomp} model does not include relativistic
effects properly; its output spectrum drops more sharply above
the rollover than the that of the real Comptonization spectrum
\citep[see][]{don10}.
Previous observations of H 1743$-$322 with {\it INTEGRAL} and
{\it Suzaku} at similar luminosities reported the detection of
the cutoff energy \citep[e.g.,][]{cap05, blu10}.
The anti-correlation between the electron temperature and the
X-ray luminosity in the bright ($\lesssim 0.1 L_{\rm Edd}$)
low/hard state was suggested by \citet{miy08} with the {\it RXTE}
\citep[Rossi X-ray Timing Explorer;][]{bra93} spectra of GX 339$-$4.
\citet{chi10} also found in the {\it RXTE} data of Swift J1753.5$-$0127
that the electron temperature is relatively low in the bright low/hard
state like in our case, and that it moves to higher energies in
accordance with the decline of an outburst.
It would be explained if inverse Compton cooling in the corona is
more efficient by a relatively higher input rate of seed photons
in brighter periods \citep{miy08}.
We find that the X-ray spectra are dominated by the Comptonized
component and that the direct disk emission is very weak in the
{\it Suzaku} bandpass. The disk temperature estimated from the
time-averaged spectra is much smaller than typical values in
the high/soft state \citep[$\approx$1 keV; e.g.,][]{mcc06}.
We calculate the inner disk radius from the intrinsic disk flux
including the disk photons consumed by Compton-scattering in the
corona in each epoch. The radii are found to be 1.3--2.3 times larger
than that in the high/soft state, supporting the idea that the standard
disk does not extend to the ISCO in the low/hard state. This is
consistent with the somewhat smaller solid angle of the reflection
component than those obtained in the very high state
by using similar spectral models \citep{tam12, hor14}.
We note, however, the calculated inner radii should be
taken with caution, since this is somewhat dependent on spectral
modeling. We considered only the statistical uncertainties of
the inner disk temperatures, but additional errors may be
produced in modeling the effects of dust scattering.
Also, the spectral hardening factor of disk emission for
deriving the actual inner radii could be different between
the high/soft and low/hard states \citep[e.g.,][]{shi95}.
Moreover, we assumed that all the seed photons are originated
from the disk, but if other emission components contribute,
the inner radii could become smaller.
For more precise estimation of the inner disk radii, we
have to assess these possible uncertainties thoroughly
by using higher-quality wide-band spectra, particularly
from the sources with a much smaller absorption column.
This is left for future work.
We successfully detect the weak, cool disk component independently of
the time-averaged spectral modeling, using the spectral variability on a
short ($\approx$1 sec) timescale. The overall profile of spectral ratio
between high- and low-intensity phases is very similar to what
\citet{yam13} found in Cyg X-1 data during the low/hard state. The power
of short-term variability clearly declines below 1--2 keV, suggesting
that the constant standard disk component contributes to the soft X-ray
flux. Modeling the profile of the spectral ratio in the same manner as
\citet{yam13}, we estimate the inner radius and temperature of the disk
component. Although the resultant parameters have large uncertainties
due to poor statistics, they are consistent with those obtained from the
time-averaged spectra.
The mass accretion rate of the disk can be derived through
$L_{\rm disk} = GM_{\rm BH} \dot{m}_{\rm disk}/(2R_{\rm in})$, where
$L_{\rm disk}$ and $\dot{m}_{\rm disk}$ represent the luminosity and
mass accretion rate of the disk component, respectively. We estimate
$L_{\rm disk} = 2.5 \times 10^{36}$ $D_{8.5}^2$ erg sec$^{-1}$ (the
unabsorbed disk luminosity in the 0.01--500 keV band) and
$R_{\rm in} = 120$ $D_{8.5} (\cos i/\cos 75^\circ)^{-1/2}$ km (for
Epoch-1) and thus we have $\dot{m}_{\rm disk} = 4.5 \times 10^{16}$
$D_{8.5}^3$ $M_{10}^{-1} (\cos i/\cos 75^\circ)^{-1/2}$ g s$^{-1}$.
The accretion rate of corona can be calculated as $L_{\rm c}
= \eta \dot{m}_{\rm c} c^2$, where $\eta$, $L_{\rm c}$, and
$\dot{m}_{\rm c}$ are the radiative efficiency,
the luminosity and the mass accretion rate of the corona, respectively.
Using the best-fit {\tt nthcomp} component for Epoch-1 as $L_{\rm c}$
($5.1 \times 10^{37}$ $D_{8.5}^2$ erg sec$^{-1}$),
we obtain $\dot{m}_{\rm c} = 5.7 \times 10^{17}$ $D_{8.5}$ g s$^{-1}$
if the corona is radiatively efficient ($\eta = 0.1$ is assumed).
It can be much more larger in the case of a radiatively inefficient
flow. Thus, $\dot{m}_{\rm c}$ is much larger than
$\dot{m}_{\rm disk}$, regardless of the radiative efficiency. This is
inconsistent with the simple truncation disk model, in which the
standard disk makes transition at the truncated radius into the
hot inner flow with the same mass accretion rate.
A possible explanation is that there may be a separate coronal
accretion flow with a large scale height extending from the outer
edge of the disk to the black hole, and that its mass accretion rate
is much larger than that of the thin disk.
There is another possibility, which reconciles the mass accretion
rates of the disk and corona. If the soft X-ray component is not the
disk itself, but is instead originated from a much smaller, hot clumps
torn off the inner edge of a much larger disk \citep{chi10, yam13}, the
true disk component has larger normalization but is truncated at a
larger radius. In this case, $\dot{m}_{\rm disk}$ can be comparable to
$\dot{m}_{\rm c}$, and the true inner disk temperature can be much lower,
making impossible to detect the disk component in the soft X-ray bandpass.
The cool disk component peaking in the EUV band was actually detected in a
low/hard state spectrum of XTE J1118$+$480 \citep{esi01}, along
with a much smaller soft X-ray component \citep{rei09}, which would
be from the clumps \citep{chi10}.
\subsection{Ionized Absorber and Disk Wind}
\citet{mil06} discovered blue-shifted, ionized absorption lines at 6.7
keV and 7.0 keV (corresponding to He-like and H-like iron K$\alpha$
lines, respectively) in the Chandra high-resolution spectra of
H~1743$-$322 in the high/soft and very high (or soft intermediate)
states. These lines, often seen in high inclination sources
\citep[e.g.,][]{ued98, kot00, lee02}, are likely to be originated from
the disk wind. They show ``state-dependent'' behavior and are rarely
detected in the low/hard state \citep[e.g.,][]{nei09,pon12}. For H
1743$-$322, non detection of the iron absorption lines in the low/hard
state was previously reported by \citet{blu10} and \citet{mil12}.
Our {\it Suzaku} data do not exhibit any significant absorption features
apparently. To check the presence of the ionized iron absorption lines,
we add two negative Gaussian components to the best-fit model for the
time-averaged spectra. Here we assume a wind velocity of
300 km sec$^{-1}$ and a 1-$\sigma$ line width of 20 eV (which can be
regarded as representative values of \citealt{mil06}) and fix all the
parameter except for the normalizations of continuum and line
components. We estimate the 90\% upper limits of the equivalent widths
in Epoch-2 to be 22~eV for both H-like and He-like iron lines, which
are similar or slightly larger values than those obtained in \citet{mil06}.
Thus, we cannot exclude the presence or disappearance of the ionized
iron absorption lines.
\subsection{Origin of QPO}
A weak low-frequency (LF) QPO is detected at 0.1--0.2~Hz in each
epoch. We find that the QPO frequency becomes $\approx$30\% lower in
accordance with the $\approx$20\% decrease of the hard X-ray luminosity
in the 15--50~keV band.
LF QPOs (below 10~Hz) were detected in many BHXBs during the low/hard
and intermediate states, such as XTE J150$-$564
\citep[e.g.,][]{cui99,sob99} and GX 339$-$4
\citep[e.g.,][]{miy91,mot11,tam12}. Previous studies suggested that
their frequencies are positively correlated with the photon index of the
Comptonization below $\approx$10~Hz, at which the correlation saturates
\citep{vig03,tit04,tit05}. Our results follow this ubiquitous
correlation between the QPO frequency and the photon index; the photon
index becomes by $\approx$0.03
lower in the $\approx$30\% decrease of the QPO frequency.
The correlations of the observed LF-QPO frequencies with the X-ray flux
and the photon index lead us to invoke the idea that they are associated
with the inner disk structure. \citet{ing09} suggested the possibility that
the LF QPOs are originated in the Lense-Thirring precession of the hot
inner flow extended between the black hole and the inner edge of the
disk. Their model can describe both the QPO frequencies and the
observed X-ray spectrum, and predicts the anti-correlation with the
inner disk radius and the QPO frequency.
In the truncation disk model, the inner radius recedes as the mass
accretion rate is decreased \citep[see][]{don07}. At the same time,
the power of seed photons illuminating the hot flow becomes weaker,
making the observed X-ray spectrum harder. The decrease of the LF-QPO
frequency can be explained if it reflects the timescale at the inner
radius, where the standard disk is replaced to the hot flow \citep{ing09}.
\citet{axe14} extracted the spectrum of the pure QPO component from
the RXTE data of XTE J1550$-$564 and found that it can indeed be
interpreted by the Comptonization in the hot inner flow. Although the
inner disk radii that we estimated from the {\it Suzaku} data in each
epoch have large uncertainties, they are compatible to the
anti-correlation trend between the QPO frequency and the inner disk.
The QPO frequencies correspond to 40--50 $R_{\rm g}$ in the
Lense-Thirring precession model of \citet{ing09}, if a black hole
mass of $10 M_\sun$ is assumed. This is about several to ten times
larger than what we estimated from the time-averaged spectra.
The possible reason of this inconsistency would be that the QPO
was generated by a more complicated mechanism, or that we are
underestimating the true inner disk radius, as is also suggested
in Section 6.2 by assuming that the mass accretion rate through
the disk and that in the coronal flow are the same.
A separate, small variable soft component could simultaneously
allow the QPO to be consistent
with a Lense-Thirring origin, and for the corona to be fed directly
by the mass accreting through the disk.
\section{Summary and Conclusion}
We observed the black hole transient H 1743$-$322
with {\it Suzaku} on three occasions during the outburst in 2010
October, which provided one of the best quality broad band X-ray
spectra of this source in the low/hard state.
The results are summarized as follows:
\begin{enumerate}
\item We find that the observed X-ray spectra are significantly affected
by dust scattering in the interstellar medium with $N_{\rm H} \approx 2
\times 10^{22}$ cm$^{-2}$, which must be corrected for accurate spectral
analyses.
\item The time-averaged spectra are dominated by a strong Comptonization
in the corona with an electron temperature of $\approx$60~keV.
The disk reflection component with $\Omega/2\pi \approx 0.6$ is
detected. The estimated inner disk radius is larger than that
in the high/soft state, suggesting that the standard disk is truncated
at 1.3--2.3 times larger radii than the ISCO during the {\it Suzaku}
observations.
\item We estimate the cool disk component by investigating short-term
spectral variability on the $\sim$1-sec timescale, independently of the
time-averaged spectral modeling. We find that the spectral ratio between
high- and low-intensity phases has a very similar shape to those of Cyg
X-1 in the low/hard state, indicating the presence of stable disk
emission at low energies.
\item A weak low-frequency QPO was detected at 0.1--0.2~Hz. We find that
the QPO frequency becomes lower as the X-ray luminosity and photon index
decrease. This may be explained by the evolution of the disk truncation
radius.
\end{enumerate}
\acknowledgments
We appreciate the {\it Suzaku} operation team for arranging and carrying out
the ToO observations. We thank Takayuki Yuasa for his help in examining the
contamination of Galactic ridge X-ray emission. We are grateful to the
{\it Kanata} team, Hiroshima University, for performing the optical
observations. We also thank
Kohji Tsumura, Mai Shirahata, Sudhanshu Barway, Daisuke Suzuki, and
Fumio Abe for carrying out the {\it IRSF} observations. We
appreciate Tatsuhito Yoshikawa for his help in analysing the {\it IRSF} data.
This research has made use of the MAXI data provided by RIKEN, JAXA and
the MAXI team. We utilized data products from the Two Micron All Sky Survey,
which is a joint project of the University of Massachusetts and the Infrared
Processing and Analysis Center/California Institute of Technology, funded by
the National Aeronautics and Space Administration and the National Science
Foundation. This work is partly supported by a Grant-in-Aid for JSPS Fellows
for Young Researchers (MS) and for Scientific Research 23540265 (YU).
{\it Facilities:} \facility{{\it Suzaku}}, \facility{{\it IRSF}},
\facility{{\it Kanata}}.
|
\section{Supplementary Information}
As introduced in the main text, we use RF induced
evaporation on magnetically trapped $^{87}$Rb atoms to reach the
BEC. The magnetic trap is provided by the chip structures and,
before its loading, we perform optical pumping in the $m_F=2$
magnetic sublevel. This condensation scheme imply that the initial
state of the condensate is the $|F=2,m_F=2\rangle$ state, so as
initial condition every state preparation pulse uses a system in
that pure state. After the BEC creation we let expand the atomic
cloud in the homogeneous magnetic field for $0.7$ ms and then, using
a programmable function generator, we apply to a wire on the chip
the signal $V(t)=V_0 \sin[t \omega(t)]$, which build up the driving
field. In the frequency range $[4000,4700]$ Hz in which we constrain
$\omega(t)$ the radiation efficiency of the used wire can be
considered constant, so the amplitude $V_0$ of the signal can be
determined with high accuracy by performing a series of atomic
population oscillation to have a Rabi frequency $\Omega=2 \pi \;60$
kHz.
To calibrate the homogeneous magnetic field applied by the Helmholtz
coils, we close a multi-state atom interferometer with constant
amplitude pulses as in \cite{PetrovicNJP} from which we deduce the
right current to apply to the coils to produce the value
$B_{th}=6.179 \ G$. To compensate for very slow variations of the
magnetic field we repeat the calibration procedure before the
sequence of ten measurements of a given state preparation pulse.
From a statistical analysis of the interferometer output we also
derive information on the residual noise present in the field. In
particular, by the reduction of the fringe contrast we deduce the
dephasing rate interval $\gamma \in 2 \pi \; \left[20,200\right]$,
while by the variations of the fringe phase we can estimate the
amount of shot-to-shot offset field fluctuations, to $1$ mG.
Introducing these parameters in the theoretical model it is possible
to derive also a maximal theoretical error by taking the worst and
the best output of the model evolution. Those errors are reported in
table \ref{table2}.
To observe the evolution during the optimization pulse, we splice it
in subpulses, measuring the output at the end of each of them.
Figure \ref{SI_Fig1} shows an example of experimental points and
theoretical prediction, reported in solid line, relative to the five
population $m_F=+2 ... -2$ driven from the initial state to the
ground state of the Hamiltonian $H_0+H_1+\overline{V}(t)$ with
$\overline{V}(t)$ being the harmonic potential that oscillates at
$f(t)=\overline{f} = 2\pi \; 4323$ kHz. The optimized preparation
pulse is $20 \mu s$ long and, after the preparation, we keep the
frequency of the potential at the constant target value. The
stability of the populations shows that the system is really in the
target state, with the right coherence between the $|F,m_F\rangle$
components.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth,angle=0]{SI_Fig2}
\caption{
Time evolution of the optimally shaped $f(t)/(2\pi)$ for the preparation of the ground state and the excited state of the Hamiltonian $H_0+H_1+\overline{V}(t)$ with $\overline{V}(t)$ being the harmonic potential that oscillates at $f(t)=\overline{f}=\ 2\pi\; 4323$ kHz. The optimally shaped pulse (gray area) is followed by a constant one for $t \geq T = 20 \ \mu s$.}
\label{SI_Fig2}
\end{figure}
Finally we show in Fig. \ref{SI_Fig2} the optimal pulses for the preparation of the ground and excited states as in Fig. \ref{SI_Fig1}.
Similar pulses (with $n_f=7$) have been obtained as a result of our optimization scheme to prepare the other states discussed in the main text.
|
\section{Introduction}
It is believed that primordial black holes (PBHs) could have formed in the early universe from the collapse of large density fluctuations, and if so, could have observational implications - either from their gravitational effects, or the effects of their Hawking radiation (see \cite{Carr:2009jm,Josan:2009qn} for recent lists of the constraints). They have not been observed, but this fact is enough that they can be used to constrain the early universe (i.e. \cite{Bugaev:2012ai,Young:2013oia,Green:1997sz,Byrnes:2012yx,Shandera:2012ke}) - and provide the only known tool for probing the primordial universe on extremely small scales (i.e. \cite{Scott:2012kx}). However, the constraints from PBHs on small scales are much weaker than those on cosmological scales, for example, the constraints from the cosmic microwave background from Planck.
During inflation, the Hubble horizon shrinks on a comoving scale, and quantum fluctuations become classical density perturbations once they exit the horizon. Once inflation ends, the horizon begins to grow and perturbations begin to reenter the horizon. If a perturbation is large enough, it will collapse to form a PBH almost immediately after horizon reentry - and there has been extensive research into the nature of this collapse and how large a perturbation must be in order to collapse \cite{Shibata:1999zs,Hawke:2002rf,Niemeyer:1999ak,Musco:2004ak,Musco:2008hv,Musco:2012au,Niemeyer:1997mt}.
Calculations for the critical value of the density contrast, $\Delta$, or comoving curvature perturbation, $\mathcal{R}_{c}$, above which a region will collapse to form a PBH are typically of order 0.5 or 1 respectively - and so an insignificant number of PBHs will form unless the power spectrum on small scales is much larger than on large scales, by several orders of magnitude. This is possible in several models, such as the running mass model \cite{Drees:2011hb}, axion inflation \cite{Bugaev:2013fya}, a waterfall transition during hybrid inflation \cite{Bugaev:2011wy}, from passive density fluctuations \cite{Lin:2012gs}, or during inflation with small field excursions \cite{Hotchkiss:2011gz}. For a recent summary of PBH forming models see \cite{Green:2014faa}. Alternatively, the constraint on the formation criteria can be relaxed during a phase transition in the early universe, causing PBHs to form preferentially at that mass scale (i.e. \cite{Jedamzik:1999am}).
In this paper, we will review the calculation of the PBH abundance. The calculation typically computes the fraction of the universe which is above the critical value - in terms of $\Delta$ or $\mathcal{R}_{c}$. This is typically done using the theory of peaks, which calculates the number density of peaks above the critical value, or a Press-Schechter approach, which computes the volume of the universe above the critical value. in order to calculate the abundance of PBHs on different scales, the distribution is convolved with a smoothing function to smooth out modes smaller than the horizon, whilst leaving the horizon and super-horizon modes. When $\mathcal{R}_{c}$ is used to do the calculation in this manner, the super-horizon modes have a large impact on the calculation - we will argue that they should not affect the calculation and that using $\mathcal{R}_{c}$ can be misleading and give errors of many orders of magnitude compared to using $\Delta$.
In Section 2, we will discuss the formation criteria for PBHs explaining these arguments, and in Section 3, we will briefly review the calculation of the mass of a PBH dependant on the scale it forms at. In Section 4 we discuss the different ways the abundance of PBHs and constraints on the early universe can be calculated for different models. We conclude our arguments in Section 5.
\section{Formation criteria}
The abundance of PBHs is normally stated in terms of $\beta$, the mass fraction of the Universe contained within PBHs at the time of their formation. Typically, $\beta$ is given as a function of their mass (which, we will see later, is a function of the time at which they form) - so that $\beta$ can be used to describe the mass spectrum of PBHs. In order to determine whether a region of the early universe will collapse to form a PBH, then typically either the density or curvature of that region is compared to a threshold value, which itself is typically calculated from numerical simulations.
Traditionally, the density contrast $\Delta=\frac{\delta\rho-\rho}{\rho}$ had been used to calculate $\beta$. However, following the paper by Shibata and Sasaki in 1999 \cite{Shibata:1999zs} which calculated the threshold value in terms of a metric perturbation $\psi$, and the paper by Green, Liddle, Malik and Sasaki (GLMS) in 2004 \cite{Green:2004wb}, it became more common to use the comoving curvature perturbation $\mathcal{R}_{c}$ (for example, \cite{Bugaev:2012ai,Shandera:2012ke})\footnote{The comoving curvature perturbation $\mathcal{R}_{c}$ is equal to the curvature perturbation on uniform density slices $\zeta$ on super-horizon scales, and because sub-horizon modes are smoothed out, it is common to use $\zeta$ instead of $\mathcal{R}_{c}$.}.
In figure \ref{dontusezeta} we demonstrate the danger of using $\mathcal{R}_{c}$ to calculate $\beta$. By simply comparing the height of either peak, one would be drawn to the conclusion that the first (left hand) peak will collapse to form a PBH and the second (right hand) peak will not. However, because the long wavelength mode is well outside the horizon, it is unobservable at the expected time of collapse and invoking the separate universe approach (see \cite{Wands:2000dp}) means that it should not affect the local evolution of the universe. Therefore, the universe looks locally identical to observers at either peak - either both peaks should collapse to form a PBH or neither should\footnote{Note that we are assuming that a PBH will form shortly after entering the horizon, or not at all. It is possible for the PBH formation process to last several e-foldings after horizon entry \cite{Musco:2012au} in which case the long wavelength mode will become important, but only for values extremely close to the threshold value - although this is thought to be rare, see equation (\ref{close to critical}) (however, the effect of a perturbation sitting inside a much larger scale perturbation has not been well studied).}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{dontusezeta}
\caption{Here, as an example, we show a universe with two sharp (gaussian) peaks in $\mathcal{R}_{c}$ which sit on top of a long wavelength mode. The two thick black boxes represent the size of the visible universe to an observer at the centre of the peaks at the time of PBH formation, whilst the dotted red line represents the hypothetical threshold value for collapse. Both universes appear the same locally to each observer, and so the evolution of each patch should be identical (until the long wavelength becomes observable).}
\label{dontusezeta}
\end{figure}
It should be noted that papers which have calculated a critical value in terms of $\mathcal{R}_{c}$
(i.e. \cite{Shibata:1999zs,Nakama:2013ica}) assume that $\mathcal{R}_{c}$ drops quickly to zero outside of the perturbation - so these values can be used if one assumes that there are no super-horizon perturbations affecting your calculation. Therefore it may be possible to use $\mathcal{R}_{c}$ to calculate $\beta$ if one takes care to exclude super-horizon modes from the calculation (one possibility is to simply subtract the long wavelength modes - although this is strongly dependant on what is considered to be a long wavelength.), and in Section 4.5 we will consider an approximation where only the value of the power spectrum at horizon entry is used.
A more formal way to consider this to investigate the effect of super-horizon modes on local observables, such as the density contrast and the spatial 3-curvature. Figure \ref{curvature density} shows the same universe as figure \ref{dontusezeta} but in terms of the spatial curvature and density contrast.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{curvaturedensity}
\caption{The same universe as shown in figure \ref{dontusezeta}, but this time showing the spatial curvature and the density contrast at the time the scale of the small peaks enter the horizon. We now see that both peaks look identical - and so should evolve in the same manner. We see that the peaks in the spatial curvature and density contrast are very similar, both having a Mexican hat profile (rather than the gaussian shape in the comoving curvature perturbation) - note that the difference in the height of the peaks is due to the arbitrary scaling we have used in the figure.}
\label{curvature density}
\end{figure}
\emph{Spatial curvature} - consider the perturbed, spatially flat FRW metric
\begin{equation}
\label{metric}
ds^{2}=-N^{2}dt^{2}+g_{ij}\left(dx^{i}+N^{i}dt\right)\left(dx^{j}+N^{j}dt\right); g_{ij}=e^{2\alpha}\delta_{ij},
\end{equation}
where we have chosen a comoving slicing, and
\begin{equation}
\alpha=\ln a(t) + \mathcal{R}_{c},
\end{equation}
with $a(t)$ the scale factor of some flat background and $\mathcal{R}_{c}$ the comoving curvature perturbation. A constant value of $\mathcal{R}_{c}$ can be absorbed into the scale factor by defining
\begin{equation}
\bar{a}(t)=a(t)e^{\mathcal{R}_{c}},
\end{equation}
and so a constant $\mathcal{R}_{c}$ corresponds only to a rescaling of the spatial coordinates, as perhaps clear from the form of the metric (\ref{metric}). The spatial curvature is given by
\begin{equation}
R^{(3)}=-\frac{2}{e^{2\alpha}}\delta^{ij}\left(2\alpha,_{ij}+\alpha,_{i}\alpha,_{j}\right),
\end{equation}
and the spatial curvature of the metric is then
\begin{equation}
\label{spatial curvature}
R^{(3)}=-\frac{2}{e^{2\alpha}}\left(2\nabla^{2}\mathcal{R}_{c}+(\vec{\nabla}\mathcal{R}_{c})^{2}\right).
\end{equation}
If we consider a very long wavelength $\mathcal{R}_{c}$ mode, which appears constant on horizon scales, we see that the spatial curvature due to this mode is negligible due to the derivatives in Eq. (\ref{spatial curvature}).
\emph{Density contrast} - on comoving slices, there is a simple relation at linear order between the comoving curvature perturbation and the density contrast \cite{Green:2004wb}
\begin{equation}
\label{density contrast}
\Delta(t,k)=\frac{2(1+\omega)}{5+3\omega}\left(\frac{k}{aH}\right)^{2}\mathcal{R}_{c}(k),
\end{equation}
where $\omega$ is the equation of state $\omega=p/\rho$, which during radiation domination is $\frac{1}{3}$.\footnote{Josan, Green and Malik \cite{Josan:2009qn} derive an alternative formula valid on super- and sub-horizon scales during radiation domination,
\begin{equation}
\Delta(t,k)=-\frac{4}{\sqrt{3}}\left(\frac{k}{aH}\right)j_{1}\left(\frac{k}{\sqrt{3}aH}\right)\mathcal{R}_{c}(k),
\end{equation}
where $j_{1}$ is a spherical Bessel function. However, after smoothing, there is little difference between this and equation (\ref{density contrast}).} In real space this is
\begin{equation}
\Delta(t,x)=\frac{2(1+\omega)}{5+3\omega}\left(\frac{1}{aH}\right)^{2}\nabla^{2}\mathcal{R}_{c}(x).
\end{equation}
Again, we see that this depends on the second derivative of $\mathcal{R}_{c}$ - and so the effect of super-horizon $\mathcal{R}_{c}$ modes is negligible. At linear order, the density contrast is therefore equivalent to the spatial curvature. However, there has been extensive research into the threshold value of $\Delta$ but not for $R^{(3)}$, we therefore advocate the use of the density contrast in order to calculate the mass fraction, $\beta$.
There has been extensive research on the threshold value for the density contrast above which a PBH will form. Carr \cite{Carr:1975qj} was the first to derive a threshold value for the formation of PBHs, $\Delta_{c}\approx\omega$ where $\omega$ is the equation of state, by calculating the density necessary for gravity to overcome pressure forces. In recent years, numerical simulations of gravitational collapse have been used to investigate the collapse of different shapes of the initial density profile. Niemeyer and Jedamzik \cite{Niemeyer:1999ak} studied initial shapes including gaussian, Mexican hat, and polynomial, finding $\Delta_{c}\approx0.7$. Musco et al \cite{Musco:2004ak,Musco:2008hv,Musco:2012au}\footnote{Musco et al note that the difference in value obtained by Niemeyer and Jedamzik can be explained because they only considered a pure density perturbation imposed at the time of horizon crossing. Later work included only growing modes accounting for the effect of the perturbation in the velocity field.} later studied PBH formation, finding $\Delta_{c}\approx0.45$. More recently, Harada et al \cite{Harada:2013epa} studied a top hat shape, finding an analytic formula $\Delta_{c}=\sin^{2}[\pi\sqrt{\omega}/(1+3\omega)]=0.41$ during radiation domination, and Nakama et al \cite{Nakama:2013ica} studied generalised shapes to determine the crucial parameters in the shape and size of an overdensity. See also \cite{Hawke:2002rf}. \footnote{It was previously thought that there was an upper bound above which density perturbations would form a separate closed universe rather than a PBH, however, this has been shown not to be the case \cite{Kopp:2010sh}. This is relatively unimportant in practice, as the effect of an upper bound is negligible because higher peaks are exponentially suppressed.}.
\section{Primordial black hole mass}
In order to calculate the mass spectrum, or mass function, of PBHs, it is necessary to relate the horizon scale at the time of formation to the mass of PBH formed. We will first review the calculation of the horizon mass carried out by GLMS \cite{Green:2004wb}. The horizon mass is
\begin{equation}
M_{H}=\frac{4\pi}{3}\rho(H^{-1})^{3}.
\end{equation}
In co-moving units, the horizon scale during radiation domination is $R=(aH)^{-1}\propto a$, and expansion at constant entropy gives $\rho\propto g_{*}^{-1/3}a^{-4}$ (where $g_{*}$ is the number of relativistic degrees of freedom, which is expected to be of order 100 in the early universe). This allows the horizon mass at a given reentry scale to be related to the horizon mass at matter radiation equality,
\begin{equation}
M_{H}=\frac{3}{2}M_{H,eq}(k_{eq}R)^{2}\left(\frac{g_{*,eq}}{g_{*}}\right)^{1/3},
\end{equation}
where we use $k_{eq}=0.07\Omega_{m}h^{2}$Mpc$^{-1}$, $g_{*,eq}\approx 3$ and $g_{*}\approx 100$. $M_{H,eq}$ is given by
\begin{equation}
M_{H,eq}=\frac{4\pi}{3}2\rho_{rad,eq}H_{eq}^{-3}=\frac{8\pi}{3}\frac{\rho_{rad,0}}{k_{eq}^{3}a_{eq}},
\end{equation}
where we take $a_{eq}^{-1}=24 000\Omega_{m}h^{2}$ and $\Omega_{rad,0}h^{2}=4.17\times10^{-5}$. Taking $\Omega_{m}h^{2}=0.14$ gives $M_{H,eq}=7\times10^{50}g$ (for this calculation, we have used the same numbers as GLMS \cite{Green:2004wb}).
Now that the horizon mass has been calculated, it remains to determine fraction of the horizon mass which goes into the PBH, $f_{H}$. Several papers (for example, \cite{Hawke:2002rf,Niemeyer:1999ak}) have noted that, when the density is close to the critical value, the mass of PBH formed depends on the size of the over-density, obeying a simple power law,
\begin{equation}
f_{H}=C\left(\Delta-\Delta_{c}\right)^{\gamma},
\end{equation}
where $C$ and $\gamma$ are constants - although the values calculated depend on the shape of the initial over-density. Chisholm \cite{Chisholm:2006qc} summarises the different measurements, as well as discussing a minimum bound on the PBH mass from entropy constraints. Typical values for these parameters which we will consider here are $C=3$, $\Delta_{c}=0.5$, and $\gamma=0.3$. For these values, the mass of PBH formed is only significantly smaller than than the horizon mass, $M_{PBH}<0.1M_{H}$, for values of $\Delta$ in the range
\begin{equation}
0.5<\Delta<0.500012,
\label{close to critical}
\end{equation}
and so we will assume that PBHs form with a mass approximately equal to the horizon mass for the remainder of this paper. PBHs of significantly larger mass could form in regions where $\Delta$ is substantially larger than 0.5, but the abundance of these regions is exponentially suppressed, and are thus extremely rare.
\section{Primordial black hole abundance}
We will now discuss the calculation of the PBH mass fraction, $\beta$. The density contrast on a comoving slicing, $\Delta$, is smoothed on a given scale $R$, and the fraction of the universe with a density contrast above the critical value is calculated. The smoothed density contrast $\Delta(R,x)$ is calculated by convolving the density contrast with a window function $W(R,x)$:
\begin{equation}
\Delta(R,x)=\int_{-\infty}^{\infty}d^{3}x' W(R,x-x')\Delta(x').
\end{equation}
The variance of $\Delta(R,x)$ is given by
\begin{equation}
\label{density variance}
\langle\Delta^{2}\rangle=\int_{0}^{\infty}\frac{dk}{k}\tilde{W}^{2}(R,k)\mathcal{P}_{\Delta}(k),
\end{equation}
where $\tilde{W}(R,k)$ is the fourier transform of the window function, and $\mathcal{P}_{\Delta}(k)$ is the density power spectrum. Using equation (\ref{density contrast}) this can be related to the comoving curvature perturbation power spectrum as,
\begin{equation}
\langle\Delta^{2}\rangle=\int_{0}^{\infty}\frac{dk}{k}\tilde{W}^{2}(R,k)\frac{4(1+\omega)^{2}}{(5+3\omega)^{2}}\left(kR\right)^{4}\mathcal{P}_{\mathcal{R}_{c}}(k).
\end{equation}
Throughout this paper, we will use a volume-normalised gaussian window function, such that the fourier transform is given by
\begin{equation}
\label{window function}
\tilde{W}(R,k)=\exp\left(-\frac{k^{2}R^{2}}{2}\right).
\end{equation}
In the remaining portion of this section, we discuss the difference between using a peaks theory or Press-Schechter approach, and the predicted mass spectra of PBHs for a scale invariant curvature spectrum, a power law spectrum and for a spectrum with a running of the spectral index.
\subsection{Peaks theory vs Press-Schechter}
The initial mass fraction of the Universe $\beta$, that went into PBHs can be calculated either using a peaks theory approach, or a Press-Schechter approach. A comparison of these two methods was carried out by GLMS \cite{Green:2004wb} who compared the mass spectra calculated using the curvature perturbation, with peaks theory, and the density contrast, using a Press-Schechter approach. In their calculation it was necessary to assume a blue primordial power spectrum, $n_{s}>1$, and they found the two to be in close agreement\footnote{In the appendix, we correct their calculation, finding that calculating $\beta$ in the different methods disagree strongly.}. We will repeat the calculation here for the density contrast only - finding that using peaks theory or a Press-Schechter are not in as close agreement previously found in \cite{Green:2004wb} but still similar, to within a factor of order 10.
To investigate the difference between the two methods, we will use a variable $\nu=\Delta/\sigma$\footnote{We note here that with peaks theory, the critical value is stated in terms of the peak value of a fluctuation, but in a Press-Schechter approach, it is the average value of the fluctuation. The relationship between the peak value and the average depends on the shape of the fluctuation - but typically, these are expected to differ only by a factor of order unity, with the peak value being higher. The difference in the critical value of the peak value and the average is therefore within the error of the predicted critical value from different sources. We also note the fact that looking for peaks above a certain value in a smoothed distribution is equivalent to looking for patches with an average density above that value - and so the distinction here is only a technical note.}, where $\sigma$ is the square root of the variance $\langle\Delta^{2}\rangle$ given by equation (\ref{density variance}) and is a function of the form of the power spectrum and the smoothing scale (the calculation of $\sigma$ is the same for either method).
In the theory of peaks, the number density of peaks above a height $\nu_{c}$ is given by \cite{Bardeen:1985tr}
\begin{equation}
n_{peaks}(\nu_{c},R)=\frac{1}{(2\pi)^{2}}\left(\frac{\langle k^{2}\rangle(R)}{3}\right)^{\frac{3}{2}}\left(\nu_{c}^{2}-1\right)\exp\left(-\frac{\nu_{c}^{2}}{2}\right),
\label{peaks}
\end{equation}
where $\langle k^{2}\rangle$ is the second moment of the smoothed density power spectrum
\begin{equation}
\langle k^{2}\rangle(R)=\frac{1}{\langle\Delta^{2}\rangle(R)}\int_{0}^{\infty}\frac{dk}{k}k^{2}\tilde{W}^{2}(k,R)\mathcal{P}_{\Delta}(k).
\end{equation}
If we assume a power law spectrum $\mathcal{P}_{\mathcal{R}_{c}}=A_{\mathcal{R}_{c}}(k/k_{0})^{n_{s}-1}$, and a gaussian window function (equation (\ref{window function})), we obtain
\begin{equation}
\langle k^{2}\rangle(R)=\frac{n_{s}+3}{2R^{2}},
\end{equation}
assuming that $n_{s}>-3$. The number density of peaks above the threshold can be related to the density parameter $\Omega_{PBH,peaks}$ (which is equal to the mass fraction $\beta$ for a flat universe) by $\Omega_{PBH,peaks}(\nu_{c})=n_{peaks}(\nu_{c},R)M(R)/\rho$, where $M(R)$ is the mass of PBH associated with the horizon size $R$, $M(R)=(2\pi)^{3/2}\rho R^{3}$. Finally, we have
\begin{equation}
\beta_{peaks}(\nu_{c})=\Omega_{PBH,peaks}(\nu_{c})=\frac{(n_{s}+3)^{3/2}}{6^{3/2}(2\pi)^{1/2}}\nu_{c}^{2}\exp\left(-\frac{\nu_{c}^{2}}{2}\right).
\label{beta peaks}
\end{equation}
By contrast, the Press-Schechter calculation simply integrates the probability distribution function (PDF),
\begin{equation}
P(\nu)=\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{\nu^{2}}{2}\right),
\end{equation}
over the range of values that form a PBH:
\begin{equation}
\beta_{PS}(\nu_{c})=2\int_{\nu_{c}}^{\infty}P(\nu)d\nu=2\int_{\nu_{c}}^{\infty}\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{\nu^{2}}{2}\right)d\nu.
\end{equation}
This can be written in terms of the complimentary error function simply as
\begin{equation}
\beta_{PS}(\nu_{c})=\textrm{erfc}\left(\frac{\nu_{c}}{\sqrt{2}}\right),
\end{equation}
and using the asymptotic expansion of erfc$(\nu_{c})$ this can be written as
\begin{equation}
\beta_{PS}(\nu_{c})\approx\sqrt{\frac{2}{\pi}}\frac{1}{\nu_{c}}\exp\left(-\frac{\nu_{c}^{2}}{2}\right).
\end{equation}
Figure \ref{peaks vs PS} shows the difference in the predicted values of $\beta$ for either calculation - the two are in relatively close agreement (differing by a factor of order 10), whilst $\nu$ is not too large\footnote{However, this uncertainty in $\beta$ has little effect on the uncertainty of $\nu_{c}$ which would be calculated, as it depends only on $\log(\beta)$ (see \cite{Young:2013oia}).}. For larger values of $\nu_{c}$, $\beta_{peaks}$ is systematically higher than $\beta_{PS}$. However, the difference between these methods is small compared to the error due to uncertainties in the threshold value $\Delta_{c}$ (see figure \ref{spectral index} for an example).
\begin{figure}[t]
\centering
\includegraphics[width=0.7\linewidth]{peaksPS}
\caption{Here we compare the value of $\beta$ calculated using peaks theory or Press-Schechter against $\nu_{c}=\frac{\Delta_{c}}{\sigma}$.}
\label{peaks vs PS}
\end{figure}
\subsection{Scale invariant power spectrum}
In the case where the primordial curvature power spectrum is scale invariant, $\mathcal{P}(k)=A_{\mathcal{R}_{c}}$, where $A_{\mathcal{R}_{c}}$ is a constant, then the variance of the smoothed density field during radiation domination, $\omega=1/3$, is
\begin{equation}
\langle\Delta^{2}\rangle=\int_{0}^{\infty}\frac{dk}{k}\tilde{W}^{2}(k,R)\frac{4(1+\omega)^{2}}{(5+3\omega)^{2}}A_{\mathcal{R}_{c}}
=\frac{8}{81}A_{\mathcal{R}_{c}}.
\label{scale invariant density variance}
\end{equation}
Note that, as expected for a scale invariant spectrum, this is now independent of the smoothing scale $R$
- and so predicts that $\beta$ is independent of the mass of the PBHs\footnote{It is also worth noting that for either a red or scale invariant power spectrum $\langle\mathcal{R}_{c}^{2}\rangle\rightarrow\infty$.}. Using peaks theory:
\begin{equation}
\beta=\frac{1}{2^{3/2}(2\pi)^{1/2}}\frac{81\Delta_{c}^{2}}{8A_{\mathcal{R}_{c}}}\exp\left(-\frac{81\Delta_{c}^{2}}{16A_{\mathcal{R}_{c}}}\right).
\label{scale invariant beta}
\end{equation}
\subsubsection{Constraints on the power spectrum}
Using the relation between the (scale invariant) comoving curvature perturbation power spectrum and $\beta$, equation (\ref{scale invariant beta}), it is simple to calculate a constraint on the power spectrum from the constraint on $\beta$ at a given scale. We will here consider a constraint of size $\beta<10^{-20}$, with $\Delta_{c}=0.5$, and give the constraints one would calculated from peaks theory and Press-Schechter, seeing that the two are in very close agreement:
\begin{align}
\mathcal{P}_{\mathcal{R}_{c},peaks}<0.026, \nonumber \\
\mathcal{P}_{\mathcal{R}_{c},PS}<0.029.
\label{constraints}
\end{align}
\subsection{Power law power spectrum}
In order to compare with the GLMS paper \cite{Green:2004wb}, we will consider a power law spectrum (see also Drees and Erfani \cite{Drees:2011hb}). The form of the power spectrum is given by
\begin{equation}
\mathcal{P}_{\mathcal{R}_{c}}(k)=A_{0}\left(\frac{k}{k_{0}}\right)^{n_{s}-1},
\end{equation}
where $A_{0}$ is the amplitude of the power spectrum defined on some pivot scale $k_{0}$, and we will consider only blue spectra, $n_{s}>1$. In this case, the variance of the smoothed density field during radiation domination, given by equation (\ref{density variance}) is
\begin{equation}
\langle\Delta^{2}\rangle=\frac{8}{81}\frac{A_{0}}{(k_{0}R)^{n_{s}-1}}\Gamma\left(\frac{n_{s}+3}{2}\right),
\end{equation}
and $\beta$ is given by equation (\ref{beta peaks}). For the purposes of making a specific calculation we will take $A_{0}=2.2\times10^{-9}$ and $k_{0}=0.05$ Mpc$^{-1}$, loosely based on observations. Figure \ref{spectral index} shows the predicted mass spectra for a range of different spectral indexes $n_{s}$, and threshold values of the density contrast $\Delta_{c}$ - here, we only consider a blue spectrum (it is possible to consider a red spectrum on small scales in which case $\beta$ is larger for more massive PBHs, but a complicated model is needed to produce a significant number of PBHs and be consistent with observations).
We can place a limit on the spectral index from the observational constraints on the abundance of PBHs - as has been done previously (for example, \cite{Green:1997sz}). Taking $\Delta_{c}=0.5$ and using the constraint $\beta<10^{-20}$ for PBHs in the mass range $10^{8}$g$<M_{PBH}<10^{10}$g \cite{Josan:2009qn}, the constraint on the spectral index is $n_{s}<1.34$. Because there is a minimum mass of PBHs, at the Planck mass, then we can also place a minimum value on $n_{s}$ which is required to form a significant number of PBHs. Approximately 70 efoldings of inflation are required after todays horizon scale exited during inflation in order for the horizon to reach a sufficiently small scale corresponding to the Planck mass. Typical inflationary models predict that the current horizon scale exited the Hubble scale during inflation about 55 efoldings before the end of inflation \cite{Liddle:2003as}. In that case, the mass contained in the horizon scale at the end of inflation is approximately $e^{30}M_{\rm Planck}\sim 10^8$g. If we require that $\beta>10^{-20}$ for PBHs of mass $M_{PBH}=10^{-5}$g then the spectral index must be $n_{s}>1.26$. In order for a significant number of PBHs to form, then $n_{s}$ must lie in the range
\begin{equation}
1.26<n_{s}<1.34.
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{spectralindex}
\caption{This figure shows the predicted PBH mass spectra for different values of $n_{s}$ and $\Delta_{c}$. A smaller spectral index produces PBHs of smaller masses. Note that the calculation has been artificially cut off when $\beta$ becomes large as it is only valid for rare peaks (where $\beta$ is small), as well as for PBHs smaller than the Planck mass ($M\approx 10^{-5}g$).}
\label{spectral index}
\end{figure}
\subsection{Running of the spectral index}
Over the large range of scales considered here, the spectral index is unlikely to be a constant. We will therefore consider a running of the spectral index, $\alpha$, defined as
\begin{equation}
\alpha=\frac{dn_{s}}{d\ln(k)},
\end{equation}
leading to an expression for the comoving curvature perturbation power spectrum given by
\begin{equation}
\mathcal{P}_{\mathcal{R}_{c}}(k)=A_{0}\left(\frac{k}{k_{0}}\right)^{n_{0}-1+\frac{1}{2}\alpha\ln(k/k_{0})},
\end{equation}
where $A_{0}$ and $n_{0}$ are the values of the power spectrum and spectral index respectively, defined at a pivot scale $k_{0}$. If values are given for parameters $k_{0}$, $A_{0}$, $n_{0}$ and $\alpha$ then the PBH mass spectra can be calculated as before, calculating the variance of the smoothed density contrast using equation (\ref{density variance}) and finding $\beta$ using equation (\ref{beta peaks}).
The same as in the previous section, we will take $A_{0}=2.2\times10^{-9}$ and $k_{0}=0.05$ Mpc$^{-1}$. The Planck collaboration \cite{Ade:2013uln} found a spectral index $n_{s}=0.9603\pm0.0073$, but no statistically significant running of the spectral index, $\alpha=-0.0134\pm0.0090$. We will therefore take $n_{0}=0.96$ and allow $\alpha$ to vary - see figure \ref{running}. A positive running is necessary to produce a significant number of PBHs, and the smallest value we will consider is $\alpha=0.01$.
PBHs of masses greater than $M_{PBH}\approx 10^{8}$g are well constrained by observations \cite{Josan:2009qn,Carr:2009jm}, and we see from figure \ref{running} that these values of the running produce too many PBHs, and would be ruled out by observational constraints. We therefore state an upper bound on the running of the spectral index, $\alpha<0.0162$ (again, using the constraint $\beta<10^{-20}$ for PBHs in the mass range $10^{8}$g$<M_{PBH}<10^{10}$g \cite{Josan:2009qn}). Although, again, we note that there is no reason to assume the running of the spectral index will be constant over a large range of scales.
We will not consider the running of the running in this paper, although it has been considered by Erfani \cite{Erfani:2013iea}, who places an upper limit on the running of the running by considering the non-production of (long lived) PBHs.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{running}
\caption{This figure shows the predicted PBH mass spectra for different values of the running of the spectral index $\alpha$. Again, the calculation has been artificially cut off when $\beta$ becomes large.}
\label{running}
\end{figure}
\subsection{Approximation using the comoving curvature perturbation power spectrum}
The power spectrum is, formally, the variance of the amplitude of the Fourier modes at a certain scale. Less formally, one can consider it to be the characteristic size of perturbations at that scale. We show in this section that one can quickly find an approximate value for the PBH mass fraction using the comoving curvature perturbation by only considering perturbations at the exact scale of horizon crossing, without using window functions - this is the approach used in previous papers \cite{Byrnes:2012yx, Young:2013oia}. At horizon crossing, the relation between the density contrast and the comoving curvature perturbation becomes even simpler, as the factor $(k/aH)=1$:
\begin{equation}
\Delta(t_{H},k)=\frac{2(1+\omega)}{5+3\omega}\mathcal{R}_{c}(k)=\frac{4}{9}\mathcal{R}_{c}(k),
\end{equation}
where $t_{H}$ is the time at horizon entry, and $\omega=1/3$ is the equation of state during radiation domination. As $\Delta$ is proportional to $\mathcal{R}_{c}$ at horizon entry, it is reasonable to assume that peaks in the smoothed density contrast correspond to peaks in the comoving curvature perturbation (ignoring other scales).
We will assume that the power spectrum at a given scale gives the variance of the comoving curvature perturbation at that scale and use a Press-Schechter approach to calculate $\beta$:
\begin{equation}
\beta=2\int_{\mathcal{R}_{c,crit}}^{\infty}P(\mathcal{R}_{c})d\mathcal{R}_{c},
\end{equation}
where $P(\mathcal{R}_{c})$ is the (gaussian) probability distribution function. Writing this in terms of the complimentary error function gives
\begin{equation}
\beta=\textrm{erfc}\left(\frac{\mathcal{R}_{c,crit}}{\sqrt{2 \mathcal{P}_{\mathcal{R}_{c}}}}\right).
\label{zeta approx}
\end{equation}
Compare this to the expression one would derive using the density contrast for a scale invariant power spectrum, where $\langle\mathcal{R}_{c}^{2}\rangle$ is given by equation (\ref{scale invariant density variance}),
\begin{equation}
\beta=\textrm{erfc}\left(\frac{9\Delta_{c}}{4\sqrt{\mathcal{P}_{\mathcal{R}_{c}}}}\right).
\label{beta spectral index}
\end{equation}
These two expressions will be exactly equal if $\Delta_{c}\approx\frac{2\sqrt{2}}{9}\mathcal{R}_{c,crit}$. However, these methods cannot be considered identical, which is evident if a power law spectrum is considered, $\mathcal{P}_{\mathcal{R}_{c}}(k)=A_{0}(k/k_{0})^{n_{s}-1}$. Equation (\ref{zeta approx}) is unchanged, but equation (\ref{beta spectral index}) becomes
\begin{equation}
\beta=\textrm{erfc}\left(\frac{9\Delta_{c}}{4\sqrt{\mathcal{P}_{\mathcal{R}_{c}}\Gamma{\left(\frac{3+n_{s}}{2}\right)}}}\right).
\end{equation}
However, provided that $\Gamma\left(\frac{3+n_{s}}{2}\right)\approx 1$ (which is satisfied if $n_{s}\approx 1$) and $\Delta_{c}=\frac{2\sqrt{2}}{9}\mathcal{R}_{c,crit}$, these two expressions will be approximately equal. Figure \ref{zeta approx and density} shows a specific example of these calculations, showing that they still agree closely.
We now compare the constraints on the power spectrum calculated in this method to the constraints calculated earlier (equation (\ref{constraints})). Using $\mathcal{R}_{c,crit}=1.2$ \cite{Shibata:1999zs,Green:2004wb}, and $\beta<10^{-20}$ gives the constraint
\begin{equation}
\mathcal{P}_{\mathcal{R}_{c}}<0.024,
\end{equation}
which is in close agreement with the previously calculated bound, equation (\ref{constraints}).
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{zetaapproximation}
\caption{We show the mass spectra of PBHs calculated, for a power law power spectrum $\mathcal{P}_{\mathcal{R}_{c}}(k)=A_{0}(k/k_{0})^{n_{s}-1}$, using the density contrast (method described in Section 4.3) and the comoving curvature perturbation (method described in Section 4.5). The values we have used in this figure are $A_{0}=2.2\times10^{-9}$, $k_{0}=0.05$Mpc$^{-1}$, $n_{s}=1.3$, $\Delta_{c}=0.4$ and $\mathcal{R}_{c,crit}=1.2$}
\label{zeta approx and density}
\end{figure}
\section{Conclusions}
We have placed the calculation of the PBH abundance on a more solid grounding. Using the comoving curvature perturbation $\mathcal{R}_{c}$ can be misleading and care needs to be taken if one wishes to use $\mathcal{R}_{c}$ to perform this calculation, due to the effect of super-horizon modes. The problem with using $\mathcal{R}_{c}$ is most easily seen when one considers either a red or scale-invariant power spectrum, which causes the variance of $\mathcal{R}_{c}$ to diverge (it is possible to complete the calculation when a blue spectrum is considered but the results differ drastically from using $\Delta$, see Appendix). We therefore advocate the use of the density contrast to perform the calculation, which does not suffer from the same problem due to the $k^{2}$ dependance of super-horizon modes. In addition, calculations and simulations to calculate the critical threshold for collapse most often use $\Delta$. However, it is more convenient to calculate $\mathcal{R}_{c}$ when studying inflationary models, and finding the constraints on the small scale power spectrum from PBHs - an approximation for $\beta$ can be quickly calculated using $\mathcal{R}_{c}$ if the power spectrum, $\mathcal{P}_{\mathcal{R}_{c}}$, is used rather than using the variance, $\langle\mathcal{R}_{c}^{2}\rangle$ (although this can only ever be an approximation as modes of a similar scale can affect the production of PBHs - which this calculation ignores). It is therefore important that calculations using $\Delta$ or $\mathcal{R}_{c}$ give the same results, and we have provided a method for doing so.
We have considered both a Press-Schechter approach and a peaks theory approach, finding that there is a significant discrepancy between the two - however, this is dwarfed by the error due to uncertainty in the critical value of the density contrast above which PBHs are assumed to form, $\Delta_{c}$. In this paper, we use the peaks theory method, which has a better theoretical grounding. The implications of this paper will be explored further in future papers.
\section{Acknowledgements}
SY is supported by an STFC studentship, and would like to thank Yukawa Institute for Theoretical Physics for its hospitality during a month long stay which was supported by the Bilateral International Exchange Program (BIEP). CB was supported by a Royal Society University Research Fellowship. The authors would like to thank Will Watson, Aurel Schneider, David Seery, Shaun Hotchkiss, Anne Green, Andrew Liddle, John Miller and Ilia Musco for useful discussion which brought about the production of this paper.
|
\section{Introduction}
\label{intro}
In the 1958 International Conference on High Energy Physics,
Oppenheimer \cite{OPP} gave the concluding remarks. He said, among other
points:
"There are areas where we know very little -- extremely high energy\\
\vskip-0.5cm
collisions, for instance -- where little can be done by anyone."
Physics is basically an experimental science. At that time over half a
century ago when Oppenheimer gave his speech, there were no high-energy
accelerator where the collisions can be called even remotely "extremely
high energy", in the sense that the incident particles are extremely
relativistic in the center-of-mass system.
This situation began to change about a decade later: the 200-GeV proton
accelerator was to be built at National Accelerator Laboratory, now called
the Fermilab, and the Intersecting Storage Ring (ISR) was to be built at
CERN. At the laboratory energy of 200 GeV, the center-of-mass energy of
each proton is about ten times the proton mass (times the velocity of
light squared), which is extremely relativistic. The proton energy of ISR
is even higher. That these accelerators would be soon available gave
motivation, indeed urgency, to study theoretically the Oppenheimer
challenge of understanding extremely high energy collisions.
A main role of theoretical physics is to give quantitative predictions
that can be either confirmed or refuted by later experiments. In this way
our understanding of physics can be deepened.
What quantity should be studied first at such extremely high energies? At
any energy, the overall property of a collision process is provided by the
total cross section, which is the sum of the integrated cross sections of
all possible scattering and production processes. In turn, the cross
sections of these various processes are necessarily affected by the
behavior of the total cross section.
Motivated by the Oppenheimer challenge and the accelerators to be
available soon, we embarked on the task of studying theoretically the
proton-proton total cross section at the energies of these accelerators
and beyond.
In the nineteen sixties, there were two distinct schools of thought on the
high-energy scattering of strongly interacting particles: the droplet
model of Yang and collaborators \cite{ttwu3} and the Regge pole model \cite{regge}. The
droplet model has numerous successes for many processes at high energies,
and is based on the following observations from experimental data: the
elastic differential cross sections appear to approach limiting values as
the incident energy $E \to \infty$, above about 300 MeV excitation energy, the
nucleon has many excited states and the large-angle proton-proton elastic
cross section drops spectacularly with energy. In contrast, the Regge
pole model deals mostly with scattering processes where some quantum
number is exchanged. The intuition obtained from these two models guided
our thinking on total cross sections at high energies.
Shortly after Professor Yang developed the droplet model for hadron-hadron
collisions, the question was asked: What are the experimental evidences
for or against the existence of limiting values for total cross section?
By studying the experimental data on the ratio
$$ \rho = \frac{\mbox{Real part of pp} \rightarrow \mbox{pp in the forward direction}}
{\mbox{Imaginary part of pp} \rightarrow \mbox{pp in the forward direction}}$$
in the case of proton-proton elastic scattering and using dispersion
relation, there was a weak and preliminary indication that there may not be
such a limiting value in this case. This weak indication actually played
an important role in our decision in the nineteen sixties to study the
total cross section at high energies. This episode has been described in \cite{ttwu4} and we will come back to it in Section 3. In the above expression, the amplitude for $pp \to pp$ does not include the contribution from the Coulomb interaction, (See Eq. (\ref{eq:nc}) in Section 3).
For the purpose of this study, we began by writing down a list of the
basic features of the interaction between elementary particles, features
that are valid not only at high energies but at all energies:\\
- three spatial dimensions and one time dimension\\
- relativistic kinematics\\
- unitarity and\\
- particle production.\\
Even though each of these four features may be considered to be "trivial",
it is nevertheless not easy to have a model with all these four features.
The simplest way to have all features is to have a relativistic quantum
field theory.
As emphasized by Yang and Mills \cite{YM1,YM2}, "... the usual principle of
invariance under isotopic spin rotation is not consistent with the concept
of localized fields." This inconsistency is avoided in theories with a
gauge invariance of the second kind. For this reason, the decision was
made to study the high-energy behavior of a relativistic quantum gauge
field theory in 3+1 dimensions. At that time in the nineteen sixties,
there was no choice because the only such theory that was understood was
the Abelian case. It has turned out that this was most
fortunate: even now nearly half a century later, the high-energy behavior
of non-Abelian Yang-Mills theory remains unknown, even for the simplest
case with the gauge group {\it SU(2)}.
The theoretical investigation of the high-energy behavior, with fixed
transverse momentum, for this Abelian relativistic quantum gauge
field theory began in the late nineteen sixties. The main
theoretical result, obtained in 1970, is that the total cross section
increases without limit for higher and higher energies; this result is
then interpreted to imply that the total cross sections of
strongly interaction particles all have to increase this way \cite{ttwu1}. This is
in agreement with the preliminary indication from the
experimental measurements on the ratio $\rho$ mentioned above.
This theoretical development, together with some of the early
phenomenology has been describe in detail \cite{ttwu2}. This book \cite{ttwu2} is
already over quarter of a century old; reference \cite{ttwu9} gives a later summary
of the situation.
\section{Basic Idea of Phenomenology}
\label{sec2}
It is an interesting and important theoretical statement that hadronic
total cross sections all increase without bound for higher and higher
energies. Nevertheless, such a statement by itself is not of great help
for us to gain a better understanding of nature.
It is the foremost purpose of theoretical physics to make quantitative
predictions. The most important predictions are those that open up new
directions of investigation and are verified accurately by subsequent
experiments; they allow us to have a novel way of understanding nature.
The next type of important predictions are those that open up new
directions but are refuted by subsequent experiments; they tell us that
the novel view needs suitable modification \footnote{To the best of the knowledge of the authors, this wisdom is due to Niels Bohr}.
In the present context of scattering processes at extremely high energies,
the step after the asymptotic calculation of the perturbation series for
the Abelian gauge theory is therefore to develop a quantitatively accurate
phenomenology. What does this mean? Since the internal structure of a
proton is exceedingly complicated and not understood quantitatively from
first principles, this phenomenology must satisfy the following two
conditions:
On the one hand, the proton-proton and proton-antiproton total cross
sections must increase without bound at very high energies, and
On the other hand, this phenomenology must give a reasonable description
of these two cross sections at the relatively high-energy region of the
existing experimental data.
{\it The development of such a phenomenology requires a great deal of physical
intuition and is perhaps the most difficult step for the present approach}
to scattering at extremely high energies, an attempt to response to the
Oppenheimer challenge. More precisely, the difficulty was as follows.
While the field-theoretic calculation \cite{ttwu2,ttwu1} leading to the prediction of
increasing total cross section played a central role, it was physical
intuition that determined which field theoretic results should be
incorporated into the phenomenology and which ones must be rejected.
Here are the two most important and difficult decisions for the development of
our phenomenology, the first one making use of one of the results from
field theory, while the second one purposely contradicting another result
from field theory.
(I) A choice must be made of the variable for describing the increasing
total cross sections. For this purpose, the Regge language is most
helpful. It was known that the leading Regge singularity should be above
1, but the question is: what is the nature of this singularity? The
possible choices are:\\
- moving Regge pole, as in the usual Regge theory\\
- fixed Regge pole\\
- moving Regge cut or\\
- fixed Regge cut.
Our initial inclination was to follow the prevailing thinking in the
nineteen sixties and to use a moving Regge pole. However, our intuitive
feeling was that the leading singularity being a Regge pole is intimately
related to the nature of the underlying quantum field theory. Regge poles
are obtained from the summation of the ladder diagrams in $\phi^3$ theory,
which is a super renormaliable field theory. In general, this property of
being super renormalizable is connected with the ladder-like diagrams
being described by a Fredholm integral equation, and this is the origin of
the moving Regge pole. In contrast, quantum gauge theories in four
dimensions are never super renormalizable, and the explicit calculation
for the Abelian case gives a fixed Regge cut. This was our reason in 1978
to use a fixed Regge cut above 1 \cite{BSW79}. The other two cases of a fixed
Regge pole and a moving Regge cut were considered too unnatural to be used
for our phenomenology.
(II) We also need to make a decision how to describe the absorption in the
scattering of strongly interacting particles. The specific issue is: when
the impact distance between the two incident particles is small, what is
the amount of absorption that should be incorporated into the
phenomenology?
In the field-theoretic calculation for the quantum Abelian gauge field
theory, this absorption is only partial even at extremely high energies.
Since this lack of complete absorption has a rather complicated origin, a
major issue was whether this property was likely to hold for the
interactions between strongly interacting particles. This was an
agonizing choice, and we finally decided to go against this field-theoretic
result and take the absorption in this limit to be complete. The basic
reason for this choice was that it seemed to be against physical intuition
that, in the language of Yang and collaborators, some of the "stuff" is
absorbed while others not \cite{ttwu3}.
It has been gratifying that our phenomenology has worked out well after
these two decisions based on our physical intuition. While we have to
make a number of other choices to complete our phenomenology, the above
two are the most difficult and far-reaching ones.
To describe the experimental data taken at the relatively low energies available
to experiments forty years ago, a new model was proposed in 1978 \cite{BSW79}, including Regge backgrounds. Besides those pertaining to the Regge terms, there is a total of six
parameters for $pp$ and $\bar p p$ elastic scattering. From the overall fit to
the existing data in 1978, the values of these six parameters \cite{BSW79} are
given in the left column of Table 1.
Six years later, when there were significantly more experimental data
at high energies, the overall fit was repeated \cite{BSW84}. The revised
values of these six parameters are given in the right column of Table
1. It should be emphasized that, in these six years from 1978 to
1984, the expressions used to describe the model is not altered at
all; these formulas are given explicitly later in this Section 2.
It is interesting to compare the two columns of values in Table 1:
the six values have not changed much due to the additional
information. This implies that this new model, sometimes referred to
as the BSW Model, is quite robust. There has been no further change
of these parameter values in the thirty years since.
Both for the energies of the present-day colliders and for the purpose of studying
the asymptotic behavior of the model at high energies, all the Regge backgrounds
can be neglected. The BSW model is given by the
following matrix element for elastic scattering
\begin{equation}
\label{eq:one}
\mathcal{M}(s,\mbf{\Delta}) = \frac{is}{2\pi}\int d\mbf{x}_\perp
e^{-i\mbf{\Delta}\cdot \mbf{x}_\perp}D(s,\mbf{x}_\perp)~,
\end{equation}
where $s$ is the square of the center-of-mass energy, $\mbf{\Delta}$ is the
momentum transfer, $\mbf{x}_\perp$ is the impact parameter and all
spin variables have been omitted. For this model we use for the opacity
\begin{equation}
D(s,\mbf{x}_\perp) = 1-e^{-\Omega(s,\mbf{x}^{2}_\perp)}~,
\label{eq:opac}
\end{equation}
with
\begin{equation}
\label{eq:omDef}
\Omega(s,\mbf{x}^{2}_\perp)=\mathcal{S}(s)F(x^{2}_\perp)~,
\end{equation}
where $x_\perp \equiv \mbf{x}_\perp|$~.
The function $\mathcal{S}(s)$ is given by the complex symmetric expression, obtained from the high energy behavior
of quantum field theory \cite{ttwu2,ttwu1}
\begin{equation}
\label{eq:Sdef}
\mathcal{S}(s)= \frac{s^c}{(\ln s)^{c'}} + \frac{u^c}{(\ln u)^{c'}}~,
\end{equation}
with $s$ and $u$ in units of $\mbox{GeV}^2$, where $u$ is the
third Mandelstam variable \cite{SM}. In this Eq. (\ref{eq:Sdef}), $c$ and $c'$ are two
dimensionless constants given in Table 1. That they
are constants implies that the Pomeron is a fixed Regge cut as discussed above. For the asymptotic behavior at high energy and modest momentum
transfers, we have to a good approximation
\begin{equation}
\label{eq:uAp}
\ln u = \ln s -i\pi~,
\end{equation}
so that
\begin{equation}
\label{eq:S2}
\mathcal{S}(s)= \frac{s^c}{(\ln s)^{c'}} +
\frac{s^ce^{-i\pi c}}{(\ln s-i\pi)^{c'}}~.
\end{equation}
Because $F$ depends on $\mbf{x}_\perp$ only through $x^{2}_\perp$, the Fourier transform in Eq. (\ref{eq:one}) simplifies to
\begin{equation}
\label{eq:M2nd}
\mathcal{M}(s,\Delta) = is\int_0^\infty dx_\perp\,x_\perp\,J_0(x_\perp\Delta)
\left[1-e^{-\mathcal{S}(s)F(x^{2}_\perp)}\right]~,
\end{equation}
where $\Delta \equiv |\mbf{\Delta}|$.
The function $F(x_\perp)$ is taken to be related to the electromagnetic form factor
$G(t)$ of the proton, where $t= -\mbf{\Delta}^2$ is the Mandelstam variable for
the square of the momentum transfer. Specifically, $F(x^{2}_\perp)$ is defined as in
\cite{BSW79} via its Fourier transform $\tilde{F}(t)$ by
\begin{equation}
\label{eq:FtildDef}
\tilde{F}(t) = f[G(t)]^2\frac{a^2+t}{a^2-t}~,
\end{equation}
with
\begin{equation}
\label{eq:Gdef}
G(t)=\frac{1}{(1- t/m_1^2)(1- t /m_2^2)}~.
\end{equation}
The remaining four parameters of the model, $f$, $a$, $m_1$ $\mbox{and}$ $m_2$, are given in Table 1.\\
In the next section we will recall some of the early successes of our approach at the CERN $\bar {p} p$ collider
and at the FNAL Tevatron and the next section will be devoted to a preliminary discussion of the sitution at the Large Hadron Collider.
We define the ratio of the real to imaginary parts of the forward amplitude, mentioned earlier in the introduction,
$\rho(s) = \frac{\mbox{Re}~\mathcal{M}(s, t=0)}{\mbox{Im}~\mathcal{M}(s, t=0)}$,
the total cross section
$\sigma_{tot} (s) = (4\pi/s)\mbox{Im}~\mathcal{M}(s, t=0)$,
the differential cross section
${d\sigma (s,t)/dt} = \frac{\pi}{s^2}|\mathcal{M}(s,t)|^2$,
and the integrated elastic cross section
$\sigma_{el} (s) = \int dt \frac{d\sigma (s,t)} { dt}$. One important feature of the BSW model is, as a consequence of Eq. (\ref{eq:S2}), the fact that the phase of the amplitude is built in. Therefore real and imaginary parts of the amplitude cannot be chosen independently and we will show now how to test them.
\begin{table}[htb]
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
Year & 1979 & 1984\\
\hline\raisebox{0pt}[12pt][6pt]{$c$}&0.151 & 0.167 \\[4pt]
\hline\raisebox{0pt}[12pt][6pt]{$c'$}&0.756 & 0.748 \\[4pt]
\hline\raisebox{0pt}[12pt][6pt]{$m_1$}&0.619 & 0.586 \\[4pt]
\hline\raisebox{0pt}[12pt][6pt]{$m_2$}&1.587 & 1.704 \\[4pt]
\hline\raisebox{0pt}[12pt][6pt]{$f$} &8.125 & 7.115 \\[4pt]
\hline\raisebox{0pt}[12pt][6pt]{a} &2.257 & 1.953 \\[4pt]
\hline
\end{tabular}
\caption {Pomeron fitted parameters for $pp (\bar pp)$. Comparison of the
1979 and 1984 solutions..}
\label{tab:table1}
\end{center}
\end{table}
\section{Early successes}
Before showing some early successes of our phenomenology we would like to
come back to an important observation which gave the first hint against
the possibility that the total cross section would remain constant at very high energies.
This is indeed related to the behavior of $\rho(s)$ which is shown in Fig. \ref{fig:rho}. The forward scattering amplitude is expected to satisfy a dispersion relation. This means that the real part of this amplitude can be written as an integral over the imaginary part, which is essentially the total cross section. If the $pp$ total cross section does approach a finite limit, then this ratio $\rho(s)$ must approach zero at high energies. In the mid-sixties the experimentally measured values of $\rho(s)$ in the low energy region, below $\sqrt{s} \sim$ 10GeV, were negative and increasing toward zero. However, the rate of increase seemed to have a tendency to overshoot to become positive and this was the first indication for a possible increasing total cross section at high energies. Indeed positive $\rho(s)$ values were later obtained as displayed in Fig. 1, which also shows the BSW results and predictions up to high energies.
\vskip 1.cm
\begin{figure}[h]
\vspace*{-3.5ex}
\begin{center}
\includegraphics[width=9.0cm]{rho.eps}
\caption[*]{\baselineskip 1pt
The ratio $\rho(s) = \frac{\mbox{Re}~\mathcal{M}(s, t=0)}{\mbox{Im}~\mathcal{M}(s, t=0)}$ for $pp$ (black points) and $\bar p p$ (open points)
elastic scattering and the BSW predictions up to high energies (Taken from Ref. \cite{BSW84}).}
\label{fig:rho}
\end{center}
\end{figure}
Let us now turn to some successful BSW predictions for near-forward $\bar p p$ elastic scattering at the FNAL and CERN colliders energies. For this kinematic region near the forward direction, one must consider the contribution of the Coulomb amplitude $a^C$, so Eq. (\ref{eq:one}) is replaced by
\begin{equation}
\label{eq:nc}
\mathcal{M}(s,\mbf{\Delta}) \pm a_{\pm}^{C}(-\Delta^2),
\end{equation}
the upper sign is for $\bar p p$ while the lower one is for $pp$ scattering and
\begin{equation}
\label{eq:ac}
a_{\pm}^{C}(t) = 2\alpha s\frac{G^{2}(t)}{|t|}\mbox{exp}[\mp i\alpha\phi(t)],
\end{equation}
where $\alpha$ is the fine structure constant, $\phi(t)$ is the West-Yennie phase \cite{wy} given by $\phi(t) = \mbox{ln}(t_0/|t|) - \gamma$, with
$t_0 = 0.08 \mbox{GeV}^2$ and $\gamma \sim 0.577$ is the Euler constant. \\
At the FNAL-Tevatron, the E710 experiment running at $\sqrt{s}$=1.8TeV, has obtained $\sigma_{tot}=72.8 \pm 3.1$ mb and $\sigma_{el}/\sigma_{tot}= 0.23 \pm 0.012$ \cite{amos}, whereas the BSW predictions are 74.8 mb and 0.230 respectively. They were also able to extract the following $\rho$ value, $\rho = 0.140 \pm 0.069$ \cite{amos2}. This important measurement is in agreement with the BSW prediction, but has unfortunately little significance because of its large error. These data are reported in Fig. 2 together with the results of the CDF experiment at two different Tevatron energies $\sqrt{s}$=1.8TeV and $\sqrt{s}$=546GeV \cite{cdf}
and the results of UA(4) at the CERN $\bar p p$ collider at $\sqrt{s}$=541GeV \cite{augier2}. At $\sqrt{s}$=1.8TeV CDF found $\sigma_{tot}=80.03 \pm 2.24$ mb and $\sigma_{el}/\sigma_{tot}= 0.246 \pm 0.004$, at variance with the E710 results.
\begin{figure}[ht]
\vspace*{-3.5ex}
\begin{center}
\includegraphics[width=10.0cm]{tot.eps}
\caption[*]{\baselineskip 1pt
Pre LHC era, data on the total cross section $\sigma_{tot}$ in barns, the ratio $\sigma_{el}/\sigma_{tot}$ of the integrated elastic cross section to the total cross section and the ratio $\rho(s)$, as function of $\sqrt{s}$, together with the BSW predictions from Ref. \cite{BSW84}.}
\label{fig:tot}
\includegraphics[width=10.0cm]{ua4a.eps}
\caption[*]{\baselineskip 1pt
$d\sigma/dt$ for near-forward $\bar p p$ elastic scattering at $\sqrt{s}$= 541GeV data from Ref. \cite{augier}, the curve is the BSW prediction \cite{bsw87} (Taken from Ref. \cite{bsw93}).}
\label{fig:dsigua4}
\end{center}
\end{figure}
However at $\sqrt{s}$=546GeV, the CDF results $\sigma_{tot}=61.26 \pm 0.93$ mb and $\sigma_{el}/\sigma_{tot}=0.210 \pm 0.002$ agree well with those of UA(4), $\sigma_{tot}=63.0 \pm 2.1$ mb and $\sigma_{el}/\sigma_{tot}=0.208 \pm 0.007$. The UA(4) experiment has obtained a very precise value for
the parameter $\rho$, $\rho=0.135 \pm 0.015$, from the measurement of $d\sigma/dt$ in the Coulomb-nuclear interference region \cite{augier}, as shown
in Fig. \ref{fig:dsigua4}. One notices the rapid rise of the cross section in the very low $t$ region and the remarquable agreement with the BSW prediction.
\begin{figure}[h]
\vspace*{-3.5ex}
\begin{center}
\includegraphics[width=11cm]{FNAL.eps}
\caption[*]{\baselineskip 1pt
$d\sigma/dt$ for near-forward $\bar p p$ elastic scattering at $\sqrt{s}$= 1.8TeV data from Ref. \cite{amos}, the curve is the BSW prediction \cite{bsw88} (Taken from Ref. \cite{bsw90}).}
\label{fig:dsige710}
\end{center}
\end{figure}
The BSW model predicts the correct $\rho(s)$ which appears to have a flat energy dependence in the high energy region (see Fig. \ref{fig:rho}) and for $s \to \infty$, one expects $\rho(s) \to 0$. Another specific feature of the BSW model is the fact that it incorporates the theory of expanding protons \cite{ttwu2,ttwu1}, with the physical consequence that the ratio $\sigma_{el}/\sigma_{tot}$ increase with energy. This is precisely in agreement with the data and when $s \to \infty$ one expects $\sigma_{el}/\sigma_{tot} \to 1/2$, which is the black disk limit.
Finally we show in Fig. \ref{fig:dsige710} the $t$-dependence of the elastic cross section measured by the E710 experiment, which again confirms the BSW prediction. It may be worth emphasizing that in this $t$ domain, the $t$-behavior is definitely not a straight line.
\begin{figure}[ht]
\vspace*{-8.5ex}
\begin{center}
\includegraphics[width=10.0cm]{an-200-rhic.eps}
\vspace*{-10mm}
\caption{ The analyzing
power $A_N$ versus $t$ at RHIC energy. The data from Ref. \cite{star} are in excellent agreement with the BSW prediction.}
\label{fig:asy}
\end{center}
\end{figure}
Before moving to the LHC energy it is worth mentioning another independent test of the BSW amplitude, by means of the analyzing power $A_N$, in $pp$
eleastic scattering near the very forward direction. In addition to the non-flip component Eq. (\ref{eq:ac}), the Coulomb amplitude has also a single-flip component $a^{C}_5$, which
involves the proton anomalous magnetic moment. In this kinematic region, the CNI region, $A_N$ results only from the interference of $a^{C}_5$, which is purely real, with the imaginary part of the hadronic non-flip amplitude, if one assumes that there is no contribution from the single-flip hadronic amplitude. This is what we have done in the calculation of the curve displayed in Fig. \ref{fig:asy} compared to some new data from STAR \cite{star}. It confirms the absence of single-flip hadronic amplitude and the right determination of $\mbox{Im}~\mathcal{M}(s, t)$ in the CNI region.
\section{ The LHC energy}
At the moment, the situation with the experimental data at the energies of
the Large Hadron Collider is not completely clear. Here is a description
of the present data.
There are two experiments measuring the proton-proton total cross section:
TOTEM associated with the CMS detector and the ALFA associated with the
ATLAS Collaboration. The center-of-mass energy chosen by both experiments
are 7 TeV.
The ALFA has not published yet any result from their measurements,
although such publications are expected in the very near future. Thus the
discussion here has to be limited to those from TOTEM. The total
proton-proton cross section given by TOTEM is
$\sigma_{tot} (TOTEM) = (98.0 \pm 2.5)$ mb ,
which is 1.7 $\sigma$ above the BSW prediction at 7 TeV. It should be
emphasized that the BSW parameters, shown in Table 1, was determined in
1984 and has not changed in thirty years..
As usual, the total cross section is obtained by extrapolating the
differential cross section to the forward direction. It is therefore of
interest to compare the TOTEM measurement of the proton-proton
differential cross section directly with the BSW prediction as discussed
in Sec. 2. Such a comparison is shown in Fig. \ref{fig:lhc}.
The meaning of this comparison of Fig. \ref{fig:lhc} is not straightforward and has
not yet been completely clarified, and we look forward to future
developments, especially at even higher energies of LHC.
\begin{figure}[hp]
\vspace*{1.0ex}
\begin{center}
\includegraphics[width=13.00cm]{dsigtotem.eps}
\caption[*]{\baselineskip 1pt
$d\sigma/dt$ for near-forward $pp$ elastic scattering.
The data are from TOTEM \cite{totem} and the curve is the BSW prediction. }
\label{fig:lhc}
\end{center}
\end{figure}
\section{Conclusion}
The basic idea for the development of the present phenomenology has been
described in Sec. 2. On the one hand, it is essential to incorporate
suitably chosen results from field theory; however, some other
results from field theory have to be purposely contradicted in the
phenomenology. Physics has played an essential role in the choices, and
is largely responsible for the success in the many predictions of the
present phenomenology, which is entirely unchanged in nearly thirty years.
When the present phenomenology was first worked out, the highest
center-of-mass energy of available experimental data was 62 GeV. As shown
in Sec. 3, the predictions of this phenomenology are in good agreement
with later experimental data up to the center-of-mass energy of 1.96 TeV.
This is an increase of energy by a factor of more than thirty -- an
extraordinary success as discussed in the previous sections.
The conclusion is therefore reached that there is a good understanding of
proton-proton elastic scattering near the forward direction.
It will be of interest to push this phenomenology to higher energies by
comparing its predictions with the data from the Large Hadron Collider at
the center-of-mass energy of 7 TeV. Some such comparison has already been
presented in Sec. 4, and some more data are expected from the ALFA
Collaboration in the near future. It is even more interesting to compare
the present predictions with the data after an upgrade of the
center-of-energy of the Large Hadron Collider next year to 13 or 14 TeV,
both for the differential elastic cross section and the parameter $\rho$
discussed above, whose relevance of its measurement has been strongly emphasized \cite{bkmsw}.
It may be of some importance to add the following comment to the present
development on the increasing total cross section. It is the production
of relatively low-energy particles in the center-of-mass system that leads
to this increasing total cross section; such production processes are
usually referred to as "pionization". Also it was known from the
beginning \cite{ttwu2,ttwu1} that, at extremely high energies, half of this increase
in the total cross section is due to the integrated elastic cross section.
The obvious question, already raised more than forty years ago, is: what
processes are responsible for the other half of this increase at extremely
high energies?
This question has been answered recently through the application of
geometrical optics to production processes \cite{gww}: the same processes, i.e.,
pionization, are not only responsible for the increasing total cross
section, but are also responsible for the other half of the increasing
total cross section at extremely high energies -- an answer that we find
to be philosophically satisfying.
\vskip 0.5cm
{\bf Acknowledgments}
We thank Andr\'{e} Martin and Maurice Haguenauer for several fruitful discussions.
|
\section{Introduction}
Complex networks have garnered much recent attention for their intrinsic mathematical interest and their many applications to everyday life~\cite{albert02,newman03,barrat08,dorogovtsev08}. A particular area of research is the study of simple growth models that capture the networks' more salient features~\cite{watts98,barabasi99,dorogovtsev00,jin01,growthbook}. A well known example of this is
the Krapivsky-Redner growth redirection model (KR)~\cite{kr}, where each new node is connected to a randomly selected node, with probability $1-r$, or else the connection is ``redirected" to the ancestor of that node, with probability $r$ (the {\it ancestor} of $x$ is the node that $x$ connects to upon joining the network).
The redirection events select preferentially for existing nodes of higher degree (the rich-get-richer effect), giving rise to a {\em scale-free} degree distribution with degree exponent $\gamma_{\rm KR}=1+1/r$ --- arguably, the most central characteristic of real-life complex networks. Moreover, the KR model achieves this with a simple, {decentralized} algorithm (only local knowledge about the target node is required at each step) that in all likelihood captures an essential ingredient in the growth of real complex networks: befriending a friend of a friend, in social networks, or finding a new reference through an already cited paper, in a network of citations, are obvious analogs of the KR redirection mechanism. On the other hand, it is unrealistic to expect, in real-life examples, that each joining agent makes only {\em one} connection.
A result of the restriction to one connection is that the KR model yields loopless (acyclic) graphs, or {\em trees}, which simplifies their theoretical analysis, but is also in disparity with real-life complex networks that tend to have abundant loops and a high degree of clustering~\cite{ebel02,hidalgo08,bullmore09,kaluza10,loopy}. In~\cite{rb} we explored the possibility of connecting each new joining node to $m$ of the existing network nodes, with probability $p_m$, where each of the $m$ connections follows the original KR redirection recipe. If the support of $m$ is finite, $m=1,2,\dots,m_{max}$, then the resulting networks are still scale-free,
with the same degree exponent $\gamma_{\rm KR}=1+1/r$, but they now contain loops and their degree distributions for small $k$ may be conveniently manipulated through an appropriate choice of the $p_m$'s~\cite{rb}. Our interest here is in the case that the $p_m$ do {\em not} have finite support, but $m_{max}$ grows along with the size of the network, and we focus on power-law distributions of the form $p_m\sim m^{-\alpha}$, such that the probability for {\it superjoiners} --- new nodes that connect to most of the nodes already in the network --- is substantial.
This power-law choice is motivated by several practical situations. For example, in social networks it is well known that the distribution of the number of acquaintances of a person is governed by a power-law, and it is then plausible that the number of friends $m$ a person would make upon joining a new social network would be power-law distributed as well;
when a company introduces a new router to the Internet, we may expect that it would be connected to $m$ routers, with a power-law distribution, since larger companies can afford more connections in proportion to their resources, and companies' sizes follow a power-law distribution, etc.
There is a potential tension between the exponent $\alpha$ governing the distribution of {\em outgoing} links of each new node (we regard the links formed from each new node as {\em directed} from the node to their target), and the exponent $\g_{\rm in}=1+1/r$ that one expects for the {\em incoming} links, that form under the KR rich-get-richer mechanism with redirection probability~$r$. How does this tension play out and which of the two effects dominates the {\em total} degree distribution, where all the links of a node are counted, regardless of their directionality? Below we shall see that rich behavior results from this interesting tug of war.
We show that the degree distribution of outgoing links is $P_{\rm out}(l)\sim l^{-\g_{\rm out}}$, $\gamma_{out}=\alpha$, whereas that of incoming links is $P_{\rm in}(k)\sim k^{-\g_{\rm in}}$, where $\g_{\rm in}=1+1/r$ for $\alpha>2$, and $\g_{\rm in}=0$ (a flat distribution) for $\alpha<2$.
The distribution of the total degree (in- {\em and} out-links) is $P(q)\sim q^{-\gamma}$, where $\gamma=\min\{\g_{\rm in},\g_{\rm out}\}$ for $\alpha>1$, and $\gamma=-(1-\alpha)$ for $\alpha<1$. In the latter case, $\gamma<0$, with nodes of larger degree becoming more abundant, in complete reversal from the usual state of affairs for everyday complex nets.
One result of the fat-tailed distribution for the number of new links, $m$, is that the total number of links in the network, $M(t)$, may grow faster than the number of nodes, $N(t)$. We find that $M(t)\sim N(t)^\beta$, with $\beta=1$ for $\alpha>2$, $\beta=3-\alpha$ ($>1$) for $1<\alpha<2$, and $\beta=2$ for $\alpha<1$. The case in which $\beta>1$ and $M(t)$ outpaces the growth of $N(t)$ is known as the {\em non-sustainable} regime. In fact, due to the fast growth of $M$ conflicts might arise on introducing a new node, when some of its random connections call to be directed to a common target. For $\alpha<2$ and $r>1/(3-\alpha)$ we find a regime of {\em extreme non-sustainability}, where the number of conflicts becomes a {\em finite} fraction of the total number of links and our model becomes ill-defined.
The rich behavior of the superjoiners model is summarized in Fig.~\ref{phase_diagram.fig}. Throughout the remainder of this paper we present the theoretical and numerical considerations that led us to these conclusions.
An infinite support for $p_m$ was already briefly considered in~\cite{rb}, for the ``self-consistent" case that it is dictated by the network's own degree distribution, $p_m=P(m)$. In that case the total degree distribution is scale-free, with $\gamma=1+1/r$. Krapivsky and Redner~\cite{copying} have made a detailed study of network growth by ``copying," where new nodes connect to a randomly selected node and to all its ancestors, so that there too $m$ may grow without bound. Power-law distribution for new connections (with unbounded $m$) were studied by Tessone et al.,~\cite{tessone}, in the context of dependency networks in Open Source Software projects, where they observed $\alpha$ values in the range of $2.2$--$3.5$. Although their theoretical analysis focuses around strictly linear growth kernels (yielding BA-like models limited to $\gamma=3$), many of our findings here are closely related to theirs.
\begin{figure}[h]
\includegraphics[width=0.45\textwidth]{fig1}
\caption{(Color online) ``Phase diagram" for the superjoiners model: The model exhibits different
asymptotic behavior and divides into four regions in the $(\alpha,r)$ space, with regards to the exponents characterizing the asymptotic power-law behavior of in-, out-, and total-degree distribution ($\g_{\rm in}$, $\g_{\rm out}$, and $\gamma$, respectively). For $\alpha>2$, the networks are sustainable, $M\sim N^{\beta}$, $\beta=1$, and $\gamma=\min\{\g_{\rm in},\g_{\rm out}\}$, yielding two distinct ``phases" separated by the curve $\g_{\rm out}=\g_{\rm in}$, or $\alpha=1+1/r$. For $\alpha<2$, the networks are unsustainable, with $\beta>1$, and the region divides into two distinct ``phases" separated by the line $\alpha=1$. Additionally, the superjoiners model is ill-defined in the region marked as ``extreme non-sustainability," for $r>1/(3-\alpha)$, as conflicts arise in the implementation of the superjoiners growth model. The various results summarized in this diagram are derived throughout the paper.}
\label{phase_diagram.fig}
\end{figure}
\section{Model description}
Our network model is constructed by adding one node at a time. A newly added node, $y$, is connected to $m$ of the existing nodes in the network, $x_1,x_2,\dots,x_m$. The connections are {\it directed}, pointing from $y$ into each of the $m$ target nodes. The $m$ target nodes are called the {\it ancestors} of $y$. Each of the $m$ target nodes is selected by the Krapivsky-Redner recipe: A node $x$ is selected at random and it is identified as $x_1$ with probability $1-r$. Otherwise, with probability $r$, $x_1$ is a randomly selected ancestor of $x$. In either case, a directed link is created from $y$ to $x_1$. This process is repeated $m$ times, until all target nodes are determined (and $y$ has $m$ out-links and $m$ ancestors). For now, we assume that {\it conflicts}---when the same target node is selected for more than one of the $m$ connections---do not arise, or at least are rare enough to be neglected. Conflicts are addressed in Section~\ref{conflicts.sec}.
The number of nodes in the system grows by $1$ with the addition of each new node, so that the total number of nodes
after step $t$ is
\begin{equation}
N(t)=N(0)+t\,,
\end{equation}
where $N(0)$ is the initial number of nodes at the starting configuration, at time $t=0$. The range of $m$, therefore, may increase linearly with time. In this work, we are interested in what happens when $m$ is selected from a power-law distribution, such as
\begin{equation}
\label{pm.eq}
p_m(t)=C(t)m^{-\alpha}\,,\qquad m=1,2,\dots,N(t)-1\,,
\end{equation}
where both the normalization factor,
\begin{equation}
\label{ct.eq}
C(t)=1\Big/\sum_{m=1}^{N(t)-1}m^{-\alpha}\,,
\end{equation}
and the range of $m$ adapt to the growing network~\cite{remark1}.
Although $m$ can grow without bound, it is still bounded at any {\it finite} time. This allows us, in principle, to consider any value for the exponent $\alpha$, including cases in which $C(t)^{-1}\to\infty$, as $t\to\infty$. We nevertheless limit our study to $\alpha>0$, since unsustainability becomes extreme as $\alpha$ decreases, and it is hard to think of practical examples for $\alpha<0$.
The normalization factor $C(t)$ converges to a finite value for $\alpha>1$, but not for $\alpha<1$, resulting in the asymptotic behavior
\begin{equation}
\label{pm_asym.eq}
p_m(t)\sim m^{-\alpha}\times\begin{cases}
1 & \alpha>1,\\
t^{-(1-\alpha)} & \alpha<1.
\end{cases}
\end{equation}
The starting configuration, at time $t=0$, has no significant effect on the network at large times. For concreteness, however, our simulations (and numerical integrations of the theory's equations) start with two nodes, $v_A$ and $v_B$, connected to one another via directed links: $v_A\to v_B$ and $v_B\to v_A$. From now on, we specialize our discussion to this case, where $N(0)=2$, $M(0)=2$, and $N(t)=t+2$.
\section{theory and results}
Let $N_{kl}(t)$ denote the number of nodes with $k$ links in and $l$ links out, at time $t$. The total number of nodes with in-degree $k$ (regardless of the out degree) is $F_k(t)=\sum_lN_{kl}(t)$, and the number of nodes with out-degree $l$ (regardless of the in-degree) is $G_l(t)=\sum_kN_{kl}(t)$. The number of nodes with {\em total} degree $q$ is
\begin{equation}
\label{Hq.eq}
H_q(t)=\sum_{k,l\atop k+l=q}N_{kl}(t)\,.
\end{equation}
The total number of nodes in the network at time $t$ is
\begin{equation}
\sum_kF_k=\sum_lG_l=\sum_qH_q=N(t)\,.
\end{equation}
The {\it average number of outgoing links} added to the net at time $t$ is
\begin{equation}
\label{kappa.eq}
\kappa(t)=\sum_{m=1}^{t+1}mp_m(t)\,,
\end{equation}
or, in the asymptotic limit of $t\gg1$,
\begin{equation}
\label{kappa_asym.eq}
\kappa(t)\sim\begin{cases}
\kappa_{\infty}, & \alpha>2,\\
t^{2-\alpha}, & 1<\alpha<2,\\
t, & \alpha<1,
\end{cases}
\end{equation}
where we have made use of (\ref{pm_asym.eq}), and $\kappa_{\infty}$ is a constant (dependent on $\alpha$).
The {\it expected total number of outgoing links} at time $t$ is then
\begin{equation}
\label{M.eq}
M(t)=M(0)+\sum_{t'=1}^t\kappa(t')\,,
\end{equation}
where $M(0)$ is the number of outgoing links in the initial configuration. Note that the total number of incoming links and outgoing links is equal at all times (and equal to the total number of links, since each link is simultaneously an in-, and out-link). Using Eq.~(\ref{kappa_asym.eq}) and the fact that $N(t)\sim t$ for all $\alpha$, we get
\begin{equation}
\label{M_asym.eq}
M(t)\sim N(t)^{\beta}\,;\qquad \beta=\begin{cases}
1, & \alpha>2,\\
3-\alpha, & 1<\alpha<2,\\
2, &\alpha<1.
\end{cases}\end{equation}
Thus, for $\alpha<2$, the growth of links outpaces that of the nodes, and our model then yields {\em unsustainable} networks, with $\beta>1$.
$N_{kl}$ satisfies the rate equation:
\begin{equation}
\label{Nkl.eq}
N_{kl}(t+1)-N_{kl}(t)=\kappa(t)\frac{1-r}{N(t)}[N_{k-1,l}(t)-N_{kl}(t)]+\kappa(t)\frac{r}{M(t)}[(k-1)N_{k-1,l}(t)-kN_{kl}(t)]
+p_l(t)\delta_{k,0}\,.
\end{equation}
The first term on the right-hand side denotes the changes to $N_{kl}$ due to direct connections: a node is selected for such a process with probability $(1-r)/N(t)$, and this takes place $\kappa(t)$ times during each step (on average). The second term denotes redirected connections: a node of in-degree $k$ (and out-degree $l$) is selected with probability $kN_{kl}/M(t)$. The last term accounts for the new node added to the net: the new node has $l$ outgoing links (and zero incoming links), with probability $p_l(t)$.
Equation~(\ref{Nkl.eq}) is supplemented with the boundary condition
\begin{equation}
N_{kl}(t)=0 \qquad {\rm for\ }k<0{\rm\ or\ }l<1\,.
\end{equation}
\subsection{Outgoing links}
\label{outgoing.sec}
Summing (\ref{Nkl.eq}) over $k$, we obtain $G_l(t+1)-G_l(t)=p_l(t)$,
or
\begin{equation}
\label{Gl.eq}
G_l(t)=G_l(0)+\sum_{t'=0}^{t-1}p_l(t')=2\delta_{l,1}+\sum_{t'=l-1}^{t-1}p_l(t')\,.
\end{equation}
[The lower limit of $t'=l-1$ in the last sum is explained by the fact that $p_l(t)=0$ for $l>t+1$.]
This can be obtained more directly upon realizing that the out-degree of each node is determined as it is added to the network, and never changes thereafter. If $\alpha>1$, the normalization factor for $p_l(t)$ converges with time to a finite value, while it vanishes as $C(t')\sim t'^ {\alpha-1}$ for $\alpha<1$, leading to
\begin{equation}
\label{Gl_asymp.eq}
G_l(t)\sim l^{-\alpha}\times\begin{cases}
(1-l/t), & \alpha>1\,,\\
(1-(l/t)^{\alpha}), & \alpha<1\,.
\end{cases}
\end{equation}
In either case, $G_l(t)\sim l^{-\alpha}$ for finite $l/t$, as illustrated in Fig.~\ref{Gl.fig}. This means that $\g_{\rm out}=\alpha$
for all values of $\alpha$ (and $r$).
\begin{figure}[h]
\includegraphics[width=0.45\textwidth]{fig2}
\caption{(Color online) Out-degree $G_l(t)$ for $\alpha=0.5$ (upper curve) and $\alpha=2$ (lower curve), for $t=10000$. The symbols represent a direct summation of Eq.~(\ref{Gl.eq}) and the solid curves are fit from Eq.~(\ref{Gl_asymp.eq}). Slopes of $l^{-\alpha}$ are shown as broken lines, for comparison.}
\label{Gl.fig}
\end{figure}
\subsection{Incoming links}
\label{incoming.sec}
Summing (\ref{Nkl.eq}) over $l$, we obtain
\begin{equation}
\label{Fk.eq}
F_k(t+1)-F_k(t)=\kappa(t)\frac{1-r}{N(t)}[F_{k-1}(t)-F_k(t)]+\kappa(t)\frac{r}{M(t)}[(k-1)F_{k-1}(t)-kF_k(t)]+\delta_{k,0}\,.
\end{equation}
The outcome now depends on the time-asymptotic behavior of $N(t)$, $\kappa(t)$, and $M(t)$. For~$\alpha>2$, $\kappa(t\to\infty)=\kappa_{\infty}$ converges to a constant, and $M(t)\sim\kappa_{\infty}t$. Using these asymptotic relations, along with $N(t)\sim t$ and the {\it ansatz} $F_k(t)\sim f_kt$ (where $f_k$ is a constant independent of time), one gets
\begin{equation}
\label{fk.eq}
f_k=\kappa_{\infty}(1-r)[f_{k-1}-f_k]+r[(k-1)f_{k-1}-kf_k]+\delta_{k,0}\,.
\end{equation}
This can be analyzed exactly~\cite{kr}, leading to $f_k\sim k^{-(1+1/r)}$, or $\g_{\rm in}=1+1/r$ for $\alpha>2$.
For $\alpha<2$, the term with $\kappa(t)/N(t)$ dominates Eq.~(\ref{Fk.eq}), asymptotically, and determines the outcome, which now becomes $F_k(t)\to{\rm const.}$, independent of $k$. In other words,
$\g_{\rm in}=0$ for $\alpha<2$. The two regimes for $\g_{\rm in}$ ($\alpha>2$ and $\alpha<2$) are demonstrated in Fig.~\ref{Fk.fig}.
\begin{figure}[h]
\includegraphics[width=0.45\textwidth]{fig3}
\caption{(Color online) In-degree $F_k(t)$ for $\alpha=1$, $r=0.3$~(a) and $\alpha=3$, $r=0.5$~(b), as computed from Eq.~(\ref{Fk.eq}) (solid line) and from simulations ($\circ$) for $t=10000$, averaged over 50 runs. The predicted asymptotic slopes, of $\g_{\rm out}=0$ and $\g_{\rm out}=1+1/r=3$, respectively, are shown for comparison (broken line).}
\label{Fk.fig}
\end{figure}
\subsection{All links}
\label{all.sec}
To derive $H_q(t)$, we need to deal fully with Eq.~(\ref{Nkl.eq}). This can be done exactly by exploiting the fact that the index $l$ is fixed and it effectively enters the equation only as a boundary condition. First, set $k=0$ to obtain
\begin{equation}
\label{Nk0.eq}
N_{0l}(t+1)-N_{0l}(t)\left[1-\kappa(t)\frac{1-r}{N(t)}\right]=p_l(t)\,,
\end{equation}
which can be solved with the help of the integrating factor
\begin{equation}
\label{A.eq}
A_0(t)=\prod_{j=1}^t\left[1-\kappa(j)\frac{1-r}{N(j)}\right]^{-1}\,,
\end{equation}
to yield
\begin{equation}
\label{N0lsol.eq}
N_{0l}(t)=A_0(t-1)^{-1}\sum_{i=0}^{t-1}A_0(i)p_l(i)\,.
\end{equation}
In general, for $k\geq1$, Eq.~(\ref{Nkl.eq}) can be rewritten as
\begin{equation}
\label{Nkla.eq}
N_{kl}(t+1)-N_{kl}(t)\left[1-\kappa(t)\left(\frac{1-r}{N(t)}+\frac{rk}{M(t)}\right)\right]=\kappa(t)\left[\frac{1-r}{N(t)}+
\frac{r(k-1)}{M(t)}\right]N_{k-1,l}(t)\,,
\end{equation}
which can be solved for $N_{kl}(t)$, in terms of $N_{k-1,l}(t)$, using an integrating factor similar to $A_0(t)$,
\begin{equation}
\label{Ak.eq}
A_k(t)=\prod_{j=1}^t\left[1-\kappa(j)\left(\frac{1-r}{N(j)}+\frac{rk}{M(j)}\right)\right]^{-1}\,,
\end{equation}
leading to
\begin{equation}
\label{Nklsol.eq}
N_{kl}(t)=A_k(t-1)^{-1}\sum_{i=0}^{t-1}A_k(i)\kappa(i)\left[\frac{1-r}{N(i)}+
\frac{r(k-1)}{M(i)}\right]N_{k-1,l}(i)\,.
\end{equation}
Thus, one can obtain $N_{1l}$ from the known $N_{0l}$, then $N_{2l}$ from $N_{1l}$, etc.
This procedure, however, is cumbersome and does not allow for a simple analysis of asymptotic properties.
Instead, we observe that the weak interaction between $k$ and $l$ should lead to a near absence of correlations between the in-and out-degrees, so that
\begin{equation}
\label{correlations.eq}
N_{kl}(t)\approx F_k(t)G_l(t)/N(t)\,,
\end{equation}
where the denominator is dictated by normalization.
Beyond the theoretical insights gained by this simplification, the uncorrelated in- and out-degrees allow for numerical integration of much larger networks: from algorithms that grow as $N^3$, for integration of $N_{kl}$ directly [Eq.~(\ref{Nkl.eq})], to algorithms of order $N^2$ for the integration of both $G_l$ and $F_k$ [Eqs.~(\ref{Gl.eq}) and (\ref{Fk.eq}].
Assuming noncorrelation and in view of~(\ref{Hq.eq}), the probability for total degree $q$ can be written as a convolution,
\begin{equation}
\label{Hq_approx.eq}
H_q(t)\approx\sum_{k}F_k(t)G_{q-k}(t)/N(t)\,.
\end{equation}
If we further use the long-time asymptotic results,
$F_k\sim k^{-\g_{\rm in}}$ and $G_l\sim l^{-\g_{\rm out}}$ then, approximating the sum with an integral and analyzing the divergences at the lower and upper limits, we find
\begin{equation}
\label{Hq_slopes.eq}
H_q\sim q^{-\gamma}\sim\begin{cases}
q^{-\min\{\g_{\rm in},\g_{\rm out}\}}, & \g_{\rm in}>1,\>\g_{\rm out}>1,\\
q^{-\g_{\rm in}}, & \g_{\rm in}<1,\>\g_{\rm out}>1,\\
q^{-\g_{\rm out}}, & \g_{\rm in}>1,\>\g_{\rm out}<1,\\
q^{1-\g_{\rm in}-\g_{\rm out}}, & \g_{\rm in}<1,\>\g_{\rm out}<1.
\end{cases}
\end{equation}
The first case corresponds to $\alpha>2$, where $\g_{\rm in}=1+1/r\,(>1)$ and $\g_{\rm out}=\alpha\,(>1)$. Thus, for $\alpha>2$ we have
$\gamma=\min\{1+1/r,\alpha\}$ and the region includes two phases (with distinct $\gamma$) separated by the curve $\alpha=1+1/r$. On that curve,
$\gamma=\g_{\rm in}=\g_{\rm out}$. The second case corresponds to the region $1<\alpha<2$, where $\g_{\rm in}=0\,(<1)$ and $\g_{\rm out}=\alpha\,(>1)$. Hence, for $1<\alpha<2$ we have $\gamma=0$ and the line $\alpha=2$ demarcates a sharp transition for both the values of $\g_{\rm in}$ (from $1+1/r$ for $\alpha>2$, to $0$ for $\alpha<2$) and of $\gamma$ (from $\alpha$ to $0$). The third case does not occur in our model. Finally, the fourth case corresponds to $\alpha<1$, where $\g_{\rm in}=0\,(<1)$ and $\g_{\rm out}=\alpha\,(<1)$. In this region, we get the strange result of a negative $\gamma$ exponent: $\gamma=-(1-\alpha)$. An accessible
summary of the different regimes for the predicted values of $\g_{\rm in}$, $\g_{\rm out}$, and $\gamma$, as a function of $\alpha$ and $r$, is presented in the ``phase diagram" of Fig.~\ref{phase_diagram.fig}.
We have two theoretical approaches for the computation of $H_q(t)$: (i)~Compute $N_{kl}(t)$ {\em directly} from integration of Eq.~(\ref{Nkl.eq}) and then use Eq.~(\ref{Hq.eq}), and (ii)~Integrate Eqs.~(\ref{Gl.eq}) and (\ref{Fk.eq}) numerically to obtain $G_l(t)$ and $F_k(t)$, respectively; then assuming {\em noncorrelation}, use Eq.~(\ref{Hq_approx.eq}). We next wish to compare these theoretical approaches to one another as well as to simulation results. In Fig.~\ref{Hq.fig} we show results for representative $(\alpha,r)$ pairs in each of the four regions of the phase diagram (Fig.~\ref{phase_diagram.fig}). The direct theoretical approach~(i) is shown as a solid curve, and the noncorrelation approach~(ii) is shown as dash-dotted curve, while the simulation results are denoted by symbols. Because (i) can be carried out only for relatively small nets, we limit ourselves to $t=1000$. It is encouraging that the long-time asymptotic prediction for $\gamma$ (shown as dashed lines) provides a reasonable description for even such small nets. In fact, we find that the range where the asymptotic behavior applies increases with the size of the net.
The good agreement between (i) and (ii) (the curves are practically indistinguishable, other than in the first panel) supports the noncorrelation approximation, which we used for the derivation of $\gamma$. The results also demonstrate that either theoretical approach, (i) or (ii), provides a very apt description of the transient behavior observed in small nets.
\begin{figure}[h]
\includegraphics[width=0.45\textwidth]{fig4}
\caption{(Color online) Total degree distribution $H_q(t)$ in small networks of $t=1000$ for (a)~$\alpha=0.5$, $r=0.2$; (b)~$\alpha=1.5$, $r=0.4$; (c)~$\alpha=3$, $r=0.4$; and (d)~$\alpha=3$, $r=0.8$. The solid curves represent the direct analytical approach~(i), while dashed-dotted curves are computed from approach (ii). Simulation results, averaged over 1000 runs, are denoted by open circles ($\circ$). For comparison, the theoretical long-time asymptotic prediction of the exponent $\gamma$ is indicated by dashed lines.}
\label{Hq.fig}
\end{figure}
\subsection{Conflicts and extreme unsustainability}
\label{conflicts.sec}
We now address the possibility of {\it conflicts}, when the same target node is selected more than once and multiple links between pairs of nodes may result (for $m>1$). Conflicts may be dealt with in several practical ways: (a)~When a conflict arises, that is, when a target is selected twice, repeat the random selection process until a new (non-conflicting) target is found. (b)~When a conflict arises, do not connect. This means that the actual number of connections for the new node might be smaller than $m$. (c)~When a conflict arises, make the connection regardless of the conflict, thus allowing for multiple directed links between pairs of vertices.
\begin{figure}[htbp]
\includegraphics[width=0.45\textwidth]{fig5}
\caption{(Color online) Conflicts. The fraction of conflicts that arise in the construction of networks up to time $t=5000$ for each of the three strategies described in the text is indicated in different shadings, according to the scheme of the bars at the right. The theoretical curve $r=1/(3-\alpha)$ reasonably separates between regions of many conflicts (extreme unsustainability) and almost no conflicts in all three cases. }
\label{conflicts.fig}
\end{figure}
In the original KR model, where $m$ is always 1, conflicts never arise. If $\kappa(t)$ tends to a {\it finite} number as $t\to\infty$, as is the case for $\alpha>2$, conflicts are rare in our model and they have a negligible influence on the outcome: it does not matter then which strategy one adopts to deal with conflicts.
For $\alpha<2$, however,
conflicts are no longer rare and deviations between the results from the various conflict strategies might be expected.
Generally, we expect conflicts to be relevant if they occur as a {\em finite fraction} of all attempts to connect new links. If the ratio of conflicts to the total number of attempts tends to zero, in the thermodynamic limit of $t\to\infty$, then they can be neglected. Indeed, our principal equation~(\ref{Nkl.eq}) for the evolution of $N_{kl}(t)$ does not take into account the possibility of conflicts, so it is valid only in the latter case. Its domain of applicability is suggested by the iterative solution of Eq.~(\ref{Nklsol.eq}): A necessary requirement for the solution to make sense is that $\left[1-\kappa(t)\left(\frac{1-r}{N(t)}+\frac{rk}{M(t)}\right)\right]>0$ for all $k\leq t$.
Putting $k=t$, and using the asymptotic expressions for $N(t)$, $\kappa(t)$, and $M(t)$, we get
\begin{equation}
1-\kappa(t)\left(\frac{1-r}{N(t)}+\frac{rt}{M(t)}\right)\sim
\begin{cases}
1-r, &\alpha>2,\\
1-(3-\alpha)r, &1<\alpha<2,\\
1-\frac{1-\alpha}{2-\alpha}(1-r)-2r, &\alpha<1.
\end{cases}
\end{equation}
The first case, for $\alpha>2$, satisfies the requirement for all values of $r$, consistent with the expectation that conflicts are negligible for $\alpha>2$. Surprisingly, the two other cases,
for $1<\alpha<2$ and for $\alpha<1$, yield the very same condition:
\begin{equation}
\label{demarcate.eq}
r<\frac{1}{3-\alpha}\qquad {\rm for\ }\alpha<2\,.
\end{equation}
Thus the regime of $r>1/(3-\alpha)$, $\alpha<2$ is not only {\em unsustainable} ($\beta>1$), but one expects a finite fraction of conflicts there. We term this phenomenon {\em extreme unsustainability}. In Fig.~\ref{conflicts.fig} we plot the fraction of conflicts in simulations, using all three strategies. The boundary suggested by
Eq.~(\ref{demarcate.eq}) seems to capture the transition to extreme sustainability quite adequately.
\subsection{Graph spectra of superjoiners networks}
Finally, having established the effect of superjoiners on the degree distribution, we address the question of how they affect the dynamics of processes on such networks. Toward that end, we focus on the spectral properties of the adjacency matrix $A=[A_{ij}]_{N\times N}$ and the Laplacian matrix $L=[L_{ij}]_{N\times N}$, as these play a key role in the interplay between structure and dynamics in networks in general~\cite{barrat08,dorogovtsev08,arenas08,nishikawa10,skardal14prl}. The adjacency matrix is defined as
\begin{equation}
A_{ij}=
\begin{cases}
1 & \mbox{if node $i$ connects to node $j$},\\
0 & \mbox{otherwise,}
\end{cases}
\end{equation}
and the Laplacian matrix $L$ is given by $L_{ij}=\delta_{ij}d_i-(1-\delta_{ij})A_{ij}$, where $d_i=\sum_{k}A_{ik}$ is the out-degree of node $i$.
Based on the construction of the superjoiners networks, $A$ has the lower-triangular block form
\begin{equation}\label{Adj.eq}
A=\begin{pmatrix}
\begin{array}{c|c}
A^{(0)} & O \\ \hline
A^{(10)} & A^{(1)}
\end{array}
\end{pmatrix},
\end{equation}
where $A^{(0)}$ is the adjacency matrix of the initial network of $N(0)$ nodes,
$A^{(1)}$ is the adjacency matrix among the added nodes, and $A^{(10)}$ encodes the directed links between those two groups. Since no self-loops are allowed, both $A^{(0)}$ and $A^{(1)}$ have zero diagonals. Moreover, since each new node can only connect to nodes that joined the network before it, the matrix $A^{(1)}$ is in fact lower-diagonal. Consequently, the spectrum of $A$ is given by
\begin{equation}
\Lambda(A) = \Lambda(A^{(0)})\cup\{0,0,\dots,0\},
\end{equation}
where $\Lambda(M)$ denotes the set of eigenvalues of a matrix $M$.
The same argument applies to the Laplacian matrix $L$, which yields the spectrum
\begin{equation}
\Lambda(L) = \Lambda(L^{(0)})\cup\{d_{N(0)+1},d_{N(0)+2},\dots,d_N\},
\end{equation}
where $L^{(0)}$ is the Laplacian matrix of the initial network.
In the examples considered in this paper, the initial network contains $N(0)=2$ interconnected nodes. It follows that $\Lambda(A^{(0)})=\{-1,1\}$ and consequently $\Lambda(A)=\{-1,0,0,\dots,0,1\}$. The eigenvalues of $A$ have been theoretically hypothesized and experimentally confirmed to be a determining factor for neuronal activity~\cite{kinouchi06,larremore11}. In fact, the dynamic range, quantifying the range of stimuli that results in distinguishable responses, was shown to be maximized when the largest eigenvalue of $A$ (in magnitude) equals 1~\cite{kinouchi06,larremore11}, as in our case.
Turning to the Laplacian matrix $L$, it follows that for the initial network, $\Lambda(L^{(0)})=\{0,2\}$ while the rest of the eigenvalues of $L$ are given by the out-degrees of the remaining nodes in the network. Since the out-degrees of added nodes are bounded by $1\leq d\leq N-1$, Eq.~(\ref{pm.eq}), the eigenvalues of $L$ can be ordered as
\begin{equation}
0=\lambda_1(L)<1=\lambda_2(L)\leq\dots\leq\lambda_N\leq N-1.
\end{equation}
As an example of the use of $\Lambda(L)$, consider the ratio $\lambda_N(L)/\lambda_2(L)$: The synchronizability of a network of identical dynamical units is generally enhanced the smaller this ratio~\cite{pecora98}.
For our superjoiners networks, $\lambda_2(L)=1$, as the first node added always has a single outgoing link.
On the other hand, the value of $\lambda_N(L)$ generally decreases as a function of $\alpha$ and shows essentially no dependence on the reduction probability $r$. In fact, the dependence of $\lambda_N(L)$ on $\alpha$ can be estimated, approximately, from extreme-value statistics. A new node that joins at time $t$ can connect to up to $N(t)-1$ nodes. If we make the simplifying assumption that each node can connect to $N$, the final (larger) network size, then we would be overestimating $\lambda_N(L)$. With that assumption, we get
\begin{equation}\label{loglambda.eq}
\lambda_N(L)\approx \left(\frac{1}{N}-\frac{1}{N^\alpha}+\frac{1}{N^{\alpha-1}}\right)^{1/(1-\alpha)}\sim
\begin{cases}
N,&\alpha\ll1,\\
N^{1/(\alpha-1)},&\alpha\gg1.
\end{cases}
\end{equation}
Since the outgoing links statistics is not affected by $r$, neither is the value of $\lambda_N(L)$. In Fig.~\ref{spectra.fig} we show simulation results for $\lambda_N(L)$ of superjoiners nets of different $\alpha$, along with the prediction from extreme statistics. The figure confirms our analysis, and it shows that the synchronizability is better the higher the value of $\alpha$ is.
\begin{figure}[htbp]
\includegraphics[width=0.45\textwidth]{fig6}
\caption{(Color online) Largest Laplacian eigenvalue $\lambda_N(L)$ as a function of $\alpha$ in the superjoiners model averaged over $100$ network simulations per parameter combination. The dashed curve shows the analytical estimate given by Eq.~\eqref{loglambda.eq}.}
\label{spectra.fig}
\end{figure}
\section{Discussion}
Traditional network growth models allow each incoming node to connect to a capped number of existing nodes.
In this paper, we explored the possibility that the number of new connections $m$ is uncapped, but instead grows with the network size $N(t)$, and where $m$ is taken from a power-law distribution $P(m)\sim m^{-\alpha}$. Keeping our network model design close to the KR growth model~\cite{kr} has allowed us to explore analytically the competition between the exponent $\alpha$, characterizing the degree of new nodes, and the expected degree exponent $\gamma_{\rm KR}=1+1/r$ that arises from the rich-get-richer bias in the original KR model. For $\alpha>2$, our network model is sustainable ($M\sim N$),
the in-degree exponent is $\g_{\rm in}=\alpha$, the out-degree exponent is $\g_{\rm out}=1+1/r$, and the total-degree exponent $\gamma=\min\{\g_{\rm in},\g_{\rm out}\}$. For $\alpha<2$, the superjoiners network model is unsustainable ($M\sim N^{\beta}$, $\beta>1$) and the various exponents compete in interesting ways, as summarized in Fig.~\ref{phase_diagram.fig}. The superjoiners network model is extremely unsustainable, in the sense that conflicts that arise when trying to connect new nodes become common-place and analytically intractable, for $\alpha<2$ and $r>1/(3-\alpha)$.
Aside from the long-time asymptotic power-law distributions of the in-, out-, and total-degree, which we were able to derive analytically, the model exhibits rich transient behavior. Numerical integration of the master equation~(\ref{Nkl.eq}) predicts very nicely the results from simulations, but because of the three independent variables ($k$, $l$, and $t$) and computer time and memory constraints, this procedure is limited to rather small nets. On the other hand, we have shown that the in- and out-degrees correlate only weakly, and $N_{kl}(t)$ can then
be obtained from $F_k(t)$ and $G_l(t)$, allowing for easier analysis and numerical integration of substantially larger nets. A most important open question regarding transient behavior is deriving the specific ranges (of $k$, $l$, and $q$) where the long-time asymptotic predictions are valid. We expect that the noncorrelation phenomenon would be a great help in finding the answer.
\acknowledgements
This work is partially supported by the Simons Foundation Grant No.~318812 (J.~S.).
|
\section{Introduction}
In recent years the field of radio transients has seen much interest, most of which has been focused on $>$ 1 GHz radio emission \citep{Frail12,Lorimer07,Keane12,Thornton13,Wayth12}. Only a handful of blind searches have been carried out below 100 MHz \citep{Cutchin11,Lazio10,Kardashev77} leaving this a relatively unexplored region of the spectrum.
A recent study with the LWA1 yielded two promising transients below 40 MHz \citep{Obenberger14}. This search was focused on placing limits on prompt emission from gamma ray bursts and therefore only the times shortly after the occurrence of GRBs were searched. In total $<$ 200 hours of data were analyzed. The two detected transients were not associated with GRBs, but at the time of publication their origins were unknown. However the need to conduct further investigation was evident.
In this letter we present an analysis of over 11,000 hours of all-sky images recorded by the LWA1. Including the two previously mentioned events, a total of 49 transients have been detected in this data and there is strong correlation to large meteors known as fireballs.
\section{Observations}
The LWA1 is a radio telescope located in central New Mexico, operating between 10 and 88 MHz. It consists of 256 cross-dipole antennas spread over an ellipse of 100 $\times$ 110 m. There are also 5 additional antennas at distances ranging from 200 to 500 m from the center of the array \citep{Ellingson13,Taylor12}. The telescope is collocated with the Karl Jansky Very Large Array (VLA), at a latitude of 34.070$^{\circ}$ N and a longitude of 107.628$^{\circ}$ W.
PASI, a backend to the LWA1, correlates the signals from all antennas in real time, integrating for 5 s with 75 kHz bandwidth tunable to a center frequency anywhere within the 78 MHz over which the LWA1 operates. PASI also produces dirty images of the entire $\sim 2\pi$ sr sky above the LWA1 \citep{Obenberger14}, and in April 2012 began saving the images to a permanent archive\footnote{Prior to this the images were deleted after the generation of a movie comprising an entire day of images.}. This archive now contains 11,000 hours of all-sky images at various center frequencies. The vast majority of the data has been recorded at center frequencies of 37.8, 37.9, 52.0, and 74.0 MHz, with 30\%, 31\%, 17\%, and 13\% recorded at those frequencies, respectively.
\section{Discovery of Transients and Correlations with Fireballs}
Image subtraction algorithms have been developed to search for transients in the full 11,000 hours of PASI data. The underlying method is that from every image, the image 15 seconds prior is subtracted. Pixels above 6 $\sigma$ of the image noise are reported as candidate events and the equatorial coordinates are calculated. All four Stokes parameters and total spectra are saved and used to check if the candidate event is radio frequency interference (RFI). The coordinates of events are also checked against known bright sources which can be focused by the ionosphere at low frequencies \citep{Obenberger14}.
This method has resulted in the discovery of 22 transients at 37.8 MHz, 20 at 37.9 MHz, 1 at 29.9 MHz, 1 at 25.6 MHz, and none at either 52.0 MHz or 74.0 MHz. The majority of these transients display a fast rise and exponential decay, lasting for tens of seconds to up to a few minutes, and have flux densities ranging from 500 to 3500 Jy (Fig. 1). They also show constant power across the 75 kHz band and contain low polarization levels consistent with instrumental leakage of unpolarized sources.
\begin{figure}
\centering
\includegraphics[width = 7in]{lccomp_4.eps}
\caption{Light curves of the 9 brightest transients at 5 s integrations (Bottom). One of the transients shown on bottom (green) was also recorded using 1 s integrations (top), we show this light curve, offset by 2000 Jy, to illustrate that they this transient was smooth over 1 s timescales.}
\end{figure}
Most of the transients appear as point sources, meaning they are limited to $\leq4.4^{\circ}$ at 38 MHz. However some are extended over several degrees across the sky, most of which span less than ten degrees. In one case on January 21 2014, a source leaves a trail covering 92$^{\circ}$ in less than ten seconds (Fig. 2). The trail then slowly recedes to the end point which glows for $\sim$90 seconds. The only known source that could cover this distance across the sky in less than 10 seconds and leave a persistent trail is a fireball. To investigate this further, the transients were compared with the detections from NASA's All Sky Fireball Network\footnote{http://fireballs.ndc.nasa.gov}. The network consists of 12 all-sky cameras, two of which are situated in Southern New Mexico and share a portion of the sky with the LWA1. The cameras are used to determine the 3 dimensional position, speed, absolute magnitude\footnote{Stellar magnitude at zenith}, and mass of the of the fireballs.
\begin{figure}
\centering
\includegraphics[width = 7in]{PASI_sub}
\caption{Image of the sky, after subtraction, of the fireball which covered 92$^{\circ}$. The edge of the circle marks a cutoff of 25$^{\circ}$ above the horizon. Strong constant sources Cassiopeia A and the Virgo A, Taurus A, and the Galactic Plane have left weak residual signals in the image. }
\end{figure}
\begin{figure}
\centering
\includegraphics[width = 7in]{hist_hist}
\caption{(Top) A histogram showing the number of events per day of year. The full 365 days of the year are grouped into 40 bins, each bin representing 9.125 days. The red lines show the dates of several major meteor showers. The large clump which occurs near January 10 - 25 does not have a major meteor shower associated with it and may be evidence of a previously unknown fireball stream. (Bottom) Shown here are the normalized results of conducting 10$^{6}$ iterations of simulated data spread over the 7028 hours of data recorded at 25.6, 37.8, and 37.9 MHz. For every iteration 45 events were generated within the actual observing times, and the mean delay of all the events to the closest meteor shower was calculated. The actual measured delay of 2.3 is shown in red.}
\end{figure}
While the two New Mexico stations are too far South East of the LWA1 to have detected the extended transient candidate, 5 of the other 44 transients correlate in both space and time to fireballs. The fireball network could not have seen the remaining 39 events because they were either too far North West or occurred during the day. Figure 3 shows a histogram of the events, demonstrating that groups of transient detections appear around the times of meteor showers. This implies that a large fraction of the events not seen by the network are most likely meteors as well.
\begin{figure}
\centering
\includegraphics[width = 7in]{PASI_sub_FBN}
\caption{Comparison images from the NASA All-Sky Fireball Network station located in Mayhill, NM (Left) and PASI (Right). The red rectangles outline the fireballs in the Fireball Network images. The red dots on the PASI images show the final location of the fireball as provided by the Fireball Network. For the PASI images, the edge of the circle marks a cutoff of 25$^{\circ}$ above the horizon. East-West orientation of the PASI images have been reversed from their normal appearance to have the same orientation as the Fireball Network Images. The large bright disk in the lower left image is the moon.}
\end{figure}
The mean delay between an event and the nearest previous major meteor shower in Figure 3 is 2.3 time bins\footnote{Each time bin represents 9.125 days}. In order to quantify the probability of this distribution with respect to the meteor showers, we performed a Monte Carlo simulation. Random dates were generated and checked if they occurred during actual PASI observing time at 25.6, 37.8, or 37.9 MHz. If they did occur during times at these frequencies, they were saved into an array, otherwise they were discarded. Once 44 events were compiled, the average delay between the simulated events and the closest of the four major meteor showers was calculated. This was performed 10$^{6}$ times and a normalized histogram of the results are shown in Figure 3. The actual data lies between 2 and 3 bins which has a probability of 0.2\%.
Older events from the fireball network were used to search locations and times for our oldest data, which was taken at 25.6 MHz and was only saved as movie files which are not searchable with the image subtraction algorithm\footnote{Figure 3 only represents the 44 events found with the image subtraction algorithm, since the other 5 events were found by looking for specific fireballs rather than by a blind search. The inclusion of these would therefore skew the statistics.}. In this data an additional five extremely bright correlated events were found, one of which covered $\sim$ 65$^{\circ}$ passing through zenith and was of comparable brightness to the galactic plane. These additional events bring the total number to 49, 10 of which are correlated directly to fireballs.
The fireballs themselves are of the more energetic variety seen by the fireball network. The velocities of all 10 were above 50 km s$^{-1}$, with an average velocity of 68 km s$^{-1}$. The observed optical emission of each event lasted for $\sim$ 1 s, and was followed shortly (usually within 1 PASI integration) by the onset of the radio emission. The four seen near 38 MHz had peak absolute visual magnitudes of $\sim$ $-$4.8, $-$6.1, $-$6.3, and $-$6.4. The six seen at 25.6 MHz had peak absolute visual magnitudes of $\sim$ $-$3.3, $-$4.1, $-$4.3, $-$4.7, $-$6.6, and $-$7.1. Meteors with visual magnitudes brighter than $-$4 occur about once every 20 hours \citep{Halliday96}. The chance overlap in space and time of 5 events\footnote{The 5 events found by using times and locations from the fireball network were not used in the calculation because they were not found randomly and therefore cannot be added to the random population of PASI transients.} found randomly in PASI data with fireballs brighter than magnitude $-$4 within the 11,000 hours of data is about 1 in 10$^{28}$. Therefore it is clear that these correlations are not mere coincidence but, indeed, connected to fireballs. Figure 4 shows images of two fireballs observed simultaneously by the Fireball Network and the LWA1.
\section{Reflection vs Emssion} The ionized trails left by meteors have long been known to scatter radio waves, and since the 1940s the use of radar echoes became a popular method for observing their trails allowing the detection of very dim as well as daytime meteors unobservable by optical instruments. A large fraction of these observations have been of specular trails, which occur when a line perpendicular to the meteor's path satisfies the reflection requirement that the angle to the transmitter equals the angle to the receiver \citep{Wislez95}. The majority of these reflections last for $<$ 1 s, have electron line densities of $\alpha < 2.4 \times 10^{12}$ cm$^{-1}$, and are known as underdense meteors. A less common variety of specular trail last for several seconds, have electron line densities of $\alpha > 2.4 \times 10^{12}$ cm$^{-1}$, and are known as overdense meteors. These are associated with larger meteors, the largest of which are associated with optical fireballs \citep{Ceplecha98,Bronshten83,McKinley61}.
A smaller subset of meteor trail echoes are non-specular (i.e. due to scattering), which are not yet well understood. They most often occur when the radar is pointed perpendicular to the geomagnetic field, but do not necessarily satisfy the specular reflection requirement. These types of trails can last from seconds to several minutes, but are typically much weaker than specular trails and therefore high power large-aperture radars are required to detect them \citep{Bourdillon05,Sugar10,Close11}.
Meteor trail echoes are a well studied phenomenon, and their characteristics are well documented. Moreover a recent study by \citet{Helmboldt14} using 55.25 MHz analog TV broadcasting stations for meteor scatter provides a direct example of how to detect meteor echoes with the LWA1 and how to identify them in the all-sky data. In comparison, several key differences arise between the transients reported in this paper and what is expected from trail echoes. The differences are as follows:
First, typical transmitters are strongly polarized, resulting in reflections that are strongly polarized \citep{Helmboldt14,Close11,Wislez95}. However no significant amount of linear nor circular polarization has been detected from the observed transients.
Secondly, a large portion of the RFI seen by LWA1 is narrower than the 75 kHz PASI band and is easily identifiable by its spectra \citep{Obenberger11}. Yet none of the observed transients contain any spectral features.
Thirdly, the light curves of the observed transients are consistent with each other, ranging from 30 to 150 s, showing a linear rise and a long exponential decay, and and otherwise show a smooth evolution (Figure 1). A typical reflection from an overdense trail reaches maximum brightness in just a few seconds, maintains a relatively constant average brightness while undergoing sporadic dimming and rebrightening. It then quickly decays away once it expands to the point that the density reaches the underdense criteria \citep{Ceplecha98,Wislez95,Helmboldt14}. The observed transients are also inconsistent with light curves from non-specular echoes, which vary greatly from one to the next, following no particular pattern. More importantly, however, non-specular echoes are weaker and more rare than overdense specular echoes \citep{Bourdillon05,Close11}. Therefore if the LWA1 were seeing non-specular reflections it should also see many more bright specular reflections scattering from the same transmitters.
Finally, the observed transients have azimuths and elevations consistent with the uniform distribution convolved with the LWA1 power pattern (Figure 5). This distribution implies that the sources appear in random locations with no preferable sky position. This is inconsistent with what is expected from specular echoes of man-made radio frequency interference (RFI), which should increase towards the horizon due to the increased number of incident angles with distant transmitters required for forward scattering \citep{Wislez95}. The observed pattern could be consistent with nearby transmitters. However, because the signal strength depends on the inverse-cube of the distance to the meteor, there should be many very bright nearby RFI sources on the horizon but these are not observed. This pattern is also inconsistent with non-specular reflections, which are preferentially located in a relatively small region of the sky that satisfies the requirement that the pointing vector is perpendicular to the geomagnetic field \citep{Bourdillon05,Close11}.
\begin{figure}[!]
\centering
\includegraphics[width = 7in]{zhist}
\caption{A histogram showing the number of events as a function of zenith angle. The blue bars show the actual measured distribution of the transients. The red bars show a model based on a uniform distribution convolved with the power pattern of the LWA1. }
\end{figure}
For these reasons it seems unlikely that forward scattering is responsible for the signals detected from fireballs. It is therefore our conclusion that fireball trails radiate at low frequencies.
\section{Physical Constraints on the Emission}
Since PASI can only record at one center frequency at a time, there is limited spectral information for each event. Nevertheless the similarity of the light curves of fireballs recorded at 25.6, 29.9, 37.8, and 37.9 MHz implies that this emission is broad band. It also appears that they are brighter at 25.6 than at 38 MHz, suggesting a spectral slope favorable to lower frequencies. Also, there have been no detections at 52.0 or 74.0 MHz which may be the result of a sharp cutoff at higher frequencies or a steep spectrum. In either case it is clear from these observations that the emission is non-thermal, yet the exact mechanism is currently unknown.
If a magnetic field of 10 to 15 G were present within the trail, it follows that cyclotron radiation would be emitted at the observed frequencies by the electrons in the plasma. However, the surface geomagnetic field is only 0.5 G, so this would require the generation of a strong magnetic field by a fireball, an effect that has never been observed.
Brightness temperature for a radio source can be calculated using the equation:
\begin{equation}
T_{b} = \frac{S_{\nu} \lambda^{2}}{2 k \Omega}
\end{equation}
Where $S_{\nu}$ is the flux density of the source, $\lambda$ is the observed wavelength, $k$ is the Boltzmann constant, and $\Omega$ is the solid angle of the source. The solid angle of the emitting portion of the fireball is estimated by using typical size scales reported for bright meteors trails.
A typical fireball observed by the LWA1 has a peak flux density of 1 kJy, 20 s after first light, and most are point sources with a measured beam size of $ 4.4^{\circ}$ at 37.8 MHz. Meteor plasma trails are typically modeled as long cylinders with radii much smaller than their length. Assuming a height of 110 km and zenith angle of 40$^{\circ}$, an angular size of $ 4.4^{\circ}$ corresponds to 12 km, which is a reasonable value for the length of the emitting part of the plasma trail. However for large plasma trails a typical initial radius is 10 m, which quickly expands due to diffusion. It is estimated that at 20 s after entry, the radius of the expanding trail would be $\sim$100 m \citep{Ceplecha98,Bronshten83,McKinley61}. Given these dimensions the estimated brightness temperature is $3 \times 10^{5}$ K. This temperature is two orders of magnitude higher than the typical peak temperature of a fireball trail, and this provides further evidence that the emission is non thermal.
While the spectral slope has yet to be fully measured, the total energies of the radio emission are estimated by assuming a flat spectrum from 10 to 50 MHz and heights of 110 km. These assumptions yield total radio energy estimates ranging from 10$^{-4}$ to 10$^{-2}$ J, which is one part in $10^{12}$ to $10^{10}$ of the kinetic energy of a typical fireball. Further observations are in progress to better characterize the spectral properties of this emission, and to get better estimates of the total amount of energy radiated at radio frequencies.
\section{Discussion}
Decametric radio emission from meteors has not been previously detected, but this is not the first time its existence has been discussed. \citet{Hawkins58} conducted a search for radio emission, but reported only upper limits with a 5 $\sigma$ sensitivity of $\sim 10^{8}$ Jy at 30 MHz for 1 s bursts. It is also interesting to note that in the last several decades detections of extremely low frequency (ELF\footnote{3 Hz to 3 kHz}) and very low frequency (VLF\footnote{3 kHz to 30 kHz}) emission have been reported coincident with large meteors \citep{Guha12,Keay80,Beech95}. The physical mechanism responsible for this emission is not well understood, but might be related to our detections of higher frequency emission.
Given the vast range in energies and size scales of meteors and their corresponding plasma trails, this emission may exist at a wide range of frequencies, timescales, and energies. Investigating this emission further will yield new insights into the physics of meteors and their interaction with our atmosphere. Moreover fireballs are now a known radio transient foreground source and need to be taken into account when searching for cosmic transients. It is interesting to note that transient atmospheric phenomena, unknown to emit at radio frequencies, have been proposed as the possible source of Perytons and perhaps even Fast Radio Bursts \citep{Kulkarni14,Katz14,Burke-Spolaor11,Thornton13,Lorimer07}. In fact Katz (2014) suggests meteors as a possible source of Perytons. In this paper we reported observations of high energy meteors generating bright radio emission.
\section{Acknowledgments} Construction of the LWA1 has been supported by the Office of Naval Research under Contract N00014-07-C-0147. Support for operations and continuing development of the LWA1 is provided by the National Science Foundation under grants AST-1139963 and AST-1139974 of the University Radio Observatory program.
This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
Part of this research was conducted at the Jet Propulsion Laboratory, California Institute of Technology, under contract to NASA
\bibliographystyle{plainnat}
|
\section{Introduction and main results}
We consider an open quantum system consisting of a small, finite-dimensional part interacting with a heat bath, modeled by a spatially infinitely extended free Bose gas in thermal equilibrium. The analysis of such systems, and especially of their dynamics, has a long tradition. The reduced dynamics of the small system alone is described in the theoretical physics literature primarily using master equation techniques, which rely on approximations that are not controlled mathematically, but are very popular and successful nevertheless \cite{JZKGKS,Schl,GZ, BreuerPetruccione,NC}. A rigorous approach is the van Hove, or weak coupling limit \cite{Davies1,Davies2,AL}. It describes the dynamics of the small system for times up to the order of $\lambda^{-2}$, where $\lambda$ is the strength of the system-environment coupling. Given a fixed $\lambda$, the time-asymptotics, $t\rightarrow\infty$, cannot be resolved with the weak coupling method. It is shown in \cite{DK} however that, for a class of open systems, if the conditions for the weak coupling limit are satisfied, then the small subsystem converges to a final state in the long time limit.
The analysis of the {\em total} system -- the small system plus the reservoir -- is more delicate than that of the small subsystem alone. Over the last decade and a half, a perturbation theory based on quantum resonance methods has been developed to deal with this problem, see \cite{JP,BFS,M,JP1,DJ,FM,MMS,MMS1,MSB}. It is implemented in various forms, using spectral deformation, positive commutator and renormalization group techniques and permits a mathematically rigorous treatment of the full dynamics (system plus reservoir), for {\em fixed, small} coupling $\lambda$ and for {\em all} times $t\geq 0$. Other than the spectral approach of the above references, the polymer expansion method of \cite{DK} allows the analysis the total system as well, see \cite{DK2}.
\medskip
The techniques of the above works are based on a perturbation theory in the system-reservoir coupling parameter $\lambda$. The latter is assumed to be small relative to the spacing $\sigma>0$ between the energy levels of the small system: $|\lambda|<\!\!<\sigma$. This is the {\em isolated resonances regime}. However, there are many physical systems for which this condition is not valid. For instance in {\em complex open systems}, the small system itself is composed of many individual parts (particles) and the energy level spacing may become very small. Take the Hamiltonian of a system of $N$ spins, having $2^N$ eigenvalues. The total energy of the spins is of the order of $N$. The generic energy spacing is thus of the order of $\sigma\sim N/2^N$, which is exponentially small in $N$. For such systems, the condition $|\lambda|<\!\!<\sigma$ is not reasonable.
In the present work, we develop the resonance method in the {\em overlapping resonances regime} $\sigma<\!\!< |\lambda|$. We study here the simplest case, in which all the system energies lie close together relative to $|\lambda|$. Our results hold for a fixed, finite (but arbitrary) dimension $N$ of the small system and for small coupling constants, $|\lambda|\leq\lambda_0$, for some $\lambda_0>0$.
The $N$-level system coupled to a thermal reservoir is described by the Hamiltonian
$$
H^\Lambda(\sigma,\lambda) = \sigma H_{\rm S} +H_\r^\Lambda +\lambda G\otimes\Phi^\Lambda(g),
$$
acting on the Hilbert space ${{\mathbb C}}^N \otimes{\cal F}(L^2(\Lambda,\d^3x))$, where the second factor is the Fock space over the one-particle Hilbert space of wave functions localized in a finite box $\Lambda\subset{\mathbb R}^3$. The system Hamiltonian $H_{\rm S}$ is an arbitrary self-adjoint operator on ${{\mathbb C}}^N$. The reservoir Hamiltonian $H^\Lambda_\r$ is the second quantization of the single Boson energy, the self-adjoint Laplace operator with periodic boundary conditions. The system-reservoir interaction is the product of a self-adjoint $G$ acting on the system and the field operator $\Phi^\Lambda(g)=\frac{1}{\sqrt 2}(a^*(g)+a(g))$, where $a^*, a$ are the creation and annihilation operators on ${\cal F}(L^2(\Lambda,\d^3x))$, smoothed out with the form factor $g$ supported in $\Lambda$. The Hamiltonian contains the two parameters $\sigma\geq 0$ and $\lambda\in{\mathbb R}$, the system energy level splitting parameter and the interaction strength, respectively. The bosonic field is initially in its {\em thermal equilibrium} state at positive temperature $1/\beta$, given by the density matrix $\rho^\Lambda_{\r,\beta}\propto{\rm e}^{-\beta H_\r^\Lambda}$. In order to have a true open system, one performs the infinite-volume limit of the reservoir, in which the box $\Lambda$ grows to all of ${\mathbb R}^3$. More precisely, the expectation values of observables (Weyl operators) of the reservoir, in the thermal state, have a limit as $\Lambda\rightarrow{\mathbb R}^3$. This defines the infinite-volume equilibrium state $\omega_{\r,\beta}$ by its expectation values $\omega_{\r,\beta}(W(f))$ on the Weyl operators. A Hilbert space on which that state is represented by a vector can then be reconstructed using the Gelfand-Naimark-Segal (GNS) construction \cite{AW}. This procedure leads to the description of the coupled system as a $W^*$-dynamical system \cite{BR,OQS1}. It consists of a Hilbert space
\begin{equation}
{\cal H} = {\cal H}_{\rm S}\otimes{\cal H}_\r,
\label{m2}
\end{equation}
of a von Neumann algebra of observables
\begin{equation}
{\frak M} = {\frak M}_{\rm S}\otimes {\frak M}_\r,
\label{m3}
\end{equation}
and of a Heisenberg dynamics of ${\frak M}$,
\begin{equation}
A\mapsto \alpha_{\sigma,\lambda}^t(A) = {\rm e}^{\i tL(\sigma,\lambda)} A{\rm e}^{-\i tL(\sigma,\lambda)}, \qquad A\in{\frak M}.
\label{fullLiouville}
\end{equation}
The Liouvillian $L(\sigma,\lambda)$ is a self-adjoint operator on ${\cal H}$.
The small system is an $N$-level system having a Hamiltonian $H_{\rm S}$. In the GNS (Gelfand-Naimark-Segal) representation, the Hilbert space is ${\cal H}_{\rm S}={\mathbb C}^N\otimes{\mathbb C}^N$ and the algebra of observables is given by ${\frak M}_{\rm S} = {\cal B}({ {\mathbb C}^N})\otimes\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{{\mathbb C}^N}$ (bounded linear operators). The dynamics is implemented as
\begin{equation}
A_{\rm S}\mapsto {\rm e}^{\i t L_{\rm S}} (A_{\rm S}\otimes\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{{\mathbb C}^N}) {\rm e}^{-\i tL_{\rm S}},\qquad A_{\rm S}\in {\cal B}({\mathbb C}^N),
\label{m10}
\end{equation}
where
\begin{equation}
L_{\rm S} = H_{\rm S}\otimes\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{{\mathbb C}^N} - \mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{{\mathbb C}^N}\otimes {\cal C}H_{\rm S}{\cal C}
\label{m11}
\end{equation}
is the self-adjoint system Liouville operator. Here, $\cal C$ is the operator taking the complex conjugate of components of vectors represented in the orthonormal eigenbasis $\{\varphi_a\}_{a=1}^N$ of the interaction operator,
\begin{equation}
G\varphi_a=g_a\varphi_a,\quad a=1,\ldots,N.
\label{m22}
\end{equation}
The procedure of doubling of the Hilbert space is well known in the physics literature, also called the `Liouville Representation', see e.g. \cite[Chapter 3]{Muk}.
The reservoir state is the thermodynamic (infinite volume) limit of a free Bose gas in equilibrium at inverse temperature $\beta$. Its Hilbert space representation has first been constructed in \cite{AW} and a unitarily equivalent representation, suitable for the use of spectral translation techniques, has been given in \cite{JP}. The GNS Hilbert space is ${\cal H}_\r = {\cal F}\big( L^2({\mathbb R}\times S^2,\d u\times\d\vartheta)\big) = \oplus_{n\geq 0} L_{{\rm symm}}^2(({\mathbb R}\times S^2)^n,(\d u\times\d\vartheta)^n)$,
the symmetric Fock space over the one-particle function space $L^2({\mathbb R}\times S^2,\d u\times\d\vartheta)$. Here, $\d\vartheta$ is the uniform measure on $S^2$. The thermal field operator is given by
\begin{equation}
\Phi(f_\beta)=\frac{1}{\sqrt 2}\big( a^*(f_\beta) +a(f_\beta)\big),
\label{m5}
\end{equation}
where $a^*(f_\beta)= \int_{{\mathbb R}\times S^2} f_\beta(u,\vartheta) a^*(u,\vartheta)\ \d u\d\vartheta$ is the creation operator acting on the Fock space ${\cal H}_\r$ and $a(f_\beta)$ is its adjoint, smoothed out with $f_\beta\in L^2({\mathbb R}\times S^2,\d u\times\d\vartheta)$ defined by
\begin{equation}
f_\beta(u,\vartheta):=\sqrt{\frac{u}{1-{\rm e}^{-\beta u}}}\ |u|^{1/2}\left\{
\begin{array}{ll}
f(u,\vartheta), & \hbox{if}\,\, u\geq0, \\
\overline{f}(-u,\vartheta), & \hbox{if}\,\, u<0.
\end{array}
\right.
\label{glue}
\end{equation}
Here, $f\in L^2({\mathbb R}^3,\d^3k)$ is represented in polar coordinates (and in Fourier space). The thermal Weyl CCR algebra ${\frak M}_\r\subset{\cal B}({\cal H}_\r)$ is the von Neumann algebra generated by the unitary Weyl operators $W(f_\beta):={\rm e}^{\i\Phi(f_\beta)}$. The dynamics on ${\frak M}_\r$ is given by the Bogoliubov transformation $t\mapsto W({\rm e}^{\i t u}f_\beta) = {\rm e}^{\i t L_\r} W(f_\beta) {\rm e}^{-\i tL_\r}$.
It is implemented by the self-adjoint reservoir Liouvillian
\begin{equation}
L_\r = \d\Gamma(u):= \int_{{\mathbb R}\times S^2} u\ a^*(u,\vartheta) a(u,\vartheta) \d u\d\vartheta,
\label{m8}
\end{equation}
the second quantization of the operator of multiplication by $u\in\mathbb R$.
The vacuum vector $\Omega_\r\in{\cal H}_\r$ represents the $\beta$-KMS state w.r.t. the dynamics generated by \fer{m8}.
The Liouville operator $L(\sigma,\lambda)$ determining the full dynamics, \fer{fullLiouville}, has the form
\begin{equation}
L(\sigma,\lambda)=L_0(\sigma)+\lambda V,
\label{m11'}
\end{equation}
with a free part
\begin{equation}
L_0(\sigma) = \sigma L_{\rm S}+L_\r
\label{m12}
\end{equation}
(see \fer{m11}, \fer{m8}) and where the system-reservoir interaction is
\begin{equation}
\lambda V=\lambda G\otimes\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{{\mathbb C}^N}\otimes\Phi(g_\beta).
\label{m13}
\end{equation}
Here, $\sigma$ and $\lambda$ are two real parameters, $G$ is a self-adjoint matrix on ${\mathbb C}^N$ and $g_\beta\in L^2({\mathbb R}\times S^2)$ is obtained from a {\it form factor} $g\in L^2({\mathbb R}^3)$ using the relation \fer{glue}. It is well known that $L(\sigma,\lambda)$ is self-adjoint for all $\lambda,\sigma\in\mathbb R$ (this can be proven by the Glimm-Jaffe-Nelson commutator theorem, see e.g. \cite[Theorem A.2]{M}). We assume the following regularity of the form factor.
\medskip
{\bf Assumption A1.} (Analyticity) There is a $\theta_0>0$ such that $\theta\mapsto
g_\beta(u+\theta,\vartheta)$ has an analytic extension to the domain $\{\theta\in{\mathbb C}\ :\ |\theta|<\theta_0\}$, as a map from
$\mathbb C$ to $L^2({{\mathbb R}}\times S^2,\d u\times \d\vartheta)$.
\medskip
{\bf Assumption A2.} (Ultra-violet decay) There is an $\epsilon > 0$ such that ${\rm e}^{a|k|} g(k)\in L^2({\mathbb R}^3,\d^3k)$ for an $a>(1/2+\epsilon)\beta$, where $\beta$ is the inverse temperature.
\medskip
Examples of form factors satisfying this condition are $g(r,\vartheta)=r^p{\rm e}^{-a r^m}g_1(\vartheta)$ (polar coordinates in ${\mathbb R}^3$), where $p=-1/2+n$, $n=1,2,\ldots$, $m=1,2$, and $g_1(\vartheta)\in{\mathbb R}$ (see also \cite{FM} for more general classes of admissible $g$). More generally, we charaterize the infrared behaviour of the form factor by $p\geq -\frac{1}{2}$ satisfying $0<\lim_{|k|\rightarrow 0}\frac{|g(k)|}{|k|^p}=C<\infty$. The value of $p$ depends on the physical model considered. For quantum optical systems, $p=1/2$, for the quantized electromagnetic field, $p=-1/2$. We define the complex numbers
\begin{equation}
\delta_{a,b}=-\textstyle\frac{1}{2}(g_a^2-g_b^2)\scalprod{g}{|k|^{-1}g}+\i \frac{\pi}{2} (g_a-g_b)^2
\left
\{\begin{array}{lll}
0 & \mbox{\rm if $p>-1/2$}\\
\xi(0)>0 & \mbox{\rm if $p=-1/2$}
\end{array},
\right.
\label{delta}
\end{equation}
for $a,b=1,\ldots,N$ and where
\begin{equation}
\xi (0)=\lim_{\epsilon\downarrow
0}\frac{1}{\pi}\int_{{\mathbb{R}}^3}\coth(\frac{\beta|k|}{2})|g(k)|^2\frac{\epsilon}{|k|^2+\epsilon^2}\d^3k.
\label{xinot}
\end{equation}
The $\lambda^2\delta_{a,b}$ are the {\em resonance energies} for $\sigma=0$, see Theorem \ref{thm3} below. The following assumption simplifies the presentation of our results.
\medskip
{\bf Assumption A3.} (Non-degeneracy) The spectrum $\{g_a\}_{a=1}^N$ of $G$ is such that all non-zero $\delta_{a,b}$ are distinct.
\medskip
Our analysis is readily generalized to the case of degenerate resonances (see the proof of Theorem \ref{thm4}). Indeed, we do this for the spin-boson model, in which the two non-zero resonances are given by $\delta_{1,2}=\delta_{2,1}=\i\frac\pi2 \xi(0)$.
The following is a well-coupledness condition which we will assume for some results. It implies that the coupled system has a unique stationary state (the coupled equilibrium).
\medskip
{\bf Assumption A4.} (Fermi Golden Rule Condition) For all $a,b$, $a\neq b$, we have ${\rm Im}\delta_{a,b}>0$ and $\scalprod{\varphi_a}{H_{\rm S}\varphi_b}\neq 0$.
\medskip
We show in Appendix A that the manifold of normal $\alpha_{0,\lambda}^t$-invariant states on $\frak M$ is the convex span of the states $\omega_{a}=\omega_{{\rm S},a}\otimes\omega_{\r,a}$, $a=1,\ldots,N$. Here, $\omega_{{\rm S},a}$ is given by the rank-one density matrix $|\varphi_a\rangle\langle\varphi_a|$ (spectral projection associated to $G$), and $\omega_{\r,a}$ is a normal perturbation of the reservoir equilibrium state, explicitly given in \fer{mexpl}. When $\sigma>0$ is small, then there is a unique (normal) $\alpha_{\sigma,\lambda}^t$-invariant state on $\frak M$, namely, the coupled system-reservoir equilibrium state $\omega_{\beta,\sigma,\lambda}$, which is an $(\alpha^t_{\sigma,\lambda},\beta)$-KMS state.
\medskip
Our main result, summarized in Theorem \ref{dynthm} below, concerns the dynamics of initial conditions and observables taken from sets ${\cal S}_0$ and ${\frak M}_0$, respectively. ${\cal S}_0$ is a set of bounded linear functionals on ${\frak M}$ (defined in \fer{anstates}), dense in the set of all states of ${\frak M}$. All states of the form $\omega_{\rm S}\otimes\omega_{\r,\beta}$ are in ${\cal S}_0$, where $\omega_{\rm S}$ is an arbitrary state on ${\frak M}_{\rm S}$ and $\omega_{\r,\beta}$ is the equilibrium state of the reservoir. ${\frak M}_0$ is the collection of translation analytic elements of ${\frak M}$, a dense set in ${\frak M}$, see \fer{anstates}. All observables $A=A_{\rm S}\otimes\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_\r$ of the system alone are in ${\frak M}_0$. To express the details of the evolution, we introduce the following. For $a,b=1,\ldots,N$, $a\neq b$, set
\begin{eqnarray}
\lefteqn{
\eta_{a,b}(\sigma,\lambda) = \lambda^2\delta_{a,b} + \sigma\big([H_{\rm S}]_{a,a}-[H_{\rm S}]_{b,b}\big)} \nonumber\\
&& \qquad\qquad- \frac{\sigma^2}{\lambda^2}\Big(\sum_{c=1,\ldots,N; c\neq a}\frac{|[H_{\rm S}]_{a,c}|^2}{\delta_{c,b}-\delta_{a,b}}+ \sum_{c=1,\ldots,N; c\neq b}\frac{|[H_{\rm S}]_{b,c}|^2}{\delta_{a,c}-\delta_{a,b}}\Big),
\label{pertevnon-zero}
\end{eqnarray}
where $[H_{\rm S}]_{b,c}=\scalprod{\varphi_b}{H_{\rm S}\,\varphi_c}$ are the matrix elements of the system Hamiltonian. For $a=1,\ldots,N$, set
\begin{equation}
\eta_{a,a}(\sigma,\lambda) = 2\i\frac{\sigma^2}{\lambda^2}\xi_a,
\label{pertevzero}
\end{equation}
where $\xi_a\geq 0$ are the eigenvalues of the real symmetric $N\times N$ matrix $T$ with matrix elements
\begin{equation}
{}[T]_{a,b}=
\left\{
\begin{array}{cl}
\displaystyle -\frac{{\rm Im}\delta_{a,b}}{|\delta_{a,b}|^2} |[H_{\rm S}]_{a,b}|^2, & \mbox{if $a\neq b$}\\
& \\
\displaystyle \sum_{c=1,\ldots,N; c\neq a} \frac{{\rm Im}\delta_{a,c}}{|\delta_{a,c}|^2} |[H_{\rm S}]_{a,c}|^2, & \mbox{if $a=b$}.
\end{array}
\right.
\label{mmm6}
\end{equation}
The vector $\frac{1}{\sqrt N}(1,\ldots,1)$ is in the null space of $T$. We enumerate the eigenvalues of $T$ s.t. $\xi_1=0$. Under Assumption A4, zero is a simple eigenvalue of $T$ (see after \fer{mmm5} for a proof). We show in Theorem \ref{thm4} that, for $\sigma<\!\!<|\lambda|$, the resonances are given by
\begin{eqnarray}
\varepsilon_{a,b}( \sigma,\lambda)&=& \eta_{a,b}(\sigma,\lambda) + O\left(\sigma^2 |\lambda|^{-1}\right)+ O_\lambda(\sigma^3) \label{l7}\\
\varepsilon_a(\sigma,\lambda)&=& 2\i\frac{\sigma^2}{\lambda^2} \xi_a + O\left(\sigma^2 |\lambda|^{-1}\right)+ O_\lambda(\sigma^3).\label{17'}
\end{eqnarray}
Here, $O_\lambda(\sigma^3)$ is a term $f(\lambda,\sigma)$ satisfying $\limsup_{\sigma\rightarrow 0}\sigma^{-3}\|f(\lambda,\sigma)\|=C_\lambda<\infty$.
\begin{thm}[Dynamics in the overlapping resonances regime.]
\label{dynthm}
Assume A1-A4. There is a constant $\lambda_0>0$, such that for $0<|\lambda|<\lambda_0$, the following holds. There is a $\sigma_0>0$ (depending on $\lambda$) such that for $0\leq \sigma< \sigma_0$ and for any $\omega_0\in{\cal S}_0$, $A\in{\frak M}_0$, $t\geq 0$, we have
\begin{equation}
\omega_0\big(\alpha^t_{\sigma,\lambda}(A)\big) = \omega_{\beta,\sigma,\lambda}(A)+\sum_{a=2}^N {\rm e}^{\i t\varepsilon_a(\sigma,\lambda)} \chi_a(A) + \sum_{a,b=1\atop a\neq b}^N {\rm e}^{\i t\varepsilon_{a,b}(\sigma,\lambda)} \chi_{a,b}(A) +O({\rm e}^{-\gamma t}).
\label{1}
\end{equation}
The $\chi_a$, $\chi_{a,b}$ in \fer{1} are linear functionals on ${\frak M}_0$. They depend on $\sigma,\lambda$ and the initial condition $\omega_0$, but not on $t$. The decay rate $\gamma>0$ is independent of $\lambda,\sigma$ and satisfies $\gamma>\max\{{\rm Im}\varepsilon_a,{\rm Im}\varepsilon_{a,b}\}$.
\end{thm}
{\bf Discussion.\ } The imaginary parts ${\rm Im}\varepsilon_{a,b}\propto \lambda^2$ and ${\rm Im}\varepsilon_a\propto \sigma^2/\lambda^2$ (to leading order) have the associated decay times $t_1\propto \lambda^{-2}<\!\!< t_2\propto\lambda^2/\sigma^2$. The representation \fer{1} thus paints the following picture. In the non-degenerate situation, $\sigma>0$, the remainder term becomes negligible very quickly, for $t> t_0=1/\gamma$. Then, for $t> t_1$ the sum over the $\chi_{a,b}$ becomes small as well. Finally, for $t> t_2$, the first sum becomes negligible and in the limit $t\rightarrow\infty$, the system is in the coupled equilibrium $\omega_{\beta,\sigma,\lambda}$. In the degenerate situation, $\sigma=0$, the remainder term is small again after times $t> t_0$, and again after times $t> t_1$, the second sum in \fer{1} is negligible. However, since $\varepsilon_a(0,\lambda)=0$, the first sum is independent of time and does not decay. The initial state $\omega_0$ (applied to ${\frak M}_0$) converges thus to the final state $\omega_\infty=\omega_{\beta,0,\lambda}+\sum_{a\geq 2}\chi_a$. The final state $\omega_\infty$ depends on the initial state $\omega_0$. It belongs to the manifold of $\alpha^t_{0,\lambda}$-invariant states on ${\frak M}$, i.e., it is a convex combination $\sum_a\mu_a(\omega_0)\,\omega_{{\rm S},a}\otimes\omega_{\r,a}$, with initial state dependent mixing parameters $\mu_a$.
Therefore, two time-scales emerge for the dynamics of systems in the overlapping resonances regime. On a time-scale $t_1\propto \lambda^{-2}$, which is very short with respect to $t_2
\propto \lambda^2/\sigma^2$, the initial state approaches a quasi-stationary manifold given by the first two terms on the r.h.s. of \fer{1}. For $\sigma=0$, this manifold is exactly stationary, but for $\sigma>0$ small, the manifold is only approximately stationary and it decays (into a the single equilibrium) for times exceeding $t_2\propto\lambda^2/\sigma^2$.
The appearence of different time-scales in open systems (albeit in somewhat different situations) has been observed before. The paper \cite{Dnew} examines the dynamics of a particle attracted by two widely separated potential wells and interacting with an infinite reservoir. The spacing of the wells, $1/\mu$, and the particle-reservoir interaction $\lambda$ are related by $\mu=\lambda^\beta$. It is shown that the dynamics of the particle in the weak coupling limit exists. The interaction between the wells has no effect for times of order $1/\lambda^2$ for $\beta>2$. However, for $0<\beta<2$ it has a direct effect on the particle dynamics and modifies the decay induced by the reservoir alone. The set of invariant states in the two regimes for $\beta$ are different.
In \cite{A}, various master equations for the dynamics of a nonlinear oscillator interacting with a reservoir are investigated. It is found that different generators yield more accurate descriptions of the reduced oscillator dynamics for different time-scales. In particular, different generators should be used for times shorter than, and longer than, the inverse of the system level-spacing. We mention that our analysis is valid for the {\em total} system-reservoir dynamics and for {\em all} times $t\geq 0$.
\medskip
{\bf Reduced dynamics.\ } Consider initial states of the form $\omega_0=\omega_{{\rm S},0}\otimes\omega_{\r,\beta}$, where $\omega_{{\rm S},0}$ is a state given by an arbitrary density matrix $\rho_0$ on ${{\mathbb C}}^N$, $\omega_{{\rm S},0}(A)={\rm Tr}_{\rm S}(\rho_0 A)$. The {\em reduced density matrix} $\rho_t$ of the system at time $t\geq0$ is defined by
$$
{\rm Tr}_{\rm S}(\rho_t A) = \omega_{{\rm S},0}\otimes\omega_{\r,\beta}\big(\alpha^t_{\sigma,\lambda}(A)\big), \qquad\forall A\in{\cal B}({{\mathbb C}}^N),
$$
where the trace is taken over the system space ${{\mathbb C}}^N$. We denote the reduced evolution of the system by
$$
T_{\sigma,\lambda}(t)\rho_0 = \rho_t,
$$
and the manifold of initial system states which are invariant under the evolution, by
$$
{\cal M}_{\sigma,\lambda}=\{\rho_0\ :\ T_{\sigma,\lambda}(t)\rho_0= \rho_0 \mbox{\quad $\forall t\geq 0$} \}.
$$
{}For $\sigma=0$ one can find the dynamics of the reduced density matrix exactly \cite{PSE,MSB,Privman} (see \fer{m31}). The manifold ${\cal M}_{0,\lambda}$ is the set of all system density matrices which are {\em diagonal in the eigenbasis of the interaction operator $G$}. Moreover, we show in Appendix A that there is a constant $C$ such that, for all initial system states $\rho_0$ and all times $t\geq 0$,
\begin{equation}
\label{*}
{\rm dist}\big( {\cal M}_{0,\lambda}, T_{0,\lambda}(t)\rho_0 \big) \leq C {\rm e}^{-\lambda^2\gamma_G\Gamma(t)} {\rm dist}\big( {\cal M}_{0,\lambda}, \rho_0 \big).
\end{equation}
The distance ${\rm dist}( {\cal M}_{0,\lambda}, \rho)=\inf\{\|\tau-\rho\|_1\ :\ \tau\in{\cal M}_{0,\lambda}\}$ is measured in trace norm, $\|x\|_1={\rm Tr}\sqrt{xx^*}$ for linear operators $x$ on ${\mathbb C}^N$. Here, $\Gamma(t)\geq 0$ is the {\em decoherence function} (see \fer{m31.1}) and $\gamma_G=\min\{(g_a-g_b)^2 : a\neq b\}$, where $\{g_a\}_{a=1}^N$ is the spectrum of $G$. Relation \fer{*} shows that
the manifold ${\cal M}_{0,\lambda}$ is {\em orbitally stable}, meaning that a state initially close to ${\cal M}_{0,\lambda}$ remains so for all times. If $\gamma_G>0$ and $\Gamma(t)\rightarrow\infty$ as $t\rightarrow\infty$, then the system undergoes full decoherence in the eigenbasis of $G$ (off-diagonal density matrix elements converge to zero as $t\rightarrow\infty$). In this case, \fer{*} shows that the manifold ${\cal M}_{0,\lambda}$ is dynamically attractive, or {\em asymptotically stable}. One shows that for suitable infra-red behaviour of the interaction form factor $g(k)$, the decoherence function satisfies $\lim_{t\rightarrow\infty}\Gamma(t)/t=\Gamma_\infty$, with $\Gamma_\infty>0$. The manifold ${\cal M}_{0,\lambda}$ is then approached exponentially quickly, at the rate $\lambda^2\gamma_G\Gamma_\infty$. We give further detail in Appendix \ref{appb}.
As the degeneracy is lifted, for small $\sigma>0$, the manifold of invariant initial system states becomes empty, ${\cal M}_{\sigma,\lambda}=\emptyset$. All initial states approach a single asymptotic state, which is the reduction to the small system of the joint system-reservoir equilibrium state (which is not a product state, see Appendix \ref{appb}). In the regime $\sigma<\!\!<|\lambda|<\!\!<1$, the approach of the asymptotic state, and hence the dissolution of the manifold ${\cal M}_{0,\lambda}$, takes place at a rate proportional to $\sigma^2/\lambda^2$, as we now show.
The density matrix elements of the small system are given by
\begin{equation}
{}[\rho_t]_{a,b}\equiv\scalprod{\varphi_a}{\rho_t\,\varphi_b}, \qquad a,b=1,\ldots,N.
\label{dmatel}
\end{equation}
\begin{thm}[Reduced dynamics]
\label{thmreddynfinal}
Assume A1-A4. There is a $\lambda_0>0$ such that for fixed $\lambda$ satisfying $0<|\lambda|<\lambda_0$, the following holds. There is a $\sigma_0>0$ (depending on $\lambda$) s.t. if $0\leq\sigma<\sigma_0$, then we have, uniformly in $t\geq 0$:
-- For $a,b=1,\ldots,N$, $a\neq b$,
\begin{equation}
{}[\rho_t]_{a,b} = {\rm e}^{\i t\varepsilon_{b,a}(\sigma,\lambda)}[\rho_0]_{a,b} +O_\lambda(\sigma) +O(\lambda).
\label{l5}
\end{equation}
-- For $a=1,\ldots,N$,
\begin{equation}
{}[\rho_t]_{a,a} = \frac 1N+\sum_{b=2}^N D_{a,b}(t)[\rho_0]_{b,b} +O_\lambda(\sigma) +O(\lambda).
\label{l9}
\end{equation}
Let $\{\varphi^T_a\}_{a=1}^N$ be an orthonormal basis of eigenvectors of $T$, with $T\varphi_a^T=\xi_a\varphi_a^T$ and denote by $[\varphi^T_a]_c$, $c=1,\ldots,N$, the components of $\varphi^T_a$ (in the canonical basis). Then
$$
D_{a,b}(t) = \sum_{c=2}^N {\rm e}^{\i t\varepsilon_{c,c}(\sigma,\lambda)} \ \overline{[\varphi^T_c]_b} \ [\varphi^T_c]_{a}.
$$
\end{thm}
{\bf Discussion.\ } 1. The resonance energies governing the dynamics of {\em off-diagonals} are of the form (see \fer{l7})
$$
\varepsilon_{a,b}(\sigma,\lambda) = \lambda^2\delta_{a,b} +\sigma r_{a,b}+ \frac{\sigma^2}{\lambda^2}\,z_{a,b}+ O\left(\frac{\sigma^2}{\lambda}\right)+ O_\lambda(\sigma^3).
$$
We have the following interpretation:
$\bullet$ $\lambda^2\delta_{a,b}$ is a resonance energy for $\sigma=0$. The imaginary part of $\delta_{a,b}$ is proportional to $(g_a-g_b)^2$. All off-diagonal density matrix elements tend to zero (modulo an error term) as $t\rightarrow\infty$ if $g_a\neq g_b$ for $a\neq b$ and infra-red behaviour $p=-1/2$. The system exhibits then {\em decoherence in the eigenbasis of $G$}, regardless of whether the system energy is degenerate or not. The contribution to the decoherence rate of this term is proportional to $\lambda^2$.
$\bullet$ The term linear in $\sigma$ is real, with $r_{a,b}=[H_{\rm S}]_{a,a}-[H_{\rm S}]_{b,b}$. The decay rates of matrix elements do not depend on the first order in the energy splitting parameter $\sigma$.
$\bullet$ The second order term in $\sigma$ has generally non-vanishing real and imaginary parts. The complex $z_{a,b}$ are determined by the ratio of matrix elements $[H_{\rm S}]_{c ,d}$ and differences of $\delta_{c,d}$ (see \fer{pertevnon-zero}). The factor $1/\lambda^2$ is due to the presence of the reduced resolvent in second order perturbation theory in $\sigma$ (here, the `non-degenerate energies' are $\lambda^2\delta_{a,b}$). The sign of ${\rm Im}\, z_{a,b}$ can be positive or negative, depending on the model.
2. The resonance energies driving the dynamics of the {\em diagonal} density matrix elements have the form
$$
\varepsilon_{c,c} (\sigma,\lambda)= 2\i\frac{\sigma^2}{\lambda^2} \xi_c + O\left(\frac{\sigma^2}{\lambda}\right)+ O_\lambda(\sigma^3).
$$
The $\xi_c$, $c=2,\ldots,N$, are strictly positive if, for instance, $[H_{\rm S}]_{a,b}{\rm Im}\delta_{a,b}\neq 0$ for all $a,b$ with $a\neq b$ (see after \fer{mmm5}). Then $D_{a,b}(t)$ decays exponentially quickly in time. Contrary to the off-diagonals, the diagonal entries of the density matrix evolve as a group: the value of a given diagonal entry depends on the initial condition of all of them. While the convergence rate of off-diagonals is proportional to $\lambda^2$, that of the diagonal is proportional to $\sigma^2/\lambda^2$. Hence the convergence of the diagonal, the part of the density matrix in the manifold ${\cal M}_{0,\lambda}$, is driven by the level splitting, while that of the off-diagonals is driven by the system-reservoir interaction.
\medskip
{\bf Transition between regimes for the spin-boson model. } We consider the small system to be a spin with Hamiltonian and interaction operator given by
\begin{equation*}
H_{\rm S}=S^z\equiv\frac{1}{2}\left(
\begin{array}{cc}
1 & 0 \\
0 & -1
\end{array}
\right) \quad \mbox{and}\quad G=S^x\equiv\frac{1}{2}\left(
\begin{array}{cc}
0 & 1 \\
1 & 0
\end{array}
\right),
\end{equation*}
respectively. The parameters $\sigma,\lambda$ are now considered to be small but independent of each other. We analyze the decoherence properties of the spin {\em in the energy basis}. Let $\phi^z_\pm$ be the normalized energy eigenvectors, satisfying $H_{\rm S}\phi^z_\pm=\pm\frac12\phi_\pm^z$, and denote the spin density matrix elements in this basis by $[\rho_t]^z_{+,-}:=\scalprod{\phi^z_+}{\rho_t\phi^z_-}$ (and similarly for other matrix elements). We show in Section \ref{sectspinboson} that
\begin{eqnarray*}
{}[\rho_t]^z_{+,+}&\doteq&\textstyle\frac{1}{2}+\frac{1}{2}{\rm e}^{\i t
w_2}([\rho_0]^z_{+,+}-[\rho_0]^z_{-,-}),\\
{}[\rho_t]^z_{+,-}&\doteq& \textstyle\frac{r}{r^2+1}\left((1+r){\rm e}^{\i tw_3}+(1/r-1){\rm e}^{\i
tw_4}\right) [\rho_0]^z_{+,-},
\end{eqnarray*}
where $\doteq$ means that terms of order $O(\lambda^2)$ are disregarded (see \fer{u1}). It is assumed here that $[\rho_0]^z_{+,-}\in\mathbb R$ (see \fer{1.106} for the general expression) and we have set
$$
r= \frac{-4\i \gamma -\sqrt{\pi^2\xi(0)^2 -16\gamma^2}}{\pi\xi(0)}\quad \mbox{with}\quad \gamma=\frac{\sigma}{\lambda^2}.
$$
Here, the square root is the principal branch with branch cut on the negative real axis and $\xi(0)>0$ is a constant proportional to the reservoir spectral density at zero (see \fer{xinot}). The system has four resonance energies, one is zero and the other three are
$$
w_2=\i \frac{\lambda^2}{2}\pi \xi(0), \quad w_{3,4}=\i\frac{\lambda^2}{4}\pi\xi(0)\pm\i \sqrt{\frac{\lambda^4}{16}\pi^2\xi(0)^2-\sigma^2}.
$$
These expressions interpolate the values of the previously known, isolated regime (lowest order in $\lambda$ for $\sigma$ fixed) and the overlapping resonances values derived here ($\sigma$ small, $\lambda$ fixed; see also the remark after Theorem \ref{thm4}).
The diagonal converges to $\frac12$ at the rate ${\rm Im}w_2\propto\lambda^2$, independently of $\sigma$. The decoherence rate (decay of the off-diagonal in the energy basis) is obtained as follows.
\begin{itemize}
\item[-] {\em Overlapping resonances regime}: $\gamma<\!\!<1$ and $r\approx -1$. Thus, $[\rho_t]^z_{+,-}\approx {\rm e}^{\i tw_4}[\rho_0]^z_{+,-}$, which has decay rate ${\rm Im}w_4\approx \frac{2}{\pi\xi(0)}\frac{\sigma^2}{\lambda^2}$.
\item[-] {\em Isolated resonances regime}: $1/\gamma<\!\!<1$ and $r\approx-\i\infty$. Thus, $[\rho_t]^z_{+,-}\approx {\rm e}^{\i tw_3}[\rho_0]^z_{+,-}$, which has decay rate ${\rm Im}w_3\approx \frac{\pi\xi(0)}{4}\lambda^2$.
\end{itemize}
In the isolated resonances regime, the decoherence rate is given by the system-reservoir coupling constant $\lambda$ alone, while in the overlapping case, it depends also on the level splitting parameter $\sigma$. For a fixed $\lambda$, the decoherence rate increases quadratically in $\sigma$ (for small $\sigma$). The further its energy levels lie apart, the quicker the spin decoheres.
We define the critical value $\gamma_*$ for which the square root in $w_{3,4}$ vanishes,
$$
\gamma_* := \textstyle\frac14 \pi\xi(0).
$$
This critical value separates two regimes with different qualitative behaviour of the resonances $w_3$ and $w_4$. As $\gamma$ increases from zero to $\gamma_*$, the resonance $w_3$ moves down the imaginary axis, decreasing from the initial value $\frac12\i\pi\xi(0)\lambda^2$ to $\frac14\i\pi\xi(0)\lambda^2$, while $w_4$ moves up the imaginary axis, from the origin to $\frac14\i\pi\xi(0)\lambda^2$. The two resonances meet for $\gamma=\gamma_*$. As $\gamma>\gamma_*$ increases further, the resonances $w_3$ and $w_4$ move horizontally away from the imaginary axis, their imaginary parts stay constant, equal to $\frac14\pi\xi(0)\lambda^2$. This motivates the sharp definition of the overlapping resonances regime, in the spin-boson model, to be given by $\gamma<\gamma_*$ and of the isolated resonances regime to be given by $\gamma>\gamma_*$.
It is interesting to note that in nuclear physics, there is a (to our knowledge not rigorously defined) notion of overlapping resonances, used in the description of processes involving unstable nuclei by non-hermitian Hamiltonians \cite{SZ,CIZB}. It is observed that in the overlapping regime, the resonance widths (imaginary parts of resonance energies) segregate into two clusters, one located close to the origin (slow channels), the other at a much larger value (fast channels). The same occurs in our system: in the overlapping regime, we have one resonance at zero and another one, $w_4$, close to it. The other two, $w_2$ and $w_3$, are much larger, both close to $\frac12\i\pi\xi(0)\lambda^2$. As the system transitions into the isolated resonances regime, the two clusters mix.
\section{Resonances and dynamics}
\subsection{Resolvent representation}
The main result of this section is Theorem \ref{thm2}. For $\theta\in{\mathbb R}$ let $U_\theta$ be the unitary (translation) on ${\cal H}_\r$ defined by sector-wise action $U_\theta \Omega_\r=\Omega_\r$ and $U_\theta\psi_n(u_1,\vartheta_1,\cdots,u_n,\vartheta_n)=\psi_n(u_1+\theta,\vartheta_1,\cdots,u_n+\theta,\vartheta_n)$. A vector $\psi\in{\cal H}_\r$ is called $U_\theta$-analytic if the map $\theta\mapsto U_\theta\psi$ is ${\cal H}_\r$-valued analytic in $\{\theta\in{\mathbb C} :\ |\theta|<\theta_0\}$ (the $\theta_0$ is that of assumption A1). All vectors of the form $\psi\otimes\Omega_\r$, for arbitrary $\psi\in{\cal H}_{\rm S}$, are $U_\theta$-analytic. We introduce the {\em reference state}
\begin{equation}
\Omega=\Omega_{\rm S}\otimes\Omega_\r,
\label{m20}
\end{equation}
where $\Omega_\r$ is the vacuum in ${\cal H}_\r$ and $\Omega_{\rm S}$ is the trace state
\begin{equation}
\Omega_{\rm S} = \frac{1}{\sqrt{N}}\sum_{a=1}^N\varphi_a\otimes\varphi_a.
\label{m-1}
\end{equation}
$\Omega$ is cyclic and separating for ${\frak M}$ and we denote the associated modular operator and modular conjugation by $\Delta$ and $J$, respectively \cite{BR}. We have $\Delta=\Delta_{\rm S}\otimes\Delta_\r$, where $\Delta_\r={\rm e}^{-\beta L_\r}$ and $\Delta_{\rm S}=\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l$ (the trace state is KMS with inverse temperature $\beta=0$). The modular conjugation is $J=J_{\rm S}\otimes J_\r$. We have $J_{\rm S}\phi\otimes\chi = \bar\chi\otimes\bar\phi$ for $\phi,\chi\in{\mathbb C}^N$, and where the bar means complex conjugation of vector components in the basis $\{\varphi_a\}_{a=1}^N$. Furthermore, $J_\r\psi_n(u_1,\vartheta_1,\ldots,u_n,\vartheta_n) = \overline{\psi_n(-u_1,\vartheta_1,\ldots,-u_n,\vartheta_n)}$. A suitable generator of the dynamics is constructed as follows, see \cite{JP1} and also \cite{MMS}. On the dense set ${\frak M}\,\Omega$ we define the group ${\cal U}(t)$ by
\begin{equation}
{\cal U}(t) A\Omega = {\rm e}^{\i tL(\sigma,\lambda)} A{\rm e}^{-\i tL(\sigma,\lambda)}\Omega,\qquad A\in{\frak M}, \ t\in{\mathbb R},
\label{mm1}
\end{equation}
where $L(\sigma,\lambda)$ is the Liouvillian \fer{m11'}. We introduce the linear space
\begin{equation}
{\cal D}_0 = {\cal D}(L_\r)\cap{\cal D}(N^{1/2})\cap{\frak M}\, \Omega \subset{\cal H},
\label{m1}
\end{equation}
where $N=\d\Gamma(\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l)$ is the number operator.
\begin{prop}
\label{propone}
{\bf (a)} ${\cal U}(t)$ is strongly differentiable on ${\cal D}_0$ and its
generator is given by
\begin{equation}
\i\frac{\d}{\d t}|_{t=0} \ {\cal U}(t) = K(\sigma,\lambda) := L_0(\sigma)+\lambda V -\lambda J\Delta^{1/2} V J\Delta^{1/2}.
\label{mm0}
\end{equation}
{\bf (b)} $\theta\mapsto U_\theta K(\sigma,\lambda) U_\theta^*$ has an analytic continuation from $\theta\in{\mathbb R}$ to $\{\theta\in{\mathbb C}\ :\ |\theta|<\theta_0\}$,
in the strong sense on ${\cal D}_0$. This continuation is given by
\begin{equation}
K_\theta(\sigma,\lambda) = L_{0,\theta}(\sigma)+\lambda I_\theta,
\label{b1}
\end{equation}
where
\begin{eqnarray}
L_{0,\theta}(\sigma) &=& L_0(\sigma)+\theta N\\
I_\theta&=& V_\theta -V'_\theta\\
V_\theta&=&\frac{1}{\sqrt 2}G\otimes
\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l\otimes\Big(a^*\big(g_\beta(\cdot+\theta)\big)+a\big(g_\beta(\cdot+\bar\theta)\big)\Big)\\
V'_\theta&=&\frac{1}{\sqrt 2}\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l\otimes
G\otimes\Big(a^*\big(e^{\frac{\beta}{2}(\cdot+\theta)}\overline{g}_\beta(-\cdot-\bar\theta)\big)+a\big(e^{-\frac{\beta}{2}(\cdot+\bar\theta)}
\overline{g}_\beta(-\cdot-\theta)\big)\Big)\qquad \label{b2}
\end{eqnarray}
(Here, we use the convention $\overline g(u)=\overline{g(u)}$.)
\end{prop}
\bigskip
\noindent
{\bf Proof.\ } We do not write the dependence of operators on $(\sigma,\lambda)$ in this proof, which follows \cite{JP1} (see also \cite{MMS}).
{\bf (a)} Let $A\Omega\in{\cal D}_0$. Then
\begin{equation}
\frac{\d}{\d t}|_{t=0} \ {\cal U}(t) A\Omega =-\i AL\Omega+\i LA\Omega =-\i A(L_0+\lambda V)\Omega+\i (L_0+\lambda V)A\Omega.
\label{m24}
\end{equation}
Since $L_0\Omega=0$ and $AV\Omega= J\Delta^{1/2}V^*A^*\Omega= J\Delta^{1/2}VJ\Delta^{1/2}A\Omega$, the right side of \fer{m24} equals $\i L_0A\Omega+ \i\lambda (V-J\Delta^{1/2} VJ\Delta^{1/2})A\Omega$. This shows part (a).
\smallskip
{\bf (b)} For real $\theta$, we have
\begin{equation*}
\begin{split}
U_\theta K(\lambda)U_\theta^*=&L_0+\theta N+\frac{\lambda}{\sqrt
2}G\otimes
\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l\otimes\Big(a^*(g_\beta(\cdot+\theta))+a(g_\beta(\cdot+\theta))\Big)\\
-&\frac{\lambda}{\sqrt 2}\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l\otimes
G\otimes\Big(a^*(e^{\frac{\beta}{2}(\cdot+\theta)}\overline{g}_\beta(-\cdot-\theta))+a(e^{-\frac{\beta}{2}(\cdot+\theta)}\overline{g}_\beta(-\cdot-\theta))\Big).
\end{split}
\end{equation*}
By assumption (A) we obtain the analytic extension \fer{b1}-\fer{b2}. Note that in the argument of the annihilation operators, the analytic extension has the complex conjugate $\bar\theta$, since the annihilation operators are anti-linear in their argument. {\hfill $\square$}
\begin{thm}
\label{thm2}
Assume A1 and A2. Let $\theta$ with $0<{\rm Im} \theta<\theta_0$ be fixed. There is a $\lambda_0>0$ such that for all $|\lambda|<\lambda_0$ and all $\sigma\in\mathbb R$, we have the following. Let $\phi\in {\cal H}$ and $A\in{\frak M}$ be such that $\phi$ and $A\Omega$ are $U_\theta$-analytic vectors, and such that $\phi_{\bar\theta}\in{\cal D}(|L_\r|^{\frac14+\eta})$, for some $\eta>0$. Then we have for all $t\geq 0$
\begin{equation}
\scalprod{\phi}{{\rm e}^{\i tL(\sigma,\lambda)}A{\rm e}^{-\i tL(\sigma,\lambda)}\Omega} =
\frac{-1}{2\pi\i}\int_{{\mathbb R}-\i} {\rm e}^{\i
tz}\scalprod{\phi_{\overline\theta}}{(K_\theta(\sigma,\lambda)-z)^{-1}
(A\Omega)_\theta}\d z.
\label{mm2}
\end{equation}
\end{thm}
We give a proof of this result in Appendix \ref{appa}.
{\bf Remarks.\ } 1. Vectors representing product states of an arbitrary small system state and the equilibrium reservoir states are of the form $\phi=B\Omega$, where $B\in{\frak M}_{\rm S}$ (and, recall, $\Omega$ is given in \fer{m20}). The proof of \fer{mm2} for such $\phi$ and $A\in{\frak M}_{\rm S}$ is easier than that of the full result. This is the situation of \cite{MSB}.
2. In \cite{MMS} a spectral dilation deformation is performed simultaneously with the translation (see also \cite{BFS,MMS1}). In this doubly-deformed situation, the analogue of Theorem \ref{thm2} is proven in Section 8 of \cite{MMS}. The dilation deforms the spectrum of $K$ in a `sectorial way' (a $V$-shape), leading to useful decay estimates of the (deformed) resolvent $(K-z)^{-n}$, as $|{\rm Re}z|\rightarrow\infty$. However, in the present work, we only use spectral translation and such decay estimates do not hold (as the distance between the spectrum of $K_\theta$ and the real axis does not grow now when $|{\rm Re} z|\rightarrow\infty$). We therefore need a new proof of this result. The advantage of only performing the translation deformation is that less restrictive conditions on the form factor are needed only.
\subsection{Resonances of $K(\sigma=0,\lambda)$}
The operator $K_\theta(0,\lambda)$ is defined in Proposition \ref{propone}, with $L_0=L_\r$. Recall that $\varphi_a$, $a=1,\ldots, N$, is the orthonormal eigenbasis of $G$, \fer{m22}. The operator $K_\theta(0,\lambda)$ is reduced by the decomposition
$$
{\cal H} = \bigoplus_{a,b=1}^N {\, \rm Ran\,}\Big(|\varphi_a\rangle\langle \varphi_a|\otimes|\varphi_b\rangle\langle \varphi_b|\Big)\otimes{\cal H}_\r.
$$
Namely,
\begin{equation}
K_\theta(0,\lambda) = \bigoplus_{a,b=1}^N K_{a,b},
\label{kab}
\end{equation}
where $K_{a,b}$ acts on ${\cal H}_\r$ as
\begin{equation}
K_{a,b}=L_\r+\theta N+\lambda(g_a\Phi_\theta-g_b\widetilde\Phi_\theta),
\label{01.33}
\end{equation}
with
\begin{equation}
\begin{split}
\Phi_\theta=&\frac{1}{\sqrt 2}\big( a^*(g_\beta(\cdot+\theta))+a(g_\beta(\cdot+\bar\theta))\big)\\
\widetilde \Phi_\theta=&\frac{1}{\sqrt
2}\big(a^*({\rm e}^{\frac{\beta}{2}(\cdot+\theta)}\overline{g}_\beta(-\cdot-\theta))+a({\rm e}^{-\frac{\beta}{2}(\cdot+\bar\theta)}\overline{g}_\beta(-\cdot-\bar\theta))\big).
\end{split}
\end{equation}
To alleviate the notation, we do not display $\theta$ and $\lambda$ in $K_{a,b}$.
\begin{thm}[Spectrum of $K_{a,b}$]
\label{thm3}
Assume A1 and A2. Let $\theta$ with $0<{\rm Im \theta}<\theta_0$ be fixed.
There is a $\lambda_0>0$ such that if $0\leq |\lambda|<\lambda_0$, then for all $a,b=1,\ldots,N$, the operator $K_{a,b}$ has a simple eigenvalue $\lambda^2\delta_{a,b}$, where $\delta_{a,b}$ is given in \fer{delta}. All other spectrum of $K_{a,b}$ lies in $\{z\in{\mathbb C}\ :\ {\rm Im}z>\frac34{\rm Im}\theta\}$.
\end{thm}
{\bf Remarks.\ } 1. It follows from Theorem \ref{thm3} and the decomposition \fer{kab} that the spectrum of $K_\theta(0,\lambda)$ in the strip $\{z\in{\mathbb C}\ :\ {\rm Im}z<\frac34{\rm Im}\theta\}$ consists precisely of the eigenvalues $\{\lambda^2\delta_{a,b}\}_{a,b=1}^N$ (there are no higher order terms in $\lambda$). A simple expression for the eigenvectors associated to the non-zero eigenvalues is not available, only a perturbation series is. However, it is readily seen that the eigenvalue zero has the eigenvectors $\varphi_a\otimes\varphi_a\otimes\Omega_\r$, $a=1,\ldots,N$. Indeed, if $a=b$, then it follows directly from \fer{01.33} that
\begin{equation}
K_{a,a}\Omega_\r =\lambda g_a U_\theta (\Phi-J\Delta^{1/2} \Phi J\Delta^{1/2})\Omega_\r =0,
\label{remark}
\end{equation}
since $J\Delta^{1/2}\Phi J\Delta^{1/2}\Omega_\r=\Phi\Omega_\r$.
2. If the form factor $g$ satisfies $\|g_\beta/u\|^2_2<\infty$, then the operator $K_{a,b}$, \fer{01.33}, is unitarily equivalent to the operator $L_\r+{\rm const.}$ The condition on the form factor implies the infra-red behaviour $g(k)\sim|k|^p$ for small $k$, with $p>-1/2$. Then $K_{a,b}$ has a simple real eigenvalue, as also predicted by \fer{delta}, saying that ${\rm Im}\delta_{a,b}=0$. In the infra-red singular case, $p=-1/2$, the unitary transformation ceases to exist and the eigenvalue becomes complex.
\medskip
\noindent
{\bf Proof of Theorem \ref{thm3}.\ } The spectrum of $K_{a,b}$ for $\lambda=0$ consists of a single simple eigenvalue at zero, with eigenvector $\Omega_\r$, and of horizontal lines of continuous spectrum $\{ x+{\rm Im} \theta \,n\ :\ x\in{\mathbb R}, n=1,2,\ldots\}$. The operators $\Phi_\theta$ and $\widetilde\Phi_\theta$ are infinitesimally small w.r.t. $N$ (relatively bounded with arbitrarily small relative bound). Analytic perturbation theory implies that there exists a $\lambda_0>0$ such that if $0\leq |\lambda|<\lambda_0$, then the only spectrum of $K_{a,b}$ in $\{ z\in{\mathbb C}\ :\ {\rm Im}z<{\rm Im}\theta/2\}$ is a single, simple eigenvalue. We show that this eigenvalue is $\lambda^2\delta_{a,b}$, with $\delta_{a,b}$ given in \fer{delta}.
The dynamics of the {\em reduced density matrix} of the small system has been calculated explicitly in Proposition 7.4 of \cite{MSB}. Let $\psi_0=B\Omega_{\rm S}\otimes \Omega_\r$ be an initial state, where $B\in{\frak M}_{\rm S}'$ (the commutant) is arbitrary (see also \fer{m20}). The reduced system density matrix at time $t$, in the basis $\{\varphi_a\}$, is given by $[\rho_t]_{a,b}= \langle \psi_0, {\rm e}^{\i tL(0,\lambda)}(|\varphi_b\rangle \langle\varphi_a|\otimes \mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{\rm S}) {\rm e}^{-\i tL(0,\lambda)}\psi_0\rangle$.
It is shown in the above reference that
\begin{equation}
[\rho_t]_{a,b}=[\rho_0]_{a,b}\ {\rm e}^{\i\lambda^2\alpha_{a,b}(t)},
\label{m31}
\end{equation}
with
$\alpha_{a,b}(t)=(g_a^2-g_b^2)S(t)+\i(g_a-g_b)^2\Gamma(t)$, where
\begin{equation}
\Gamma(t)=\textstyle\int_{{\mathbb R}^3}|g(k)|^2\coth
(\frac{\beta|k|}{2})\frac{\sin ^2(\frac{|k|
t}{2})}{|k|^2}\d^3k,\quad
S(t)=\frac{1}{2}\int_{{\mathbb R}^3}|g(k)|^2\frac{|k| t-\sin
|k| t}{|k|^2}\d^3k.
\label{m31.1}
\end{equation}
For large times, $\alpha_{a,b}(t)$ becomes linear,
\begin{equation}
\lim_{t\rightarrow\infty}\frac{\alpha_{a,b}(t)}{t}=\delta_{a,b},
\label{m32}
\end{equation}
with $\delta_{a,b}$ given in \fer{delta}. We express the reduced density matrix alternatively, using Theorem \ref{thm2}, as
\begin{equation}
[\rho_t]_{a,b}=\frac{-1}{2\pi \i}\int_{ \mathbb{R}-\i}{\rm e}^{\i tz}\scalprod{
B^*B\Omega_{\rm S}\otimes\Omega_\r}{(K_\theta-z)^{-1}\big(|\varphi_b\rangle
\langle\varphi_a|\otimes \mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{\rm S}\big) \Omega_{\rm S}\otimes\Omega_\r} \d z.
\label{m33}
\end{equation}
We use that ${\rm e}^{\i tL(0,\lambda)}(|\varphi_b\rangle \langle\varphi_a|\otimes \mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{\rm S}) {\rm e}^{-\i tL(0,\lambda)}B = B{\rm e}^{\i tL(0,\lambda)}(|\varphi_b\rangle \langle\varphi_a|\otimes \mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{\rm S}) {\rm e}^{-\i tL(0,\lambda)}$, which holds since $B\otimes\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_\r$ belongs to the commutant ${\frak M}'$. It follows from the definition \fer{m-1} that $(|\varphi_b\rangle \langle\varphi_a|\otimes \mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{\rm S})\Omega_S=\frac{1}{\sqrt{N}}\varphi_b\otimes \varphi_a$. Therefore, we obtain from \fer{m33} that
\begin{eqnarray}
[\rho_t]_{a,b} &=&\frac{1}{\sqrt N}\scalprod{B^*B\Omega_{\rm S}}{\varphi_b\otimes \varphi_a}\ \frac{-1}{2\pi \i}\int_{\mathbb R-\i}{\rm e}^{\i tz}\scalprod{\Omega_\r}{(K_{b,a}-z)^{-1}\Omega_\r} \d z\nonumber\\
&=&[\rho_0]_{a,b}\ \frac{-1}{2\pi \i}\int_{{\mathbb R}-\i}{\rm e}^{\i tz}\scalprod{\Omega_\r}{(K_{b,a}-z)^{-1}\Omega_\r}\d z.
\label{01.16}
\end{eqnarray}
Comparing \fer{01.16} and \fer{m31} yields the identity
\begin{equation}
{\rm e}^{\i\lambda^2\alpha_{a,b}(t)}=\frac{-1}{2\pi \i}\int_{\mathbb
R-\i}{\rm e}^{\i tz}\scalprod{\Omega_\r}{(K_{b,a}-z)^{-1}\Omega_\r}\d z.
\label{01.39}
\end{equation}
Denote the unique eigenvalue of $K_{a,b}$ in $\{z\in{\mathbb C}\ :\ {\rm Im} z< {\rm Im}\theta/2\}$ by $\zeta_{a,b}(\lambda)$ and let ${\mathcal C}_{a,b}$ be a small circle around $\zeta_{a,b}(\lambda)$ not including any other point of the spectrum of $K_{a,b}$. By deforming the contour of integration, we have
\begin{equation}
\frac{-1}{2\pi \i}\int_{\mathbb R-\i}{\rm e}^{\i tz}\scalprod{\Omega_\r}{(K_{a,b}-z)^{-1}\Omega_\r}\d z = \frac{-1}{2\pi
\i}\oint_{{\mathcal C}_{a,b}}{\rm e}^{\i tz}\scalprod{\Omega_\r}{(K_{a,b}-z)^{-1}\Omega_\r}\d z +R_\lambda(t),
\label{01.40}
\end{equation}
with a remainder term small in $\lambda$ and decaying to zero exponentially quickly as $t\rightarrow\infty$. This follows from the following result, proven in \cite{MSB}, Proposition 4.2:
\begin{prop}[\cite{MSB}]
\label{prop7}
Let $\psi_0\in{\cal H}_{\rm S}$. Then
$$
\left|\int_{{\mathbb R}+\i\frac34{\rm Im}\theta}{\rm e}^{\i t z}\scalprod{\psi_0\otimes\Omega_\r}{(K_\theta(\sigma,\lambda)-z)^{-1}\psi_0\otimes\Omega_\r}\d z\right|\leq C \lambda^2 {\rm e}^{-\frac34t\,{\rm Im\theta}},
$$
uniformly in $\sigma$ varying in compact sets. The same bound holds if $K_\theta(\sigma,\lambda)$ is replaced by $K_{a,b}$.
\end{prop}
This result implies that $|R_\lambda(t)|\leq C\lambda^2{\rm e}^{-\frac{3{\rm Im \theta}}{4}t}$ for some constant $C$.
Since $\zeta_{a,b}(\lambda)$ is a simple pole of the resolvent $(K_{a,b}-z)^{-1}$ we can replace ${\rm e}^{\i tz}$ by ${\rm e}^{\i t \zeta_{a,b}(\lambda)}$ in \fer{01.40} and we obtain
\begin{equation}
\frac{-1}{2\pi \i}\int_{{\mathbb R}-\i}{\rm e}^{\i tz}\scalprod{\Omega_\r}{(K_{a,b}-z)^{-1}\Omega_\r}\d z = {\rm e}^{\i t\zeta_{a,b}(\lambda)} c_{a,b}(\lambda)+R_\lambda(t),
\label{m35}
\end{equation}
where $c_{a,b}(\lambda)=\frac{-1}{2\pi
\i}\oint_{{\mathcal{C}}_{a,b}}\scalprod{\Omega_\r}{(K_{a,b}-z)^{-1}\Omega_\r}\d z$. Combining \fer{01.39} and \fer{m35} gives
$$
{\rm e}^{\i\lambda^2\alpha_{a,b}(t)-\i t\zeta_{a,b}(\lambda)} = c_{a,b}(\lambda)+{\rm e}^{-\i t\zeta_{a,b}(\lambda)} R_\lambda(t).
$$
As ${\rm Im}\zeta_{a,b}(\lambda)<\frac12{\rm Im}\theta$, we have $\lim_{t\rightarrow\infty}{\rm e}^{-\i t\zeta_{a,b}(\lambda)} R_\lambda(t)=0$. Thus the exponent on the left hand side converges to a finite number, as $t\rightarrow\infty$, and so this exponent, divided by $t$, tends to zero as $t\rightarrow\infty$. (Note that $c_{a,b}(\lambda)$ is not zero for small $\lambda$, by perturbation theory.) Then, due to \fer{m32}, we have $\zeta_{a,b}(\lambda)=\lambda^2\delta_{a,b}$. The proof of Theorem \ref{thm3} is complete. \hfill $\square$
\subsection{Resonances of $K(\sigma,\lambda)$}
We now examine the operator $K_\theta(\sigma,\lambda)$, defined in Proposition \ref{propone}, \fer{b1}-\fer{b2}, with $L_0$ given in \fer{m12}. We consider $K_\theta(\sigma,\lambda)$ as an unperturbed part, $K_\theta(0,\lambda)$, plus a perturbation $\sigma L_{\rm S}$ (see \fer{m11}). Since the eigenvalues of $K_\theta(0,\lambda)$ are isolated (Theorem \ref{thm3}), we can apply analytic perturbation theory to follow them as the perturbation is switched on ($\sigma\neq 0$).
\begin{thm}[Spectrum of $K_\theta(\sigma,\lambda)$]
\label{thm4}
Assume A1-A3. Let $\lambda$ be fixed, satisfying $0<|\lambda|<\lambda_0$, where $\lambda_0$ is given in Theorem \ref{thm3}. There is a $\sigma_0>0$ (depending on $\lambda$) s.t. if $0\leq \sigma<\sigma_0$, then the spectrum of $K_\theta(\sigma,\lambda)$ in the region $\{z\in{\mathbb C}\ :\ {\rm Im}z<\frac12{\rm Im}\theta\}$ consists of simple eigenvalues $\varepsilon_{a,b}(\sigma,\lambda)$. Those eigenvalues are analytic functions of $\sigma$, given by \fer{l7}. Zero is an eigenvalue of $T$, \fer{mmm6}. It is simple if $[H_{\rm S}]_{a,b}\neq 0$ for all $a\neq b$.
\end{thm}
{\bf Remark. } The theorem assumes the non-degeneracy condition A3. An analysis in presence of degenerate non-zero resonances $\lambda^2\delta_{a,b}$ can be carried out along the same lines. We have done this for the spin-boson model. We have checked that the values for the resonances thus obtained coincide with those obtained in Section \ref{sectspinboson} (to order two in $\sigma$).
\medskip
\noindent
{\bf Proof of Theorem \ref{thm4}.\ } {\bf (A) Non-zero eigenvalues.} The non-zero eigenvalues of $K_\theta(0,\lambda)$ are simple, given by $\varepsilon_{a,b}(0,\lambda)=\lambda^2\delta_{a,b}$, for $a\neq b$. We denote by $\varphi_{a,b}\otimes X_{a,b}$ the eigenvector associated to $\varepsilon_{a,b}(0,\lambda)$, where $\varphi_{a,b}=\varphi_a\otimes\varphi_b$ and $X_{a,b}$ is a normalized vector in ${\cal H}_\r$, depending on $\lambda$ and $\theta$. The adjoint operator satisfies $K_\theta(0,\lambda)^*\varphi_{a,b}\otimes X_{a,b}^*=\lambda^2\overline{\delta}_{a,b}\varphi_{a,b}\otimes X_{a,b}^*$ for a vector $X^*_{a,b}$ satisfying $\scalprod{X_{a,b}}{X^*_{a,b}}=1$. We denote the Riesz projection of $K_\theta(0,\lambda)$ associated to $\varepsilon_{a,b}(0,\lambda)$ by
\begin{equation}
P_{a,b} = |\varphi_{a,b}\otimes X_{a,b}\rangle\langle \varphi_{a,b}\otimes X^*_{a,b}|.
\label{pab}
\end{equation}
By analytic perturbation theory, $K_\theta(\sigma,\lambda)$ has a simple eigenvalue in the vicinity of $\lambda^2\delta_{a,b}$, for small $\sigma$. It is given by
\begin{equation}
\varepsilon_{a,b}(\sigma,\lambda) =\lambda^2\delta_{a,b}+\sigma \varepsilon_{a,b}^{(1)}+\sigma^2 \varepsilon_{a,b}^{(2)} +O_\lambda(\sigma^3),
\end{equation}
where (see \cite[Sect. II.2.2]{Kato} and also \cite[Thm. XII.12]{RS4})
\begin{equation}
\varepsilon_{a,b}^{(1)} = {\rm Tr} (L_{\rm S} P_{a,b}) = [H_{\rm S}]_{a,a}-[H_{\rm S}]_{b,b}.
\label{m50}
\end{equation}
Here, we have set $[H_{\rm S}]_{a,b}=\scalprod{\varphi_a}{H_{\rm S}\varphi_b}$. The second order correction is
\begin{equation}
\varepsilon_{a,b}^{(2)}= -{\rm Tr} \big( L_{\rm S} (K_\theta(0,\lambda)-\lambda^2\delta_{a,b})^{-1}\bar P_{a,b} L_{\rm S} P_{a,b}\big).
\label{m51}
\end{equation}
We write $\bar P$ for $\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l-P$ for general projections $P$. We set $P^{\rm S}_{a,b}=|\varphi_{a,b}\rangle\langle\varphi_{a,b}|$ and $P^\r_{a,b}=|X_{a,b}\rangle\langle X_{a,b}^*|$. Then $P_{a,b}=P^{\rm S}_{a,b}\otimes P^\r_{a,b}$ and $\bar P_{a,b} = \bar P^{\rm S}_{a,b}\otimes\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_\r + P^{\rm S}_{a,b}\otimes\bar P^\r_{a,b}$. It follows that
$\bar P_{a,b} L_{\rm S}\ (\varphi_{a,b}\otimes X_{a,b}) = (\bar P^{\rm S}_{a,b} L_{\rm S}\varphi_{a,b})\otimes X_{a,b}$.
Using this and $\bar P^{\rm S}_{a,b}=\sum_{(c,d)\neq(a,b)}P^{\rm S}_{c,d}$ in expression \fer{m51} yields
\begin{equation*}
\varepsilon_{a,b}^{(2)} =-\sum_{(c,d)\neq (a,b)}\scalprod{\varphi_{a,b}\otimes X_{a,b}^*}{L_{\rm S}\, (K_\theta(0,\lambda)-\lambda^2\delta_{a,b})^{-1} \varphi_{c,d}\otimes X_{a,b}} \scalprod{\varphi_{c,d}}{L_{\rm S}\varphi_{a,b}}.
\end{equation*}
`Replacing' $\varphi_{c,d}\otimes X_{a,b}$ by the eigenvector $\varphi_{c,d}\otimes X_{c,d}$, we obtain
\begin{equation}
\varepsilon_{a,b}^{(2)} =-\sum_{(c,d)\neq (a,b)}\frac{1}{\lambda^2(\delta_{c,d}-\delta_{a,b})} |\scalprod{\varphi_{a,b}}{L_{\rm S} \varphi_{c,d}}|^2 \scalprod{X^*_{a,b}}{X_{c,d}} +\xi,
\label{mmm1}
\end{equation}
where
\begin{equation}
\xi =\sum_{(c,d)\neq (a,b)} \scalprod{\varphi_{a,b}\otimes X^*_{a,b}}{L_{\rm S} (K_\theta(0,\lambda)-\lambda^2\delta_{a,b})^{-1}\varphi_{c,d}\otimes(X_{c,d}-X_{a,b})}\scalprod{\varphi_{c,d}}{L_{\rm S}\varphi_{a,b}}.
\label{m51''}
\end{equation}
By perturbation theory, we have $X_{a,b}=\Omega_\r+O(\lambda)$. Therefore, $X_{c,d}-X_{a,b}=O(\lambda)$ and $\scalprod{X^*_{a,b}}{X_{c,d}}=1+O(\lambda)$. Together with the bound \fer{resbnd} of Corollary \ref{corollary1} below, we obtain
\begin{equation}
|\xi|\leq \frac{C}{|\lambda|}.
\label{m51.1}
\end{equation}
Finally,
\begin{equation}
\scalprod{\varphi_{a,b}}{L_{\rm S}\varphi_{c,d}} = \chi_{b=d}\, [H_{\rm S}]_{a,c}-\chi_{a=c}\, [H_{\rm S}]_{d,b}.
\label{bnd3}
\end{equation}
Relation \fer{l7} for $a\neq b$ follows from \fer{mmm1}, \fer{m51.1} and \fer{bnd3} and a little algebra.
\begin{prop}[Bound on the resolvent]
\label{resbndprop}
There are constants $C$ and $\lambda_0$ (depending on ${\rm Im}\theta$ only) such that if $0<|\lambda|<\lambda_0$, then we have the following. Fix any $\alpha>0$ and take complex $z$ satisfying $|z|<C\alpha$, ${\rm Im}z<\frac14{\rm Im}\theta$, and ${\rm dist}({\cal E},z)\geq\alpha\lambda^2$, where ${\cal E}=\{\lambda^2\delta_{a,b}\, : \, a,b=1,\ldots,N\}$ is the set of eigenvalues of $K_\theta(0,\lambda)$. Then we have
\begin{equation}
\|(K_\theta(0,\lambda)-z)^{-1}\|\leq C_1 \left(\frac{1}{{\rm Im}\theta} +\frac{1}{{\rm dist}({\cal E},z)}\right),
\label{resbnd1}
\end{equation}
where $C_1$ is a constant depending only on ${\rm Im}\theta$.
\end{prop}
Knowing the bound on the resolvent we can obtain a bound on the {\em reduced resolvent}.
\begin{cor}
\label{corollary1}
{}For any $a,b=1,\ldots,N$ we have
\begin{equation}
\|(K_\theta(0,\lambda)-\lambda^2\delta_{a,b})^{-1}\bar P_{a,b}\|\leq C_2\left( \frac{1}{{\rm Im}\theta} + \frac{1}{\lambda^2}\right),
\label{resbnd}
\end{equation}
for some constant $C_2$ depending on ${\rm Im}\theta$.
\end{cor}
\noindent
{\bf Proof of Corollary \ref{corollary1}.\ } The reduced resolvent has the representation
$$
(K_\theta(0,\lambda)-\lambda^2\delta_{a,b})^{-1}\bar P_{a,b} = \frac{-1}{2\pi\i}\oint_{\Gamma_{a,b}(\lambda)} (z-\lambda^2\delta_{a,b})^{-1} (K_\theta(0,\lambda)-z)^{-1}\bar P_{a,b} \d z,
$$
where $\Gamma_{a,b}(\lambda)=\{z=\lambda^2\delta_{a,b}+\lambda^2 r{\rm e}^{\i\phi}\, :\, \phi\in[0,2\pi]\}$, with an appropriate radius $r$ (independent of $\lambda$) such that $\Gamma_{a,b}(\lambda)$ encircles only the eigenvalue $\lambda^2\delta_{a,b}$ and such that $\Gamma_{a,b}(\lambda)$ lies within the region of $z$ for which the bound \fer{resbnd1} holds, according to Proposition \ref{resbndprop}. Then ${\rm dist}({\cal E},z)$ is a constant times $\lambda^2$. It follows that
$$
\|(K_\theta(0,\lambda)-\lambda^2\delta_{a,b})^{-1}\bar P_{a,b}\|\leq C \left(\frac{1}{{\rm Im}\theta} +\frac{1}{\lambda^2}\right)(1+\|P_{a,b}\|),
$$
for some constant $C$. The bound \fer{resbnd} follows from $\|P_{a,b}\|=1+O(\lambda)$. \hfill $\square$
\noindent
{\bf Proof of Proposition \ref{resbndprop}.\ } Let $P_\r=|\Omega_\r\rangle\langle\Omega_\r|$, $\bar P_\r=\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l-P_\r$, and $R(z)=(K_\theta(0,\lambda)-z)^{-1}$.
{\em Step 1.} For any $\psi\in{\cal H}$ we have
\begin{eqnarray}
\lefteqn{
\left|\scalprod{\psi}{\bar P_\r (K_\theta(0,\lambda)-z)\bar P_\r\psi}\right|
\geq{\rm Im}\scalprod{\psi}{\bar P_\r (K_\theta(0,\lambda)-z)\bar P_\r\psi}}\nonumber\\
&=& \scalprod{\psi}{\bar P_\r\big(N^{1/2}\{ {\rm Im}\theta+ \lambda{\rm Im}N^{-1/2}I_\theta N^{-1/2}\}N^{1/2} -{\rm Im}z \big)\bar P_\r\psi}\nonumber\\
&\geq&({\rm Im}\theta -C |\lambda| - {\rm Im}z)\|\bar P_\r\psi\|^2\nonumber\\
&\geq&\textstyle\frac12{\rm Im}\theta\,\|\bar P_\r\psi\|^2.\nonumber
\end{eqnarray}
By the Cauchy-Schwartz inequality, it follows that $\|\bar P_\r(K_\theta(0,\lambda)-z)\bar P_\r\psi\|\geq \frac 12{\rm Im}\theta\,\|\bar P_\r\psi\|$ and therefore
\begin{equation}
\|\bar P_\r R(z) \bar P_\r\|\leq \frac{2}{{\rm Im}\theta}.
\label{bnd1}
\end{equation}
{\em Step 2.} Consider the Feshbach map
\begin{eqnarray}
{\cal F}_z&=& P_\r(-z-\lambda^2 I_\theta \bar P_\r R(z)\bar P_\r I_\theta)P_\r\nonumber\\
&=& P_\r(-z-\lambda^2 I_\theta \bar P_\r R(0)\bar P_\r I_\theta)P_\r +O(\lambda^2|z|).
\label{diffg}
\end{eqnarray}
Let
\begin{equation}
{\cal G}_z = -\lambda^2 P_\r I_\theta \bar P_\r R(z)\bar P_\r I_\theta P_\r.
\label{np1}
\end{equation}
By the isospectrality property of the Feshbach map (see e.g. \cite[Theorem IV.1]{BFS2}) we know that
$$
{\cal G}_{\lambda^2\delta_{a,b}}\,\varphi_{a,b}\otimes\Omega_\r = \lambda^2\delta_{a,b} \,\varphi_{a,b}\otimes\Omega_\r,
$$
for all $a,b=1,\ldots,N$. We also have ${\cal G}_z-{\cal G}_\zeta = O(\lambda^2|z-\zeta|)$, as long as ${\rm Im}z, {\rm Im}\zeta <\frac14{\rm Im}\theta$. It follows that ${\cal G}_0\,\varphi_{a,b}\otimes\Omega_\r = \lambda^2\delta_{a,b}\,\varphi_{a,b}\otimes\Omega_\r +O(\lambda^4)$, for all $a,b=1,\ldots,N$. Therefore, ${\cal G}_0=\sum_{a,b=1}^N\lambda^2\delta_{a,b} |\varphi_{a,b}\rangle\langle\varphi_{a,b}|\otimes P_\r +O(\lambda^4)$,
and so
\begin{equation}
{\cal G}_z = \sum_{a,b=1}^N\lambda^2\delta_{a,b} |\varphi_{a,b}\rangle\langle\varphi_{a,b}|\otimes P_\r +O(\lambda^4+\lambda^2|z|).
\label{np2}
\end{equation}
Using \fer{np2} and \fer{np1} in \fer{diffg} shows that
\begin{equation}
{\cal F}_z = \sum_{a,b=1}^N (\lambda^2\delta_{a,b} -z) |\varphi_{a,b}\rangle\langle\varphi_{a,b}|\otimes P_\r +O(\lambda^4+\lambda^2|z|).
\label{np3}
\end{equation}
The sum on the right side is an invertible operator, the norm of the inverse being
$$
\max_{a,b=1,\ldots,N}|\lambda^2\delta_{a,b}-z|^{-1}=[{\rm dist}({\cal E},z)]^{-1}.
$$
Therefore, there is a constant $C$ s.t. if
\begin{equation}
\lambda^4+\lambda^2|z|< C\,{\rm dist}({\cal E},z),
\label{np5}
\end{equation}
then ${\cal F}_z$ is invertible and
\begin{equation}
\|{\cal F}_z^{-1}\|\leq \frac{2}{{\rm dist}({\cal E},z)}.
\label{np4}
\end{equation}
Let $\alpha>0$ be fixed, and take $z$ s.t. ${\rm dist}({\cal E},z)\geq\alpha\lambda^2$. Then \fer{np5} is satisfied provided $\lambda$ is small enough and $|z|< C\alpha$.
\medskip
{\em Step 3.} The resolvent $R(z)$ is related to $\bar P_\r R(z)\bar P_\r$ and ${\cal F}_z^{-1}$ by (see e.g. \cite[Eqn. (IV.14)]{BFS2})
$$
R(z) = \big( P_\r -\bar P_\r R(z)\bar P_\r K_\theta(0,\lambda)P_\r\big){\cal F}_z^{-1} \big( P_\r - P_\r K_\theta(0,\lambda)\bar P_\r R(z)\bar P_\r\big) +\bar P_\r R(z)\bar P_\r.
$$
We combine this equation with the bounds $\|\bar P_\r K_\theta(0,\lambda)P_\r\|$, $\|P_\r K_\theta(0,\lambda)\bar P_\r\|\leq C|\lambda|$ and \fer{bnd1}, \fer{np4} to arrive at the estimate \fer{resbnd1}. This completes the proof of Proposition \ref{resbndprop}.\hfill $\square$
{\bf (B) Zero eigenvalue.} Let $P(\sigma)$ be the group projection associated to the eigenvalues of $K_\theta(\sigma,\lambda)$ bifurcating out of the origin as $\sigma\neq 0$. Here, we consider $\lambda$ fixed and $\sigma$ small. The null space of $K_\theta(0,\lambda)$ is known exactly, see \fer{remark}. Let $X^*_{a,a}\in{\cal H}_\r$ be the vector satisfying $K^*_{a,a}X^*_{a,a}=0$ and $\scalprod{\Omega_\r}{X^*_{a,a}}=1$. We have $X^*_{a,a}=\Omega_\r+O(\lambda)$. Then $P(0)=\sum_{a=1}^N|\varphi_{a,a}\rangle\langle\varphi_{a,a}|\otimes|\Omega_\r\rangle\langle X^*_{a,a}|$. Note that $P(0) L_{\rm S} P(0)=0$. Analytic perturbation theory gives
\begin{eqnarray}
K_\theta(\sigma,\lambda) P(\sigma)&=& \sigma^2 T_2 +O_\lambda(\sigma^3)
\nonumber\label{mmm1'}\\
T_2&=& -P(0) L_{\rm S} K_\theta(0,\lambda)^{-1} L_{\rm S} P(0).
\label{mmm2}
\end{eqnarray}
We have $ L_{\rm S} P(0) = \sum_{a=1}^N\sum_{c,d=1,\ldots,N; c\neq d} |\varphi_{c,d}\rangle\langle\varphi_{a,a}| \otimes P_\r\scalprod{\varphi_{c,d}}{ L_{\rm S} \varphi_{a,a}} +O(\lambda)$. Next,
\begin{eqnarray}
K_\theta(0,\lambda)^{-1} \varphi_{c,d}\otimes \Omega_\r &=& K_\theta(0,\lambda)^{-1} \varphi_{c,d}\otimes (X_{c,d}+\Omega_\r-X_{c,d})\nonumber\\
&=& \frac{1}{\lambda^2\delta_{c,d}} \varphi_{c,d}\otimes \Omega_\r +O(\lambda^{-1}),
\label{mnm1}
\end{eqnarray}
where we use Corollary \ref{corollary1} in the last step. Starting from \fer{mmm2} and using \fer{mnm1}, we arrive at
\begin{equation}
T_2 = \frac{2\i}{\lambda^2} T+O(\lambda^{-1}),
\label{mmm5}
\end{equation}
where the operator $T$ has matrix elements $[T]_{a,b}=\scalprod{\varphi_{a,a}\otimes\Omega_\r}{T\,\varphi_{b,b}\otimes\Omega_\r}$ given by \fer{mmm6}. In this derivation, we also use that $\delta_{b,a}=-\overline{\delta_{a,b}}$, see \fer{delta}. Note that $T$ is a real symmetric matrix, $[T]_{a,b}<0$ for $a\neq b$, and $[T]_{a,a}=-\sum_{b\neq a}[T]_{a,b}$. These properties imply that for $x=(x_1,\ldots,x_N)\in {\mathbb C}^N$, $\scalprod{x}{T x} = \sum_{a,b=1}^N | [T]_{a,b}|\, |x_a-x_b|^2\geq 0$. Therefore, if $[T]_{a,b}\neq 0$ for all $a\neq b$, then zero is a simple eigenvalue of $T$, with eigenvector proportional to $(1,\ldots,1)$ and all other eigenvalues of $T$ are strictly positive.
This completes the proof of Theorem \ref{thm4}. \hfill $\square$
\subsection{Proof of Theorem \ref{dynthm}}
The proof of these two theorems is based on the resolvent representation, Theorem \ref{thm2}, and on the spectral data given in Theorem \ref{thm4}. The procedure follows \cite{JP,BFS,MMS} (for the path integration deformation argument) and \cite[Theorem 3.1]{MSB} (for the reduced dynamics).
Let $\Psi_0\in{\cal H}$ (initial state). Given $\epsilon>0$, we can find a vector $\Psi_\epsilon$ such that (a) $|\scalprod{\Psi_0}{A\Psi_0}-\scalprod{\Psi_\epsilon}{A\Omega}|<\|A\| \epsilon$, for all $A\in{\frak M}$, where $\Omega$ is the reference state \fer{m20}, and (b) $\Psi_\epsilon$ is $U_\theta$-analytic and $U_{\bar\theta}\Psi_\eta\equiv(\Psi_\eta)_{\bar\theta}$ is in the domain of ${\rm e}^{|L_\r|/2}$. To produce $\Psi_\epsilon$, one may first find $B\in{\frak M}'$ (commutant of ${\frak M}$) s.t. $\|\Psi_0-B\Omega\|<\epsilon/2$ (this can be done by the cyclicity of $\Omega$) and set $\Psi_{1,\epsilon}=B^*B\Omega$. Then (a) is verified. Next, one regularizes this vector to satisfy (b), e.g. by forming $\Psi_{2,\epsilon} ={\rm e}^{-\eta L^2_\r}{\rm e}^{-\eta D^2} {\rm e}^{-4 \eta\theta_0^2 N^2}\Psi_{1,\epsilon}$, where $D=\d\Gamma(-\i\partial_u)$ is the generator of spectral deformation and $N=\d\Gamma(\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l)$ is the number operator. Taking $\eta>0$ small enough gives $\Psi_\epsilon$ satisfying (a) and (b). The set of translation-analytic functionals
\begin{equation}
\label{anstates}
{\cal S}_0 = \{\scalprod{\Psi}{\,\cdot\,\Omega} : \Psi \mbox{\ satisfies (a) and (b)}\}
\end{equation}
is hence dense in the set of all states on ${\frak M}$. The translation-analytic observables are defined by
\begin{equation}
\label{anobs}
{\frak M}_0 = \{ A\in{\frak M} : A\Omega \mbox{\ is $U_\theta$-analytic}\}.
\end{equation}
Let $\omega_0\in{\cal S}_0$ and $A\in{\frak M}_0$. Theorem \ref{thm2} gives
$$
\omega_0(\alpha^t_{\sigma,\lambda}(A)) =
\frac{-1}{2\pi\i}\int_{{\mathbb R}-\i} {\rm e}^{\i
tz}\scalprod{\Psi_{\overline\theta}}{(K_\theta(\sigma,\lambda)-z)^{-1}
(A\Omega)_\theta}\d z.
\label{mm2'}
$$
We deform the contour of integration into the upper half-plane, as in \cite{JP,BFS,MMS}, to pick up the contributions of the poles at the resonance energies of the resolvent by means of the residue theorem. The integral over the path ${\mathbb R}-\i$ equals the integral over the path ${\mathbb R}+\frac34\i{\rm Im}\theta$ plus the sum of the integrals around circles $\Gamma_{a,b}$, each enclosing exactly one eigenvalue $\varepsilon_{a,b}$ of $K_\theta(\sigma,\lambda)$. While the integral over ${\mathbb R}+\frac34\i{\rm Im}\theta$ is $O({\rm e}^{-\frac34 t\,{\rm Im}\theta})$, the integral around a given eigenvalue $\varepsilon_{a,b}$ is
\begin{equation}
\label{new1}
\frac{-1}{2\pi \i}\oint_{\Gamma_{a,b}} {\rm e}^{\i tz}\scalprod{\Psi_{\bar\theta}}{(K_\theta(\sigma,\lambda)-z)^{-1} A \Omega}\d z = {\rm e}^{\i t\varepsilon_{a,b}(\sigma,\lambda)} \scalprod{\Psi_{\bar\theta}}{{\widetilde Q}_{a,b}(A\Omega)_\theta},
\end{equation}
where ${\widetilde Q}_{a,b} = \frac{-1}{2\pi\i}\oint_{\Gamma_{a,b}} (K_\theta(\sigma,\lambda)-z)^{-1}\d z$ is the Riesz spectral projection.
The KMS state of the uncoupled system ($\lambda=0$) is given by the standard vector $\Omega_0=\Omega_{{\rm S},\beta}\otimes\Omega_\r$. Here, $\Omega_{{\rm S},\beta}$ is the unique vector in the standard natural cone, the closure of $\{AJ_{\rm S} A\Omega_{\rm S}: A\in{\frak M}_{\rm S}\}$ (recall the definition of $\Omega_{\rm S}$ and $J_{\rm S}$ given in and after \fer{m-1}), representing the system Gibbs equilibrium state (which is determined by the density matrix $\propto{\rm e}^{-\beta\sigma H_{\rm S}}$). Perturbation theory of KMS states (see \cite{BR,DJP,BFS}) tells us that
\begin{equation}
\label{coupledkms}
\Omega_{{\rm S}\r}={\rm e}^{-\beta L(\sigma,\lambda)/2}\Omega_0/\|{\rm e}^{-\beta L(\sigma,\lambda)/2}\Omega_0\|,
\end{equation}
where $L(\sigma,\lambda)$ is given in \fer{m11'}, is the KMS state for the interacting system.
Consider $\sigma>0$. Since ${\rm Im}\varepsilon_{a,b}>0$ for all $a\neq b$ and ${\rm Im}\varepsilon_{a,a}>0$ for $a=2,\ldots,N$ and since $\Omega_{{\rm S}\r}$ is an invariant state, it follows by taking $t\rightarrow\infty$ that the quantity \fer{new1} for $a=b=1$ is $\scalprod{\Omega_{{\rm S}\r}}{A\Omega_{{\rm S}\r}}=\omega_{\beta,\sigma,\lambda}(A)$. The remaining contributions to the right side of \fer{1} come from the resonances bifurcating out of the origin (first sum) and those bifurcating out of $\varepsilon_{a,b}(0,\lambda)$, as $\sigma$ becomes nonzero. We have $\chi_a(A)= \langle{\Psi_{\bar\theta}},{{\widetilde Q}_{a,a}(A\Omega)_\theta}\rangle$, and a similar definition for $\chi_{a,b}$.
Consider $\sigma=0$. Then $\varepsilon_{a,a}(0,\lambda)=0$ for all $a=1,\ldots,N$. The first two terms on the right side of \fer{1} arise from the projection onto the kernel of $K_\theta(0,\lambda)$. This defines the $\chi_a$ for $\sigma=0$. The $\chi_{a,b}$ are again given by the the scalar products on the right side of \fer{new1}.
Note that the $\chi_a$ are not continuous as $\sigma\rightarrow 0$, as only the total group projection associated to the eigenvalues bifurcating out of the origin is continuous (actually analytic), but not the individual projections.
\subsection{Proof of Theorem \ref{thmreddynfinal}}
\begin{thm}[Reduced dynamics]
\label{reduceddyn}
Let $\chi_1$ be an arbitrary normalized vector in ${\cal H}_{\rm S}$ and let $A\in{\frak M}_{\rm S}$ be a system observable. Then we have
\begin{eqnarray}
\lefteqn{
\scalprod{\chi_1\otimes\Omega_\r}{{\rm e}^{\i tL(\sigma,\lambda)} A {\rm e}^{-\i tL(\sigma,\lambda)}\Omega}}\nonumber\\
&=&\sum_{a,b=1}^N{\rm e}^{\i t\varepsilon_{a,b}(\sigma,\lambda)}\scalprod{\chi_1}{Q_{a,b}A\Omega_{\rm S}} \big( 1+O_\lambda(\sigma)+O(\lambda)\big) +O\big(\lambda^2{\rm e}^{-\frac34 t\theta_0}\big),
\label{l1}
\end{eqnarray}
where the $\varepsilon_{a,b}(\sigma,\lambda)$ are given in \fer{l7}. Here,
\begin{equation}
Q_{a,b} =
\left\{
\begin{array}{ll}
|\varphi_{a,b}\rangle\langle\varphi_{a,b}| & \mbox{if $a\neq b$}\\
|\varphi^T_a\rangle\langle\varphi^T_a| & \mbox{if $a=b$,}
\end{array}
\right.
\label{l3}
\end{equation}
where $\{\varphi_a^T\}_{a=1}^N$ is the orthonormal basis of eigenvectors of $T$, \fer{mmm6}, so that $T\varphi_a^T=\xi_a\varphi_a^T$.
\end{thm}
\noindent
{\bf Proof of Theorem \ref{reduceddyn}.\ } Take the representation \fer{mm2} for a fixed $\theta$. The integral over the path ${\mathbb R}-\i$ equals the integral over the path ${\mathbb R}+\frac34\i{\rm Im}\theta$ plus the sum of the integrals around circles $\Gamma_{a,b}$, each enclosing exactly one eigenvalue $\varepsilon_{a,b}$ of $K_\theta(\sigma,\lambda)$. While the integral over ${\mathbb R}+\frac34\i{\rm Im}\theta$ is $O(\lambda^2{\rm e}^{-\frac34 t\,{\rm Im}\theta})$ (see Proposition \ref{prop7}), the integral around a given eigenvalue $\varepsilon_{a,b}$ is
$$
\frac{-1}{2\pi \i}\oint_{\Gamma_{a,b}} {\rm e}^{\i tz}\scalprod{\chi_1\otimes\Omega_\r}{(K_\theta(\sigma,\lambda)-z)^{-1} A \Omega}\d z = {\rm e}^{\i t\varepsilon_{a,b}} \scalprod{\chi_1\otimes\Omega_\r}{{\widetilde Q}_{a,b}A\Omega},
$$
where ${\widetilde Q}_{a,b} = \frac{-1}{2\pi\i}\oint_{\Gamma_{a,b}} (K_\theta(\sigma,\lambda)-z)^{-1}\d z$ is the Riesz spectral projection. By perturbation theory, we have, for $a\neq b$,
$$
\widetilde{Q}_{a,b}=|\varphi_{a,b}\rangle\langle\varphi_{a,b}|\otimes|X_{a,b}\rangle\langle X_{a,b}^*| +O_\lambda(\sigma)=|\varphi_{a,b}\rangle\langle\varphi_{a,b}|\otimes|\Omega_\r\rangle\langle \Omega_\r| +O_\lambda(\sigma)+O(\lambda).
$$
Similarly, we have $\widetilde{Q}_{a,a}=|\varphi_a^T\rangle\langle\varphi^T_a|\otimes|\Omega_\r\rangle\langle\Omega_\r|+O_\lambda(\sigma)$. (Note that $T$ is self-adjoint.) This completes the proof of Theorem \ref{reduceddyn}.\hfill $\square$
We now prove Theorem \ref{thmreddynfinal}. Let $\rho_0$ be the initial density matrix of the small system. It is represented by a normalized vector $\chi$ in the GNS space ${\cal H}_{\rm S}$. By the cyclicity of $\Omega_{\rm S}$ there is a unique element $B'$ in the commutant ${\frak M}_{{\mathbb C}^N}'=\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{{\mathbb C}^N}\otimes{\cal B}({\cal H}_{\rm S})$ such that $\chi=B'\Omega_{\rm S}$. The evolution of the reduced density matrix elements $[\rho_t]_{a,b}=\scalprod{\varphi_a}{\rho_t\varphi_b}$ is given by
\begin{eqnarray}
{}[\rho_t]_{a,b} &=& \scalprod{\chi\otimes\Omega_\r}{{\rm e}^{\i t L(\sigma,\lambda)} (|\varphi_b\rangle\langle\varphi_a|\otimes\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{{\mathbb C}^N}){\rm e}^{-\i t L(\sigma,\lambda)}\chi\otimes\Omega_\r}\nonumber\\
&=& \scalprod{\chi\otimes\Omega_\r}{B'{\rm e}^{\i t L(\sigma,\lambda)}( |\varphi_b\rangle\langle\varphi_a|\otimes\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{{\mathbb C}^N}){\rm e}^{-\i t L(\sigma,\lambda)}\Omega}.
\end{eqnarray}
We can thus use Theorem \ref{reduceddyn}. The main term on the right side of \fer{l1} is
\begin{equation}
\sum_{c,d=1}^N{\rm e}^{\i t\varepsilon_{c,d}(\sigma,\lambda)}\scalprod{\chi}{B'Q_{c,d} ( |\varphi_b\rangle\langle\varphi_a|\otimes\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{{\mathbb C}^N}) \Omega_{\rm S}}=\frac{1}{\sqrt N} \sum_{c,d=1}^N{\rm e}^{\i t\varepsilon_{c,d}(\sigma,\lambda)}\scalprod{\chi}{B'Q_{c,d} \, \varphi_{b,a}},
\label{l2}
\end{equation}
by the definition \fer{m-1} of $\Omega_{\rm S}$. If $a\neq b$ then, according to \fer{l3}, $Q_{c,d} \varphi_{b,a}$ vanishes, except when $c=b$ and $d=a$, in which case it equals $\varphi_{b,a}$. Then we have $\scalprod{\chi}{B'\varphi_{b,a}}=\sqrt N \scalprod{\chi}{B'(|\varphi_b\rangle\langle\varphi_a|\otimes\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{{\mathbb C}^N}|)\Omega_{\rm S}}=\sqrt N\, [\rho_0]_{a,b}$. We conclude that for $a\neq b$, the main term of $[\rho_t]_{a,b}$ is ${\rm e}^{\i t\varepsilon_{b,a}(\sigma,\lambda)}[\rho_0]_{a,b}$. This shows \fer{l5}. Relation \fer{l9} is proven in the same way. \hfill $\square$
\subsection{Using the Feshbach map}
Zero is an eigenvalue of $K_\theta(0,0)$ of multiplicity $N^2$. By a simple Riesz projection argument, one shows that, for $\sigma$ and $\lambda$ small, $K_\theta(\sigma,\lambda)$ has $N^2$ eigenvalues in the vicinity of the origin. The size of the eigenvalues can be estimated as follows. Suppose that $z\neq 0$ and ${\rm Im}z<\frac12{\rm Im}\theta$, so that $z$ is in the resolvent set of $K_\theta(0,0)$. If the series
\begin{equation}
( K_\theta(0,0)-z)^{-1}\sum_{n\geq 0} \big[(\sigma L_{\rm S} +\lambda I_\theta)( K_\theta(0,0)-z)^{-1}\big]^n
\label{fesh0}
\end{equation}
converges, then $z$ belongs to the resolvent set of $K_\theta(\sigma,\lambda)$ and \fer{fesh0} equals $(K_\theta(\sigma,\lambda)-z)^{-1}$. Therefore, if $z$ is a (non-zero) eigenvalue of $K_\theta(\sigma,\lambda)$, then we must have
\begin{equation}
\|(\sigma L_{\rm S} +\lambda I_\theta)(K_\theta(0,0)-z)^{-1}\|\geq 1.
\label{fesh0.1}
\end{equation}
Using standard bounds on the interaction, we see that \fer{fesh0.1} implies that there are constants $C,c>0$ s.t. if $\sigma,|\lambda|<c$, then
\begin{equation}
|z| < C(\sigma+|\lambda|).
\label{fesh0.2}
\end{equation}
Estimate \fer{fesh0.2} is a bound on the eigenvalues of $K_\theta(\sigma,\lambda)$ in the vicinity of the origin. The eigenvalues can be tracked using the Feshbach map. Namely, $z\in\mathbb C$, ${\rm Im}z<\frac12{\rm Im}\theta$ is an eigenvalue of $K_\theta(\sigma,\lambda)$ if and only if it is an eigenvalue of the operator
\begin{equation}
{\cal F}_z = P_\r\left( \sigma L_{\rm S} -\lambda^2 I_\theta( K_\theta(\sigma,\lambda)-z)^{-1}I_\theta\right) P_\r
\label{fesh1}
\end{equation}
which acts on the smaller space ${\rm Ran}P_\r={\mathbb C}^N\otimes{\mathbb C}^N$. Recall that $P_\r=|\Omega_\r\rangle\langle\Omega_\r|$. By expanding the resolvent around $z=0$, $\sigma=0$ and $\lambda=0$, taking into account \fer{fesh0.2}, we have
\begin{equation}
{\cal F}_z= P_\r\left( \sigma L_{\rm S} -\lambda^2I_\theta K_\theta(0,0)^{-1}I_\theta\right) P_\r +O\Big(\lambda^2\big(\sigma+|\lambda|\big)\Big),
\label{fesh2}
\end{equation}
provided $z$ is an eigenvalue of $K_\theta(\sigma,\lambda)$ and $\sigma, |\lambda|<c$. An elementary calculation shows that the operator ${\cal F}_z$, viewed as acting on ${\rm Ran}P_\r$, has the form
\begin{equation}
{\cal F}_z=\sigma L_{\rm S} -\lambda^2\big(\alpha G^2\otimes \mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l -\alpha G\otimes G+\overline{\alpha}G\otimes G -\overline{\alpha}\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l\otimes G^2 \big)+ O\Big(\lambda^2\big(\sigma+|\lambda|\big)\Big),
\label{fesh3}
\end{equation}
where ${\cal C}$ is defined after \fer{m11} and $\alpha= \frac12 \scalprod{g}{|k|^{-1}g}- \frac{\i}{2}\pi\xi(0)$, with $\xi(0)$ given in \fer{xinot}. Note that the quadratic term in $\lambda$ is diagonal in the basis $\varphi_{a,b}$,
\begin{eqnarray}
\lefteqn{
-\lambda^2\big(\alpha G^2\otimes \mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l -\alpha G\otimes G +\overline{\alpha}G\otimes G -\overline{\alpha}\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l\otimes G^2 \big)}\nonumber\\
&& = -\frac{\lambda^2}{2}\sum_{a,b=1}^N \big(\scalprod{g}{|k|^{-1}g}(g^2_a-g^2_b) -\i\pi\xi(0)(g_a-g_b)^2\big) |\varphi_{a,b}\rangle\langle\varphi_{a,b}|.
\label{fesh4}
\end{eqnarray}
We conclude from the isospectrality of the Feshbach map and \fer{fesh3}, \fer{fesh4} that the eigenvalues of $K_\theta(0,\lambda)$ are given by
$ -\frac{\lambda^2}{2}\big(\scalprod{g}{|k|^{-1}g}(g^2_a-g^2_b) -\i\pi\xi(0)(g_a-g_b)^2\big)$,
modulo a remainder $O(\lambda^2(\sigma+|\lambda|))$. This is compatible with the result of Theorem \ref{thm3}. However, from that Theorem, we know in addition that the remainder actually vanishes.
\subsection{The spin-boson system}
\label{sectspinboson}
The Feshbach operator \fer{fesh3} is represented in the energy basis $\{\phi_{+,+},\ \phi_{+,-},\ \phi_{-,+}\ \phi_{-,-}\}$, where $\phi_{+,-}=\phi_+\otimes\phi_-$ (etc) and $S^z\phi_\pm=\pm\frac12\phi_\pm$, by the matrix
\begin{eqnarray}
{\cal F}_z&=& W + O\Big(\lambda^2\big(\sigma+|\lambda|\big)\Big),\\
W&=& \left(
\begin{array}{cccc}
\i\frac{\lambda^2}{4}\pi\xi(0) & 0 & 0 & -\i\frac{\lambda^2}{4}\pi \xi(0)\\
0 & \sigma+\i\frac{\lambda^2}{4}\pi \xi(0)& -\i\frac{\lambda^2}{4}\pi \xi(0) & 0 \\
0 & -\i\frac{\lambda^2}{4}\pi \xi(0) & -\sigma+\i\frac{\lambda^2}{4}\pi \xi(0) & 0 \\
-\i\frac{\lambda^2}{4}\pi \xi(0)& 0 &0 & \i\frac{\lambda^2}{4}\pi
\xi(0)
\end{array}
\right).
\end{eqnarray}
The four eigenvalues of $W$ are
\begin{equation}
w_1=0,\quad w_2=\i \frac{\lambda^2}{2}\pi \xi(0), \quad w_{3,4}=\i\frac{\lambda^2}{4}\pi\xi(0)\pm\i \sqrt{\frac{\lambda^4}{16}\pi^2\xi(0)^2-\sigma^2},
\label{evals}
\end{equation}
where the square root is the principal branch with branch cut on the negative real axis. The corresponding eigenvectors of $W$ are
\begin{equation}
\chi_1=\frac{1}{\sqrt 2}\left[\begin{array}{c} 1\\
0\\0\\1\end{array}\right],\ \
\chi_2=\frac{1}{\sqrt 2}\left[\begin{array}{c} 1\\
0\\0\\-1\end{array}\right],\ \
\chi_3= \frac{1}{1+r^2}\left[\begin{array}{c} 0\\
1\\
r\\0
\end{array}\right],\ \
\chi_4=\frac{1}{1+r^2}\left[\begin{array}{c} 0\\
-r\\1\\0\end{array}\right],
\label{evect}
\end{equation}
where $r=\frac{-4\i \gamma -\sqrt{\pi^2\xi(0)^2 -16\gamma^2}}{\pi\xi(0)}$ with $\gamma=\frac{\sigma}{\lambda^2}$.
The eigenvalues of the adjoint $W^*$ are the complex conjugates $\overline w_j$ and the corresponding eigenvectors are
\begin{equation}
\chi_1^*=\frac{1}{\sqrt 2}\left[\begin{array}{c} 1\\
0\\0\\1\end{array}\right],\ \
\chi_2^*=\frac{1}{\sqrt 2}\left[\begin{array}{c} 1\\
0\\0\\-1\end{array}\right],\ \
\chi_3^*=\left[\begin{array}{c} 0\\
1\\
\overline r \\0\end{array}\right],\ \
\chi_4^*=\left[\begin{array}{c} 0\\
-\overline r\\1\\0\end{array}\right].
\label{evect*}
\end{equation}
The eigenvectors are normalized as $\scalprod{\chi_i}{\chi_i^*}=1$ and $\langle\chi_i, \chi_j^*\rangle=0$ if $i\neq j$. The reduced spin density matrix, represented in the energy basis $\phi_\pm$, is given by (proceed as for Theorem \ref{thmreddynfinal} or see \cite[Theorem 2.1]{mjmp} and \cite{MSB})
\begin{equation}
[\rho_t]^z_{m,n}\doteq\sum_{j=1}^4{\rm e}^{\i t w_j}\sum_{k,l=\pm}[\rho_0]^z_{l,k}\scalprod{\phi_{k,l}}{\chi_j}\scalprod{
\chi_j^*}{\phi_{n,m}}.
\label{u1}
\end{equation}
Here, we take $m,n,k,l$ to stand for either $+$ or $-$, and $\doteq$ means that we approximate the true resonances $\varepsilon$ (the eigenvalues of ${\cal F}_z$) by the $w$ and we neglect additive $O(\lambda^2)$ terms (uniform in $t\geq 0$) on both sides. Using the explicit formulas \fer{evect}, \fer{evect*} for the eigenvectors $\chi_j$, $\chi^*_j$, we arrive at
\begin{equation}\label{1.106}
\begin{split}
[\rho_t]^z_{+,+}\doteq&\textstyle\frac{1}{2}+\frac{1}{2}{\rm e}^{\i t
w_2}([\rho_0]^z_{+,+}-[\rho_0]^z_{-,-}),\\
[\rho_t]^z_{+,-}\doteq &\textstyle \frac{r}{r^2+1}{{\rm e}^{\i
tw_3}(r[\rho_0]^z_{+,-}+[\rho_0]^z_{-,+})+\frac{1}{r^2+1}{\rm e}^{\i
tw_4}([\rho_0]^z_{+,-}-r[\rho_0]^z_{-,+})}.
\end{split}
\end{equation}
|
\section*{Introduction}
The study of the existence of stochastic models with some prescribed distributional properties has a long tradition in the theory of probability and various fields of application. Let $\{X_t\}_{t \in T}$ be a stochastic process on some index set $T$ (which may be finite or infinite with some topological structure). Typically, a real-valued summary statistic $\kappa^{(X)}(s,t)$ of the distribution of $(X_s,X_t)$ is of particular interest for all pairs $(s,t) \in T \times T$.
The question is whether for some prescribed function $\kappa$ on $T \times T$ a stochastic model $\{X_t\}_{t \in T}$ exists that \emph{realizes} $\kappa$, i.e. if $\kappa^{(X)}=\kappa$.
Recent accounts and surveys on such \emph{realization problems} with an emphasis on $\{0,1\}$-valued processes
(or random sets, two-phased media, binary processes)
include \cite{torquato02}, \cite{quint08}, \cite{emery10}, \cite{lm15} and \cite{la15}.
Also from a statistical point of view realization problems are important, namely for consistent inference.
As pointed out by \cite{lm15}, the question of realizability usually leads to a (possibly infinite and even in finite setups huge) set of positivity conditions for the quantity of interest, and secondly, to a set of regularity conditions if the topology of the underlying space is of interest as well.
These positivity conditions are needed in statistical applications to correct estimators $\widehat{\kappa}^{(X)}$ for $\kappa^{(X)}$ to an admissible function.
Let us consider a classical example. Assuming that the second moments
of a real-valued stochastic process $\{X_t\}_{t \in T}$ exist at each locaction $t \in T$, the process possesses a covariance function $C^{(X)}(s,t)=\text{Cov}(X_s,X_t)$. It is well-known that $C=C^{(X)}$ must be \emph{positive semi-definite}, i.e. $C(s,t)=C(t,s)$ and
\begin{align}\label{eq:psd}
\sum_{i=1}^m \sum_{j=1}^m a_ia_j C(t_i,t_j) \geq 0 \quad \forall \, (t_1,\dots,t_m) \in T^m, \, (a_1,\dots,a_m) \in \RR^m, \, m \in \NN.
\end{align}
Conversely, for any such function $C$, there exists a stochastic process $\{X_t\}_{t \in T}$ with covariance function $C^{(X)}=C$. The stochastic process $\{X_t\}_{t \in T}$ is not unique, but it may be chosen to be a centered Gaussian process as can be easily checked by Kolmogorov's extension theorem. Such a process on the space $T$ (if additionally equipped with some topology), may have very uncomfortable regularity properties. Several authors have established connections between the regularity of the covariance function $C$ and the existence of a corresponding stochastic process with a certain sample path regularity, \cf e.g.\ \cite{adler90} for an overview in case of continuity.
In statistical applications, the development of efficient non-parametric estimators for the covariance function that ensure positive semi-definiteness can be a challenging task, cf.\ e.g.\ \cite{halletal} and \cite{politis11}.
When it comes to the extreme values in the upper quantile regions of a real-valued stochastic process $\{X_t\}_{t \in T}$, summary measures like the covariance function often do not exist and would be genuinely inappropriate to characterize dependence.
Instead, among several other summary statistics that have emerged in an extreme value context (cf.\ for instance \cite{beirlantetal04} Section 8.2.7), the following bivariate quantity
\begin{align*}
\chi^{(X)}(s,t):=\lim_{\tau \to \tau_{\text{up}}} \PP(X_s > \tau \,|\, X_t > \tau), \qquad s,t \in T,
\end{align*}
which we call \emph{tail correlation function} (TCF) \citep{strokorbballanischlather_15},
has received particular attention.
As commonly done and in accordance with stationarity assumptions, we assume here and hereafter that $\{X_t\}_{t \in T}$ has identical one-dimensional marginal distributions with upper endpoint $\tau_{\text{up}}$ (which may be $\infty$).
Dating back to \cite{geffroy_5859}, \cite{sibuya_60} and \cite{tiagodeoliveira_62}, the TCF enjoys steady popularity among practitioners and scholars in order to account for tail dependence, albeit frequently reported
under different names.
The insurance, finance, economics and risk management literature knows it mainly as \emph{(upper) tail dependence coefficient} \citep{frahmetal05}, \emph{coefficient of (upper) tail dependence} \citep{mcneil2003modelling} or simply as \emph{(upper) tail dependence} \citep{patton06}.
In environmental contexts it has been additionally addressed as \emph{$\chi$-measure} \citep{colesheffernantawn_99}. Spatial environmental applications tend to prefer the equivalent quantity $2-\chi$, referred to as \emph{extremal coefficient function}. Among many others, the references \cite{blanchetdavison_11}, \cite{emks_15} and \cite{thibaudopitz_15} use it as an exploratory tool for testing the goodness of fit. In the context of stationary time series, the TCF constitutes a special case of the \emph{extremogram} \citep{davismikosch_09}. Moreover, the standard classification of the random pair $(X_s,X_t)$ as exhibiting either \emph{asymptotic/extremal independence} (when $\chi^{(X)}(s,t)=0$) or \emph{asymptotical/extremal dependence} (when $\chi^{(X)}(s,t)\in (0,1]$) is based on the TCF $\chi$.
Even though the TCF is a ubiquitous quantity within the extremes literature, surprisingly little is known about the class of TCFs and even less when it comes to the interplay of TCFs and their realizing models.
This is the central theme of the present text.
That is,
we are aiming at giving at least partial answers to the following questions:
{\it
\begin{enumerate}[(A)]
\item Can we decide if a given real-valued function $\chi: T\times T \rightarrow \RR$ is the TCF of a stochastic process $\{X_t\}_{t \in T}$?
\item If this is the case, can we find a specific stochastic process $\{X_t\}_{t \in T}$ with $\chi^{(X)}=\chi$?
\end{enumerate}
}
We also address the following regularity question.
{\it
\begin{enumerate}[(C)]
\item Does the continuity of a TCF $\chi$ imply the existence of a stochastic process realizing $\chi$ that additionally satisfies some regularity property?
\end{enumerate}
}
A satisfactory answer to Question (A) is desirable in a statistical context
in order to decide whether estimators of the TCF produce admissible TCFs as an outcome.
This concerns specifically spatial applications where one is bound to encounter very high-dimensional observations and therefore only partial low-dimensional information (such as the TCF) can be taken into account for inference. A first attempt to include properties of the class of TCFs to improve statistical inference can be found in \cite{schlathertawn_03}.
The TCF $\chi=\chi^{(X)}$ is a \emph{non-negative correlation function}. That is, $\chi$ is positive semi-definite in the sense of \eqref{eq:psd} with $\chi(s,t) \geq 0$ and $\chi(t,t)=1$ for all $s,t \in T$ (cf.\ e.g.\ \cite{schlathertawn_03}, \cite{davismikosch_09} and \cite{fasenetalii_10}).
However, even though TCFs are non-negative correlation functions, not all such functions are TCFs. For instance, $\eta:=1-\chi$ has to satisfy the \emph{triangle inequality}
\begin{align}\label{eqn:triangle}
\eta(s,t) \leq \eta(s,r) + \eta(r,t) \qquad r,s,t \in T
\end{align}
\citep{schlathertawn_03}. In the context of $\{0,1\}$-valued stochastic processes, it is well-known that the respective covariance functions obey this triangle inequality and implications are addressed \eg in \cite{matheron_87}, \cite{markov_95} and \cite{jiaoetalii_07}. If $T=\RR^d$ and the underlying process is stationary, then the function $h \mapsto \chi(o,h)$ (with $o \in \RR^d$ being the origin) cannot be differentiable unless it is constant.
The simplest TCFs are the constant function $\chi(s,t) = 1$ realized by a process of identical random variables, and the function $\chi(s,t)=\delta_{st}:=\Eins_{s=t}$ realized by a process of independent random variables.
Another example for $\chi^{(X)}(s,t)=\delta_{st}$ is a {Gaussian process} $X$ on $T$, whose correlation function $\rho$ on $T \times T$ attains the value $1$ only on the diagonal $\{(t,t) : t \in T\}$ \citep[Theorem~3]{sibuya_60}.
While Gaussian processes do not exhibit tail dependence, the class of \emph{max-stable processes} naturally provides rich classes of non-trivial TCFs.
For instance, any function of the form $\chi(s,t)=\int_{[0,\infty)} \exp\left(-\lambda \lVert s - t \rVert \right) \Lambda(d\lambda)$ will be the TCF of a max-stable process on $T=\RR^d$, if $\Lambda$ is a probability measure on $[0,\infty)$ \citep{strokorbballanischlather_15}.
Beyond the realizability question, \cite{kabluchkoschlather_10} and \cite{wangroystoev_13} establish some connections between mixing properties of $X$ and decay properties of its TCF $\chi^{(X)}$ when $T=\RR^d$ and $X=\{X_t\}_{t \in \RR^d}$ is stationary and {max-stable}.
It is natural to ask whether even further TCFs will arise if we do not restrict ourselves to the max-stable class, since an affirmative answer would imply a first important reduction for the questions $(A)$-$(C)$.
{\it
\begin{enumerate}[(D)]
\item Is the set of TCFs stemming from max-stable processes properly contained in the set of all TCFs or do these sets coincide?
\end{enumerate}
}
Finally, realization problems are usually intimately connected with the question of admissible operations on the quantities of interest. To illustrate this again by means of covariance funcions, note that the product and convex combination of two covariance functions and the pointwise limit of a sequence of covariance functions is again a covariance function. We ask the same question for TCFs.
{\it
\begin{enumerate}[(E)]
\item Is the set of TCFs closed under basic operations such as\\ taking (pointwise) products, convex combinations and limits?
\end{enumerate}
}
In order to deal with the questions above, we establish close connections with $\{0,1\}$-valued processes, polytopes, partitions of sets and combinatorics.
Recent developments indicate that such tools may appear more frequently in the analysis of extremes, cf.\ \cite{molchanov_08}, \cite{yuenstoev14}, \cite{wangstoev11}, \cite{ehw15} and \cite{thibaudetal15}.
We divide the text into two parts.
\textbf{Part I} deals with the realization problem of TCFs of stochastic processes $\{X_t\}_{t \in T}$ on arbitrary base spaces $T$.
Close connections with $\{0,1\}$-valued processes will be established and enter the subsequent considerations.
We give answers to Questions (D) and (E), partial answers to the Questions (A), (B) and (C) and reduce Question (A) to infinitely (countably) many finite-dimensional problems (in case our base space $T$ is countable).
\textbf{Part II} deals with these finite-dimensional problems, that is, the realization problem of TCFs of random vectors $\{X_t\}_{t \in T}$ on finite base spaces $T$ with $\lvert T \rvert=n$ for some $n \in \NN$. We are aiming at establishing a (reduced) system of necessary and sufficient conditions for deciding whether a given function is a TCF or not and study the geometry of the set of TCFs. Arguments used in this part will be related to the study of polytopes, are often of combinatorial nature or are based on additional software computations. The latter is a typical phenomenon for realization problems of this kind.
More detailed descriptions are given at the beginning of each part. Finally, we end with a discussion of our results.
The appendix contains all tables.\\[4mm]
{\noindent \textbf{\large Part I\\[2mm] The realization problem for TCFs on arbitrary sets $T$}}\\[3mm]
\noindent To start with, Section~\ref{sect:MST_ECF_TM} reviews some preparatory results on max-stable processes, extremal coefficient functions and a particular subclass of max-stable processes, which we called Tawn-Molchanov (TM) processes \citep{strokorbschlather_15}. These processes are important for our analysis, since it turns out that any TCF can be realized by (at least one) TM process, our main result in Section~\ref{sect:realizedTM} and a substantial reduction of the realization problem of TCFs. Section~\ref{sect:realizedTM} also reveals a close connection between the class of TCFs and the class of correlation functions of $\{0,1\}$-valued stochastic processes and addresses the existence of stochastic processes for a prescribed TCF $\chi$ with some minimal regularity properties if $\chi$ is at least continuous. Subsequently, Section~\ref{sect:props} collects some immediate consequences concerning closure properties of the set of TCFs and the characterization of the set of TCFs by means of finite-dimensional inequalities, our starting point for Part II.
\section{{Max-stable processes, extremal coefficients and TM processes}}\label{sect:MST_ECF_TM}
A stochastic process $X=\{X_t\}_{t \in T}$ is \emph{simple max-stable}, if it has unit Fr{\'e}chet margins (meaning $\PP(X_t \leq x)=\exp(-1/x)$ for all $t \in T$ and $x > 0$), and if the maximum process $\Max_{i=1}^n X^{(i)}$ of independent copies of $X$ has the same finite dimensional distributions (f.d.d.) as the process $n X$ for each $n \in \NN$.
The crucial point in the realization problem for TCFs will be the close connection of the TCF $\chi^{(X)}$ of a simple max-stable process $X=\{X_t\}_{t \in T}$ to the \emph{extremal coefficient function (ECF)} $\theta^{(X)}$ of the respective process $X$. Therefore, let $\finite(T)$ denote the \emph{set of finite subsets} of the space $T$. The ECF $\theta^{(X)}$ of a simple max-stable process $X$ on $T$ is a function on $\finite(T)$ that is given by $\theta^{(X)}(\emptyset):=0$ and
\begin{align*}
\theta^{(X)}(A):= - \tau \log \PP\Big(\Max_{t \in A} X_t \leq \tau \Big), \qquad \tau > 0,
\end{align*}
in case $A \neq \emptyset$. The \rhs is indeed independent of $\tau>0$ and lies in the interval $[1,|A|]$, where $|A|$ denotes the number of elements in $A$. In fact, the value $\theta^{(X)}(A)$ can be interpreted as the effective number of independent random variables in the collection $\{X_t\}_{t \in A}$ (\cf \cite{smith_90,schlathertawn_02}). We call the set of all possible ECFs of simple max-stable processes
\begin{align}\label{eqn:Theta}
\Theta(T)=\left\{\,\theta^{(X)}: \finite(T) \rightarrow \RR \pmid
X \text{ a simple max-stable process on } T\,
\right\}.
\end{align}
The bounded ECFs will be denoted
\begin{align}\label{eqn:Theta_bounded}
\Theta_b(T)=\left\{\,\theta \in \Theta(T) \pmid \theta \text{ is bounded}\,\right\}.
\end{align}
In fact, the set of ECFs $\Theta(T)$ can be completely characterized by a property called \emph{complete alternation} (\cf Theorem~\ref{thm:ECF_CA} below). Using the notation and definition from \cite{molchanov_05}, we set for a function $f:\finite(T) \rightarrow \RR$ and elements $K,L \in \finite(T)$
\begin{align*}
\left(\Delta_{K}f\right)(L):= f(L)-f(L\cup K).
\end{align*}
Then a function $f: \finite(T) \rightarrow \RR$ is called \emph{completely alternating} on $\finite(T)$ if for all $n \geq 1$, $\{K_1,\dots, K_n\} \subset \finite(T)$ and $K \in \finite(T)$
\begin{align}\label{eqn:defn_CA}
\left(\Delta_{K_1}\Delta_{K_2} \dots \Delta_{K_n} f \right)(K) =
\sum_{I \subset \{1, \dots, n\}} (-1)^{|I|} \,f\left(K \cup \bigcup_{i \in I} K_i\right)
\leq 0.
\end{align}
This condition can be slightly weakened as in Lemma~\ref{lemma:CA_finite} below. Its proof uses the following auxiliary argument.
\begin{lemma}\label{lemma:subsetinduction}
Let $M$ be a finite set and $f:\finite(M) \rightarrow \RR$ be a function on the subsets of $M$. Let $K,L \subset M$ with $K \cap L =\emptyset$. Then
\begin{align}\label{eqn:assertionM}
\sum_{I \subset L} (-1)^{|I|+1} f(K \cup I) = \sum_{J \subset (K \cup L)^c} \Big(\sum_{I \subset L \cup J} (-1)^{|I|+1} f((L \cup J)^c \cup I)\Big).
\end{align}
\end{lemma}
\begin{proof}
Each set $(L\cup J)^c\cup I$ occuring on the r.h.s. can be written as a disjoint union $K \cup A \cup B$, with $A \subset L, B \subset (K\cup L)^c$. Let us consider the terms on the r.h.s.\ with fixed $A \subset L$ and fixed $B \subset (K\cup L)^c$.
If $B = \emptyset$, the only possible $I$ and $J$ leading to such a situation are $I=A$ and $J = (K\cup L)^c$, i.e., one obtains the term on the l.h.s.\ with $I = A$.
If $B \neq \emptyset$, the possibilities can be listed as $I = A \cup (B\setminus C)$ and $J = (K\cup L\cup C)^c$ for some $C \subset B$. Summing these terms over all $C \subset B$ yields $\sum_{C\subset B}(-1)^{|A|+|B\setminus C|+1} f(K \cup A \cup B)= (-1)^{|A|+1}(1-1)^{|B|}f(K \cup A \cup B) = 0$.
\end{proof}
It follows that for finite sets $M$ (instead of arbitrary $T$) complete alternation can be formulated by bounding the value $f(M)$ by lower order values $f(L)$ for $L \subset M$ as follows (\cf also \cite{schlathertawn_02}, Ineq.~(12)).
\begin{lemma}\label{lemma:CA_finite}
\begin{enumerate}[a)]
\item
A function $f: \finite(T) \rightarrow \RR$ is {completely alternating} on $\finite(T)$ if and only if for all $\emptyset \neq L \in \finite(T)$ and $K \in \finite(T)$ with $K \cap L = \emptyset$
\begin{align}\label{eqn:short_CA}
\sum_{I \subset L} (-1)^{|I|+1} f\left(K \cup I\right) \geq 0.
\end{align}
\item Let $M$ be a non-empty finite set. Then $f: \finite(M) \rightarrow \RR$ is completely alternating if and only if (\ref{eqn:short_CA}) holds for all $\emptyset \neq L \subset M$ and $K=L^c$, which is equivalent to
\begin{align}
\label{eqn:finite_CA} \Max_{\substack{L \subset M\\|L| \text{ \normalfont odd}}} \, \sum_{\substack{I \subset L\\ I \neq L}} (-1)^{|I|} f\left(L^c \cup I\right)
& \leq f(M)
\leq
\Min_{\substack{\emptyset \neq L \subset M \\ |L| \text{ \normalfont even}}} \, \sum_{\substack{I \subset L\\ I \neq L}} (-1)^{|I|+1} f\left(L^c \cup I\right).
\end{align}
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}[a)]
\item Note that $\finite(T)$ forms an abelian semigroup \wrt the union operation that is generated already by the singletons $\{t\}$ for $t \in T$ and that $\Delta_{\{t\}}\Delta_{\{t\}}=\Delta_{\{t\}}$. Therefore, it suffices already to require (\ref{eqn:defn_CA}) only for $K_i=\{t_i\}$ for pairwise distinct elements $t_i \in T$ ($i=1,\dots,n$) (\cf \cite{bcr_84}, Proposition 4.6.6). Set $L=\{t_1,\dots,t_n\}$. Hence $f$ is completely alternating on $\finite(T)$ if and only if for all $\emptyset \neq L \in \finite(T)$ and $K \in \finite(T)$ the inequality (\ref{eqn:short_CA}) holds. Secondly, the expression on the l.h.s.\ of (\ref{eqn:short_CA}) equals automatically $0$ if $K \cap L \neq \emptyset$.
\item
Because of \eqref{eqn:assertionM},
it suffices to check (\ref{eqn:short_CA}) for $\emptyset \neq L \subset M$ and $K = L^c$. Separating $f(M)$ and summarizing the cases where $|L|$ is odd and where $|L|$ is even yields the second equivalence.
\end{enumerate}
\end{proof}
The following example shows that the concept of complete alternation is closely linked to the distributions of $\{0,1\}$-valued processes.
\begin{example}[\cite{molchanov_05}, p.~52] \label{example:cap_CA}
Let $Y=\{Y_t\}_{t \in T}$ be a stochastic process with values in $\{0,1\}$ and let the function $C^{(Y)}:\finite(T) \rightarrow [0,1]$ be given by $C^{(Y)}(\emptyset) = 0$ and $C^{(Y)}(A) = \PP(\exists \, t \in A \text{ such that } Y_t = 1)$.
Then $C^{(Y)}$ is completely alternating. Conversely, if $C:\finite(T) \rightarrow [0,1]$ is completely alternating with $C(\emptyset) = 0$, then $C$ determines the \fdd of a stochastic process $Y=\{Y_t\}_{t \in T}$ with values in $\{0,1\}$, such that $C^{(Y)}=C$.
\end{example}
\begin{remark}[\cite{molchanov_05}, p.~10]
From the perspective of the theory of random sets it is more natural to define a functional $C^{\Xi}(K)=\PP(\Xi \cap K \neq \emptyset)$ for a random closed set $\Xi$ on compact sets $K$. In this case, $C^{\Xi}$ will be termed the \emph{capacity functional} of the random closed set $\Xi$ and is not only completely alternating on compact sets, but also upper semi-continuous in the sense that $C^{\Xi}(K_n) \downarrow C^{\Xi}(K)$ for $K_n \downarrow K$. These properties ensure that $\Xi$ can be defined on a sufficiently regular probability space.
A priori our considerations below do not include any regularity constraints. However, we will come back to {Question~(C)} in Corollary~\ref{cor:chicty} and Remark~\ref{rk:chicty}.
\end{remark}
\begin{theorem}[\cite{strokorbschlather_15}, Theorem 8] \label{thm:ECF_CA}
$\phantom{a}$\\
Let $\theta: \finite(T) \rightarrow \RR$ be a function on the finite subsets of $T$. Then
\begin{align*}
\theta \in \Theta(T) \qquad \Longleftrightarrow \qquad
\left\{\begin{array}{l}
\theta \text{ is completely alternating, } \\
\theta(\emptyset)=0,\\ \theta(\{t\})=1 \text{ for } t \in T.
\end{array}\right.
\end{align*}
If $\theta \in \Theta(T)$, then there exists a simple max-stable process $X=\{X_t\}_{t \in T}$ on $T$ with ECF $\theta^{(X)}=\theta$, whose \fdd are given by
\begin{align*}
&-\log \, \PP\left(X_{t_i} \leq x_i\,,\, i=1,\dots,m\right) = \\
&\sum_{k=1}^m\,\sum_{1\leq i_1 < \dots < i_k \leq m} \hspace{-3mm} -\Delta_{\{t_{i_1}\}} \dots \Delta_{\{t_{i_k}\}} \theta \, (\{t_1,\dots,t_m\}\setminus \{t_{i_1},\dots,t_{i_k}\}) \Max_{j \in \{i_1,\dots,i_k\}} \hspace{-3mm}{x_j^{-1}}.
\end{align*}
\end{theorem}
If a process $\{X_t\}_{t \in T}$ has the f.d.d.\ stated in Theorem~\ref{thm:ECF_CA}, then it is called \emph{Tawn-Molchanov process (TM process) associated with the ECF $\theta$} henceforth. Note that this convention and the notation from \cite{MolchanovStrokorb} differ in the sense that \cite{MolchanovStrokorb} consider TM processes with at least upper-semi continuous sample paths.
By construction, the class of f.d.d.'s of TM processes on a space $T$ is in a one-to-one cor\-res\-pondence with the set of ECFs $\Theta(T)$.
In fact, if $\theta \in \Theta(T)$ and $X$ is an associated TM process, the process $X$ takes a unique role among simple max-stable processes sharing the same ECF $\theta$ in that it provides a sharp lower bound for the \fdd \citep[Corollary 33]{strokorbschlather_15}.
\begin{corollary}[\cite{strokorbschlather_15}, Corollaries 13 and 14] \label{cor:Theta_convex_compact}
The set of ECFs $\Theta(T)$ is convex and compact \wrt the topology of pointwise convergence on $\RR^{\finite(T)}$.
\end{corollary}
\noindent The connection of the TCF $\chi^{(X)}$ to the second-order extremal coefficients of a simple max-stable process $X$ is given by
\begin{align}
\notag \chi^{(X)}(s,t) &= 2- \lim_{\tau \to \infty} \frac{1-\PP\left(X_s \leq \tau, X_t \leq \tau \right)}{1-\PP\left(X_t \leq \tau \right)} \\
\label{eqn:TCFfromECF}
&= 2- \frac{\log \PP\left(X_s \leq \tau, X_t \leq \tau \right)}{\log \PP\left(X_t \leq \tau \right)}
=2-\theta^{(X)}(\{s,t\}).
\end{align}
Therefore, it will be convenient to introduce the following map
\begin{align}\label{eqn:psi}
\psi:\RR^{\finite(T)} \rightarrow \RR^{T \times T}, \qquad \psi(F)(s,t):=2-F(\{s,t\}),
\end{align}
such that (\ref{eqn:TCFfromECF}) reads as $\chi^{(X)}=\psi(\theta^{(X)})$. Note that $\psi$ is continuous if we equip both spaces $\RR^{\finite(T)}$ and $\RR^{T \times T}$ with the topology of pointwise convergence.
Finally, we restate a continuity result from \cite{strokorbschlather_15} in terms of TCFs (instead of ECFs as in the reference).
\begin{corollary}[\cite{strokorbschlather_15}, Theorem 25] \label{cor:TMcty}
$\phantom{a}$\\
Let $X=\{X_t\}_{t \in T}$ be a TM process and $\chi^{(X)}$ its TCF.
Then the following statements are equivalent:
\begin{enumerate}[(i)]
\item $\chi^{(X)}$ is continuous.
\item $\chi^{(X)}$ is continuous on the diagonal $\{(t,t):t \in T\}$.
\item $X$ is stochastically continuous.
\end{enumerate}
\end{corollary}
\begin{remark}
In fact, a TM process $X=\{X_t\}_{t \in T}$ is always stochastically continuous with respect to the semimetric $\eta^{X}(s,t)=1-\chi^{(X)}(s,t)$.
\end{remark}
\section{TCFs are realized by TM processes}\label{sect:realizedTM}
In order to simplify the realization problem for TCFs (termed as Questions (A) to (E) in the introduction) it is desirable to find a subclass of stochastic processes which can realize any given TCF $\chi$. We denote the set of all TCFs and certain subclasses as follows:
\begin{align*}
\tcf(T)&:=\left\{\chi^{(X)} \pmid \begin{array}{l}\text{$X$ a stochastic process on $T$ with identical} \\ \text{one-dimensional margins and existing \ $\chi^{(X)}$} \end{array}\right\},\\
\tcf_\infty(T)&:=\left\{\chi^{(X)} \in \tcf(T) \pmid \text{$X$ with essential supremum $\infty$}\right\},\\
\maxstab(T)&:=\left\{\chi^{(X)} \in \tcf(T) \pmid \text{$X$ simple max-stable}\right\},\\
\tm(T)&:=\left\{\chi^{(X)} \in \tcf(T) \pmid \text{$X$ a TM process}\right\}.
\end{align*}
\begin{remark}
The class $\tcf_\infty(T)$ represents the TCFs of processes whose margins have no jump at the upper endpoint. To see this, first note that a distribution function $F:\RR\rightarrow [0,1]$ has no jump at its upper endpoint $u \in (-\infty,\infty]$ if and only if there exists a continuous strictly increasing transformation $f:(-\infty,u)\rightarrow \RR$ such that $F \circ f^{-1}$ is a distribution function with upper endpoint $\infty$, and secondly, $\chi^{(X)} = \chi^{(f \circ X)}$ if $X$ is a stochastic process with marginal distribution $F$ and TCF $\chi^{(X)}$.
\end{remark}
\noindent A priori it is clear that
\begin{align}\label{eqn:MAXinTCF}
\tm(T) \subset \maxstab(T)\subset \tcf_\infty(T)\subset \tcf(T).
\end{align}
Further, let us introduce the class of uncentered and normalized covariance functions of binary processes
\begin{align}\label{eqn:BIN_defn}
\bin(T)&:=\left\{(s,t) \mapsto \PP(Y_s = 1 | Y_t = 1) \pmid \begin{array}{l}\text{$Y$ a stochastic process on $T$ with} \\ \text{identical one-dimensional margins}\\ \text{with values in $\{0,1\}$ and $\EE Y_t \neq 0$} \end{array}\right\},
\end{align}
which is closely related to the above classes. By definition of $\tcf(T)$ and considering the processes $Y_t = \Eins_{X_t > \tau}$ indexed by $\tau>0$, we observe
\begin{align}\label{eqn:TCFinSCBIN}
\tcf(T)\, \subset \text{ sequential closure of \,} \bin(T),
\end{align}
where the sequential closure is meant \wrt pointwise convergence. The following theorem gives an affirmative answer to the question whether $\tcf(T)$ and $\maxstab(T)$ coincide \emph{(Question (D) in the Introduction)} and yields also the connection to the other classes. In fact, the class of TM processes can realize already any given TCF.
\begin{theorem}\label{thm:TCFisMAX}
\begin{enumerate}[a)]
\item For arbitrary sets $T$ the following classes coincide
\begin{align}
\bin(T)&=\psi(\Theta_b(T)),\\[1mm]
\notag \tcf(T)&=\tcf_\infty(T)=\maxstab(T)=\tm(T)=\psi(\Theta(T))\\
\label{eqn:TCFisMAX} &=\text{ sequential closure of \,} \bin(T) = \text{ closure of \,} \bin(T),
\end{align}
where the map $\psi$ is from (\ref{eqn:psi}), $\Theta(T)$ and $\Theta_b(T)$ are from (\ref{eqn:Theta}) and (\ref{eqn:Theta_bounded}), respectively, and
the (sequential) closure is meant \wrt pointwise convergence.
\item For infinite sets $T$ the inclusion $\bin(T) \subsetneq \tcf(T)$ is proper.
\item For finite sets $M$ the equality $\bin(M)=\tcf(M)$ holds.
\end{enumerate}
\end{theorem}
\begin{proof}
\begin{enumerate}[a)]
\item
First, we establish $\bin(T)=\psi(\Theta_b(T))$:
Let $f \in \bin(T)$ and let $Y$ be a corresponding process with values in $\{0,1\}$ as in the definition of $\bin(T)$ (\cf (\ref{eqn:BIN_defn})).
Let the function $C^{(Y)}:\finite(T) \rightarrow [0,1]$ be given by $C^{(Y)}(\emptyset) = 0$ and $C^{(Y)}(A) = \PP(\exists \, t \in A \text{ such that } Y_t = 1)$
as in Example~\ref{example:cap_CA}. Then $C(\{t\})=\EE Y_t$ lies in the interval $(0,1]$ and is independent of $t \in T$ due to identical one-dimensional margins.
Further, the function $f$ is given by $f(s,t)=\PP(Y_s = 1 \mid Y_t = 1) = 2-C(\{s,t\})/C(\{t\})$. Now, set $\theta(A):=C(A)/C(\{t\})$ for $A \in \finite(T)$. Then $\theta$ satisfies $\psi(\theta)(s,t)=2-\theta(\{s,t\})=f(s,t)$ and $\theta$ is clearly bounded by $1/C(\{t\})$. It follows from Example~\ref{example:cap_CA} and Theorem~\ref{thm:ECF_CA} that $\theta$ lies in $\Theta(T)$. Hence, $f \in \psi(\Theta_b(T))$.
Conversely, let $\theta \in \Theta_b(T)$ be bounded, say by $\kappa$. Clearly, $\kappa \geq \theta(\{t\})=1$. Set $C(A):=\theta(A)/\kappa$. Then $C$ satisfies all requirements of Example~\ref{example:cap_CA} to determine the f.d.d.\ of a binary process $Y$ with values in $\{0,1\}$ with $C^{(Y)}=C$. The process $Y$ has identical one-dimensional margins since $\theta(\{t\})=1$ for $t \in T$, and $\EE Y_t = 1/\kappa > 0$. So $Y$ fulfills the requirements of a process in the definition of $\bin(T)$. Finally, note that the corresponding function in $\bin(T)$ is given by $\PP(Y_s = 1 \mid Y_t = 1)=2-C(\{s,t\})/C(\{t\})=\psi(\theta)(s,t)$ as desired.
Secondly, the equality $\maxstab(T)=\tm(T)=\psi(\Theta(T))$ follows directly from Theorem~\ref{thm:ECF_CA}. On the one hand this implies
\begin{align*}
\bin(T) = \psi(\Theta_b(T)) \subset \psi(\Theta(T))=\tm(T),
\end{align*}
and on the other hand, we obtain that $\tm(T)$ is compact, as it is the image of the compact set $\Theta(T)$ (Corollary~\ref{cor:Theta_convex_compact}) under the continuous map $\psi$. Now, the assertion (\ref{eqn:TCFisMAX}) follows from
\begin{align*}
& \tcf(T) \stackrel{(\ref{eqn:TCFinSCBIN})}{\subset} \text{ sequential closure of \,} \bin(T) \subset \text{ closure of \,} \bin(T) \\
& {\subset} \text{ closure of \,} \tm(T) \subset \, \tm(T) \stackrel{(\ref{eqn:MAXinTCF})}{\subset} \, \maxstab(T)
\stackrel{(\ref{eqn:MAXinTCF})}{\subset} \, \tcf_\infty(T)
\stackrel{(\ref{eqn:MAXinTCF})}{\subset} \, \tcf(T).
\end{align*}
\item Let $T$ be an infinite set and let $\chi(s,t):=\delta_{st}$. Indeed $\chi$ is an element of $\maxstab(T)$ realized by the simple max-stable process $X$ on $T$, where the variables $\{X_{t}\}_{t \in T}$ are \iid standard Fr{\'e}chet random variables. Suppose that $\chi \in \bin(T)$. Then $\PP(Y_s=1,Y_t=1)=0$ for all $s,t \in T$ with $s \neq t$. Thus, $\PP(\bigcup_{s \in S} \{Y_s=1\})=\sum_{s \in S} \PP(Y_s=1)=\infty$ for any countably infinite subset $S \subset T$, a contradiction.
\item If $M$ is finite, elements of $\Theta(M)$ are automatically bounded by $|M|$ and thus, $\Theta(M)=\Theta_b(M)$.
\end{enumerate}
\end{proof}
The latter result does not include any regularity considerations beyond the product topology that is somewhat unnatural in infinite-dimensional stochastic contexts. However, in view of Corollary~\ref{cor:TMcty}, it is possible to identify the role of continuous TCFs in this realization problem and hence address {Question (C)} as follows.
\begin{corollary}\label{cor:chicty}
Let $\chi \in \tcf(T)$. Then the following statements are equivalent.
\begin{enumerate}[(i)]
\item $\chi$ is continuous.
\item $\chi$ is continuous on the diagonal $\{(t,t) \,:\, t \in T\}$.
\item There exists a stochastically continuous stochastic process $\{X_t\}_{t \in T}$\\ with TCF $\chi^{(X)}=\chi$.
\end{enumerate}
\end{corollary}
\begin{remark}\label{rk:chicty}
In fact, any TM process $X$ with continuous TCF $\chi^{(X)}$ is stochastically continuous. It follows from \citeauthor{dehaan_84}'s (\citeyear{dehaan_84}) construction that any simple max-stable process on $\RR^d$ (or any other locally compact second countable Hausdorff space) that is continuous in probability, can be realized on a sufficiently regular probability space. Hence, this applies to TM processes with continuous TCFs, since they are simple max-stable and continuous in probability by the preceding corollary.
\end{remark}
\begin{remark}
\cite{lm15} discuss regularity conditions on the two-point covering function of a random set, or equivalently, a unit covariance function (cf.\ Section \ref{sect:corcut}) that ensure the existence of a realizing closed set, or equivalently, a realizing $\{0,1\}$-valued process with upper semi-continuous paths. Here, we do not know which regularity conditions on the TCF ensure the existence of a realizing upper semi-continuous process.
\end{remark}
\section{Basic closure properties and characterization by inequalities}\label{sect:props}
Finally, we collect some immediate and important consequences concerning operations on the set of TCFs and the characterization of the set of TCFs by means of finite-dimensional projections.
Even though not all non-negative correlation functions are TCFs, both classes have some desirable properties in common as we shall see next.
Well-known operations on (non-negative) correlation functions include convex combinations, products and pointwise limits.
Interestingly, the same operations are still admissible for TCFs (answering Question E).
\begin{corollary}\label{cor:basicoperations}
The set of tail correlation functions $\tcf(T)$ is convex, closed under pointwise multiplication and compact \wrt pointwise convergence.
\end{corollary}
\begin{proof} These closure properties follow from Theorem~\ref{thm:TCFisMAX}. Convexity and compactness of $\tcf(T)=\psi(\Theta(T))$ are immediate taking additionally Corollary~\ref{cor:Theta_convex_compact} into account. Moreover, let $\chi_1$ and $\chi_2$ be in $\tcf(T)=\tcf_{\infty}(T)$ with corresponding processes $X^{(1)}$ and $X^{(2)}$ with upper endpoint $\tau_{\text{up}}=\infty$. We choose them to be independent and set $X^{(3)}:=X^{(1)} \wedge X^{(2)}$, which then also has upper endpoint $\infty$ and satisfies
\begin{align*}
\PP(X^{(3)}_s \geq x \,|\, X^{(3)}_t \geq x) = \PP(X^{(1)}_s \geq x \,|\, X^{(1)}_t \geq x) \cdot \PP(X^{(2)}_s \geq x \,|\, X^{(2)}_t \geq x).
\end{align*}
Consequently, the TCF $\chi_3$ of $X^{(3)}$ is the product $\chi_3=\chi_1 \cdot \chi_2$.
\end{proof}
\noindent Secondly, the set of TCFs can be characterized through finite-dimensional projections.
\begin{corollary}\label{cor:TCFextension}
A real-valued function $\chi: T\times T \rightarrow \RR$ belongs to $\tcf(T)$ if and only if the restriction $\left.\chi\right|_{M \times M}$ belongs to $\tcf(M)$ for all non-empty finite subsets $M$ of $T$.
\end{corollary}
\begin{proof}
If $\chi \in \tcf(T)$, then necessarily $\chi|_{S\times S}\in \tcf(S)$ for any subset $S \in T$. To show the reverse implication, let $\chi|_{M\times M} \in \tcf(M)$ for all $M \in \mathcal{F}(T)\setminus \{\emptyset\}$. Since $\tcf(T) \subset [0,1]^{T\times T}$ is closed, to prove $\chi \in \tcf(T)$ it suffices to show that $U \cap \tcf(T) \neq \emptyset$ for any open neighborhood $U$ of $\chi$ in $[0,1]^{T\times T}$. Given $U$, there is a finite subset of $T\times T$, which we may assume to be of the form $M \times M$, and open sets $A_{(i,j)}\subset [0,1]$, $(i,j) \in M\times M$, such that $\chi \in \bigcap_{(i,j) \in M \times M} \pr_{(i,j)}^{-1}\left(A_{(i,j)}\right) \subset U$ (where $\pr_{(s,t)}:[0,1]^{T \times T} \rightarrow [0,1]$ denotes the natural projection). Since $\chi|_{M\times M}$ trivially extends to an element $\tilde\chi \in \tcf(T)$ (e.g.\ copy one of the random variables), we have $\tilde\chi \in U \cap \tcf(T) \neq \emptyset$.
\end{proof}
In Part II of this exposition, we will see that for a finite set $M$, the set of TCFs $\tcf(M)$ constitutes a convex polytope in $\RR^{\lvert M \rvert \times \lvert M \rvert}$ that can be described by means of a finite system of (affine) inequalities.
In this regard Corollary~\ref{cor:TCFextension} shows that for an arbitrary set $T$, the class $\tcf(T)$ may also be completely characterized by a system of (affine) inequalities. This is not evident since elements of $\tcf(T)$ are defined a priori through a limiting procedure.
\vspace{7mm}
{\noindent \textbf{\large Part II\\[2mm]
The realization problem for TCFs on finite sets}}\\[3mm]
\noindent In view of Corollary~\ref{cor:TCFextension} it suffices to study $\tcf(M)$ for finite sets $M$ if one is interested in a complete characterization of the space $\tcf(T)$ for arbitrary $T$. Therefore, we focus on a non-empty finite set $M=\{1,\dots,n\}$ in this section and set
\begin{align*}
\tcf_n:=\tcf(\{1,\dots,n\}).
\end{align*}
To begin with, we show that $\tcf_n$ can be viewed as a convex polytope in Section~\ref{sect:TCFpolytope}. Its geometry will be studied subsequently. Here, we start off with some basic observations and low-dimensional results in Section~\ref{sect:basicfacts}.
Section~\ref{sect:generalresults} collects more sophisticated results on $\tcf_n$ with deeper insights into the rapidly growing complexity of $\tcf_n$ as $n$ grows, including connections between $\tcf_n$ and $\tcf_{n'}$ for $n' > n$.
Thereby, some
obervations from Section~\ref{sect:basicfacts} will be uncovered as low-dimensional phenomena.
At least, it is possible to identify the precise relation of $\tcf_n$ to the so-called cut- and correlation-polytopes as well as to the polytope of unit covariances.
To complement these general observations, Section~\ref{sect:compresults} reports all results relying on software computations and, in particular, all combinatorial considerations that were necessary in order to push the entire description of the vertices and facets of $\tcf_n$ up to $n \leq 6$.
Finally, we pursue some open questions on the geometry on $\tcf_n$ in Section~\ref{sect:openquestions}.
\section{TCF$_n$ is a convex polytope}\label{sect:TCFpolytope}
Elements of $\tcf_n$ are functions on $\{1,\dots,n\} \times \{1,\dots,n\}$, that is to say, they are $n\times n$ matrices. Since TCFs are symmetric and take the value $1$ on the diagonal, we may regard $\tcf_n$ for $n \geq 2$ as a subset of
\begin{align*}
\RR^{E_n} \cong \RR^{n \choose 2}=\RR^{n(n-1)/2},
\end{align*}
where $E_n$ is the set of edges of the \emph{complete graph} $K_n$ with vertices $V_n=\{1,\dots,n\}$. It will be convenient to interpret elements of $\tcf_n$ as an edge labelling of $K_n$, which is why we call $K_n$ the \emph{support graph} for $\tcf_n$. Due to Theorem~\ref{thm:TCFisMAX} and (\ref{eqn:BIN_defn}) we know already
\begin{align}\label{eqn:TCFisBIN}
\tcf_n=\bin_n:=\left\{\chi \in \RR^{E_n} \pmid
\begin{array}{l}
\chi_{ij} = {\EE(Y_i Y_j)}/{ \EE Y_j} \text{ where } \\
\text{$Y_1,\dots,Y_n$ take values in $\{0,1\}$} \\
\text{and $\EE Y_1 = \ldots = \EE Y_n > 0$}
\end{array}
\right\}.
\end{align}
The following lemma is a reformulation of this fact and will be useful later on.
\begin{lemma}\label{lemma:TCFevents}
An element $\chi \in \RR^{E_n}$ belongs to $\tcf_n$ if and only if it can be written as
\begin{align*}
\chi_{ij}=\PP(A_i|A_j), \quad 1 \leq i<j \leq n
\end{align*}
for some (finite) probability space $(\Omega,\A,\PP)$ and measurable subsets $A_1,\dots,A_n \in \A$ which satisfy $\PP(A_1)=\dots=\PP(A_n)>0$.
\end{lemma}
\begin{remark}\label{remark:TCFevents}
In Lemma~\ref{lemma:TCFevents} we may assume that $\PP(A_1)=\dots=\PP(A_n)=c$ for any constant $0 < c \leq 1/n$: Otherwise enlarge $\Omega$, such that $A :=\bigcup_{i=1}^n A_{i} \neq \Omega$. On $A$ define the measure $\QQ :=c/\PP(A_{1})\cdot \PP|_{A}$. Then $\QQ(A) \leq 1$ and, thus, $\QQ$ extends to a probability measure on $\Omega$ with $\QQ(A_{i}) = c$ and $\QQ(A_{i} | A_{j})=\chi_{ij}$.
\end{remark}
\noindent Likewise, we set
\begin{align*}
\Theta_n:=\Theta(\{1,\dots,n\})
\end{align*}
and, since $\theta_\emptyset = 0$ and $\theta_i = 1$ for $i=1,\dots,n$, we may regard $\Theta_n$ for $n \geq 2$ as a subset of
\begin{align*}
\RR^{\finite^{(2)}_n} \cong \RR^{2^n-n-1},
\end{align*}
where $\finite^{(2)}_n$ is the set of subsets of $V_n$ with at least two elements. Remember from (\ref{eqn:psi}) that
\begin{align}\label{eqn:psin}
\tcf_n = \psi_n(\Theta_n) \qquad \text{where} \qquad \psi_n: \RR^{\finite^{(2)}_n} \rightarrow \RR^{E_n}, \quad \psi_n(\theta)_{ij} = 2 - \theta_{ij},
\end{align}
and note that $\psi_n=2-\pr_{E_n}$ is essentially a projection onto the $n \choose 2$ coordinates of $\RR^{E_n}$. Before we proceed, we need to revise some notation for convex polytopes.
\paragraph{\textbf{\upshape Notation and facts concerning convex polytopes}}
(\cf \cite{ziegler_95}).\\
A subset $P \subset \RR^p$ is a \emph{convex polytope} if $P$ is bounded and can be represented as $P=\{x \in \RR^p \,:\, Cx \leq c\}$ for a $q \times p$ matrix $C$ and a $q$-vector $c$ for some $q \in \NN$ (where $\leq$ is meant componentwise). The rows of $C$ and $c$ represent hyperplanes in $\RR^d$ and the inequality $\leq$ determines the corresponding halfspace to which $P$ belongs. The system $Cx \leq c$ will be called an \emph{$\mathcal{H}$-representation} (or \emph{halfspace representation}) of $P$.
An $\mathcal{H}$-representation will be called a \emph{facet representation} if it is minimal in the sense that none of the rows in $C$ and $c$ can be deleted in order to define $P$, \ie $P \neq \{x \in \RR^p \,:\, C_{-i} x \leq c_{-i}\}$ for all $i=1,\dots,q$, where $C_{-i}$ and $c_{-i}$ are the modified versions of $C$ and $c$ with the $i$-th row removed. In fact, an $\mathcal{H}$-representation $Cx\leq c$ is a {facet representation} if every row of $C$ and $c$ yields in fact a \emph{facet inducing} inequality of $P$, where an inequality $C_{i} x \leq c_{i}$ is facet inducing if $\dim(P \cap \{x \in \RR^p \,:\, C_{i} x = c_{i}\}) = \dim(P) - 1$. The latter is equivalent to the existence of $\dim(P)$ affinely independent points $x^1,\dots,x^{\dim(P)} \in P$ solving the equation $C_{i} x = c_{i}$.
By a slight abuse of notation, we will usually refer to the inequality $C_{i} x \leq c_{i}$ as a \emph{facet} of $P$ if it induces a facet (instead of calling the set $P \cap \{x \in \RR^p \,:\, C_{i} x = c_{i}\}$ a facet).
Equivalently, a subset $P \subset \RR^p$ is a convex polytope if $P$ equals the convex hull of a finite subset $S \subset \RR^p$. Then $S$ will be called a \emph{$\mathcal{V}$-representation} of $P$. A minimal $\mathcal{V}$-representation, with respect to set inclusion, will be called a \emph{vertex representation}. In fact, the vertex representation is unique and given by the set $\Ex(P)$ of extremal points, or \emph{vertices}, of $P$, \ie the points of $P$ that cannot be decomposed non-trivially as a convex combination of two other points of $P$. Note that in general a $\mathcal{V}$-representation of $P$ may consist of more points than the vertex set $\Ex(P)$.
Moreover, if $P \subset \RR^p$ is a convex polytope and $\pi:\RR^p \rightarrow \RR^{p'}$ is an affine map $x \mapsto Ax+b$, then the image $\pi(P)$ is again a convex polytope and secondly, any intersection of $P$ with an affine subspace of $\RR^p$ is a convex polytope.
\begin{corollary}\label{cor:finitepolytope}
For all $n \in \NN$ the sets $\Theta_n$ and $\tcf_n$ are convex polytopes.
\end{corollary}
\begin{proof}
For $\Theta_n$ this property is evident from Theorem~\ref{thm:ECF_CA} and (\ref{eqn:finite_CA}).
But then the affine map $\psi_n$ maps $\Theta_n$ to the convex polytope $\tcf_n=\psi_n(\Theta_n)$.
\end{proof}
\noindent Now, that we know that $\tcf_n$ is a convex polytope, we seek to understand its geometric structure. At best, we would like to determine its vertex and facet representation (and we will indeed do so in Section~\ref{sect:compresults} up to $n \leq 6$).
To repeat the terminology adopted from convex geometry in this context, note that an $\mathcal{H}$-representation of $\tcf_n$ (and in particular, a facet representation) allows one to check whether a given matrix is indeed a TCF, since any $\mathcal{H}$-representation of $\tcf_n$ constitutes a set of necessary and sufficient conditions for being a TCF. In a facet representation no condition is obsolete.
Complementary, a $\mathcal{V}$-representation (and in particular, a vertex representation) of $\tcf_n$ is more useful if one wants to generate valid TCFs. Any TCF can be obtained as a convex combination of the elements of a $\mathcal{V}$-representation.
In a vertex representation no point is obsolete.
\section{Basic observations and low-dimensional results for TCF$_{n}$}\label{sect:basicfacts}
This section comprises two first general observations. First, every polytope $\tcf_n$ satisfies a certain system of inequalities (to be called \emph{hypermetric inequalities}) and, second, we identify its $\{0,1\}$-valued vertices as so-called \emph{clique partition points}. With regard to the explicit vertex and facet structure of $\tcf_n$ in low dimensions, both findings might lead to tempting conjectures on the geometry of $\tcf_n$ eventually refuted by the more sophisticated methods applied in Section~\ref{sect:generalresults}.
\paragraph{\textbf{\upshape Hypermetric inequalities}}
Remember that we identified the set of all TCFs on $V_n=\{1,\dots,n\}$ with a subset of $\RR^{E_n} = \RR^{\binom{n}{2}}$ while it originally was interpreted as a set of symmetric $n \times n$ matrices with 1's on the diagonal. In the sequel we will identify points $x=(x_{ij})_{1\leq i < j\leq n} \in \RR^{E_n}$ with $n \times n$ matrices $(x_{ij})_{1\leq i,j \leq n}$ via $x_{ji} =x_{ij}$ and $x_{ii} := 1$.
Let $b =(b_{1},\ldots,b_{n}) \in \ZZ^n$. The point $(x_{ij})_{1\leq i < j\leq n}\in \RR^{E_n}$ satisfies the \emph{hypermetric inequality defined by} $b$ if
\begin{align}
\notag \sum_{1\leq i,j\leq n}b_{i}b_{j}x_{ij} &\geq \sum_{i=1}^{n}b_{i}\\
\text{or, equivalently,} \qquad
\label{eq:hypermetric} \sum_{1\leq i < j \leq n} (-b_{i}b_{j}) x_{ij} &\leq \frac{1}{2}\sum_{i=1}^n b_{i}(b_{i} -1).
\end{align}
\begin{remark}
In \cite{dezalaurent_97}
the inequalities $\sum_{1\leq i<j\leq n}b_{i}b_{j}x_{ij} \leq 0$ with $\sum_{1\leq i \leq n}b_{i} = 1$ are termed hypermetric. All these inequalities are valid for the cut polytope $\cut_n$ to be introduced here in Section~\ref{sect:corcut} \cite[Lemma 28.1.3]{dezalaurent_97}. For $\tcf_n$ the variant \eqref{eq:hypermetric} is an appropriate ``counterpart''.
\end{remark}
\begin{lemma}\label{lemma:hypervalid} All hypermetric inequalities (in the sense of \eqref{eq:hypermetric}) are valid for elements of $\tcf_{n}$.
\end{lemma}
\begin{proof}
Let $Y_{1},\ldots,Y_{n}$ be a $\{0,1\}$-valued stochastic model for $\chi \in \tcf_{n}$. Set $a:=\EE(Y_1)>0$. Then for $b \in \ZZ^n$
\begin{align*}
\sum_{1\leq i,j\leq n}\hspace{-1mm} b_{i}b_{j}\chi_{ij} =
\hspace{-1mm} \sum\limits_{1 \leq i,j\leq n} \hspace{-1mm} b_{i}b_{j}\frac{\EE(Y_{i}Y_{j})}{a} = \frac{1}{a} \EE\bigg[\sum\limits_{i=1}^n b_{i}Y_{i}\bigg]^2 \geq \frac{1}{a} \EE\bigg[\sum\limits_{i=1}^n b_{i}Y_{i}\bigg] = \sum\limits_{i=1}^n b_{i}, \qquad
\end{align*}
as for any integer $k$ we have $k^2 \geq k$.
\end{proof}
\paragraph{\textbf{\upshape Clique partition polytopes}}
A subset $\{C_1,\dots,C_k\}$ of the powerset of $V_n=\{1,\dots,n\}$ is a \emph{partition} of $V_n$ if $k \geq 1$, $C_r\cap C_s = \emptyset$ for $r \neq s$ and $\bigcup_{r=1}^k C_r = V_n$. A partition of $V_n$ defines a \emph{clique partition point} $\gamma(\{C_1,\dots,C_k\}) \in \{0,1\}^{E_n}$ by
\begin{align*}
\gamma(\{C_1,\dots,C_k\})_{ij} = \sum_{r=1}^{k} \Eins_{\{i,j\} \subset C_r}, \quad 1 \leq i < j \leq n.
\end{align*}
The \emph{clique partition polytope} is defined as the convex hull of the clique partition points \citep{groetschelwakabayashi_90} in $\RR^{E_n}$
\begin{align*}
\cpp_n := \conv\left(\{ \gamma(\{C_1,\dots,C_k\}) \,:\, \{C_1,\dots,C_k\} \text{ partition of $V_n$} \}\right).
\end{align*}
Being $\{0,1\}$-valued, the clique partition points are automatically the extremal points of their convex hull:
\begin{align*}
\Ex\left(\cpp_n\right)=\left(\{ \gamma(\{C_1,\dots,C_k\}) \,:\, \{C_1,\dots,C_k\} \text{ partition of $V_n$} \}\right).
\end{align*}
It turns out that all $\{0,1\}$-valued vertices of $\tcf_n$ are precisely the clique partition points.
\begin{proposition}\label{prop:cpp}
$\tcf_n \cap \{0,1\}^{E_n} = \Ex(\cpp_n)$ for all $n \in \NN$. In particular $\cpp_n \subset \tcf_n$.
\end{proposition}
\begin{proof}
Since $\tcf_{n} \cap \{0,1\}^{E_n} \subset \Ex(\tcf_{n})$ it suffices to show the first statement. For $n=2$ we have $\tcf_{2} = [0,1]$ and $\{0,1\} = \Ex(\cpp_{2})$. For $n\geq3$ the points in $\tcf_{n}$ have to satisfy the triangle-inequalities (all permutations of
$\chi_{1,2}+\chi_{2,3}-\chi_{1,3} \leq 1$,
see (\ref{eqn:triangle}) and also \eqref{eq:hypermetric} with $b=(1,-1,1,0,\ldots,0)$). For points $\chi \in \tcf_{n} \cap \{0,1\}^{E_n}$, viewed via the support graph $K_n$, this implies for any triple of nodes $i,j,k$, where the edges $\{i,j\}$ and $\{j,k\}$ have value 1, that also the edge $\{i,k\}$ has value 1. Thus, a simple inductive argument shows: for any pair of nodes $i,j$, which are connected by a path of edges with value 1, the edge from $i$ to $j$ has also value 1. This shows that the points in $\tcf_{n} \cap \{0,1\}^{E_n}$ are clique partition points.
In order to see that any clique partition point $\gamma(\{C_1,\dots,C_k\})$ belongs to $\tcf_{n} \cap \{0,1\}^{E_n}$ choose $\Omega = \{1,\ldots,k\}$ with uniform distribution $\PP$ and $A_{i} = \{r_{i}\}$, $1 \leq i \leq n$, with $r_{i}$ uniquely determined by $i \in C_{r_{i}}$ and apply Lemma~\ref{lemma:TCFevents}.
\end{proof}
For $n \leq 4$ the clique partition polytope and $\tcf_n$ even coincide.
\begin{proposition}\label{prop:cpplowdim}
$\tcf_n = \cpp_n$ for $n \leq 4$.
\end{proposition}
\begin{proof}
For $n\leq 4$ we computed explicitly that $\Ex(\tcf_n)=\Ex(\cpp_n)$ from the characterization (\ref{eqn:psin}) \citep[Tables~3.1 and 3.3]{strokorb_13} and confirmed this result using the software \texttt{polymake}. This implies $\tcf_n = \cpp_{n}$ for $n \leq 4$.
\end{proof}
\paragraph{\textbf{\upshape Low-dimensional phenomena}}
Even though for $n \leq 4$ the polytope $\tcf_n$ and the clique partition polytope $\cpp_n$ coincide, the property $\tcf_n = \cpp_n$ will turn out to be a low-dimensional phenomenon. Starting from $n=5$ the vertices of $\tcf_n$ are not $\{0,1\}$-valued anymore (see Corollary \ref{cor:nocpp} in Section~\ref{sect:generalresults}), in particular $\cpp_n \subsetneq \tcf_n$ for $n \geq 5$. Still, up to $n \leq 5$ all facet inducing inequalities of $\tcf_n$ turn out to be hypermetric and one might be tempted to believe that certain hypermetric inequalities provide an $\mathcal{H}$-representation for $\tcf_n$ also in higher dimensions. Again, this property constitutes only another low-dimensional phenomenon. Starting from $n=6$ not all facets of $\tcf_n$ are hypermetric anymore (see Proposition \ref{prop:nonhypfacets} in Section~\ref{sect:generalresults}).
\section{Sophisticated results on the geometry of $\tcf_n$}\label{sect:generalresults}
A fundamental observations in this section concerns the lifting of vertices and facets to higher dimensions (Section~\ref{sect:lifting}). It means that vertices (and facets) of $\tcf_n$ will also appear as vertices (and facets) of $\tcf_{n'}$ for $n'>n$ if the coordinates (or coefficients) are filled up with zeros at appropriate places. Note that both statemenents are not evident, but a deep structural result only revealed by some delicate combinatorial arguments. Subsequently, we prove that every rational number in the interval $[0,1]$ will appear as coordinate value in the vertex set of $\tcf_n$ starting from a sufficiently large $n$ (Proposition~\ref{prop:unbounded} in Section~\ref{sect:unbounded}) and that $\tcf_n$ possesses non-hypermetric facets starting from $n \geq 6$ (Proposition~\ref{prop:nonhypfacets} in Section~\ref{sect:nonhypfacets}).
Taken together, these results give insights into the rapidly growing complexity of $\tcf_n$ as $n$ grows and confound the aim of a full description of vertices and facets of $\tcf_n$ for arbitrary $n$.
Finally, Section~\ref{sect:corcut} provides an alternative (``dual'') description of the polytope $\tcf_n$ (which we recognized already as the \emph{projection} of the polytope $\Theta_n$) as an \emph{intersection} with the so-called correlation polytope or, equivalently, with the so-called cut-polytope.
\subsection{\textbf{\upshape Lifting of vertices and facets to higher dimensions}}\label{sect:lifting}
First, we deal with connections between $\tcf_{n}$ and $\tcf_{n+1}$.
A particularly important feature is the lifting property. That is every vertex of $\tcf_n$ will appear again in the list of vertices of $\tcf_{n+1}$ with some zeros added.
\begin{lemma}[Projections and liftings of points and vertices]\label{lemma:zeroliftingvertices}
$\phantom{a}$\\
For $\chi \in \tcf_{n+1}$ let $\chi|_{K_{n}}$ denote the restriction of $\chi$ to the subgraph $K_{n} \subset K_{n+1}$ (delete all $\chi_{i,n+1}$, $1 \leq i \leq n$). Conversely, let $\chi^0 \in \RR^{E_{n+1}}$ denote the extension of a point $\chi \in \tcf_{n}$ by
\begin{align*}
\chi^0_{i,n+1} = 0, \quad 1 \leq i \leq n.
\end{align*}
\begin{enumerate}[a)]
\item The assignment $\chi \mapsto \chi|_{K_n}$ maps $\tcf_{n+1}$ onto $\tcf_{n}$.
\item The assignment $\chi \mapsto \chi^0$ embeds $\tcf_{n}$ into $\tcf_{n+1}$ and $\Ex(\tcf_{n})$ into $\Ex(\tcf_{n+1})$.
\item If $ \chi \in \Ex(\tcf_{n+1})$ and $\chi_{i,n+1}=0$ for all $1 \leq i \leq n$, then $\chi|_{K_{n}} \in \Ex(\tcf_{n})$.
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}[a)]
\item Let $Y_{1},\ldots, Y_{n+1}$ be a binary process that models $\chi$. Simply deleting $Y_{n+1}$ gives a model for $\chi|_{K_{n}} \in \tcf_{n}$. Surjectivity follows from b).
\item Let $Y_{1},\ldots, Y_{n}$ be a binary process that models $\chi$. Let $a = \EE(Y_{1})$. Add a disjoint point $\omega_{0}$ to the underlying probability space $\Omega$ and replace the probability measure $\PP$ by $\frac{1}{1+a}\cdot \PP|_{\Omega} + \frac{a}{1+a}\cdot \delta_{\omega_{0}}$. Extend $Y_{1},\ldots,Y_{n}$ by 0 on $\omega_{0}$, let $Y_{n+1} = \eins_{\{\omega_{0}\}}$. Now, $Y_{1},\ldots,Y_{n+1}$ is a model for $\chi^0$, since $Y_{i}Y_{n+1} = 0$, $1 \leq i \leq n$.
If $\chi^0 \notin \Ex(\tcf_{n+1})$, there is a representation $\chi^0 = \lambda y + (1-\lambda)z$, with $y,z\in \tcf_{n+1}$, $0 < \lambda<1, y \neq z$.
Since $\chi^0$ is zero on the new edges, the points $y,z$ also have to be zero on the new edges, so $y|_{K_{n}} \neq z|_{K_{n}}$ and $y|_{K_{n}},z|_{K_{n}} \in \tcf_{n}$ by a). Thus, $\chi = \chi^0|_{K_{n}} \notin \Ex(\tcf_{n})$.
\item If $\chi|_{K_{n}} \notin \Ex(\tcf_{n})$, then $\chi|_{K_{n}} = \lambda y + (1-\lambda)z$, with $y,z\in \tcf_{n}$, $0 < \lambda<1, y \neq z$. By b) we know $y^0, z^0 \in \tcf_{n+1}$.
Since $\chi_{i,n+1}=0$ for all $1 \leq i \leq n$, we have $\chi = (\chi|_{K_{n}})^{0} = \lambda y^0 + (1-\lambda)z^0 \notin \Ex(\tcf_{n+1})$.
\end{enumerate}
\end{proof}
\noindent We call $\chi^0$ a \emph{lifting} of $\chi$. The following lemma generalizes the lifting of vertices and will be applied to deduce Proposition~\ref{prop:unbounded}.
\begin{lemma}[Lifting of vertices arising from partitions] \label{lemma:unionsofvertices}
$\phantom{a}$\\
Let $C_{1},\ldots,C_{k}\subset V_n$ be disjoint subsets of the vertex set $V_n=\{1,\ldots,n\}$ each containing at least two elements of $V_n$.
For $1\leq r \leq k$ let $\chi^{r} \in \Ex(\tcf(C_r))$. Similarly to the interpretation of TCFs on $V_n=\{1,\dots,n\}$ as elements of $\RR^{E_n}$, we interpret $\chi^r$ as an element of $\RR^{E(C_r)}$, where $E(C_r)$ is the set of edges of the complete graph with vertex set $C_r \subset V_n$.
Define $\chi \in \RR^{E_{n}}$ by
\begin{align*}
\chi_{ij}
= \left\{
\begin{array}{ll}
\chi^r_{ij} & \quad \text{\normalfont if } \{i,j\} \subset C_r \text{ \normalfont for some } 1 \leq r \leq k, \\
0 & \quad \text{\normalfont else}.
\end{array}
\right.
\end{align*}
Then $\chi \in \Ex(\tcf_{n})$.
\end{lemma}
\begin{proof}
Because of the lifting property (Lemma~\ref{lemma:zeroliftingvertices}), it suffices to consider the case $V_n=\bigcup_{r=1}^k C_r$, where $C_r=\{i^{(r)}_1,\dots,i^{(r)}_{|C_r|}\}$.
First, we show that $\chi \in \tcf_{n}$.
To this end, choose (finite) set models
\begin{align*}
(\Omega_{r}, \PP_{r}), \quad A^r_{i^{(r)}_1},\ldots,A^r_{i^{(r)}_{|C_{r}|}} \subset \Omega_{r} , \quad 1 \leq r \leq k
\end{align*}
for $\chi^{r}$ as in Lemma~\ref{lemma:TCFevents} such that $\chi^{(r)}_{ij} = \PP(A^r_{i} \,|\, A^r_{j})$.
By Remark~\ref{remark:TCFevents} these models can be chosen such that $\PP_{r}(A^r_{i})$ does not depend on $r$. Then a stochastic model for $\chi$ is obtained through the normalized disjoint union of these models, \ie where $\Omega=\bigcup_{r=1}^k \Omega_r$,
$\PP=\frac{1}{k}\sum_{r=1}^k \PP_r(\cdot \cap \Omega_r)$ and $A_i=A^r_i$ if $i \in C_r$. (Note that for each $i\in V_n$ there exists a unique $r$ with $i \in C_r$, since the sets $C_r$ are disjoint and cover $V_n$.)
Now, we show that $\chi \in \Ex(\tcf_{n})$. Suppose not. Then $\chi = \lambda y + (1-\lambda)z$ with $1 < \lambda <0$ and $ y,z \in \tcf_{n}$ with $y \neq z$. Necessarily $y_{ij} = 0$ and $z_{ij}= 0$ whenever $\chi_{ij}=0$. Thus, $y|_{K_n^{r}} \neq z|_{K_n^{r}}$ for some $1 \leq r \leq k$ when $K^r_n$ denotes the complete subgraph of $K_n$ defined by $C_r$.
Since $y|_{K_n^{r}}, z|_{K_n^{r}} \in \tcf(C_r)$ by Lemma~\ref{lemma:zeroliftingvertices}, we obtain $\chi^r=\chi|_{K_n^{r}} = \lambda y|_{K_n^{r}} + (1-\lambda)z|_{K_n^{r}}$ contradicting $\chi^{r} \in \Ex(\tcf(C_r))$.
\end{proof}
In order to deduce the lifting property also for inequalities and facets, we adapt ideas from \cite[Lemma~26.5.2]{dezalaurent_97}.
We show that, starting from $n = 3$, no facet inducing inequality will ever become obsolete as $n$ grows. For instance, the triangle inequality $(\ref{eqn:triangle})$ cannot be deduced from a set of other valid inequalities for $\tcf_n$.
One needs $n \geq 3$, since the inequality $\chi_{12} \leq 1$, although facet-inducing for $n=2$, is no longer facet-inducing for $n\geq3$, see Table~\ref{table:hypermetricTCFfacets} and Proposition~\ref{prop:zeroliftingfacets} b).
\begin{proposition}[Lifting of valid inequalities and facets]\label{prop:zeroliftingfacets}
$\phantom{a}$\\
Suppose that
\begin{align}\label{eqn:tcfnfacet}
a_{0} + a_{1,2}\chi_{1,2}+ \ldots +a_{n-1,n}\chi_{n-1,n} \geq 0
\end{align}
is a valid inequality for $\tcf_{n}$. The lifting of this inequality to $\RR^{E_{n+1}}$ is the corresponding inequality which is extended by
\begin{align*}
a_{i,n+1}= 0, \quad 1 \leq i \leq n.
\end{align*}
\begin{enumerate}[a)]
\item Every lifting of a valid inequality of $\tcf_{n}$ defines a valid inequality of $\tcf_{n+1}$.
\item For $n \geq 3$, the lifting of a facet of $\tcf_{n}$ defines a facet of $\tcf_{n+1}$.
\end{enumerate}
\end{proposition}
\begin{proof}
\begin{enumerate}[a)]
\item The lifting of a valid inequality for $\tcf_n$ is always valid for $\tcf_{n+1}$, even for $n=2$, since the lifted equation applied to $\chi \in \tcf_{n+1}$ returns the same value as the orginal equation applied to $\chi|_{K_{n}}$, which is a point of $\tcf_{n}$, see Lemma~\ref{lemma:zeroliftingvertices}.
\item Now suppose that (\ref{eqn:tcfnfacet}) is a facet for $\tcf_{n}$. By the above, its lifting is a valid inequality for $\tcf_{n+1}$. We show that it defines a facet if $n \geq 3$. First, note that there has to be a coefficient $a_{i,j} \neq 0$. Since $n\geq 3$, there is some index $k \notin \{i,j\}$. To simplify notation, we assume $k=1 < i < j \leq n$.
Further, let $m := {\binom{n}{2}}$ and let $a = (a_{0},a_{1,2},a_{1,3},\ldots,a_{n-1,n}) \in \RR^{m+1}$ denote the vector of coefficients that appear in the inequality (\ref{eqn:tcfnfacet}). Since (\ref{eqn:tcfnfacet}) induces a facet of $\tcf_n$, there exist $m$ affinely independent points $\chi^k \in \tcf_{n} \subset \RR^{m}$, $ 1 \leq k \leq m$ that solve the inequality (\ref{eqn:tcfnfacet}) as an equation. Affine independence of the $m$ points $\chi^k$ means that the $m$ points $(1,\chi^k) \in \RR^{m+1}$ are linearly independent in $\RR^{m+1}$. By assumption, they solve $\langle (1,\chi^k),a \rangle = 0$, $1 \leq k \leq m$. Let $W \subset \RR^{m+1}$ denote the vector space spanned by $(1,\chi^k), 1 \leq k \leq m$. Then $\dim(W) = m$ and $W \perp a$.
Since $a_{i,j} \neq 0$ for some $1<i<j$, a non-zero entry occurs after the $n^{th}$ entry of $a$. Thus, a suitable unit vector shows $U_{n} := \{0\}^n \oplus \RR^{m+1-n} \not \subset \{a\}^\perp$. Since $W \perp a$, the inclusion $W \cap U_{n} \subset U_{n}$ is necessarily strict, which entails $\dim (W \cap U_{n}) \leq m-n$. Let $\pr: W \to \RR^n$ denote the projection onto the first $n$ coordinates.
By elementary linear algebra and since $\text{Ker}(\pr) = W \cap U_{n}$ by definition, $\dim(\text{Im}(\pr)) = \dim W - \dim(\text{Ker}(\pr)) \geq m - (m-n) = n$. Thus, $\pr(W) = \RR^n$ and the set $\{\pr((1,\chi^k))\}_{1\leq k \leq m} = \{(1,\chi^k_{1,2},\ldots ,\chi^k_{1,n})\}_{1 \leq k \leq m}$ contains $n$ linearly independent vectors, which we may assume to be indexed by $1 \leq k \leq n$ (reordering the $\chi^k$ if necessary).
Finally, we construct $\binom{n+1}{2}$ affinely independent solutions in $\tcf_{n+1}$ for the lifted equation
\begin{align*}
a_{0} + a_{1,2}\chi_{1,2}+ \ldots +a_{n,n+1}\chi_{n,n+1} = 0, \quad \text{with}\quad a_{i,n+1}= 0,\quad 1 \leq i \leq n.
\end{align*}
To simplify notation, assume that the new coordinates $\chi_{1,n+1},\ldots,\chi_{n,n+1}$ are added to the right of the previous coordinates $\chi_{1,2},\ldots,\chi_{n-1,n}$. We show that the $m+n=\binom{n+1}{2}$ points (recall $m := \binom{n}{2}$)
\begin{align*}
&\text{(a)} \quad (\chi^k,0,\ldots,0) \in \RR^{m+n}, \quad 1 \leq k \leq m, \quad \text{(with $n$ $0$'s added)},\\
&\text{(b)} \quad (\chi^k,\pr((1,\chi^k)))\in \RR^{m+n}, \quad 1 \leq k \leq n,
\end{align*}
solve the lifted equation, belong to $\tcf_{n+1}$ and are affinely independent.
The first statement follows from the choice of the $\chi^k$. The points in (a) belong to $\tcf_{n+1}$ by Lemma~\ref{lemma:zeroliftingvertices}. For (b), let $Y_{1},\ldots,Y_{n}$ be a stochastic model for $\chi^{k}$. Extend this model to $n+1$ variables $Y_{1},\ldots,Y_{n},Y_{n+1}$ by $Y_{n+1}:= Y_{1}$. Since $\pr((1,\chi^k)) = (1,\chi^k_{1,2},\ldots ,\chi^k_{1,n})$, this yields $(\chi^k,\pr((1,\chi^k))) \in \tcf_{n+1}$.
Linear independence of the $m+n$ points
\begin{align*}
\{(1,\chi^k,0,\ldots,0)\}_{1 \leq k \leq m} \cup \{(1,\chi^k,\pr((1,\chi^k)))\}_{1 \leq k \leq n}
\end{align*}
follows from the independence of $\pr((1,\chi^k)),\, 1 \leq k \leq n$ and the choice of the $\chi^k$.
\end{enumerate}
\end{proof}
\begin{remark}
By a slight abuse of notation, we will also call any vertex in the permutation orbit of $\chi^0$ a lifting of the vertex $\chi$ and any facet in the permutation orbit of a lifted facet a lifting of the respective facet.
\end{remark}
\subsection{\textbf{\upshape Unboundedness of denominators}}\label{sect:unbounded}
The following proposition shows that every rational number in the interval $[0,1]$ will appear as coordinate value in the vertex set of $\tcf_n$ starting from a sufficiently large $n$. The result is even sharper in that it detects a single vertex, whose coordinate values comprise a given finite subset of $[0,1]$-valued rational numbers.
\begin{proposition}[Unboundedness of denominators]\label{prop:unbounded}
$\phantom{a}$\\
For each finite subset $Q \subset \QQ \cap [0,1]$ of rational numbers in the interval $[0,1]$ there exists an $n \in \NN$ and a point $\chi \in \Ex(\tcf_n)$ whose coordinate-values $(\chi_{ij})_{1\leq i < j \leq n}$ include the set $Q$.\\
(By the lifting property, this holds for all $n' \geq n$, too.)
\end{proposition}
\begin{proof} By Lemma~\ref{lemma:unionsofvertices} it suffices to consider singletons $Q = \{q\}, q \in \mathbb{Q} \cap [0,1]$. The proof only uses the following properties of $\chi \in \tcf_{n}$:
\begin{itemize}
\item ``Positivity'' $\chi_{ij} \geq 0$ and
\item the permutations of the valid inequalities
\begin{align*}
\sum\limits_{i=1}^r\chi_{i,r+1} - \sum\limits_{1\leq i < j \leq r}\chi_{i,j} \leq 1, \quad r \geq 2
\end{align*}
which are hypermetric with $b$-vector $b =(1,\ldots,1,-1,0,\ldots,0)$ (with $r \geq 2$ times the entry 1), in particular permutations of the ``triangle inequality'' $\chi_{1,3}+\chi_{2,3}-\chi_{1,2} \leq 1$. The validity of these inequalities has been shown in Lemma~\ref{lemma:hypervalid}.
\end{itemize}
The cases $q= 0$ and $q=1$ are trivial.
(I) We show that for rationals $q = \frac{1}{m}$ and $q = \frac{m-1}{m}$ it suffices to choose $n=2m+1$. Let $\Omega = \{\omega_{1}, \omega_{2,1},\ldots,\omega_{2,m},\omega_{3,1},\ldots,\omega_{3,m}\}$ be a set with $2m+1$ elements
and define a positive function $g$ on $\Omega$ by
\begin{align*}
g(\omega_{1})=\frac{1}{m}; \quad g(\omega_{2,i})=\frac{m-1}{m} \quad \text{and} \quad g(\omega_{3,i})=\frac{1}{m}, \quad 1 \leq i \leq m.
\end{align*}
Normalizing $g$ by $c:=\frac{m^2+1}{m}$ yields a probability measure $\PP$ on $\Omega$ by $\PP(\{\omega\})=g(\omega)/c$.
Now, we define $2m+1$ subsets of $\Omega$ as follows:
\begin{align*}
A_{1,i}= \{\omega_{1},\omega_{2,i}\}, \quad A_{2,i}= \{\omega_{2,i},\omega_{3,i}\}, \quad 1 \leq i \leq m, \quad A_{3,1} = \{\omega_{3,1},\ldots,\omega_{3,m}\}.
\end{align*}
Since all of these $2m+1$ sets have the same probability $1/c$, they define a point $\chi \in \tcf_{2m+1}$ as in Lemma~\ref{lemma:TCFevents}.
When viewed as an edge labelling $\chi$ can be described as follows: \\
Let $\{v_{1,1},\ldots,v_{1,m},v_{2,1},\ldots,v_{2,m},v_{3,1}\}$ denote the nodes of the support graph of $\chi$. A pair of nodes $v_{i_{1},i_{2}},v_{j_{1},j_{2}}$ is connected by an edge with label $\chi_{(i_{1},i_{2}),(j_{1},j_{2})} = \PP(A_{i_{1},i_{2}} \, | \,A_{j_{1},j_{2}})$.
Draw the nodes $\{v_{1,1},\ldots,v_{1,m}\}$ at the bottom level, they form a complete subgraph, all edges labelled by $\frac{1}{m}$. Above them draw the nodes $v_{2,1},\ldots,v_{2,m}$, where $v_{2,i}$ is connected to $v_{1,i}$ with an edge labelled $\frac{m-1}{m}$. Finally, the top node $v_{3,1}$ is connected to each $v_{2,1},\ldots,v_{2,m}$ with an edge labelled $\frac{1}{m}$.
We show now that $\chi \in \Ex(\tcf_{2m+1})$.
To this end, consider a representation $\chi = \lambda y +(1-\lambda) z$, $0 < \lambda < 1$, $y,z \in \tcf_{2m+1}$.
Whenever $\chi$ satisfies a valid inequality as an equality, the same has to be true for $y$ and $z$.
Consider $y$. All $\chi$-edges with label 0 have label 0 for $y$, too. Denote the unknown label $y_{(1,1),(2,1)}$ of the $y$-edge from $v_{1,1}$ to $v_{2,1}$ by $1-a \in [0,1]$. Note that $\chi$ satisfies a triangle inequality as an equality at $v_{2,1},v_{1,1},v_{1,2}$, since $\frac{m-1}{m} + \frac{1}{m} - 0 = 1$. This enforces $y_{(1,1),(1,2)} = a$. Now the triangle $v_{1,1},v_{1,2},v_{2,2}$ enforces $y_{(1,2),(2,2)} = 1-a$. Repeating this argument gives $y_{(1,i),(2,i)} = 1-a$ for all $1 \leq i \leq m$. From this, again just using triangles, it follows $y_{(1,i),(1,j)} = a$ for all $1 \leq i < j \leq m$ and $y_{(2,i),(3,1)} = a$ for all $1 \leq i \leq m$. Finally, observe that $\chi$ satisfies the hypermetric inequality given by $b=(0,\ldots,0,1,1,\ldots,1,-1)$, with $m$ 1's, as an equality $\sum_{i=1}^{m}\chi_{(3,1),(2,i)} - \sum_{1 \leq i <j \leq m}\chi_{(2,i),(2,j)} = m\cdot \frac{1}{m} - 0 = 1$. Applied to $y$, this forces $m\cdot a = 1$, thus $a=1/m$. This shows $y=\chi$. The same argument applies to $z$. Hence $y= \chi = z$ and $\chi \in \Ex(\tcf_{2m+1})$.
(II) Now let $q = \frac{k}{m}$ for some $1 \leq k \leq m-1$. We modify the above construction to obtain a $\chi \in \Ex(\tcf_{2m+3})$ with some coordinate value equal to $q$. Extend $\Omega$ by two points to $\Omega' := \Omega \cup \{\omega_{3,m+1},\omega_{3,m+2}\}$.
Extend $g$ by
\begin{align*}
g(\omega_{3,m+1}) = \frac{k}{m} \quad \text{and} \quad
g(\omega_{3,m+2}) = \frac{m-k}{m}.
\end{align*}
Normalizing $g$ defines now $\PP'$. Use the same definitions for the sets $A_{i,j}$ as above and add the two sets
\begin{align*}
A_{3,2} = \{\omega_{3,1},\ldots,\omega_{3,m-k},\omega_{3,m+1}\}
\quad \text{and} \quad A_{3,3}= \{\omega_{3,m+1},\omega_{3,m+2}\}.
\end{align*}
All sets have the same probability (the inverse of the normalizing constant) and thus, they define a point $\chi \in \tcf_{2m+3}$.
Its support graph has two more nodes $v_{3,2},v_{3,3}$, corresponding to $A_{3,2}$ and $A_{3,3}$.
The new edges are
\begin{align*}
\chi_{(2,i),(3,2)}=\frac{1}{m}, \quad 1 \leq i \leq m-k, \quad \chi_{(3,1),(3,2)} = \frac{m-k}{m}, \quad \chi_{(3,2),(3,3)} = \frac{k}{m}.
\end{align*}
Repeating the arguments from the first part shows $y = \chi$ on the ``old'' edges. Now, using the new triangles at $v_{3,2},v_{2,i},v_{1,i}$ for $1 \leq i \leq m-k$, we get
\begin{align*}
y_{(2,i),(3,2)}=\frac{1}{m},\quad 1 \leq i \leq m-k.
\end{align*}
Note that a permutation of the hypermetric inequality $b=(1,\ldots,1,-1,0,\ldots,0)$ with $m-k+1$ leading 1`s is fulfilled by $\chi$ as an equality, if the $-1$ corresponds to $v_{3,2}$ and the 1's correspond to $v_{2,1},\ldots,v_{2,m-k},v_{3,3}$. Applied to $y$, this yields $(m-k)\cdot \frac{1}{m} + y_{(3,2),(3,3)} = 1$, thus $y_{(3,2),(3,3)} = \frac{k}{m}$. Finally, the triangle at $v_{3,1},v_{3,2},v_{3,3}$ implies $y_{(3,1),(3,2)} = \frac{m-k}{m}$. Thus, $y=\chi$ and the same argument applies to $z$. Hence, $\chi \in \Ex(\tcf_{2m+3})$.
\end{proof}
For $n \leq 4$ we have seen that $\cpp_{n} = \tcf_{n}$ (Proposition \ref{prop:cpplowdim}). This is complemented by the following result.
\begin{corollary}\label{cor:nocpp}
$\phantom{a}$\\
For $n \geq 5$ we have $\Ex(\tcf_n) \not \subset \{0,1\}^{E_n}$ and, in particular, $\cpp_n \subsetneq \tcf_n$.
\end{corollary}
\begin{proof}
By the lifting of extremal points (Lemma~\ref{lemma:zeroliftingvertices}) it suffices to prove this for $n=5$. For $q = \frac{1}{2}$ the construction (I) in the proof of Proposition~\ref{prop:unbounded} yields an example with $n=5$.
\end{proof}
\begin{remark}
For $q = \frac{1}{2}$ the above construction (I) is optimal: it gives the smallest possible $n$ for the occurence of $q$ as the coordinate value of a vertex of $\tcf_{n}$. To realize $q=\frac{1}{3}$ the construction (I) uses $n=7$, but a coordinate value $\frac{1}{3}$ already occurs for $n=6$, as the computation of $\Ex(\tcf_{6})$ in Section~\ref{sect:compresults} shows.
\end{remark}
\subsection{\textbf{\upshape Non-hypermetric facets of $\tcf_{n}$ for $n \geq 6$}}\label{sect:nonhypfacets}
We give a proof for the existence of non-hypermetric facets.
First, we provide two simple necessary conditions for hypermetricity. Of course, multiplying a given (affine) inequality by some constant $q \neq 0$ does not change the halfspace it describes. Thus, one is often interested, if a given inequality is hypermetric up to a suitable multiplication.
\begin{lemma}
\label{lemma:condforhypermetricity}
Suppose that an inequality $\sum_{1\leq i < j \leq n} c_{ij}x_{ij} \leq c_{0}$ (with rational coefficients) is equivalent to a hypermetric inequality, i.e., it becomes a hypermetric inequality defined by some $b \in \ZZ^n$ after multiplication with a suitable constant $q \in \QQ \setminus \{0\}$. Then we have:
\begin{enumerate}[a)]
\item The edges $\{i,j\} \subset E_n$ with $c_{ij} \neq 0$ form a complete subgraph of the support graph $K_n$.
\item The vectors $v_{1}:=(c_{1,3},\ldots,c_{1,n})$ and $v_{2}:=(c_{2,3},\ldots,c_{2,n})$ are linearly dependent.
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}[a)]
\item By assumption $c_{ij} = -q^{-1}\cdot b_{i}b_{j}$ for some $q \in \QQ \setminus \{0\}$. Thus, the non-zero $c_{ij}$ correspond to the edges of the complete subgraph with nodes $\{1\leq i \leq n \, | \, b_{i} \neq 0\}$.
\item Again, $c_{ij} = -q^{-1}\cdot b_{i}b_{j}$. If $b_{2}=0$, then $v_{2} = 0$, thus, $v_{1},v_{2}$ are dependent. If $b_{2}\neq 0$, then $v_{1} = (b_{1}/b_{2})\cdot v_{2}$.
\end{enumerate}
\end{proof}
\begin{remark}\label{rk:condforhyp}
Note that criterion a) of Lemma~\ref{lemma:condforhypermetricity} also implies: if there is at least one 0-coefficient, there have to be at least $n$ 0-coefficients, and if the first $n-1$ coefficients $c_{1,2},\ldots,c_{1,n}$ are positive, all have to be positive.
\end{remark}
\noindent The following proposition shows the existence of non-hypermetric facets of $\tcf_n$ starting from $n \geq 6$.
It was inspired by the 2nd inequality of Generator 7 in Table~\ref{table:TCFsixfacets}.
\begin{proposition}[Non-hypermetric facets of $\tcf_n$ for $n \geq 6$]
\label{prop:nonhypfacets}
$\phantom{a}$\\
For $n \geq 6$ there are non-hypermetric facets of $\tcf_{n}$.\\ An example, for arbitrary $n\geq 6$, is given by the facet inducing inequality
\begin{align*}
\sum_{i=1}^{5}x_{i,6} - \sum_{i=1}^{4}x_{i,i+1} - x_{1,5}\leq 2.
\end{align*}
\end{proposition}
\begin{proof}
By the lifting of facets (Proposition~\ref{prop:zeroliftingfacets}), it suffices to consider the case $n = 6$. We start with a simple observation for 0-1-vectors of even length: For $y \in \{0,1\}^{2k}, k \in \NN,$ the inequality
\begin{align}\label{eqn:cyclicinequality}
\sum_{i=1}^{2k-1}y_{i}\cdot (y_{2k}-y_{\pi(i)}) \leq (k-1)\cdot y_{2k}
\end{align}
holds, where $\pi$ is the cyclic permutation of $1,\ldots,2k-1$, i.e., $\pi(i)=i+1$, $i < 2k-1$ and $\pi(2k-1)=1$. The observation is trivial if $y_{2k} = 0$. To handle the case $y_{2k} = 1$ observe that $y_{i}(1-y_{\pi(i)}) = 1$ if and only if $y_{i}=1$ and $y_{\pi(i)}=0$. There can be at most $k-1$ occurrences of the word ``10'' in the string $y_{1},\ldots,y_{2k-1},y_{1}$.
Applying (\ref{eqn:cyclicinequality}) to arbitrary binary random variables $Y_{1},\ldots,Y_{2k}$ and taking expectations yields
\begin{align*}
\sum_{i=1}^{2k-1} \EE(Y_{i}Y_{2k}) - \sum_{i=1}^{2k-1}\EE(Y_{i}Y_{\pi(i)}) \leq (k-1) \EE(Y_{2k}).
\end{align*}
If, additionally, $a:=\EE(Y_{1})=\ldots=\EE(Y_{2k}) > 0$, dividing by $a$ gives the following valid inequality for $\tcf_{2k}$, where $x_{i,j} := \frac{1}{a}\EE(Y_iY_j)$,
\begin{align}\label{eqn:nonhypinequality}
\sum_{i=1}^{2k-1}x_{i,2k} - \sum_{i=1}^{2k-1}x_{i,\pi(i)} \leq (k-1)
\end{align}
(which has a very simple supporting graph when we identify $x_{2k-1,1}$ with $x_{1,2k-1}$).
Assume now $k \geq 3$. Since the coefficients of $x_{1,2}$ and $x_{2,3}$ are $-1$ and the coefficient of $x_{1,3}$ is $0$, the non-zero coefficients do not define a complete subgraph of the support graph. Thus, Lemma~\ref{lemma:condforhypermetricity} a) shows that the above inequality is not hypermetric for $k\geq3$.
Finally, we show that for $k=3$, the inequality (\ref{eqn:nonhypinequality})
defines a facet for $\tcf_{6} \subset \RR^{E_6}$: To this end, we define $\lvert E_6\rvert=15$ points $x^r,y^r,z^r \in \{0,1\}^{E_{6}}$, $1 \leq r \leq 5$ by
\begin{align*}
&(a) \quad x^r_{i,j}=1 \quad :\Leftrightarrow \quad \{i,j\} \subset A_{r} :=\{r,\pi^2(r),6\},\\
&(b) \quad y^r_{i,j}=1 \quad :\Leftrightarrow \quad \{i,j\} \subset B_{r} := \{r,\pi(r),\pi^3(r),6\},\\
&(c) \quad z^r_{i,j}=1 \quad :\Leftrightarrow \quad (\{i,j\} \subset B_{r} \quad \text{or} \quad \{i,j\} = \{\pi^2(r),\pi^4(r)\}).
\end{align*}
Note that these points are clique partion points and thus belong to the set $\tcf_6$ by Proposition~\ref{prop:cpp}.
Using the support graph of (\ref{eqn:nonhypinequality}) for $k=3$, it can be easily seen that they solve (\ref{eqn:nonhypinequality}) for $k=3$ as an equality. Moreover, these 15 points are affinely independent, since they are even linearly independent as the determinant of the corresponding $15 \times 15$ $0$-$1$-matrix is $-2\neq 0$.
\end{proof}
\subsection{\textbf{\upshape Embedding $\tcf_{n}$ into the Correlation and Cut polytopes}}\label{sect:corcut}
We saw already in the proof of Corollary~\ref{cor:finitepolytope} that the polytope $\tcf_n$ can be viewed essentially as a \emph{projection} of the convex polytope $\Theta_n$ onto several coordinates as in (\ref{eqn:psin}). In this section we will see that the polytope $\tcf_n$ can be embedded into the so-called \emph{correlation polytope} (or, equivalently, the so-called \emph{cut polytope}, see Proposition~\ref{prop:TCFembedding} below). Thereby, we obtain a ``dual'' description of $\tcf_n$ as the \emph{intersection} of a polytope with an affine subspace.
To this end, we need to review some notation and results from \cite{dezalaurent_97}.
Remember that $E_n$ denotes the set of edges of the complete graph $K_n$ with vertices $V_n=\{1,\dots,n\}$. For $R \subset V_n$ we define a \emph{correlation vector} $\pi(R) \in \{0,1\}^{V_n \cup E_n}$ by
\begin{align*}
\pi(R)_i = \Eins_{i \in R}, \quad 1\leq i \leq n \qquad \text{and} \qquad \pi(R)_{ij} = \Eins_{i \in R} \Eins_{j \in R}, \quad 1 \leq i < j \leq n.
\end{align*}
The \emph{correlation polytope} is then defined as the convex hull of these $2^n$ correlation vectors in $\RR^{V_n \cup E_n}$
\begin{align*}
\cor_n := \conv\left(\left\{\pi(R) \,:\, R \subset V_n\right\}\right).
\end{align*}
\begin{lemma}[\cite{dezalaurent_97} Prop.\ 5.3.4] \label{lemma:CORevents}
$\phantom{a}$\\
A point $p \in \RR^{V_n\cup E_n}$ belongs to $\cor_n$ if and only if it can be written as $p_i = \PP(A_i)$, $1\leq i \leq n$ and $p_{ij} = \PP(A_i \cap A_j)$, $1\leq i < j \leq n$ for some probability space $(\Omega,\A,\PP)$ and measurable subsets $A_1,\dots,A_n \in \A$.
\end{lemma}
\noindent Secondly, let $S \subset V_{n+1}$. A \emph{cut vector} $\delta(S) \in \{0,1\}^{E_{n+1}}$ is defined through
\begin{align*}
\delta(S)_{ij} = \Eins_{\lvert S \cap \{i,j\} \rvert = 1}, \qquad 1 \leq i < j \leq n+1.
\end{align*}
Since $\delta(S)=\delta(S^c)$, there are, in fact, $2^{n+1}/2= 2^n$ different points $\delta(S)$. The \emph{cut polytope} is defined as the convex hull of these cut vectors in $\RR^{E_{n+1}}$
\begin{align*}
\cut_{n+1} := \conv\left(\left\{\delta(S) \,:\, S \subset V_{n+1}\right\}\right).
\end{align*}
Being $\{0,1\}$-valued, the correlation vectors and the cut vectors are automatically the extremal points of their convex hulls
\begin{align*}
\Ex\left(\cor_n\right)=\{ \pi(R) \,:\, R \subset V_n \} \quad \text{and} \quad
\Ex\left(\cut_{n+1}\right)=\{ \delta(S) \,:\, S \subset V_{n+1} \}.
\end{align*}
It is a well-known result that $\cor_n \subset \RR^{V_n \cup E_n}$ and $\cut_{n+1} \subset \RR^{E_{n+1}}$ can be transformed into each other by a linear bijection.
\begin{proposition}[\cite{dezalaurent_97}, Section~5.2)] \label{prop:covariancemapping}
$\phantom{a}$\\
The {covariance mapping} $\zeta_n: \RR^{V_n \cup E_n} \rightarrow \RR^{E_{n+1}}$,
which maps $p \in \RR^{V_n \cup E_n}$ to $\zeta_n(p)=x \in \RR^{E_{n+1}}$ via
\begin{align*}
x_{i,n+1} = p_i, \quad 1 \leq i \leq n \qquad \text{and} \qquad x_{ij}=p_i + p_j - 2p_{ij}, \quad 1 \leq i < j \leq n,
\end{align*}
induces a linear bijection
\begin{align*}
\zeta_n:\cor_{n} \rightarrow \cut_{n+1}.
\end{align*}
\end{proposition}
\begin{remark}
In \citep{dezalaurent_97} the inverse $\xi_n:=\zeta_n^{-1}$ is termed covariance mapping. For us, it was more instructive to work with $\zeta_n$ instead of $\xi_n$.
\end{remark}
A probabilistic description of $\cut_{n+1}$ is as follows.
Here the symmetric difference between sets $A$ and $B$ will be denoted by $A \triangle B = (A \setminus B) \cup (B \setminus A)$.
\begin{lemma} \label{lemma:CUTevents}
A point $x \in \RR^{E_{n+1}}$ belongs to the cut polytope $\cut_{n+1}$ if and only if one of the following two equivalent statements holds true:
\begin{enumerate}[(i)]
\item $x_{i,n+1} = \PP(A_i)$, $1 \leq i \leq n$ and $x_{ij} = \PP(A_i \triangle A_j)$, $1 \leq i < j \leq n$ for some probability space $(\Omega,\A,\PP)$ and measurable subsets $A_1,\dots,A_n \in \A$.
\item $x_{ij} = \PP(B_i \triangle B_j)$, $1 \leq i < j \leq n+1$ for some probability space $(\Omega,\A,\PP)$ and measurable subsets $B_1,\dots,B_{n+1} \in \A$.
\end{enumerate}
\end{lemma}
\begin{proof} The equivalence to (i) is an immediate consequence of Lemma~\ref{lemma:CORevents} and Proposition~\ref{prop:covariancemapping}. The equivalence of (i) and (ii) can be seen as follows: (i) $\Rightarrow$ (ii): Set $B_i=A_i$, $1 \leq i \leq n$ and $B_{n+1}=\emptyset$. (ii) $\Rightarrow$ (i): Set $A_i=B_i \triangle B_{n+1}$, $1\leq i \leq n$ and use that $(C \triangle D) \triangle ( E \triangle D)= C \triangle E$ for any triplet of sets $C,D,E$.
\end{proof}
\noindent Finally, this enables us to interpret $\tcf_n$ as an intersection of $\cor_n$ (\resp $\cut_{n+1}$) with an affine subspace of $\RR^{V_n \cup E_n}$ (\resp $\RR^{E_{n+1}}$) in the following sense.
\begin{proposition}[Embedding $\tcf_n$ into the correlation polytope]\label{prop:TCFembedding}
$\phantom{a}$\\
The injective affine map $\iota_n:\RR^{E_n} \rightarrow \RR^{V_n \cup E_n}$ which maps $\chi \in \RR^{E_n}$ to $\iota_n(\chi)=p \in \RR^{V_n \cup E_n}$ via
\begin{align*}
p_i=\frac{1}{n}, \quad 1 \leq i \leq n \qquad \text{and} \qquad p_{ij}= \frac{\chi_{ij}}{n}, \quad 1 \leq i < j \leq n,
\end{align*}
induces a bijection
\begin{align*}
\iota_n:\tcf_{n} \rightarrow \cor_{n} \cap \left\{ p \in \RR^{V_n \cup E_n} \,:\, p_i=\frac{1}{n}, \, i=1,\dots,n\right\}.
\end{align*}
\end{proposition}
\begin{proof}
The map $\iota_n$ is injective by definition. First, we show that $\iota_n(\tcf_n) \subset \cor_n$. Because of Lemma~\ref{lemma:TCFevents} and Remark~\ref{remark:TCFevents}, a point $\chi \in \tcf_n$ has a stochastic model $A_1,\dots,A_n$ with $\PP(A_1)=\dots =\PP(A_n)=1/n$ and $\chi_{ij}=\PP(A_i \cap A_j)/\PP(A_j)$. Lemma~\ref{lemma:CORevents}, applied to $A_{1},\ldots,A_{n}$ and $\PP$, shows that $\iota_{n}$ maps $\tcf_n$ to $\cor_{n}$. Now, suppose that $p \in \cor_{n}\cap \bigcap_{i=1}^n\{p_{i} = 1/n\}$. By Lemma~\ref{lemma:CORevents} there is a stochastic model with sets $A_{1},\ldots,A_{n}$, $\PP(A_{1})=\ldots=\PP(A_{n}) = 1/n$, $\PP(A_{i}\cap A_{j}) = p_{ij}$. Thus, $\chi = (n\cdot p_{ij})_{1 \leq i < j \leq n} $ is a preimage of $p$ in $\tcf_{n}$.
\end{proof}
\noindent Note that we just established the following equivalences
\begin{align*}
\chi \in \tcf_n \quad \Leftrightarrow \quad \iota_n(\chi) \in \cor_n \quad \Leftrightarrow \quad \zeta_n \circ \iota_n (\chi) \in \cut_{n+1}.
\end{align*}
In particular, one can pull back facets from $\cut_{n+1}$ to $\cor_{n}$ with the covariance mapping $\zeta_{n}$, and further, we obtain an $\mathcal{H}$-representation for $\tcf_{n}$ using $\zeta_{n} \circ \iota_{n}$. Thus, any $\mathcal{H}$-representation of $\cor_n$ or $\cut_{n+1}$ yields an $\mathcal{H}$-representation of $\tcf_n$ as follows.
\begin{proposition}[Pulling back $\mathcal{H}$-representations]\label{prop:cutcorfacets}
$\phantom{a}$
\begin{enumerate}[a)]
\item {\normalfont (\cite{dezalaurent_97} Prop.\ 26.1.1, p.~402)}
$\phantom{a}$\\
The covariance mapping $\xi_n:=\zeta_n^{-1}$ maps a valid inequality for $\cut_{n+1}$ (resp.\ facet of $\cut_{n+1}$)
\begin{align}\label{eqn:cuthalfsp}
\sum_{1 \leq i < j \leq n+1} c_{ij}x_{ij} \leq c_{0}
\end{align}
to the following valid inequality $\cor_{n}$ (resp.\ facet of $\cor_{n}$)
\begin{align}\label{eqn:corhalfsp}
\sum_{1\leq i \leq n} b_{i}p_{i} + \sum_{1 \leq i < j \leq n} (-2c_{ij})p_{ij} \leq c_{0}
\quad \text{with} \quad
b_{i} = \sum_{1 \leq s < i}c_{si} + \sum_{i < s \leq n+1}c_{is}.
\end{align}
\item The above valid inequality (resp.\ facet) of $\cut_{n+1}$ induces the following valid inequality for $\tcf_{n}$ via
$\zeta_n \circ \iota_n$
\begin{align}\label{eqn:tcfhalfsp}
\sum_{1\leq i < j \leq n} (-2c_{ij}) \chi_{ij} \leq n\cdot c_{0} - 2 \sum_{1\leq i < j \leq n} c_{ij} - \sum_{i=1}^n c_{i,n+1}.
\end{align}
If applied to all elements of an $\mathcal{H}$-representation of $\cut_{n+1}$ (e.g.\ all facets of $\cut_{n+1}$), this gives an $\mathcal{H}$-representation for $\tcf_{n}$.
\end{enumerate}
\end{proposition}
\begin{proof}
\begin{enumerate}[b)]
\item
It suffices to replace $x_{ij}$ in Inequality~(\ref{eqn:cuthalfsp}) by
\begin{align*}
(\zeta_n \circ \iota_n(\chi))_{ij}
= \left\{\begin{array}{ll}
\frac{1}{n} & \qquad j=n+1,\smallskip\\
\frac{2}{n} - \frac{2}{n} \chi_{ij} & \qquad 1 \leq i < j \leq n.
\end{array}\right.
\end{align*}
and to reorder the resulting terms.
\end{enumerate}
\end{proof}
\paragraph{\textbf{\upshape Dual views on $\tcf_n$}} Summarizing, we obtain two complementary views on the polytope $\tcf_n$ which may be illustrated as follows.
\begin{center}
\begin{minipage}{0.45\textwidth}
\xymatrix{
\Theta_n \ar@{>>}[d]_{\psi_n} \\
\tcf_n \ar@{^{(}->}[r]_{\iota_n} & \cor_n \ar[r]_{\zeta_n} & \cut_{n+1}
}
\end{minipage}
\hspace{2mm}
\begin{minipage}{0.45\textwidth}
\xymatrix{
\RR^{\finite_n^{(2)}} \ar@{>>}[d]_{\psi_n} \\
\RR^{E_n} \ar@{^{(}->}[r]_{\iota_n} & \RR^{V_n \cup E_n} \ar[r]_{\zeta_n} & \RR^{E_{n+1}}
}
\end{minipage}
\end{center}
Here $\psi_n$ is given by the ``projection'' map (\ref{eqn:psin}), the map $\iota_n$ is the embedding from Proposition~\ref{prop:TCFembedding} and $\zeta_n$ the covariance mapping from Proposition~\ref{prop:covariancemapping}.
While any $\mathcal{V}$-representation of $\Theta_n$ easily yields a $\mathcal{V}$-representation of $\tcf_n$ essentially by a projection, any $\mathcal{H}$-representation of $\cut_{n+1}$ easily yields an $\mathcal{H}$-representation of $\tcf_n$ essentially by an intersection. Unfortunately, $\Theta_n$ is a priori given by its facets (an $\mathcal{H}$-represenation), while $\cut_{n+1}$ is a priori given by its vertices (a $\mathcal{V}$-representation) and not the other way around, such that both views come along with certain drawbacks.
At least the facets of $\cut_{n+1}$ are classified to some extent.
\paragraph{\textbf{\upshape The facets of $\cut_{n+1}$ and their generators}} \citep[Part V]{dezalaurent_97}
Let us consider the following two kinds of actions on $\RR^{E_{n+1}}$. On the one hand the symmetric group $S_{n+1}$ acts on $\RR^{E_{n+1}}$ by node permutations: $(\sigma(x))_{ij}:= x_{\sigma(i)\sigma(j)}$ for $\sigma \in S_{n+1}$. These actions are simply called \emph{permutations}. On the other hand each of the $2^n$ cut vectors $\delta(S)$ acts on $\RR^{E_{n+1}}$ by
\begin{align*}
(\delta(S)(x))_{ij}
= \left\{
\begin{array}{ll}
1-x_{ij} & \text{if } \delta(S)_{ij}=1,\\
x_{ij} & \text{otherwise},
\end{array}
\right.
\end{align*}
for any $S \subset V_{n+1}=\{1,\dots,n+1\}$, \ie coordinates $x_{ij}$ corresponding to the edges of the cut beween $S$ and $S^c$ are replaced by $1-x_{ij}$. These actions are called \emph{switchings}. Note that $\delta(S) \circ \delta(R) = \delta(S \triangle R)$ and that $\delta(S) \circ \sigma = \sigma \circ \delta(\sigma(S))$.
In fact, both kinds of actions can be restricted to the cut polytope $\cut_{n+1}$. For any $\sigma \in S_{n+1}$ and any $S \subset V_{n+1}$
\begin{align*}
\sigma(x) \in \cut_{n+1} \quad \Leftrightarrow \quad x \in \cut_{n+1} \quad \Leftrightarrow \quad \delta(S)(x) \in \cut_{n+1}.
\end{align*}
These permutations and switchings on the polytope $\cut_{n+1}$ induce, of course, corresponding actions on its facets.
First, it is not surprising that (\ref{eqn:cuthalfsp}) is a facet inducing inequality of $\cut_{n+1}$ if and only if
\begin{align*}
\sum_{1 \leq i < j \leq n+1} c_{\sigma(i)\sigma(j)}x_{ij} \leq c_{0}
\end{align*}
is facet inducing for $\cut_{n+1}$. Second, any facet inducing inequality (\ref{eqn:cuthalfsp}) can be \emph{switched} by a cut vector $\delta(S)$
to another facet inducing inequality of $\cut_{n+1}$ which is given by
\begin{align*}
\sum_{1 \leq i < j \leq n+1} (1-2\delta(S)_{ij})c_{ij}x_{ij} \leq c_{0} - \sum_{1 \leq i < j \leq n+1} \delta(S)_{ij}c_{ij}.
\end{align*}
Let $O^{SP}(g,c_{0})$ denote the full orbit of a facet $g(x) \leq c_{0}$ under all possible finite applications of switchings and permutations to $g(x) \leq c_{0}$. The set of all facets of $\cut_{n+1}$ splits into finitely many such orbits, say $O^{SP}_{i}$, $i\in I$. Choosing one facet $g^{(i)}(x) \leq c^{(i)}_{0}$ from each orbit $O^{SP}_{i}$ yields a set of representatives $g^{(i)}(x) \leq c^{(i)}_{0}$, $i\in I$, of the facets of $\cut_{n+1}$, up to switchings and permutations. In this way \emph{generators} for the facets of $\cut_{n+1}$ are given in the literature. It is a feature of the cut polytope that it always has a set of \emph{homogeneous generators}, i.e. with $c^{(i)}_{0} = 0$, $i \in I$ \citep[Section 26.3.2]{dezalaurent_97}.
The facets of $\cut_{n+1}$ and corresponding generators are known for $n \leq 7$ \citep[p.~504]{dezalaurent_97}. In Table~\ref{table:cutgenerators} (Appendix~\ref{sect:tables}) we list the 11 generators of the $116\phantom{.}764$ facets of $\cut_7$ that will be used to derive the facets of $\tcf_6$.
\paragraph{\textbf{\upshape Relations to unit covariances}}
In their works on {\citeauthor{mcmillan1955history}'s} ({\citeyear{mcmillan1955history}}) realization problem concerning covariances of binary processes
{\cite{quint08}, \cite{la13,la15} and \cite{shepp1963positive,shepp1967covariances}} considered $\{-1,1\}$-valued random vectors $(U_1,\dots,U_n)$ (instead of $\{0,1\}$-valued vectors) and studied the set of \emph{unit covariances}
\begin{align*}
\mathcal{U}_n := \left\{ u \in \RR^{E_n} \pmid
\begin{array}{l}
\text{$u_{ij}=\EE (U_iU_j)$ where}\\
\text{$U_1,\dots,U_n$ take values in $\{-1,1\}$}
\end{array}
\right\}.
\end{align*}
As a consequence of Lemma~\ref{lemma:CUTevents} (ii) (set $B_i=\{U_i=1\}$ therein) the cut polytope $\cut_n$ and the set of unit covariances $\mathcal{U}_n$ are affine equivalent via the bijective mapping $g_n:\RR^{E_n} \rightarrow \RR^{E_n}$, $g_n(x)=\frac{1}{2}(1-x)$ through
\begin{align}\label{eq:cutunit}
\cut_n = g_n(\mathcal{U}_n).
\end{align}
Let us further denote for $c \in [0,1]$ as in {\cite{shepp1963positive}}
\begin{align*}
\mathcal{U}_n(c) := \left\{ u \in \RR^{E_n} \pmid
\begin{array}{l}
\text{$u_{ij}=\EE (U_iU_j)$ where}\\
\text{$U_1,\dots,U_n$ take values in $\{-1,1\}$} \\
\text{and $\PP(U_1=1)=\dots=\PP(U_n=1)=c$}
\end{array}
\right\}.
\end{align*}
It is immediate that $\mathcal{U}_n(c)=\mathcal{U}_n(1-c)$ and
repeating an argument from {\cite{shepp1963positive}}, p.~10,
it is not difficult to see that
$\mathcal{U}_n(c)$, $0\leq c \leq 1/2$ are increasing towards $\mathcal{U}_n(1/2)=\mathcal{U}_n$.
The latter equality follows from the fact that the unit covariance of a $\{-1,1\}$-valued random vector remains unchanged after multiplication with an independent $\{-1,1\}$-valued zero mean variable.
The affine equivalence \eqref{eq:cutunit} can be refined to
\begin{align}\label{eq:cutunitrefined}
\cut_n(c) = g_n(\mathcal{U}_n(c)), \quad c \in [0,1]
\end{align}
if we set
\begin{align*}
\cut_n(c):=\pr_n(\cut_{n+1} \cap \{x \in \RR^{E_{n+1}} \,:\, x_{i,n+1}=c, \, 1\leq i \leq n\})
\end{align*}
and $\pr_n:\RR^{E_{n+1}} \rightarrow \RR^{E_n}$ is the projection onto the edges not containing the vertex $n+1$.
A probabilistic description of the polytopes $\cut_n(c)$, $c \in[0,1]$
follows from the equivalence (i) in Lemma~\ref{lemma:CUTevents} (set $A_i=\{U_i=1\}$) and thereby proves the refinement \eqref{eq:cutunitrefined} as follows.
\begin{lemma}\label{lemma:CUTCevents}
A point $x \in \RR^{E_n}$ belongs to $\cut_n(c)$ if and only if can be written as $x_{ij}=\PP(A_i \triangle A_j)$, $1 \leq i < j \leq n$ for some probability space $(\Omega,\A,\PP)$ and measurable subsets $A_1,\dots,A_n \in \A$ satisfying $\PP(A_i)=c$, $1 \leq i \leq n$.
\end{lemma}
A direct connection of unit covariances to $\tcf_n$ can be obtained
from Lemma~\ref{lemma:TCFevents} and Remark~\ref{remark:TCFevents} (set $U_i=2 \cdot \eins_{A_i}-1$ therein) as
\begin{align*}f_n(\tcf_n) = \mathcal{U}_n(1/n),\end{align*}
where $f_n:\RR^{E_n} \rightarrow \RR^{E_n}$ is the bijective affine mapping $f_n(x)=\frac{4}{n}x-\frac{4}{n}+1$.
It can be easily checked that the following diagram commutes
if $\zeta_n$ and $\iota_n$ are the respective affine mappings from Propositions~\ref{prop:covariancemapping} and \ref{prop:TCFembedding}.
\begin{center}
\begin{minipage}{0.45\textwidth}
\xymatrix{
\tcf_n \ar@{^{(}->}[r]^{\zeta_n \circ \iota_n} \ar[d]_{f_n} & \cut_{n+1} \ar@{>>}[d]^{\pr_n}\\
\mathcal{U}_n(1/n) \ar[r]_{g_n} & \cut_n(1/n)
}
\end{minipage}
\hspace{2mm}
\begin{minipage}{0.45\textwidth}
\xymatrix{
\RR^{E_{n}} \ar@{^{(}->}[r]^{\zeta_n \circ \iota_n} \ar[d]_{f_n} & \RR^{E_{n+1}} \ar@{>>}[d]^{\pr_n}\\
\RR^{E_{n}} \ar[r]_{g_n} & \RR^{E_{n}}
}
\end{minipage}
\end{center}
We remark the simple form of the mapping $(g_n \circ f_n)(x)=\frac{2}{n}(1-x)$. The following lemma shows that the polytopes $\cut_n(c)$, $0 < c \leq 1/n$ and $\mathcal{U}_n(c)$, $0 < c \leq 1/n$ are also affine isomorphic.
\begin{lemma} For $\lambda \in [0,1]$ we have
\begin{align*}
\cut_n(\lambda/n)&=\lambda \cdot \cut_n(1/n)\\
\mathcal{U}_n(\lambda/n)&=\lambda \cdot \mathcal{U}_n(1/n) + (1-\lambda).
\end{align*}
\end{lemma}
\begin{proof}
The second relation follows from the first by applying the map $g_n$. We prove the first statement using Lemma~\ref{lemma:CUTCevents} for both inclusions (``$\subset$'' and ``$\supset$''), where we may assume that $A := \bigcup_{i=1}^n A_i \neq \Omega$ (otherwise add a point to $\Omega$).
An element $x \in \cut_n(\lambda/n)$ admits the representation $x_{ij}=\PP(A_i \triangle A_j)$ for sets $A_1,\dots,A_n$ with $\PP(A_i)=\lambda/n$.
It follows that $\PP(A)\leq \lambda$ and we can extend $\PP':=(1/\lambda) \cdot \left.\PP\right|_A$ to a probability measure on $\Omega$ which gives $x_{ij}=\lambda \PP'(A_i\triangle A_j)$ with $(\PP'(A_i\triangle A_j))_{1 \leq i<j \leq n} \in \cut_n(1/n)$.
Conversely, $x \in \cut_n(1/n)$ admits the representation $x_{ij}=\PP(A_i \triangle A_j)$ for sets $A_1,\dots,A_n$ with $\PP(A_i)=1/n$ and we can extend the measure $\PP':= \lambda \cdot \left.\PP\right|_A$ to $\Omega$ which gives
$\lambda \cdot x_{ij}= \PP'(A_i\triangle A_j)$ with $(\PP'(A_i\triangle A_j))_{1 \leq i<j \leq n} \in \cut_n(\lambda/n)$.
\end{proof}
Together with $(g_n \circ f_n)^{-1}(y)=1-\frac{n}{2}y$ this identifies the polytope $\tcf_n$ as
\begin{align}\label{eq:tcfcutunit}
\tcf_n = 1 - \frac{1}{2c} \cut_n(c) = 1- \frac{1}{4c} \left(1-\mathcal{U}_n(c)\right), \quad \text{ for any } c \in (0,1/n].
\end{align}
Hence, any better understanding on one of the polytopes in \eqref{eq:tcfcutunit}
will automatically transfer to all the other ones.
\section{Computational results}\label{sect:compresults}
We computed the vertices and facets of $\tcf_n$ for $n \leq 6$ using the software \texttt{R} \citep{R} and \texttt{polymake} \citep{polymake}.
Their explicit representatives are documented in the tables of Appendix~\ref{sect:tables}.
In order to obtain the vertices and facets of $\tcf_6$,
we had to use both views on $\tcf_{6}$ described at the end of Section~\ref{sect:corcut}: a $\mathcal{V}$-representation of $\tcf_6$ was obtained via the polytope $\Theta_{6}$, the reduction to the vertex representation Ex($\tcf_{6}$) took extra efforts. An $\mathcal{H}$-representation for $\tcf_{6}$ was obtained via the embedding into $\cut_{7}$ (using the known facet-representation), from which we extracted a facet-representation of $\tcf_{6}$ using the previously computed vertices Ex($\tcf_{6}$). Below we give a detailed description of our methods.
\paragraph{\textbf{\upshape The vertices and facets of $\tcf_n$ for $ 2 \leq n \leq 5$}}
$\phantom{a}$\\
For $n \leq 4$ the vertices and facets of $\tcf_n$ were computed already in \cite{strokorb_13} p.~62. In particular, all vertices are $\{0,1\}$-valued, hence clique partition points (cf.\ Proposition~\ref{prop:cpplowdim}).
The vertices and facets of $\tcf_5$ have been obtained directly using \texttt{R} and \texttt{polymake} via the two different approaches presented in Section~\ref{sect:corcut} (leading to the same result): via the polytope $\Theta_5$ and the embedding into the correlation polytope $\cor_5$ (defined by its vertices).
Here, for $n=5$, the software \texttt{R} was simply used to generate the input for \texttt{polymake}. From these computations we see that $\tcf_5$ has 214 vertices in 11 permutation orbits as listed in Table~\ref{table:TCFfivevertices} (Appendix~\ref{sect:tables}). While 52 vertices in 7 permutation orbits are $\{0,1\}$-valued (the expected clique partition points), for the first time also $\{0,\OH\}$-valued vertices turn up (162 in 4 permutation orbits).
Representatives for the permutation orbits of the facets of $\tcf_{n}$ for each $2\leq n \leq 5$ are listed in Table~\ref{table:hypermetricTCFfacets} in the Appendix~\ref{sect:tables}. Since all facets turned out to be hypermetric, we describe them by their defining vectors $b \in \ZZ^n$. In particular, we obtain the following result.
\begin{proposition}\label{prop:nleq5hypermetric} For $n \leq 5$ all facets of $\tcf_{n}$ are hypermetric.
\end{proposition}
\noindent Let us now turn to the case $n=6$, which needed additional arguments to reduce the computational burden.
\paragraph{\textbf{\upshape The vertices of $\tcf_6$}}
$\phantom{a}$\\
According to our computational results, the polytope $\tcf_{6}$ possesses 28895 vertices in 88 permutation orbits, whose representatives are listed in Table~\ref{table:TCFsixvertices} in the Appendix~\ref{sect:tables}.
For the first time, also $\{0,\OT,\TT\}$-valued vertices occur, more precisely,
\begin{itemize}
\item 203 vertices in 11 orbits are $\{0,1\}$-valued,
\item 4662 vertices in 16 orbits are $\{0,\OH\}$-valued
\item 2430 vertices in 11 orbits are $\{0,\OH,1\}$-valued,
\item 21600 vertices in 50 orbits are $\{0,\OT,\TT\}$-valued.
\end{itemize}
\noindent It was not feasible to use the simple embedding of $\tcf_6$ into $\cut_6$
from Section~\ref{sect:corcut} and \texttt{polymake} to compute the vertices by common standard hardware in reasonable time. Instead, we used the projection of the $\Theta_{6}$ polytope in (\ref{eqn:psin}) to obtain a $\mathcal{V}$-representation for $\tcf_{6}$, from which - with some additional efforts - we extracted the vertex representation
\begin{description}
\item[1st step:]\emph{Computing a $\mathcal{V}$-representation of $\tcf_6$.} \\ With \texttt{R} we generated the input for \texttt{polymake} (63 inequalities with 58 coefficients each) to define the polytope $\Theta_{6}$ in $\RR^{57}$. Then \texttt{polymake} computed the 200$\phantom{.}$214 extremal points of $\Theta_{6}$ in less then 20 minutes by standard hardware. We projected the extremal points of $\Theta_{6}$ onto the 15 coordinates for $\tcf_{6}$, applied the coordinatewise $2-x$-transformation, and removed duplicates. This gave us 168$\phantom{.}$894 points in $[0,1]^{15}$ with convex hull $\tcf_{6}$ (a $\mathcal{V}$-representation of $\tcf_6$).
Their coordinate values were all fractions $\frac{a}{b}, 0 \leq a \leq b \leq 9$.
\item[2nd step:] \emph{Reduction to a vertex representation of $\tcf_{6}$.}\\
It was not feasible to extract the subset of extremal points directly by \texttt{polymake}. Using \texttt{R} we determined the 521 permutation orbits of these 168$\phantom{.}$894 convex hull points and chose 521 representatives. These representatives included the 11 well-known representatives for $\Ex(\tcf_{6}) \cap \{0,1\}^{15}$ (i.e., the clique partition points of the complete graph $K_{6}$, see Proposition~\ref{prop:cpp}), and the 4 liftings of the 4 representatives for $\Ex(\tcf_{5}) \cap \{0,1/2\}^{10}$ described
above (see Table~\ref{table:TCFfivevertices} in the Appendix~\ref{sect:tables}). This gave us a list of 15 representatives known to be extremal and 506 undecided ones.\\
The extremal ones among them were identified as follows.\\
a) First, we took the union of the full permutation orbits of the 15 known representatives, a set of 1175 points,
and added the undecided 506 candidates. The resulting list of 1681 points was handed over to \texttt{polymake}, which computed the 1259 extremal points of their convex hull (among them the previously mentioned set of 1175 points). Any candidate from the 506-list not appearing among these 1259 extremal points is a strict convex combination of points from $\tcf_{6}$, thus not extremal. This left us with the 15 representatives known to extremal plus only $84 = 1259-1175$ undecided representatives from the previous list of 506.\\
b) For each of the remaining 84 undecided representatives we computed with {\tt polymake}, if there is a hyperplane positively separating this selected representative from the union of all orbits of the 15 representatives known to extremal and the 83 other undecided representatives (in each case roughly 30000 points). If so, the selected representative is extremal, otherwise not. For a proof of this statement see the following Lemma~\ref{lemma:decidingextremality}.
In this way we found $73$ extremal representatives among the $84$ undecided ones, which led to the 15+73= 88 representatives for $\Ex(\tcf_{6})$ in Table~\ref{table:TCFsixvertices} (Appendix~\ref{sect:tables}).
\end{description}
In order to justify the last step, the following lemma is needed.
\begin{lemma}\label{lemma:decidingextremality}
Let $A \subset \RR^n$ and $B\subset \RR^n$ be two disjoint finite sets with the property that
either $B \subset \Ex(A\cup B)$
or $B \cap \Ex(A \cup B) = \emptyset$ (property {\upshape ($*$)} in the proof).
Let $x$ be a point from $B$.
Then
$x \in \Ex(A\cup \{x\})$ if and only if $x \in \Ex(A \cup B)$.
\end{lemma}
\noindent (Our application in mind is $A \subset \RR^{\binom{n}{2}}$, a union of $S_{n}$-orbits, $B \subset \RR^{\binom{n}{2}}$ another $S_{n}$-orbit. Then the above condition $(*)$ holds, since $S_{n}$ acts via invertible linear maps.)
\begin{proof}
Note that the following identities hold trivially for a finite set $A \subset \RR^n$:\\
$\Ex(A) \subset A$ ($*$1) and $\conv(A) = \conv(\Ex(A))$ ($*$2).
Hence, the assertion is a consequence of the following.\\
``$\Leftarrow$'': $x \not\in \Ex(A\cup \{x\}) \overset{(*1)}{\Rightarrow} \Ex(A\cup \{x\}) \subset A \overset{(*2)}{\Rightarrow} x \in \conv(A)$, thus $x$ is a convex combination of points from $A$ (which are different from $x$, since $x \in B$, $A \cap B = \emptyset$), thus $x \not\in \Ex(A \cup B)$.\\
``$\Rightarrow$'': $x \not\in \Ex(A \cup B) \overset{(*)}{\Rightarrow} B \cap \Ex(A \cup B) = \emptyset \overset{(*1)}{\Rightarrow} \Ex(A \cup B) \subset A \overset{(*2)}{\Rightarrow} \conv(A\cup B) \subset \conv(A) \Rightarrow x \in \conv(A)$, as above now $x \not\in \Ex(A\cup\{x\})$ follows.
\end{proof}
\paragraph{\textbf{\upshape The facets of $\tcf_6$}}
$\phantom{a}$\\
It turned out that $\tcf_{6}$ has 18720 facets which split into 67 permutation orbits. For an annotated complete list see Table~\ref{table:TCFsixfacets} in the Appendix~\ref{sect:tables}. The 67 representatives for $\tcf_{6}$ are grouped into 11 classes, according to their ``ancestral cut polytope generator'' (see below). The first 6 generators led to 6 classes with 17 representatives for $\tcf_{6}$, which are all hypermetric. A list of the corresponding 17~$b$-vectors is given in Table~\ref{table:hypermetricTCFsixfacets} (Appendix~\ref{sect:tables}). The remaining 5 generators induced 50~representatives and all of them are non-hypermetric (this is easily checked using Lemma~\ref{lemma:condforhypermetricity} and Remark~\ref{rk:condforhyp} for all but the 7th inequality derived from generator 9, for this one the vectors $c_{2,4},c_{2,5},c_{2,6}$ and $c_{3,4},c_{3,5},c_{3,6}$ are independent and the same reasoning as for criterion (b) of Lemma~\ref{lemma:condforhypermetricity} works). Thus, the number of hypermetric orbits is 17 out of 67 ( $\approx 25.4\%$), with 858 hypermetric facets out of 18720 (just $\approx 4.6\%$). \\
\\
We obtained this list of representatives for the facets of $\tcf_6$ in using known results about the cut polytope $\cut_{7}$ (Section~\ref{sect:corcut}), the previously computed vertex set $\Ex(\tcf_{6})$ and the software \texttt{R}:
\begin{description}
\item[1st step:]
Choose one of the 11 homogeneous generators $g_{i} \leq 0$, $i \in \{1,\ldots,11\}$ for the facets of the cut polytope $\cut_{7}$ (see Table~\ref{table:cutgenerators} (Appendix~\ref{sect:tables}) and Section~\ref{sect:corcut}). Compute the list of all facets of $\cut_{7}$ generated by $g_{i}$ w.r.t.\ switchings and permutations (cf. Section~\ref{sect:corcut}). This results in an $a_{i} \times 22$ matrix with $a_{i} \leq 40320$ for all $i$ (see also \cite{dezalaurent_97} Figure~30.6.1).
\item[2nd step:]
Apply the simple map from Proposition~\ref{prop:cutcorfacets} to all rows of the matrix from step 1. This yields a set of valid inequalities for $\tcf_{6}$ (an $a_{i} \times 16$ matrix), which is permutation invariant by construction.
Choose representatives of the permutation orbits (the largest count was 93 representatives).
\item[3rd step:]
Use the $28\phantom{.}895$ precomputed vertices in $\Ex(\tcf_{6})$ to decide for each representative from step 2, if it defines a facet of $\tcf_{6}$. For that, first determine which vertices from $\Ex(\tcf_{6})$ solve the inequality as an equality. Then check if the rank of the matrix of solutions with an added 1-column in front is at least 15. We used the vertex set $6 \cdot \Ex(\tcf_{6})$ to make all computations integer valued, so the rank-checking procedure should be computationally reliable in this case. This gives a list of representatives for certain permutation orbits of $\tcf_{6}$-facets ``stemming from the cut polytope generator $g_{i}$''.
\item[4th step:] If done for all 11 generators, the union of the 11 lists obtained in step 3 gives a complete list of representatives of the facets of $\tcf_{6}$. This holds true, since the set of all valid inequalities obtained in the second step for all $1 \leq i \leq 11$ defines $\tcf_{6}$ by Proposition~\ref{prop:cutcorfacets}, thus we know that the facets of $\tcf_{6}$ are a subset.
Finally, we checked that representatives from different lists have different permutation orbits. Thus, the 11 lists partition a minimal set of facet representatives for $\tcf_{6}$ according to the unique ``ancestral cut polytope generator''.
\end{description}
\begin{remark}
It is feasible to generate all 116 764 facets of $\cut_{7}$ in step 1, and go through steps 2 and 3 (testing 391 representatives from step 2), to just obtain the 67 facet representatives, but then relating them to the different cut polytope generators needs extra bookkeeping.
\end{remark}
\begin{remark}
One can exploit the interaction of the permutation group actions on $\cut_{n+1}$ and $\tcf_{n}$ to avoid the large row counts in step 1 and 2. Starting from a list $h_{j} \leq c_{j}, j \in J$ of facet representatives for the cut polytope $\cut_{n+1}$ w.r.t. \emph{permutations} ($|J| =108$ in the case $n=6$) there is a way to immediately compute a list of at most $(n+1)\cdot |J|$ valid inequalities for $\tcf_{n}$ that contains a complete collection of facet representatives for $\tcf_{n}$ as a sublist (details omitted). This might get interesting if one wants to investigate $\tcf_{n}$ for $n \geq 7$ using knowledge about $\cut_{n+1}$.
\end{remark}
\section{Some open questions on the geometry of $\tcf_n$}\label{sect:openquestions}
Finally, we pursue some questions which arose while studying the convex polytope $\tcf_n$ that remained open to us. To this end, let $\psd_n \subset \RR^{E_n}$ be the space of symmetric and positive semi-definite $n\times n$ matrices in the sense of \eqref{eq:psd}. As mentioned in the introduction, it is well-known that all elements of $\tcf_n$ are positive semi-definite, that is
\begin{align*}
\tcf_n \subset \psd_n.
\end{align*}
It is natural to ask whether certain subsets of inequalities from facets of $\tcf_n$ imply already positive semi-definiteness. A simple candidate for such a question could be all facets at the exposed vertices $v_0=(0,0,\ldots,0)$ and $v_1=(1,1,\ldots,1)$ of $\tcf_n$. Let us denote the polytope which is defined by these facets by $\tcf_n(v_0,v_1)$. The following problem can be seen in a similar vein to Matheron's conjecture \citep{matheron_93}.
{\it
\begin{enumerate}[(F)]
\item For which values of $n$ does $\tcf_n(v_0,v_1) \subset \psd_n$ hold?
\end{enumerate}
}
\noindent Therefore, let us take a closer look at the facets of $\tcf_n$ at the exposed vertices $v_0$ and $v_1$. The facets at $v_{0}=(0,0,\ldots,0)$ are just the positivity inequalities $\chi_{ij}\geq 0$, which are hypermetric with $b = \mathbbm{1}_{\{i,j\}}$. To investigate the facets of $\tcf_n$ at $v_{1} = (1,1,\ldots,1)$, the following simple lemma is helpful.
\begin{lemma} \label{lemma:hypequalityat1} A hypermetric inequality given by $b \in \ZZ^n$ is satisfied as an equality by $v_{1}$ if and only if $\sum_{i=1}^n b_{i} \in \{0,1\}$.
\end{lemma}
\begin{proof}
$\sum_{1\leq i,j\leq n}b_{i}b_{j}\cdot 1=\sum_{i=1}^{n}b_{i}$ if and only if $(\sum_{i=1}^n {b_{i}})^2 = \sum_{i=1}^n {b_{i}}$.
\end{proof}
\noindent
A hypermetric inequality is \emph{pure hypermetric} if its corresponding $b$-vector satisfies $b \in \{-1,0,1\}^n$.
Using this lemma and inspecting Tables~\ref{table:hypermetricTCFfacets}, \ref{table:TCFsixfacets} and \ref{table:hypermetricTCFsixfacets} we derive the following proposition.
\begin{proposition}[Facets of $\tcf_n$ at $v_1$] \label{prop:facets_at1} $\phantom{a}$\\[-6mm]
\begin{enumerate}[a)]
\item For $n=2$ the (exceptional) facet at $v_{1}$ is pure hypermetric\\ with $\sum_{i=1}^n b_{i} = 0$.
\item For $3 \leq n \leq 5$ the facets at $v_{1}$ are pure hypermetric with $\sum_{i=1}^n b_{i} = 1$.
\item For $ n = 6$ the facets at $v_{1}$ are hypermetric with $\sum_{i=1}^n b_{i} = 1$.\\ Some are not pure.
\end{enumerate}
\end{proposition}
\noindent
Thus, the pure hypermetricity of the hypermetric facets at $v_{1}$ is another low-dimensional phenomenon: For $n=6$ there exist non-pure hypermetric facets at $v_1$. By the lifting property for $n \geq 3$ (Proposition~\ref{prop:zeroliftingfacets}), the same holds true for $n\geq 7$.
However, we may ask (\cf also Lemma~\ref{lemma:hypequalityat1}):
{\it
\begin{enumerate}[(G)]
\item[(G)] Are there non-hypermetric facets at $v_1$ for $n \geq 7\,$?\\
Are there hypermetric facets at $v_1$ with $\sum_{i=1}^n b_{i} = 0$ for $n \geq 7\,$?
\end{enumerate}
}
\noindent Let $\tcf^\hy_n(v_0,v_1)$, resp. $\tcf^\pu_n(v_0,v_1)$, be given by the hypermetric, resp. pure hypermetric, facets at $v_{0}$ and $v_{1}$.
\begin{remark}\label{remark:polytopv0v1}
By definition
$\tcf_n(v_0,v_1) \subset \tcf^\hy_n(v_0,v_1) \subset \tcf^\pu_n(v_0,v_1).$
All these sets are polytopes, since
positivity (= the facets of $\tcf_n$ at $v_0$) and triangle inequalities (which are certainly among the pure hypermetric facets of $\tcf_n$ at $v_1$ for $n\geq 3$) suffice already to imply $\tcf^\pu_n(v_0,v_1) \subset [0,1]^{E_n}$ (which also holds true for $n=2$), i.e., all of these sets are bounded and thus indeed polytopes. These polytopes are called {\it spindles}, since each facet contains one of the two vertices $v_{0},v_{1}$.
\end{remark}
\noindent The following proposition collects some partial answers to Question (F)
\begin{proposition}[Partial answers to Question (F)]
\label{prop:tcfpsd}$\phantom{a}$\\[-6mm]
\begin{enumerate}[a)]
\item For $2 \leq n \leq 5$ we have\\
$\tcf_n(v_0,v_1)=\tcf^\hy_n(v_0,v_1)=\tcf^\pu_n(v_0,v_1) \subset \psd_n$.
\item For $n=6$ we have $\tcf_6(v_0,v_1) = \tcf^\hy_6(v_0,v_1) \neq \tcf^\pu_6(v_0,v_1)$.
\item For $n \geq 6$ we have $\tcf^\hy_n(v_0,v_1) \not \subset \psd_n$.\\
In particular $\tcf_6(v_0,v_1) \not\subset \psd_6$.
\end{enumerate}
\end{proposition}
\begin{proof}
\begin{enumerate}[a)]
\item
The equalities follow from Table~\ref{table:hypermetricTCFfacets}. The inclusion $\tcf^\pu_n(v_0,v_1) \subset \psd_n$ has been solved by hand in \cite{strokorb_13} Proposition~3.6.5. for the cases $n \leq 4$. The idea for $n=4$ was to compute the extremal points of the polytope defined by positivity and triangle inequalities and to check p.s.d.\ for them. This suffices since $\psd_{n}$ is convex. For $n=5$ we used \texttt{polymake} to compute the extremal points of the polytope defined by positivity, triangle and pentagonal inequalities (see Table \ref{table:hypermetricTCFfacets}), and \texttt{R} to check p.s.d.
\item This follows from Proposition \ref{prop:facets_at1}.
\item For $n\geq6$ consider the point $x \in \RR^{E_n}$ with $x_{in}=0.5$, $1 \leq i \leq n-1$; $x_{ij}=0$ otherwise. Let $X$ denote the associated matrix.
For $b=(b_{1},\ldots,b_{n}) \in \ZZ^n$ and $s := \sum _{i=1}^n b_{i}$ we have
\begin{align*}
bXb^t = \sum_{i=1}^n b_{i}^2 + b_{n}\sum_{i=1}^{n-1}b_{i} = \sum_{i=1}^{n-1}b_{i}^{2} + b_{n}\cdot s.
\end{align*}
This shows $bXb^t \geq s$ for $s\in\{0,1\}$, for $s=1$ use $\sum_{i=1}^{n-1}b_{i}^{2} + b_{n} \geq \sum_{i=1}^{n-1}b_{i} + b_{n} = 1$. Thus, {all} hypermetric inequalities with $\sum_{i=1}^n b_{i} \in \{0,1\}$ are satisfied, in particular those at $v_{1}$ (c.f. Lemma \ref{lemma:hypequalityat1}), and $x$ is non-negative. On the other hand, $x$ is not positive semi-definite: For $a=(1,\ldots,1,-2)$ the above formula shows $aXa^t = (n-1) - 2(n-3)=5-n \leq -1$.
\end{enumerate}
\end{proof}
Thus, we expect ``if and only if $n \leq 5$'' to be the answer to Question F.\\
\noindent Our final question is motivated by the following observation.
Let $\hyp_{n}$ denoted the set of points $x=(x_{ij})_{1\leq i < j \leq n} \in \RR^{E_n}$ that satisfy all hypermetric inequalities (cf.\ Section~\ref{sect:basicfacts}).
\begin{lemma} \label{lemma:tcfhyp}
For all $n \geq 2$ the inclusions $\tcf_n \subset \hyp_n \subset \psd_n$ hold.
\end{lemma}
\begin{proof}
The first inclusion is a reformulation of Lemma \ref{lemma:hypervalid}. Now let $x \in \hyp_n$. By assumption we have $\sum_{1\leq i,j \leq n}b_{i}b_{j}x_{ij} \geq \sum_{i=1}^n b_{i}$ for all $b =(b_{1},\ldots,b_{n}) \in \ZZ^n$. This holds for $b$ and $-b$. Thus, $\sum_{1\leq i,j \leq n}b_{i}b_{j}x_{i,j} \geq 0$ for all $b \in \ZZ^n$. Division by integers extends this to $\QQ^n$, and continuity to $\RR^n$.
\end{proof}
\begin{remark}
By Proposition \ref{prop:nleq5hypermetric} the sets $\tcf_{n}$ and $\hyp_{n}$ even coincide for $n \leq 5$. This is no longer true for $n\geq 6$. A point $x \in \hyp_{6} \setminus \tcf_{6}$ is given by $x_{i,6} = 1/2, 1 \leq i \leq 5$, $x_{1,2}=x_{2,3}=x_{3,4}=x_{4,5}=x_{1,5}= 1/2$, and $x_{i,j} = 0$ otherwise. Indeed, since $(\sum_{i=1}^{5} x_{i,6}) - x_{1,3} - x_{3,5}-x_{2,5}-x_{2,4}-x_{1,4} = 2.5 > 2$, the point $x$ does not satisfy a permutation of the $\tcf_{6}$-facet from Proposition~\ref{prop:nonhypfacets}. We omit the computations showing $x \in \hyp_{6}$. By lifting, this extends to examples $x^0 \in \hyp_{n} \setminus \tcf_{n}$ for all $n\geq6$.
\end{remark}
{\it
\begin{enumerate}[(H)]
\item Do the inequalities of all hypermetric facets of $\tcf_n$ define a polytope, say $\tcf_{n}^{\hy}$, already contained in $\psd_n$?
\end{enumerate}
}
\noindent This holds true for $n \leq 5$ by Proposition \ref{prop:tcfpsd}, and remains open for $n \geq 6$. Note that $\tcf_{n}^{\hy}$ is a polytope by $\tcf_{n}^{hyp} \subset \tcf_{n}^{hyp}(v_{0},v_{1})$ and Remark \ref{remark:polytopv0v1}.
\section*{Discussion}\label{sect:discussion}
In this article, we deal with the realization problem for the tail correlation function (TCF), which is an omnipresent bivariate tail dependence measure in the extremes literature.
We make this specific by formulating Questions~(A)-(E) in the introduction.
Here, we discuss our contribution to these questions.
In doing so we address Questions~(A)-(E) partially in reversed order according to their growing complexity.
Questions~(E) and (D) can be answered fully and affirmatively by
Corollary~\ref{cor:basicoperations} and Theorem~\ref{thm:TCFisMAX}, respectively. That is, convex combinations, products and pointwise limits are admissible operations on the set of TCFs and Theorem~\ref{thm:TCFisMAX} shows that the class of TM processes, a subclass of max-stable processes, is rich enough to realize any given TCF. Concerning the regularity of the corresponding TM process, we identify continuity of its TCF as a necessary and sufficient condition for its stochastic continuity (Corollary~\ref{cor:chicty}), which contributes to Question~(C). Theorem~\ref{thm:TCFisMAX} also opens up links to binary ($\{0,1\}$-valued) processes and thereby provides a substantial reduction of Questions~(A) and (B). Corollary~\ref{cor:TCFextension} reduces them even further to the study of TCFs on finite base spaces. Together with Corollary~\ref{cor:finitepolytope}, we reveal that membership in the set of TCFs (even on infinite spaces) can be completely characterized by a system of affine inequalities, which -- if known -- would provide a complete answer to Question~(A).
To identify and classify these affine inequalities, a better understanding of the geometry of the polytope $\tcf_n$ of $n\times n$ tail correlation functions (matrices) for arbitrary $n$ is needed. Its facet inducing inequalities constitute such a list (actually, an $\mathcal{H}$-representation would suffice already).
Lemma~\ref{lemma:hypervalid} contributes to Question~(A) in that it provides a rich class of necessary conditions (all hypermetric inequalities) for membership in $\tcf_n$, whereas Proposition~\ref{prop:cpp} identifies any clique partition point to be an admissible TCF. In Section~\ref{sect:corcut}, we discuss that the polytope $\tcf_n$ can be viewed either as an affine \emph{projection} of the polytope $\Theta_n$ (whose facets are well-understood) or as an affine \emph{intersection} with the \emph{correlation polytope} (whose vertices are well-understood). Both views immediately suggest algorithms that can be easily implemented in order to obtain the vertices and facets of $\tcf_n$ that in theory would work ``for arbitrary $n$''. This would solve Question~(A) computationally.
Due to the complexity of the problem, software computations lead to a full description of facets and vertices of $\tcf_n$ only up to $n=6$ (Section~\ref{sect:compresults}).
Indeed, several of our results reveal the rapidly growing complexity of Question~(A) as $n$ grows.
Starting from $n=3$, no facet inducing inequality of $\tcf_n$ will ever become obsolete (Proposition~\ref{prop:zeroliftingfacets}).
For instance, the triangle inequality $(\ref{eqn:triangle})$ cannot be deduced from any other set of valid inequalities for $\tcf_n$.
By contrast, \emph{all} facet inducing inequalities that define the polytope of ECFs $\Theta_{n}$ become obsolete for $\Theta_{n'}$ for higher $n'>n$,
and still $\Theta_n$ has $2^n$ facets in dimension $n$.
Starting from $n \geq 6$ there exist (actually plenty of) non-hypermetric facets of $\tcf_{n}$ (Proposition~\ref{prop:nonhypfacets}).
Moreover, we derived the facets of $\tcf_6$ from the facets of the cut polytope $\cut_7$ which had 11 generators for 116$\phantom{.}$764 facets. The next step would take into account the polytope $\cut_{8}$, which has already more than 217 million facets which can be subdivided into 147 orbits under permutations and switchings \citep[p.~505]{dezalaurent_97}.
It is even possible to choose $n$ sufficiently large, such that a given finite set of rational numbers from the interval $[0,1]$ turns up as coordinate values of a single vertex of $\tcf_n$ (Proposition~\ref{prop:unbounded}).
Altogether, these results confound the aim of a full answer to Question~(A).
Finally, if Question~(A) is already so difficult to answer, what more can be eventually said about Question~(B)? That is, given a TCF $\chi$, say on a finite space, how to construct a specific stochastic model that realizes $\chi$? Again, from our dual views on $\tcf_n$ as affine ``projection of'' or ``intersection with'' other polytopes, it is easy to formulate naive ad-hoc algorithms providing an entire convex polytope of solutions to such a problem, cf.\ \cite{strokorb_13}, p.~65. Perhaps more interestingly, in case of $T=\RR^d$, \cite{strokorbballanischlather_15} characterize subclasses of radially symmetric and monotonously decreasing TCFs with some sharp bounds on membership in the class of TCFs on $\RR^d$ (cf.\ Table 2 therein) and recover realizing max-stable models. Surprisingly often, it is possible to obtain explicitly several such realizing models sharing an identical TCF, but with rather different spectral profiles.
In this sense, the reader should not overrate the finding of a specific model meeting a given TCF even though the TM models helped us here to approach the realization problem.
To conclude with,
independently of our research \citep{fss_14} and motivated from an insurance context, \cite{ehw15} dealt with almost the same questions (in particular Questions A and B) for random vectors with an emphasis on the construction of realizing copulas as we learned on the EVA 2015 in AnnArbor.
Our approach offers (at least theoretically) an algorithm that can solve Questions A and B for random vectors completely (even though the feasibilty of such an algorithm breaks down very quickly as the dimension grows and we have doubts on its practical use in higher dimensions). This answers one of the questions raised in the discussion of \cite{ehw15}.
\paragraph{Acknowledgements}
The authors would like to thank two referees and an AE who provided many thoughtful comments leading to substantial improvements in the presentation of this material. We are also thankful to be made aware of the regularity question and the early works of Shepp on unit covariances.
K.~Strokorb undertook part of this work as part of her PhD thesis \cite{strokorb_13} as a member of the Research Training Group 1023 and gratefully acknowledges financial support by the German Research Foundation DFG.
{\small
\bibliographystyle{spbasic}
|
\section{Introduction}
A gauge invariant regularization method for Wilson loop variables appears
to be an unavoidable necessity for
construction of quantized Yang-Mills fields. The standard methods
of regularizing a quantum field, that have been successful in studying scalar field theories,
are inapplicable to gauge fields.
Thus
a simple weighted average $\int_{\mathbb R^3} f(x-y) A(y) d^3y$ destroys gauge invariance
of the gauge potential $A$. Similar expressions, such as $\int_{R^n} f(x-y) F(y) d^n y$,
with $F(y)$ the curvature of $A$, also destroy gauge invariance,
both for a space average, with $n=3$, or a Euclidean space-time average, with $n=4$.
For a closed curve $C$ in $R^3$
the Wilson loop variable, $W_C(A) \equiv trace\, T(\exp(\int_C A(x)\cdot dx))$,
where $T$ denotes time ordering around the loop, is gauge invariant but highly singular
as a function of $A$ when $A$ varies over the very large space of typical gauge fields required
in the quantized theory.
The lattice regularization of
these functions of the gauge fields has been the only useful gauge invariant
regularization procedure so far but has not produced a continuum limit.
Polyakov \cite{Pol1,Pol2} already observed that the vacuum expectation
of continuum Wilson loop variables are likely to be zero for a non-commutative gauge group.
They are zero in the electromagnetic case.
Nevertheless
it has been hoped that the informal symbol, defined as $W_C(A)/ \<W_C\>_{VEV}$,
which nominally is identically infinite in absolute value, could play a central role in
a gauge invariant formulation of some future internally consistent quantized Yang-Mills theory.
Such a program was outlined by E. Seiler, \cite[Pages 163-181]{Sei}.
Many steps toward carrying this out were made by a renormalization group approach
in a series of papers by T. Balaban.
See e.g. \cite{Bal}.
In \cite{CG1} we began a regularization program based on use of the Yang-Mills heat
equation for regularizing gauge fields. The magnetic energy of a classical gauge field
$A$ over three dimensional space is $\int_{R^3}|B(x)|^2 d^3x$, where $B$ is the magnetic field
($\equiv$ curvature) of $A$. The Yang-Mills heat equation
flows $A$ in the direction of the negative of the gradient of the magnetic energy.
It is a non-linear, weakly parabolic equation with difficulties of its own.
But it is fully gauge invariant: if one transforms the initial data $A_0$ by a gauge transformation
on $\mathbb R^3$ and then propagates, one arrives at the same gauge field as if one first propagates
$A_0$ and then gauge transforms. Moreover the flow regularizes the initial data
well enough so that
the Wilson loop function $W_C(A(s))$ is meaningful for any fixed time $s >0$,
even when $W_C(A_0)$ itself is meaningless. Most importantly, $W_C(A(s))$
is gauge invariant under gauge transforms of the initial data $A_0$.
Here $A(s)$ is the solution to the Yang-Mills heat flow equation at time $s$.
The Yang-Mills heat flow has also been used for regularization as part of a
method for implementing a Monte Carlo computational protocol for lattice gauge theory,
\cite{L1,L2,L3,LW} .
Like other heat equations, the Yang-Mills heat equation propagates information
instantly. This would cause problems for local quantum field theory because
one wants $W_C(A(s))$ to capture information about $A_0$ just in a small neighborhood
of the curve $C$, not over all of $\mathbb R^3$. This issue can be resolved by using
the Yang-Mills heat equation regularization over a bounded open
set $M$ in $\mathbb R^3$ that contains $C$.
For this procedure one must prove existence and uniqueness of the solution
when the initial data is specified only in $M$. Of course for uniqueness one needs then to specify
boundary conditions on
the solution $A(s)$ for $ s >0$.
These in turn
must be gauge invariant and must allow use of initial data which are the restrictions to $M$
of a typical gauge field $A_0$ on $\mathbb R^3$. The classical Neumann and Dirichlet boundary
conditions will be vital boundary conditions for us for technical use. But in the end
Marini boundary conditions, which simply set the normal
component of
the magnetic field $B(s)$ to zero on the boundary of $M$, are the only ones that
are fully gauge invariant. We will explore all three boundary conditions in this paper.
We used, in \cite{CG1}, the Zwanziger-Donaldson-Sadun
\cite{Z}, \cite{Do1}, \cite{Sa} method for proving existence
of solutions to the Yang-Mills heat equation, which consists of adding a gauge symmetry breaking
term to the equation and then removing it from the solution by gauge transformation.
The ZDS procedure does not enter directly into the present paper since our goal
is to establish further properties of a solution whose existence we already know.
Instead we will use the fact that the absolute values $|B(s, x)|$ and $|(d/ds) A(s, x)|$ satisfy
parabolic inequalities with Neumann-like boundary conditions.
Our goal is to get detailed information about the behavior of these two
functions as $s\downarrow 0$
in order to help pass, eventually, to more general initial data.
Some of the initial steps in this technique will be carried out
over a compact manifold with boundary
rather than just over a bounded open set in $\mathbb R^3$ because they provide illumination
as to what the techniques depend on, and there is little extra cost.
In \cite{CG1} we established existence and uniqueness of solutions
in case the initial data $A_0$ is in the
Sobolev space $H_1(M)$. This corresponds to initial data of finite magnetic energy.
In order to get this program to work we anticipate that it will be necessary
to extend the results in \cite{CG1} so as to allow the initial data to lie in the
larger space $H_{1/2}(M)$, which corresponds to initial data of finite magnetic action.
In the present paper we will still focus on initial data in $H_1$. However this is already
broad enough to include gauge fields that need to be regularized before their
Wilson loop functional can be defined.
We will give an example in Section \ref{secLTB} of a current distribution in $\mathbb R^3$ whose magnetic
field has finite energy but nevertheless gives infinite magnetic flux through certain loops,
rendering the Wilson loop functional for these loops meaningless.
Although our main concern is the behavior of the solution for small time,
we are also going to prove that the
Wilson loop functions $W_C(A(s))$ converge as $s \rightarrow \infty$ for any initial
gauge potential $A_0$ in $H_1$.
\section{Neumann domination} \label{secNdom}
\begin{notation} \label{not2.1}{\rm $M$ will denote a compact
Riemannian 3-manifold with smooth boundary. We will be concerned
with a product bundle $M\times {\mathcal V} \rightarrow M$, where ${\mathcal V}$ is a finite dimensional
real or complex vector space with an inner product.
$K$ will denote
a compact connected subgroup of the orthogonal, respectively, unitary group of
the space $End\ {\mathcal V}$, of operators on ${\mathcal V}$ to ${\mathcal V}$.
The Lie algebra of $K$, denoted $\frak k$, may then be identified with a
real subspace of $End\ {\mathcal V}$.
We denote by $\<\cdot, \cdot\>$ an $Ad\ K$
invariant inner product on $\frak k$ and denote its associated norm
by $|\xi |_{\frak k}$ for $\xi \in \frak k$. We will not distinguish between
$|\xi|_{\frak k}$ and $|\xi|_{End {\mathcal V}}$, which are equivalent norms.
If $\omega$ and $\phi$ are $\frak k$ valued $p$-forms define
$(\omega, \phi) = \int_M\<\omega(x), \phi(x)\>_{\Lambda^p\otimes \frak k} d\, \text{Vol}$
and $\|\omega \|_2^2 = (\omega, \omega)$.
Define also
$\|\omega\|_\infty = \sup_{x \in M}|\omega(x)|_{\Lambda^p \otimes \frak k}$ and
\begin{equation}
\| \omega \|_{W_1(M)}^2
= \int_M |\nabla \omega|_{\Lambda^p\otimes\frak k}^2 d\, \text{Vol}\ \ + \| \omega \|_2^2 \label{ymh2}
\end{equation}
where $\nabla$ is the Riemannian gradient on forms and $\nabla \omega$ refers to
the weak derivative. Define
$W_1 = W_1(M) = \{ \omega: \|\omega\|_{W_1(M)} <\infty\}$.
Since we are concerned only with a product bundle, a connection form
can be identified with a $\frak k$ valued 1-form.
For a connection form $A$, given in local coordinates by
$A = \sum_{j=1}^3 A_j(x) dx^j$, its curvature (magnetic field) is given by
\begin{equation}
B = dA +(1/2) [A\wedge A] \label{ymh3}
\end{equation}
where $[A\wedge A] = \sum_{i,j} [A_i, A_j] dx^i\wedge dx^j$ and
$[A_i(x), A_j(x)]$
is the commutator in $\frak k$. $B$ is a $\frak k$ valued 2-form.
For $\omega \in W_1$
we define $d_A \omega = d \omega + (ad\ A) \wedge \omega$ and
$d_A^* \omega = d^*\omega + (ad\ A\wedge)^* \omega$. Boundary conditions
will be imposed on these operators later.
}
\end{notation}
We recall from \cite{CG1} the definition of a strong solution of the
Yang-Mills heat equation.
\begin{definition}\label{defstrsol} {\rm Let $0 < T \le \infty$.
By a {\it strong solution} to the Yang-Mills heat
equation over $[0, T)$
we mean a continuous function
\begin{equation}
A(\cdot): [0,T) \rightarrow W_1 \subset \frak k\text{-valued 1-forms}
\end{equation}
such that
\begin{align}
a)& \ B(t) \in W_1 \ \text{for each}\ \ t\in (0,T),\
\text{where}\ B(t) = \text{ curvature of}\ A(t), \label{ymh8}\\
b)& \ \text{the strong $L^2(M)$ derivative $A'(t) \equiv dA(t)/dt $}\
\text{exists on}\ (0,T), \nolinebreak \label{ymh9} \\
c)& \ A'(t) = - d_{A(t)}^* B(t)\ \ \text{for each}\ t \in(0, T). \label{ymh10}
\end{align}
A strong solution will be called {\it locally bounded} if
\begin{align}
&d) \ \| B(t)\|_\infty\ \text{is bounded on each
bounded interval $ [a,b) \subset (0, T)$ and}\ \label{ymh11}\\
&e) \ t^{3/4} \| B(t)\|_\infty\
\text{is bounded on some interval $(0, b)$ with $0<b <T$.} \label{ymh12}
\end{align}
}
\end{definition}
\bigskip
\noindent
We are interested in three kinds of boundary conditions.
\bigskip
\noindent
{\it Neumann boundary conditions:}
\begin{align}
&i)\ \ \ A(t)_{norm} =0\ \ \text{for}\ t \ge 0 \ \text{and} \label{N1}\\
&ii)\ \ B(t)_{norm}=0\ \ \text{for}\ t >0. \label{N2}
\end{align}
\noindent
{\it Dirichlet boundary conditions:}
\begin{align}
&i)\ \ \ A(t)_{tan} =0\ \ \ \text{for}\ t \ge 0 \ \text{and} \label{D1}\\
&ii)\ \ B(t)_{tan} =0 \ \ \ \text{for}\ t >0. \label{D2}
\end{align}
{\it Marini boundary conditions:}
\begin{equation}
B(t)_{norm}=0\ \ \text{for}\ \ t >0. \label{M1}
\end{equation}
Given a solution of the Yang-Mills heat equation, \eqref{ymh10},
we are going to make pointwise estimates of
$|B(s,x)|_{\Lambda^2\otimes \frak k}$ and
$|A'(s,x)|_{\Lambda^1\otimes \frak k}$
based on parabolic inequalities that these functions satisfy
for the Neumann Laplacian on real valued functions over $M$.
The final step in our method will require that $\partial M$ be convex in the sense that the
second fundamental form be non-negative on $\partial M$.
\subsection{Sub-Neumann boundary conditions} \label{secNdom1}
In this section $M$ will denote a compact, $n$-dimensional, Riemannian
manifold with a smooth, not necessarily convex, boundary.
\begin{proposition} \label{propsubNeu}
$($Sub-Neumann boundary conditions.$)$
Denote the extended shape operator by $Q(x) \in End( \Lambda (T^* ( \partial M))$
$($the extension by derivation of the adjoint of the usual shape operator.
See e.g. {\rm \cite[Notation 4.6]{CG1}$)$.}
Denote by $\nabla_{\bf n}$
the outward drawn normal derivative operator.
Let $A$ be a continuous $\frak k$ valued 1-form on $M$ and let $\omega$ be a $\frak k$ valued $p$-form on $M$ of class $C^1$.
a$)$ If
\begin{equation}
\omega_{norm} = 0 \ \text{and}\ (d_A \omega)_{norm} =0 \label{neu50}
\end{equation}
then
\begin{equation}
\nabla_{\bf n} | \omega|^2 = -2\< \{I_{\frak k} \otimes Q\} \omega, \omega\>\
\ \text{on}\ \ \partial M. \label{neu51}
\end{equation}
b$)$ If
\begin{equation}
\omega_{tan} =0 \ \ \text{and}\ \ (d_A^* \omega)_{tan} = 0 \label{neu52}
\end{equation}
then
\begin{equation}
\nabla_{\bf n} | \omega|^2 = -2\< \{I_{\frak k} \otimes(*^{-1} Q*)\} \omega, \omega\>\ \
\text{on}\ \ \partial M. \label{neu53}
\end{equation}
\end{proposition}
\begin{proof} In a neighborhood $U$ of the boundary choose
an adapted coordinate system $(U, x^1,\dots, x^n)$. See e.g. \cite[Notation 4.2]{CG1}.
We can write a $\frak k$ valued $p$-form in $U$ as
\begin{equation}
\omega = \beta + \gamma \wedge dx^n \label{neu54}
\end{equation}
with $\beta = \sum_{J} \beta_J dx^J$ and
$\gamma = \sum_{I} \gamma_I dx^I$.
The multi-indices will signify generically $J =(j_1,\dots, j_p)$
with $ j_1<\cdots < j_p < n$ and
$I = (i_1, \dots, i_{p-1})$ with $ i_1 < \cdots < i_{p-1} < n$.
Since
$\< dx^j, dx^n\> = 0$ for all $j <n$ we have
$$
|\omega(x)|^2 = |\beta(x)|^2 + | \gamma(x) \wedge dx^n|^2 \ \ \text{in}\ U.
$$
Suppose first that $\omega_{norm} =0$ on $U\cap \partial M$.
That is, $\gamma|(U\cap \partial M) =0$. Then, writing $\partial_j^A = \partial/\partial x^j +ad\ A_j$
on $\frak k$ valued functions, and $\nabla_j^A = \nabla_j +ad\ A_j$ for the Riemann covariant derivative on $\frak k$ valued forms, we have, at $U\cap \partial M$,
\begin{align}
(1/2) \partial_n |\omega(x)|^2
&= \< \nabla_n^A \beta(x), \beta(x) \>
+ \< \nabla_n^A( \gamma(x) \wedge dx^n), \gamma(x) \wedge dx^n \> \notag \\
&= \< \nabla_n^A \beta(x), \beta(x)\>, \label{neu55}
\end{align}
because $\gamma(x) = 0$ on $U\cap \partial M$.
Now in $U$,
\begin{equation}
\nabla_n^A \beta = \sum_J (\partial_n^A \beta_J) dx^J
+ \sum_J \beta_J(\nabla_n dx^J). \notag
\end{equation}
But
$\sum_J \beta_J(\nabla_n dx^J)
=-\sum_J \beta_JQ dx^J
= - (I_{\frak k} \otimes Q) \beta$ at $U\cap \partial M$.
So
\begin{align}
\nabla_n^A \beta = \sum_J(\partial_n^A \beta_J) dx^J
- (I_{\frak k} \otimes Q) \beta\ \ \text{at}\ \ U\cap \partial M. \label{neu56}
\end{align}
Further,
\begin{align}
d_A(\gamma \wedge dx^n)
& = (d_A \gamma)\wedge dx^n \notag \\
&= \sum_{j=1}^{n-1}
\sum_I(\partial_j^A \gamma) dx^j\wedge dx^I \wedge dx^n \notag \\
&=0 \ \ \text{on}\ \ U\cap \partial M , \label{neu56.1}
\end{align}
because $\gamma$, as well as any tangential derivative of $\gamma$, is zero
when $\omega_{norm} =0$.
Hence, if both equations in \eqref{neu50} hold, then, by \eqref{neu54}
and \eqref{neu56.1}, we find
\begin{align*}
0 = (d_A\omega)_{norm} &= (d_A \beta)_{norm}\\
&= \sum_J (\partial_n^A \beta_J) dx^n \wedge dx^J.
\end{align*}
Thus $ \partial_n^A \beta_J =0$ on $U\cap \partial M$ and consequently \eqref{neu56}
reduces to
\begin{equation}
\nabla_n^A \beta = - (I_{\frak k} \otimes Q) \beta
\ \ \text{on}\ \ U\cap \partial M. \label{neu57}
\end{equation}
Therefore \eqref{neu55} yields
$(1/2) \partial_n | \omega |^2 = - \<(I_{\frak k} \otimes Q) \beta , \beta\>
= -\<(I_{\frak k} \otimes Q)\omega, \omega \>$,
which is \eqref{neu51} since $\nabla_{\bf n} = \partial_n$ in this coordinate system.
In the last step we have used $\beta = \omega$ on $U\cap \partial M$.
To prove the assertion in case b) one need only observe that
if \eqref{neu52} holds for $\omega$ then \eqref{neu50} holds for $* \omega$.
Consequently
\begin{align*}
\partial_n | \omega|^2 & = \partial_n |*\omega|^2 \\
&= -2 \< (I_{\frak k} \otimes Q) *\omega, , * \omega \>,
\end{align*}
which is \eqref{neu53}.
\qed
\end{proof}
\begin{corollary} \label{corsubneu} In addition to the hypotheses of Proposition
\ref{propsubNeu}, suppose
that $M$ is convex in the sense that its
second fundamental form is everywhere non-negative on $\partial M$.
If \eqref{neu50} or \eqref{neu52} hold then
\begin{equation}
\nabla_{\bf n} | \omega |^2 \le 0. \label{neu58}
\end{equation}
If, in addition, $\partial M$ is totally geodesic then
\begin{equation}
\nabla_{\bf n} |\omega|^2 =0 . \label{neu59}
\end{equation}
\end{corollary}
\begin{proof}
If $M$ is convex then $Q(x) \ge 0$ on $\partial M$ as is its unitary
transform $*^{-1} Q(x) *$.
The inequality \eqref{neu58} now follows from \eqref{neu51} and \eqref{neu53}.
If $\partial M$ is totally geodesic then $Q(x) =0$ on $\partial M$ as is
its transform $*^{-1} Q(x) *$. \eqref{neu59} now follows in the same way.
\end{proof}
\begin{Remark}
{\rm In the context of a Riemannian $n$-manifold without further vector
bundle structure, the first author found, in \cite{Cha4}, that the Neumann boundary condition \eqref{neu59} follows from either of the
boundary conditions $(a) \ \omega_{norm} = 0$ and $(d \omega)_{norm} =0$ or
$(b) \ \omega_{tan} =0$ and $(d^* \omega)_{tan} = 0$, in the presence of a slightly weaker
condition on the boundary than used in this paper. Namely, it was shown that if
$\omega$ is an $(n-1)$-form then it suffices that the trace of the second fundamental form be
zero. But for lower order forms the condition that the boundary be totally geodesic was needed. In the present paper the weakened hypothesis would be applicable, in case dim $M =3$,
to the curvature $B$ but not to $A'$.
In \cite{Cha4}, in addition to the Hodge Laplacian, the first author considered the
Bochner Laplacian and showed that, for the relevant notion of Dirichlet boundary
condition on the $k$-form $\omega$, no conditions on the boundary are needed to conclude
\eqref{neu59}.
}
\end{Remark}
\subsection{Domination by the Neumann heat kernel} \label{secNdom2}
In this section $M$ will denote the closure of a bounded open subset of $\mathbb R^n$
with smooth boundary.
$\Delta$ will denote the Laplacian on real valued functions on $M$
with domain $W_2(M)$ and $\Delta_N$ will denote the Neumann version.
Some aspects of the techniques we are exploring have been used
for manifolds without boundary in \cite{Do1} and \cite{Sa} for the Yang-Mills heat equation.
\begin{lemma}\label{lemNdom1}
Let $\psi$ be a real valued function in $W_2(M)$ whose outer
normal derivative satisfies
\begin{equation}
\nabla_{\bf n} \psi \le 0 \ \ \ \text{a.e. on }\ \partial M. \label{neu60}
\end{equation}
Then
\begin{equation}
e^{t\Delta_N} \Delta \psi
\le \Delta_N e^{t\Delta_N} \psi\ \ a.e. \ \ \text{for all}\ \ t>0. \label{neu61}
\end{equation}
\end{lemma}
\begin{proof}
If $ 0 \le f \in {\mathcal D}(\Delta_N)\cap C^\infty(M)$ then $\nabla_{\bf n} f =0$ on $\partial M$.
Hence
\begin{align*}
(\Delta \psi, f) & = \int_M (div\ grad\ \psi) f dx \\
& = \int_{\partial M} ( {\bf n}\cdot grad\ \psi) f - \int_M\< (grad\ \psi), (grad\ f) \> dx\\
&\le - \int_M\< (grad\ \psi), (grad\ f) \> dx,
\end{align*}
in view of \eqref{neu60}.
Since $f \in {\mathcal D}(\Delta_N)$ we may integrate by parts once more to find
\begin{equation}
(\Delta \psi, f) \le ( \psi, \Delta_N f). \label{neu65}
\end{equation}
Now let $ 0 \le \phi \in C_c^\infty(M^{\text{int}})$. We may
apply \eqref{neu65} to $f \equiv e^{t\Delta_N} \phi$ because
$e^{t\Delta_N}$ is positivity preserving.
It follows that
$$
( e^{t\Delta_N} \Delta \psi, \phi) = (\Delta \psi, e^{t\Delta_N} \phi )
\le (\psi, \Delta_N e^{t\Delta_N} \phi) = (\psi, e^{t\Delta_N}\Delta_N \phi).
$$
Hence
$
( e^{t\Delta_N} \Delta \psi, \phi) \le ( \Delta_N e^{t\Delta_N} \psi, \phi)
$
for all non-negative $\phi \in C_c^\infty(M^{\text{int}})$.
This proves \eqref{neu61}.
\end{proof}
\begin{proposition}\label{propDomination}
Suppose that $M$ is convex in the sense that the
second fundamental form is non-negative on $\partial M$.
Let $T>0$. Suppose that
$A(\cdot) : [0,T) \rightarrow C^1( M; \Lambda^1\otimes \frak k)$
is a time dependent, 1-form on $M
which is continuous in the time variable.
Let $\omega(\cdot):[0,T) \rightarrow C^2( M; \Lambda^p\otimes \frak k)$
be a time dependent, $\frak k$ valued, $p$-form on $M
which is continuously differentiable in the time variable and
satisfies the equation
\begin{equation}
\omega'(s, x) = \sum_{j=1}^n (\nabla_j^{A(s)} )^2 \omega(s,x) + h(s,x), \label{5.51}
\end{equation}
where $h \in C([0,T)\times M; \Lambda^p\otimes \frak k)$.
Assume also that $\omega$ satisfies either the boundary conditions
\begin{equation}
\omega(s)_{norm} =0, \ \ \text{and}\ \ (d_{A(s)} \omega(s))_{norm} =0
\ \ \text{for all}\ \ s\in [0,T) \label{5.53}
\end{equation}
or
\begin{equation}
\omega(s)_{tan} = 0 \ \ \text{and} \ \ (d_{A(s)}^* \omega(s))_{tan} = 0
\ \ \text{for all}\ \ s\in [0,T). \label{5.54}
\end{equation}
Then, for all $(t,x) \in [0,T)\times M$, there holds
\begin{equation}
|\omega(t, x)| \le \{e^{t\Delta_N} |\omega(0)|\}(x)
+ \int_0^t \{e^{(t-s)\Delta_N} |h(s)|\}(x) ds . \label{5.55}
\end{equation}
Here the norm denotes $|\cdot |_{\Lambda^p\otimes \frak k}$.
\end{proposition}
\begin{proof}
Given $\omega$ as specified, let $\epsilon >0$ and define
\begin{equation}
\psi(s,x) = ( |\omega(s,x)|^2 + \epsilon^2)^{1/2}. \label{5.41}
\end{equation}
For fixed $s$ (and suppressing $s$) we assert that
\begin{equation}
\<\sum_{j=1}^n (\nabla_j^A)^2 \omega(x), \omega(x)\>
\le \psi(x) (\Delta \psi)(x)\ \text{for all}\ x \in M^{int}. \label{5.43}
\end{equation}
The proof of this well known pointwise inequality follows a standard
pattern and does not depend on the boundary conditions.
Thus for any real valued function $\psi \in C^2(M)$ one verifies easily
the identity
\begin{equation}
\psi(x) \Delta \psi (x)
= (1/2) \Delta \psi^2(x) - |grad\, \psi(x)|^2,\ \notag
\end{equation}
and then, with $\psi$ defined now by \eqref{5.41}, one computes, for each $s$, that
\begin{equation}
(1/2) \Delta \psi^2(x) = (1/2)\Delta |\omega(x)|^2
= \< \sum_{j=1}^n (\nabla_j^A)^2 \omega(x), \omega(x) \>
+ \sum_{j=1}^n |\nabla_j^A \omega(x)|^2 . \notag
\end{equation}
But
$
|\partial_j \psi(x)| = |(1/2)(\partial_j |\omega(x)|^2)|/\psi(x)
= \<\nabla_j^A \omega(x), \omega(x)\>/\psi(x) \le |\nabla_j^A \omega(x)|.
$
Combining this with the previous two equations
yields \eqref{5.43}.
Suppose now that $\omega(s,x)$ satisfies the differential equation \eqref{5.51}.
Take the pointwise inner product of \eqref{5.51} with $\omega(s, x)$ to find,
with the help of \eqref{5.43},
\begin{align*}
\psi(s, x) \psi'(s,x) &= (1/2) (d/ds) \psi^2(s, x)\\
& = \< \omega'(s,x), \omega(s,x)\>\\
& = \<\sum_{j=1}^3 (\nabla_j^{A(s)} )^2 \omega(s,x) + h(s,x), \omega(s,x)\>\\
&\le \psi(s,x) \Delta \psi(s, x) + |h(s,x)| |\omega(s,x)|.
\end{align*}
Divide by $\psi(s,x)$ to deduce
\begin{equation}
\psi'(s,x) \le \Delta \psi(s,x) +|h(s,x)| \ \text{for all}\ x \in M^{int}. \label{5.53a}
\end{equation}
In view of \eqref{5.53} and \eqref{5.54}, it follows from
Corollary \ref{corsubneu} that
$\nabla_{\bf n} \psi^2= \nabla_{\bf n} |\omega|^2 \le 0$ on $\partial M$.
And, since $\psi \ge \epsilon >0$ on $M$, it follows that
\begin{equation}
\nabla_{\bf n} \psi \le 0\ \ \text{on}\ \ \partial M. \label{5.53b}
\end{equation}
Let $\phi(s,x) = \psi(s,x) - \epsilon$. Then \eqref{5.53a} and \eqref{5.53b} hold
also with $\psi$
replaced by $\phi$. Thus
$\phi'(s,x) - \Delta \phi(s,x) - |h(s, x)| \le 0$ for each $x \in M^{int}$ and moreover
$\nabla_n \phi \le 0$ on $\partial M$.
For $0 \le s \le t <T$ define at each $x\in M$ (and suppressing $x$)
$$
u(s) = e^{(t-s)\Delta_N} \phi(s)
+ \int_s^t e^{(t-\sigma)\Delta_N} |h(\sigma)| d \sigma.
$$
Then, by virtue of Lemma \ref{lemNdom1},
\begin{align*}
(d/ds)u(s) &= -\Delta_N e^{(t-s)\Delta_N}\phi(s)
+e^{(t-s)\Delta_N} \{ \phi'(s) - |h(s)|\}\\
& \le e^{(t-s)\Delta_N}\{ -\Delta \phi(s) + \phi'(s) - |h(s)|\}\\
& \le 0,
\end{align*}
wherein we have used once more the fact that $e^{t\Delta_N}$ is positivity preserving
for $t\ge 0$. Thus $u(t) \le u(0)$. That is,
\begin{equation}
\phi(t)
\le e^{t \Delta_N} \phi(0)
+\int_0^t e^{(t-\sigma)\Delta_N} |h(\sigma)| d \sigma. \label{5.66}
\end{equation}
Now observe that $ 0 \le \phi(t,x) \le |\omega(t,x)|$ and
$\lim_{\epsilon\downarrow 0} \phi(s,x) = |\omega(s,x)|$ for all $s$ and $x$.
Using the dominated convergence theorem on the first term on the right of \eqref{5.66} (which is an integral of a heat kernel), we may now let $\epsilon\downarrow 0$ in \eqref{5.66} to arrive at \eqref{5.55}.\qed
\end{proof}
\begin{Remark} \label{remHunSim} {\rm If, in Proposition \ref{propDomination},
one assumes that $A$ is independent of time and that $h \equiv 0$ then the inequality
\eqref{5.55} asserts that
\begin{equation}
|e^{t\Delta^A} \omega(0) | \le e^{t\Delta_N} |\omega(0)|, \label{5.68}
\end{equation}
where $\Delta^A$ is the gauge covariant Laplacian on $\frak k$ valued
$p$-forms ($p \ge 1$)
associated to either relative or absolute boundary conditions and
$\Delta_N$ is the Neumann Laplacian on real valued functions.
This is a diamagnetic inequality (see \cite[Section 1.3]{CFKS}) for a region with boundary.
It seems quite feasible to derive our results from such an inequality
by writing the time dependent propagator as a limit of short time propagators
for $A(t)$ with different $t$. This would entail some regularity on the $t$
dependence of $A(t, \cdot)$. We have not explored this approach.
For a recent paper extending and reviewing diamagnetic inequalities
of the form \eqref{5.68} when $\frak k$ is abelian see \cite{HS}.
}
\end{Remark}
\subsection{Pointwise bounds on solutions} \label{secNdom3}
Henceforth $M$ will denote the closure of a bounded open set in $\mathbb R^3$
with smooth boundary. We will assume $M$ to be convex in the sense
that its second fundamental form is everywhere non-negative. For the Neumann
heat operator $e^{t\Delta_N}$ over $M$, the constant
\begin{equation}
c_N = \sup_{0<t \le 1} t^{3/4} \|e^{t\Delta_N}\|_{2\rightarrow \infty} \label{AAhk}
\end{equation}
is finite, \cite[page 274]{Tay3}. As in \cite{CG1}, we will take
$c:= \sup\{ |\,[\xi, \eta]\,|_\frak k: |\xi|_\frak k \le 1, |\eta|_\frak k \le 1\}$ as a measure of the non-commutativity of $\frak k$.
\begin{theorem}\label{thmAA} There exist strictly positive
constants $a$ and $\gamma$ such that
for any number $\tau \in (0, 1/2]$ and any smooth solution $A(\cdot)$ to
the Yang-Mills heat equation \eqref{ymh10} over the interval $[0,\infty)$
satisfying either Neumann boundary conditions \eqref{N1} and \eqref{N2},
or Marini boundary conditions \eqref{M1},
or Dirichlet boundary conditions \eqref{D1} and \eqref{D2},
the inequality
\begin{equation}
(2\tau)^{1/4} c\|B_0\|_2 \le a\ \ \ \label{AA0}
\end{equation}
implies that
\begin{align}
\|B(t)\|_\infty &\le 2c_N \|B_0\|_2 t^{-3/4},\ \ \text{for} \ \ 0 <t \le 2\tau,\label{AA1} \\
\|B(t)\|_\infty &\le 2c_N \|B_0\|_2 \tau^{-3/4}, \ \ \text{for} \ \ \tau \le t <\infty \ \ \text{and} \label{AA2}\\
\|A'(t)\|_\infty & \le \gamma \|A'(0)\|_2 t^{-3/4}, \ \ \text{for} \ \ 0 < t \le 2\tau . \label{AA3}
\end{align}
In particular $A(\cdot)$ is a locally bounded strong solution.
Moreover
\begin{align}
&\tau^{5/4} \|A'(t)\|_\infty \le \gamma \|B_0\|_2, \ \text{for} \ \ 2\tau \le t <\infty \ \ \text{and} \label{AA4} \\
&\|A'(t)\|_\infty \to 0 \ \text{as}\ t \to \infty . \label{AA5}
\end{align}
\end{theorem}
The proof depends on the following lemma.
\begin{lemma}\label{lemAA1} $($Differential identities$)$ For a smooth solution to
\eqref{ymh10} there hold
\begin{equation}
d B(t)/dt = \sum_{j=1}^3 (\nabla_j^{A(t)})^2 B(t) + B(t)\# B(t) \ \ \text{and} \label{AA11}
\end{equation}
\begin{equation}
(d/dt) A'(t) =\sum_{j=1}^3 (\nabla_j^{A(t)})^2 A'(t)
+ B(t)\# A'(t) - [A'(t)\lrcorner B(t)], \label{AA12}
\end{equation}
where $\#$ denotes a pointwise product of forms arising from the Bochner-Weitzenboch formula.
\end{lemma}
\begin{proof}
Bianchi's identity and the Bochner-Weitzenbock formula yield
\begin{align*}
B'(t)& = d_{A(t)} A'(t) \\
& = -( d_{A(t)} d_{A(t)}^* + d_{A(t)}^* d_{A(t)})B(t) \\
& = \sum_{j=1}^3 (\nabla_j^{A(t)})^2 B(t) + B(t)\# B(t),
\end{align*}
which is \eqref{AA11}.
Differentiating \eqref{ymh10}
with respect to $t$ gives
\begin{align*}
A''(t) &= - d_{A(t)}^*B'(t) - [ A'(t) \lrcorner B(t)] \\
&= -d_{A(t)}^*d_{A(t)} A'(t) - [ A'(t) \lrcorner B(t)].
\end{align*}
Since $ d_{A(t)}^*A'(t) = -d_{A(t)}^* d_{A(t)}^* B(t) = 0$ we find
\begin{align*}
A''(t) &= -(d_{A(t)}^*d_{A(t)} +d_{A(t)} d_{A(t)}^*)A'(t) - [A'(t) \lrcorner B(t)]\\
& = \sum_{j=1}^3 (\nabla_j^{A(t)})^2 A'(t) + B(t)\# A'(t) - [ A'(t)\lrcorner B(t)],
\end{align*}
which is \eqref{AA12}.
\end{proof}
\bigskip
\noindent
\begin{proof}[Proof of Theorem \ref{thmAA}]
Both of the equations \eqref{AA11} and \eqref{AA12} have the form
specified in \eqref{5.51} with different choices of the form $\omega$ and the
function $h$. We need to verify the boundary conditions \eqref{5.53}
or \eqref{5.54} in each case.
First choose $\omega(s,x) = B(s,x)$ and $h(s, x) = B(s,x)\# B(s,x)$.
If $A(\cdot)$ satisfies Marini boundary conditions, then
$\omega_{norm} = B_{norm} = 0$ by \eqref{N2}, while $d_{A(t)} B(t) = 0$ by the Bianchi
identity. So \eqref{5.53} holds if $A(\cdot)$ satisfies Marini boundary conditions.
Since Neumann boundary conditions
are a special case of Marini boundary conditions, \eqref{5.53} holds in that case also.
If $A(\cdot)$ satisfies Dirichlet boundary conditions then $\omega_{tan} = B_{tan} = 0$
by \eqref{D2},
while $(d_{A(t)}^* \omega(t))_{tan} = (d_{A(t)}^* B(t))_{tan} = -A'(t)_{tan} =0$
by \eqref{ymh10} and \eqref{D1}. So \eqref{5.54} holds for Dirichlet boundary conditions also.
In either case we may therefore apply Proposition \ref{propDomination}
to the choice $\omega= B$.
The inequality \eqref{5.55} then gives the following pointwise inequality
\begin{align}
|B(t, x)| &\le \{e^{t\Delta_N} |B(0)|\}(x)
+ \int_0^t \{e^{(t-s)\Delta_N} |B(s) \# B(s)|\}(x) ds. \label{AA30}
\end{align}
By \eqref{AAhk},
\begin{align}
\|B(t)\|_\infty &\le \| e^{t\Delta_N} |B_0| \|_\infty
+ \int_0^t \|e^{(t-s)\Delta_N} c |B(s)|^2\|_\infty ds \notag\\
&\le c_N \{ t^{-3/4} \|B_0\|_2 + \int_0^t(t-s)^{-3/4} c\| \ |B(s)|^2 \|_2 ds\} \label{5.75}
\end{align}
for $0 < t \le 1$.
Define
\begin{equation}
\beta(t) = \sup_{0\le s \le t} s^{3/4} \|B(s)\|_{L^\infty(M)}.
\end{equation}
This is finite for each number $t\in (0,\infty)$ because $A(\cdot)$ is smooth.
Then
$
\|B(s)\|_\infty \le s^{-3/4} \beta(t) \ \ \text{whenever}\ \ 0 < s \le t,
$
and therefore
$$
\||B(s)|^2\|_2 \le s^{-3/4}\beta(t) \|B(s)\|_2 \le s^{-3/4}\beta(t) \|B_0\|_2.
$$
Using this to estimate the integrand in \eqref{5.75} we find
\begin{align}
\int_0^t(t-s)^{-3/4} c\| \ |B(s)|^2 \|_2 ds
&\le \beta(t) c\| B_0\|_2 \int_0^t (t-s)^{-3/4} s^{-3/4} ds.
\end{align}
The last integral is
$t^{-1/2} \int_0^1 (1-\sigma)^{-3/4} \sigma^{-3/4} d\sigma \equiv t^{-1/2} a_4$
for a constant $a_4$.
Therefore \eqref{5.75} yields
\begin{equation}
t^{3/4} \|B(t)\|_\infty \le c_N\Big\{ \|B_0\|_2
+ \beta(t) c \|B_0\|_2 t^{1/4} a_4 \Big\} \ \text{for}\ 0< t \le 1. \label{5.79}
\end{equation}
Replace $t$ by $t' <t$ in this inequality and take
the supremum over $t' <t$ to find, using the monotonicity of $\beta(t)t^{1/4}$,
\begin{equation}
\beta(t) \le c_N\|B_0\|_2
+ \beta(t) \{t^{1/4} c \|B_0\|_2 c_N a_4\} \ \text{for}\ 0< t \le 1. \label{5.80}
\end{equation}
Let $a= (2 c_N a_4)^{-1}$. Then the inequality \eqref{AA0} may
be written
\begin{equation}
(2\tau)^{1/4} c\|B_0\|_2 c_Na_4 \le 1/2 . \label{5.81}
\end{equation}
Hence, for $0 < t \le 2\tau \le 1$, \eqref{5.80} yields
$\beta(t) \le c_N \|B_0\|_2 +(1/2) \beta(t)$ and thus
$ \beta(t) \le 2c_N \|B_0\|_{L^2(M)}$ if $ 0 < t \le 2\tau \le 1$.
This proves the inequality \eqref{AA1}.
Now \eqref{AA2} follows from \eqref{AA1} easily thus. Write $R = 2c_N \|B_0\|_2$.
If $\tau \le t \le 2\tau$ then, from \eqref{AA1}, it follows that
$\tau^{3/4} \|B(t)\|_\infty \le t^{3/4} \|B(t)\|_\infty \le R$,
which is \eqref{AA2} on the interval $[\tau, 2\tau]$.
We may repeat the previous argument over an interval whose time origin is $\tau$.
Since $\|B(\tau)\|_2 \le \|B_0\|_2$ the definition \eqref{AA0}
shows that we
we may take the ``new $\tau$'' to be the same as the old $\tau$.
Apply the inequality \eqref{AA1} over the interval $(\tau, 3\tau]$.
Taking $t$ in the second half of this interval, i.e. in $[2\tau, 3\tau]$, we find
$\tau^{3/4} \|B(t)\|_\infty \le (t - \tau)^{3/4} \|B(t)\|_\infty \le 2c_N \|B(\tau)\|_2
\le 2c_N \|B_0\|_2$,
which is \eqref{AA2} over the interval $[2\tau, 3\tau]$. Proceeding in this way,
$\tau$ units at a time, we find that \eqref{AA2} holds over the whole interval $[\tau, \infty)$.
Turning to the proof of \eqref{AA3}, take $\omega(s, x) = A'(s,x)$ in
Proposition \ref{propDomination} and take
$h(s) = B(s)\# A'(s) - [ A'(s)\lrcorner B(s)]$
over the interval $[0,2 \tau)$. Once again one needs to verify
the boundary conditions \eqref{5.53} or \eqref{5.54}.
If $A(\cdot)$ satisfies Marini boundary conditions then
$\omega_{norm} = (A')_{norm} = - (d_{A}^* B)_{norm} = 0$ by \eqref{M1}
and \cite[Equ (3.20)]{CG1}.
Moreover $(d_A \omega)_{norm} = (d_A A')_{norm} = (B')_{norm} =0$.
Therefore \eqref{5.53} holds for $\omega = A'$.
Since Neumann boundary conditions are stronger than
Marini boundary conditions, \eqref{5.53} holds in that case also.
If $A(\cdot)$ satisfies Dirichlet boundary conditions then
$\omega_{tan} = (A')_{tan} =0$ by \eqref{D1}. Moreover
$d_A^*\omega = d_A^* A' = - (d_A^*)^2 B =0$
by \cite[Equ (3.24)]{CG1}.
In either case we may therefore apply
Proposition \ref{propDomination}
to the choice $\omega= A'$.
The inequality \eqref{5.55} then gives the estimate
\begin{align*}
\|A'(t)&\|_\infty \le \| e^{t\Delta_N} A'(0) \|_\infty
+ \| \int_0^t e^{(t-s) \Delta_N} |B(s) \# A'(s) + A'(s) \lrcorner B(s)| ds \|_\infty \\
&\le c_N\Big\{ t^{-3/4} \|A'(0)\|_2 + \int_0^t (t-s)^{-3/4} \|B(s) \# A'(s) + A'(s) \lrcorner B(s)\|_2 ds\Big\}.
\end{align*}
But, using \eqref{AA1} combined with Lemma \ref{lemfa37} below, we have
\begin{align*}
\|B(s) \# A'(s) - [ A'(s) \lrcorner B(s)]\|_2 &\le 2c \|B(s)\|_\infty \|A'(s)\|_2 \\
&\le 4c_N s^{-3/4} c\|B_0\|_2 \|A'(s)\|_2 \\
&\le 4c_N s^{-3/4} c\|B_0\|_2 \|A'(0)\|_2 e^{8c_Nc\|B_0\|_2 t^{1/4}}.
\end{align*}
Hence, for $0 < t \le 2\tau$, we find
\begin{align*}
t^{3/4}\|A'(t)\|_\infty&\le c_N\Big\{ \|A'(0)\|_2 \\
&+ t^{3/4}4c_N c\|B_0\|_2 \| A'(0)\|_2 e^{8c_Nc\|B_0\|_2 t^{1/4}}
\int_0^t (t-s)^{-3/4} s^{-3/4} ds \Big\} \\
& = c_N\Big\{ \|A'(0)\|_2
+4c_N c\|B_0\|_2 \| A'(0)\|_2 e^{8c_Nc\|B_0\|_2 t^{1/4}} t^{1/4} a_4\Big\}\\
&\le \|A'(0)\|_2\Big( c_N
+ 4c_N^2 \{(2\tau)^{1/4} c \|B_0\|_2\}
e^{8c_N \{c \|B_0\|_2 (2\tau)^{1/4}\}} a_4\Big) \\
&\le \|A'(0)\|_2 \Big( c_N
+ 4c_N^2 a e^{8c_N a} a_4\Big).
\end{align*}
This proves \eqref{AA3} with $\gamma = c_N
+ 4c_N^2 a e^{8c_N a} a_4$.
We may apply \eqref{AA3} beginning at time $\sigma \ge 0$ instead of time zero to find
\begin{equation}
(t-\sigma)^{3/4} \|A'(t)\|_\infty
\le \gamma \|A'(\sigma)\|_2\ \ \text{if}\ \ \sigma < t \le \sigma + 2\tau.
\end{equation}
In particular, if $ \sigma + \tau \le t$ then $(t-\sigma)^{3/4} \ge \tau^{3/4}$ and therefore
\begin{equation}
\tau^{3/4} \|A'(t)\|_\infty \le \gamma \|A'(\sigma)\|_2\ \ \text{if}\ \ t -2\tau
\le \sigma \le t -\tau . \label{AA34}
\end{equation}
Keeping $t$ fixed and integrating the square of this inequality over the interval
$t -2\tau \le \sigma \le t -\tau$ we find
\begin{equation}
\tau \tau^{3/2} \|A'(t)\|_\infty^2
\le \gamma^2 \int_{t-2\tau}^{t-\tau} \|A'(\sigma)\|_2^2 d\sigma. \label{AA35}
\end{equation}
In view of the bound $\int_0^\infty \|A'(\sigma)\|_2^2 d\sigma \le \|B_0\|_2^2$,
established in \cite[Equ (6.5)]{CG1},
the bound \eqref{AA4} follows and at the same time
the integrability over $[0, \infty)$ of the integrand on the right of
\eqref{AA35} proves \eqref{AA5}.
\end{proof}
\begin{lemma}\label{lemfa37} Let $\psi_\infty (t) = 2c \int_0^t \|B(s)\|_\infty ds$.
Then
\begin{equation}
\|A'(t)\|_2^2 + 2\int_0^t e^{\psi_\infty(t)-\psi_\infty (s)} \|B'(s)\|_2^2 ds
\le e^{\psi_\infty(t)} \|A'(0)\|_2^2. \label{A33}
\end{equation}
In particular, if \eqref{AA1} holds, then
\begin{equation}
\|A'(t)\|_2 \le e^{8c_Nc \|B_0\|_2 t^{1/4}} \| A'(0)\|_2 \ \
\text{for}\ \ 0 <t \le 2\tau \le 1. \label{AA35}
\end{equation}
\end{lemma}
\begin{proof} In the identity (see \cite[Equ (5.8)]{CG1})
$$
(d/ds) \|A'(s)\|_2^2 + 2\|B'(s) \|_2^2 = -2 ([A'(s) \wedge A'(s)], B(s)),
$$
use the inequality
$ 2| ([A'(s) \wedge A'(s)], B(s))| \le 2c \|B(s)\|_\infty \|A'(s)\|_2^2 $
to dominate the last term.
One arrives at
$(d/ds) \|A'(s)\|_2^2 + 2\|B'(s)\|_2^2 \le 2c \|B(s)\|_\infty \|A'(s)\|_2^2. $
Hence
\begin{equation}
(d/ds) (e^{-\psi_\infty (s)} \|A'(s)\|_2^2) +2 e^{-\psi_\infty (s)} \|B'(s)\|_2^2 \le 0. \notag
\end{equation}
Integrate from $0$ to $t$ to get
\begin{equation}
e^{-\psi_\infty(t)} \|A'(t)\|_2^2 - \|A'(0)\|_2^2
+2\int_0^t e^{-\psi_\infty (s)} \|B'(s)\|_2^2 ds \le 0 \notag
\end{equation}
which gives \eqref{A33}.
Now if \eqref{AA1} holds then
\begin{align*}
\psi_\infty(t) & \le 2c \int_0^t 2c_N \|B_0\|_2 s^{-3/4} ds
= 16 cc_N \|B_0\|_2 t^{1/4}.
\end{align*}
Using just the first term in \eqref{A33} we find therefore that \linebreak
$\|A'(t)\|_2^2 \le e^{16c_Nc \|B_0\|_2 t^{1/4}} \| A'(0)\|_2^2 $,
which is \eqref{AA35}.
\end{proof}
\begin{Remark}{\rm Our proof of uniqueness of solutions to the Yang-Mills heat equation
\eqref{ymh10} required use of the allowed initial singularity of $\|B(t)\|_\infty$
specified in the definition \eqref{ymh12} of ``locally bounded''. As to whether uniqueness
holds without such an assumption, we have not been able to decide.
J. R{\aa}de, \cite{Ra}, has proven uniqueness of solutions if one defines
a solution to be a limit of smooth solutions. The following corollary shows that
such a limiting solution is automatically locally bounded and therefore
our uniqueness proof applies to such limiting solutions
when $M$ is a bounded, smooth, convex subset of $\mathbb R^3$.
}
\end{Remark}
\begin{corollary} \label{corAA2} Suppose that, for some $T\le \infty$,
$A(\cdot)$ is a strong solution on
$[0,T)$ satisfying Neumann, Dirichlet, or Marini boundary conditions.
Assume that there is a sequence $A_n$ of smooth solutions on
$[0,T)$, satisfying the same boundary conditions as $A$,
such that $A_n \rightarrow A$ in the $C_{loc}( [0,T), W_1)$ topology.
That is,
\begin{equation}
\sup_{0 \le t \le t_0} \| A_n(t) - A(t) \|_{W_1} \rightarrow 0 \ \ \text{for each}\ \ t_0 \in (0,T)
\end{equation}
Then $A(\cdot)$ is locally bounded.
\end{corollary}
\begin{proof} Each function $A_n$ is clearly a locally bounded strong solution
on $[0,T)$. Since $\|A_n(0)\|_{W_1} $ is uniformly bounded in $n$ there exists
a constant $R >0$ such that $\|B_n(0)\|_2 \le R$ for all $n$.
By Theorem \ref{thmAA} there exists $\tau >0$, depending only on $R$, such that
\begin{equation}
t^{3/4} \| B_n(t)\|_\infty \le 2c_N R\ \ \text{for } 0 < t \le 2\tau.
\end{equation}
For each $t >0$ the $W_1$ convergence of $A_n(t)$ to $A(t)$ implies that
$B_n(t) \rightarrow B(t)$ in $L^2(M)$. Hence
\begin{equation}
t^{3/4} \| B(t)\|_\infty \le 2c_N R \ \ \text{for } 0 < t \le 2\tau.
\end{equation}
The same argument applies on any interval
$ [\alpha, \alpha +2\tau] \subset [0,T)$ because $\|B_n(\alpha) \| _2 \le \| B_n(0)\|_2$.
Therefore $
\| B(t)\|_\infty \le 2c_N R \tau^{-3/4}$ for $\alpha+\tau \le t \le \alpha +2 \tau$. Hence
$\|B(t)\|_\infty$ is bounded on any interval $[b,c] \subset [\tau, T)$.
\end{proof}
\begin{theorem} \label{thmAA3} Suppose that $A(\cdot)$ is a locally
bounded strong solution on an interval
$[0,T)$ satisfying Neumann or Dirichlet boundary conditions.
Then \eqref{AA1} and \eqref{AA2} hold for a number $\tau$ depending
only on $\|B(0)\|_2$. Moreover if $\|A'(0)\|_2 <\infty$ then \eqref{AA3} holds also.
\end{theorem}
\begin{proof} For a given locally bounded strong solution $A(\cdot)$ we know
from the gauge invariant regularization lemma, \cite[Lemma 9.1]{CG1}
that if $T_0 <T$ then there exists $\epsilon >0$, depending on
$\gamma \equiv \sup_{0\le s \le T_0}\|A(s)\|_{W_1}$, such that,
for any interval $[a,b] \subset (0, T_0]$
of length at most $\epsilon$, there is a
sequence, $A_n$ of
smooth solutions over $[a,b]$ which approximate $A$ over this interval in the strong
sense given in \cite[Equ (9.1)]{CG1}.
We need to modify the simple argument of
Corollary \ref{corAA2} to take into account the possibility that $\epsilon$, which
depends on $\gamma$ and therefore on $\|A(\cdot)\|_{W_1}$ over the interval
$[0,T_0]$, may be much smaller than the desired number $\tau$,
which we hope will depend only on $\|B(0)\|_2$.
To this end we will have to derive \eqref{AA30} for non-smooth solutions
to \eqref{ymh10}.
If $[a,b] \subset (0, T_0]$ is an interval of length at most $\epsilon$ and $A_n$
denotes the sequence of smooth approximations of $A$ over $[a,b]$, then
\eqref{AA30} shows that
\begin{align}
|B_n(b, x)| &\le \{e^{(b-a)\Delta_N} |B_n(a)|\}(x)
+ \int_a^b \{e^{(b-s)\Delta_N} |B_n(s) \# B_n(s)|\}(x) ds. \notag
\end{align}
Since $B_n$ converges to $B$ uniformly over $[a,b]\times M$ by
\cite[Equ (9.1)]{CG1}, and since
$e^{t\Delta_N}$ is bounded on $L^\infty(M)$, we may pass to the limit in
the last inequality to find
\begin{align}
|B(b, x)| &\le \{e^{(b-a)\Delta_N} |B(a)|\}(x)
+ \int_a^b \{e^{(b-s)\Delta_N} |B(s) \# B(s)|\}(x) ds \label{AA50}
\end{align}
for any interval $[a,b] \subset (0, T_0]$ of length at most $\epsilon$.
We will show in Lemma \ref{lemAA4} that the validity of the pointwise inequality
\eqref{AA50} over these small intervals implies its validity over large intervals.
Assuming then that \eqref{AA50} holds over any interval $[a,b] \subset (0, T_0]$
we will show that
the derivation leading from \eqref{AA30} to \eqref{5.79} now goes through exactly
as before, provided we replace the interval $[0,t]$ by the interval $[a,t]$, with
the number $a$ necessarily greater
than zero. Thus, defining
$\beta_a(t) =\sup_{a\le s \le t}(s-a)^{3/4} \|B(s)\|_\infty$,
the derivation of \eqref{5.79} shows that
\begin{equation}
(t-a)^{3/4} \|B(t)\|_\infty
\le c_N \Big\{ \|B(a)\|_2 + c \beta_a(t) \|B(a)\|_2 (t-a)^{1/4} a_4\Big\}, \label{AA53}
\end{equation}
for $a \le t \le T_0$. Now fix $t >0$ and let $a\downarrow 0$. Each term in
\eqref{AA53} converges to the corresponding term in \eqref{5.79}. Moreover
the hypothesis that $A(\cdot)$ is locally bounded shows that
$\beta(t) < \infty $ for all $t >0$. The remainder of the proof that \eqref{AA1}
holds is now exactly the same as in the proof of Theorem \ref{thmAA}. The proof of
\eqref{AA2} also follows as before.
The proof of \eqref{AA3} is similar: Taking $\omega(t) = A'(t)$ in
Proposition \ref{propDomination},
one finds the pointwise bound
\begin{equation}
|A'(b, x)|
\le \{e^{(b-a)\Delta_N} |A'(a)|\}(x) +\int_a^b \{e^{(b-s)\Delta_N} |h(s)| \}(x) ds \label{AA54}
\end{equation}
for smooth solutions over an interval $[a,b]$. We may apply the gauge
invariant regularization lemma, \cite[Lemma 9.1]{CG1},
to the given locally bounded
strong solution $A(\cdot)$ and conclude that \eqref{AA54} holds for small intervals
$[a,b] \subset (0, T)$, and therefore, by the next lemma, holds
for all intervals $[a,b] \subset (0, T)$. The limiting procedure for letting
$a\downarrow 0$,
used in the proof of \eqref{AA1}, applies now equally well to the proof of \eqref{AA3}.
\end{proof}
\begin{lemma}\label{lemAA4} $($Time dependent semigroup inequality.$)$ Let $u(t,x)$ and $g(t,x)$ be non-negative bounded measurable functions on
$[0,T) \times M$. Let $\epsilon >0$. Suppose that
\begin{equation}
u(b,x)
\le \{e^{(b-a)\Delta_N} u(a,\cdot)\}(x) + \int_a^b \{e^{(b-s)\Delta_N}g(s)\}(x) ds
\ \ \text{for}\ \ a.e.\ x \label{AA60}
\end{equation}
whenever $ 0 < a < b <T$ and
\begin{equation}
b-a < \epsilon. \label{AA61}
\end{equation}
Then \eqref{AA60} holds for all intervals $[a,b] \subset (0, T)$.
\end{lemma}
\begin{proof} Let $0 < a < b < c <T$ and suppose
$c-b < \epsilon$. For an induction proof, suppose that \eqref{AA60} holds for
this $a$ and $b$. Then
\begin{align*}
u(c) &\le e^{(c-b) \Delta_N} u(b) +\int_b^c e^{(c-s)\Delta_N} g(s) ds \\
&\le e^{(c-b)\Delta_N} \Big\{e^{(b-a)\Delta_N} u(a) + \int_a^b e^{(b-s)\Delta_N}g(s) ds\Big\} + \int_b^c e^{(c-s)\Delta_N} g(s) ds \\
&= e^{(c-a)\Delta_N}u(a)
+\int_a^b e^{(c-s)\Delta_N}g(s) ds + \int_b^c e^{(c-s)\Delta_N} g(s) ds \\
&= e^{(c-a)\Delta_N}u(a) + \int_a^c e^{(c-s)\Delta_N} g(s) ds.
\end{align*}
Therefore, given any interval $[a,b] \subset (0,T)$, one can partition it into small
subintervals $ a =a_0 < a_1 < \cdots < a_n = b$ of length less than $\epsilon$
and arrive at \eqref{AA60} by induction.
\end{proof}
\begin{Remark}{\rm We have not included Marini boundary conditions in
the hypothesis of Theorem \ref{thmAA3} because the gauge invariant
regularization procedure used in the proof has not yet been proven
for Marini boundary conditions.
}
\end{Remark}
\section{Long time behavior} \label{secLTB}
It has been shown in several different contexts \cite{Ra,HT1,HT2} that
over a manifold without boundary, a
solution to the Yang-Mills heat equation over $(0,\infty)$ converges to a limit
as time goes to infinity through some sequence, if one counts only the gauge equivalence class at each time.
Moreover, if one assumes a solution which is smooth for all time then the limiting
connection is also gauge equivalent to a smooth connection \cite{Ra,HT1,HT2}, at least on an open dense set.
One can expect the same kind of behavior for a manifold with boundary.
In this section we are going to prove a version of
such limiting behavior, but only in dimension three. It is aimed partly at showing how
Wilson loop functions can be used to formulate such a convergence procedure
and partly at showing how our gauge invariant regularization procedure smooths
finite energy initial data enough to give meaning to such ``regularized Wilson loops''.
Given a connection on a vector bundle, it is well known that the
associated parallel transport operators along curves determine
the connection. See e.g. \cite[Theorem 2.28]{Po}.
We are going to prove convergence of the parallel transport operators
rather than convergence of the connection forms themselves.
This is analogous to proving, for some sequence of unbounded self-adjoint
operators $C_n$ on a Hilbert space, convergence of the unitary operators
$e^{itC_n}$ instead of convergence of the $C_n$ themselves.
Our main interest is in the regularization of rough gauge potentials, adequate for giving
meaning to the Wilson loop function.
We will begin with an example of a gauge potential with finite energy but
which produces an infinite magnetic flux through some loops.
The Wilson loop function is meaningless for such loops.
In the example we will take the gauge group to be the circle group.
In Section \ref{secLTB1}
we will review how a parallel transport function on loops gives rise to a parallel
transport function on paths, with the help of homotopies. In Section \ref{secloops}
we will show that, for a solution to the Yang-Mills heat equation,
there is a sequence of times going to infinity for which the associated parallel
transport operators around loops converge.
\subsection{Magnetic field of a current carrying washer} \label{secwash}
\
A wire in $\mathbb R^3$ of zero thickness, carrying current, produces a magnetic field of infinite energy.
We are going to describe a slightly smoother current distribution
which produces a magnetic field of finite energy and yet gives
an infinite magnetic flux through certain loops. For such loops the Wilson
loop functional is undefined.
We will show in subsequent sections how
our gauge invariant regularization procedure, via the Yang-Mills heat
equation, applies to the Wilson loop function for finite energy gauge fields.
Consider a washer of zero thickness lying in the $x,y$ plane in $\mathbb R^3$ with center at the origin. We take
the outer radius of the washer to be one and the inner radius to be $1/2$.
A current circulates counterclockwise (viewed from above)
through the washer in concentric circles centered at the origin.
For a point ${\bf x}$ on such a circle, the current vector ${\bf J}({\bf x})$ is tangent to the circle.
See Figure \ref{fig:Wash7}.
\begin{figure}[htbp]
\centering
\includegraphics[width=3in]{Washer7.pdf}
\caption{Current carrying washer}
\label{fig:Wash7}
\end{figure}
We take the current density to vary with the distance from the origin and
to be heavily weighted toward the outer rim of the washer.
We can write the planar current density explicitly as
\begin{align}
{\bf J}({\bf x}) = \lambda(r) \Big( -{\bf i} \sin \phi + {\bf j}\cos \phi\Big)\ \ \text{when}\ \ \
{\bf x} = ( r \cos \phi, r \sin \phi, 0).
\end{align}
$\lambda(r)$ is the profile of the current strength as one moves from
the inner rim at $r = 1/2$ to the outer rim at $r= 1$. By this we mean that $\lambda(r)dr$
is the total current passing through a small radial interval $dr$ at distance $r$.
We will take
\begin{align}
\lambda(r)=\frac{1}{(1-r)(\log\frac{1}{1-r})^2}, \ \ \ 1/2 \le r < 1 \label{m7}
\end{align}
The intensity of current is therefore quite large near the outer rim. But the total current
circulating around the washer is $\int_{1/2}^1 \lambda(r) dr$, which is finite.
The magnetic potential produced by the current is given by ${\bf A} = (-\Delta)^{-1} {\bf J}$.
Thus
\begin{align}
4\pi {\bf A}({\bf x}) &= \int_{wash} \frac{{\bf J}({\bf x}')}{|{\bf x} - {\bf x}'|}d^2 {\bf x}' \notag \\
&=\int_{1/2}^1 dr \lambda(r) \int_{-\pi}^{\pi}
\frac{ -{\bf i} \sin \phi + {\bf j} \cos \phi}{ |{\bf x} - ( r \cos \phi, r \sin \phi, 0)|} d\phi \label{m9}
\end{align}
The energy of this field is given by (See e.g. \cite[Equ. 5.153]{Jac}.)
\begin{align}
W = \int_{wash} \int_{wash} \frac{ {\bf J}({\bf x}) \cdot {\bf J}({\bf x}')}{|{\bf x} - {\bf x}'|} d^2{\bf x} d^2{\bf x}' \label{m10}
\end{align}
where $\int_{wash}$ means the two dimensional integral over the washer. This is equivalent to the $H_1$ norm of ${\bf A}$ because
$\|{\bf A}\|_{H_1}^2 = \int_{\mathbb R^3}\sum_{j=1}^3 \partial_j {\bf A} \cdot\partial_j {\bf A} d^3x
= (-\Delta {\bf A}, {\bf A}) = ({\bf J}, (-\Delta)^{-1} {\bf J})$.
\begin{theorem}\label{washer}\
1$)$ The gauge potential ${\bf A}$ has finite energy.
2$)$ There are piecewise smooth curves of finite length in the $x,y$ plane
through which the magnetic flux is infinite. In particular
the holonomy (Wilson loop) $ W_C({{\bf A}}) = e^{i\infty}$ is undefined for such a curve $C$.
\end{theorem}
\begin{proof} To bound the energy \eqref{m10} take $ {\bf x} = (r \cos \phi, r \sin \phi, 0)$ and
${\bf x} ' = (r' \cos \phi', r' \sin \phi', 0)$ in \eqref{m10}. Then
$|{\bf x} - {\bf x}'|^2 = (r\cos \phi - r'\cos \phi')^2 + (r\sin \phi - r'\sin \phi')^2 = r^2 + (r')^2 - 2rr' \cos(\phi -\phi') =(r-r')^2 +2rr'(1-\cos(\phi -\phi'))$ and
${\bf J}({\bf x})\cdot {\bf J}({\bf x}') = \lambda(r) \lambda(r') (-{\bf i} \sin\phi + {\bf j} \cos\phi)\cdot(-{\bf i} \sin \phi' + {\bf j} \cos \phi') = \lambda(r) \lambda(r') \cos (\phi - \phi')$. Hence
\begin{align}
W &= \int_{1/2}^1 \int_{1/2}^1 drdr' \lambda(r) \lambda(r')
\int_{-\pi}^\pi \int_{-\pi}^\pi
\frac{ \cos(\phi - \phi')\ d\phi d\phi'}{ \Big( (r-r')^2 +2rr'(1-\cos(\phi -\phi'))\Big)^{1/2}} \notag\\
&= 2\pi \int_{1/2}^1 \int_{1/2}^1drdr' \lambda(r) \lambda(r')\int_{-\pi}^\pi
\frac{\cos \theta\ d\theta }{\Big((r-r')^2 + 2rr'(1- \cos\theta)\Big)^{1/2} }. \label{m12}
\end{align}
Since $\lambda(r)$ has singular behavior near $ r =1$ we will need to bound
the $\theta$ integral above to prove finite energy. We will also need a lower bound
later to prove that ${\bf A}({\bf x})$ is unbounded.
For these purposes we will show in Section \ref{seculbds} that
there are strictly positive constants $c_1, c_2, C_1, C_2$ such that
\begin{align}
c_1 + c_2 \log\frac{1}{u} &\le \int_{-\pi/4}^{\pi/4}
\frac{\cos\theta}{\Big( u^2 + 2v^2 (1-\cos \theta)\Big)^{1/2}} d \theta
\le C_1 + C_2 \log\frac{1}{u} \label{m51}\\
&\text{for}\ \ 0 < u <1\ \ \text{and}\ \ 1/2 \le v \le 2 \label{m52}
\end{align}
Now let $u = |r - r'|$ and $v = (rr')^{1/2}$. Then \eqref{m52} is satisfied for all $r$ and $r'$
entering the integrals in \eqref{m12}. The contribution
to the $\theta$ integral in \eqref{m12} from $|\theta| \ge \pi/4$ is a bounded function of $ r$
and $r'$ and since $\lambda(r)$ is integrable the contribution to \eqref{m12} from
$|\theta| \ge \pi/4$ is finite. In view of the second inequality in \eqref{m51} it suffices therefore
to show that
\begin{equation}
\int_{1/2}^1dr \int_{1/2}^1dr' \lambda(r) \lambda(r') \log \frac{1}{|r-r'|} < \infty.
\end{equation}
Since the possibly non-integrable singularity is near $r= r' =1$ it will be
more perspicuous to change variables
to $s = 1-r$ and $s' = 1 - r'$. Thus we need to show that
\begin{align}
\int_0^{1/2} \int_0^{1/2} \mu(s) \mu(s') \log\frac{1}{|s - s'|} ds ds' < \infty \label{m32}
\end{align}
when
$\mu(s) = (s (\log s)^2)^{-1}$. The value of this double integral over the two triangles
$s \le s'$ and $ s' \le s$ is the same. So it suffices to show that one of them is finite.
In fact we will show that
\begin{equation}
\int_0^{s'} \mu(s)\log\frac{1}{|s - s'|} ds
\end{equation}
is bounded for $ 0 \le s' \le 1/2$, which will prove \eqref{m32} because $\mu(s')$ is integrable.
Let $c = s'/2$. Now $\log\frac{1}{s' -s}$ is an increasing function of $s$ on $(0,c)$ while
$\mu(s)$ is a decreasing function of $s$ on $(c, s')$. Hence
\begin{align}
\int_0^{s'} \mu(s) \log\frac{1}{s' -s} ds = \int_0^c \mu(s) \log\frac{1}{s' -s} ds + \int_c^{s'} \mu(s) \log\frac{1}{s' -s} ds \notag\\
\le \log\frac{1}{s' -c} \int_0^c \mu(s) ds + \mu(c) \int_c^{s'} \log\frac{1}{s' -s} ds \label{m34}
\end{align}
Both integrals can be done explicitly. One finds
$\int_0^c \mu(s) ds = (\log (c^{-1}))^{-1} =(\log (2/s'))^{-1}$
and $\int_c^{s'} \log\frac{1}{s' -s} ds =(s'/2) ( 1 +\log (2/s'))$.
Hence the right side of \eqref{m34} equals
\begin{align*}
\Big(\log\frac{2}{s'}\Big) \frac{1}{\log\frac{2}{s'}}
+ \frac{1}{(s'/2)(\log\frac{2}{s'})^2} \cdot (s'/2) \Big(1+ \log\frac{2}{s'}\Big)
= 1 +\frac{1+\log\frac{2}{s'}}{(\log\frac{2}{s'})^2},
\end{align*}
which is bounded on $ 0 < s' \le 1/2$. This proves Part 1) of Theorem \ref{washer}.
For Part 2 we need to understand the behavior of the magnetic potential
${\bf A}({\bf x})$ as ${\bf x}$ approaches the outer rim of the washer.
Because of the cylindrical symmetry it will suffice to do this when ${\bf x}$ lies in the
$x,z$ plane. In fact it suffices to consider just $ {\bf x} = (x_1,0, x_3)$ with $x_1\ge 0$.
The distance from ${\bf x}$ to a current element is
$ | {\bf x} - (r \cos \phi, r\sin \phi, 0)|^2 = (x_1 - r\cos \phi )^2 + r^2 \sin^2 \phi + x_3^2
= x_1^2 + x_3^2 +r^2 - 2 x_1 r \cos \phi = (x_1 - r)^2 + x_3^2 + 2x_1 r (1 -\cos\phi)$.
Inserting this into \eqref{m9} we see that the denominator is an even function of $\phi$.
The contribution of ${\bf i} \sin \phi$ in the integral is therefore zero. Hence
\begin{equation}
4\pi {{\bf A}}({\bf x}) = {\bf j} \int_{1/2}^1 dr \lambda(r) \int_{-\pi}^\pi
\frac{\cos \phi}{\Big( (x_1 -r)^2 + x_3^2 + 2x_1 r (1 - \cos \phi)\Big)^{1/2}} d\phi \label{m43}
\end{equation}
for ${\bf x}$ in the $x,z$ plane. From the cylindrical symmetry we see that ${\bf A}({\bf x})$ is horizontal for all ${\bf x} \in \mathbb R^3$ and in fact is tangent to the horizontal circle which is centered
on the $z$ axis and passes through ${\bf x}$. (On the $z$ axis ${\bf A}({\bf x})$ is zero, as
one sees by putting $x_1=0$ in \eqref{m43}.)
Of course ${\bf A}$ is a smooth function on the complement of the closed washer
because the denominator in \eqref{m9} is locally bounded away from zero there.
We need only focus attention on the behavior of ${\bf A}({\bf x})$ for ${\bf x}$ in a
small neighborhood of $(1,0,0)$ in the $x,z$ plane. For such ${\bf x}$ the contribution
to the integral from points in the washer where $|\phi| \ge \pi/4$
produces a smooth function of $x_1, x_3$ for $x_1 >0$. We therefore need only
to analyze the behavior of the function $f$ defined by
\begin{equation}
f(x_1, x_3) = \int_{1/2}^1 dr \lambda(r) \int_{|\phi| \le \pi/4}
\frac{\cos \phi\ d\phi}{\Big( (x_1 -r)^2 + x_3^2 + 2 x_1 r(1- \cos \phi) \Big)^{1/2}} .
\end{equation}
The first inequality in \eqref{m51} will give
a lower bound on this integral just outside the outer rim of the washer as follows.
Suppose that
$1 \le x_1 \le 5/4$ and $|x_3| \le 1/2$. Let $u^2 = (x_1 - r)^2 + x_3^2$ and $ v^2 = x_1 r$.
The reader can verify that \eqref{m52} is satisfied.
Hence
\begin{align}
f(x_1, x_3) \ge \int_{1/2}^1 \lambda(r)(c_1 + c_2 \log \frac{1}{u}) dr .
\end{align}
If $x_1 \downarrow 1$
and $|x_3| \downarrow 0$ then $u\downarrow 1-r$ and the monotone convergence theorem shows that,
for some finite constant $C_6$, one has
\begin{align}
\liminf_{x_1\downarrow 0\ |x_3|\downarrow 0} f(x_1,x_3) & \ge c_2 \int_{1/2}^1 dr \lambda(r) \log\frac{1}{1-r} + C_6 \notag\\
&=c_2\int_{1/2}^1 dr \frac{1}{(1-r)(\log\frac{1}{1-r})^2} \log\frac{1}{1-r} + C_6 \notag\\
&= \infty. \label{m48}
\end{align}
Therefore ${\bf A}({\bf x})\cdot {\bf j}$ is infinite at the point $(1,0,0)$ and also goes
to $\infty$ as ${\bf x} \rightarrow (1,0,0)$ in the $x,z$ plane.
Consider now the loop shown in Figure \ref{fig:Wash7}. It is given by a closed curve $C$
lying in the $x,y$ plane and forming the boundary of an annular sector centered at the origin and whose inner radius is one. In Figure \ref{fig:Wash7} the curve is shown separated from
the rim of the washer for clarity. But we are interested in the circumstance in which the inner circle
of the annular sector coincides with a portion of the outer rim of the current carrying washer.
The outer circle segment of $C$
is concentric with the inner one and is joined to it by radial lines.
Since ${\bf A}$ is tangential to the outer rim
of the washer it is also tangential to the inner circle of $C$.
However this tangential component of ${\bf A}$ is infinite, as we have seen.
Thus the integral of ${\bf A}$ along the inner circle of $C$
is infinite. The integral of
${\bf A}$ along the two radial lines
is zero because ${\bf A}$ is perpendicular to these radial lines.
The integral of ${\bf A}$ along the outer circle is finite because ${\bf A}$ is smooth
in the vicinity of the outer circle. Thus $\int_C {\bf A}({\bf x}) \cdot d{\bf x} = \infty$.
This proves Part 2) of Theorem \ref{washer}.
There is another sense in which this loop integral is infinite: keep the outer circle
of the curve $C$ fixed
and shift the inner circle
away from the outer rim of the washer by a small amount,
say $\epsilon >0$, as is shown in Figure \ref{fig:Wash7}.
For this curve $C_\epsilon$ the contour integral $ \int_{C_\epsilon} {\bf A}({\bf x}) \cdot dx$
is finite, but increases to infinity as $\epsilon \downarrow 0$ because, as \eqref{m48}
shows, for $x_3=0$ the tangential component of ${\bf A}({\bf x})$ increases to $\infty$
as $x_1 \downarrow 1$. In particular, by Stokes' theorem, the magnetic flux through
the planar surface bounded by $C_\epsilon$ increases
to $\infty$ as $\epsilon \downarrow 0$. (The right hand rule
also shows that the magnetic field $B$ points downward everywhere on the ring.)
\end{proof}
\begin{remark}\label{remrestrict}{\rm Trace
theorems assert that a function lying in a Sobolev space
$H_s$ over a manifold will, upon restriction to a co-dimension one submanifold $N$,
lie in $H_{s -(1/2)}(N)$. This theorem notoriously breaks down if $s = 1/2$.
In our case
the magnetic field ${\bf A}$ lies in $H_1(\mathbb R^3)$ and therefore restricts to a function
in $H_{1/2}(N)$ for any reasonable surface $N$. One can try to restrict it once more to
a curve $C$ contained in $N$ and ask what properties the restriction to $C$ has.
Since $s$ is now equal to $1/2$ the trace theorem breaks down. One can not infer from it
that the restriction to the curve $C$ is an almost everywhere finite
function. Our example shows that the very
worst can happen: the restriction of the magnetic field to an arc of the outer rim of the washer is
identically infinite. For further discussion of trace theorems see \cite[Theorem 9.5]{LM}
and \cite{EL}.
}
\end{remark}
\subsubsection{Upper and lower bounds for an integral} \label{seculbds}
\begin{lemma}
Let $0 < \theta_0 < \pi/2$ and let
$a^2 = (\sin\theta_0)/\theta_0$ with $a >0$. Then, for $u>0$ and $v >0$, there holds
\begin{align}
\frac{1}{v} \log (1 + \frac{v\theta_0}{u})
\le \int_0^{\theta_0} \frac{1}{\Big( u^2 + 2 v^2 (1 - \cos\theta)\Big)^{1/2}} d \theta
\le \frac{\sqrt{2}}{va} \log (1 + \frac{va\theta_0}{u}) . \label{m25}
\end{align}
In particular \eqref{m51} holds.
\end{lemma}
\begin{proof} For $s \ge 0$ the inequalities $ 1+s^2 \le (1+s)^2 \le 2(1+s^2)$ imply that
\begin{align}
\int_0^b (1+s^2)^{-1/2} ds \ge \log(1+b)
\ge 2^{-1/2} \int_0^b (1+s^2)^{-1/2} ds\ \text{for}\ b >0. \label{m26}
\end{align}
Since $a^2$ is the slope of a line segment lying below $\sin(\cdot)$, integration gives
$ a^2 \theta^2 \le 2(1-\cos\theta) \le \theta^2$ for $0 \le \theta \le \theta_0$. Hence
\begin{align*}
\int_0^{\theta_0}\frac{d\theta}{\Big( u^2 + v^2 a^2 \theta^2\Big)^{1/2} }
\ge \int_0^{\theta_0}\frac{d\theta}{\Big(u^2 + 2v^2 ( 1- \cos\theta)\Big)^{1/2}}
\ge \int_0^{\theta_0}\frac{d\theta}{\Big( u^2 + v^2 \theta^2\Big)^{1/2} }
\end{align*}
Change variables in the left-most integral to $s = (va/u)\theta$ to find
$(va)^{-1}\int_0^{va\theta_0/u} (1+ s^2)^{-1/2} ds$, which, by \eqref{m26}, is at most
$(va)^{-1}\sqrt{2} \log(1 + (va\theta_0/u))$. The other half of \eqref{m25} follows similarly.
For the proof of \eqref{m51} we can ignore the factor $\cos\theta$ in the numerator
of \eqref{m51} because it is bounded and bounded away from zero on $[-\pi/4, \pi/4]$.
We are assuming now that $1/2\le v \le 2$, from which
it follows that the left side of \eqref{m25} dominates the left side of \eqref{m51} for some constants
$c_1, c_2$ and for all $u \in (0,1)$. Moreover, since $va\pi/4 \le2$, the right side of
\eqref{m25} is at most $(2\sqrt{2}/a)\log(1 +(2/u)) \le (4\sqrt{2}/a)\log(2/u)$ because
$\log (1+x) \le 2 \log x$ when $x \ge 2$.
\end{proof}
\subsection{From loops to paths} \label{secLTB1}
\begin{notation}\label{notpaths}
{\rm Let $M$ be the closure of a bounded open set in $\mathbb R^3$
with smooth convex boundary. Denote by $\Gamma$ the set of piecewise
$C^1$ functions from $[0,1]$ into $M^{int}$.
If $\gamma$ and $\mu$ are two elements of $\Gamma$
such that $\gamma(1) = \mu(0)$ then their concatenation $\gamma\mu$ is defined by
\begin{equation}
(\gamma \mu)(s) =
\begin{cases} \gamma(2s), &0\le s \le 1/2 \\
\mu(2s -1), &1/2 \le s \le 1.
\end{cases} \label{ltp3}
\end{equation}
The curve $\gamma\mu$ is clearly again in $\Gamma$.
The inverse path is defined as usual
by $\gamma^{-1} (s) = \gamma(1-s), 0 \le s \le 1$. The path $\gamma \gamma^{-1}$
retraces itself.
By a {\it parallel transport system} in the bundle ${\mathcal V}\times M \rightarrow M$
we mean a map
$\Gamma \ni \gamma \mapsto //_\gamma \in End\ {\mathcal V}$ such that
i) $//_{\gamma\circ \phi} = //_\gamma$ for any homeomorphism $\phi:[0,1]\rightarrow [0,1]$, which, together with its inverse, is piecewise $C^1$,
ii) $//_{\gamma\mu} = //_\gamma\ \ //_\mu$,
iii) $//_{\gamma \gamma^{-1}} = I_{\mathcal V}$.
Taking $\epsilon_0$ to be the trivial curve, $\epsilon_0(s) \equiv x_1$, it follows from
ii) and iii) that $//_{\epsilon_0} = I$ and $(//_\gamma)^{-1} = //_{\gamma^{-1}}$.
Under further technical assumptions such a parallel transport system always
comes from a connection on the bundle ${\mathcal V}\times M \rightarrow M$. This has been discussed for example in \cite[Theorem 2.28]{Po}.
We are going to show that, for a solution $A(\cdot)$ to the Yang-Mills heat equation,
and for any sequence $t_k$ going to infinity,
the parallel transport operators $//_\gamma^{A(t_k)}$ converge to such a parallel transport
system after suitable gauge transformations.
Choose a point $x_0 \in M^{int}$ and denote the set of loops at $x_0$ by
\begin{equation}
\Gamma_0 = \{ \gamma \in \Gamma : \gamma(0) = \gamma(1) = x_0 \}. \label{ltp5}
\end{equation}
A parallel transport system can be recovered, up to gauge transformation,
from its restriction to $\Gamma_0$
by choosing a homotopy of $M$ with $x_0$.
The well known procedure for doing this will be described
in the following algebraic lemma.
Let $X$ be a manifold and let $x_0$ be a point in $X$. By a piecewise $C^1$ homotopy of
$X$ with $\{x_0\}$ we mean a continuous map
$h: [0,1]\times
X \rightarrow X$ with $h(0,x) = x_0$, $ h(1,x) = x$,
and, for all $x \in X$, the curve $s \mapsto h_x(s) := h(s,x)$ is
piecewise $C^1$. We will assume also that $h(s, x_0)= x_0$ for all $s \in [0,1]$.
Our limit results can easily be extended to non-contractible manifolds,
but the analytic idea is already well illustrated in the contractible case,
to which we will restrict our attention.
\begin{lemma} \label{loops-paths}
Let $X$ be a finite dimensional pathwise connected manifold. Let
$x_0 \in X$. Denote by $\Gamma$ the set of piecewise $C^1$ functions
from $[0,1]$ into $X$ and define
$\Gamma_0 = \{ \gamma \in \Gamma : \gamma(0) = \gamma(1) = x_0 \}$.
Suppose that $P: \Gamma_0 \rightarrow End\ {\mathcal V}$ is a map
with the following properties
$1)\ ($parametrization invariance$)$ $P(\gamma) = P(\gamma\circ \phi)$ for any
piecewise $C^1$ homeomorphism $\phi: [0,1]\rightarrow [0,1]$ with
piecewise $C^1$ inverse.
$2)$ $P(\gamma \mu) = P(\gamma) P(\mu)$ for all $\gamma$ and
$\mu$ in $\Gamma_0$.
$3)$ $P(\gamma \gamma^{-1}) = I_{\mathcal V}$ for any path $\gamma \in \Gamma$ with
$\gamma(0) = x_0$.
\noindent
Let $h:[0,1]\times X \rightarrow X$ be a piecewise $C^1$ homotopy of $X$ to $x_0$.
Then there is a unique parallel transport system
$//_\gamma,\ \gamma \in \Gamma$, which is the identity along all homotopy
paths $h_x(\cdot)$ and agrees with $P$ on $\Gamma_0$.
\end{lemma}
\begin{proof} Suppose that $\gamma \in \Gamma$ with
$\gamma(0) = x$ and $\gamma(1) = y$. Then the path
$h_x \gamma h_y^{-1}$ lies in $\Gamma_0$.
Define $//_\gamma = P(h_x \gamma h_y^{-1})$.
If $\mu \in \Gamma$ also and $\gamma(1) = \mu(0)$ and $\mu(1) =z$ then
\begin{equation}
//_{\gamma \mu} = P(h_x \gamma \mu h_z^{-1})
= P((h_x \gamma h_y^{-1})(h_y \mu h_z^{-1})) = //_\gamma \ //_\mu
\end{equation}
by 2). Moreover $//_{\gamma \gamma^{-1}}
= P((h_x\gamma)( \gamma^{-1}h_x^{-1}))
= P(h_x\gamma)(h_x\gamma)^{-1})= I_{\mathcal V}$ by 3).
Thus items ii) and iii) are verified. Item i) is clear. Moreover
$//_{h_x} = P(h_{x_0}h_x (h_x)^{-1}) = I_{\mathcal V}$. For any other parallel transport system
$///$ with the stated properties one has $///_\gamma = ///_{h_x}\ ///_\gamma\ ///_{h_y^{-1}}
=///_{h_x \gamma h_y^{-1}} = P(h_x \gamma h_y^{-1}) =//_\gamma$.
\end{proof}
}
\end{notation}
\subsection{Convergence on loops} \label{secloops}
\begin{notation}{\rm Let $M$ be the closure of a bounded open set in $\mathbb R^3$
with smooth convex boundary. Let $x_0 \in M^{int}$ and define
$\Gamma $ and $\Gamma_0$ as in Notation \ref{notpaths}.
In our simple setting a tangent vector to $M$ at a point $\gamma(s)$ is just
a vector $u(s) \in \mathbb R^3$. We will denote by $T_\gamma(\Gamma)$ any
piecewise $C^1$ function $u:[0,1]\rightarrow \mathbb R^3$. If $\gamma \in \Gamma_0$
then we will write $u \in T_{\gamma}(\Gamma_0)$ if $u$ is in piecewise
$C^1([0,1]; \mathbb R^3)$ and $u(0) = u(1) =0$. For $u \in T_\gamma (\Gamma)$
define
\begin{equation}
\| u \| = \sup_{0\le s \le 1} |u(s)|_{\mathbb R^3} + \sup_{0\le s \le 1} |u'(s)|_{\mathbb R^3}. \label{LTB1}
\end{equation}
For a curve $[a,b] \ni t\rightarrow \gamma_t(\cdot) \in \Gamma$
we take its length to be
$\int_a^b \|\partial_t \gamma_t\| dt$ as usual and define the distance $d_1(\gamma_1, \gamma_2)$
to be the infimum of lengths of curves joining $\gamma_1$ to $ \gamma_2$
in the manifold $\Gamma$.
$\Gamma$ and $\Gamma_0$ are (incomplete) metric spaces in this metric.
}
\end{notation}
\begin{definition}{\rm For a smooth End ${\mathcal V}$ valued
connection form $A$ on $M^{int}$ and a piecewise
$C^1$ path $\gamma$ in $M$ the parallel transport operator along $\gamma$
is defined by the solution to the ordinary differential equation
\begin{equation}
g(t)^{-1} dg(t)/dt = A\< d \gamma(t)/dt \>,\ \ \ g(0) = I_{{\mathcal V}}.
\end{equation}
We put $//_\gamma^A = g(1)$. Properties i), ii), iii) of Notation \ref{notpaths}
are well known for this map.
}
\end{definition}
In this section we are going to prove that for any locally bounded strong solution
of the Yang-Mills heat equation satisfying Neumann or Dirichlet boundary conditions,
and for any sequence of times
going to infinity,
there is a subsequence $t_j$ and gauge transforms $k_j$
such that the connection forms $A(t_j)^{k_j}$ are smooth and the
parallel transport operators $//_\gamma^{A(t_j)^{k_j}} $
converge, as operators from ${\mathcal V}$ to ${\mathcal V}$, to a map $P$ on $\Gamma_0$
satisfying all the conditions listed in Lemma \ref{loops-paths}.
\begin{theorem} \label{LTB} Suppose that $M$ is a compact convex subset of $\mathbb R^3$
with smooth boundary.
Let $A(\cdot)$ be a locally bounded strong solution of the
Yang-Mills heat equation \eqref{ymh10} over $[0,\infty)$ satisfying Dirichlet or
Neumann boundary conditions. Choose $x_0 \in M^{int}$.
Suppose that $\{t_k\}$ is a sequence of times going to $\infty$.
There is a function $P:\Gamma_0 \rightarrow End\ {\mathcal V}$ satisfying
conditions {\rm 1), 2), 3)}
of Lemma \ref{loops-paths},
a subsequence $t_j$
and functions $k_j \in W_1(M;K)$ such that
a$)$\ \ $k_j^{-1} dk_j \in W_1(M;\frak k)$ for all $j$,
b$)$\ \ $\alpha_j \equiv A(t_j)^{k_j}$ is in $C^\infty(M;\Lambda^1\otimes \frak k)$ and,
c$)$ for each $\gamma \in \Gamma_0$
the operators $//_\gamma^{\alpha_j}$
converge to $P(\gamma)$ as $j\rightarrow \infty$.
\noindent
Moreover
d$)$ $P$ is continuous on $\Gamma_0$ in the metric $d_1$.
In particular, given a piecewise $C^1$ homotopy of $M^{int}$ onto $x_0$,
there is a parallel transport system on $\Gamma$ that extends $P$.
\end{theorem}
\begin{Remark}{\rm If $\gamma$ is a closed curve in $M^{int}$
beginning at $x_0$, and
$A$ is a smooth connection form, then for any smooth function $k: M \rightarrow K$
one has the well known identity.
\begin{equation}
//_\gamma^{A^k} = k(x_0)^{-1} (//_\gamma^A)k(x_0)
\end{equation}
Consequently
\begin{equation}
trace\ //_\gamma^{A^k} = trace //_\gamma^A.
\end{equation}
The function $A \mapsto trace\ //_\gamma^A$ is therefore fully gauge invariant
and in particular is independent of the choice of gauge transformation
$k$. Theorem \ref{LTB} implies then that there exists a sequence of times
going to infinity for which the functions $\;trace //_\gamma^{A(t_j)}$ converge for all piecewise
$C^1$ loops $\gamma$ starting at $x_0$. One need not specify gauge transformations $k_j$ for this convergence.
}
\end{Remark}
The proof of Theorem \ref{LTB} depends on the following lemmas.
\begin{lemma} \label{lem1}
Let $A(\cdot)$ be a locally bounded strong solution satisfying Dirichlet or Neumann
boundary conditions and let $t_1 >0$.
Then there exists a continuous function $k: M \rightarrow K$ such that
a$)$ $k^{-1} dk \in W_1(M)$ and
b$)$ $\alpha\equiv A(t_1)^{k} \in C^\infty(M; \Lambda^1\otimes \frak k)$.
\end{lemma}
\begin{proof} The proof depends heavily on results in \cite{CG1}.
From \cite[Corollary 9.3]{CG1}
it follows that
$\sup_{0<t \le t_1} \|A(t)\|_{H_1} < \infty$. Therefore, by \cite[Theorem 2.13]{CG1} there exists $T>0$ such
that, for any $t_0 \in (0, t_1)$, the parabolic equation
\begin{equation}
(\partial/\partial t)C = -(d_C^* B_C + d_C d^*C), t >0, \ \
C(0) = A_0 . \label{ST11}
\end{equation}
\cite[Equ (2.14)]{CG1} has a solution $C(\cdot)$ on the interval $[t_0, t_0 + T]$,
with $C(t_0) = A(t_0)$. Pick $t_0 \in (0,t_1)$ such that $t_1 < t_0 + T$. \cite[Corollary 8.4]{CG1}
then ensures that there exists a continuous function $ g: M\rightarrow K$ such that $g^{-1} dg \in W_1$ and for which $A(t_1) = C(t_1)^g$. Since $C(t_1) \in C^{\infty}$ we may take $k = g^{-1}$.
Take note here that the equality $A(t_1) = C(t_1)^g$ relies on the uniqueness
theorem, \cite[Theorem 8.15]{CG1}, which is applicable to the
restriction of $A(\cdot)$ to $[t_0,t_1]$.
\end{proof}
\begin{lemma} \label{lem2}
Let $\gamma : [0,1]\rightarrow M$ be a piecewise $C^1$ closed curve starting at $x_0$.
Let $ u:[0,1]\rightarrow T(M)$ be a $C^1$ vector field along $\gamma$ for which
$u(0) = u(1) =0$.
That is, $u(s) \in T_{\gamma(s)}(M), 0 \le s \le 1$.
Let $\alpha$ be a smooth connection form on $M$ with bounded curvature $B$.
Then
\begin{equation}
\| \partial_u //_\gamma^\alpha \|_{End \ {\mathcal V}}
\le \| B \|_\infty \sup_{0\le s \le1} |u(s)|\ \ \text{Length}(\gamma). \label{LTB5}
\end{equation}
\end{lemma}
\begin{proof} Since $u(0) = u(1) = 0$ the identity \cite[Equ (2.6)]{G2}
shows that
\begin{align*}
\Big\|\partial_u //_\gamma^\alpha\Big\|_{End \ {\mathcal V}}
&= \Big\|\int_0^1 //_{\gamma|_0^s}^\alpha
\< B(\gamma(s)), \gamma'(s) \wedge u(s) \> ds \Big\|_{End \ {\mathcal V}}\\
& \le \int_0^1 \| //_{\gamma|_0^s}^\alpha \|_{End \ {\mathcal V}}
\|B \|_\infty |\gamma'(s) \wedge u(s) \> |_{\Lambda^2(\mathbb R^3)} ds \\
& \le \|B \|_\infty \int_0^1|\gamma'(s)| | u(s)| ds \\
& \le \|B \|_\infty (\sup_{0 \le s \le 1} |u(s)| ) \int_0^1 |\gamma'(s)| ds,
\end{align*}
which is \eqref{LTB5}.
\end{proof}
\bigskip
\noindent
\begin{proof}[Proof of Theorem \ref{LTB}]
For each $\;t \ge 1$ we have constructed a gauge
function $\;k(t): M \rightarrow K$
such that $\alpha(t) \equiv A(t)^{k(t)}$ is a $C^\infty$ connection form. Denote by $B_{\alpha}(t)$ the curvature of the connection $\alpha(t)$.
Let $\gamma$ and $\eta$ be in $\Gamma_0$ and of length at most $L$.
Define $u(s) = \gamma(s) - \eta(s)$ and let
$\gamma_\sigma(s) = \eta (s) + \sigma u(s)$.
Then $\gamma_\sigma$ lies in $M^{int}$ for small $\sigma$
and $\partial_\sigma \gamma_\sigma = u$. Let $b = \sup_{t\ge 1} \|B(t)\|_\infty$.
We know that $ b < \infty$ by Theorem \ref{thmAA}. Since
$ \| B_{\alpha}(t)\|_ \infty = \| B(t)\|_\infty \le b$,
Lemma \ref{lem2} shows that
\begin{align*}
\|\partial_\sigma //_{\gamma_\sigma}^{\alpha(t)} \|_{End\ {\mathcal V}}
&\le b \sup_{0\le s \le 1}|\gamma(s) - \eta(s)|\ \cdot \text{Length}(\gamma_\sigma) \\
&\le b \sup_{0\le s \le 1}|\gamma(s) - \eta(s)| \
\cdot [\text{Length}(\gamma) + \text{Length}(\eta) ] \\
&\le 2bL \sup_{0\le s \le 1}|\gamma(s) - \eta(s)|.
\end{align*}
Hence
\begin{align}
\| //_\gamma^{\alpha(t)} - //_{\eta}^{\alpha(t)} \|_{End\ V}
&\le \int_0^1 \| \partial_\sigma //_{\gamma_\sigma}^{\alpha(t)} \|_{End\ V} d\sigma \notag \\
&\le 2bL \sup_{0\le s \le 1}|\gamma(s) - \eta(s)| \label{LTB10}.
\end{align}
An Arzela-Ascoli type diagonalization argument
shows that a pointwise \linebreak
bounded, equicontinuous sequence of
functions on a separable metric space $S$ to a compact subset of $End \ V$
contains a subsequence that converges
pointwise to a continuous function (and of course the convergence is uniform on compact subsets.) Taking the metric space to be the set
$\mathcal C_L \equiv \{\gamma \in \Gamma_0: \text{Length}( \gamma) \le L\}$
with the metric
$d_0(\gamma, \eta) = \sup_{0\le s \le1} |\gamma(s) - \eta(s)|$, and taking the
functions to be $\{ //_\gamma^{\alpha(t)}\}$ with ranges
contained in $K \subset End\ V$,
the estimate \eqref{LTB10} shows that we may apply this Arzela-Ascoli
argument and conclude that for any sequence of times increasing to $\infty$
there is a subsequence $t_j\uparrow \infty$
for which $//_{\gamma}^{\alpha(t_j)}$ converges in operator norm for each
curve $\gamma \in \mathcal C_L$.
We may allow $L \uparrow \infty$ through a sequence and use
diagonalization again to conclude
that there is a function $ P:\Gamma_0 \rightarrow End\ V$
such that $P(\gamma) = \lim_{j\to \infty} //_\gamma^{\alpha(t_j)}$ in operator norm
for all $\gamma \in \Gamma_0$. By \eqref{LTB10} $P|\mathcal C_L$ is continuous
in the norm $d_0$ for each $L <\infty$ and therefore is continuous on
$\Gamma_0$ in the metric $d_1$. The properties 1), 2), 3) of
Lemma \ref{loops-paths} follow from the corresponding properties of the
maps $\gamma \mapsto //_{\gamma}^{\alpha(t_j)}$. The map $P$
therefore extends, by Lemma \ref{loops-paths}, to a parallel transport
system on paths when a piecewise $C^1$ homotopy of $M^{int}$ to $x_0$
is specified. The extension is unique in the sense given in Lemma \ref{loops-paths}.
\end{proof}
\begin{acknowledgement}
N. Charalambous would like to thank the Asociaci\'{o}n Mexicana de Cultura A.C.
\end{acknowledgement}
\begin{bibdiv}
\begin{biblist}
\bib{Bal}{article}{
AUTHOR = {Balaban, T.},
TITLE = {Convergent renormalization expansions for lattice gauge
theories},
JOURNAL = {Comm. Math. Phys.},
FJOURNAL = {Communications in Mathematical Physics},
VOLUME = {119},
YEAR = {1988},
NUMBER = {2},
PAGES = {243--285},
ISSN = {0010-3616},
CODEN = {CMPHAY},
MRCLASS = {81E25 (81E08 81E15)},
MRNUMBER = {968698 (90b:81104)},
MRREVIEWER = {Claus Montonen},
URL = {http://projecteuclid.org/getRecord?id=euclid.cmp/1104162401},
}
\bib{Cha4}{article}{
AUTHOR = {Charalambous, Nelia},
TITLE = {Eigenvalue estimates for the {B}ochner {L}aplacian and
harmonic forms on complete manifolds},
JOURNAL = {Indiana Univ. Math. J.},
FJOURNAL = {Indiana University Mathematics Journal},
VOLUME = {59},
YEAR = {2010},
NUMBER = {1},
PAGES = {183--206},
ISSN = {0022-2518},
CODEN = {IUMJAB},
MRCLASS = {58J60 (35P05 35R01 58J50)},
MRNUMBER = {2666477 (2011j:58060)},
MRREVIEWER = {Julie Rowlett},
DOI = {10.1512/iumj.2010.59.3770},
URL = {http://dx.doi.org/10.1512/iumj.2010.59.3770},
}
\bib{CG1}{article}{
AUTHOR = {Charalambous, Nelia},
AUTHOR = {Gross, Leonard},
TITLE = {The {Y}ang-{M}ills heat semigroup on three-manifolds with
boundary},
JOURNAL = {Comm. Math. Phys.},
FJOURNAL = {Communications in Mathematical Physics},
VOLUME = {317},
YEAR = {2013},
NUMBER = {3},
PAGES = {727--785},
ISSN = {0010-3616},
CODEN = {CMPHAY},
MRCLASS = {58J35 (58J32 81T13)},
MRNUMBER = {3009723},
MRREVIEWER = {Thomas Krainer},
DOI = {10.1007/s00220-012-1558-0},
URL = {http://dx.doi.org/10.1007/s00220-012-1558-0},
}
\bib{CFKS}{book}{
AUTHOR = {Cycon, H. L.},
AUTHOR = {Froese, R. G.},
AUTHOR = { Kirsch, W. },
AUTHOR = {Simon, B.},
TITLE = {Schr\"odinger operators with application to quantum mechanics
and global geometry},
SERIES = {Texts and Monographs in Physics},
EDITION = {Study},
PUBLISHER = {Springer-Verlag},
ADDRESS = {Berlin},
YEAR = {1987},
PAGES = {x+319},
ISBN = {3-540-16758-7},
MRCLASS = {35-02 (35J10 47F05 58G40 81C10)},
MRNUMBER = {MR883643 (88g:35003)},
MRREVIEWER = {M. Demuth},
}
\bib{Do1}{article}{
AUTHOR = {Donaldson, S. K.},
TITLE = {Anti self-dual {Y}ang-{M}ills connections over complex
algebraic surfaces and stable vector bundles},
JOURNAL = {Proc. London Math. Soc. (3)},
FJOURNAL = {Proceedings of the London Mathematical Society. Third Series},
VOLUME = {50},
YEAR = {1985},
NUMBER = {1},
PAGES = {1--26},
ISSN = {0024-6115},
CODEN = {PLMTAL},
MRCLASS = {58E15 (14F99 53C05 57R99)},
MRNUMBER = {MR765366 (86h:58038)},
MRREVIEWER = {S. Ramanan},
}
\bib{EL}{article}{
AUTHOR = {Einav, Amit},
AUTHOR = {Loss, Michael},
TITLE = {Sharp trace inequalities for fractional {L}aplacians},
JOURNAL = {Proc. Amer. Math. Soc.},
FJOURNAL = {Proceedings of the American Mathematical Society},
VOLUME = {140},
YEAR = {2012},
NUMBER = {12},
PAGES = {4209--4216},
ISSN = {0002-9939},
CODEN = {PAMYAR},
MRCLASS = {35A23 (26D10)},
MRNUMBER = {2957211},
MRREVIEWER = {Jean Van Schaftingen},
DOI = {10.1090/S0002-9939-2012-11380-2},
URL = {http://dx.doi.org/10.1090/S0002-9939-2012-11380-2},
}
\bib{G2}{article}{
AUTHOR = {Gross, Leonard},
TITLE = {A {P}oincar\'e lemma for connection forms},
JOURNAL = {J. Funct. Anal.},
FJOURNAL = {Journal of Functional Analysis},
VOLUME = {63},
YEAR = {1985},
NUMBER = {1},
PAGES = {1--46},
ISSN = {0022-1236},
CODEN = {JFUAAW},
MRCLASS = {53C80 (53C05 58A10 81E20)},
MRNUMBER = {795515 (87a:53110)},
MRREVIEWER = {Ng{\^o} Van Qu{\^e}},
DOI = {10.1016/0022-1236(85)90096-5},
URL = {http://dx.doi.org/10.1016/0022-1236(85)90096-5},
}
\bib{HT1}{article}{
AUTHOR = {Hong, Min-Chun},
AUTHOR = {Tian, Gang},
TITLE = {Global existence of the {$m$}-equivariant {Y}ang-{M}ills flow
in four dimensional spaces},
JOURNAL = {Comm. Anal. Geom.},
FJOURNAL = {Communications in Analysis and Geometry},
VOLUME = {12},
YEAR = {2004},
NUMBER = {1-2},
PAGES = {183--211},
ISSN = {1019-8385},
MRCLASS = {53C44 (53C07 58E15)},
MRNUMBER = {MR2074876 (2005e:53103)},
MRREVIEWER = {J{\"u}rgen Eichhorn},
}
\bib{HT2}{article}{
AUTHOR = {Hong, Min-Chun},
AUTHOR = {Tian, Gang},
TITLE = {Asymptotical behaviour of the {Y}ang-{M}ills flow and singular
{Y}ang-{M}ills connections},
JOURNAL = {Math. Ann.},
FJOURNAL = {Mathematische Annalen},
VOLUME = {330},
YEAR = {2004},
NUMBER = {3},
PAGES = {441--472},
ISSN = {0025-5831},
CODEN = {MAANA},
MRCLASS = {53C44 (53C07)},
MRNUMBER = {MR2099188 (2006h:53063)},
}
\bib{HS}{article}{
AUTHOR = {Hundertmark, Dirk},
AUTHOR = {Simon, Barry},
TITLE = {A diamagnetic inequality for semigroup differences},
JOURNAL = {J. Reine Angew. Math.},
FJOURNAL = {Journal f\"ur die Reine und Angewandte Mathematik},
VOLUME = {571},
YEAR = {2004},
PAGES = {107--130},
ISSN = {0075-4102},
CODEN = {JRMAA8},
MRCLASS = {47D08 (35J10 47F05 81Q10)},
MRNUMBER = {MR2070145 (2005d:47078)},
MRREVIEWER = {Michael J. Gruber},
}
\bib{Jac}{book}{
AUTHOR = {Jackson, John David},
TITLE = {Classical electrodynamics},
EDITION = {Third},
PUBLISHER = {John Wiley \& Sons Inc.},
ADDRESS = {New York},
YEAR = {1999},
PAGES = {xxii+808},
MRCLASS = {78.00},
MRNUMBER = {MR0436782 (55 \#9721)},
MRREVIEWER = {T. Kahan},
}
\bib{LM}{book}{
AUTHOR = {Lions, J.-L.},
AUTHOR = {Magenes, E.},
TITLE = {Non-homogeneous boundary value problems and applications.
{V}ol. {I}},
NOTE = {Translated from the French by P. Kenneth,
Die Grundlehren der mathematischen Wissenschaften, Band 181},
PUBLISHER = {Springer-Verlag, New York-Heidelberg},
YEAR = {1972},
PAGES = {xvi+357},
MRCLASS = {35JXX (35KXX 35LXX 46E35)},
MRNUMBER = {0350177 (50 \#2670)},
}
\bib{L1}{article}{
AUTHOR = {L{\"u}scher, Martin},
TITLE = {Trivializing maps, the {W}ilson flow and the {HMC} algorithm},
JOURNAL = {Comm. Math. Phys.},
FJOURNAL = {Communications in Mathematical Physics},
VOLUME = {293},
YEAR = {2010},
NUMBER = {3},
PAGES = {899--919},
ISSN = {0010-3616},
CODEN = {CMPHAY},
MRCLASS = {81T25 (81T13 81T17 81T80)},
MRNUMBER = {2566166 (2011d:81217)},
MRREVIEWER = {Axel Maas},
DOI = {10.1007/s00220-009-0953-7},
URL = {http://dx.doi.org/10.1007/s00220-009-0953-7},
}
\bib{L2}{article}{
AUTHOR = {L{\"u}scher, Martin},
TITLE = {Properties and uses of the {W}ilson flow in lattice {QCD}},
JOURNAL = {J. High Energy Phys.},
FJOURNAL = {Journal of High Energy Physics},
YEAR = {2010},
NUMBER = {8},
PAGES = {071, 18},
ISSN = {1029-8479},
MRCLASS = {81V05 (81T25)},
MRNUMBER = {2756058 (2011m:81308)},
DOI = {10.1007/JHEP08(2010)071},
URL = {http://dx.doi.org/10.1007/JHEP08(2010)071},
}
\bib{L3}{article}{
AUTHOR = {L{\"u}scher, Martin},
TITLE = {Chiral symmetry and the {Y}ang-{M}ills gradient flow},
JOURNAL = {J. High Energy Phys.},
FJOURNAL = {Journal of High Energy Physics},
YEAR = {2013},
NUMBER = {4},
PAGES = {123, front matter + 39},
ISSN = {1126-6708},
MRCLASS = {81T15 (81T25 81V05)},
MRNUMBER = {3065842},
}
\bib{LW}{article}{
AUTHOR = {L{\"u}scher, Martin},
AUTHOR = {Weisz, Peter},
TITLE = {Perturbative analysis of the gradient flow in non-abelian
gauge theories},
JOURNAL = {J. High Energy Phys.},
FJOURNAL = {Journal of High Energy Physics},
YEAR = {2011},
NUMBER = {2},
PAGES = {051, i, 22},
ISSN = {1029-8479},
MRCLASS = {81T25 (81T13 81V05)},
MRNUMBER = {2820807},
DOI = {10.1007/JHEP02(2011)051},
URL = {http://dx.doi.org/10.1007/JHEP02(2011)051},
}
\bib{Pol1}{article}{
AUTHOR = {Polyakov, A. M.},
TITLE = {String representations and hidden symmetries for gauge fields},
JOURNAL = {Phys. Lett.},
FJOURNAL = {Physics Letters},
VOLUME = {82B},
YEAR = {1979},
NUMBER = {2},
PAGES = {247--250},
ISSN = {0029-5582},
CODEN = {NUPBBO},
MRCLASS = {81G05 (81E99)},
}
\bib{Pol2}{article}{
AUTHOR = {Polyakov, A. M.},
TITLE = {Gauge fields as rings of glue},
JOURNAL = {Nuclear Phys. B},
FJOURNAL = {Nuclear Physics. B},
VOLUME = {164},
YEAR = {1980},
NUMBER = {1},
PAGES = {171--188},
ISSN = {0029-5582},
CODEN = {NUPBBO},
MRCLASS = {81G05 (81E99)},
MRNUMBER = {561638 (81c:81060)},
MRREVIEWER = {Jorge Andr{\'e} Swieca},
DOI = {10.1016/0550-3213(80)90507-6},
URL = {http://dx.doi.org/10.1016/0550-3213(80)90507-6},
}
\bib{Po}{book}{
AUTHOR = {Poor, Walter A.},
TITLE = {Differential geometric structures},
PUBLISHER = {McGraw-Hill Book Co.},
ADDRESS = {New York},
YEAR = {1981},
PAGES = {xiii+338},
ISBN = {0-07-050435-0},
MRCLASS = {53-01 (53-02 53C21)},
MRNUMBER = {647949 (83k:53002)},
MRREVIEWER = {N. J. Hitchin},
}
\bib{Ra}{article}{
AUTHOR = {R{\aa}de, Johan},
TITLE = {On the {Y}ang-{M}ills heat equation in two and three
dimensions},
JOURNAL = {J. Reine Angew. Math.},
FJOURNAL = {Journal f\"ur die Reine und Angewandte Mathematik},
VOLUME = {431},
YEAR = {1992},
PAGES = {123--163},
ISSN = {0075-4102},
CODEN = {JRMAA8},
MRCLASS = {58E15 (53C07 58G11)},
MRNUMBER = {MR1179335 (94a:58041)},
MRREVIEWER = {Dennis M. DeTurck},
}
\bib{Sa}{article}{
AUTHOR = {Sadun, Lorenzo Adlai},
TITLE = {Continuum regularized Yang-Mills theory},
JOURNAL = {Ph. D. Thesis, Univ. of California, Berkeley},
FJOURNAL = {Communications in Mathematical Physics},
YEAR = {1987},
PAGES={67+ pages}
}
\bib{Sei}{book}{
AUTHOR = {Seiler, Erhard},
TITLE = {Gauge theories as a problem of constructive quantum field
theory and statistical mechanics},
SERIES = {Lecture Notes in Physics},
VOLUME = {159},
PUBLISHER = {Springer-Verlag},
ADDRESS = {Berlin},
YEAR = {1982},
PAGES = {v+192},
ISBN = {3-540-11559-5},
MRCLASS = {81E08 (81-02 81E25 82A05)},
MRNUMBER = {MR785937 (86g:81084)},
MRREVIEWER = {Claus Montonen},
}
\bib{Tay3}{book}{
AUTHOR = {Taylor, Michael E.},
TITLE = {Partial differential equations. {III}},
SERIES = {Applied Mathematical Sciences},
VOLUME = {117},
NOTE = {Nonlinear equations,
Corrected reprint of the 1996 original},
PUBLISHER = {Springer-Verlag},
ADDRESS = {New York},
YEAR = {1997},
PAGES = {xxii+608},
ISBN = {0-387-94652-7},
MRCLASS = {35-01 (46N20 47N20 58Gxx)},
MRNUMBER = {MR1477408 (98k:35001)},
MRREVIEWER = {Luigi Rodino},
}
\bib{Z}{article}{
AUTHOR = {Zwanziger, Daniel},
TITLE = {Covariant quantization of gauge fields without {G}ribov
ambiguity},
JOURNAL = {Nuclear Phys. B},
FJOURNAL = {Nuclear Physics. B},
VOLUME = {192},
YEAR = {1981},
NUMBER = {1},
PAGES = {259--269},
ISSN = {0029-5582},
CODEN = {NUPBBO},
MRCLASS = {81E10 (53C05 58D30)},
MRNUMBER = {MR635216 (82k:81062)},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
\paragraph{The Feasibility Problem:}
The existence of integer solutions for a certain system of equations has been discussed
as one of the fundamental problems in the theory of computation.
A prominent example is the Hilbert 10th problem on Diophantine equations~\cite{Matiyasevich93}.
In this paper, we study the feasibility problem of 0-1 integer programs whose constraints are only linear equalities as follows:
\begin{problem}[Feasibility of 0-1 Integer Programs with Linear Equalities] \
\label{feasibility}
\begin{center}
Find $x \in \{0,1\}^n$ which satisfies a given set of linear equalities $A x = b$.
\end{center}
\end{problem}
We give an exact algorithm running in $O(1.415^n)$-time and $O(1.190^n)$-space, which achieves a quadratic speedup compared to exhaustive search running in $O(2^n)$-time.
Our algorithm can store the data of all the feasible solutions in $O(1.415^n)$-space, even if the number of solutions is more than $O(1.415^n)$.
As a similar problem, which achieves quadratic speedup,
there is a quantum algorithm known as Grover's algorithm for unstructured database search problems~\cite{Grover97}.
It gives a correct answer with high probability, but our algorithm do not use randomness and always gives a correct answer.
Recently, probabilistic polynomial algorithms solving a system of linear equations has been discussed by Raghavendra~\cite{Raghavendra12} and Fliege~\cite{Fliege12}.
If we eliminate the 0-1 constraints, we can give a polynomial time algorithm by the Gaussian elimination.
\paragraph{The Optimization Problem:}
Then, we extend our algorithm for the following standard optimization problem running in $O(1.415^n)$-time and $O(1.190^n)$-space:
\begin{problem}[Optimization of 0-1 Integer Programs with Linear Equalities]
\label{optimization}
\begin{eqnarray*}
\label{IP_primal}
\begin{array}{ll}
\min & \displaystyle c^T x \\
s.t. & \displaystyle A x = b,\\
& x \in \{0,1\}^n.
\end{array}
\end{eqnarray*}
\end{problem}
We know that there are many sophisticated ideas (e.g., the branch-and-bound method and the cutting-plane method) improving algorithms
and implementations for computing 0-1 integer programs~\cite{GG11,Wolsey98}.
However, we don't know any improvements of worst-case time complexity for such a general setting in which elements of $A$ and $b$ can be arbitrary real numbers.
\paragraph{Exact Algorithms for NP-hard Problems:}
Since there are no polynomial time algorithms for NP-hard problems unless P=NP,
many researchers have studied exact exponential time algorithms which are faster than exhaustive search for NP-hard problems~\cite{FK10,IP01,IPZ01,Woeginger03}.
Integer programs include many NP-hard problems as special cases.
For instance, the subset sum problem is a special case of Problem~\ref{feasibility} in which the number of constraints is exactly one.
Among several such problems whose exact algorithms have been studied,
some problems (e.g., the subset sum problem~\cite{HS74,SS81}) have the same time complexity as Problem~\ref{feasibility}, and
some other problems (e.g., the exact satisfiability problem~\cite{BMS05} and the exact hitting set problem~\cite{DP02}) have algorithms faster than $O(1.415^n)$-time.
In particular, the exact satisfiability problem, which is also a special case of Problem~\ref{feasibility}, has been intensively studied~\cite{BMS05,DP02}.
On the other hand, it seems to be difficult to improve the time complexity of our algorithms due to a similar reason of NP-hardness.
In other words, if we can improve our algorithms, then we simultaneously improve the time complexity of exact algorithms
for many NP-hard problems which can be reduced to Problem~\ref{feasibility}.
\paragraph{Circuit Lower Bounds from Moderately Exponential Algorithms:}
Very recently, Impagliazzo, Lovett, Paturi and Schneider~\cite{ILPS14} studied the feasibility problem for the inequality version of 0-1 integer programs stated as follows:
\begin{problem}[Feasibility of 0-1 Integer Programs with Linear Inequalities] \
\label{inequality}
\begin{center}
Find $x \in \{0,1\}^n$ which satisfies a given set of linear inequalities $A x \geq b$.
\end{center}
\end{problem}
Impagliazzo, Lovett, Paturi and Schneider~\cite{ILPS14} gave an algorithm solving Problem~\ref{inequality} in $O(2^{(1 - \mathrm{poly}(1/d))n})$-time where $dn$ is the number of constraints.
It improves an algorithm for Problem~\ref{inequality} by Impagliazzo, Paturi and Schneider~\cite{IPS13}, which is faster than
$O(2^n)$-time only when the number of inequalities is smaller than $0.136 n$.
These results are motivated from the challenge initiated by Williams~\cite{Williams13}
for proving lower bounds for certain circuit models.
In this context, it is important to give only a modest improvement of the exponential factor from the $O(2^n)$-time exhaustive search.
\paragraph{Our Algorithms:}
Our algorithms are built on a simple combination of basic techniques on exact algorithms for NP-hard problems.
In particular, we use a classic technique called the $k$-table method studied in \cite{HS74,SS81}.
This method splits $n$-variables into the $k$ sets of $n/k$-variables,
and lists all possible $2^{n/k}$-assignments for each set.
This preprocessing enables us to give algorithms which run faster than $O(2^n)$-time for certain problems.
On the other hand, it was unclear how we construct the $k$-table
for exact algorithms to compute 0-1 integer programs with linear equalities.
The ideas introduced in the two papers~\cite{ILPS14,IPS13} give us inspiration
to overcome technical problems including analysis to bound the time complexity.
In this paper, we introduce a notion of the vector equality problem, a variation of
the vector domination problem studied by Impagliazzo, Paturi and Schneider~\cite{IPS13}, to construct the 2-table for our problems.
In Section~\ref{vector_section}, we show two algorithms solving the vector equality problem and give analysis of its time complexity.
In Section~\ref{IP_section}, we describe how we compute the feasibility problem and the optimization problem for 0-1 integer programs with linear equalities by reducing them to the vector equality problem.
A novel point here is an extension of the feasibility problem to the optimization problem by using the 2-table method.
This is achieved by post-processing after solving the feasibility problem with extra storage of the objective function.
There are no extra blow-up of the exponential time complexity.
In Section~\ref{space_section}, we improve the space complexity of our algorithms to $O(1.190^n)$ following the idea of Shroeppel and Shamir~\cite{SS81}.
\section{Notation and Definition}
Throughout the paper, we use the following notations.
We denote $m \times n$ constant matrices by $A$, and $i,j$-th element of a matrix $A$ by $A_{i,j}$.
We use $b$ and $c$ as constant vectors. $c^T$ is the transpose of $c$.
We also use $x$, $u$ and $v$ as variable vectors.
We denote $j$-th element of a vector $x$ by $x_j$.
The same notation applies for other constant and variable vectors.
The function $\mathrm{poly}(n)$ is some polynomial for $n$.
Following the convention in the theory of exact algorithms, we measure the time complexity by the function of $n$, which is the number of variables.
We assume $m \in O(\mathrm{poly}(n))$ since otherwise the input size is super-polynomial to $n$.
In this paper, we will give algorithms for 0-1 integer programs with linear equality constraints by reducing them to following problem.
\begin{definition}[Vector Equality]
Two vectors $u=(u_1,u_2,\cdots,u_m)$ and $v=(v_1,v_2,\cdots,v_m)$ are equal (i.e., $u=v$) if and only if $u_i=v_i$ for any $i~(0 \leq i \leq m)$.
Given two sets of $m$-dimensional vectors $U$ and $V$, the {\it vector equality problem} is
a problem to output information (e.g., a list of subsets of $U$ and $V$) of all the pairs of two vectors $u \in U$ and $v \in V$ such that $u = v$.
\end{definition}
Note that elements of vectors can be real numbers.
Therefore, we cannot combine a set of elements into one element unlike the case of integers whose absolute values are bounded~\cite{CD99}.
We will use the lexicographical order to compare two $m$-dimensional vectors.
\begin{definition}[Lexicographical Order]
A vector $u = (u_1,u_2,\cdots,u_m)$ is larger than another vector $v = (v_1,v_2,\cdots,v_m)$ (i.e., $u > v$) if and only if there is an index $l~(0 \leq l \leq m)$ such that
$u_l > v_l$ and $u_i=v_i$ for any $i~(0 \leq i < l)$.
\end{definition}
\section{Algorithms for the Vector Equality Problem}
\label{vector_section}
\subsection{Overview and Comparison of Two Algorithms}
In this section, we present two algorithms (Algorithm~\ref{sort_algo} and Algorithm~\ref{recursive_algo})
to solve the vector equality problem efficiently.
Algorithm~\ref{sort_algo} is simple and uses a sort routine by lexicographical order between two $m$-dimensional vectors,
while Algorithm~\ref{recursive_algo} is recursive and uses the idea of measure-and-conquer~\cite{FK10}.
Although its theoretical time complexity is the same as the other one,
Algorithm~\ref{recursive_algo} has a practical merit when $m$ (the number of constraints) is large.
Moreover, we can see the quadratic difference between
the equality and inequality versions of integer programs by looking at the recursive algorithm
compared to the algorithms for Problem~\ref{inequality}~\cite{ILPS14,IPS13}.
To analyze the time complexity of Algorithm~\ref{recursive_algo},
we need to incorporate an idea using the weighted median,
which is introduced very recently by Impagliazzo, Lovett, Paturi and Schneider~\cite{ILPS14}.
Furthermore, we complete our analysis of the time complexity by setting a suitable choice of complexity measure, which is a novel point of this paper, for the search space in the 2-table method.
Algorithm~\ref{recursive_algo} can be regarded as a vector version of the quicksort algorithm
with the weighted median as a pivot.
Actually, it is not necessary to use the weighted median in practice to select the pivot.
A heuristical choice of the pivot may be faster in many cases,
while there are no certificates of its worst-case time complexity.
After the completion of the first draft including Algorithm~\ref{recursive_algo},
we noticed that it can be simplified as Algorithm~\ref{sort_algo}.
However, we consider Algorithm~\ref{recursive_algo} is still beneficial to
present due to several reasons as mentioned above.
\subsection{A Simple Algorithm by Sorting}
\begin{algorithm}[ht]
\caption{SortVectorEquality$(U,V)$}
\begin{algorithmic}
\REQUIRE Two sets of $m$-dimensional vectors
\ENSURE A list of two sets of $m$-dimensional vectors
\STATE Sort $U$ and $V$ in the ascending lexicographical order, respectively.\\
($u^k$ and $v^k$ denote the $k$-th vectors in $U$ and $V$, respectively)
\STATE Initialize two indices $\alpha = 1$ and $\beta = 1$ for $U$ and $V$.
\WHILE{$\alpha \leq |U|$ \AND $\beta \leq |V|$}
\IF{$u^\alpha > v^\beta$}
\STATE Increment $\beta$.
\ELSIF{$u^\alpha < v^\beta$}
\STATE Increment $\alpha$.
\ELSE
\STATE Set $\alpha' := \alpha$, $\beta' := \beta$ and $w := u^\alpha$ ($:= v^\beta$).
\WHILE{$w = u^\alpha$}
\STATE Increment $\alpha$.
\ENDWHILE
\WHILE{$w = v^\beta$}
\STATE Increment $\beta$.
\ENDWHILE
\STATE Output $(\alpha',\alpha-1)$ and $(\beta',\beta-1)$ as representation of two subsets of $U$ and $V$.
\ENDIF
\ENDWHILE
\end{algorithmic}
\label{sort_algo}
\end{algorithm}
We describe a simple algorithm by sorting for solving the vector equality problem in Algorithm~\ref{sort_algo}.
Since the sort of $m$-dimensional vectors can be done in $O(m N \log N)$-time, we can also run Algorithm~\ref{sort_algo} in $O(m N \log N)$-time.
\begin{lemma}
\label{vector_lemma1}
The vector equality problem can be computed in $O(m N \log N)$-time where $|U| = |V| = N$ by Algorithm~\ref{sort_algo}.
\end{lemma}
Our algorithm can enumerate all the possible solutions.
It may sound strange that we can store the data of all possible solutions within $O(m N \log N)$-space, even if the number of all possible solutions is $\omega(m N \log N)$.
This is just because we store the data as a collection of two sets of elements.
If the number of solutions is bounded by $O(m N \log N)$, then the time complexity to enumerate all the solutions is also $O(m N \log N)$.
If the number of solutions is $\omega(m N \log N)$, then the time complexity depends on the number of possible solutions.
\subsection{A Recursive Algorithm by Measure-and-Conquer}
Another way to solve the vector equality problem is to use
a notion of the weighted median to bound the time complexity of our algorithms
in the way of measure-and-conquer.
\begin{definition}[Weighted Median]
The weighted median for a set of weighted numbers is a number such that both the total weight of numbers smaller than the weighted median
and the total weight of numbers larger than the weighted median are at most half of the total weight of all the numbers.
\end{definition}
\begin{algorithm}[ht]
\caption{RecursiveVectorEquality$(U,V,i,m)$} \label{recursive_algo}
\begin{algorithmic}
\REQUIRE Two sets of $m$-dimensional vectors and an index $i$ and the dimension $m$.
\ENSURE A list of two sets of $m$-dimensional vectors.
\IF{$U = \emptyset$ or $V = \emptyset$}
\RETURN an empty list
\ELSIF{$i > m$}
\RETURN a singleton list of $(U,V)$
\ELSE
\STATE
\begin{enumerate}
\item Find the weighted median $k$ of the $i$-th coordinates of $U \cup V$ with weight $|V|$ and $|U|$ for each element in $U$ and $V$, respectively.
\item Partition $U$ into three sets:
\begin{enumerate}
\item $U^+ = \{ u \mid u_i > k \}$,
\item $U^= = \{ u \mid u_i = k \}$,
\item $U^- = \{ u \mid u_i < k \}$.
\end{enumerate}
\item
Partition $V$ into three sets:
\begin{enumerate}
\item $V^+ = \{ u \mid v_i > k \}$,
\item $V^= = \{ u \mid v_i = k \}$,
\item $V^- = \{ u \mid v_i < k \}$.
\end{enumerate}
\item
Solve the following three subproblems:
\begin{enumerate}
\item L1 = VectorEquality$(U^+,V^+,i,m)$
\item L2 = VectorEquality$(U^=,V^=,i+1,m)$
\item L3 = VectorEquality$(U^-,V^-,i,m)$
\end{enumerate}
\end{enumerate}
\RETURN the concatenation of the three lists L1, L2, and L3
\ENDIF
\end{algorithmic}
\end{algorithm}
Then, we consider a recursive algorithm (Algorithm~\ref{recursive_algo}) computing the vector equality problem.
Following a linear time algorithm for the unweighted median problem~\cite{Blum72},
we can give a linear time algorithm for the weighted median problem~\cite{BO83}, which is also indicated in \cite{ILPS14}.
\begin{lemma}[\cite{BO83,ILPS14}]
The weighted median of $N$ numbers can be computed in $O(N)$-time.
\end{lemma}
We analyze the time complexity of Algorithm~\ref{recursive_algo} for the vector equality problem in the following lemma.
\begin{lemma}
\label{vector_lemma2}
The vector equality problem can be computed in $O(m N \log N)$-time where $|U| = |V| = N$ by starting Algorithm~\ref{recursive_algo} at RecursiveVectorEquality$(U,V,1,m)$.
\end{lemma}
\begin{proof}
In Algorithm~\ref{recursive_algo}, we find the weighted median $k$ of the $i$-th coordinates of $U \cup V$ where all the elements in $U$ and $V$ have weight $|V|$ and $|U|$, respectively.
Then, we partition each of $U$ and $V$ into three sets, respectively.
Two vectors $u \in U$ and $v \in V$ can be equal in at most one of the following three cases:
\begin{center}
(1)~ $u \in U^+$ and $v \in V^+$,\\
(2)~ $u \in U^=$ and $v \in V^=$,\\
(3)~ $u \in U^-$ and $v \in V^-$.\\
\end{center}
We solve smaller subproblems of the vector equality problem for the three cases.
In particular, we decrease the dimension $m$ to $m-1$ in the case of (2).
The rule of the partition immediately gives the following equation:
\begin{align*}
|V| \cdot (|U^+| + |U^=| + |U^-|) + |U| \cdot (|V^+| + |V^=| + |V^-|)
= |V| \cdot |U| + |U| \cdot |V|.
\end{align*}
Dividing it by $|U| \cdot |V|$, we have
\[
\frac{|U^+| + |U^=| + |U^-|}{|U|} + \frac{|V^+| + |V^=| + |V^-|}{|V|} = 2.
\]
For some constants $s$ and $t$ such that $0 \leq s \leq 1$ and $0 \leq t \leq 1$, we have
\begin{align*}
&\frac{|U^+|}{|U|} + \frac{|V^+|}{|V|} = 1 - s,\\
&\frac{|U^-|}{|U|} + \frac{|V^-|}{|V|} = 1 - t,\\
&\frac{|U^=|}{|U|} + \frac{|V^=|}{|V|} = s + t
\end{align*}
because we partitioned $U$ and $V$ at the weighted median.
Since $\alpha + \beta \geq 2\sqrt{\alpha \beta} $ for any $\alpha,\beta \geq 0$, we have
\begin{align*}
&\frac{|U^+|}{|U|} \cdot \frac{|V^+|}{|V|} \leq \frac{1}{4} \cdot (1 - s )^2,\\
&\frac{|U^-|}{|U|} \cdot \frac{|V^-|}{|V|} \leq \frac{1}{4} \cdot (1 - t )^2,\\
&\frac{|U^=|}{|U|} \cdot \frac{|V^=|}{|V|} \leq \frac{1}{4} \cdot (s + t )^2.
\end{align*}
Collecting these inequalities, we have
\begin{align*}
&|U^+| \cdot |V^+| \cdot 2^m + |U^-| \cdot |V^-| \cdot 2^m + |U^=| \cdot |V^=| \cdot 2^{m-1} \\
\leq &\frac{1}{4} \cdot \{ (1 - s )^2 \cdot 2^m + (1 - t )^2 \cdot 2^m + (s + t )^2 \cdot 2^{m-1} \} \cdot |U| \cdot |V|\\
= &\frac{1}{4} \cdot \{ (1 - s )^2 + (1 - t )^2 + \frac{1}{2} \cdot (s + t )^2 \} \cdot |U| \cdot |V| \cdot 2^m.
\end{align*}
It means that the search space $|U| \cdot |V| \cdot 2^m$ decreases by the factor of
\begin{align*}
f (s,t) &= \frac{1}{4} \cdot \{ (1 - s )^2 + (1 - t )^2 + \frac{1}{2} (s + t )^2 \}\\
&= 0.5 - 0.5 s - 0.5 t + 0.375 {s}^2 + 0.375 {t}^2 + 0.25 s t
\end{align*}
at each recursion.
We can conclude $f (s,t) \leq \frac{1}{2}$ in the domain of $0 \leq s \leq 1$ and $0 \leq t \leq 1$ by the following argument.
By taking the partial derivatives, we have
\begin{align*}
\frac{\partial f (s,t)}{\partial s} = - 0.5 + 0.75 s + 0.25 t,\\
\frac{\partial f (s,t)}{\partial t} = - 0.5 + 0.25 s + 0.75 t.
\end{align*}
If $\frac{\partial f (s,t)}{\partial s} > 0$ (equivalently, $t > 2 - 3 s$), then the function $f(s,t)$ is monotonically increasing in the direction of $s$.
If $\frac{\partial f (s,t)}{\partial s} < 0$ (equivalently, $t < 2 - 3 s$), then the function $f(s,t)$ is monotonically decreasing in the direction of $s$.
The same thing applies for $t$ instead of $s$.
Therefore, we can verify that it is maximized at two edges $(s,t) = (0,0), (1,1)$ as $f (s,t) = 0.5$
and minimized at the middle point $(s,t) = (0.5, 0.5)$ as $f (s,t) = 0.25$.
Moreover, maximal points except the two edges are only two points $(s,t) = (0,1), (1,0)$
as $f (s,t) = 0.375$.
The recursions occur at most $\log_2 (|U| \cdot |V| \cdot 2^m) \in O(m\log N)$ depth.
At each depth $d$ of the recursion, we need to solve at most $3^d$ ($< N$) subproblems of
the vector equality problem, but the total number of elements is at most $2N$.
Therefore, we can solve the weighted median in linear time $O(|U|+|V|) = O(N)$ as a whole
at each depth of the recursion.
As a consequence, we conclude that the total time complexity of Algorithm~\ref{recursive_algo} is $O(m N \log N)$.
\qed
\end{proof}
\section{Exact Algorithms for 0-1 Integer Programs}
\label{IP_section}
In this section, we give an exact algorithm for solving the feasibility and optimization problem of 0-1 integer programs with linear equality constraints
by reducing it to the vector equality problem described in the previous section.
\begin{theorem}
\label{feasibility_thm}
The feasibility and optimization problem of 0-1 integer programs with linear equalities (Problem~\ref{feasibility} and Problem~\ref{optimization}) can be computed in $O(m \cdot 2^{n/2} \mathrm{poly}(n))$-time.
\end{theorem}
\begin{proof}
We solve the feasibility problem of $Ax=b$ by reducing it to the vector equality problem.
First, we partition the set of variables $X = \{x_1, \cdots, x_n\}$ into two disjoint subsets $X_1$ and $X_2$.
Here, we assume the number of variables $n$ is even without loss of generality.
Let $\varphi(x_j)$ be assignments of $x_j$.
Then, we define vectors $u$ and $v$ by
\begin{align*}
&u_i = \sum_{x_j \in X_1} A_{ij} \cdot \varphi(x_j),
&v_i = b_i - \sum_{x_j \in X_2} A_{ij} \cdot \varphi(x_j).
\end{align*}
for each assignment of $X_1$ and $X_2$. Let $U$ and $V$ be two sets of $2^{n/2}$ such vectors $u$ and $v$, respectively.
Taking into account the linearity of the objective function $c^T x$ of Problem~\ref{optimization}, we can extend the algorithm
for the feasibility problem to one for the optimization problem.
For this purpose, we additionally calculate weight
\begin{align*}
&w(u) = \sum_{x_j \in X_1} c_j \cdot \varphi(x_j),
&w(v) = \sum_{x_j \in X_2} c_j \cdot \varphi(x_j)
\end{align*}
for each of $u \in U$ and $v \in V$, respectively.
From the construction of $U$ and $V$, there is a 0,1-vector $x \in \{0,1\}^n$ satisfying $Ax = b$ if and only if
there is a pair of two vectors $u \in U$ and $v \in V$ satisfying $u_i = v_i$ for all $i$ ($1 \leq i \leq m$).
We can solve the vector equality problem for $U$ and $V$ in $O(m N \log N)$-time by using algorithms in Section~\ref{vector_section}.
After the algorithms terminates, we can get a list of submatrices which contains information of all the possible solutions.
\[
(U^1,V^1), (U^2,V^2), \cdots, (U^k,V^k), \cdots , (U^l,V^l)
\]
There are at most $N = 2^{n/2}$ submatrices.
From the construction of the algorithms, each row and column of submatrices has no intersection.
Let $U^k \times V^k$ ($U^k \subseteq U$ and $V^k \subseteq V$) be one of such submatrices.
Then we would like to solve the following optimization problem for each $k$.
\begin{eqnarray*}
\begin{array}{ll}
\min & w(u) + w(v) \\
s.t. & u \in U^k \mbox{ and } v \in V^k.
\end{array}
\end{eqnarray*}
From the linearity of $c^T$, $w(u)$ and $w(v)$ are independent.
Therefore , the above minimization problem is solvable separately for $u$ and $v$.
Hence, $O(|U^k| + |V^k|)$-time is sufficient to optimize.
We solve the same problem for each submatrices and take the minimum of all the problems.
The total time complexity is $O(m \cdot 2^{n/2} \mathrm{poly}(n))$.
\qed
\end{proof}
\section{Improved Space Complexity}
\label{space_section}
Shroeppel and Shamir~\cite{SS81} studied the $k$-table method, which is a generalization of the 2-table method,
and showed an $O(2^{n/4})$-space exact algorithm for the subset sum problem (a special case of the Problem~\ref{feasibility}) by using the 4-table method.
Following the idea of Shroeppel and Shamir~\cite{SS81} using the priority queue,
we can reduce the space complexity of our algorithms from $O(2^{n/2})$ to $O(2^{n/4})$ as in the following theorem.
\begin{theorem}
The feasibility and optimization problem of 0-1 integer programs with linear equality constraints
(Problem~\ref{feasibility} and Problem~\ref{optimization}) can be computed
in $O(m \cdot 2^{n/2} \mathrm{poly}(n))$-time and $O(m \cdot 2^{n/4} \mathrm{poly}(n))$-space.
\end{theorem}
\begin{proof}
We partition the set of variables $X = \{x_1, \cdots, x_n\}$ into four disjoint subsets $X_1$, $X_2$, $X_3$ and $X_4$.
Here, we assume the number of variables $n$ can be divided by 4 without loss of generality.
Let $\varphi(x_j)$ be an assignment of $x_j$.
Then, we define vectors $u$, $v$, $s$ and $t$ by
\begin{align*}
&u_i = \sum_{x_j \in X_1} A_{ij} \cdot \varphi(x_j),
&v_i = \sum_{x_j \in X_2} A_{ij} \cdot \varphi(x_j),\\
&s_i = - \sum_{x_j \in X_3} A_{ij} \cdot \varphi(x_j),
&t_i = b_i - \sum_{x_j \in X_4} A_{ij} \cdot \varphi(x_j).
\end{align*}
for each assignment of $X_1$, $X_2$, $X_3$ and $X_4$.
Let $U$, $V$, $S$ and $T$ be four sets of $2^{n/4}$ such vectors $u$, $v$, $s$ and $t$, respectively.
We additionally calculate weight
\begin{align*}
&w(u) = \sum_{x_j \in X_1} c_j \cdot \varphi(x_j),
&w(v) = \sum_{x_j \in X_2} c_j \cdot \varphi(x_j),\\
&w(s) = \sum_{x_j \in X_3} c_j \cdot \varphi(x_j),
&w(t) = \sum_{x_j \in X_4} c_j \cdot \varphi(x_j).
\end{align*}
for each of $u \in U$, $v \in V$, $s \in S$ and $t \in T$, respectively.
For each vector, we can store data of its corresponding assignment and weight within $O(n)$-space.
From the construction of the four sets, there is a 0,1-vector $x \in \{0,1\}^n$ satisfying $Ax = b$ if and only if
a quartet of four vectors $u \in U$, $v \in V$, $s \in S$ and $t \in T$ such that $u + v = s + t$.
We can search such quartets by Algorithm~\ref{pq_algo} in
$O(m N^2 \log N)$-time and $O(m N)$-space where $|U| = |V| = |S| = |T| = N$.
In the algorithm, we use priority queues in which we can push and pop any element
in the logarithmic time to the number of elements.
We can compute the minimum objective value and the corresponding assignment of the original 0-1 integer programs by Algorithm~\ref{pq_algo}
where the inputs are given as four sets of $m$-dimensional vectors with their assignments $\varphi$ and weights $w$.
The return value of $\infty$ means that the problem is infeasible.
If it is feasible, we can retrieve the corresponding assignment $\varphi$ of variables from the values of $\mathbf{SOL}$ in Algorithm~\ref{pq_algo}.
\qed
\end{proof}
\begin{corollary}
The feasibility and optimization problems of 0-1 integer programs with linear equality constraints (Problem~\ref{feasibility} and Problem~\ref{optimization})
can be computed in $O(1.415^n)$-time and $O(1.190^n)$-space.
\end{corollary}
\begin{algorithm}
\caption{VectorSumEquality$(U,V,S,T)$}
\begin{algorithmic}
\STATE Sort $U$, $V$ $S$ and $T$ in the ascending lexicographical order, respectively. \\
($u^k$, $v^k$, $s^k$, and $t^k$ denote the $k$-th vectors in $U$, $V$ $S$ and $T$, respectively)
\STATE Set $\mathbf{MIN} := \infty$ and initialize two priority queues $Q_1$ and $Q_2$ as empty sets.
\FOR{$k$ = 1 \TO $|V|$}
\STATE Push $(u^k, v^1)$ to the priority queue $Q_1$.
\ENDFOR
\FOR{$k$ = 1 \TO $|T|$}
\STATE Push $(s^k, t^1)$ to the priority queue $Q_2$.
\ENDFOR
\WHILE{Both of $Q_1$ and $Q_2$ are not empty}
\STATE Take the top elements $(u^\alpha,v^\beta)$ and $(s^\gamma,t^\delta)$ from $Q_1$ and $Q_2$, respectively.
\IF{$u^\alpha + v^\beta < s^\gamma + t^\delta$}
\STATE Pop $(u^\alpha,v^\beta)$.
\IF{$\beta + 1 \leq |V|$}
\STATE Push $(u^\alpha,v^{\beta+1})$.
\ENDIF
\ELSIF{$u^\alpha + v^\beta > s^\gamma + t^\delta$}
\STATE Pop $(s^\gamma,t^\delta)$.
\IF{$\delta + 1 \leq |T|$}
\STATE Push $(s^\gamma,t^{\delta+1})$.
\ENDIF
\ELSE
\STATE $w := u^\alpha + v^\beta (:= s^\gamma + t^\delta)$;
\STATE $\mathbf{MIN_1} := \infty$;~ $\mathbf{MIN_2} := \infty$;
\WHILE{$Q_1 \neq \emptyset$ and $w = u^{\alpha'} + v^{\beta'}$ where $(u^{\alpha'},v^{\beta'})$ is the top element of $Q_1$}
\STATE Pop $(u^{\alpha'},v^{\beta'})$.
\IF{$\beta' + 1 \leq |V|$}
\STATE Push $(u^{\alpha'},v^{\beta'+1})$.
\ENDIF
\IF{$\mathbf{MIN_1} > w(u^{\alpha'}) + w(v^{\beta'})$}
\STATE $\mathbf{MIN_1} := w(u^{\alpha'}) + w(v^{\beta'})$;~ $\mathbf{SOL_1} := (\alpha',\beta')$;
\ENDIF
\ENDWHILE
\WHILE{$Q_2 \neq \emptyset$ and $w = s^{\gamma'} + t^{\delta'}$ where $(s^{\gamma'},t^{\delta'})$ is the top element of $Q_2$}
\STATE Pop $(s^{\gamma'},t^{\delta'})$.
\IF{$\delta' + 1 \leq |V|$}
\STATE Push $(s^{\gamma'},t^{\delta'+1})$.
\ENDIF
\IF{$\mathbf{MIN_2} > w(s^{\gamma'})+w(t^{\delta'})$}
\STATE $\mathbf{MIN_2} := w(s^{\gamma'})+w(t^{\delta'})$;~ $\mathbf{SOL_2} := (\gamma',\delta')$;
\ENDIF
\ENDWHILE
\IF{$\mathbf{MIN} > \mathbf{MIN_1} + \mathbf{MIN_2}$}
\STATE $\mathbf{MIN} := \mathbf{MIN_1} + \mathbf{MIN_1}$;~ $\mathbf{SOL} := (\mathbf{SOL_1}, \mathbf{SOL_2})$;
\ENDIF
\ENDIF
\ENDWHILE
\STATE Return $\mathbf{MIN}$
\end{algorithmic}
\label{pq_algo}
\end{algorithm}
\section{Conclusions}
In this paper, we have presented $O(1.415^n)$-time and $O(1.190^n)$-space exact algorithms for 0-1 integer programs with linear equality constraints.
We can apply our algorithms to the optimization problem as well as the feasibility problem.
We can also extend our algorithms to integer programs where their variables are constrained by any finite set of integers.
There are several recent progress on the subset sum problem such as time-space tradeoff results~\cite{AKKM13} and
improved algorithms for a certain important class of the subset sum problem~\cite{BCJ11,HG10}.
It would be interesting to investigate in these directions with connection to our results concerned with 0-1 integer programs.
Our computational experiments show that our algorithms can solve 0-1 integer programs with around 60 variables which are generated in a random way.
Some of well-known IP solvers cannot solve these instances because they do not have any favorable structures to cut down the search space.
By connecting our algorithms to existing techniques for 0-1 integer programs (e.g., the branch-and-bound method),
we hope that our algorithms will be useful from the practical point of view as well as theoretical analysis.
|
\section{Introduction}
Ever since the discovery of massive tidal tails emerging from the globular
cluster Palomar\,5 \citep{ode01}, the Milky Way field population has been
the subject of intense scrutiny to find more of these cold stellar streams.
The main incentives for finding streams are their potential use to constrain
the shape and mass of the Milky Way dark matter halo
\citep[e.g.][]{joh05,kop10,var11,pen12,ver13,lux13,deg14}, their ability to
constrain the existence of mini-halos \citep{iba02,sie08,yoo11,car12}, as
well as the fossil record they provide of the mass assembly history of the
Milky Way.
\begin{figure*}
\includegraphics[width=13.5cm]{f1.eps}
\caption{{\it Top}: Density of stars with dereddened colours and magnitudes
consistent with the main-sequence turn-off of an old and metal-poor
population at heliocentric distances of 8--12~kpc (see selection box
in Figure~\ref{fig:cmd}), shown in equatorial ({\it left}) and Galactic
({\it right}) coordinates. Darker areas indicate higher stellar density.
{\it Bottom}: Reddening maps of the corresponding fields, derived from
Pan-STARRS1 stellar photometry \citep[see][]{sch14}. The colour scale
is logarithmic; white (black) corresponds to E(B$-$V)~=~0.17 (0.58).
The stellar density and reddening maps have been smoothed
with a gaussian kernel with full-width at half maximum of 12$\arcmin$ and
6$\arcmin$, respectively. The stream is located close to the centre of each
panel. The thick lines in the right panels trace the best-fitting great circle
containing the stream.}
\label{fig:map}
\end{figure*}
While a few streams have been revealed by kinematics alone
\citep[e.g.][]{hel99,new09,wil11}, the vast majority were found by searching
for coherent stellar over-densities in the homogeneous, wide-field photometric
catalogue provided by the {\it Sloan Digital Sky Survey} \citep[SDSS;][]{yor00}
\citep[e.g.][]{gri06a,gri06b,gri06c,gri06d,gri09,gri12,gri13,bel06,bel07,bon12},
and more recently in VST ATLAS \citep{kop14}.
As a result of SDSS's predominant high-latitude, Northern hemisphere coverage,
the known streams are located far from the Galactic disc and bulge that were
mostly avoided by these surveys.
The \protect \hbox {Pan-STARRS1}\ \citep[PS1;][]{kai10} 3$\pi$ Survey, observing the whole sky visible
from Hawaii, has the advantage of providing some of the first deep imaging
of the dense inner regions of our Galaxy. With a sky coverage spanning twice
that of the SDSS footprint, it offers the possibility to survey these less
well-studied areas to seek new tidal streams as well as further extensions
of already known ones.
In this paper, we report the discovery of a very thin stellar stream located
in the inner Galaxy close to the Galactic bulge. It was found serendipitously
when analysing PS1 data of wide areas around nearby globular clusters for
the presence of tidal debris.
We briefly describe the PS1 survey in Section~\ref{ps1}, and present the
new stream and measurements of its properties in Section~\ref{phot}. A
summary is given in Section~\ref{summ}.
\section{The \protect \hbox {Pan-STARRS1}\ 3$\pi$ Survey}\label{ps1}
The PS1 3$\pi$ Survey \citep[K.\,C.\ Chambers et al., in preparation]{kai10}
is being carried out with the 1.8~m optical telescope installed on the peak of
Haleakala in Hawaii. Thanks to the 1.4-Gigapixel imager \citep{ona08,ton09}
covering a 7 square degree field-of-view ($\sim3.3\degr$ diameter), it is
observing the whole sky north of $\delta>-30 \degr$ in five optical to
near-infrared bands \citep[\gps\rps\ips\zps\yps;][]{ton12b} up to four times per year.
The exposure time ranges from 30 to 45 seconds, leading to median 5$\sigma$
limiting AB magnitudes of 21.9, 21.8, 21.5, 20.7, and 19.7 for individual
exposures in the \gps\rps\ips\zps\yps bands, respectively \citep{mor12}. At the end of
the survey, the 12 or so images per band will be stacked, increasing the
depth of the final photometry by $\sim$1.2~mag \citep{met13}.
The individual frames are automatically processed with the Image Processing
Pipeline \citep{mag06} to produce a photometrically and astrometrically
calibrated catalogue. A detailed description of the general PS1 data
processing is given in \citet{ton12a}. The analysis presented in this paper
is based on the photometric catalogue obtained by averaging the magnitudes
of objects detected in individual exposures \citep{sch12}. At the end of the
survey, the point source catalogue will be based on detections in the stacked
images, leading to a significantly deeper photometry and better constraints
on the tidal streams.
The catalogue used here was first corrected for foreground reddening by
interpolating the extinction at the position of each source using the
\citet{sch14} dust maps with the extinction coefficients of \citet{sch11}.
In this part of the sky, E(B$-$V) ranges from 0.17 to $\sim$1.2 (see
Figure~\ref{fig:map}).
We then cleaned the catalogue by rejecting non-stellar objects using the
difference between PSF and aperture magnitudes, as well as poorly measured
stars by keeping only objects with a signal-to-noise ratio of 10 or higher.
\begin{figure}
\includegraphics[width=8.5cm]{f2.eps}
\caption{Extinction corrected CMDs of the stream (a), a nearby field (b),
the difference between the two (c), and the significance (d). The MSTO
selection box used to produce Figure~\ref{fig:map} is shown by a red line in
the top panels. The red lines in the last panel show the fiducial and HB of
NGC\,5904 (M\,5) calculated from the PS1 photometry by \citet{ber14}. A few
blue HB stars are visible in panel (a) at ($\ensuremath{g_{\rm P1}}-\ensuremath{i_{\rm P1}}$)$_0\sim-$0.5 and
$i_{P1,0}\sim$~16.}
\label{fig:cmd}
\end{figure}
\section{A New Stream in Ophiuchus}\label{phot}
The stream appears as a coherent structure in the maps showing the stellar
density of objects with colours and magnitudes corresponding to the old,
metal-poor main-sequence turn-off (MSTO) of nearby globular clusters. The
colour and magnitude cuts were then refined based on the distribution of
stars in the colour-magnitude diagram (CMD) of the stream region (see below).
The resulting map is shown
in Figure~\ref{fig:map} in both equatorial and Galactic coordinates, along
with the corresponding reddening maps from \citet{sch14}. While there is a
little residual substructure due to the strong differential reddening in
this part of the sky, the stream is obvious at $\alpha \sim 242\degr$ and
$\delta \sim -7\degr$ ($l\sim 5\degr$, $b\sim +32\degr$). There are no
features in the dust map that could produce such an artifact.
In fact, we checked that we recover the stream whether or not we apply the
reddening correction.
Figure~\ref{fig:cmd} shows the CMD of a 9$\arcmin$ wide box centred on the
stream and extending between 241\fdg5~$<\alpha<$~243\fdg5, as well as a nearby
comparison region of the same area. The bottom panels show the residuals and
the significance of the residuals in Poissonian sigmas. The MSTO is visible
as an over-density at ($\ensuremath{g_{\rm P1}}-\ensuremath{i_{\rm P1}}$)$_0\sim$~0.2 and $i_{P1,0}\sim$~19, and is
highlighted in the top panels by a red box. A few blue
horizontal-branch (HB) stars are also detected at ($\ensuremath{g_{\rm P1}}-\ensuremath{i_{\rm P1}}$)$_0\sim-$0.5 and
$i_{P1,0}\sim$~16. As expected from such a sparse HB, no RR~Lyrae star could be
found at the location and distance of the stream in the catalogue of
\citet{dra13}. The lack of obvious sub-giant and red giant branches does not
allow tight constraints on the constituent stellar populations, although the
presence of a blue HB suggests an old and metal-poor population ($\ga$~10~Gyr
old, [Fe/H]~$\la-$~1.0). The features of the stream CMD are well fitted
by the fiducial of the old globular cluster NGC\,5904 (M\,5, [Fe/H]~$\sim-$1.3)
shown as red lines \citep[from][]{ber14}. The fiducial was first corrected for
the reddening along the line of sight to NGC\,5904
\citep[E(B$-$V)~$\sim$~0.09;][]{sch14}, then shifted by $+$0.5~mag to match
the luminosity of the stream MSTO. A small colour shift (0.07~mag to the blue)
improves the fit to the MSTO, possibly due to the high and varying reddening
in this region. The HB of the fiducial also provides a good fit to the
observed blue HB stars, giving further support that the estimated distance is
reliable. Assuming (m$-$M)$_0$~=~14.44$\pm$0.02 for NGC\,5904 \citep{cop11}
yields a true distance modulus of (m$-$M)$_0$~=~14.9$\pm$0.2
(i.e.\ 9.5$\pm$0.9~kpc) for the Ophiuchus stream. Finally, if we assume
that the Sun is located 8~kpc from the Galactic centre (GC), we find a
Sun--GC--stream angle of $\sim$89 degrees, placing the stream almost directly
above the Galactic bulge at a Galactocentric distance of 5.0~$\pm$~1.0~kpc.
\input{tab1}
To estimate the total luminosity of the stream, we used {\sc IAC-Star}
\citep{apa04} to generate the CMD of an old population (11.5--12.5~Gyr)
containing 10$^5$ stars with a metallicity comparable to that of NGC\,5904
and with the same depth as our observations -- stars fainter than our
detection limit have a negligible contribution to the total magnitude.
In Figure~\ref{fig:cmd}, we find that there are about 500 more stars in
panel (a) containing the stream than in the comparison field (panel (b)).
Summing the luminosity of a sample of 500 stars extracted randomly from the
synthetic CMD gives the total flux. We repeated this step 10$^4$ times,
extracting between 300 and 700 stars each time, and found a total
magnitude and luminosity of $M_V=-3.0\pm0.5$ and
$L_V=1.4 \pm 0.6 \times 10^3 L_{\sun}$, respectively.
These are comparable to some of the fainter halo clusters (e.g.\ Whiting\,1,
Terzan\,9, Palomar\,1, Palomar\,13: \citealt[2010 Edition]{har96}),
although a significant fraction of the stars may have already been stripped
off and lie further out along the stream orbit.
\begin{figure*}
\includegraphics[width=13.5cm]{f3.eps}
\caption{{\it Left}: MSTO star density as a function of stream latitude.
The stream is detected as a strong over-density that is well fitted
by a gaussian with $\sigma$~=~2.99$\arcmin\pm$0.33$\arcmin$.
{\it Right}: Stellar density as a function of stream longitude. The blue
line and shaded area represent the mean background contamination and its
dispersion (see text).}
\label{fig:lenwid}
\end{figure*}
The extent of the stream on the sky was calculated by reprojecting the
stellar maps to a new spherical coordinate system ($\Lambda$,$B$) with
the pole located at ($\alpha$,$\delta$) = (184\fdg32, +77\fdg25)
[($l$,$b$) = (125\fdg37, +39\fdg72)]. In this system the stream approximately
lies along the equator -- shown by the red lines in Figure~\ref{fig:map} --
making it easier to measure its width and length. In Figure~\ref{fig:lenwid},
we show the distribution of MSTO stars across the $\Lambda$ and $B$
dimensions. The left panel shows the stellar density across the stream,
where all the MSTO stars in the range $-1\degr<\Lambda<1\degr$ have been
used. The stream is detected as a significant over-density at $B\sim0\degr$.
The profile is best fit by a gaussian with $\sigma$~=
2.99$\arcmin\pm$0.33$\arcmin$, corresponding to a full-width at half maximum
(FWHM) of 7.0$\arcmin\pm$0.8$\arcmin$, i.e.\ 19~$\pm$~2~pc at the distance of
the stream. However, we note that in Figure~\ref{fig:map} the stream appears
slightly more curved than the great circle shown, implying that the
intrinsic width may be even smaller. The right panel shows the profile along
the length of the stream, for which we used the
MSTO stars in a narrow strip (12$\arcmin$) centred on the stream. To estimate
the background, we selected MSTO stars in 6 identical strips at higher and
lower $B$; their average and dispersion are shown as the blue line and
the shaded area. The on-stream histogram shows a significant over-density
between $-1\fdg2\la\Lambda\la1\fdg3$, and is therefore about 2\fdg5
long (i.e.\ $\sim$400~pc in projection).
Table~\ref{tab:prop} summarizes the estimated properties of the stream:
we list the approximate coordinates of its centre, the heliocentric and
Galactocentric distances, the extent on the sky, and the estimated luminosity.
\section{Discussion and Conclusions}\label{summ}
We have identified a new stellar stream in the constellation of Ophiuchus,
a part of the sky that has rarely been searched for streams because of the
high stellar density and significant differential and foreground reddening.
Both the morphology of the MSTO and the presence of a blue HB are typical
of an old and metal-poor population ($\ga$~10~Gyr old, [Fe/H]~$\la-$~1.0).
These properties, along with the small width of the stream and absolute
magnitude suggest a globular cluster as progenitor.
We find that the stream is exceptionally short ($\sim$2\fdg5,
i.e.\ $\sim$400~pc, in projection) compared
to all the other streams found so far, that are usually several tens of
degrees long. For comparison, the shortest stream known to date is the
Pisces--Triangulum stream, which has been traced over $\sim$15\degr
(i.e.\ $\sim$7~kpc) on the sky \citep[e.g.][]{bon12,mar14}.
We have explored shifting the MSTO selection box in magnitude to account
for possible effects of differential reddening residuals and extension
of the stream along the line of sight, but failed to detect any over-density
beyond the current extent. This experiment did reveal a possible distance
gradient along the stream, with the eastern tail being closer to the Sun,
although the apparent change in MSTO magnitude could simply be a consequence
of the differential reddening.
Surprisingly, neither the stellar density along the stream nor a careful
visual inspection of the images reveals a potential remnant of the
progenitor. This suggests that it has already been completely disrupted.
Given that the length of a tidal stream is a function of the time since the
stars became unbound, a short stream may indicate that the progenitor has
been disrupted only recently. However, this scenario is hard to reconcile
with the lack of an obvious progenitor in the vicinity of the stream.
Another possibility is that we are observing the stars of a fully disrupted
cluster at apogalacticon on a highly elliptical orbit: at this point of the
orbit, unbound stars tend to clump together because of the slower orbital
velocity \citep[e.g.][]{deh04,kup12}. The data currently available are not
sufficient to reliably trace the orbit of the progenitor, which would help us
understand its past evolution and likely fate; radial velocities and
proper motions of a sample of stream members will be crucial for this purpose.
Spectroscopic metallicities of stream stars will also shed light on the
nature of the progenitor, and help understand how such events may contribute
to the stellar populations in the central regions of the Galaxy.
For example, recent spectroscopic surveys of the outer Galactic bulge have
revealed the presence of a significant number of metal-poor stars
([Fe/H]~$<-$1; e.g.\ \citealt{gon11,gar13}). Their origin may be linked to
tidal stripping events such as the one we are witnessing with this stream.
\section*{Acknowledgments}
E.J.B.\ and A.M.N.F.\ are grateful to Douglas C.\ Heggie and Anna~Lisa Varri
for enlightening discussions. The authors would like to thank the anonymous
referee for a prompt report and useful comments. This research was supported
by a consolidated grant from the Science Technology and Facilities Council.
E.F.S. and N.F.M. acknowledge support from the DFG's grant SFB881 (A3) ``The
Milky Way System". N.F.M. gratefully acknowledges the CNRS for support through
PICS project PICS06183. The research leading to these results has received
funding from the European Research Council under the European Union's Seventh
Framework Programme (FP 7) ERC Grant Agreement n.\ [321035].
The PS1 Surveys have been made possible through contributions of
the Institute for Astronomy, the University of Hawaii, the Pan-STARRS Project
Office, the Max-Planck Society and its participating institutes, the Max Planck
Institute for Astronomy, Heidelberg and the Max Planck Institute for
Extraterrestrial Physics, Garching, The Johns Hopkins University, Durham
University, the University of Edinburgh, Queen's University Belfast, the
Harvard-Smithsonian Center for Astrophysics, the Las Cumbres Observatory Global
Telescope Network Incorporated, the National Central University of Taiwan, the
Space Telescope Science Institute, the National Aeronautics and Space
Administration under Grant No.\ NNX08AR22G issued through the Planetary Science
Division of the NASA Science Mission Directorate, the National Science
Foundation under Grant No.\ AST-1238877, and the University of Maryland.
This work has made use of the IAC-STAR Synthetic CMD computation code.
IAC-STAR is supported and maintained by the computer division of the Instituto
de Astrof\'isica de Canarias.
|
\section{Methods}
Our experiments were carried out in a \textsc{Specs} JT-STM, an ultra-high vacuum scanning tunneling microscope operating at a base temperature of 1.2 K.
Spectra of the differential conductance $dI/dV(V)$ were acquired under open-feedback conditions with standard lock-in technique using a modulation frequency of $f=912$~Hz and an amplitude of $V_{rms}=30 - 50~\mu$V.
The Pb(111) surface (critical temperature $T_c=7.2$~K) was cleaned by repeated sputter/anneal cycles until a clean, superconducting surface was obtained.
The Pb tip was prepared by indenting the tip into the Pb surface while applying a voltage of 100 V. To check the quality of the as prepared tips we record $dI/dV$ spectra on the bare Pb(111) surface at 4.8~K. At this temperature, thermal excitations of the quasi-particles across the superconducting energy gap lead to a finite number of hole-like states at $-\Delta$ and electron-like states at $+\Delta$. If the superconducting gaps of tip and sample are of same width, \textit{i.e.}, $\Delta_{tip} = \Delta_{sample}$, this finite state occupation results in a small conductance peak exactly at zero bias \cite{franke11}. If the gaps are of different width, \textit{i.e.}, $\Delta_{tip} \neq \Delta_{sample}$, we find peaks at $eV=\Delta_{tip}-\Delta_{sample}$ and $eV=-(\Delta_{tip}-\Delta_{sample})$. Throughout the experiment we only used tips that fulfilled $\Delta_{tip}=\Delta_{sample}=\Delta$.
The superconducting state of the tip leads to an increase in energy resolution beyond the intrinsic Fermi-Dirac broadening of a normal metal tip \cite{franke11}, because the substrate's density of states is sampled with the sharp quasi-particle peaks of the tip.
Fe-OEP-Cl was sublimated from a crucible at 490~K onto the clean Pb(111) surface held at 120~K. To enhance self-assembly into ordered domains, the sample was subsequently annealed to 240~K for 180~s, prior to cooling down and transferred into the STM.
Ordered monolayer islands of quasi-hexagonal structure can be identified, with the ethyl-groups clearly visible in the STM images (Fig. 1a). About 30$ \%$ of the molecules show a bright protrusion in their center. Annealing to higher temperatures after deposition reduces the number of protrusions.
We can thus identify the protrusion as the central Cl ligand which is present when the molecules are evaporated from the powder. We can exclude any impurities adsorbed from the background pressure by a series of different preparations. The fraction of chlorinated molecules does not depend
on the time for which the prepared sample is kept in the preparation chamber. The only observable influence is an elevated annealing temperature after molecule deposition. Reference~\cite{wende07} reports complete dechlorination upon adsorption on thin Ni and Co layers at 300~K, while Fe-OEP-Cl deposited on Au(111) at $240$~K, retains the Cl ligand almost completely~\cite{bheinrich13}. We focus our study on the molecules retaining their central chlorine ligand.
For the analysis of the excitation lifetime, $A_{r1}$ and $A_{r2}$ are quantified as the relative amplitude of the peaks appearing at the two inelastic onsets with respect to the amplitude of the BCS peak. They are defined as the ratio of amplitudes of the inelastic peak and the BCS peak in the $dI/dV$ spectra: $A_{r1} = \frac{A_{1}}{A_{BCS}}$ and $A_{r2} = \frac{A_{2}}{A_{BCS}}$, for the first and second excitation, respectively. Due to the peaked nature of the superconducting density of states, this corresponds to the commonly employed measure of the increase of the differential conductance at the threshold of the excitation.
$A_{1}$ ($A_{2}$), and $A_{BCS}$ are determined as the mean amplitudes of positive and negative bias side of the first (second) excitation and of the BCS peak. For details see the Supplementary Information.
|
\section{INTRODUCTION}
It has been suggested recently that the linear $O(N)$ scalar models
may be conformal and unitary for sufficiently large $N$ in dimensions $4<d<6$ \cite{fgk}.
The argument is based on a model with a potential that
is unbounded from below and the conjectured fixed point has a field
anomalous dimension of order $1/N$.
The presence of a non trivial UV fixed point would imply an unexpected
asymptotically safe behavior of such theories.
Here we consider the $O(N)$ models from the point of view of the functional renormalization group,
which has proved very successful in analyzing critical phenomena in dimensions $d<4$.
We find by means of analytic and numerical methods
that within certain approximations described below, for large $N$,
scaling solutions (fixed points of the action functional)
are either unbounded from below or singular and not globally defined.
The tool we use is the 1-PI generating functional, or Effective Average Action (EAA) $\Gamma_k$, which is defined as
the ordinary effective action but with an additional term in the action (infrared cutoff)
suppressing the contribution of modes with momentum lower than $k$.
It is convenient to choose a quadratic cutoff operator, whose kernel is usually denoted $R_k$,
so that the exact flow equation for the EAA in the RG time $t=\log(k/k_0)$ is given by
$\partial_t \Gamma_k = \frac{1}{2} {\rm Tr}[( \Gamma_k^{(2)}+R_k)^{-1} \partial_t R_k]$
\cite{wett1,morris,ber}.
This equation can be taken as an alternative definition of a QFT:
given that $R_k\to0$ for $k\to0$, the solution of the equation
starting from a given bare action at a UV scale,
gives the usual quantum effective action at $k=0$.
For applications to critical phenomena one expects an expansion in derivatives
to be a good approximation.
The most general form of the (Euclidean) EAA containing up to two derivatives is
\begin{equation}
\int d^dx\left[\frac{Z_k(\rho)}{2}\partial_\mu\phi^a\partial^\mu\phi^a
+\frac{Y_k(\rho)}{4}\partial_\mu\rho\partial^\mu\rho
+V_k(\rho)\right]\ ,
\end{equation}
where $Z_k$, $Y_k$ and $V_k$ are functions of $\rho=\phi^a\phi^a/2$ depending
on the external cutoff scale $k$.
The zeroth order of the derivative expansion,
called the Local Potential Approximation (LPA) consists in setting $Z_k=1$
(and the anomalous dimension $\eta=0$), $Y_k=0$
and studying only the running of $V_k$.
The first order consists in keeping the running of all three functions,
treating the anomalous dimension $\eta$ as a free parameter.
In $d=3$ this has been studied, with a power-law cutoff, in \cite{morris3}.
From these earlier studies we know that the
contribution of $Y_k$ to the running of $V_k$ and $Z_k$ is of order $1/N$,
so that it can be neglected in the large $N$ limit.
Moreover, in the same limit, the only known solutions of the fixed point equations for
$V_k$ and $Z_k$ have a field--independent $Z_k$ with anomalous dimension $\eta=0$.
These clearly should correspond to the large-$N$ limit of an $N$-dependent family of solution with $\eta$ suppressed for large $N$.
No other solutions, with a non vanishing $\eta$ at $N=\infty$, are known up to now.
The next order, where one keeps terms with four derivatives,
has been studied only in the case $N=1$ \cite{canet,LZ}.
In most other works the functions $Z_k$ and $Y_k$ are expanded around a constant
field configuration $\rho_0$ , typically at the minimum of the potential, and only the first term is retained.
The wave function renormalization constants of the Goldstone and radial modes
are defined by $Z_k=Z_k(\rho_0)$ and $\tilde Z_k=Z_k(\rho_0)+\rho_0 Y_k(\rho_0)$
and the anomalous dimensions are defined as
$\eta=-\partial_t Z_k/Z_k$ and $\tilde\eta=-\partial_t\tilde Z_k/\tilde Z_k$ \cite{wett2,tw,tl,btw}.
This setup, which is sometimes called LPA', makes contact with
the perturbative formulae for the anomalous dimension and extends them
somewhat in a non-perturbative way.
In this work we confine ourselves mostly to the LPA.
\section{CALCULATIONS}
In order to have a beta functional for $V_k$ in closed form we shall
consider the piecewise linear cutoff discussed in \cite{optimized},
which allows the loop integrals to be performed exactly.
Defining the dimensionless field $\tilde\phi^a=k^{(1-d/2-\eta/2)}\phi^a$, the dimensionless
potential $\tilde V(\tilde\phi)=k^{-d}V$ and rescaling
$\rho=N c_d\tilde\phi^2/2$, $u=N c_d\tilde V$, with $c_d=(1-\eta/(d+2))(4\pi)^{-d/2}/\Gamma(d/2+1)$,
the flow equation for the potential assumes the simple form
\begin{equation}
\label{udot}
\dot u=-d u(\rho)+(d-2+\eta)\rho u'(\rho)+\frac{1}{1+u'(\rho)}+A(N)\ ,
\end{equation}
where
\begin{equation}
\label{udotA}
A(N)=\frac{1}{N}\left(\frac{1}{1+u'(\rho)+2\rho u''(\rho)}-\frac{1}{1+u'(\rho)}\right)\ .
\end{equation}
The dot and prime denote partial derivatives with respect to $t$ and $\rho$ respectively.
In order to first establish a link with the work of \cite{fgk} we can make
a quartic ansatz for the potential $u=\lambda_2\rho+\lambda_4\rho^2$
and insert it in (\ref{udot}) to obtain beta functions for $\lambda_2$ and $\lambda_4$.
These turn out to have a fixed point at
\begin{eqnarray}
\small
\lambda_{2*}&=&\frac{-(d-4)(N+2)}{(d-8)N+2d-40}\ ;
\nonumber\\
\lambda_{4*}&=&\frac{16(d-4)N(N+8)^2}{((d-8)N+2d-40)^3}\ .
\label{polysol}
\end{eqnarray}
Taking into account the rescalings, $\lambda_{4*}$ agrees with Eq.(2.7) of \cite{fgk}
to first order in $\epsilon=d-4$.
An approximate expression for the anomalous dimension is obtained from the flow of $\Gamma^{(2)}$:
one has $\eta=c_d\frac{4\bar\rho u''(\bar\rho)^2}{(1+2\bar\rho u''(\bar\rho))^2}$
where $\bar\rho=-\lambda_2/(2\lambda_4)$ is the stationary point of the potential.
Using the fixed point couplings,
\begin{equation}
\eta=\frac{128(d-4)^2(N+2)(N+8)^2}{((d-8)N+2d-40)^2((3d-16)N+6d-56)^2}
\end{equation}
This is twice the anomalous dimension of the fields $\phi^a$,
and agrees with Eq. (2.4) of \cite{fgk} to first order in $\epsilon$.
One notices that $\eta$ is of order $1/N$.
From this and partial results of the second order derivative expansion quoted above,
we shall assume that the anomalous dimension is negligible in the large $N$ limit.
We discuss this assumption further in the concluding remarks.
Our strategy is to solve Eq.~(\ref{udot}) without expanding the potential.
The fixed point equation $\dot u=0$ is a second order differential equation
with a fixed singularity in $\rho=0$, so that analytic solution are parametrized by
a single integration constant $\sigma=u'(0)$.
Around the origin analytic solutions can be written as
$
u(\rho)=\frac{1}{d (1+\sigma)}+ \sigma \rho
+O\left(\rho^2\right).
$
Such an expansion is valid only within a certain radius of convergence.
One can integrate the equation numerically with this
boundary condition at a very small $\rho$.
For any fixed $\sigma$, integration fails at some maximum value $\rho_{max}(\sigma)$.
In $d=3$ this function has a sharp spike at a negative value of $\sigma$ that corresponds to
the Wilson-Fisher fixed point, and althought it is in practice not possible
to continue this solution numerically to infinity, it can be matched smoothly
to a solution satisfying the right asymptotic behavior at large $\rho$~\cite{morris3}.
\begin{figure}[htp]
\centering
\resizebox{0.8\columnwidth}{!}{
\includegraphics{peaks_5_d.eps}}
\caption{Top to bottom: the function $\rho_{max}(\sigma)$ for $N=2000$, $4000$, $8000$, $16000$.}
\end{figure}
One can repeat this analysis for any $d$ and in particular in $d=5$.
For $N\approx10^3$ the function $\rho_{max}(\sigma)$ has a smooth bump at $\rho\approx0.12$.
The bump becomes more pronounced as $N$ increases, and it moves to slightly
larger $\rho$. This behavior is shown in Fig.~1.
One could suspect that the peak corresponds to a scaling solution as $N\to\infty$.
Unfortunately, when $N\approx 10^5$ it becomes increasingly difficult to study numerically
what happens near the peak.
However the case $N=\infty$ can be treated analytically.
It is convenient to derive the fixed point equation once and to define $u'(\rho)=w(\rho)$ to obtain
\begin{equation}
\label{eqlargeN}
-2 w(\rho)+(d-2)\rho w'(\rho)-\frac{w'(\rho)}{(1+w(\rho))^2}=0\,.
\end{equation}
From this equation one immediately sees that apart from the trivial solution ($w=w'=0$)
the potential can have a stationary point only at $\rho_0=1/(d-2)$.
Deriving repeatedly Eq.(\ref{eqlargeN}) one determines the Taylor coefficients
of the expansion of the solution around this point:
\begin{equation}
w(\rho)=-\frac{d\!-\!4}{2} (\rho-\rho_0)+\frac{3}{8}\frac{(d\!-\!4)^3}{d\!-\!6}(\rho-\rho_0)^2+O\left((\rho\!-\!\rho_0)^3\right)
\nonumber
\end{equation}
The general solution is given by the implicit relation
\begin{equation}
\!\rho=C' w^{\frac{d}{2}-1}
\!+\frac{1}{(d\!+\!2)(1\!+\!w)^2}\!\!\phantom{a}_2F_1\left(\!\!1,2,2\!+\!\frac{d}{2},\frac{1}{1\!+\!w}\!\right).
\label{sol1}
\end{equation}
For $d\not=2n$ the solution of Eq.~\eqref{eqlargeN} is most conveniently written in the form
\cite{marchais}
($C'=C-\frac{d \pi}{4}/\sin(d \pi/2)$).
\begin{equation}
\rho(w)=C w^{\frac{d}{2}-1}
+\frac{1}{d-2}\!\!\phantom{a}_2F_1\left(2,1-\frac{d}{2},2-\frac{d}{2},-w\right)
\label{sol2}
\end{equation}
\begin{figure}[htp]
\centering
\resizebox{0.7\columnwidth}{!}{
\includegraphics{portrait.eps}}
\caption{The solutions of Eq.(\ref{eqlargeN}) with $C$ real ($w>0$, green), $C$ imaginary ($-1<w<0$, red), $C'$ imaginary ($w<-1$, blue),
and the special solutions $C=0$ ($w>-1$, thick black) and $C'=0$ ($w>0$, thick blue).}
\end{figure}
The only real solution extending continuously through $w=0$ is the one with $C=0$,
which has the Taylor expansion given above.
For $d=5$ the solution is the thick black curve in Fig.~2, enlarged in Fig.~3.
It intersects the $\rho=0$ axis at $w(0)\approx 0.1392$,
which corresponds to the accumulation point of the
peaks of Fig.~1, and at $w(0)\approx -0.5776$.
We see that $u''$ has a singularity at $\rho_{max}\approx 0.621$,
where the implicit solution cannot be inverted,
and the potential does not exist beyond this value.
The solutions corresponding to different values of the integration constant are also shown in Fig.~2.
The ones with real $C\not=0$ exist only for positive $w$ (green curves),
those for imaginary $C\not=0$ exist only in the range $-1<w<0$ (red curves)
and those of the form given in Eq.~\eqref{sol1} with $C'$ purely imaginary exist only for $w<-1$ (blue curves).
If ${\rm Im}(C')<0$ the latter form a continuum of global solutions covering the whole range
$-\infty<\rho<\infty$, therefore including the whole physical region.
However, these solutions have $u'<0$ everywhere so that the potentials $u$ are unbounded from below
and therefore are physically unacceptable.
The green and red curves also have
$w'(\rho_0)=-(d-4)/2$ for $d>4$,
but some higher derivative diverges at this point,
for example in $d=5$, $u'''(\rho_0)=w''(\rho_0)$ is singular.
In conclusion, we have shown analytically
that there is no acceptable scaling solution in the LPA at $N=\infty$.
\begin{figure}[htp]
\centering
\resizebox{0.7\columnwidth}{!}{
\includegraphics{mixed.eps}}
\caption{The exact solution for $N=\infty$ with $C=0$ (black curve) with superimposed (top to bottom):
the polynomial solution for N=2000 at order $\rho^2$ of Eq.(\ref{polysol}) (green, dashed) and $\rho^3$ (blue, dot-dashed);
numerical solutions for
$N=8000$ (with $\sigma=0.137$, blue curve)
$N=2000$ (with $\sigma=0.124$, red curve),
corresponding to the "peaks" in Fig.~1;
numerical solutions for $N=2000$ and $\sigma=-0.567,-0.577,-0.587$ (red, continuous).
}
\end{figure}
For large but finite $N$ we have studied
the solutions numerically integrating the flow equation
starting either from $\rho=0$ or from large $\rho$.
The solutions corresponding to the peaks in Fig.1 are very close to
the exact solution with $C=0$ at infinite $N$
up to some maximal value for $\rho<\rho_{max}$, see Fig.~3.
The numerical evolution then starts to deviate and stops where $u''$ diverges.
For $N\approx 8000$ the curves are nearly indistinguishable.
Similarly one can analyse the numerical evolution from initial conditions
near $w\approx -0.577$. Nothing special happens here:
the function $\rho_{max}(\sigma)$ shows no bumps or peaks for negative $-1<\sigma<0$.
The evolution for varying initial conditions produces curves that are close to the
large-$N$ solution, all terminating in a singularity of $u''$, see Fig.~3.
For large $\rho$ the solutions of Eq.(\ref{udot}) have an asymptotic expansion
$u =A \rho^\alpha+\frac{1}{d A(2\alpha-1)}\left(1-\frac{1}{N} \frac{2(\alpha-1)}{2\alpha-1}\right)\rho^{1-\alpha}+O\left(\rho^{2-2\alpha}\right)$
where $\alpha=d/(d\!-\!2\!+\!\eta)$ and $A$ is an arbitrary constant.
If a global solution existed,
one should be able to find it from a numerical integration with this asymptotic behavior.
As in the exact large-$N$ case,
we find it more convenient to analyze this problem in terms of $\rho(w)$ which satisfies
the fixed point differential equation
\begin{equation}
(d-2+\eta)\rho -(2-\eta) w \rho'-\frac{1\!-\!\frac{1}{N}}{(1\!+\!w)^2}
+\frac{1}{N}\frac{2\rho\,\rho''-3\rho'{}^2}{(2\rho\!+\!(1\!+\!w)\rho')^2}=0
\nonumber
\end{equation}
which determines the asymptotic behavior:
\begin{equation}
\label{asbe}
\rho=A w^\beta +\frac{1-\frac{2}{N(2+\beta)}}{d+2-\eta} w^{-2}+O\left( w^{-2} \right)
\end{equation}
where $\beta=(d-2+\eta)/(2-\eta)$.
In the large-$N$ limit this series correspond exactly to the one given by the expansion of expressions in Eqs.~\eqref{sol1} and \eqref{sol2}.
Solving numerically with $\eta=0$
we find that the function $\rho(w)$ never becomes zero.
Instead, it turns around and crosses $w=0$ without singularities near $w=1/(d-2)$,
then tends to $+\infty$ at $w=-1$.
Inverting this relation, the solutions for $w(\rho)$ become
singular before reaching $\rho=0$, so they do not correspond to globally defined scaling solutions
(see Fig.~4).
These numerical studies show that the conclusion of the $N=\infty$
case extends also to large but finite $N$.
\begin{figure}[htp]
\centering
\resizebox{0.7\columnwidth}{!}{
\includegraphics{asymptotic.eps}}
\caption{Two exact solution for $N=\infty$ ($w>0$, green curves) and the corresponding numerical solution
matching the asymptotic behavior Eq.(\ref{asbe}) for $N=8000$ (extending to negative $w$, blue curves).}
\end{figure}
We worked throughout with a specific choice of the cutoff kernel $R_k$
and one may wonder whether different cutoff shapes could lead to qualitatively different results.
We have checked that similar results are obtained with a sharp cutoff.
For an arbitrary cutoff the fraction in the third term of Eq.~(\ref{udot}) can be approximated by
$C_1-C_2u'(\rho)$ for small $u'(\rho)$ and by $C_3/u'(\rho)$ for large $u'(\rho)$,
where $C_1$, $C_2$, $C_3$ are cutoff--dependent constants of order one
(the terms in $A(N)$ also behave in a similar way).
One cannot write the solutions in closed form for general cutoff, but for large and small
$\rho$ they have the same general behavior as the solutions discussed above,
so we expect our results to hold for any cutoff shape.
We have also repeated the analysis in other dimensions $d>4$ with similar results.
Finally, as a preliminary step towards a more complete analysis,
we have studied the solutions of Eq.(\ref{udot}) for $N=1000$
with nonvanishing anomalous dimension $\eta$ ranging between $-0.05$ and $0.05$.
The dependence of the solutions on $\eta$ is smooth and one always finds
the same qualitative behavior as with $\eta=0$.
\smallskip
\section{DISCUSSION}
While the polynomial expansion of the potential is a useful tool,
it may sometimes be misleading.
Truncating the potential to a finite polynomial generally leads to spurious fixed points
and one has to carefully sift for the physically significant ones~\cite{morris-last}.
In the present case, Taylor expanding the potential
at second order in $\rho$ leads to (\ref{polysol}),
which corresponds to the dashed green curve in Fig.3.
Polynomials of higher order yield, among others, solutions that
become arbitrarily good approximations of
the exact solution $N=\infty$, $C=0$ (black curve in Fig.2 and 3).
These polynomial approximations exist for all $\rho$ but we have seen
that the real solution that they approximate does not.
If one keeps the whole Taylor series, and a recursive method is used to find
the solution, one has to check the radius of convergence of the series.
For example, if one considers a general scalar potential and
writes the beta functions for the coefficients of its Taylor expansion
around the origin, then given an arbitrary value to $\sigma=\lambda_2$,
the beta functions can be solved iteratively yielding a fixed point potential \cite{hh}.
The existence of such solutions in $d=4$ would go against the expectation that linear scalar
theory is trivial in dimension $d\geq4$.
Indeed it was shown by Morris \cite{morris1} that in $d=4$ these fixed point potentials
become singular at some finite value of the field.
On the other hand in $2<d<4$, only for a discrete set of values of $\sigma$
does the solution extend to infinity.
These global scaling solutions correspond to multicritical models
(see \cite{morris2,codello} for the $Z_2$-invariant case,
\cite{cood} for some results in the general $O(N)$ case).
The functional method we have used here avoids expanding the solution.
We have used the minimal truncation that yields the desired information,
keeping the running of the potential.
In the large $N$ limit we have shown analytically that no global solution
exists that is bounded from below. For the class of theories with $\eta=0$ at $N=\infty$ this result goes beyond the LPA.
For large but finite $N$ we have reached the same conclusion by studying
numerical solutions starting from $\rho=0$ or $\rho=+\infty$.
In both cases either $u''$ becomes singular at some finite $\rho$
or the potential has everywhere $u'<0$ and is therefore unbounded from below.
Note that this is also the case for the QFT model analysed in~\cite{fgk}.
Even though we concentrated on large $N$,
we have found no evidence of a qualitative change of behavior as $N$ decreases.
This is in agreement with the analysis of \cite{nakayama},
which was based on the conformal bootstrap.
We have focused primarily on the case $d=5$ but a least in the LPA there is evidence that the conclusions hold for all $d\ge4$.
For a complete analysis at first order of the derivative expansion
one would have to study the flow of the three functions $V_k$, $Z_k$, $Y_k$ as in \cite{morris3}.
The LPA' formulas indicate that $\eta$ is of order $1/N$,
making the existence of solutions with $\eta\not=0$ for $N=\infty$ seem unlikely.
In any case, if such a solution existed it would be in another
universality class from the one conjectured in \cite{fgk}.
For finite $N$ the system of equations for $V_k$, $Z_k$, $Y_k$ is completely coupled
and one cannot rule out the existence of solutions, but this we also consider to be unlikely,
since in lower dimensions all the qualitative features of the fixed point
can already be seen in the LPA.
A possible exception may occur if, in presence of non global but singular solutions, the function $Z_k(\rho)$ has a singularity at the
same location as the potential. Then it may be legitimate to redefine the quantum field
in such a way as to shift the singularity to infinity.
We leave this to future investigations.
{\it Acknowledgements}. We would like to thank A. Codello, S. Giombi, I. Klebanov and D. Litim
for discussions.
\goodbreak
\medskip
|
\section{Introduction}
Nowadays, there is a need for efficient data mining techniques. The human readability of the model is an important factor of a good data mining algorithm. Among various data mining methods very popular are decision trees \cite{quinlan}, \cite{marcin}. Although, they have an easy interpretable model, a single tree does not always obtain the highest accuracy. To overcome this problem, various ensemble methods were proposed. Among them, the popular is Random Forest (RF) proposed by Leo Breiman \cite{breiman}. The RF builds a set of trees using bagging and random subspace methods. The final output is a mode of responses from all individual trees. The RF can be used for classification and regression tasks. Despite the high accuracy of the RF, the human readability of the model is lost. There exist some methods to look inside RF black-box, like: examining variable importance \cite{var_sel}, parallel cooridinate plots by variable \cite{breiman} or visualizing the RF proximity distance matrix with Multidimensional Scaling \cite{liaw}. Herein, we propose a novel method for visualizing the RF proximity matrix based on Self-Organising Maps (SOM) \cite{kohonen}. The SOM is an artificial neural network model that maps high-dimensional input data space onto usually two-dimensional lattice of neurons in an unsupervised way. Although, the SOM is an originally unsupervised algorithm there exist supervised extensions \cite{lasso}, \cite{mh_som}, \cite{sto}, \cite{fuzzy_SOM}. The SOM has been proved as an efficient data mining tool in many real life applications \cite{dominik1}, \cite{dominik2}, \cite{julek}, \cite{julek2}. In this paper, we focus on using the labeled SOM model for mapping the RF used in classification tasks. The RF proximity matrix will be used for the SOM learning. It was shown that using more sophisticated distance metric than Euclidean can improve the accuracy of the SOM \cite{lasso_dml}. The RF proximity matrix was used earlier for improving clustering accuracy. Horvath et al. \cite{horvath} presented method for building clusters from the RF learned with unlabeled data and successfully used it for tumor detection \cite{horvath2}. Moosmann et al. \cite{moosmann} used the RF for efficient segmentation of images, where leaves were assigned to distinct image regions rather than to specific class. Gray et al. \cite{gray} used the RF proximity matrix and the MDS for classification of medical images of different types of dementia.
The paper is organized as follow: firstly, we describe the SOM and the RF algorithms; secondly, proposed approach for learning the SOM with the RF proximity matrix is presented; then, the MDS vs the SOM visualization and the accuracy of the SOM learned with Euclidean metric and the RF proximity matrix are compared.
\section{Methods}
Let's denote data set as $D = \{ ( \vec{x_i}, c_i)\}$, where $\vec{x_i}$ is an attribute vector, $\vec{x} \in \mathcal{R}^M$, $M$ is attribute vector length and $c_i$ is a discrete class number of \textit{i}-th sample, $i=[1,2, ... , N]$ and $c=[1,2, ..., C]$.
\subsection{Self Organising Maps} \label{som}
In this paper, we used the SOM as a two-dimensional grid of neurons. Each neuron is represented by a weights vector $W_{pq}$, where $(p,q)$ are indices of the neuron in the grid. It is important to notice, that neuron's weights vector has the same length as sample's attribute vector in the data set. The neuron's weights directly corresponds to attributes in the data set. In the training phase, for each sample we search for a neuron which is the closest to the \textit{i}-th sample. In the original SOM algorithm the distance is computed with squared Euclidean distance by following equation:
\begin{equation}\label{distance}
Dist_{train}(D_i, W_{pq}) = (\vec{x_i} - W_{pq})^T(\vec{x_i} - W_{pq}).
\end{equation}
The neuron $(p,q)$ with the smallest distance to \textit{i}-th sample is so-called the Best Matching Unit (BMU), and we note its indicies as $(r, v)$. Once the BMU is found, the weights update step is executed. The weights of each neuron are updated with formula:
\begin{equation}\label{learning}
W_{pq}(t+1) = W_{pq}(t) + \eta h_{pq}^{(i)} (\vec{x_i} - W_{pq}(t)),
\end{equation}
where $t$ is an iteration number and $\eta$ is a learning coefficient and $h_{pq}^{(i)}$ is a neighbourhood function. We assume that one iteration is a presentation of one training sample, whereas a presentation of all training samples is one learning epoch. The learning coefficient $\eta$ is decreased between consecutive epochs to improve network's ability to remember patterns. It is described by:
\begin{equation}
\eta = \eta_{0} exp(- e \lambda_{\eta}),
\end{equation}
where $\eta_{0}$ is the initial step size, $e$ is the current epoch number and $\lambda_{\eta}$ is responsible for regulating the speed of the decrease. The neighbourhood function controls changing of weights with respect to the distance to the BMU in the grid. It is noted as:
\begin{equation}
h_{pq}^{(i)} = exp(-\alpha((r - p)^2 + (v - q)^2)),
\end{equation}
where $\alpha$ describes the neighbourhood function width. This parameter is increasing during learning $\alpha=\alpha_0 exp(-(e_{stop}-e)\lambda_{\alpha})$ - it assures that neighbourhood becomes narrower during the training. The network is trained till chosen number of learning procedure epochs $e_{stop}$ is exceeded.
In the described algorithm the class label information is not used. The simplest approach of using the SOM as a classifier is to label the neurons after the unsupervised training. For each neuron we remember the overal sum of neighbourhood values $h_{pq}^{(i)}$ from each class over all samples. The label of major class is assigned to the neuron. In the testing phase, the input sample's class is designated based on the class of the found BMU.
\subsection{Random Forest}
In the RF algorithm a set of single trees is built. The process of constructing one tree can be described in the following steps:
\begin{enumerate}
\item Draw a bootstrap data set $D'$ by choosing $n$ times with replacement from all $N$ training samples.
\item Determine a decision at node using only $m$ attributes, where $m$ is smaller than $M$. The split is selected based on maximal Information Gain. \label{dec}
\item Move the data through the node with respect to decision from the step \ref{dec}. \label{mov}
\item Repeat steps \ref{dec}, \ref{mov} till full tree is grown.
\end{enumerate}
At the end of tree constructing, the class label is assigned to each leaf based on a class of samples in it. In the testing phase, a new sample is pushed down through all trees. From each tree a class label is remembered based on class of the reached leaf. The final response is the mode of votes from all trees. The proximity matrix $Prox$, with size $N$x$N$, can be easily obtained by putting all samples down the all trees. If two samples $i$ and $j$ are in the same terminal node in the tree, their proximity is increased by one $Prox(i,j) = Prox(i,j)+1$. After the presentation of all samples the proximities are divided by the number of trees in the RF. The greater proximity value is, the more similar samples are. The dissimilarity measure can be formulated as $Dis(i,j) = 1 - Prox(i,j)$.
\subsection{Self Organising Maps learned with Random Forest (RF-SOM)}
In the proposed approach we assume that the RF is already learned. The learning of the network in one epoch can be summarized in the following steps:
\begin{enumerate}
\item Build a data set $H$ as a union of all network's weights $W$ and attribute vector $\vec{x}_j$ of $j$-th sample, $H = W \cup \vec{x}_j$. The matrix $W$ size is $L$x$M$, where the $L$ is a total number of neurons in the network. In this matrix each row contains weights from one neuron, the mapping from neurons's 2D grid to matrix $W$ is assumed. \label{fst}
\item For set $H$ compute dissimilarity matrix $Dis_H$ using the RF. The $Dis_H$ size is $(L+1)$x$(L+1)$.
\item Find the smallest distance to the neurons in dissimilarity matrix $Dis_H$, in distances corresponding to $j$-th sample:
\begin{equation}
v = \arg\min_h Dis_H (j,h), h \neq j,
\end{equation}
where $v$ is an index of BMU in the matrix $W$, which can be mapped into $r,v$ indices in the network 2D grid. \label{bmu}
\item Update the network weigths with formula \ref{learning}. \label{lst}
\item Repeat steps \ref{fst}-\ref{lst} for all samples in the training set.
\end{enumerate}
After the end of the SOM learning, it is labeled as described in Section \ref{som}. In the testing phase, for input sample the BMU search is performed by taking the steps \ref{fst}-\ref{bmu}. The output class label is the same as the BMU class. We will denote the proposed method as RF-SOM. The computational complexity of using the Euclidean distance in the SOM is $\mathcal{O}(N * L)$, because we need to compute for each sample, from $N$ samples, a distance between sample and $L$ neurons. Whereas, the using of the RF proximity matrix in the proposed RF-SOM has complexity $\mathcal{O}(N*L*T*log_2(tree_{size}))$\footnote{We omit cost of constructing the RF in the complexity assessment.}, where $tree_{size}$ is a number of nodes in the tree. In the RF-SOM for each sample we propagate $L+1$ input vectors through $T$ trees in the RF, and passage through a tree has complexity $\mathcal{O}(log_2(tree_{size}))$. The complexity of distance computation using the RF proximity matrix is worse than using Euclidean distance. Although, it is beneficial when compared to the memory complexity of the MDS used for RF proximity visualization, which uses $\mathcal{O}(N^2)$ memory. The RF-SOM requires only $\mathcal{O}(L^2)$. This discards the MDS as a method of the RF proximity matrix visualization for large data sets.
\section{Results}
We used 6 real data sets to examine properties of the RF-SOM. There were used data sets: 'Glass', 'Wine', 'Iris', 'Sonar', 'Ionosphere', 'Pima' from the 'UCI Machine Learning Repository' \cite{uci}. In the Table \ref{sets} are presented data sets properties. In all experiments we used following parameters values for each SOM type: $e_{stop} = 200$, $\eta_0 = 0.1$, $\lambda_{\eta}=0.0345$, $\alpha_0=0.1$, $\lambda_{\alpha}=0.008$. We will denote a network learned with Eculidean distance as SOM. The network sizes for the SOM and the RF-SOM for each data set are equal, they are presented in the Table \ref{sets}. The network sizes were chosen arbitrarily because selecting optimal network size is not in the scope of this paper. In all cases the SOM and the RF-SOM starts learning from the same initial weights values. The RF was constructed with 100 trees and $m=\sqrt{M}$ for all data sets.
\begin{table}
\begin{center}
\begin{tabular}{c| >{\centering}m{1.7cm} | >{\centering}m{1.7cm} | >{\centering}m{1.7cm} | c |}
\cline{2-5}
& Samples & Attributes & Classes & Network size \\ \hline
\multicolumn{1}{|c|}{Glass} & 214 & 9 & 6 & 7x7 \\ \hline
\multicolumn{1}{|c|}{Wine} & 178 & 13 & 3 & 4x4 \\ \hline
\multicolumn{1}{|c|}{Iris} & 150 & 4 & 3 & 5x5 \\ \hline
\multicolumn{1}{|c|}{Sonar} & 208 & 60 & 2 & 8x8 \\ \hline
\multicolumn{1}{|c|}{Ionosphere} & 351 & 34 & 2 & 8x8 \\ \hline
\multicolumn{1}{|c|}{Pima} & 768 & 8 & 2 & 7x7 \\ \hline
\end{tabular}
\end{center}
\caption{The description of data sets used in experiments and network size used for each data set.}\label{sets}
\end{table}
To present visulization properties of the proposed method we used 'Pima' data set. In the Fig.\ref{mds_som} there are presented: the SOM (Fig.\ref{som_pi}) and the MDS (Fig.\ref{mds_pi}) both learned with Euclidean distance; and RF-SOM (Fig.\ref{rfsom_pi}) and RF-MDS (Fig.\ref{rfmds_pi}) constructed using the RF proximity matrix. The SOM networks were presented as a 2D grid of neurons, where for each neuron, its weigths are presented by a polar area diagram (sometimes called coxcomb plot). In the MDS and the RF-MDS plots information about points distribution in reduced 2D space is available. However, information about how point's position is affected by combination of attributes values is missing. What is more, there is hard to find crisp border to distinguish two classes on the MDS neither on the RF-MDS. In contrary to MDS technique, the SOM and RF-SOM plots do not provide information about explicit point distribution but rather a mapping of attributes combination onto 2D grid of neurons. After network labeling the class labels are assigned to neurons, therefore distinguishing specific combination of attributes in each class is possible. As expected, although, the SOM and the RF-SOM started learning from the same initial weights values they have different final distribution of attributes combinations across the network.
\begin{figure}
\captionsetup[subfigure]{aboveskip=-13pt,belowskip=-3pt}
\centering
\begin{subfigure}[b]{0.49\textwidth}
\caption{SOM} \label{som_pi}
\includegraphics[trim=1.5cm 0cm 1.5cm 1.1cm, clip=true, width=\textwidth]{som_pima_2.pdf}
\end{subfigure}
\begin{subfigure}[b]{0.49\textwidth}
\caption{RF-SOM} \label{rfsom_pi}
\includegraphics[trim=1.5cm 0cm 1.5cm 1.1cm, clip=true, width=\textwidth]{rfsom_pima_2.pdf}
\end{subfigure}
\begin{subfigure}[b]{0.47\textwidth}
\caption{MDS} \label{mds_pi}
\includegraphics[clip=true, width=\textwidth]{pima_mds_euc.pdf}
\end{subfigure}
\begin{subfigure}[b]{0.47\textwidth}
\caption{RF-MDS} \label{rfmds_pi}
\includegraphics[clip=true, width=\textwidth]{pima_mds.pdf}
\end{subfigure}
\caption{The visualizations of 'Pima' data set with (a) the SOM, (b) the RF-SOM, (c) the MDS with Euclidean distance, (d) the RF-MDS. The class information is coded in red color for the first class and blue color for the second class in all plots. In the (a) and (b) for polar area diagram the attribute's number corresponding to the neuron's weight is shown in the legend.} \label{mds_som}
\end{figure}
To compare the accuracy of the SOM learned with Euclidean distance and RF-SOM learned with RF proximity matrix we use information about samples class label. We measure the accuracy of classification. The results of comparison are presented in the Table \ref{results}. All of the results are mean over 10-fold cross validation. In the Table \ref{results} we also include the accuracy of the alone RF as the reference. The RF-SOM on 4 data sets ('Glass', 'Sonar', 'Ionosphere', 'Pima') obtained better results than the SOM. The greatest improvement over the SOM was achived on 'Sonar' set, it was $12.57\%$. The same performance of the SOM and the RF-SOM was on 'Wine' and 'Iris', when on the latter the RF-SOM has higher standard deviation value. It is worth noting, that the improvement of the RF-SOM over the SOM depends on the accuracy of the RF. On data sets (like 'Sonar' or 'Glass') where the accuracy difference between the RF and the SOM is high, the RF-SOM noted large improvement over the SOM. Whereas, on data sets ('Wine' and 'Iris') with small accuracy difference between the RF and the SOM, the RF-SOM obtained the same mean accuracy as the SOM.
\begin{table}
\begin{center}
\begin{tabular}{c|c|c|c|c|c|c|c|}
\cline{2-7} & Glass & Wine & Iris & Sonar & Ionosphere & Pima \\ \cline{1-7}
\multicolumn{1}{|c|}{RF} & 77.96$\pm$7.82 & 98.85$\pm$2.29 & 95.33$\pm$6.70 & 85.05$\pm$4.74 & 93.44$\pm$2.88 & 76.14$\pm$7.09 \\ \hline
\multicolumn{1}{|c|}{SOM} & 61.73$\pm$6.18 & 96.60$\pm$3.73 & 94.67$\pm$4.00 & 67.69$\pm$8.97 & 84.33$\pm$6.43 & 71.73$\pm$5.16
\\ \hline
\multicolumn{1}{|c|}{RF-SOM} & 67.27$\pm$6.04 & 96.60$\pm$3.73 & 94.67$\pm$5.81 & 80.26$\pm$5.17 & 89.72$\pm$6.42 & 74.71$\pm$6.80 \\ \hline
\end{tabular}
\end{center}
\caption{The classification accuracy for RF, SOM and RF-SOM. The results are mean and std. over 10-fold cross validation.}\label{results}
\end{table}
We measure classfication accuracy for different number of trees in the RF to examine the influence of a number of trees in the RF to performance of the RF-SOM. The accuracy for the RF and the RF-SOM for a number of trees in the RF, $T=\{10, 20, 50, 100, 200, 500\}$, is presented in Fig.\ref{diff_rf_som}. It can be observed that for all data sets except 'Sonar' and 'Glass' the accuracy of the RF does not depend on $T$. The good results of classification are obtained even for 10 trees in the RF. For 'Sonar' and 'Glass' sets the accuracy of the RF increases with increasing a number of trees. The very similar behaviour can be observed for the RF-SOM. The accuracy for 'Wine', 'Iris', 'Ionosphere' and 'Pima' slightly varies with $T$ growing. However, for 'Sonar' and 'Glass' sets the RF-SOM obtains better results if more trees are used in the RF. The observed behaviour can be explained by more complex data relationships in 'Sonar' and 'Glass' sets which are better modeled with greater number of trees in the RF.
\begin{figure}
\centering
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[clip=true, width=\textwidth]{diff_rf_rf_leg.pdf}
\caption{RF} \label{rf_rf}
\end{subfigure}
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[clip=true, width=\textwidth]{diff_rf_som.pdf}
\caption{RF-SOM} \label{rf_som}
\end{subfigure}
\caption{The classification accuracy for (a) RF and (b) RF-SOM, for a different number of trees in the RF. The result are mean over 10-fold cross validation.} \label{diff_rf_som}
\end{figure}
\section{Conclusions}
The novel method for visualizing the RF proximity matrix by the SOM was proposed. The RF-SOM method uses the RF to compute distances between input sample and neurons. The proposed method of visualization provide better understanding of relationship between data in the RF structure than the MDS. The RF-SOM contrary to the MDS provides a mapping of data onto 2D neurons grid. In case of new coming samples there is no need to recompute the whole RF proximity matrix like in the MDS method. Additionally, the proposed method has lower memory complexity than the MDS, which for large data sets is not applicable. What is more, the experimental results show that the RF-SOM learned with the RF dissimilarity gained better or the same accuracy than the SOM learned with Euclidean distance. As pointed in \cite{horvath} the RF dissimilarity has attractive features: it can handle mixed variable types well, is invariant to monotonic transformations of the input variables and is robust to outliers. The attractiveness of RF dissimilarity and obtained results with RF-SOM encourage to focus our future work on using the RF proximity matrix in other clustering algorithms.
\section{Acknowledgements}
PP has been supported by the European Union in the framework of European Social Fund through the Warsaw University of Technology Development Programme.
|
\section{Introduction}
It is well-known that many different notions of integral were introduced in the last century, for real-valued functions, in order to generalize the Riemann one. An exhaustive discussion of the various definitions for real-valued functions can be found in \cite{KS}, where the Henstock-Kurzweil,
the Lebesgue, and
the McShane notions have been finely compared, and the equivalence between
the McShane and
the Lebesgue integrals
is
clearly described.
The situation changes deeply in the case of Banach space-valued functions: in this case, it is well-known that the stronger type of integral is the Bochner one, which implies
the Birkhoff integrability, which in turn is stronger than
the McShane and the latter is stronger than both Henstock and Pettis integrals. The wide literature in this topic witnesses the great interest for these problems: see for example
\cite{bbs,BS2004,BS2011,BS2004-nitra,bcs2011,bvg,cao,f1995,DPMILL,DPVM,dpp,dr,f1994a,f1994b,riecan,r2009a,r2009b,r2005,SS}.
Alternative notions of integrals have also been given, for various applications (see e.g. \cite{ims1998a,ims1998b,is1998,LW2}).
Subsequently,
the notions of order-type integrals have also been introduced and studied, for functions taking their values in ordered vector spaces, and in Banach lattices in particular: see
\cite{bvl,fvol3,mn,bd2014,bdp2012,bms,csmed,mimmoroma,dallas}.
Also the multivalued case has been intensively studied, for both types of convergences: see for example \cite{cascales2007b,bcs2014}.
In this research mainly the differences between norm- and order-type McShane integrals
have been
discussed, for single-valued functions taking values in a Banach lattice with order-continuous norm.
However this note is just an anticipation of a forthcoming extended paper, with further results and more detailed proofs.
After a section of preliminaries, in the third one norm and order-type integrals are compared, showing
a first striking difference: order integrals in general do not respect almost everywhere equality, except for order-bounded functions. Another interesting difference is that order integrals enjoy the so-called Henstock Lemma: this fact has interesting consequences in $L$-spaces,
where McShane order integrability is stronger than the Bochner (norm) one.
In the fourth section integrability in $[0,1]$ is discussed and it is proven that monotone mappings are McShane order-integrable, by using a similar procedure as in \cite{KS}.
\section{Preliminaries}
From now on, $T$ will denote a compact metric space, and $\mu:\mathcal{B}\to \erre^+_0$ any regular, nonatomic $\sigma$-additive measure on the $\sigma$-algebra $\mathcal{B}$ of Borel subsets of $T$.
A \textit{gage} is any map $\gamma: T \rightarrow \mathbb{R}^+$.
A \textit{partition} $\Pi$ of $[0,1]$ is a finite family $\Pi = \{ (E_i,t_i): i=1, \ldots,k \} $ of pairs such that the
$E_i$ are pairwise disjoint sets whose union is $T$ and the points $t_i$ are called {\em tags}. If all tags satisfy the condition $t_i \in E_i$ then the partition is said to be of {\em Henstock} type, or a {\em Henstock partition}. Otherwise, it is said to be a {\em free} or {\em McShane} partition.
Given a gage $\gamma$, a partition $\Pi$ is \textit{$\gamma$-fine} $(\Pi \prec \gamma)$ if
$d(w,t_i) < \gamma (t_i)$ for every $ w \in E_i$ and $i = 1, \ldots, k$.\\
Clearly, a gage $\gamma$ can also be defined as a mapping associating with each point $t\in T$ an open ball centered at $t$: sometimes this concept will be used, without risk of confusion.\\
Let
$X$ be any Banach lattice with an order-continuous norm.
For the sake of completeness the main notions of
gauge integral are recalled here. \\
\begin{definition}\rm \label{fnorm}
\rm A function $f:T\rightarrow X$ is \textit{norm-integrable}
if there exists $J \in X$ such that, for every $\varepsilon > 0$
there is a gage $\gamma : T \rightarrow \mathbb{R}^+$ such that for every $\gamma$-fine Henstock partition of $T$,
$\Pi=\{(E_i, t_i), i=1, \ldots, q \}$, it is:
\begin{eqnarray*}
\left\|\sigma(f, \Pi)- J \right\| \leq \varepsilon.
\end{eqnarray*}
(Here, as usual, the symbol $\sigma( f, \Pi)$ means
$\sum_{i=1}^q f(t_i) \mu(E_i)).$
In the case of integrability in the Henstock sense, this will be denoted with \textit{{\rm H}-integrability} and the integral $J$ will be denoted with $H\int f d\mu$.\\
\end{definition}
\begin{remark}\label{mettipunt}\rm
It is worth noticing here that, in the previous definition, taking more generally {\em free} $\gamma$-fine partitions does not modify the concept introduced: in other words, the same integral is obtained if all free $\gamma$-fine partitions are allowed in the previous definition: see for example \cite[Proposition 2.3]{bcs2014}).
\end{remark}
Parallel to this definition, notions of {\em order-type} integral can be given, in accordance with the following
\begin{definition}\rm \label{forder}
\rm A function $f:T\rightarrow X$ is \textit{order-integrable}
in the Henstock sense if there exist $J \in X$, an $(o)$-sequence $(b_n)_n$ in $X$ and a corresponding sequence $(\gamma_n)_n$ of gages, such that
for every $n$ and every $\gamma_n$-fine Henstock partition of $T$,
$\Pi=\{(E_i, t_i), i=1, \ldots, q \}$, one has
\begin{eqnarray*}
\left|\sigma(f, \Pi)- J \right| \leq b_n.
\end{eqnarray*}
In the case of integrability in the Henstock sense, this will be denoted with \textit{{\rm (oH)}-integrability} and the integral $J$ will be denoted with $(oH)\int f$.\\
\end{definition}
It is obvious that also in this case there is no difference in taking all {\em free} $\gamma_n$-fine partitions in the previous definition.
Thanks to the order continuity of the norm in $X$, it is easy to see that any (oH)-integrable map $f$ is also H-integrable and the integrals coincide.
Furthermore, it can be observed that (oH)-integrability of a function $f:T\to X$ implies also Pettis integrability, thanks to well-known results concerning the McShane norm-integral: see \cite[Theorem 8]{f1994a}.
However, later it will be shown that the (H)- and the (oH)-integrability are not equivalent, in general.
\section{Comparison between Norm and Order integral}
There are deep differences between order-type and norm-type integrals.
A first remarkable fact is that, in general, almost equal functions can behave in different ways with respect to the (oH)-integral, as it was proven in \cite[Example 2.8]{bms}, where a function $f:[0,1]\to c_{00}$ is
given
with the following properties: $f$ is almost everywhere null (with respect to the Lebesgue measure) and $f$ is not (oH)-integrable. So, this function is almost everywhere equal to 0, hence is Bochner-integrable, but not (oH)-integrable
or order-bounded.
Indeed, for order-bounded functions, the situation is better, as shown in the next Proposition.
\begin{proposition}
Let $f,g:T\to X$ be two bounded maps, such that $f=g$ $\mu$-almost everywhere. Then, $f$ is {\rm (oH)}-integrable if and only if $g$ is, and the integral is the same.
\end{proposition}
{\bf Proof:}\ Let $M$ be any majorant for $|f|$ and $|g|$, and assume that $f$ is
(oH)-integrable, with integral $J$. Let $(b_n)_n$ and $(\gamma_n)_n$be the sequences related to (oH)-integrability of $f$. In order to show integrability of $g$, fix $n$, and pick any open set $A_n\subset T$, with $\mu(A_n)< n^{-1}$ and
$A_n \supset N:=\{t\in T:f(t)\neq g(t)\}$.
Now, for each element $u\in N$ let $\delta_n(u)$ be any open set containing $u$ and contained in $A_n$: then define $\gamma'_n(t)=\gamma_n(t)$ when $t\notin N$, while $\gamma_n'(u)=\gamma_n(u)\cap \delta_n(u)$ when $u\in N$.\\
Now, fix any tagged $\gamma'_n$-fine partition $\Pi$, ($\Pi:=(E_i,\tau_i)_i)$ and observe that, whenever
the tag $\tau_i$ belongs to $N$, then $E_i\subset A_n$.
So, it follows easily that
\begin{eqnarray*}
\sup \left\{
\sum_{\tau_i\in N} |f(\tau_i)|\mu(E_i),
\sum_{\tau_i\in N} |g(\tau_i)|\mu(E_i) \right\} \leq \frac{M}{n},
\end{eqnarray*}
while
$$\sum_{\tau_i\notin N}f(\tau_i)\mu(E_i)=\sum_{\tau_i\notin N}g(\tau_i)\mu(E_i),$$
and finally
\begin{eqnarray*}
&& |\sigma(g,\Pi)-J| \leq\\
&&\leq |\sigma(f,\Pi)-J|+\sum_{\tau_i\in N}|f(\tau_i)|\mu(E_i)+\\
&&+\sum_{\tau_i\in N}|g(\tau_i)|\mu(E_i)\leq b_n+2\frac{M}{n}.
\end{eqnarray*}
This clearly proves that $g$ is (oH)-integrable with integral $J$.
Finally, interchanging the role of $f$ and $g$, the
assertion follows.\ \fine
\\
\medskip
Another interesting difference is in the validity of the so-called Henstock Lemma: indeed,
(oH)-integrability yields this result, contrarily to the case of the norm-integral.\\
A Cauchy-type criterion is stated first in order to prove the existence of the (oH)-integral. The proof is straightforward.
\begin{theorem}\label{ordercauchy}
Let $f:T\rightarrow X$ be any mapping. Then $f$ is {\rm (oH)}-integrable if and only if there exist an $(o)$-sequence $(b_n)_n$ and a corresponding sequence $(\gamma_n)_n$ of gages, such that for every $n$, as soon as $\Pi, \Pi'$ are two $\gamma_n$-fine Henstock partitions, the following holds:
\begin{eqnarray}\label{cauchyprimo}
|\sigma(f,\Pi)-\sigma(f,\Pi')|\leq b_n
\end{eqnarray}
\end{theorem}
\begin{comment}
{\bf Proof:} The necessity of the condition is trivial. So, assume that (\ref{cauchyprimo}) holds true: without loss of generality, suppose that the sequence $(\gamma_n)_n$ is decreasing. Now, for every $n$, define
\begin{eqnarray*}
M_n &:=& \sup\{\sigma(f,\Pi),\ \Pi\ \, \mbox{is} \,\, \gamma_n-\mbox{fine}\},\\
m_n &:=& \inf\{\sigma(f,\Pi),\ \Pi\ \, \mbox{is} \,\, \gamma_n-\mbox{fine}\}
\end{eqnarray*}
Of course it is $m_n\leq m_{n+1}\leq M_{n+1}\leq M_n$
for every $n$, and also $M_n-m_n\leq b_n$ thanks to (\ref{cauchyprimo}).
This entails that $\sup_n m_n=\inf_n M_n$:
the common value $J$
will be
the announced integral. Indeed, for the same (decreasing) sequence $(\gamma_n)_n$,
$$J-\sigma(f,\Pi)\leq J-m_n\leq M_n-m_n\leq b_n$$
and
$$\sigma(f,\Pi)-J\leq M_n-J\leq M_n-m_n\leq b_n$$
both hold, for every $\gamma_n$-fine partition $\Pi$.
This clearly suffices to prove the claim. \fine\\
\end{comment}
Now, a Henstock-type lemma is stated, for the (oH)-integral.
The proof is similar to that of \cite[Theorem 1.4]{mimmoroma}, and follows from the Cauchy criterion. However,
some details are also given here.\\
\begin{proposition}\label{henlemma}
Let $f:T\to X$ be any {\rm (oH)}-integrable function. Then, there exist an $(o)$-sequence $(b_n)_n$ and a corresponding sequence $(\gamma_n)_n$ of gages, such that, for every $n$ and every $\gamma_n$-fine Henstock partition $\Pi$ it
is
$$\sum_{E\in \Pi}Ob_n(f,E)\leq b_n,$$
where $$Ob_n(E)=\sup_{\Pi'_E,\Pi''_E}\{|\hskip-2mm\sum_{F''\in \Pi''_E}f(\tau_{F''})\mu(F'') - \hskip-2mm\sum_{F'\in \Pi'_E}f(\tau_{F'})\mu(F')|\},$$
and $\Pi'_E, \Pi''_E$ run along all $\gamma_n$-fine Henstock partitions of $E$.
\end{proposition}
{\bf Proof:}\ First observe that, thanks to the Cauchy criterion, an $(o)$-sequence $(b_n)_n$ exists, together with a corresponding sequence of gages $(\gamma_n)_n$, such that
\begin{eqnarray}\label{primocauchy}
|\sum_{F'\in \Pi'}f(\tau_{F'})\mu(F')-\sum_{F''\in \Pi''}f(\tau_{F''})\mu(F'')|\leq b_n
\end{eqnarray}
(with obvious meaning of symbols)
holds, for all $\gamma_n$-fine partitions $\Pi', \ \Pi''$. Now, take any $\gamma_n$-fine partition $\Pi$, and, for each element $E$ of $\Pi$, consider two arbitrary subpartitions $\Pi'_E$ and $\Pi''_E$.
Then, taking the {\em union} of the subpartitions $\Pi'_E$ as $E$ varies, and making the same operation with the subpartitions $\Pi''_E$, two $\gamma_n$-fine partitions of $T$ are obtained, for which (\ref{primocauchy}) holds true. From (\ref{primocauchy}), obviously it follows
\begin{eqnarray}\label{secondocauchy}
\sum_{F'\in \Pi'}f(\tau_{F'})\mu(F')- \sum_{F''\in \Pi''}f(\tau_{F''})\mu(F'')\leq b_n.
\end{eqnarray}
Now, let $E_1$ be the first element of $\Pi$.
In the summation at left-hand side, fix all the $F's$ and the $F''s$ that are not contained in $E_1$.
Taking the supremum when the remaining $F's$ and $F''s$ vary in all possible ways, it follows
\begin{comment}
$$\sup_{\Pi'_{E_1}}\sigma(f,\Pi'_{E_1})+\sum_{\substack{F'\in \Pi',\\ F'\not\subset E_1}}f(\tau_{F'})\mu(F')-\sum_{F''\in \Pi''}\sigma(f,\Pi)\leq b_n.$$
Now, by varying only the $F''s$ that are contained in $E_1$, the following inequality can be deduced:
\begin{eqnarray*}
&& \sup_{\Pi'_{E_1}}\sigma(f,\Pi'_{E_1})-\inf_{\Pi''_{E_1}}\sigma(f,\Pi''_{E_1})+\\
&&+ \sum_{\substack{F'\in \Pi', \\ F'\not\subset E_1}}f(\tau_{F'})\mu(F')-\sum_{F''\not\subset E_1}\sigma(f,\Pi)\leq b_n,
\end{eqnarray*}
namely
\end{comment}
\begin{eqnarray*}
Ob_n(f,E_1)+ \hskip-2mm \sum_{\substack{F'\in \Pi',\\ F'\not\subset E_1}}f(\tau_{F'})\mu(F') - \hskip-2mm
\sum_{\substack{F''\in \Pi'',\\F''\not\subset E_1}}f(\tau_{F''})\mu(F'')\leq b_n.
\end{eqnarray*}
In the same fashion, fixed all the $F'$ and $F''$ that are not contained in the second subset of $\Pi$, (say $E_2$), and making the same operation, it follows
\begin{eqnarray*}
\hskip-1cm&& Ob_n(f,E_1)+Ob_n(f,E_2)+\hskip-2mm \sum_{\substack{F'\in \Pi',\\ F'\not\subset E_1\cup E_2}}f(\tau_{F'})\mu(F') + \\
&&- \sum_{\substack{F''\in \Pi'',\\F''\not\subset E_1\cup E_2}}f(\tau_{F''})\mu(F'')\leq b_n
\end{eqnarray*}
Now, it is clear how to deduce the assertion. \fine\\
\begin{remark}\label{subinterval}\rm
A first consequence of the previous Proposition \ref{henlemma} is that any (oH)-integrable function $f$ is also integrable
in the same sense in every measurable subset $A$. Indeed,
taking the same $(o)$-sequence $(b_n)_n$ and the same corresponding sequence $(\gamma_n)_n$ as for integrability of $f$,
for each $n$ any $\gamma_n$-fine partition of $A$ can be extended to a $\gamma_n$-fine partition of $T$ thanks to the Cousin Lemma, and so, for any two $\gamma_n$-fine partitions $\Pi, \ \Pi'$ of $A$, it follows
$$|\sigma(f,\Pi)-\sigma(f,\Pi')|\leq Ob_n(f,A)\leq b_n.$$
Then, the Cauchy criterion yields the conclusion.\\
\end{remark}
\begin{remark}\label{additivity}\rm
By means of usual techniques, one also proves additivity of the integral:
namely
whenever $f$ is integrable in $T$, and $A,B$ are two disjoint measurable subsets of $T$, then $\int f 1_{A\cup B}d\mu=\int_A f d\mu +\int_B f d\mu$.\\
\end{remark}
The following Theorem collects some easy consequences of Proposition \ref{henlemma}.
\begin{theorem}\label{henstoc2}
Let $f:T\to X$ be any {\rm (oH)}-integrable function. Then there exist an $(o)$-sequence $(b_n)_n$ and a corresponding sequence $(\gamma_n)_n$ of gages, such that:
\begin{description}
\item[\ref{henstoc2}.1)]
for every $n$ and every $\gamma_n$-fine partition $\Pi$
one has
$$\sum_{E\in \Pi}|f(\tau_E)\mu(E)-{\rm (oH)}\int_Ef d\mu|\leq b_n.$$
\item[\ref{henstoc2}.2)]
for every $n$ and every $\gamma_n$-fine partition $\Pi$ it holds
$$\sum_{E\in \Pi}|f(\tau_E)\mu(E)-f(\tau_E')\mu(E)|\leq b_n,$$
as soon as all the tags satisfy the condition $E\subset \gamma_n(\tau_E')$ and $E\subset \gamma_n(\tau_E)$ for all $E$.
\end{description}
\end{theorem}
\begin{comment}
{\bf Proof:} Just some hints are given. To deduce \ref{henstoc2}.1) it is sufficient to fix $n$ and an arbitrary $\gamma_n$-fine partition $\Pi\equiv(E,\tau_E)$, and observe that (with the same meaning of symbols as above)
$$|f(\tau_E)\mu(E)-\sum_{F''\in \Pi''_E}f(\tau_{F''})\mu(F'')|\leq Ob_n(f,E)$$
holds true, for every $(\gamma_k)$-fine subpartition $\Pi''_E$ of $E$ and every $k>n$, and so
\begin{eqnarray*}
|f(\tau_E)\mu(E)-\int_E f | &\leq&
|f(\tau_E)\mu(E)-\sum_{F''\in \Pi''_E}f(\tau_{F''})\mu(F'')|+ \\
&+& |\sum_{F''\in \Pi''_E}f(\tau_{F''})\mu(F'')-{\rm (oH)}\int_E f| \leq\\
&\leq& Ob_n(f,E)+b_k
\end{eqnarray*}
By arbitrariness of $k$ and Proposition \ref{henlemma} the assertion follows.
\\
In order to obtain \ref{henstoc2}.1), it is enough to observe that
$$|f(\tau_E)\mu(E)-f(\tau_E')\mu(E)|\leq Ob_n(E)$$
when the tags are chosen as prescribed. \fine\\
\end{comment}
\begin{remark}\label{paracool}\rm
For further reference,
observe
that in the theorem above all partitions may also be free,
since, as already noticed, the restriction $\tau_E\in E$ does not affect the results.
\end{remark}
A consequence of this theorem is that
the (oH)-integrability of $f$ implies
the
(oH)-integrability of $|f|$.
\begin{theorem}\label{modulointegrabil}
If $f:T\to X$ is {\rm (oH)}-integrable, then also $|f|$ is.
\end{theorem}
{\bf Proof:}
The Cauchy criterion is used in the following formulation:
there exist an $(o)$-sequence $(b_n)_n$ and a corresponding sequence $(\gamma_n)_n$ of gages, such that, for each $n$, as soon as $\Pi,\Pi'$ are $\gamma_n$-fine free partitions and $\Pi'$ is finer than $\Pi$,
\begin{eqnarray}\label{cauchyy}
\big|\sum_{E\in \Pi}|f(\tau_E)|\mu(E)-\sum_{E'\in \Pi'}|f(\tau_{E'})|\mu(E')\big|\leq b_n.
\end{eqnarray}
\begin{comment}
(For the sufficiency of this formulation, one can proceed as in \cite{mimmoroma}: if the condition (\ref{cauchyy}) is satisfied, taking any two $\gamma_n$-fine partitions, $\Pi^1$ and $\Pi^2$, then it is always possible to find a $\gamma$-fine partition $\Pi'$, refining both $\Pi^1$ and $\Pi^2$: then from (\ref{cauchyy})
it follows
\begin{eqnarray*}
&& |\sigma(f,\Pi^1)-\sigma(f,\Pi^2)|\leq |\sigma(f,\Pi^1)-\sigma(f,\Pi')| +\\
+&& |\sigma(f,\Pi')-\sigma(f,\Pi^2)|\leq 2b_n).
\end{eqnarray*}
\end{comment}
Now, if $(b_n)_n$ and $(\gamma_n)_n$ are as in Theorem \ref{henstoc2}, and $\Pi$ and $\Pi'$ are as above,
it is
\begin{eqnarray*}
&& \sum_{E\in \Pi}|f(\tau_E)|\mu(E)-\sum_{E'\in \Pi'}|f(\tau_{E'})|\mu(E') =\\
=&&\sum_{E\in \Pi}\sum_{\substack{E'\in \Pi',\\ E'\subset E}}(|f(\tau_E)|-|f(\tau_{E'})|)\mu(E')= \\
=&& \sum_{E'\in \Pi'}(|f(\tau^*_{E'})|-|f(\tau_{E'})|)\mu(E'),
\end{eqnarray*}
where the tags $\tau^*_{E'}$ coincide with $\tau_E$ whenever $E'\subset E$, $E\in \Pi$.
Therefore, a simple application of \ref{henstoc2}.2) (and Remark \ref{paracool}) leads to (\ref{cauchyy}) and the proof is finished. \fine
\\
The last theorem can be compared with a previous result by Drewnowski and Wnuk, (\cite[Theorem 1]{DW}), where the Bochner integral is considered, for functions taking values in a Banach lattice $X$, and it is proven that the {\em modulus} of
the indefinite Bochner integral of $f$ is precisely the indefinite integral of $|f|$.
Here a similar result for {\rm (oH)}-integrable mappings is stated, after introducing a new definition.
\begin{definition}\label{modulus}\rm
Let $f:T\to X$ be any {\rm (oH)}-integrable mapping, and set
$$\mu_f(A)={\rm (oH)}\int_A f d\mu$$
for all Borel sets $A\in \mathcal{B}$.
Then $\mu_f$ is said to be the {\em indefinite integral} of $f$. The {\em modulus} of $\mu_f$, denoted by $|\mu_f|$, is defined for each $A\in \mathcal{B}$ as follows:
$$|\mu_f|(A)=\sup\{\sum_{B\in \pi}|\mu_f(B)|: \pi\in \Pi(A)\}$$
where $\Pi(A)$ is the family of all finite partitions of $A$.
(The {\em boundedness} of this quantity will be proven soon).\\
\end{definition}
Now
the following theorem can be stated.\\
\begin{theorem}\label{parallelo}
Assume that $f:T\to X$ is {\rm (oH)}-integrable. Then
one has
$$|\mu_f|= \mu_{|f|}.$$
\end{theorem}
{\bf Proof:} First of all, observe that $|f|$ is integrable too, thanks to Theorem \ref{modulointegrabil}. Since clearly
$$\left|{\rm(oH)}\int_B fd\mu \right|\leq {\rm (oH)}\int_B |f| d\mu$$
holds for every set $B\in \mathcal{A}$, it follows
that
$$|\mu_f|\leq \mu_{|f|}.$$
This also shows that the modulus $|\mu_f|$ is bounded.
Now, since $|\mu_f|$ and $\mu_{|f|}$ are additive, in order to obtain the reverse inequality, it will be sufficient to prove that $\mu_{|f|}(T)=|\mu_f|(T)$.
To this aim,
the Henstock Lemma, \ref{henstoc2}, will be used, and in particular \ref{henstoc2}.1). Let $(b_n)_n$ and $(\gamma_n)_n$ be an $(o)$-sequence and its corresponding sequence of gages, related to integrability of both $f$ and $|f|$.
\begin{comment}
such that, for every $n$ and every $\gamma_n$-fine partition $\pi\equiv (E_i,t_i)_i$ it holds
\begin{eqnarray}\label{prim}
\sum_i\left|f(t_i)\mu(E_i)-{\rm (oH)}\int_{E_i}f d\mu\right|\leq b_n,
\end{eqnarray}
and
\begin{eqnarray}\label{second}
\sum_{i}\left| |f(t_i)|\mu(E_i)-{\rm (oH)}\int_{E_i}|f |d\mu\right|\leq b_n.
\end{eqnarray}
From (\ref{prim})
\begin{eqnarray}\label{terz}
\sum_{i}\left|\ |f(t_i)|\mu(E_i)-\left|{\rm (oH)}\int_{E_i}f d\mu \right|\ \right|\leq b_n
\end{eqnarray}
follows. So,
\end{comment}
So, for every $n$ and every $\gamma_n$-fine partition $\pi\equiv (E_i,t_i)_i$ it holds
\begin{eqnarray*}
\hskip-.5cm
\mu_{|f|}(T)&-\hskip-.2cm&|\mu_f|(T) \leq \sum_i\left(\mu_{|f|}(E_i)-\left|{\rm (oH)}\int_{E_i}fd\mu \right|\right) \leq
\\ \leq &&
\sum_i\left (\mu_{|f|}(E_i)-|f(t_i)|\mu(E_i)\right)+\\
+&&
\sum_i\left (|f(t_i)|\mu(E_i)-\left|{\rm (oH)}\int_{E_i}fd\mu\right|\right)\leq 2 b_n
\end{eqnarray*}
Since $(b_n)_n$ is an $(o)$-sequence, then $\mu_{|f|}(T)\leq|\mu_f|(T)$, and so clearly also $\mu_{|f|}(T)=|\mu_f|(T)$. This concludes the proof. \fine
\\
Another interesting consequence is concerned with $L$-spaces. Recall that a Banach lattice $X$ is an $L$-space if its norm $\|\cdot\|$ satisfies
$$\|x+y\|=\|x\|+\|y\|$$
for all positive elements $x,y$ in $X$ (see also \cite{fvol3}).\\
The following definition, related with norm-integrability, is needed.\\
\begin{definition} \rm \cite[Definition 3]{DPVM}
$f:T \to X$ is {\em variationally {\rm H} integrable} (in short {\rm vH}-integrable) if
for every $\varepsilon>0$ there exists a gage $\gamma$ such that, for every $\gamma$-fine partition $\Pi\equiv(E,t_E)_E$ the following holds:
\begin{eqnarray}\label{variazio}
\sum_{E\in \Pi}\|f(t_E)\mu(E)-{\rm (H)}\int_E f d\mu\|\leq \varepsilon.
\end{eqnarray}
\end{definition}
For results on this setting see also \cite{DPVM}. \\
\begin{theorem}\label{Lspazio}
Let $f:T\to X$ be {\rm (oH)}-integrable, and assume that $X$ is an $L$-space. Then $f$ is Bochner integrable.
\end{theorem}
{\bf Proof:}\ In order to prove the Bochner integrability, it will suffice to show that
$f$ is vH-integrable.
Indeed, by \cite[Theorem 2]{DPMILL},
the
variational integrability implies
the
Bochner integrability.
The H-integrability of $\|f\|$ will be proved first.
In accordance with the previous Theorem \ref{henstoc2}, and with the same meanings of symbols, there exist an $(o)$-sequence $(b_n)_n$ and a corresponding sequence $(\gamma_n)_n$ of gages, such that, for every $n$ and every $\gamma_n$-fine partition $\Pi$
one has
$$\sum_{E\in \Pi}|f(\tau_E)\mu(E)-f(\tau_E')\mu(E)|\leq b_n.$$
Since the norm of $X$ is compatible with the order,
then
$\lim_n\|b_n\|=0$. So, fix $\varepsilon>0$ and pick any integer $N$ such that $\|b_N\|\leq \varepsilon$.
Then, if $\Pi$ is any $\gamma_N$-fine partition, we have
$$\left\| \, \sum_{E\in \Pi}|f(\tau_E)\mu(E)-f(\tau_E')\mu(E)| \, \right\| \leq \|b_N \| \leq \varepsilon,$$
in accordance with \ref{henstoc2}.2). Thanks to the particular nature of the norm $\|\cdot \|$, we deduce
that
\begin{eqnarray*}
\sum_{E\in \Pi}\left\| \, |f(\tau_E)\mu(E)-f(\tau_E')\mu(E)| \, \right\|\leq \|b_N\|\leq \varepsilon,
\end{eqnarray*}
and so
\begin{eqnarray}\label{normaepsilo}
\sum_{E\in \Pi} \, \big|\|f(\tau_E)\|-\|f(\tau_E')\|\big|\mu(E) \, \leq \|b_N\|\leq \varepsilon,\end{eqnarray}
as soon as $\Pi$ is $\gamma_N$-fine, both for the tags $\tau_E$ and for the tags $\tau_E'$.
\\
Now, proceeding as in the proof of Theorem \ref{modulointegrabil}, it is not difficult to prove that $\|f\|$ satisfies the Cauchy criterion for the Henstock integrability, and therefore it is integrable.
From (\ref{normaepsilo}), also thanks to \ref{henstoc2}.1), it is easy to deduce also (\ref{variazio}).
In conclusion, $f$ is variationally integrable, its norm is integrable, and then, from Pettis integrability, it follows also the Bochner integrability.
\fine\\
\begin{remark}\label{HnonoH}\rm
The previous result can be used to show that
the
(H)-integrability in general does not imply
the
(oH)-integrability: indeed, if $X$ is any infinite-dimensional Banach space, there exists a McShane (norm)-integrable map $f:[0,1]\to X$ that is not Bochner integrable (see \cite{SS}). In particular, when $X$ is an $L$-space (of infinite dimension), such function $f$ cannot be (oH)-integrable, in view of Theorem \ref{Lspazio}.
\end{remark}
\begin{comment}
\begin{remark}\label{LMspazi}\rm
Another byproduct of the previous results is that an $M$-space $X$ admits an equivalent $L$-norm only if it is of finite dimension. In fact in an $M$-space $X$
the
(H)-integrability and
the
(oH)-integrability are the same: so, if $X$ has an equivalent $L$-norm then a simple glance at the Remark \ref{HnonoH} is enough to conclude that $X$ must be of finite dimension.
\end{remark}
\end{comment}
\section{Integrability in $[0,1]$}
In this section
the (oH)-integrability of functions $f:[0,1]\to X$ is studied, where
$[0,1]$ is endowed with the usual Lebesgue measure $\lambda$ . \\
From now on,
only {\em free} partitions consisting of subintervals of $[0,1]$ are considered, rather than arbitrary measurable subsets. Indeed, in \cite{f1995} it is proven that there is equivalence between the two types: though the proof there is related only to
norm integrals, the technique is the same. For this reason, from now on
the symbol
(oM)-integral will be used rather than (oH)-integral. \\
A useful result, parallel to \cite[ Lemma 5.35]{KS}, is the following: \\
\begin{lemma}\label{come5.35}
Let $f:[0,1]\to X$ be any fixed function, and suppose that there exists an $(o)$-sequence $(b_n)_n$ such that, for every $n$
two {\rm (oM)}-integrable functions $g_1$ and $g_2$ can be found, with the same regulating $(o)$-sequence
$(\beta_n)_n$, such that $g_1\leq f\leq g_2$ and ${\rm (oM)}\int g_2 d\lambda \leq {\rm (oM)}\int g_1 d\lambda+b_n.$
Then $f$ is {\rm (oM)}-integrable.
\end{lemma}
\begin{comment}
{\bf Proof:} Fix $n$, and choose two functions $g_1$ and $g_2$ as above. Then there exists a gage $\gamma$ such that, for every $\gamma$-fine free partition $\Pi$
it is
\begin{eqnarray*}
&&{\rm (oM)}\int g_1d\lambda-\beta_n < \sigma(g_1,\Pi)\leq \sigma(f,\Pi)\leq \sigma(g_2,\Pi)<\\
&&<{\rm (oM)}\int g_2 d\lambda +\beta_n\leq {\rm (oM)}\int g_1 d\lambda +b_n+\beta_n.
\end{eqnarray*}
From this
one can easily deduce that, as soon as $\Pi, \Pi'$ are $\gamma$-fine free partitions:
$$|\sigma(f,\Pi)-\sigma(f,\Pi')|\leq 2\beta_n+b_n.$$
Now, the Cauchy criterion gives integrability of $f$. \fine\\
\end{comment}
\medskip
The fact that increasing functions are (oM)-integrable can be deduced similarly as in \cite[example 5.36]{KS}.
\begin{theorem}\label{mcmonotone}
Let $f:[0,1]\to X$ be increasing. Then $f$ is {\rm (oM)}-integrable.
\end{theorem}
\begin{comment}
{\bf Proof:}\ Fix any positive integer $n$, and set $t_i=\frac{i}{n}$, for $i=0,...,n$. Set also $E_i:=[i,i+1[$, for $i=0,...,n-1$. Now, define
$$g_1:=\sum_{i=1}^{n-1}f(t_i)1_{E_i} , \ \ g_2:=\sum_{i=1}^{n-1}f(t_{i+1})1_{E_i}.$$
Since $g_1$ and $g_2$ are step functions, they are (oM)-integrable, and obviously $g_1\leq f\leq g_2$ in $[0,1]$. Now,
it is
$${\rm (oM)}\int g_2 d\lambda- {\rm (oM)}\int g_1 d\lambda =\frac{f(1)-f(0)}{n},$$
after simple calculations.
Since $b_n=(f(1)-f(0)) n^{-1}$ is clearly an $(o)$-sequence, the conclusion follows from Lemma \ref{come5.35}. \fine \\
\end{comment}
\medskip
\begin{remark}
In a similar way, one can prove
the
(oM)-integrability more generally for (order bounded) mappings that are Riemann-integrable in the order sense.
\end{remark}
\section{Conclusion}
In this paper the notions of Henstock and McShane integrability for functions defined in a metric compact regular space and taking values in a Banach lattice with an order-continuous norm are investigated. Both the norm-type and the order-type integrals have been studied and compared. Though in general the order-type integral is stronger than the norm-one,
in $M$-spaces the two notions coincide, while in $L$-spaces the order-type Henstock integral is indeed a Bochner one. Finally, the particular case of functions defined in a real interval is considered, and it is proven that monotone mappings are always order-McShane integrable.
\vspace{1cm}
\noindent \textbf{Acknowledgment} The authors have been supported by
University of Perugia -- Department of Mathematics and Computer Sciences - Grant Nr 2010.011.0403,
Prin "Metodi logici per il trattamento dell'informazione",
Prin "Descartes"
and by the Grant prot. U2014/000237 of GNAMPA - INDAM (Italy).
|
\section{Introduction}
The Lefschetz fibrations are fundamental objects to study in $4$-dimensional topology. In the remarkable works \cite{D1, GS}, Simon Donaldson showed that every closed symplectic $4$-manifold admits a structure of Lefschetz pencil, which can be blown up at its base points to yield a Lefschetz fibration, and conversely, Robert Gompf showed that the total space of a genus $g$ Lefschetz fibration admits a symplectic structure, provided that the homology class of the fiber is nontrivial. Given a Lefschetz fibration over $\mathbb{S}^2$, one can associate to it a word in the mapping class group of the fiber composed solely of right-handed Dehn twists, and conversely, given such a factorization in the mapping class group, one can construct a Lefschetz fibration over $\mathbb{S}^2$ (see for example \cite{GS}).
Recently there has been much interest in trying to understand the topological interpretation of various relations in the mapping class group. A particularly well understood case is the daisy relation, which corresponds to the symplectic operation of rational blowdown \cite{EG, EMVHM}. Another interesting problem, which is still open, is whether any Lefschetz fibration over $\mathbb{S}^2$ admits a section (see for example \cite{Sm}). Furthermore, one would like to determine how many disjoint sections the given Lefschetz fibration admits. The later problem has been studied for the standard family of hyperelliptic Lefschetz fibrations (with total spaces $\CP\#(4g+5)\CPb$) in \cite{ko, Sinem, Tan}, using the computations in the mapping class group, and such results are useful in constructing (exotic) Stein fillings \cite{ao2, AO}.
Motivated by these results and problems, our goal in this paper is to construct new families of Lefschetz fibrations over $\mathbb{S}^{2}$ by applying the sequence of daisy substitutions and conjugations to the hyperelliptic words $(c_1c_2 \cdots c_{2g-1}c_{2g}{c_{2g+1}}^2c_{2g}c_{2g-1} \cdots c_2c_1)^2 = 1$, $(c_1c_2 \cdots c_{2g}c_{2g+1})^{2g+2} = 1$, and $(c_1c_2 \cdots c_{2g-1}c_{2g})^{2(2g+1)} = 1$ in the mapping class group of the closed orientable surface of genus $g$ for any $g \geq 3$ and study the sections of these Lefschetz fibrations (cf. Theorems \ref{4.1}-\ref{4.4}). Furthermore, we show that the total spaces of our Lefschetz fibraions given by the last two words are irreducible exotic symplectic $4$-manifolds, and compute their Seiberg-Witten invariants (cf. Theorem \ref{theorem1}). The analogues (but weaker) results for special case of $g = 2$, using the lantern substitutions only, were obtained in \cite{EG, AP}. We would like to remark that the mapping class group computations in our paper are more involved and subtle than in \cite{EG, AP}. One family of examples, obtained from the fiber sums of the Lefschetz fibrations using daisy relations, were studied in \cite{EMVHM}. However, the examples obtained in \cite{EMVHM} have larger topology, and computations of Seiberg-Witten invariants, and study of sections were not addressed in \cite{EMVHM}. Moreover, we prove non-hyperellipticity of our Lefschetz fibrations and provide some criterias for non-hyperellipticity under the daisy substitutions (cf. Theorem \ref{nonhyperelliptic}). Some of our examples can be used to produce the families of non-isomorphic Lefschetz fibrations over $\mathbb{S}^{2}$ with the same total spaces and exotic Stein fillings. We hope to return these examples in future work.
The organization of our paper is as follows. In Sections 2 and 3 we recall the main definitions and results that will be used throughout the paper. In Section 4, we prove some technical lemmas, important in the proofs of our main theorems. In Sections 5 and 6, we construct new families of Lefschetz fibrations by applying the daisy substitutions to the words given above, study the sections and prove non-hyperellipticity of our Lefschetz fibrations (Theorems \ref{4.1}, \ref{4.2}, \ref{4.7}, \ref{4.3}, \ref{4.4}, \ref{nonhyperelliptic}). Finally, in Section 7, we prove that the total spaces of some of these Lefschetz fibrations are exotic symplectic $4$-manifolds, which we veirfy by computing their Seiberg-Witten invariants and obtain infinite family of exotic $4$-manifolds via knot surgery (Theorems \ref{theorem1}, \ref{theorem3}), and make some remarks and raise questions. We would like to remark that the main technical content of our paper is more algebraic since our proofs rely heavily on mapping class group techniques. It is possible to pursue a more geometric approach (see Example~\ref{Ex}), but such approach alone does not yield the optimal results as presented here.
\section{Mapping Class Groups} Let $\Sigma_{g}^n$ be a $2$-dimensional, compact, oriented, and connected surface of genus $g$ with $n$ boundary components.
Let $Diff^{+}\left( \Sigma_{g}^n\right)$ be the group of all orientation-preserving self-diffeomorphisms of $\Sigma_{g}^n$ which are the identity on the boundary and $ Diff_{0}^{+}\left(\Sigma_{g}\right)$ be the subgroup of $Diff^{+}\left(\Sigma_{g}\right)$ consisting of all orientation-preserving self-diffeomorphisms that are isotopic to the identity.
The isotopies are also assumed to fix the points on the boundary.
\emph{The mapping class group} $\Gamma_{g}^n$ of $\Sigma_{g}^n$ is defined to be the group of isotopy classes of orientation-preserving diffeomorphisms of $\Sigma_{g}^n$, i.e.,
\[
\Gamma_{g}^n=Diff^{+}\left( \Sigma_{g}^n\right) /Diff_{0}^{+}\left(
\Sigma_{g}^n\right) .
\]
For simplicity, we write $\Sigma_g = \Sigma_g^0$ and $\Gamma_g = \Gamma_g^0$.
The hyperelliptic mapping class group $H_{g}$ of $\Sigma_{g}$ is defined as the subgroup of $\Gamma_g$ consisting of all isotopy classes commuting with the isotopy class of the hyperelliptic involution $\iota: \Sigma_{g}\rightarrow \Sigma_{g}$.
\begin{defn} Let $\alpha$ be a simple closed curve on $\Sigma_{g}^n$. A \emph{right handed} (or positive) \emph{Dehn twist} about $\alpha$ is a diffeomorphism of $t_{\alpha}: \Sigma_{g}^n\rightarrow \Sigma_{g}^n$ obtained by cutting the surface $\Sigma_{g}^n$ along $\alpha$ and gluing the ends back after rotating one of the ends $2\pi$ to the right.
\end{defn}
It is well-known that the mapping class group $\Gamma_g^n$ is generated by Dehn twists. It is an elementary fact that the conjugate of a Dehn twist is again a Dehn twist: if $\phi: \Sigma_{g}^n\rightarrow \Sigma_{g}^n$ is an orientation-preserving diffeomorphism, then $\phi \circ t_\alpha \circ \phi^{-1} = t_{\phi(\alpha)}$. The following lemma is easy to verify (see \cite{I} for a proof).
\begin{lem} \label{com&braid.lem} Let $\alpha$ and $\beta$ be two simple closed curves on $\Sigma_{g}^n$. If $\alpha$ and $\beta$ are disjoint, then their corresponding Dehn twists satisfy the commutativity relation: $t_{\alpha}t_{\beta}=t_{\beta}t_{\alpha}.$ If $\alpha$ and $\beta$ transversely intersect at a single point, then their corresponding Dehn twists satisfy the braid relation: $t_{\alpha}t_{\beta}t_{\alpha}=t_{\beta}t_{\alpha}t_{\beta}.$
\end{lem}
\subsection{Daisy relation and daisy substitution}
We recall the definition of the daisy relation (see \cite{PVHM}, \cite{EMVHM}, \cite{BKM}).
\begin{defn}\label{daisy}\rm
Let $\Sigma_0^{p+2}$ denote a sphere with $p+2$ boundary components $(p\geq 2)$. Let $\delta_0, \delta_1, \delta_2,\ldots, \delta_{p+1}$ be the $p$ boundary curves of $\Sigma_0^{p+2}$ and let $x_1, x_2,\ldots, x_{p+1}$ be the interior curves as shown in Figure~\ref{daisy}. Then, we have the \textit{daisy relation of type $p$}:
\begin{align*}
t_{\delta_0}^{p-1}t_{\delta_1}t_{\delta_2}\cdots t_{\delta_{p+1}}=t_{x_1}t_{x_2}\cdots t_{x_{p+1}}.
\end{align*}
We call the following relator the \textit{daisy relator of type $p$}:
\begin{align*}
t_{\delta_{p+1}}^{-1}\cdots t_{\delta_2}^{-1}t_{\delta_1}^{-1}t_{\delta_0}^{-p+1}t_{x_1}t_{x_2}\cdots t_{x_{p+1}} \ (=1).
\end{align*}
\end{defn}
\begin{rmk}\rm
When $p=2$, the daisy relation is commonly known as the \textit{lantern relation} (see \cite{De}, \cite{Jo}).
\end{rmk}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=.43]{daisy.eps}
\caption{Daisy relation}
\label{fig:hyper}
\end{center}
\end{figure}
We next introduce a daisy substitution, a substitution technique introduced by T. Fuller.
\begin{defn}\label{daisy substitution}\rm
Let $d_1,\ldots,d_m$ and $e_1,\ldots, e_n$ be simple closed curves on $\Gamma_g^n$, and let $R$ be a product $R=t_{d_1}t_{d_2}\cdots t_{d_l}t_{e_m}^{-1}\cdots t_{e_2}^{-1}t_{e_1}^{-1}$. Suppose that $R=1$ in $\Gamma_g^n$. Let $\varrho$ be a word in $\Gamma_g^n$ including $t_{d_1}t_{d_2}\cdots t_{d_l}$ as a subword:
\begin{align*}
\varrho=U\cdot t_{d_1}t_{d_2}\cdots t_{d_l} \cdot V,
\end{align*}
where $U$ and $V$ are words. Thus, we obtain a new word in $\Gamma_g^n$, denoted by $\varrho^\prime$, as follows:
\begin{align*}
\varrho^\prime:&=U\cdot t_{e_1}t_{e_2}\cdots t_{e_m} \cdot V.
\end{align*}
Then, we say that $\varrho^{\prime}$ is obtained by applying a $R$-\textit{substitution} to $\varrho$. In particular, if $R$ is a daisy relator of type $p$, then we say that $\varrho^{\prime}$ is obtained by applying a \textit{daisy substitution of type $p$} to $\varrho$.
\end{defn}
\section{Lefschetz fibrations}
\begin{defn}\label{LF}\rm
Let $X$ be a closed, oriented smooth $4$-manifold. A smooth map $f : X \rightarrow \mathbb{S}^2$ is a genus-$g$ \textit{Lefschetz fibration} if it satisfies the following condition: \\
(i) $f$ has finitely many critical values $b_1,\ldots,b_m \in S^2$, and $f$ is a smooth $\Sigma_g$-bundle over $\mathbb{S}^2-\{b_1,\ldots,b_m\}$, \\
(ii) for each $i$ $(i=1,\ldots,m)$, there exists a unique critical point $p_i$ in the \textit{singular fiber} $f^{-1}(b_i)$ such that about each $p_i$ and $b_i$ there are local complex coordinate charts agreeing with the orientations of $X$ and $\mathbb{S}^2$ on which $f$ is of the form $f(z_{1},z_{2})=z_{1}^{2}+z_{2}^{2}$, \\
(i\hspace{-.1em}i\hspace{-.1em}i) $f$ is relatively minimal (i.e. no fiber contains a $(-1)$-sphere.)
\end{defn}
Each singular fiber is obtained by collapsing a simple closed curve (the \textit{vanishing cycle}) in the regular fiber. The monodromy of the fibration around a singular fiber is given by a right handed Dehn twist along the corresponding vanishing cycle.
For a genus-$g$ Lefschetz fibration over $\mathbb{S}^2$, the product of right handed Dehn twists $t_{v_i}$ about the vanishing cycles $v_i$, for $i = 1,\ldots, m$, gives us the global monodromy of the Lefschetz fibration, the relation $t_{v_1} t_{v_2} \cdots t_{v_m}=1$ in $\Gamma_g$. This relation is called the \textit{positive relator}. Conversely, such a positive relator defines a genus-$g$ Lefschetz fibration over $\mathbb{S}^2$ with the vanishing cycles $v_1,\ldots, v_m$.
According to theorems of Kas \cite{Kas} and Matsumoto \cite{Ma}, if $g\geq 2$, then the isomorphism class of a Lefschetz fibration is determined by a positive relator modulo simultaneous conjugations
\begin{align*}
t_{v_1} t_{v_2} \cdots t_{v_m} \sim t_{\phi(v_1)} t_{\phi(v_2)} \cdots t_{\phi(v_m)} \ \ {\rm for \ all} \ \phi \in M_g
\end{align*}
and elementary transformations
\begin{align*}
&t_{v_1} \cdots t_{v_{i-1}} t_{v_i} t_{v_{i+1}} t_{v_{i+2}} \cdots t_{v_m}& &\sim& &t_{v_1} \cdots t_{v_{i-1}} t_{t_{v_i}(v_{i+1})} t_{v_i} t_{v_{i+2}} \cdots t_{v_m},&\\
&t_{v_1} \cdots t_{v_{i-2}} t_{v_{i-1}} t_{v_i} t_{v_{i+1}} \cdots t_{v_m}& &\sim& &t_{v_1} \cdots t_{v_{i-2}} t_{v_i} t_{t_{v_i}^{-1}(v_{i-1})} t_{v_{i+1}} \cdots t_{v_m}.&
\end{align*}
Note that $\phi t_{v_i}\phi^{-1}=t_{\phi(v_i)}$.
We denote a Lefschetz fibration associated to a positive relator $\varrho \ \in \Gamma_g$ by $f_\varrho$.
For a Lefschetz fibration $f:X\rightarrow \mathbb{S}^2$, a map $\sigma:\mathbb{S}^2\rightarrow X$ is called a \textit{section} of $f$ if $f\circ \sigma={\rm id}_{\mathbb{S}^2}$.
We define the self-intersection of $\sigma$ to be the self-intersection number of the homology class $[\sigma(\mathbb{S}^2)]$ in $H_2(X;\Z)$.
Let $\delta_1,\delta_2,\ldots,\delta_n$ be $n$ boundary curves of $\Sigma_g^n$.
If there exists a lift of a positive relator $\varrho = t_{v_1} t_{v_2} \cdots t_{v_m} = 1$ in $\Gamma_g$ to $\Gamma_g^n$ as
\begin{align*}
t_{\tilde{v}_1} t_{\tilde{v}_2} \cdots t_{\tilde{v}_m} = t_{\delta_1} t_{\delta_2} \cdots t_{\delta_n},
\end{align*}
then $f_\varrho$ admits $n$ disjoint sections of self-intersection $-1$.
Here, $t_{\tilde{v}_i}$ is a Dehn twist mapped to $t_{v_i}$ under $\Gamma_g^n \to \Gamma_g$.
Conversely, if a genus-$g$ Lefschetz fibration admits $n$ disjoint sections of self-intersection $-1$, then we obtain such a relation in $\Gamma_g^n$.
Next, let us recall the signature formula for hyperelliptic Lefschetz fibrations, which is due to Matsumoto and Endo. We will make use of this formula in Section \ref{new words}, where we prove that all our Lefschetz fibrations obtained via daisy substitutions are non-hyperelliptic.
\begin{thm}[\cite{Ma1},\cite{Ma},\cite{E}]\label{sign} Let $f:X\rightarrow \mathbb{S}^2$ be a genus $g$ hyperelliptic Lefschetz fibration. Let $s_0$ and $s=\Sigma_{h=1}^{[g/2]}s_h$ be the number of non-separating and separating vanishing cycles of $f$, where $s_h$ denotes the number of separating vanishing cycles which separate the surface of genus $g$ into two surfaces, one of which has genus $h$. Then, we have the following formula for the signature
\begin{eqnarray*}
\sigma(X)=-\frac{g+1}{2g+1}s_0+\sum_{h=1}^{[\frac{g}{2}]}\left(\frac{4h(g-h)}{2g+1}-1\right)s_{h}.
\end{eqnarray*}
\end{thm}
\subsection{Spinness criteria for Lefschetz fibrations}
In this subsection, we recall two Theorems, due to A. Stipsicz (\cite{St}), concerning the non-spinness and spinness of the Lefschetz fibrations over $\mathbb{D}^2$ and $\mathbb{S}^2$. We will use them to verify our familes of Lefschetz fibrations in Theorem \ref{theorem1} all are non-spin. Since Rohlin's Theorem can not be used to verify non-spinness when the signature of our Lefchetz fibrations is divisible by $16$, Stipsicz's results will be more suitable for our purpose.
Let $f:X \rightarrow \mathbb{D}^2$ be a Lefschetz fibration over disk, and $F$ denote the generic fiber of $f$. Denote the homology classes of the vanishing cycles of the given fibration by $v_{1}, \cdots, v_{m} \in H_{1}(F; \mathbb{Z}_{2})$.
\begin{thm}\label{Spin1} \cite{St}. The Lefschetz fibration $f:X \rightarrow \mathbb{D}^2$ is not spin if and only if there are $l$ vanishing cylces $v_{1}$, $\cdots$, $v_{l}$ such that $v = \sum_{i=1}^{l} v_{i}$ is also a vanishing cycle, and $l + \sum_{1 \leq i < j \leq l} v_{i} \cdot v_{j} \equiv 0 (\textrm{mod}\ 2)$.
\end{thm}
Note that the above theorem imples that if the Lefschetz fibration has the separating vanishing cycle then its total space is not spin. To see this, set $l = 0$ and take the empty sum to be $0$.
\begin{thm}\label{Spin2} \cite{St}. The Lefschetz fibration $f:X \rightarrow \mathbb{S}^2$ is spin if and only if $X \setminus \nu(F)$ is spin and for some dual $\sigma$ of $F$ we have $\sigma^2 \equiv 0 (\textrm{mod}\ 2)$. \end{thm}
\subsection{Three familes of hyperelliptic Lefschetz fibrations}
In this subsection, we introduce three well-known familes of hyperelliptic Lefschetz fibrations, which will serve as building blocks in our construction of new Lefschetz fibrations.
Let $c_1$, $c_2$, .... , $c_{2g}$, $c_{2g+1}$ denote the collection of simple closed curves given in Figure~\ref{fig:hyper}, and $c_{i}$ denote the right handed Dehn twists $t_{c_i}$ along the curve $c_i$. It is well-known that the following relations hold in the mapping class group $\Gamma_g$:
\begin{equation}
\begin{array}{l}
H(g) = (c_1c_2 \cdots c_{2g-1}c_{2g}{c_{2g+1}}^2c_{2g}c_{2g-1} \cdots c_2c_1)^2 = 1, \\
I(g) = (c_1c_2 \cdots c_{2g}c_{2g+1})^{2g+2} = 1, \\
G(g) = (c_1c_2 \cdots c_{2g-1}c_{2g})^{2(2g+1)} = 1.
\end{array}
\end{equation}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=.40]{hyperelliptic.eps}
\caption{Vanishing Cycles of the Genus $g$ Lefschetz Fibration on $X(g)$, $Y(g)$, and $Z(g)$}
\label{fig:hyper}
\end{center}
\end{figure}
Let $X(g)$, $Y(g)$ and $Z(g)$ denote the total spaces of the above genus $g$ hyperelliptic Lefschetz fibrations given by the monodromies $H(g) = 1$, $I(g) = 1$, and $J(g) = 1$ respectively, in the mapping class group $\Gamma_g$. For the first monodromy relation, the corresponding genus $g$ Lefschetz fibrations over $\mathbb{S}^2$ has total space $X(g) = \CP\#(4g+5)\CPb$, the complex projective plane blown up at $4g+5$ points. In the case of second and third relations, the total spaces of the corresponding genus $g$ Lefschetz fibrations over $\mathbb{S}^2$ are also well-known families of complex surfaces. For example, $Y(2) = K3\#2\CPb$ and $Z(2)$ = \emph{Horikawa surface}, respectively. In what follows, we recall the branched-cover description of the $4$-manifolds $Y(g)$ and $Z(g)$, which we will use in the proofs of our main results. The branched-cover description of $X(g)$ is well-known and we refer the reader to (\cite{GS}, Remark 7.3.5, p.257).
\begin{lem}\label{E} The genus $g$ Lefschetz fibration on $Y(g)$ over $\mathbb{S}^2$ with the monodromy $(c_1c_2 \cdots c_{2g+1})^{2g+2} = 1$ can be obtained as the double branched covering of $\CP\#\CPb$ branched along a smooth algebraic curve $B$ in the linear system $|2(g+1)\tilde{L}|$, where $\tilde{L}$ is the proper transform of line $L$ in $\CP$ avoiding the blown-up point. Furthermore, this Lefschetz fibration admits two disjoint $-1$ sphere sections.
\end{lem}
\begin{proof} We will follow the proof of Lemma 3.1 in \cite{Ar1}, where $g=2$ case have been considered (see also the discussion in \cite{AK}), and make necessary adjustments where needed. Let $D$ denote an algebraic curve of degree $d$ in $\CP$. We fix a generic projection map $\pi : \CP \setminus {pt} \rightarrow \mathbb{CP}^1$ such that the pole of $\pi$ does not belong to $D$. It was shown in \cite{MT} that the braid monodromy of $D$ in $\CP$ is given via a braid factorization. More specifically, the braid monodromy around the point at infinity in $\mathbb{CP}^1$, which is given by the central element $\Delta^2$ in $B_{d}$, can be written as the product of the monodromies about the critical points of $\pi$. Hence, the factorization $\Delta^2 = (\sigma_{1} \cdots \sigma_{d-1})^{d}$ holds in the braid group $B_{d}$, where $\sigma_{i}$ denotes a positive half-twist exchanging two points, and fixing the remaining $d-2$ points.
Now let us degenerate the smooth algebraic curve $B$ in $\CP\#\CPb$ into a union of $2(g+1)$ lines in a general position. By the discussion above, the braid group factorization corresponding to the configuration $B$ is given by $\Delta^2 = (\sigma_{1} \sigma_{2} \cdots \sigma_{2g} \sigma_{2g+1})^{2g+2}$. Now, by lifting this braid factorization to the mapping class group of the genus $g$ surface, we obtain that the monodromy factorization $(c_1c_2 \cdots c_{2g+1})^{2g+2} = 1$ for the corresponding double branched covering.
Moreover, observe that a regular fiber of the given fibration is a two fold cover of a sphere in $\CP\#\CPb$ with homology class $f = h - e_{1}$ branched over $2(g+1)$ points, where $h$ denotes the hyperplane class in $\CP$. Hence, a regular fiber is a surface of genus $g$. The exceptional sphere $e_{1}$ in $\CP\#\CPb$, which intersects $f = h - e_{1}$ once positively, lifts to two disjoint $-1$ sphere sections in $Y(g)$.
\end{proof}
The proof of the following lemma can be extracted from \cite{GS} [Ex 7.3.27, page 268]; we omit proof.
\begin{lem}\label{E1} The double branched cover $W(g)$ of $\CP$ along a smooth algebraic curve $B$ in the linear system $|2(g+1)\tilde{L}|$ can be decomposed as the fiber sum of two copies of $\CP\#(g+1)^2\CPb$ along a a complex curve of genus equal $g(g-1)/2$. Moreover, $W(g)$ admits a genus $g$ Lefschetz pencil with two base points, and $Y(g) = W(g)\#2\CPb$.
\end{lem}
\begin{exmp}\label{Ex} In this example, we study the topology of complex surfaces $W(g)$ in some details. Recall that by Lemma \ref{E1} the complex surface $W(g)$ is the fiber sum of two copies of the rational surface $\CP\#(g^{2}+2g+1)\CPb$ along the complex curve $\Sigma$ of genus $g(g-1)/2$ and self-intersection zero. Using the fiber sum decomposition, we compute the Euler characteristic and the signature of $W(g)$ as follows: $e(W(g)) = 2e(\CP\#(g^{2}+2g+1)\CPb) - 2e(\Sigma) = 4g^2 + 2g + 4$, and $\sigma(W(g)) = 2 \sigma(\CP\#(g^{2}+2g+1)\CPb) = -2(g^{2}+2g)$. Next, we recall from \cite{Fu} that $\CP\#(g^{2}+2g+1)\CPb = {\Phi}_{g(g-1)/2}(1) \cup N_{g(g-1)/2}(1)$, where ${\Phi}_{g(g-1)/2}(1)$ and $N_{g(g-1)/2}(1)$ are Milnor fiber and generalized Gompf nucleus in $\CP\#(g^{2}+2g+1)\CPb$ respectively. Notice that such decomposition shows that the intersection form of $\CP\#(g^{2}+2g+1)\CPb$ splits as $N \oplus M(g)$, where $N = \bigl(\begin{smallmatrix} 0&1\\ 1&-1\end{smallmatrix}\bigr)$ and $M(g)$ is a matrix whose entries are given by a negative definite plumbing tree in the Figure~\ref{fig:plumb}. Consequently, we obtain the following decomposition of the intersection form of $W(g)$: $2M(g) \oplus H \oplus g(g-1) H$, where $H$ is a hyperbolic pair. Let us choose the following basis which realizes the intesection matrix $M(g) \oplus N$ of $\CP\#(g^{2}+2g+1)\CPb$: $ <f = (g+1)h - e_1 - \ \cdots \ - e_{(g+1)^{2}}, \ e_{(g+1)^{2}}, \ e_1 - e_2, \ e_2 - e_3, \ \cdots, \ e_{(g+1)^{2}-2} - e_{(g+1)^{2}-1}, \ h - e_{(g+1)^{2}-(g+1)} - \cdots - e_{(g+1)^{2}-2} - e_{(g+1)^{2}-1}>$. Observe that the last $(g+1)^{2} - 1$ classes can be represented by spheres and their self-intersections are given as in the Figure~\ref{fig:plumb}. $f$ is the class of fiber of the genus $g(g-1)/2$ Lefschetz fibration on $\CP\#(g^{2}+2g+1)\CPb$ and $e_{(g+1)^{2}}$ is a sphere section of self-intersection $-1$. Using the generalized fiber sum decomposition of $W(g)$, it is not hard to see the surfaces that generate the intersection matrix $2M(g) \oplus H \oplus g(g-1) H$. The two copies of the Milnor fiber ${\Phi}_{g(g-1)/2}(1) \subset \CP\#(g^{2}+2g+1)\CPb$ are in $W(g)$, providing $2((g+1)^2 - 1)$ spheres of self-intersections $-2$ and $-g$ (corresponding to the classes $\{ e_1 - e_2, \ e_2 - e_3, \ \cdots, \ e_{(g+1)^{2}-2} - e_{(g+1)^{2}-1}, \ h - e_{(g+1)^{2}-(g+1)} - \cdots - e_{(g+1)^{2}-3} + e_{(g+1)^{2}-2} + e_{(g+1)^{2}-1} \}$ and $\{ {e_1}' - {e_2}', \ {e_2}' - {e_3}', \ \cdots, \ e_{(g+1)^{2}-2}' - e_{(g+1)^{2}-1}', \ h' - e_{(g+1)^{2}-(g+1)} - \cdots - e_{(g+1)^{2}-3}' - e_{(g+1)^{2}-2}' - e_{(g+1)^{2}-1}'\}$), realize two copies of $M(g)$. One copy of hyperbolic pair $H$ comes from an identification of the fibers $f$ and $f'$, and a sphere section $\sigma$ of self-intersection $-2$ obtained by sewing the sphere sections $e_{(g+1)^{2}}$ and ${e_{(g+1)^2}}'$. The remaining $g(g-1)$ copies of $H$ come from $g(g-1)$ rim tori and their dual $-2$ spheres (see related discussion in \cite{GS}, page 73)). These $4g^2 + 2g + 2$ classes generate $H_2$ of $W(g)$. Furthermore, using the formula for the canonical class of the generalized symplectic sum and the adjunction inequality, we compute $K_{W(g)} = (g-2)(h + h')$. Also, the class of the genus $g$ surface of square $2$ of the genus $g$ Lefschetz pencil on $W(g)$ is given by $h + h'$. As a consequence, the class of the genus $g$ fiber in $W(g)\#2\CPb$ is given by $h + h' - E_{1} - E_{2}$, where $E_{1}$ and $E_{2}$ are the homology classes of the exceptional spheres of the blow-ups at the points $p_1$ and $p_2$, the base points of the pencil. We can also verify the symplectic surface $\Sigma$, given by the class $h + h' - E_{1} - E_{2}$, has genus $g$ by applying the adjunction formula to $(W(g)\#2\CPb, \Sigma)$: $g(\Sigma) = 1 + 1/2(K_{W(g)\#2\CPb} \cdot [\Sigma] + [\Sigma]^2) = 1 + ((g-2)(h + h') + E_{1} + E_{2}) \cdot (h + h' - E_{1} - E_{2}) + (h + h' - E_{1} - E_{2})^2)/2 = 1 + (2(g-2) + 2)/2 = g$. We can notice from the intersection form of $W(g)$ that all rim tori can be chosen to have no intersections with the genus $g$ surface in the pencil given by the homology class $h + h'$. Thus, the genus $g$ fiber $\Sigma$ can be chosen to be disjoint from the rim tori that descend to $W(g)\#2\CPb$.
\end{exmp}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=.63]{plumb3.eps}
\caption{Plumbing tree for ${\Phi}_{g(g-1)/2}(1)$}
\label{fig:plumb}
\end{center}
\end{figure}
Let $k$ be any nonnegative integer, and $\mathbb{F}_{k}$ denote $k$-th Hirzebruch surface. Recall that $\mathbb{F}_{k}$ admits the structure of holomorphic $\CPS$ bundle over $\CPS$ with two disjoint holomorphic sections $\Delta_{+k}$ and $\Delta_{-k}$ with $\Delta_{\pm k} = \pm k$.
\begin{lem}\label{E3} The genus $g$ Lefschetz fibration on $Z(g)$ over $\mathbb{S}^2$ with the monodromy $(c_1c_2 \cdots c_{2g})^{2(2g+1)} = 1$ can be obtained as the $2$-fold cover of $\mathbb{F}_{2}$ branched over the disjoint union of a smooth curve $C$ in the linear system $|(2g+1)\Delta_{+2}|$ and $\Delta_{-2}$
\end{lem}
\begin{proof} The Lefschetz fibration on $Z(g) \rightarrow \CPS$ obtained by composing the branched cover map $Z(g) \rightarrow \mathbb{F}_{2}$ with the bundle map $\mathbb{F}_{2} \rightarrow \CPS$. A generic fiber is the double cover of a sphere fiber of $\mathbb{F}_{2}$ branched over $2g + 2$ points. The monodromy of this Lefschetz fibration can be derived from the braid monodromy of the branch curve $C \cup \Delta_{-2}$. The fibration admits a holomorphic sphere section $S$ with $S^2 = -1$, which is obtained by lifting $\Delta_{-2}$ to $Z(g)$.
\end{proof}
\subsection{Rational Blowdown} In this subsection, we review the rational blowdown surgery introduced by Fintushel-Stern \cite{FS1}. For details the reader is referred to \cite{FS1,P1}.
Let $p \geq 2$ and $C_p$ be the smooth $4$-manifold obtained by plumbing disk bundles over the $2$-sphere according to the following linear diagram
\begin{picture}(100,60)(-90,-25)
\put(-12,3){\makebox(200,20)[bl]{$-(p+2)$ \hspace{6pt}
$-2$ \hspace{96pt} $-2$}}
\put(4,-25){\makebox(200,20)[tl]{$u_{p-1}$ \hspace{25pt}
$u_{p-2}$ \hspace{86pt} $u_{1}$}}
\multiput(10,0)(40,0){2}{\line(1,0){40}}
\multiput(10,0)(40,0){2}{\circle*{3}}
\multiput(100,0)(5,0){4}{\makebox(0,0){$\cdots$}}
\put(125,0){\line(1,0){40}}
\put(165,0){\circle*{3}}
\end{picture}
\noindent where each vertex $u_{i}$ of the linear diagram represents a disk bundle over $2$-sphere with the given Euler number.
The boundary of $C_p$ is the lens space $L(p^2, 1 - p)$ which also bounds a rational ball $B_p$ with $\pi_1(B_p) = {\mathbb{Z}}_p$ and $\pi_1(\partial B_p) \rightarrow \pi_1(B_p)$ surjective. If $C_p$ is embedded in a $4$-manifold $X$ then the rational blowdown manifold $X_p$ is obtained by replacing $C_p$ with $B_p$, i.e., $X_p = (X \setminus C_p) \cup B_p$. If $X$ and $X \setminus C_p$ are simply connected, then so is $X_p$. The following lemma is easy to check, so we omit the proof.
\begin{lem}\label{thm:rb} $b_{2}^{+}(X_p) = {b_2}^{+}(X)$, $e(X_p) = e(X) - (p-1)$, $\sigma(X_p) = \sigma(X) + (p-1)$, ${c_1}^{2}(X_p) = {c_1}^2(X) + (p-1)$, and $\chi_{h}(X_p) = \chi_{h}(X)$.
\end{lem}
We now collect some theorems on rational blowdown for later use.
\begin{thm}\label{SW1} \cite{FS1, P1}. Suppose $X$ is a smooth 4-manifold with $b_{2}^{+}(X) > 1$ which contains a configuration $C_{p}$. If $L$ is a SW basic class of $X$ satisfying $L\cdot u_{i} = 0$ for any i with $1 \leq i \leq p-2$ and $L\cdot u_{p-1} = \pm p$, then $L$ induces a SW basic class $\bar L$ of $X_{p}$ such that $SW_{X_{p}}(\bar L) = SW_{X}(L)$.
\end{thm}
\begin{thm}\label{SW2} \cite{FS1, P1} If a simply connected smooth $4$-manifold $X$ contains a configuration $C_{p}$, then the SW-invariants of $X_{p}$ are completely determined by those of $X$. That is, for any characteristic line bundle $\bar{L}$ on $X_{p}$ with $SW_{X_{p}}(\bar{L}) \ne 0$, there exists a characteristic line bundle $L$ on $X$
such that $SW_{X}(L) = SW_{X_{p}}(\bar{L})$.
\end{thm}
\begin{thm}[\cite{EN},\cite{EG} $(p=2)$, \cite{EMVHM} $(p\geq 3)$]\label{EN}
Let $\varrho$, $\varrho^{\prime}$ be positive relators of $\Gamma_g$, and let $X_\varrho$, $X_{\varrho^{\prime}}$ be the corresponding Lefschetz fibrations over $\mathbb{S}^2$, respectively. Suppose that $\varrho^\prime$ is obtained by applying a daisy substitution of type $p$ to $\varrho$. Then, $X_{\varrho^\prime}$ is a rational blowdown of $X_\varrho$ along a configuration $C_{p}$. Therefore, we have
\begin{align*}
\sigma(X_\varrho^\prime)=\sigma(X_{\varrho})+(p-1), \ \ \ \mathrm{and} \ \ \ e(X_\varrho^\prime)=e(X_{\varrho})-(p-1).
\end{align*}
\end{thm}
\subsection{Knot Surgery} In this subsection, we briefly review the knot surgery operation, which gives rise to mutually non-diffeomorphic manifolds. For the details, the reader is referred to \cite{FS2}.
Let $X$ be a $4$-manifold with ${b_2}^{+}(X) > 1$ and contain a homologically essential torus $T$ of self-intersection $0$. Let $N(K)$ be a tubular neighborhood of $K$ in $\mathbb{S}^3$, and let $T \times D^2$ be a tubular neighborhood of $T$ in $X$. The knot surgery manifold $X_{K}$ is defined by $X_K = (X \setminus (T \times D^2)) \cup (\mathbb{S}^1 \times (\mathbb{S}^3 \setminus N(K))$ where two pieces are glued in a way that the homology class of $[pt \times \partial D^2]$ is identifed with $[pt \times \lambda]$ where $\lambda$ is the class of the longitude of knot $K$. Fintushel and Stern proved the theorem that shows Seiberg-Witten invariants of $X_{K}$ can be completely determined by the Seiberg-Witten invariant of $X$ and the Alexander polynomial of $K$ \cite{FS2}. Moreover, if $X$ and $X \setminus T$ are simply connected, then so is $X_K$.
\begin{thm}\label{thm:knotsurgery} Suppose that $\pi_{1}(X) = \pi_{1}(X \setminus T) = 1$ and $T$ lies in a cusp neighborhood in $X$. Then $X_{K}$ is homeomorphic to $X$ and Seiberg-Witten invariants of $X_{K}$ is $SW_{X_K} = SW_{X} \cdot \Delta_{K}(t^2)$, where $t = t_{T}$ (in the notation of \cite{FS2}) and $\Delta_{K}$ is the symmetrized Alexander polynomial of $K$. If the Alexander polynomial $\Delta_{K}(t)$ of knot $K$ is not monic then $X_K$ admits no symplectic structure. Moreover, if $X$ is symplectic and $K$ is a fibered knot, then $X_{K}$ admits a symplectic structure.
\end{thm}
\section{Lemmas}
In this section, we construct some relations by applying elementary transformations. These relations will be used to construct new relations obtained by daisy substitutions in Section~\ref{new words}.
Let $a_1,\ldots,a_k$ be a sequence of simple closed curves on an oriented surface such that $a_i$ and $a_j$ are disjoint if $|i-j|\geq 2$ and that $a_i$ and $a_{i+1}$ intersect at one point. For simplicity of notation, we write $a_i$, $_f(a_i)$ instead of $t_{a_i}$, $t_{f(a_i)}=ft_{a_i}f^{-1}$, respectively.
Moreover, write
\begin{align*}
&b_i = {}_{a_{i+1}}(a_i)& &\mathrm{and}& &\bar{b}_i = {}_{a_{i+1}^{-1}}(a_i).&
\end{align*}
Below we denote the arrangement using the conjugation (i.e. the cyclic permutation) and the arrangement using the relation (i) by $\xrightarrow[]{C}$ and $\xrightarrow[]{(\mathrm{i})}$, respectively.
We recall the following relation:
\begin{align*}
&a_{i+1} \cdot a_i \sim b_i \cdot a_{i+1},& &\mathrm{and}& &a_i \cdot a_{i+1} \sim a_{i+1} \cdot \bar{b}_i.&
\end{align*}
In particular, we note that
\begin{align*}
&a_i \cdot a_j \sim a_j \cdot a_i& &\mathrm{for} \ |i-j|>1.&
\end{align*}
By drawing the curves, it is easy to verify that for $m=1,\ldots,k-1$ and $i=m,\ldots,k-1$,
\begin{align*}
a_k a_{k-1} \cdots a_{m+1} a_m (a_{i+1})=a_i \ \ \ \mathrm{and} \ \ \ a_m a_{m+1} \cdots a_{k-1} a_k (a_i)=a_{i+1}.
\end{align*}
Using the relation $t_{f(c)}=ft_cf^{-1}$, we obtain the followings:
\begin{align}
&(a_k a_{k-1} \cdots a_{m+1} a_m) \cdot a_{i+1} \sim a_i \cdot (a_k a_{k-1} \cdots a_{m+1} a_m), \label{1} \\
&(a_m a_{m+1} \cdots a_{k-1} a_k) \cdot a_i \sim a_{i+1} \cdot (a_m a_{m+1} \cdots a_{k-1} a_k). \label{2}
\end{align}
\begin{lem}\label{lem3.3}
For $2\leq k$, we have the following relations:
\begin{align*}
\mathrm{(a)}\hspace{10pt}&(a_{k-1} a_{k-2} \cdots a_2 a_1) \cdot (a_k a_{k-1} \cdots a_2 a_1) \sim a_k^k \cdot \bar{b}_{k-1} \cdots \bar{b}_2 \bar{b}_1, \\[3pt]
\mathrm{(b)}\hspace{10pt}&(a_1 a_2 \cdots a_{k-1} a_k) \cdot (a_1 a_2 \cdots a_{k-2} a_{k-1}) \sim b_1 b_2 \cdots b_{k-1} \cdot a_k^k, \\[3pt]
\end{align*}
\end{lem}
\begin{proof}
The proof will be given by induction on $k$. Suppose that $k=2$. Then, we have
\begin{align*}
\red{a_1} a_2 a_1 \xrightarrow[]{(\ref{1})} a_2 a_1 \red{a_2} \sim a_2^2 \cdot \bar{b}_1.
\end{align*}
Hence, the conclusion of the Lemma holds for $k=2$.
Let us assume inductively that the relation holds for $k=j$.
Then,
\begin{align*}
&(\red{a_j} a_{j-1} \cdots a_1) \cdot (\red{a_{j+1}} a_j \cdots a_1) \\
&\sim \red{a_j a_{j+1}} \cdot (a_{j-1} \cdots a_1) \cdot (a_j \cdots a_1) \\
&\sim a_j a_{j+1} \cdot \blue{a_j^{j+1}} \cdot \bar{b}_{j-1} \cdots \bar{b}_1 \\
&\xrightarrow[]{(\ref{2})} \blue{a_{j+1}^{j+1}} \cdot a_j a_{j+1} \cdot \bar{b}_{j-1} \cdots \bar{b}_1 \\
&\sim a_{j+1}^{j+1} \cdot a_{j+1} \cdot \bar{b}_j \cdot \bar{b}_{j-1} \cdots \bar{b}_1.
\end{align*}
This proves part (a). The proof of (b) is similar, and therefore omitted.
\end{proof}
\begin{lem}\label{lem3.4}
Let $l\geq 0$. We define an element $\phi$ to be
\begin{align*}
&\phi_{l}=a_{2l+1}^{l+1} a_{2l-1}^{l} \cdots a_5^3 a_3^2 a_1.
\end{align*}
Let $D$ and $E$ be two products of right-handed Dehn twists and write them as $D=d_1\cdots d_{k_1}$ and $E=e_1\cdots e_{k_2}$, respectively.
If a word $W_1$ is obtained by applying a sequence of the conjugation and the elementary transformations to a word $W_2$, then we denote it by $\sim_C$.
For $l\geq 1$, we have the following:
\begin{align*}
\mathrm{(a)}\hspace{5pt}&D \cdot a_{2l} \cdots a_2 a_1 \cdot a_{2l+1} \cdots a_2 a_1 \cdot E \\
& \sim_C {}_{\phi_l}(D) \cdot (a_{2l+1}^{l+1} \cdot a_{2l} \cdots a_2 a_1) \cdot (b_{2l} \cdots b_4 \cdot b_2) \cdot {}_{\phi_l}(E), \\[3pt]
\mathrm{(b)}\hspace{5pt}&D \cdot a_1 a_2 \cdots a_{2l+1} \cdot a_1 a_2 \cdots a_{2l} \cdot E \\
& \sim_C {}_{\phi_l^{-1}}(D) \cdot (\bar{b}_2 \cdot \bar{b}_4 \cdots \bar{b}_{2l}) \cdot (a_1 a_2 \cdots a_{2l} \cdot a_{2l+1}^{l+1}) \cdot {}_{\phi_l^{-1}}(E),
\end{align*}
where ${}_{\phi_l}(D)={}_{\phi_l}(d_1)\cdots {}_{\phi_{l}}(d_{k_1})$ and ${}_{\phi_l}(E)={}_{\phi_l}(e_1)\cdots {}_{\phi_{l}}(e_{k_2})$.
\end{lem}
\begin{proof}
For $1\leq m\leq 2l-1$ and $m\leq i\leq 2l-1$, we have the following equalities from (\ref{1}) and (\ref{2}):
\begin{align}
&(a_{2l} \cdots a_1 \cdot a_{2l+1} \cdots a_m) \cdot a_{i+2} \sim a_i \cdot (a_{2l} \cdots a_1 \cdot a_{2l+1} \cdots a_m) \label{3}\\
&a_{i+2} \cdot (a_m \cdots a_{2l+1} \cdot a_1 \cdots a_{2l}) \sim (a_m \cdots a_{2l+1} \cdot a_1 \cdots a_{2l}) \cdot a_i. \label{4}
\end{align}
We first show (a). Since
\begin{align*}
&{}_{\phi_{l-1}}(D) \cdot (a_{2l} a_{2l-1} \cdots a_1) \cdot (b_{2l} b_{2l-2} \cdots b_2) \cdot \red{a_{2l+1}^{l+1}} \cdot {}_{\phi_{l-1}}(E) \\
&\xrightarrow[]{C} {}_{\red{a_{2l+1}^{l+1}}\phi_{l-1}}(D)\cdot \red{a_{2l+1}^{l+1}} \cdot (a_{2l} a_{2l-1} \cdots a_1) \cdot (b_{2l} b_{2l-2} \cdots b_2) \cdot {}_{\red{a_{2l+1}^{l+1}}\phi_{l-1}}(E) \\
&= {}_{\phi_l}(D)\cdot a_{2l+1}^{l+1} \cdot (a_{2l} a_{2l-1} \cdots a_1) \cdot (b_{2l} b_{2l-2} \cdots b_2) \cdot {}_{\phi_l}(E),
\end{align*}
it is sufficient to prove
\begin{align*}
&D \cdot (a_{2l} a_{2l-1} \cdots a_1) \cdot (a_{2l+1} a_{2l} \cdots a_1) \cdot E \\
&\sim {}_{\phi_{l-1}}(D) \cdot (a_{2l} a_{2l-1} \cdots a_1) \cdot (b_{2l} b_{2l-2} \cdots b_2)\cdot a_{2l+1}^{l+1} \cdot {}_{\phi_{l-1}}(E).
\end{align*}
The proof is by induction on $l$. Suppose that $l=0$. Then, we have
\begin{align*}
&D \cdot a_2 a_1 \cdot a_3 a_2 \red{a_1} \cdot E \\
&\xrightarrow[]{C} {}_{\red{a_1}}(D) \cdot \red{a_1} \cdot a_2 a_1 \cdot a_3 a_2 \cdot {}_{\red{a_1}}(E) \\
&\xrightarrow[]{(\ref{3})} {}_{a_1}(D) \cdot a_2 a_1 \cdot a_3 a_2 \cdot \red{a_3} \cdot {}_{a_1}(E) \\
&\sim {}_{a_1}(D) \cdot a_2 a_1 \cdot b_2 \cdot a_3 \cdot a_3 \cdot {}_{a_1}(E).
\end{align*}
Since $\phi_0=a_1$, the conclusion of the Lemma holds for $l=0$.
Let us assume, inductively, that the relation holds for $l=j$.
Note that since $\phi_{j-1}(a_{2j+2})=a_{2j+2}$ and $\phi_{j-1}(a_{2j+3})=a_{2j+3}$, we have
\begin{align*}
{}_{\phi_{j-1}}(D a_{2j+2} a_{2j+1} a_{2j+3} a_{2j+2})={}_{\phi_{j-1}}(D) a_{2j+2} a_{2j+1} a_{2j+3} a_{2j+2}.
\end{align*}
Since $a_{2j+3}$ is disjoint from $b_2,b_4,\ldots,b_{2j}$ and $\phi_j=a_{2j+1}^{j+1}\phi_{j-1}$, we have
\begin{align*}
&D \cdot (\red{a_{2j+2} a_{2j+1}} a_{2j} \cdots a_1) \cdot (\red{a_{2j+3} a_{2j+2}} a_{2j+1} \cdots a_1) \cdot E \\
&\sim D \cdot \red{a_{2j+2} a_{2j+1} a_{2j+3} a_{2j+2}} \cdot (a_{2j} \cdots a_1) \cdot (a_{2j+1} \cdots a_1) \cdot E \\
&\sim {}_{\phi_{j-1}}(D) \cdot a_{2j+2} a_{2j+1} a_{2j+3} a_{2j+2} \cdot (a_{2j} \cdots a_1) \cdot (b_{2j} \cdots b_2) \cdot a_{2j+1}^{j+1} \cdot {}_{\phi_{j-1}}(E) \\
&\xrightarrow[]{C} {}_{\phi_j}(D) \cdot a_{2j+1}^{j+1} \cdot \red{a_{2j+2} a_{2j+1} a_{2j+3} a_{2j+2}} \cdot (a_{2j} \cdots a_1) \cdot (b_{2j} \cdots b_2) \cdot {}_{\phi_j}(E) \\
&\sim {}_{\phi_j}(D) \cdot a_{2j+1}^{j+1} \cdot (\red{a_{2j+2} a_{2j+1}} a_{2j} \cdots a_1) \cdot \red{a_{2j+3} a_{2j+2}} \cdot (b_{2j} \cdots b_2) \cdot {}_{\phi_j}(E) \\
&\xrightarrow[]{(\ref{3})} {}_{\phi_j}(D) \cdot (a_{2j+2} a_{2j+1} a_{2j} \cdots a_1) \cdot a_{2j+3} a_{2j+2} \cdot a_{2j+3}^{j+1} \cdot (b_{2j} \cdots b_2) \cdot {}_{\phi_j}(E) \\
&\sim {}_{\phi_j}(D) \cdot (a_{2j+2} a_{2j+1} a_{2j} \cdots a_1) \cdot b_{2j+2} \cdot a_{2j+3} \cdot a_{2j+3}^{j+1} \cdot (b_{2j} \cdots b_2) \cdot {}_{\phi_j}(E) \\
&\sim {}_{\phi_j}(D) \cdot (a_{2j+2} a_{2j+1} a_{2j} \cdots a_1) \cdot (b_{2j+2} b_{2j} \cdots b_2)\cdot a_{2j+3}^{j+2} \cdot {}_{\phi_j}(E).
\end{align*}
This proves part (a) of the lemma. The proof of part (b) is similar and left to the reader.
\end{proof}
\section{A Lift of hyperelliptic relations}
In this section, we construct a relation which gives a lift of a relation, and which is Hurwitz equivalent to $I(g)$, from $\Gamma_g$ to $\Gamma_g^1$. This relation will be used in Section~\ref{new words}.
Suppose $g\geq 2$.
Let $\Sigma_g^n$ be the surface of genus $g$ with $b$ boundary components $\delta_1,\delta_2,\ldots,\delta_n$.
Let $\alpha_1,\alpha_2,\ldots,\alpha_{2g},\alpha_{2g+1}, \alpha_{2g+1}^\prime$ and $\delta, \zeta_1,\ldots,\zeta_n$ be the simple closed curves as shown in Figure~\ref{curves}.
\begin{figure}[hbt]
\centering
\includegraphics[width=7cm]{curves.eps}
\caption{The curves $\alpha_1,\ldots,\alpha_{2g+1},\alpha_{2g+1}^\prime, \delta, \zeta_1,\ldots,\zeta_n$ on $\Sigma_g^n$ and the boundary components $\delta_1,\ldots,\delta_n$ of $\Sigma_g^n$.}
\label{curves}
\end{figure}
\begin{lem}\label{lem1}
The following relation holds in $\Gamma_g^n$:
\begin{align*}
(\alpha_1 \alpha_2 \cdots \alpha_{2g})^{2g+1} = (\alpha_1 \alpha_2 \cdots \alpha_{2g-1})^{2g} \cdot \alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g}.
\end{align*}
\end{lem}
\begin{proof}
The proof follows from the braid relations, $\alpha_i \cdot \alpha_{i+1} \cdot \alpha_i = \alpha_{i+1} \cdot \alpha_i \cdot \alpha_{i+1}$ and $\alpha_i \cdot \alpha_j= \alpha_j \cdot \alpha_i$ for $|i-j|>1$.
\end{proof}
\begin{lem}\label{lem2}
The following relation holds in $\Gamma_g^n$:
\begin{align*}
\delta=(\alpha_{2g+1} \alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime)^2
\end{align*}
In particular, since $\delta_1=\delta$ for $n=1$, this relation is a lift of $H(g)$ from $\Gamma_g$ to $\Gamma_g^1$.
\end{lem}
\begin{proof}
A regular neighborhood of $\alpha_1\cup \alpha_2 \cup\cdots \cup a_{2g-1}$ is a subsurface of genus $g-1$ with two boundary components, $\alpha_{2g+1}$ and $\alpha_{2g+1}^\prime$.
Moreover, a regular neighborhood of $\alpha_1\cup \alpha_2 \cup\cdots \cup a_{2g}$ is a subsurface of genus $g$ with one boundary component $\delta$.
Then, it is well know that the following relations, callled the \textit{chain relations} hold:
\begin{align*}
&\alpha_{2g+1} \alpha_{2g+1}^\prime = (\alpha_1 \cdots \alpha_{2g-1})^{2g},
&\delta = (\alpha_1 \alpha_2 \alpha_3 \cdots \alpha_{2g})^{4g+2}.\\
\end{align*}
By applying the relations above and Lemma~\ref{lem1}, we obtain the following relation.
\begin{align*}
\delta=(\alpha_{2g+1} \alpha_{2g+1}^\prime \alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g})^2.
\end{align*}
Since $\alpha_{2g+1}^\prime$ is disjoint from $\alpha_{2g+1}$, by conjugation of $\alpha_{2g+1}^\prime$, we obtain the claim.
If $n=1$, then it is easily seen that this is a lift of $H(g)$ from $\Gamma_g$ to $\Gamma_g^1$.
This completes the proof.
\end{proof}
\begin{lem}\label{lem3}
The following relation holds in $\Gamma_g^n$.
\begin{align*}
&(\alpha_{2g+1} \alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime)^2 \\
& \ \sim (\alpha_{2g+1} \alpha_{2g} \cdots \alpha_2 \alpha_1)^2 \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime).
\end{align*}
\end{lem}
\begin{proof}
By drawing picture, we find that for each $i=1,2,\ldots,2g$,
\begin{align*}
&\alpha_{2g+1} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g+1} (\alpha_i) = \alpha_i, \\
&\alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g} (\alpha_{2g+1}^\prime) = \alpha_{2g+1}.
\end{align*}
These give the following relations.
\begin{align}
&\alpha_{2g+1} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g+1} \cdot \alpha_i
\sim \alpha_i \cdot \alpha_{2g+1} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g+1}, \label{5}\\
&\alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g} \cdot \alpha_{2g+1}^\prime
\sim \alpha_{2g+1} \cdot \alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g}. \label{6}
\end{align}
From these relations, we have
\begin{align*}
&\alpha_{2g+1}\alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g} \red{\alpha_{2g+1}^\prime} \alpha_{2g+1} \alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime \\
&\xrightarrow[]{(\ref{6})} \alpha_{2g+1} \cdot \red{\alpha_{2g+1}} \alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1} \cdot \blue{\alpha_{2g} \cdots \alpha_2 \alpha_1} \cdot \alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime. \\
&\xrightarrow[]{(\ref{5})} (\alpha_{2g+1}\blue{\alpha_{2g} \cdots \alpha_2 \alpha_1}) \cdot (\alpha_{2g+1} \alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime).
\end{align*}
This completes the proof.
\end{proof}
\section{New words in the mapping class group via daisy relation}\label{new words} We define $\phi$ in $\Gamma_g^n$ to be
\begin{align*}
&\phi=\alpha_{2g+1}^{g+1} \alpha_{2g-1}^{g} \cdots \alpha_5^3 \alpha_3^2 \alpha_1.
\end{align*}
Note that $\phi(\alpha_{2i-1}) = \alpha_{2i-1}$ for each $i=1,\ldots,g+1$ and $\phi(\alpha_{2g+1}^\prime)=\alpha_{2g+1}^\prime$.
For simplicity of notation, we write
\begin{align*}
&\beta_i={}_{\alpha_{i+1}}(\alpha_i),& &\bar{\beta}_i={}_{\alpha_{i+1}^{-1}}(\alpha_i),& &\gamma_{i+1}={}_{\alpha_i}(\alpha_{i+1}),& &\bar{\gamma}_{i+1}={}_{\alpha_i^{-1}}(\alpha_{i+1}).&
\end{align*}
We denote by $\varphi$, $d_i$, $\bar{d}_i$, $e_{i+1}$ and $\bar{e}_{i+1}$ the images of $\phi$, $\beta_i$, $\bar{\beta}_i$, $\gamma_{i+1}$ and $\bar{\gamma}_{i+1}$ under the map $\Gamma_g^n \to \Gamma_g$, that is,
\begin{align*}
&\varphi=c_{2g+1}^{g+1} c_{2g-1}^{g} \cdots c_5^3 c_3^2 c_1,
\end{align*}
and
\begin{align*}
&d_i={}_{c_{i+1}}(c_i),& &\bar{d}_i={}_{c_{i+1}^{-1}}(c_i),& &e_{i+1}={}_{c_i}(c_{i+1})& &\mathrm{and}& &\bar{e}_{i+1}={}_{c_i^{-1}}(c_{i+1})&
\end{align*}
Then, note that $\varphi(c_{2i-1}) = c_{2i-1}$ for each $i=1,\ldots,g+1$.
If a word $W_1$ is obtained by applying simultaneous conjugations by $\psi$ to a word $W_2$, then we denote it by $\xrightarrow[]{\psi}$.
Let $x_1,\ldots,x_g$, $x_1^\prime,\ldots,x_g^\prime$ and $y_1,\ldots,y_g$ be the simple closed curves on $\Sigma_g$ given in Figure~\ref{curves3}.
Moreover, we define $y_{g+1},\ldots,y_{2g-1}$ to be $x_2,\ldots,x_g$, respectively.
We take the following two daisy relators of type $g-1$ in $\Gamma_g$:
\begin{align*}
&D_{g-1} := c_1^{-1} c_3^{-1} \cdots c_{2g-1}^{-1} \cdot c_{2g+1}^{-(g-2)} \cdot x_1 x_2 \cdots x_g & &\mathrm{and}& \\
&D_{g-1}^\prime := c_1^{-1} c_3^{-1} \cdots c_{2g-1}^{-1} \cdot c_{2g+1}^{-(g-2)} \cdot x_1^\prime x_2^\prime \cdots x_g^\prime,& &&
\end{align*}
and the following daisy relator of type $2(g-1)$:
\begin{align*}
D_{2(g-1)} &:= c_{2g+1}^{-1} c_{2g-1}^{-1} \cdots c_5^{-1} c_3^{-1} \cdot c_3^{-1} c_5^{-1} \cdots c_{2g-1}^{-1} \cdot c_{2g+1}^{-2g+3} \cdot y_1 y_2 \cdots y_{2g-1}\\
&= c_3^{-2} c_5^{-2} \cdots c_{2g-1}^{-2} \cdot c_{2g+1}^{-2g+2} \cdot y_1 y_2 \cdots y_{2g-1}.
\end{align*}
\begin{figure}[hbt]
\centering
\includegraphics[width=7cm]{curves3.eps}
\caption{The curves $x_1,\ldots,x_{g}$,$x_1^\prime,\ldots,x_g^\prime$ and $y_1,\ldots,y_g$ on $\Sigma_g$.}
\label{curves3}
\end{figure}
Let $\chi_1,\ldots,\chi_g$ be the simple closed curves on $\Sigma_g^n$ given in Figure~\ref{curves2}.
We denote by $\mathcal{D}_{g-1}$ the following daisy relator of type $g-1$ in $\Gamma_g^n$:
\begin{align*}
\mathcal{D}_{g-1}= \alpha_1^{-1} \alpha_3^{-1} \cdots \alpha_{2g-1}^{-1} \cdot \alpha_{2g+1}^{-(g-2)} \cdot \chi_1 \chi_2 \cdots \chi_g.
\end{align*}
Since it is easily seen that $\alpha_i$ and $\chi_i$ are mapped to $c_i$ and $x_i$ under the map $\Gamma_g^n \to \Gamma_g$, we see that the image of this map of $\mathcal{D}_{g-1}$ is $D_{g-1}$.
\begin{figure}[hbt]
\centering
\includegraphics[width=7cm]{curves222.eps}
\caption{The curves $\chi_1,\ldots,\chi_g$ on $\Sigma_g^n$.}
\label{curves2}
\end{figure}
\begin{thm}\label{4.1}There is a positive relator
\begin{align*}
H(g,1)= {}_{\varphi^{-1}}(\bar{d}_{2g}) \cdots {}_{\varphi^{-1}}(\bar{d}_2) {}_{\varphi^{-1}}(\bar{d}_1) \bar{d}_2 \bar{d}_4 \cdots \bar{d}_{2g} e_2 e_4 \cdots e_{2g} x_1 x_2 \cdots x_g c_{2g+1}^{2g+6},
\end{align*}
which is obtained by applying once $D_{g-1}$-substitution to $H(g)$.
Moreover, the Lefschetz fibration $f_{H(g,1)}$ has $2g+6$ disjoint sections of self-intersection $-1$.
\end{thm}
\begin{proof}
Note that by Lemma~\ref{lem3.3} (a), we have
\begin{align}
(\alpha_{2g+1} \alpha_{2g}\cdots \alpha_1)^2 &= \alpha_{2g+1} \cdot (\alpha_{2g} \cdots \alpha_1) \cdot (\alpha_{2g+1} \alpha_{2g} \cdots \alpha_1) \label{7} \\
&\sim \alpha_{2g+1}^{2g+2}\bar{\beta}_{2g}\cdots \bar{\beta}_1 \notag
\end{align}
Moreover, one can check that
\begin{align}
\alpha_1 \alpha_2 \alpha_3 \cdots \alpha_{2g} \sim (\gamma_2 \gamma_4 \gamma_6 \cdots \gamma_{2g}) \cdot (\alpha_1 \alpha_3 \alpha_5 \cdots \alpha_{2g-1}). \label{8}
\end{align}
Therefore, by Lemma~\ref{lem3.4} (b), we have
\begin{align*}
&(\alpha_{2g+1} \alpha_{2g} \cdots \alpha_2 \alpha_1)^2 \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime) \\
&\xrightarrow[]{(\ref{7})} \alpha_{2g+1}^{2g+2} \bar{\beta}_{2g} \cdots \bar{\beta}_1 \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g}) \alpha_{2g+1}^\prime \\
&\sim_C {}_{\phi^{-1}}\{\alpha_{2g+1}^{2g+2} \cdot \bar{\beta}_{2g} \cdots \bar{\beta}_1\} \cdot (\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \cdot \alpha_{2g+1}^{g+1}) \cdot \alpha_{2g+1}^\prime \\
&= \alpha_{2g+1}^{2g+2} \cdot {}_{\phi^{-1}}(\bar{\beta}_{2g}) \cdots {}_{\phi^{-1}}(\bar{\beta}_1) \cdot (\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \cdot \alpha_{2g+1}^{g+1}) \cdot \alpha_{2g+1}^\prime \\
&\xrightarrow[]{C} {}_{\phi^{-1}}(\bar{\beta}_{2g}) \cdots {}_{\phi^{-1}}(\bar{\beta}_1) \cdot (\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g}) \cdot \alpha_{2g+1}^{3g+3} \cdot \alpha_{2g+1}^\prime \\
&\xrightarrow[]{(\ref{8})} {}_{\phi^{-1}}(\bar{\beta}_{2g}) \cdots {}_{\phi^{-1}}(\bar{\beta}_1) \cdot (\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot (\alpha_1 \alpha_3 \cdots \alpha_{2g-1}) \cdot \alpha_{2g+1}^{3g+3} \cdot \alpha_{2g+1}^\prime.
\end{align*}
Since $\delta$ is a central element of the group generated by $\alpha_1,\ldots,\alpha_{2g+1},\alpha_{2g+1}^\prime$, by Lemma~\ref{lem3}, the operation $\sim_C$ is Hurwitz equivalent (for example, see Lemma 6 in \cite{Auroux2}).
We have the following relation in $\Gamma_g^{2g+6}$ which is Hurwitz equivalent to the relation $\delta=(\alpha_{2g+1} \alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime)^2$:
\begin{align*}
\delta = {}_{\phi^{-1}}(\bar{\beta}_{2g}) \cdots {}_{\phi^{-1}}(\bar{\beta}_1) \cdot (\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot (\alpha_1 \alpha_3 \cdots \alpha_{2g-1}) \cdot \alpha_{2g+1}^{3g+3} \cdot \alpha_{2g+1}^\prime.
\end{align*}
By applying once $\mathcal{D}_{g-1}$-substitution to this relation, we have the following relation:
\begin{align*}
\delta &= {}_{\phi^{-1}}(\bar{\beta}_{2g}) \cdots {}_{\phi^{-1}}(\bar{\beta}_1) \cdot (\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot (\chi_1 \chi_2 \cdots \chi_g) \cdot \alpha_{2g+1}^{2g+5} \cdot \alpha_{2g+1}^\prime,
\end{align*}
Moreover, by $\alpha_{2g+1}^{2g+5} \cdot \alpha_{2g+1}^\prime \cdot \delta_1 \delta_2 \cdots \delta_{2g+6} = \alpha_{2g+1}^{2g+5} \cdot \delta_1 \delta_2 \cdots \delta_{2g+6} \cdot \alpha_{2g+1}^\prime$ and the daisy-relation $\alpha_{2g+1}^{2g+5} \cdot \delta_1 \delta_2 \cdots \delta_{2g+6} \cdot \alpha_{2g+1}^\prime = \zeta_1 \zeta_2 \cdots \zeta_{2g+6} \cdot \delta$, we obtain
\begin{align*}
&\delta \cdot \delta_1 \delta_2 \cdots \delta_{2g+6} \\
&= {}_{\phi^{-1}}(\bar{\beta}_{2g}) \cdots {}_{\phi^{-1}}(\bar{\beta}_1) \cdot (\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot (\chi_1 \chi_2 \cdots \chi_g) \cdot \\
&\alpha_{2g+1}^{2g+5} \cdot \alpha_{2g+1}^\prime \cdot \delta_1 \delta_2 \cdots \delta_{2g+6} \\
&= {}_{\phi^{-1}}(\bar{\beta}_{2g}) \cdots {}_{\phi^{-1}}(\bar{\beta}_1) \cdot (\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot (\chi_1 \chi_2 \cdots \chi_g) \cdot \\
&\zeta_1 \zeta_2 \cdots \zeta_{2g+6} \cdot \delta.
\end{align*}
Therefore, we have the following relation in $\Gamma_g^{2g+6}$:
\begin{align*}
&\delta_1 \cdot \delta_2 \cdots \delta_{2g+6} \\
&={}_{\phi^{-1}}(\bar{\beta}_{2g}) \cdots {}_{\phi^{-1}}(\bar{\beta}_1) \cdot (\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot (\chi_1 \chi_2 \cdots \chi_g) \cdot \zeta_1 \zeta_2 \cdots \zeta_{2g+6}
\end{align*}
It is easily seen that $\zeta_1,\ldots,\zeta_{2g+6}$ are mapped to $c_{2g+1}$ under the map $\Gamma_g^{2g+6} \to \Gamma_g$.
This completes the proof.
\end{proof}
\begin{thm}\label{4.2}
There is a positive relator
\begin{align*}
H(g,2)& = {}_{\varphi^{-2}}(\bar{e}_{2g}) \cdots {}_{\varphi^{-2}}(\bar{e}_4) {}_{\varphi^{-2}}(\bar{e}_2) {}_{\varphi^{-2}}(d_{2g}) \cdots {}_{\varphi^{-2}}(d_4) \cdot {}_{\varphi^{-2}}(d_2) \cdot \\
&\bar{d}_2 \bar{d}_4 \cdots \bar{d}_{2g} e_2 e_4 \cdots e_{2g} (x_1 x_2 \cdots x_g)^2 c_{2g+1}^8.
\end{align*}
which is obtained by applying twice $D_{g-1}$-substitutions to $H(g)$. Moreover, the Lefschetz fibration $f_{H(g,2)}$ has $8$ disjoint sections of self-intersection $-1$.
\end{thm}
\begin{proof}
Let $D$ be a product of Dehn twists. Then, we note that by Lemma~\ref{lem3.4} (a) we have
\begin{align}
&(\alpha_{2g+1} \alpha_{2g} \cdots \alpha_2 \alpha_1)^2 \cdot D = \alpha_{2g+1} \cdot (\alpha_{2g} \cdots \alpha_1) \cdot (\alpha_{2g+1} \alpha_{2g} \cdots \alpha_1) \cdot D \label{9} \\
&\sim_C \alpha_{2g+1}^{g+2} \cdot (\alpha_{2g} \cdots \alpha_2 \alpha_1) \cdot (\beta_{2g} \cdots \beta_4 \beta_2) \cdot {}_{\phi}(D). \notag
\end{align}
Moreover, we have
\begin{align}
\alpha_{2g}\cdots \alpha_3 \alpha_2 \alpha_1 \sim (\alpha_{2g-1} \cdots \alpha_5 \alpha_3 \alpha_1) \cdot (\bar{\gamma}_{2g} \cdots \bar{\gamma}_6 \bar{\gamma}_4 \bar{\gamma}_2). \label{10}
\end{align}
Then, by Lemma~\ref{lem3.4} (b) we have
\begin{align*}
&(\alpha_{2g+1} \alpha_{2g} \cdots \alpha_2 \alpha_1)^2 \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime) \\
&\xrightarrow[]{(\ref{9})} \alpha_{2g+1}^{g+2} \cdot (\alpha_{2g} \cdots \alpha_2 \alpha_1) \cdot (\beta_{2g} \cdots \beta_4 \beta_2) \cdot {}_{\phi}\{(\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime)\} \\
&\xrightarrow[]{(\ref{10})} \alpha_{2g+1}^{g+2} \cdot (\alpha_{2g-1} \cdots \alpha_3 \alpha_1) \cdot (\bar{\gamma}_{2g} \cdots \bar{\gamma}_4 \bar{\gamma}_2) \cdot (\beta_{2g} \cdots \beta_4 \beta_2) \cdot \\
&{}_{\phi}\{(\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime)\} \\
&\xrightarrow[]{\phi^{-1}} {}_{\phi^{-1}}\{\alpha_{2g+1}^{g+2} \cdot (\alpha_{2g-1} \cdots \alpha_3 \alpha_1) \cdot (\bar{\gamma}_{2g} \cdots \bar{\gamma}_4 \bar{\gamma}_2) \cdot (\beta_{2g} \cdots \beta_4 \beta_2)\} \cdot \\
&(\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime) \\
&\sim_C {}_{\phi^{-2}}\{\alpha_{2g+1}^{g+2} \cdot (\alpha_{2g-1} \cdots \alpha_3 \alpha_1) \cdot (\bar{\gamma}_{2g} \cdots \bar{\gamma}_4 \bar{\gamma}_2) \cdot (\beta_{2g} \cdots \beta_4 \beta_2)\} \cdot \\
&(\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \cdot \alpha_{2g+1}^{g+1}) \cdot \alpha_{2g+1}^\prime \\
&= \alpha_{2g+1}^{g+2} \cdot (\alpha_{2g-1} \cdots \alpha_3 \alpha_1) \cdot {}_{\phi^{-2}}\{(\bar{\gamma}_{2g} \cdots \bar{\gamma}_4 \bar{\gamma}_2) \cdot (\beta_{2g} \cdots \beta_4 \beta_2)\} \cdot \\
&(\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\alpha_1 \alpha_2 \cdots \alpha_{2g} \cdot \alpha_{2g+1}^{g+1}) \cdot \alpha_{2g+1}^\prime \\
&\xrightarrow[]{(\ref{8})} \alpha_{2g+1}^{g+2} \cdot (\alpha_{2g-1} \cdots \alpha_3 \alpha_1) \cdot {}_{\phi^{-2}}\{(\bar{\gamma}_{2g} \cdots \bar{\gamma}_4 \bar{\gamma}_2) \cdot (\beta_{2g} \cdots \beta_4 \beta_2)\} \cdot \\
&(\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot (\alpha_1 \alpha_3 \cdots \alpha_{2g-1}) \cdot \alpha_{2g+1}^{g+1} \cdot \alpha_{2g+1}^\prime \\
&\xrightarrow[]{C} {}_{\phi^{-2}}\{(\bar{\gamma}_{2g} \cdots \bar{\gamma}_4 \bar{\gamma}_2) \cdot (\beta_{2g} \cdots \beta_4 \beta_2)\} \cdot \\
&(\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot \alpha_1^2 \alpha_3^2 \cdots \alpha_{2g-1}^2 \cdot \alpha_{2g+1}^{2g+3} \cdot \alpha_{2g+1}^\prime \\
&=({}_{\phi^{-2}}(\bar{\gamma}_{2g}) \cdots {}_{\phi^{-2}}(\bar{\gamma}_4) \cdot \cdot {}_{\phi^{-2}}(\bar{\gamma}_2)) \cdot ({}_{\phi^{-2}}(\beta_{2g}) \cdots {}_{\phi^{-2}}(\beta_4) \cdot \cdot {}_{\phi^{-2}}(\beta_2)) \cdot \\
&(\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot \alpha_1^2 \alpha_3^2 \cdots \alpha_{2g-1}^2 \cdot \alpha_{2g+1}^{2g+3} \cdot \alpha_{2g+1}^\prime.
\end{align*}
Note that by Lemma~\ref{lem3}, the operation $\xrightarrow[]{\phi^{-1}}$ is also Hurwitz equivalent from Lemma 6 in \cite{Auroux2}.
We have the following relation in $\Gamma_g^8$ which is Hurwitz equivalent to the relation $\delta=(\alpha_{2g+1} \alpha_{2g} \cdots \alpha_2 \alpha_1 \alpha_1 \alpha_2 \cdots \alpha_{2g} \alpha_{2g+1}^\prime)^2$:
\begin{align}
&\delta = ({}_{\phi^{-2}}(\bar{\gamma}_{2g}) \cdots {}_{\phi^{-2}}(\bar{\gamma}_4) \cdot \cdot {}_{\phi^{-2}}(\bar{\gamma}_2)) \cdot ({}_{\phi^{-2}}(\beta_{2g}) \cdots {}_{\phi^{-2}}(\beta_4) \cdot \cdot {}_{\phi^{-2}}(\beta_2)) \cdot \label{11} \\
&(\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot \alpha_1^2 \alpha_3^2 \cdots \alpha_{2g-1}^2 \cdot \alpha_{2g+1}^{2g+3} \cdot \alpha_{2g+1}^\prime. \notag
\end{align}
By applying twice $\mathcal{D}_{g-1}$-substitutions to this relation, we have the following relation:
\begin{align*}
&\delta = ({}_{\phi^{-2}}(\bar{\gamma}_{2g}) \cdots {}_{\phi^{-2}}(\bar{\gamma}_4) \cdot \cdot {}_{\phi^{-2}}(\bar{\gamma}_2)) \cdot ({}_{\phi^{-2}}(\beta_{2g}) \cdots {}_{\phi^{-2}}(\beta_4) \cdot {}_{\phi^{-2}}(\beta_2)) \cdot \\
&(\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot (\chi_1 \chi_2 \cdots \chi_g)^2 \cdot \alpha_{2g+1}^7 \cdot \alpha_{2g+1}^\prime.
\end{align*}
Moreover, by $\alpha_{2g+1}^7 \cdot \alpha_{2g+1}^\prime \cdot \delta_1\delta_2\cdots \delta_8 = \alpha_{2g+1}^7 \cdot \delta_1 \delta_2 \cdots \delta_8 \cdot \alpha_{2g+1}^\prime$ and the daisy-relation $\alpha_{2g+1}^7 \cdot \delta_1 \delta_2 \cdots \delta_8 \cdot \alpha_{2g+1}^\prime = \zeta_1 \zeta_2 \cdots \zeta_8 \cdot \delta$,
we obtain
\begin{align*}
\delta \cdot \delta_1\delta_2\cdots \delta_8 &= ({}_{\phi^{-2}}(\bar{\gamma}_{2g}) \cdots {}_{\phi^{-2}}(\bar{\gamma}_4) \cdot \cdot {}_{\phi^{-2}}(\bar{\gamma}_2)) \cdot ({}_{\phi^{-2}}(\beta_{2g}) \cdots {}_{\phi^{-2}}(\beta_4) \cdot \cdot {}_{\phi^{-2}}(\beta_2)) \cdot \\
&(\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot (\chi_1 \chi_2 \cdots \chi_g)^2 \cdot \alpha_{2g+1}^7 \cdot \alpha_{2g+1}^\prime \cdot \delta_1\delta_2\cdots \delta_8 \\
&=({}_{\phi^{-2}}(\bar{\gamma}_{2g}) \cdots {}_{\phi^{-2}}(\bar{\gamma}_4) \cdot \cdot {}_{\phi^{-2}}(\bar{\gamma}_2)) \cdot ({}_{\phi^{-2}}(\beta_{2g}) \cdots {}_{\phi^{-2}}(\beta_4) \cdot \cdot {}_{\phi^{-2}}(\beta_2)) \cdot \\
&(\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot (\chi_1 \chi_2 \cdots \chi_g)^2 \cdot \zeta_1 \zeta_2 \cdots \zeta_8 \cdot \delta.
\end{align*}
Therefore, we have the following relation in $\Gamma_g^8$:
\begin{align*}
\delta_1\delta_2\cdots \delta_8 &=({}_{\phi^{-2}}(\bar{\gamma}_{2g}) \cdots {}_{\phi^{-2}}(\bar{\gamma}_4) \cdot \cdot {}_{\phi^{-2}}(\bar{\gamma}_2)) \cdot ({}_{\phi^{-2}}(\beta_{2g}) \cdots {}_{\phi^{-2}}(\beta_4) \cdot \cdot {}_{\phi^{-2}}(\beta_2)) \cdot \\
&(\bar{\beta}_2 \bar{\beta}_4 \cdots \bar{\beta}_{2g}) \cdot (\gamma_2 \gamma_4 \cdots \gamma_{2g}) \cdot (\chi_1 \chi_2 \cdots \chi_g)^2 \cdot \cdot \zeta_1 \zeta_2 \cdots \zeta_8.
\end{align*}
It is easily seen that $\alpha_i$ and $\chi_i$ are mapped to $c_i$ and $x_i$, respectively, and $\zeta_1,\ldots,\zeta_8$ are mapped to $c_{2g+1}$ under the map $\Gamma_g^8 \to \Gamma_g$.
This completes the proof.
\end{proof}
For $j=1,\ldots,2g$, we write
\begin{align*}
f_j={}_{(c_{j-1}c_{j+1})^{-1}}(\bar{d_j}),
\end{align*}
where $c_0=1$.
\begin{thm}\label{4.7}
Let $g\geq 3$. Then, the monodromy of the hypereliptic Lefschetz fibration given by the word $H(g) = 1$ can be conjugated to contain a daisy relations of type $2(g-1)$.
\end{thm}
\begin{proof}
Since $\alpha_i$ and $\delta$ are mapped to $c_i$ and $1$ under the map $\Gamma_g^n \to \Gamma_g$, respectively,
by the equation (\ref{11}) acuired in Theorem~\ref{4.2}, we obtain the following relator:
\begin{align*}
&1 = ({}_{\varphi^{-2}}(\bar{e}_{2g}) \cdots {}_{\varphi^{-2}}(\bar{e}_4) \cdot \cdot {}_{\varphi^{-2}}(\bar{e}_2)) \cdot ({}_{\varphi^{-2}}(d_{2g}) \cdots {}_{\varphi^{-2}}(d_4) \cdot {}_{\varphi^{-2}}(d_2)) \cdot \\
&(\bar{d}_2 \bar{d}_4 \cdots \bar{d}_{2g}) \cdot (e_2 e_4 \cdots e_{2g}) \cdot c_1^2 c_3^2 \cdots c_{2g-1}^2 \cdot c_{2g+1}^{2g+4}.
\end{align*}
This relator contains $D_{2(g-1)}$-relator.
This completes the proof.
\end{proof}
\begin{thm}\label{4.3} Let $g\geq 3$. Then, the monodromy of the Lefschetz fibration given by the word $I(g)= 1$ can be conjugated to contain
\begin{enumerate}[label={\upshape(\roman*)}]
\setlength{\itemsep}{5pt}
\item $g+1$ daisy relations of type $g-1$.
\item $(g+1)/2$ (resp. $g/2$) daisy relations of type $2(g-1)$ for odd (resp. even) $g$.
\end{enumerate}
\end{thm}
\begin{proof}
Let us first prove (i). Since $\varphi(c_{2i-1}) = c_{2i-1}$ for each $i=1,\ldots,g+1$, by Lemma~\ref{lem3.4} (b) we have
\begin{align}
&(c_1 c_2 \cdots c_{2g} c_{2g+1})^2 = (c_1 c_2 \cdots c_{2g} c_{2g+1}) \cdot (c_1 c_2 \cdots c_{2g}) \cdot c_{2g+1} \label{12} \\
&\sim_C (\bar{d}_2 \bar{d}_4 \cdots \bar{d}_{2g}) \cdot (c_1 c_2 \cdots c_{2g} \cdot c_{2g+1}^{g+1}) \cdot c_{2g+1}. \notag
\end{align}
Moreover, we have
\begin{align}
c_1 c_2 c_3 \cdots c_{2g} \sim (e_2 e_4 e_6 \cdots e_{2g}) \cdot (c_1 c_3 c_5 \cdots c_{2g-1}). \label{13}
\end{align}
Therefore, we have
\begin{align*}
(c_1 c_2 \cdots c_{2g+1})^{2g+2} &\xrightarrow[]{(\ref{12})} (\bar{d}_2 \bar{d}_4 \cdots \bar{d}_{2g} \cdot c_1 c_2 \cdots c_{2g} \cdot c_{2g+1}^{g+2})^{g+1} \\
&\xrightarrow[]{(\ref{13})} (\bar{d}_2 \bar{d}_4 \cdots \bar{d}_{2g} \cdot e_2 e_4 \cdots e_{2g} \cdot c_1 c_3 \cdots c_{2g-1} \cdot c_{2g+1}^{g+2})^{g+1}.
\end{align*}
Then, we see that we can apply once $D_{g-1}^\prime$-substitution and $g$ times $D_{g-1}$-substitutions to $(\bar{d}_2 \bar{d}_4 \cdots \bar{d}_{2g} \cdot e_2 e_4 \cdots e_{2g} \cdot c_1 c_3 \cdots c_{2g-1} c_{2g+1}^{g+2})^{g+1}$. The reason that we apply once $D_{g-1}^\prime$-substitution is to construct a minimal symplectic manifold $Y(g,k)$ in Theorem~\ref{theorem1}(see the proof of Theorem~\ref{theorem1} and Remark 36). This completes the proof of (i).
Next we prove (ii). By Lemma~\ref{lem3.3} (b) we have
\begin{align}
&(c_1 c_2 \cdots c_{2g} c_{2g+1})^2 = (c_1 c_2 \cdots c_{2g} c_{2g+1}) \cdot (c_1 c_2 \cdots c_{2g}) \cdot c_{2g+1} \label{14} \\
&\sim (d_1 d_2 \cdots d_{2g} \cdot c_{2g+1}^{2g+1}) \cdot c_{2g+1}. \notag
\end{align}
Moreover, we have
\begin{align}
(c_1 c_2 \cdots c_{2g+1})^2 \sim (c_1^2 c_3^2 \cdots c_{2g+1}^2) \cdot (f_2 f_4 \cdots f_{2g}) \cdot (\bar{d}_2 \bar{d}_4 \cdots \bar{d}_{2g}). \label{15}
\end{align}
From the above relations, we have
\begin{align*}
&(c_1 c_2 \cdots c_{2g} c_{2g+1})^4 \\
&\xrightarrow[]{(\ref{14})} (d_1 d_2 \cdots d_{2g} \cdot c_{2g+1}^{2g+2}) \cdot (c_1 c_2 \cdots c_{2g} c_{2g+1})^2 \\
&\xrightarrow[]{(\ref{15})} (d_1 d_2 \cdots d_{2g} \cdot c_{2g+1}^{2g+2}) \cdot (c_1^2 c_3^2 \cdots c_{2g+1}^2) \cdot (f_2 f_4 \cdots f_{2g}) \cdot (\bar{d}_2 \bar{d}_4 \cdots \bar{d}_{2g}) \\
&\sim (d_1 d_2 \cdots d_{2g}) \cdot (c_1^2 c_3^2 \cdots c_{2g-1}^2 c_{2g+1}^{2g+4}) \cdot (f_2 f_4 \cdots f_{2g}) \cdot (\bar{d}_2 \bar{d}_4 \cdots \bar{d}_{2g}).
\end{align*}
From this, we see that $(c_1 c_2 \cdots c_{2g} c_{2g+1})^4$ can be conjugated to contain a daisy relations of type $2(g-1)$.
Therefore,
\begin{align*}
&I(g) =
\left\{ \begin{array}{ll}
\displaystyle (c_1 c_2 \cdots c_{2g+1})^{4k} \cdot (c_1 c_2 \cdots c_{2g+1})^2 & \ \ (g=2k) \\
\displaystyle (c_1 c_2 \cdots c_{2g+1})^{4(k+1)} & \ \ (g=2k+1),
\end{array} \right.
\end{align*}
gives the proof of (ii).
\end{proof}
\begin{thm}\label{4.4} Let $g\geq 3$. Then, the monodromy of the Lefschetz fibration given by the word $G(g) = 1$ can be conjugated to contain $g$ daisy relations of type $2(g-1)$.
\end{thm}
\begin{proof}
By a similar argument to the proof of Theorem~\ref{4.3}, we have
\begin{align*}
&(c_1 \cdots c_{2g-1} c_{2g})^4 \\
&\sim (d_1 d_2 \cdots d_{2g-1}) \cdot (c_2^2 c_4^2 \cdots c_{2g-2}^2 c_{2g}^{2g+3}) \cdot (f_1 f_3 \cdots f_{2g-1}) \cdot (\bar{d}_1 \bar{d}_3 \cdots \bar{d}_{2g-1}).
\end{align*}
Let $h:=(a_1\cdots a_{2g} a_{2g+1})$.
Note that $h(a_i)=a_{i+1}$ for $i=1,\ldots,2g$.
Then, we have
\begin{align*}
&(c_1 \cdots c_{2g-1} c_{2g})^4 \\
&\sim (d_1 d_2 \cdots d_{2g-1}) \cdot (c_2^2 c_4^2 \cdots c_{2g-2}^2 c_{2g}^{2g+3}) \cdot (f_1 f_3 \cdots f_{2g-1}) \cdot (\bar{d}_1 \bar{d}_3 \cdots \bar{d}_{2g-1}) \\
&\xrightarrow[]{h} (d_2 d_3 \cdots d_{2g}) \cdot (c_3^2 c_5^2 \cdots c_{2g-1}^2 c_{2g+1}^{2g+3}) \cdot (f_2 f_3 \cdots f_{2g}) \cdot (\bar{d}_2 \bar{d}_4 \cdots \bar{d}_{2g}).
\end{align*}
From this, we see that $(c_1 c_2 \cdots c_{2g} c_{2g})^4$ can be conjugated to contain a daisy relations of type $2(g-1)$.
$G(g)=(c_1 \cdots c_{2g-1} c_{2g})^{4g}\cdot(c_1 \cdots c_{2g-1} c_{2g})^2$ gives the proof.
\end{proof}
\subsection{Non-hyperellipticity of our Lefschetz fibrations}
The purpose of this subsection is to prove that all the Lefschetz fibrations obtained in this paper via daisy substitutions are non-hyperelliptic. The proof will be obtained as a corollary of more general theorem given below
\begin{thm}\label{nonhyperelliptic} Let $g\geq 3$. Let $f_{\varrho_1}:X_{\varrho_1} \rightarrow \mathbb{S}^2$ be a genus-$g$ hyperelliptic Lefschetz fibration with only non-separating vanishing cycyles, and let $\varrho_1$
be a positive relator corresponding to $f$. Let $k_1$ and $k_2$ be non-negative integers such that $k_1+k_2>0$, and let $k$ be a positive integer.
Then we have the followings:
\begin{enumerate}[label={\upshape(\roman*)}]
\setlength{\itemsep}{5pt}
\item If we can obtain a positive relator, denoted by $\varrho_2$, by applying $k_1$ times $D_{g-1}$-substitutions and $k_2$ times $D_{g-1}^\prime$-substitutions to $\varrho$,
then the genus-$g$ Lefschetz fibration $f_{\varrho_2}:X_{\varrho_2} \rightarrow \mathbb{S}^2$ is non-hyperelliptic.
\item If we can obtain a positive relator, denoted by $\varrho_3$, by applying $k$ times $D_{2(g-1)}$-substitutions to $\varrho$,
then the genus-$g$ Lefschetz fibration $f_{\varrho_3}:X_{\varrho_3} \rightarrow \mathbb{S}^2$ is non-hyperelliptic.
\end{enumerate}
\end{thm}
\begin{cor}
All our Lefschetz fibrations are non-hyperelliptic.
\end{cor}
\begin{proof}[Proof of Theorem~\ref{nonhyperelliptic}]
Let $s_0$ be the number of non-separating vanishing cycles of $f_{\varrho_1}$.
Note that by Theorem~\ref{sign}, we have $\sigma(X_{\varrho_1})=-\dfrac{g+1}{2g+1}s_0$.
First, we assume that $f_{\varrho_2}$ is a hyperelliptic Lefschetz fibration. The relators $D_{g-1}$ and $D_{g-1}^\prime$ consist of only Dehn twists about non-separating simple closed curves $c_1,c_3,\ldots,c_{2g+1}$ and $x_1,x_1^\prime,\ldots,x_g,x_g^\prime$ as in Figure~\ref{curves3}. Therefore, we see that $\varrho_2$ consits of only right-handed Dehn twists about non-separating simple closed curves, so $f_{\varrho_2}$ has only non-separating vanishing cycles.
In particular, the number of non-separating vanishing cycles of $f_{\varrho_2}$ is $s_0-\{(g-1)-1\}(k_1+k_2)$. By Theorem~\ref{sign}, we have
\begin{align*}
\sigma(X_{\varrho_2})=-\frac{g+1}{2g+1}\{s_0-(g-2)(k_1+k_2)\}.
\end{align*}
On the other hand, since the relators $D_{g-1}$ and $D_{g-1}^\prime$ are daisy relators of type $g-1$,
by Theorem~\ref{EN}, we have
\begin{align*}
&\sigma(X_{\varrho_2}) = \sigma(X_{\varrho_1})+(g-2)(k_1+k_2) = -\frac{g+1}{2g+1}s_0+(g-2)(k_1+k_2).&
\end{align*}
We get a contradiction since the above equality does not hold for $g \geq 3$ and $k_1+k_2 > 0$.
Next, we assume that $f_{\varrho_3}$ is a hyperelliptic Lefschetz fibration. The relator $D_{2(g-1)}$ consits of Dehn twsts about non-separating simple closed curves $c_3,c_5,\ldots,c_{2g+1}$ and $y_2,y_3,\ldots,y_{2g-1}$ in Figure~\ref{curves3} and a Dehn twist about separating simple closed curve $y_1$ in Figure~\ref{curves3}. Note that $y_{g+1}=x_2, y_{g+2}=x_3, \ldots, y_{2g-1}=x_g$ and that $y_1$ separates $\Sigma_g$ into two surface, one of which has genus $1$. Therefore, $f_{\varrho_3}$ has $s_0-k\{2(g-1)\}$ non-separating vanishing cycles and $k$ separating vanishing cycles which are $y_1$.
By Theorem~\ref{sign}, we have
\begin{align*}
\sigma(X_{\varrho_3})&=-\frac{g+1}{2g+1}\{s_0-2k(g-1)\}+\left(\frac{4(g-1)}{2g+1}-1\right)k \\
&=-\frac{g+1}{2g+1}s_0+\frac{2g^2+2g-7}{2g+1}k
\end{align*}
On the other hand, since the relator $D_{2(g-1)}$ is a daisy relator of type $2(g-1)$,
by Theorem~\ref{EN}, we have
\begin{align*}
\sigma(X_{\varrho_3}) = \sigma(X_{\varrho_1})+(2g-3)k = -\frac{g+1}{2g+1}s_0+(2g-3)k.
\end{align*}
Since $g\geq 3$, we have
\begin{align*}
&\left(-\frac{g+1}{2g+1}s_0+(2g-3)k\right)-\left(-\frac{g+1}{2g+1}s_0+\frac{2g^2+2g-7}{2g+1}k\right) \\
&=\frac{2(g-1)(g-2)}{2g+1}k>0.
\end{align*}
This is a contradiction to the above equality.
\end{proof}
\section{Constructing exotic 4-manifolds}
The purpose of this section is to show that the symplectic $4$-manifolds obtained in Theorem \ref{4.3}, part (i), are irreducible. Moreover, by performing the knot surgery operation along a homologically essential torus on these symplectic $4$-manifolds, we obtain infinite families of mutually nondiffeomorphic irreducible smooth structures.
\begin{thm}\label{theorem1} Let $g \geq 3$ and $M$ be one of the following $4$-manifolds $(g^2 - g + 1)\CP\# (3g^{2} + 3g + 3 -(g-2)k)\CPb$ for $k = 2, \cdots, g+1$. There exists an irreducible symplectic\/ $4$-manifold $Y(g,k)$ homeomorphic but not diffeomorphic to\/ $M$ that can be obtained from the genus $g$ Lefschetz fibration on $Y(g)$ over $\mathbb{S}^2$ with the monodromy $(c_1c_2 \cdots c_{2g+1})^{2g+2} = 1$ in the mapping class group $\Gamma_g$ by applying $k$ daisy substitutions of type $g-1$.
\end{thm}
\begin{proof} Let $Y(g,k)$ denote the symplectic $4$-manifold obtained from $Y(g) = W(g)\#2\,\CPb$ by applying $k-1$ $D_{g-1}$-subsitutions and one $D_{g-1}^\prime$-substitution as in Theorem~\ref{4.3}. Applying Lemma~\ref{thm:rb}, we compute the topological invariants of $Y(g,k)$
\begin{eqnarray*}
e(Y(g,k)) &=& e(W(g)\#2\CPb) - k(g-2) = 2(2g^2 + g + 3) - k(g-2),\\
\sigma(Y(g,k)) &=& \sigma (W(g)\#2\CPb) + k(g-2) = -2(g+1)^2 + k(g-2).
\end{eqnarray*}
Using the factorization of the global monodromy in terms of right-handed Dehn twists of the genus $g$ Lefschetz on $Y(g,k)$ (see Theorem~\ref{4.3}, part (i)), it is easy to check that $\pi_1(Y(g,k))=1$.
Next, we show that $Y(g,k)$ is non-spin $4$-manifold. Let $c_{1}, \cdots, c_{2g+1}$ be the curves in Figure~\ref{fig:hyper}, and $x_{1}, \cdots, x_{g}, x'_{1}, \cdots, x'_{g}$ be the curves in Figure~\ref{curves3}. Note that $\bar{d}_i = {}_{c_{i+1}^{-1}}(c_i)$ and $e_{i+1} = {}_{c_{i}}(c_{i+1})$. The vanishing cycles of Lefschetz fibrations in Theorem~\ref{4.3} includes the curves $\bar{d}_{2i}$, $e_{2i}$ for $i = 1, \cdots, g$, and $x_j$, $x'_{j}$ for $j = 1, \cdots, g$, and $c_{2g+1}$. In $H_{1}(\Sigma_{g}; \mathbb{Z}_{2})$, we find that $\bar{d}_{2g} =c_{2g+1} + c_{2g}$, $e_{2g} = c_{2g-1} + c_{2g}$ and $x_{g} = c_{2g-1} + c_{2g+1}$.Therefore, we have $\bar{d}_{2g} + e_{2g} = x_{g}$. In the notation of Theorem~\ref{Spin1}, we have $l=2$ and $\bar{d}_{2g} \cdot e_{2g} = 0$. Therefore, $2+ \bar{d}_{2g} \cdot e_{2g} \equiv 0 (\textrm{mod}\ 2)$.
By Theorem~\ref{Spin2}, $Y(g,k)$ is non-spin, and thus have an odd intersection form. By Freedman's theorem (cf.\ \cite{freedman}), we see that $Y(g,k)$ is homeomorphic to $(g^2 - g + 1)\CP\# (3g^{2} + 3g + 3 -(g-2)k)\CPb$.
Next, using the fact that $W(g)$ is a minimal complex surface of general type with ${b_{2}}^{+} > 1$ and the blow up formula for the Seiberg-Witten function \cite{FS2}, we compute $SW_{W(g)\#2\,\CPb}$ $= SW_{W(g)} \cdot \prod_{j=1}^{2}(e^{E_{i}} + e^{-E_{i}}) = (e^{K_{W(g)}} + e^{-K_{W(g)}})(e^{E_{1}} + e^{-E_{1}})(e^{E_{2}} + e^{-E_{2}})$, where $E_{i}$ denote the exceptional class of the $i^{th}$ blow-up. By the above formula, the SW basic classes of $W(g)\#2\,\CPb$ are given by $\pm K_{W(g)} \pm E_{1} \pm E_{2}$, and the values of the Seiberg-Witten invariants on these classes are $\pm 1$. Notice that by the Corollary 8.6 in~\cite{FS1}, $Y(g,k)$ has Seiberg-Witten simple type. Furthermore, by applying Theorem~\ref{SW1} and Theorem~\ref{SW2}, we completely determine the Seiberg-Witten invariants of $Y(g,k)$ using the basic classes and invariants of $W(g)\#2\,\CPb$: Up to sign the symplectic manifold $Y(g,k)$ has only one basic classes which descends from the $\pm$ canonical class of $Y(g)$ (see a detailed explanation below). By Theorem ~\ref{SW2}, or Taubes theorem \cite{taubes} the value of the Seiberg-Witten function on these basic classes, $\pm K_{Y(g,k)}$, are ${ \pm 1}$.
In what follows, we spell out the details of the above discussion. By Theorem~\ref{SW1} and Theorem~\ref{SW2}, we can determine the Seiberg-Witten invariants of $Y(g,k)$ by computing the algebraic intersection number of the basic classes $\pm K_{W(g)} \pm E_{1} \pm E_{2}$ of $W(g)\#2\,\CPb$, with the classes of spheres of $k$ disjoint $C_{g-1}$ configurations in $Y(g)$. Notice that the leading spheres of the configurations $C_{g-1}$ are the components of the singular fibers of $Y(g)$. By looking the regions on the genus $g$ surface, where the rational blowdowns along $C_{g-1}$ are performed, and the location of the base points of the genus $g$ pencil, we compute the algebraic intersection numbers as follows: Let $S_1^j$ denote the homology class of $-(g+1)$ sphere of the $j$-th configurations $C_{g-1}$ and $S_2^j, \cdots, S_{g-2}^j$ are the homology classes of $-2$ spheres of $C_{g-1}$ in $W(g)\#2\,\CPb$, where $j = 1, 2$. These two rational blowdowns along $C_{g-1}$ chosen such that they correspond to one $D_{g-1}$-subsitutions and one $D_{g-1}^\prime$-substitution as in Theorem~\ref{4.3}, part (i).
We have $S_1^1 \cdot E_{1} = 1$, $S_1^1 \cdot E_{2} = 0$, $S_1^2 \cdot E_{2} = 1$, $S_1^2 \cdot E_{1} = 0$, $S_1^j \cdot K_{W{g}} = g-2$, and the canonical divisor does not intersect with $S_i^j$ for $2 \leq i \leq g-1$. Consequently, $S_1^j \cdot \pm (K_{W{g}} + E_{1} + E_{2}) = \pm (g-1)$ for $j = 1, 2$, and $S_1^j \cdot (\pm K_{W{g}} \pm E_{1} \mp E_{2}) \neq \pm (g-1)$ for one $j$. Observe that among the eight basic classes $\pm K_{W{g}} \pm E_{1} \pm E_{2}$, only $K_{W{g}} + E_{1} + E_{2}$ and $-(K_{W{g}} + E_{1} + E_{2})$ have algebraic intersection $\pm (g-1)$ with $-(g+1)$ spheres of $C_{g-1}$. Thus, Theorem~\ref{SW1} implies that these are only two basic classes that descend to $Y(g,2)$, and consequently to $Y(g,k)$ from $W(g)\#2\,\CPb$.
By invoking the connected sum theorem for Seiberg-Witten invariants, we see that $SW$ function is trivial for $(g^2 - g + 1)\CP\# (3g^{2} + 3g + 3 -(g-2)k)\CPb$. Since the Seiberg-Witten invariants are diffeomorphism invariants, $Y(g,k)$ is not diffeomorphic to $(g^2 - g + 1)\CP\# (3g^{2} + 3g + 3 -(g-2)k)\CPb$.
The minimality of $Y(g,k)$ is a consequence of the fact that $Y(g,k)$ has no two Seiberg-Witten basic classes $K$ and $K'$ such that $(K - K')^2 = -4$. Notice that $\pm K_{Y(g,k)}$ are only basic classes of $Y(g,k)$, and $(K_{Y(g,k)}^2 - (-K_{Y(g,k)}))^2$ = $4(K_{Y(g,k)}^{2}) \geq 0$. Thus, we conlude that $Y(g,k)$ is symplectically minimal. Furthermore, since symplectic minimality implies irreducibility for simply-connected 4-manifolds \cite{HK}, we deduce that $Y(g,k)$ is also smoothly irreducible.
\end{proof}
The analogus theorem for $g = 2$, using lantern substitution, was proved in \cite{AP}.
\begin{thm}\label{theorem3} There exist an infinite family of irreducible symplectic\/ and an infinite family of irreducible non-symplectic\/ pairwise non-diffeomorphic $4$-manifolds all homeomorphic to $Y(g,k)$.
\end{thm}
\begin{proof} $Y(g,k)$\/ contains $g(g-1)$ Lagrangian tori which are disjoint from the singular fibers of genus $g$ Lefschetz fibration on $Y(g,k)$\/. These tori descend from $W(g)$ (See Example~\ref{Ex}), and survive in $Y(g,k)$ after the rational blowdowns along $C_{g-1}$. These tori are Lagrangian, but we can perturb the symplectic form on $Y(g,k)$ so that one of these tori, say $T$ becomes symplectic. Moreover, $\pi_1(Y(g,k) \setminus T) = 1$, which follows from the Van Kampen's Theorem using the facts that $\pi_1(Y(g,k)) = 1$ and any rim torus has a dual $-2$ sphere (see Proposition 1.2 in \cite{Halic}, or Gompf \cite{gompf}, page 564). Hence, we have a symplectic torus $T$ in $Y(g,k)$\/ of self-intersection $0$ such that $\pi_1(Y(g,k) \setminus T) = 1$. By performing a knot surgery on $T$, inside $Y(g,k)$, we obtain an irreducible $4$-manifold ${Y(g,k)}_K$ that is homeomorphic to $Y(g,k)$. By varying our choice of the knot $K$, we can realize infinitely many pairwise non-diffeomorphic $4$-manifolds, either symplectic or nonsymplectic.\end{proof}
\begin{rmk} We obtain the analogous results as in Theorem~\ref{theorem1} for the genus $g$ Lefschetz fibrations obtained in Theorem~\ref{4.4}. The total space $Z(g)$ of the Lefschetz fibration with the monodromy $(c_1c_2 \cdots c_{2g+1})^{2(2g+1)} = 1$ in the mapping class group $\Gamma_g$ are complex surface of general type with ${b_{2}}^{+} > 1$ and the single blow-up of a minimal complex surface (see \cite{GS}, Section 8.4, p.320-22). The computation of the Seiberg-Witten invariants follows the same lines of argument as that of Theorems \ref{theorem1}. In fact, the SW computation is simpler, since $Z(g)$ only admits one pair of basic classes. Also, the results of our paper can be easily extendable to the Lefschetz fibrations with monodromies $(c_1c_2 \cdots c_{2g-1}c_{2g}{c_{2g+1}}^2c_{2g}c_{2g-1} \cdots c_2c_1)^{2n} = 1$, $(c_1c_2 \cdots c_{2g}c_{2g+1})^{(2g+2)n} = 1$, $(c_1c_2 \cdots c_{2g-1}c_{2g})^{2(2g+1)n} = 1$ for $n \geq 2$. Since the computations are lengthy, we will not present it here.
\end{rmk}
\begin{rmk} Note that all the Lefschetz fibrations constructed in our paper are non-spin. The fibrations obtained in Theorems~\ref{4.1} and \ref{4.2} admit a section of self-intersection $-1$. The fibrations in Theorem~\ref{4.7} and \ref{4.4} contain separating vanishing cycles (the curve $y_1$ in Figure \ref{curves3} is separating). Thus, the total spaces are non-spin. In general, if we apply the daisy substitution of type $2(g-1)$ to a positive relator in $\Gamma_g$, then the resulting Lefschetz fibration always contains separating vanishing cycles. The fibrations in Theorem~\ref{4.3} do not contain any separating vanishing cycles, but they are non-spin due to Stipsicz's criteria (see proof of Theorem~\ref{theorem1}).
\end{rmk}
\begin{rmk} It would be interesting to know if the analogue of Theorem~\ref{theorem1} holds for Lefschetz fibrations of Theorem~\ref{4.2}. Their corresponding monodromies are obtained by applying twice $D_{g-1}$-substitutions to $H(g)$. In the opposite direction, we can prove that the total spaces of the Lefschetz fibrations of Theorem~\ref{4.1}, whose monodromies obtained by applying one $D_{g-1}$-substitutions to $H(g)$, are blow-ups of the complex projective plane. Notice that by Theorem~\ref{4.1} they admit at least $2g+6$ sphere sections of self-intersection $-1$. By using the result of Y. Sato \cite{Sato} (see Theorem 1.2, page 194), we see that the total spaces of these Lefschetz fibrations are diffemorphic to $\CP\#(3g+5)\CPb$.
\end{rmk}
\begin{rmk}
In the proof of Theorem~\ref{4.3}, part (i), if we apply $k$ $D_{g-1}$-substitutions for $k=1,\ldots, g+1$ without applying an $D_{g-1}^\prime$-substitution, then the Lefschetz fibrations over $\mathbb{S}^2$ given by the resulting relations admits a section of self-intersection $-1$ (i.e. the total spaces of the fibrations are non-minimal).
\end{rmk}
\section*{Acknowledgments} The authors are grateful to the referee for valuable comments and suggestions. A. A. was partially supported by NSF grants FRG-1065955, DMS-1005741 and Sloan Fellowship. N. M. was partially supported by Grant-in-Aid for Young Scientists (B) (No. 13276356), Japan Society for the Promotion of Science.
|
\section{Introduction}\label{sec1}
The last two decades have witnessed a rapid development of quantum control
technologies, which have proved to be crucial in a large number of
quantum systems applications that require a high level of reliability, such as the
generation of on-demand quantum states, or the regulation of system performance
for quantum information processing
\cite{Wiseman09,Mabuchi05,Dong10,Altafini10,Zhang12}. Stability is central
to these quantum control systems. For example, quantum control tasks may
require stabilization of stochastic filtering process
\cite{mirra2007,Sayrin12,Handel05,Amini12}, of quantum oscillators in
optical systems \cite{Hamerly12,Serafini12}, or of a complex network
constructed via the coherent interconnection of quantum components
\cite{Yanagisawa03,James08,Nurdin09,ZhangJ12,Gough10}.
The traditional approach has been Schr\"{o}dinger picture Lyapunov techniques, that is, where
one defines a Lyapunov function as a positive function over the set of
states. This has lead to several important results on the stability of
quantum states in different control settings
\cite{BG07,mirra2007,wang2010,tico2013,Qi13}. In general, the main concern in
these control problems is the asymptotic behaviour of quantum systems in the
Schr\"{o}dinger picture.
In contrast, we wish to develop a Heisenberg picture Lyapunov approach which
exploits the fact that the Heisenberg picture more readily captures the physical dynamics,
and that this allows for a more direct extension of classical stability techniques.
We are interested primarily in open quantum systems and, in particular, the quantum stochastic calculus of Hudson and Parthasarathy \cite{HP1984}
gives the appropriate mathematical description. This moreover turns out to be
a very convenient set up for modeling open quantum coherent
networks. It has been shown in a number of references
\cite{Gough09,Zhang12} that evolution of such networks in the Heisenberg
picture can be associated with what is now sometimes referred to as the $(S,L,H)$ representation. (Here
$S$ is a scattering matrix and $L,H$ are coupling and Hamiltonian operators
of the system, respectively, and appear as coefficients in the quantum stochastic differential equation
for the unitary evolution process \cite{HP1984,Parth1992}. As we are interested in the average dynamics in
the vacuum state for the environment, we shall take $S=I$ for simplicity as only $L$ and $H$ appear in the
Lindbladian.) Ideally,
conclusions about stability properties of these operators should be made
according to their dynamics in the Heisenberg picture, particularly in the
form of conditioning on the infinitesimal generator of a Lyapunov operator
which will be introduced in Section \ref{sec3}. It is worth mentioning that the
Lyapunov operator is a generalization of the existing results about quantum
stability based on operator inequalities, which are closely related to the
dissipativity of the system \cite{JG10,PUJ1}. To fix ideas, suppose that the evolution of an observable $X\in \mathfrak{A}$ in the Heisenberg picture
is given by $:t \mapsto j_t (X)$ and that this may be described by a Langevin equation
\begin{equation*}
dj_t(X) = j_t( \mathcal{G} (X)) dt + \mathrm{Noise},
\end{equation*}
where $\mathcal{G}$ is a Lindblad generator and the ``Noise'' terms are martingale increments for the environment state.
Mirroring the approach to the classical stochastic stability
theory, the dynamics of the expectation of an operator can then be
established from the properties of the corresponding infinitesimal
generator of a Lyapunov operator $V$ and, for instance, an exponential stability criterion would be that
\begin{equation}\label{Gv}
\mathcal{G}(V)\leq -cV+dI,\quad c,d>0,
\end{equation}
see e.g., \cite{JG10,PUJ1,PUJ4a}. Here $I$ is the identity
operator, and we include a dissipation rate $d$. We point out a similarity
between (\ref{Gv}) and conditions that arise in the Lyapunov stability
theory of classical stochastic differential equations~\cite{Khas}.
Our aim in this paper is to further develop the Lyapunov stability theory
of quantum stochastic evolutions in the Heisenberg picture. First, we will
study an interplay between stability of states and stability of
operators. For example, we will employ the Lyapunov condition~(\ref{Gv}) to
infer the asymptotic behavior of quantum states from the corresponding
properties of
the operator $V$. In particular, we will present a stability analysis of
the invariant state of an open quantum system based on the Lyapunov method
and the quantum semigroup theory. In a sense, this contribution is in
parallel with the results in the classical stochastic stability theory,
such as the Foster-Lyapunov theory~\cite{MT-1993}, concerned with the
existence and stability of invariant probability measures of Markov
processes.
In addition to invariant states, we are also interested in invariant
sets of states or operators to which quantum evolutions converge. To
characterize the invariance property of the system, we develop a quantum
version of the LaSalle invariance principle in the Heisenberg
picture. Using this result, we are able to determine a limit set of state
trajectories, as well as explore the possibility of stabilizing two
non-commuting operators simultaneously. Similar conditions that employ
Lyapunov techniques have been developed for classical stochastic systems
\cite{Khas,Kushner67,Mao99}. The objective of our development is to pave the
way to the design of coherent feedback networks, as operators of a
quantum system form a non-commutative algebra. An alternative approach of
analyzing stability of state trajectories in the Schr\"{o}dinger picture is
often too difficult in this case.
The main results of this paper are formulated in the way such that the ground states of an operator $V$ or $W$ are stabilized, given that certain Lyapunov conditions are satisfied. This is directly relevant to a recent quantum information processing scheme \cite{Verstraete09}, where the task is either stabilizing the ground states of a Hamiltonian which encode the solution to a quantum computation problem, or robustly preparing entangled quantum states \cite{Verstraete09,Lin13} through engineering the dissipative property of the systems. The results obtained in this paper thus provide tools to design proper environmental couplings for these applications.
This paper is organized as follows: In Section \ref{sec2}, we briefly
review some facts about Markov dynamics of open quantum systems. In Section
\ref{sec3}, the definition of the Lyapunov operator is given. Then in
Section \ref{sec4}, we propose a Lyapunov condition to ensure the existence of an
invariant state. In Section \ref{sec5}, we prove the faithfulness and uniqueness of
an invariant state using certain non-degeneracy conditions imposed on
the diffusion coefficients of the quantum stochastic differential
equation. In Section \ref{sec6}, we derive the quantum invariance principle and
discuss its implications. In Section \ref{sec7}, we investigate the stability
within the invariant set and provide a sufficient condition to stabilize
the system to the ground state. The last section is devoted to conclusions.
\emph{Notations}. In this paper $\mathfrak{H}$ is a separable Hilbert space, and $\mathfrak{A}$
denotes the von Neumann algebra of operators acting on the Hilbert space $\mathfrak{H}$. The commutator of two operators $A$ and $B$ is written as
$[A,B]=AB-BA$, while $X^\dagger$ is the adjoint of an operator
$X$. $\mathfrak{M}^{'}=\{A\in \mathfrak{A} :[A,X]=0, X\in \mathfrak{ M} \}$ denotes the commutant of a von Neumann algebra
$\mathfrak{M}$ of operators. In particular we write $\mathbb{C}I$ for the trivial von Neumann algebra consisting of
multiples of identity. $\mathbb{C}$ is the set of complex numbers. A positive-semidefinite operator $X$ is indicated as $X\geq0$. We shall only work with normal states, and
typically write $\rho$ for the corresponding density operator so that the expectation of an
operator $X$ will be denoted as $\langle
X\rangle_{\rho}=\tr{\rho X}$, e.g. \cite{Sakurai1985}.
\section{Quantum Markov Systems}\label{sec2}
Consider the system defined on a Hilbert space $\mathfrak{H}_S$, and the
environment on a Fock space $\mathfrak{H}_B$ over $L^2 (\mathbb{R}_+ , dt ) $ corresponding to a single Boson field mode. The composite system can be
regarded as a single closed system whose dynamics are
characterized by a unitary evolution $U(t)$ on $\mathfrak{H}_S\otimes \mathfrak{H}_B$ obeying a quantum stochastic differential
equation \cite{HP1984,JG10}
\begin{equation*}
dU(t)=\{dB^\dagger L-L^{\dagger}dB-\frac{1}{2}L^\dagger Ldt-iHdt\}U(t).
\end{equation*}
Here $H$ is the Hamiltonian of the system, and $L$ describes the
coupling between the system and environment; $B(\cdot )$ and $B^\dagger (\cdot )$ are
the annihilation and creation process defined on $\mathfrak{H}_B$. In the Heisenberg picture an operator $X(t)=j_t(X)$ of the system evolves as $j_t(X)=U(t)^\dagger (X\otimes
I)U(t)$. Given the interaction Hamiltonian of the
combined system, the explicit dynamical equation for $X(t)$ can be
written as
\begin{equation}~\label{eq:dynone}
dj_t(X)=j_t(\mathcal{G}(X))dt+ j_t (\mathcal{B}(X))dW_1(t) + j_t (\mathcal{C}(X))dW_2(t).
\end{equation}
Here the following notation is used
\begin{eqnarray}~\label{eq:generator}
\mathcal{G}(X) &=& -i[X,H]+\mathfrak{L}(X),\\ \notag
\mathcal{B}(X) &=& \frac{1}{2}([X,L]+[L(t)^\dagger,X]),\\ \notag
\mathcal{C}(X) &=& \frac{i}{2}(-[X,L]+[L^\dagger,X]).
\end{eqnarray}
with
\begin{equation}~\label{eq:Lindblad}
\mathfrak{L} (X)= L^\dagger{X}L-\frac{1}{2}L^\dagger{L}X-\frac{1}{2}XL^\dagger{L}.
\end{equation}
We have written the noise increments in quadrature form, that is, in terms of
\begin{equation*}
dW_1(t)=dB(t)+dB^\dagger(t), \quad dW_2(t)=i(dB(t)^\dagger-dB(t)),
\end{equation*}
with all increments understood in the It\={o} sense.
Alternatively, we can characterize the average evolution of $X(t)$ by the semigroup $T_t$ acting as
\begin{equation*}
T_t(X)=\mathbb E_0(U(t)^\dagger(X\otimes I)U(t)).
\end{equation*}
$\mathbb
E_0$ is a conditional expectation on the given initial algebra $\mathfrak{A}$ and the
initial vacuum state $|0\rangle\langle0|$ of $\mathfrak{H}_B$. The infinitesimal generator of this Markov semigroup is then given
by $\mathcal{G}$. The dissipation functional of the semigroup is defined as \cite{Lind1976,Frigerio78}
\begin{equation}
\mathfrak{D}(X)=\mathcal{G}(X^\dagger X)-\mathcal{G}(X^\dagger)X-X^\dagger\mathcal{G}(X).
\label{dissip.functional}
\end{equation}
For a completely positive semigroup $T_t$, we have $\mathfrak{D}(X)\geq0,X\in\mathfrak{A}$. The dissipation functional characterizes the irreversible nature of the quantum Markov process, and consequently the system is dissipative if $\mathfrak{D}\neq0$.
The corresponding semigroup in the predual space of the trace
class operators (the state space) is
denoted as $T_{*t}(\rho)=\rho_t$. The support projection $P_\mathrm{sup}$ of a state $\rho$ is defined as the smallest projection (in the sense that $P_\mathrm{sup}\le P$) to which the state assigns probability 1.
\begin{Def}[\cite{fagnola2003quantum}]
A state $\rho_I$ is an invariant state, if it satisfies the condition
$T_{*t}(\rho_I)=\rho_I$ $\forall t\ge 0$. A state $\rho$ is faithful in $\mathfrak{A}$ if $\tr{\rho A}=0$ implies $A=0$ for any
positive operator $A\in\mathfrak{A}$. In other words, $\rho$ is faithful if the support
projection of $\rho$ is the identity operator in the space of bounded operators on the underlying Hilbert space.
\end{Def}
The faithfulness of an invariant state is essential to the analysis of the
asymptotic
behaviour of a quantum Markov system. For example, if the system
possesses a faithful invariant state, then its ergodic properties and the
problem of convergence to equilibrium can be studied for this system
\cite{Alberto1982}. Moreover, if this faithful invariant state is unique,
then it is the only equilibrium state of the system \cite{Fagnola2004}.
\section{Quantum Lyapunov Operators and Stability in the Heisenberg Picture}\label{sec3}
In classical theory, stability refers to the property that the trajectories of the dynamical systems will remain near an equilibrium point $x_e$ if the initial states $x_0$ are near $x_e$. A stronger notion is asymptotic stability which additionally requires that the trajectories that start near the equilibrium state $x_e$ will converge to $x_e$. In practice it
is often too complicated for nonlinear systems to solve the dynamical equations directly, and the main tool for proving stability of
these systems is Lyapunov theory \cite{Khas,Kushner67,Thygesen97asurvey} without finding the trajectories. Generally
speaking, if a given system possesses a Lyapunov function $V(x)$,
with certain conditions on $V(x)$ and its convective derivative $\dot V(x)$,
then the trajectories of the system state $x_t$ will be stable in some
sense. For example, any continuous scalar function $V:\mathbb{R}^n\rightarrow
\mathbb{R}$ having the property
\begin{equation*}
V(0)=0,\quad V(x)>0,\quad \dot{V}(x)<0,\quad x\in\mathbb{R}^n\backslash\{0\}
\end{equation*}
can be chosen as a Lyapunov function for the purpose of establishing
asymptotic stability of the zero equilibrium state of the
system. Alternatively, the Lyapunov function can be defined as a continuous
function $V(x)$ satisfying \cite{Khas,Mali09}
\begin{equation*}
a(|x|)\leq V(x)\leq b(|x|), \quad \dot{V}(x)\leq 0,
\end{equation*}
where $a(\cdot),b(\cdot)$ are strictly increasing functions with
$a(0)=b(0)=0$ and $a(\infty)=b(\infty)=\infty$. In both cases, if such a
$V$ exists, then any trajectory $x_t$ will converge to $0$.
As noted in the introduction, stability of quantum systems
may be considered within either the Schr\"{o}dinger or Heisenberg pictures. In this paper, our intention is
to develop the Heisenberg picture approach and tools for studying stability
of state trajectories $\rho_t$ within the underlying Schr\"{o}dinger
picture. To this end, we define the Lyapunov operator in the Heisenberg
picture.
\begin{Def}
A quantum Lyapunov operator $V$ is an observable (self-adjoint operator) on a Hilbert space $\mathfrak{H}$ for
which the following properties hold:
\begin{enumerate}
\item $V\in D(\mathcal G)$,
\item $V\geq 0$,
\item $\mathcal{G}(V)\leq 0$.
\end{enumerate}
\end{Def}
$D(\mathcal G)$ is the domain of the generator. Note that $\langle V\rangle_\rho\geq0$ for all states $\rho$. In the next two sections, we will show that the existence and stability of the invariant state of the system can be proved using a Lyapunov operator.
One advantage of the Heisenberg approach is that the stability of operators in the
Heisenberg picture may be studied, and not just stability of states. This is of
practical importance, especially within the framework of
quantum coherent networks. Although the stability of operators has been
studied using operator semigroup theory, to the best of our knowledge the
first approach to stabilization of quantum systems via Lyapunov methods is in \cite{JG10}.
We now introduce our concept of stability of operators in the Heisenberg picture.
\begin{Def}
Given a set of positive operators $\mathfrak{S}$, the system is $\mathfrak{S}$-stable if $\langle j_t(A) \rangle_{\rho}$ is bounded for each $A \in \mathfrak{S}$ and any initial state $\rho$.
\end{Def}
In our development of Lyapunov quantum stability, we will make use of the notion of quantum coercivity defined in terms of the spectral decomposition of a Lyapunov operator $V$.
\begin{Def}
Consider a positive operator $V$ with the spectral decomposition
$V=\sum_iv_iP_i$, $v_i$ being the eigenvalues of $V$. $V$ is coercive if
there exists a strictly increasing function $k(\cdot)$ with
$\lim_{i\to \infty}k(i)=\infty$ such that
$v_i\geq k(i), i>i_0$, for some $i_0$.
\end{Def}
Note that the definition of quantum coercivity is analogous to its
classical counterpart \cite{Renardy04,MT-1993}
\begin{equation*}
V(x)\geq c(|x|)
\end{equation*}
with $c(|x|)$ being a strictly increasing function that goes to infinity as
$|x|\rightarrow\infty$.
From Definition~$4$, it follows that if the system possesses a coercive
Lyapunov operator $V=\sum_iv_iP_i$, then the set of operators
$\mathfrak{S}=\{A\geq0:\tr{AP_i}\leq \epsilon k(i),\epsilon>0\}$ are bounded in
expectation, hence the system is $\mathfrak{S}$-stable. Also, in Section \ref{sec6} we will prove that when given a Lyapunov
operator $V$, the set $\mathfrak{S}$ of operators defined by $\mathfrak{S}=\{W\geq0:\mathcal{G} (V)\leq -W\}$ are bounded in expectation. Indeed, the expectation $\langle
W(t)\rangle_{\rho}$ will converge to zero according to quantum LaSalle
invariance principle. Hence, a conclusion about stability of the system can
be made.
As in the classical case, one may have a number of variations on the definition of a Lyapunov function,
depending on the context of stability property which one is interested
in. The definition of Lyapunov operator can be relaxed for
quantum stability analysis in different contexts. Therefore, we still call
$V$ a Lyapunov operator when the property $\mathcal{G}(V)\leq0$ is replaced
by a weaker condition (\ref{Gv}), as in the following example.
\begin{exam}\rm
Consider a quantum oscillator with the Hamiltonian given by
$H={\omega}a^\dagger{a}$, and the coupling operator $L=
\alpha{a}+\beta{a^\dagger}$, $a$ and $a^\dagger$ are annihilation and creation operators respectively, and they satisfy the commutation relation $[a,a^\dagger]=1$. Choose the candidate Lyapunov operator as the photon number operator $V=a^\dagger{a}$ which represents the energy of the system. By calculation we find $\mathcal{G}(V)=-(|\alpha|^2-|\beta|^2)V+|\beta|^2I$. If $|\alpha|^2>|\beta|^2$, $\mathcal{G}(V)$ satisfies the condition (\ref{Gv}) and $V$ becomes a Lyapunov operator in this problem. Furthermore, the system is $\mathfrak{S}$-stable with $\mathfrak{S}$ being the von Neumann algebra generated by $V$ since $\langle V(t)\rangle$ is bounded \cite{JG10}. If $|\alpha|^2<|\beta|^2$, $\langle V(t)\rangle$ is unbounded and the system is unstable in energy.
In the sequel, a Lyapunov operator $V$ for which condition (\ref{Gv}) holds is referred to as a quantum Lyapunov operator in the weak sense.
\end{exam}
\section{Quantum Tightness and the Existence of Invariant States}\label{sec4}
As a first step to study the stability of quantum states, we derive certain conditions to guarantee the existence of
invariant state.
First we present the definition of quantum tightness \cite{Meyer1995}.
\begin{Def}
A sequence $(\rho_n)_{n\geq 1}$ in the Banach space of trace-class
operators on a Hilbert space $\mathfrak{H}$ is tight if for every $\epsilon>0,$ there exists a
finite rank projection $P$ and $n_0>0$ such that
$\tr{\rho_nP}>1-\epsilon$ for all $n\geq n_0.$
\end{Def}
Obviously, trajectories of states corresponding to finite-dimensional
systems are tight. We will refer to the following lemma. \cite{Meyer1995}
\begin{lem}
A tight sequence $(\rho_n)_{n\geq 1}$ of quantum states admits a
subsequence converging to a quantum state.
\end{lem}
\begin{thm}[\cite{fagnola2003quantum}]
If the system possesses a tight family of quantum states,
$(\rho_t,t>0)$ then the system possesses at least one invariant state.
\end{thm}
\begin{proof} As $\rho_t$ is tight, any sequence of states
$\rho_{t_n}=\frac{1}{t_n}\int_0^{t_n}\rho_{t^{'}}dt^{'}$ is also tight and therefore has normalized
sequential limit points. These states are invariant because any sequential limit point of
$\frac{1}{t_n}\int_0^{t_n}\rho_{t^{'}}dt^{'}$ is invariant, according to
Proposition~$2.3$ in \cite{fagnola2003quantum}.
\end{proof}
Based on these properties, we can develop the condition on the tightness of
general quantum systems. Recall the inequality (\ref{Gv}) involving the (Lindblad)
generator $\mathcal{G}$ of a quantum Markov process
defined by Equation~\eqref{eq:generator}. Suppose $V$ is a Lyapunov operator in the weak sense, i.e., $\mathcal{G}(V)\leq -cV+dI, c>0$. By integrating (\ref{Gv})
we obtain the following inequality \cite{JG10}
\begin{equation*}
\langle V(t)\rangle\leq e^{-ct}\langle V(0)\rangle+\frac{d}{c},
\end{equation*}
which means $\langle V(t)\rangle\leq\lambda$ for any $t\geq0$ and some
positive $\lambda$. Next we will show that the condition (\ref{Gv}) not only
gives us the mean stability of $V$ but also implies tightness of the
corresponding collection of quantum states $\{\rho_t, t\ge 0\}$.
First, let us consider the following example.
\begin{exam}\rm
The photon number operator for a quantum oscillator can be written as
$V=\sum_0^{\infty}i|i\rangle\langle{i}|$ where $|i\rangle$ is the
photon number state. If $\langle V(t)\rangle\leq{c},c\geq0$, then we have
$\sum_{i=0}^{\infty}i\rho_t^{ii}\leq{c}$ for an
arbitrary sequence of states $\rho_{t}$; here
$\rho_t^{ii}=\tr{\rho_t|i\rangle\langle{i}|}$.
For an
arbitrary $\epsilon>0$, choose $m$
such that $m=\left[\frac{c}{\epsilon}\right]$, where $[x]$ is the nearest
integer to $x$ that is greater than $x$. Through
$m\sum_{i=m}^{\infty}\rho_t^{ii}\le\sum_{i=m}^{\infty}i\rho_t^{ii}\leq{c}$, we conclude $\sum_m^{\infty}\rho_t^{ii}\le\epsilon$ for any $t$. The
finite rank projection $P=\sum_{i=0}^{m}|i\rangle\langle{i}|$
then satisfies the condition $\tr{\rho_{t}P}>1-\epsilon$, which indicates
that the sequence $\rho_{t}$ is tight. Hence the corresponding state
trajectory of the quantum oscillator gives rise to an invariant state for
the oscillator.
\end{exam}
The example shows that under certain conditions, the stability of an
operator in the mean sense may
imply tightness of a corresponding state trajectory.
The inequality
$$m\sum_m^{\infty}\rho_t^{ii}\le\sum_m^{\infty}i\rho_t^{ii}\leq{c}$$
is essential in this example. In fact, the spectral property
of the above operator is the key element connecting tightness and
stability. We generalize this idea in the following theorem.
\begin{thm}
Suppose the evolution of a positive observable $V$ on a separable Hilbert
space $\mathfrak{H}$, with spectral decomposition as $V=\sum_{i=0}^{\infty}v_iP_i$, is stable
in the mean, that is, there exists a constant $c\ge0$ such that $\langle V(t)\rangle_\rho\leq{c}$ with $\rho$ as the initial state. If $V$ is coercive, then any sequence $\rho_t$ is tight which implies the existence of an invariant state.
\end{thm}
\begin{proof}
The proof is similar to the proof used in Example~$2$ to show
tightness. The condition that $\langle V(t)\rangle_\rho\leq{c}$ means
$\sum_{i=0}^{\infty}v_{i}\rho_t^{ii}\leq{c}$ for $t\geq0$. Here $\rho^{ii}=\tr{\rho P_i}$
denotes the projection $P_i$ on the state $\rho$. Since $V$ is coercive, there exists some
$N_0$ such that $v_i$ is increasing for $i\geq N_0$ and
$v_i\rightarrow\infty$ as $i\rightarrow\infty$. Choose
$m=\max\{N_0,\inf\{i:v_{i}\geq\frac{c}{\epsilon}\}\}$. Then
$v_{m}\sum_m^{\infty}\rho_t^{ii}\le\sum_m^{\infty}v_{i}\rho_t^{ii}\leq{c},$
so we find that $\sum_{i=m}^{\infty}\rho_t^{ii}\le\epsilon$. Letting
$P=\sum_{i=0}^{m-1}P_i$, we obtain $\tr{\rho_tP}>1-\epsilon$; i.e, $\rho_t$ is tight. The result of the theorem then follows from Theorem~1.
\end{proof}
It follows from Theorem~$2$ that the existence of a coercive Lyapunov operator in the weak sense (\ref{Gv}) guarantees the existence of an invariant state. This prompts the question as to under what condition such an invariant state is unique and/or faithful. This question is addressed in the next section.
\section{Stability of Invariant States}\label{sec5}
In this section, we obtain some conditions to guarantee the
faithfulness and uniqueness of an invariant state.
For a particular invariant state $\rho_I$, its support
projection is denoted as $P_I$.
We shall need the following proposition.
\begin{prop}[see e.g.,~\cite{fagnola2003quantum}]
The support projection of an invariant state is subharmonic. That is, $T_t(P_I)\geq P_I$.
\end{prop}
The above property of the support projection can be expressed
in terms of the generator $\mathcal{G}$ of the semigroup $T_t$ as
$\mathcal{G}(P_I)\geq0$.
\medskip
\subsection{Stability of invariant states of finite-dimensional systems}
For a finite-dimensional system with the underlying Hilbert space $\mathfrak{H} =\mathbb{C}^n$, the following
theorem determines faithfulness and uniqueness of an invariant
state.
\begin{Def}
A state $\rho$ is said to be globally attractive if all system trajectories asymptotically converge to $\rho$ for any initial state.
\end{Def}
\begin{thm}~\label{thm:three}
Suppose $\mathfrak{H}=\mathbb{C}^n$. If $PL^\dagger(I-P)LP\neq0$ for any non-trivial projection $P$, then the invariant state $\rho_I$ is faithful and unique.
\end{thm}
\begin{proof}
A finite dimensional system is tight by Definition~$5$ and therefore, according to Theorem 1, it admits an invariant
state $\rho_I$. Let $P_I$ be the support projection of $\rho_I$.
If we take any orthogonal projection $P$, then we have
\begin{equation*}
\mathcal{G}(P)=\mathcal{G}(P^2)=P\mathcal{G}(P)+ \mathcal{G}(P)P +\mathfrak{D} (P),
\end{equation*}
where $\mathfrak{D}$ is the dissipation functional defined in (\ref{dissip.functional}), and so
\begin{equation*}
P\mathcal{G}(P)P=-P\mathfrak{D}(P)P.
\end{equation*}
However we note that $\mathfrak{D}(X) = [X,L]^\dag [X,L] \geq 0$, and in particular,
\begin{equation*}
P \mathfrak{D}(P) P = P L^\dag (I-P) L P.
\end{equation*}
Now take the invariant state $\rho_I$ with support projection $P_I$, then from Proposition~$1$ we will have $ \mathcal{G} (P_I) \geq 0$,
and therefore $P_I \mathcal G(P_I) P_I \geq 0$. But we then must have $P_I \mathfrak{D} (P_I) P_I = 0$, as $\mathfrak{D} \geq 0$.
We thereby deduce that for the invariant state support
\begin{equation}~\label{eq:P_I}
P_I \mathfrak{D}(P_I) P_I = P_I L^\dag (I-P_I) L P_I = 0.
\end{equation}
This is automatically satisfied if $\rho_I$ is faithful, since here $I - P_I \equiv 0$.
Suppose the hypothesis of the theorem is true, namely that $ P L^\dag (I-P) L P\neq0$ for any non-trivial orthogonal projection $P$. If we also now suppose
that $\rho_I$ is not faithful, then $P_I$ is non-trivial, then setting $P=P_I$ in (\ref{eq:P_I}) leads to a contradiction. Therefore, under the hypothesis,
we see that any invariant state must be faithful.
Suppose the invariant state is not unique, then there exist non-trivial orthogonal invariant subspaces by \cite{Bau12,Schirmer10}. This leads to a contradiction since there will exist non-faithful invariant states in the subspace.
\end{proof}
\begin{rem}
Condition $PL^\dagger(I-P)LP\neq0$ means that any non-trivial projection $P$ is connected with its orthogonal complement by $L$. This property can be easily verified when the system has reduced dynamics. For example, if the quantum states maintain a diagonal form $\rho(t)=\sum_i\rho^{ii}(t)P_i,P_i=|i\rangle\langle i|$ during evolution, we only need to verify $P_iL^\dagger(I-P_i)LP_i\neq0$ for all $P_i$. To generalize, if there exists a family of projections $\{P_i\}$ such that $\sum_iP_i=I$ and $P_iL^\dagger(I-P_i)LP_i\neq0$, the marginal distribution of the invariant state will have non-vanishing probability on each projector $P_i$.
\end{rem}
\medskip
It is worth mentioning that for finite dimensional system, uniqueness of invariant state directly leads to global convergence \cite{Schirmer10}.
\begin{exam}\rm
Consider the quantum two-level system with a basis denoted as $\{|0\rangle,|1\rangle\}$. $H=\omega\sigma_z=\omega(|0\rangle\langle0|-|1\rangle\langle1|)$ and $L=\sigma_x=|0\rangle\langle1|+|1\rangle\langle0|$. The quantum state evolves according to the master equation
\begin{equation*}
\frac{d\rho(t)}{dt}=-i[H,\rho(t)]+L\rho(t)L^\dagger-\frac{1}{2}L^\dagger L\rho(t)-\frac{1}{2}\rho(t)L^\dagger L.
\end{equation*}
Obviously the density matrix of the state will remain diagonal if the
initial state is $\alpha|0\rangle\langle0|+\beta|1\rangle\langle1|$ with
arbitrary $\alpha$ and $\beta$ satisfying $|\alpha|^2+|\beta|^2=1$. As a
result, the system will possess a diagonal invariant state. We only need to consider the
projections $\{|1\rangle\langle1|,|0\rangle\langle0|\}$ in order to
conclude faithfulness of this invariant state. In fact, we have
$|1\rangle\langle1|\sigma_x|0\rangle\langle0|\sigma_x|1\rangle\langle1|=|1\rangle\langle1|\neq0$
and
$|0\rangle\langle0|\sigma_x|1\rangle\langle1|\sigma_x|0\rangle\langle0|=|0\rangle\langle0|\neq0$,
so the two-level system has a unique faithful invariant state which is globally
attractive.
\end{exam}
\medskip
\subsection{Stability of invariant states of infinite-dimensional systems}
Now we can prove the main result in this section for the quantum system defined on a separable Hilbert space $\mathfrak{H}$.
\begin{thm}\label{theorem4}
Suppose there exists a coercive Lyapunov operator in the weak sense (\ref{Gv}). If $PL^\dagger(I-P)LP\neq0$ for any non-trivial projection $P$, then any invariant state $\rho_I$ is faithful and unique. Furthermore, this faithful state $\rho_I$ is globally attractive.
\end{thm}
\begin{proof}
The proof follows along the same lines as the proof of
Theorem~$3$. However, the existence of invariant state comes from condition
(\ref{Gv}) and coercivity, and is a direct result of
Theorem~$2$. The Lyapunov operator inequality (\ref{Gv}) and the algebraic
condition $PL^\dagger(I-P)LP\neq0$ are then combined to guarantee the uniqueness and faithfulness of this invariant state which is also the
equilibrium point of the system. In addition, the unique invariant state is also globally attractive due to its faithfulness
\cite{Fagnola2004}.
\end{proof}
We also note a certain analogy between Theorem~4 and the corresponding
results from the classical theory of stochastic Markov processes; e.g.,
see~\cite{MT-1993}. In particular, our condition (\ref{Gv}) is analogous
to the positive recurrence condition (CD2) in~\cite{MT-1993}.
\medskip
\begin{exam}\rm
Consider again a quantum oscillator with the Hamiltonian $H={\omega}a^\dagger{a}$, and the coupling operator $L=
\alpha{a}+\beta{a^\dagger}$.
Consider the observable $V=a^\dagger{a}$ which has a strictly increasing and unbounded spectrum. $\mathcal{G}(V)=-(|\alpha|^2-|\beta|^2)V+|\beta|^2I$. In order to satisfy the Lyapunov
condition in Theorem~$4$, we need to set $|\alpha|>|\beta|$. In this case,
$\langle V(t)\rangle$ is bounded with respect to any initial state. Hence, according to
Theorem~2, this system admits an invariant state. Now we want to study the set of projections $\{P_i=|i\rangle\langle i|\}$. Note that since $|i\rangle$ is the photon number state, then $\sum_iP_i=I$. We have $|i\rangle\langle i|L^\dagger|i+1\rangle\langle i+1|L|i\rangle\langle i|=(i+1)|\beta|^2|i\rangle\langle i|\neq0$ and $|i\rangle\langle i|L^\dagger|i-1\rangle\langle i-1|L|i\rangle\langle i|=i|\alpha|^2|i\rangle\langle i|\neq0$. Therefore, any photon number state is connected to its two neighboring states, so by induction any non-trivial projection $P=\sum_jP_j\neq I$ is not a support projection of an invariant state $\rho_I$. Consequently, the photon-number distribution of the invariant state has non-vanishing probability on the
entire Fock basis. However, this does not imply that the invariant state is faithful because there may exist other set of projections that does not satisfy the algebraic condition of Theorem \ref{theorem4}.
\end{exam}
\section{Quantum LaSalle Invariance Principle}\label{sec6}
In the previous section, we studied the stability property of convergence
to faithful invariant
states. Other classes of stabilization problems of interest are concerned with
stability of non-commuting operators, or
require convergence to an invariant set for any state trajectories. Similar
to the classical LaSalle's invariance principle \cite{Lasalle68,Mao99} that
is used to identify the asymptotic stability of system trajectories, the
invariance theorems which we will derive here pave the way for analyzing
the underlying dynamics of general quantum states which may not be faithful in the Heisenberg picture.
The classical LaSalle theorem states the following fact \cite{Lasalle68}:
If a positive and uniformly continuous function $V(x)$ can be found on a
compact space such that $\dot V(x)\leq0$, then the limit points of any
trajectory $x_t$ are contained in the largest invariant subset of
$\{x:\dot{V}(x)=0\}$.
First we will derive the direct analogue of the classical LaSalle invariance theorem in the Heisenberg picture.
\begin{Def}
A quantum state $\rho$ is said to be the zero solution of an operator $X$ if $\rho$ solves $\langle X\rangle_\rho=0$.
\end{Def}
\begin{thm}
If there exists a coercive Lyapunov operator $V$ and a positive operator $W$ with $\mathcal{G}(W)$ bounded in the operator norm such that
\begin{equation}
\mathcal{G}(V)\leq-W,
\label{ineqlasa}
\end{equation}
then $\lim_{t\rightarrow\infty}\langle V(t)\rangle_{\rho_0}=\lim_{t\rightarrow\infty}\langle V\rangle_{\rho_t}$ exists for any initial state $\rho_0$ and
\begin{eqnarray}
&\int_0^\infty\langle W(t^{'})\rangle_{\rho_0}dt^{'}=\int_0^\infty\langle W\rangle_{\rho_{t^{'}}}dt^{'}<+\infty,&\nonumber\\
&\lim_{t\rightarrow\infty}\langle
W(t)\rangle_{\rho_0}=\lim_{t\rightarrow\infty}\langle W\rangle_{\rho_t}=0.&
\label{conlasa}
\end{eqnarray}
\end{thm}
\begin{proof}
Referring to Theorem 2, tightness of $\rho_t$ ensures the existence of a limit point (or an accumulation point) of the system evolutions. The function $\langle V(t)\rangle_{\rho_0}$ is decreasing with $t$ since $\mathcal{G}(V)\leq0$. Therefore, $\lim_{t\rightarrow\infty}\langle V(t)\rangle_{\rho_0}=\lim_{t\rightarrow\infty}\langle V\rangle_{\rho_t}$ exists because any decreasing sequence with a lower bound will converge to a limit.
Moreover, $\langle V(t)\rangle_{\rho_0}$ evolves according to
\begin{equation}
\langle V(t)\rangle_{\rho_0}-\langle
V\rangle_{\rho_0}=\int_0^t\langle\mathcal{G} (V(t^{'}))\rangle_{\rho_0}dt^{'}\leq\int_0^t\langle-W(t^{'})\rangle_{\rho_0}dt^{'}.
\label{integene1}
\end{equation}
It follows from (\ref{integene1}) that $\int_0^t\langle
W(t^{'})\rangle_{\rho_0}dt^{'}\leq\langle V\rangle_{\rho_0}$, which implies
\begin{equation*}
\int_0^\infty\langle W(t^{'})\rangle_{\rho_0}dt^{'}=\int_0^\infty\langle W\rangle_{\rho_{t^{'}}}dt^{'}<+\infty.
\end{equation*}
$\rho_t$ is tight, so the positive sequence $\langle W\rangle_{\rho_t}$
must have convergent subsequence. Suppose there exists a subsequence
$\rho_{t_k}$ such that $\lim_{k\rightarrow\infty}\langle
W\rangle_{\rho_{t_k}}=\epsilon>0$. Now we show that this leads to a
contradiction. Since $\mathcal{G}(W)$ is bounded in operator norm by $R$, we
have
\begin{equation*}
|\langle W(t_1)\rangle_{\rho_0}-\langle W(t_2)\rangle_{\rho_0}|=|\int_{t_2}^{t_1}\tr{\mathcal{G}(W)\rho_{t^{'}}}dt^{'}|\leq\int_{t_2}^{t_1}\|\mathcal{G}(W)\||\rho_{t^{'}}|dt^{'}\leq R|t_1-t_2|,
\end{equation*}
which means $\langle W\rangle_{\rho_t}$ is uniformly continuous in $t$. Here $|\rho_{t^{'}}|=1$ denotes the trace-norm of the density state $\rho_{t^{'}}$. According to the uniform continuity, we are able to find a $\delta>0$ such that the following inequality
\begin{equation*}
\langle W\rangle_{\rho_{t^{'}}}>\frac{\epsilon}{2}
\end{equation*}
holds if $|t^{'}-t_k|<\frac{\delta}{2}$ for any $t_k$. This further implies
\begin{equation*}
\int_0^\infty\langle W\rangle_{\rho_t}dt\geq\sum_k^\infty\delta\frac{\epsilon}{2}=+\infty,
\end{equation*}
which is a contradiction.
The above contradiction implies that every converging subsequence of
$\langle W\rangle_{\rho_t}$ converges to 0. Then we conclude that
$\lim_{t\rightarrow\infty}\langle W\rangle_{\rho_t}=0$.
\end{proof}
\begin{rem}
For a Lyapunov operator $V$ with $\mathcal{G}(V)\leq0$, we can always let $W=-\mathcal{G}(V)$ and thus the trajectories will converge to $\{\rho:\langle\mathcal{G}(V)\rangle_\rho=0\}$ if $\mathcal{G}(\mathcal{G}(V))$ is bounded, according to Theorem~$5$. The states from the invariant set $\{\rho:\langle\mathcal{G}(V)\rangle_\rho=0\}$ are zero solutions of $W$. This conclusion is similar to the statement of the classical LaSalle theorem.
\end{rem}
\begin{coro}
If Inequality (\ref{ineqlasa}) in Theorem~$5$ is replaced by
\begin{equation*}
\mathcal{G}(V)\leq U-W,
\end{equation*}
where $U$ is a positive operator satisfying
\begin{equation*}
\int_0^\infty\langle U(t)\rangle_{\rho_0}dt<\infty
\end{equation*}
for any initial state $\rho_0$, the conclusions of Theorem~$5$ still hold.
\end{coro}
\begin{proof}
The proof is similar to the proof of Theorem~$5$.
\end{proof}
The question is how we can characterize the pairs of operators $V$ and $W$ for which (\ref{ineqlasa}) holds. Note that $\langle\mathcal{G}(V)\rangle_{\rho}\geq0$ for any ground state $\rho$ of $V$ and consequently $\langle W\rangle_{\rho}=0$. More specifically, $W$ must have the ground states of $V$ as its zero solutions. This observation will limit the set of $W$ we can choose from. For example, if $V=a^\dagger a$ is the energy operator of a quantum oscillator, we will not be able to establish $\mathcal{G}(V)\leq-W$ for the position operator $W=(a+a^\dagger)^2$ because the ground state $|0\rangle\langle0|$ of $V$ has nonzero variance in position. In other words, it is impossible to generate states with zero variance in position by stabilizing the energy of the system. This example reveals the fundamental difficulty in stabilizing non-commuting operators, which is also the implication of the Heisenberg uncertainty principle. Nevertheless, through the stability of $V$ we can still infer the information about the non-commuting operators that are restricted in a subspace. In addition, $W$ can also be used to characterize other invariant limit sets of $\rho_t$ besides the set of the ground states of $V$. We illustrate these ideas in the following example:
\begin{exam}\rm
Consider a single-qubit system with energy operator $V=\frac{1}{2}(1+\sigma_z)$. The aim is to make the expectation of the coherence operator $W=\frac{1}{2}(1+\sigma_x)$ zero and in the same time stabilize the energy of the system. However, the non-commuting observables $V$ and $W$ cannot be stabilized simultaneously via $\mathcal{G}(V)\leq-W$ since the ground state $|1\rangle\langle1|$ of $V$ is not the zero solution of $W$. An alternative solution to this problem is to consider the augmented system with an ancillary qubit and define $W=|0\rangle\langle0|\otimes\frac{1}{2}(1+\sigma_{x_2})$. The energy of the two-qubit system is characterized by the operator $V$ as $V=\sigma_{z_1}+\sigma_{z_2}$. $\sigma_{z_i}$ is the Pauli operator $\sigma_z$ acting on the $i$th qubit. The basis of the bipartite system is chosen as the four eigenstates $\{|00\rangle,|01\rangle,|10\rangle,|11\rangle\}$, leading to the following expression of $V$ and $W$
$$
V=\left(
\begin{array}{cccc}
2&0&0&0\\
0&0&0&0\\
0&0&0&0\\
0&0&0&-2
\end{array}
\right),
W=\left(
\begin{array}{cccc}
\frac{1}{2}&\frac{1}{2}&0&0\\
\frac{1}{2}&\frac{1}{2}&0&0\\
0&0&0&0\\
0&0&0&0
\end{array}
\right).
$$
Although $V$ is not positive, Theorem~$5$ still applies to this example by shifting $V$ with a constant. We can engineer $\mathcal{G}(V)$ (See Appendix) through engineering the couplings between the eigenstates. By introducing the couplings $l|01\rangle\langle00|$ and $l|11\rangle\langle01|$ with $|l|^2=\frac{1}{2}$, $\mathcal{G}(V)$ will become
$$
\left(
\begin{array}{cccc}
-1&0&0&0\\
0&-1&0&0\\
0&0&0&0\\
0&0&0&0
\end{array}
\right).
$$
Set the Hamiltonian control $H$ as $-\frac{1}{2}i|00\rangle\langle01|+\frac{1}{2}i|01\rangle\langle00|=|0\rangle\langle0|\otimes\sigma_{y_2}$, the new $\mathcal{G}(V)$ is
$$
\left(
\begin{array}{cccc}
-1&-1&0&0\\
-1&-1&0&0\\
0&0&0&0\\
0&0&0&0
\end{array}
\right)
$$
which satisfies the required inequality $\mathcal{G}(V)\leq-W$. The system will converge to the zero solutions of $W=|0\rangle\langle0|\otimes\frac{1}{2}(1+\sigma_{x_2})$ while the energy operator is stabilized (the energy of the two-qubit system is decreasing).
Given the density matrix of $\rho$ as
$$
\left(
\begin{array}{cccc}
\rho_{00}&\rho_{01}&\rho_{02}&\rho_{03}\\
\rho_{10}&\rho_{11}&\rho_{12}&\rho_{13}\\
\rho_{20}&\rho_{21}&\rho_{22}&\rho_{23}\\
\rho_{30}&\rho_{31}&\rho_{32}&\rho_{33}
\end{array}
\right),
$$
the limit states will satisfy $\langle|0\rangle\langle0|\otimes\frac{1}{2}(1+\sigma_{x_2})\rangle_{\rho}=0$ and hence $\rho_{00}+\rho_{01}+\rho_{10}+\rho_{11}=0$. In this example we are able to infer the information about the coherence $\rho_{01}+\rho_{10}$ between $|00\rangle$ and $|01\rangle$ within the two-level subspace through the generator of the energy operator $V$. Note that $V$ and $W$ do not commute. One particular state satisfying $\rho_{00}+\rho_{01}+\rho_{10}+\rho_{11}=0$ is $\frac{1}{2}(|00\rangle\langle00|-|00\rangle\langle01|-|01\rangle\langle00|+|01\rangle\langle01|)$, which is an invariant state of the system. Note that the space spanned by $\{|10\rangle,|11\rangle\}$ also satisfies $\rho_{00}+\rho_{01}+\rho_{10}+\rho_{11}=0$. We can further narrow down the set of limit points by making $G_{22}$ negative via the methods introduced in the Appendix such that the invariant set will only contain states that are either in the space spanned by $\{|00\rangle,|01\rangle\}$ with stabilized coherence, or in the ground state $|11\rangle\langle11|$.
Moreover, if we make a projection $|0\rangle\langle0|$ on the $1$st qubit via a quantum measurement, the reduced quantum state of the $2$nd qubit will satisfy $\langle\frac{1}{2}(1+\sigma_{x_2})\rangle_{\rho_2}=0$. Interestingly, by stabilizing the energy operator $\sigma_{z_1}+\sigma_{z_2}$ of the augmented system and then making a projective measurement $|0\rangle\langle0|$, we are able to stabilize the coherence operator $\frac{1}{2}(1+\sigma_x)$ of the qubit in the end.
The interpretation of these results is as follows: Extra space is needed to store the excess noises introduced by the Heisenberg uncertainty principle. This idea is similar to the design of non-degenerate parametric amplifier, where additional channel of noise input is introduced in order to amplify the amplitude and phase quadratures simultaneously.
\qed
\end{exam}
For a positive operator $W$ with unbounded $\mathcal{G}(W)$, we have the following theorem.
\begin{thm}
If there exists a Lyapunov operator $V$ and a positive operator $W$ such that
\begin{equation*}
\mathcal{G}(V)\leq-W, \quad\mathcal{G}(W)\leq0,
\end{equation*}
then $\lim_{t\rightarrow\infty}\langle
V(t)\rangle_{\rho_0}=\lim_{t\rightarrow\infty}\langle V\rangle_{\rho_t}$
exists for any state trajectory $\rho_t$ and (\ref{conlasa}) holds.
\end{thm}
\begin{proof}
Following the same reasoning as in the proof of Theorem~$5$, we can conclude $\int_0^\infty\langle W\rangle_{\rho_{t}}dt<+\infty$. The conditions $\mathcal{G}(W)\le0$ and $\langle W\rangle_{\rho_{t}}$ is bounded from below guarantee that $\langle W\rangle_{\rho_{t}}$ is convergent. The limit of $\langle W\rangle_{\rho_{t}}$ can only be $0$ because $\int_0^\infty\langle W\rangle_{\rho_{t}}dt$ is finite.
\end{proof}
\begin{rem}
Suppose $\mathcal{G}(V)\leq-cV$, $c>0$, and $V$ is a Lyapunov operator. Let $W=cV$ and we have $\mathcal{G}(W)=c\mathcal{G}(V)\leq-c^2V\leq0$. The system will converge to the zero solutions of $V$, or equivalently speaking, to the set of ground states $Z_V=\{\rho:\langle V\rangle_\rho=0\}$.
\end{rem}
Theorem~$6$ can be extended to treat a general Hermitian operator $W$
\begin{thm}
If there exists a Lyapunov operator $V$ satisfying $\langle V(t)\rangle\leq c$ for $t>0$ and
\begin{equation*}
\mathcal{G}(V)=W,
\end{equation*}
where the generator of $W$ satisfies $\mathcal{G}(W)\leq0$, then
\begin{equation*}
\lim_{t\rightarrow\infty}\langle W(t)\rangle_{\rho_0}=\lim_{t\rightarrow\infty}\langle W\rangle_{\rho_t}=0.
\end{equation*}
\end{thm}
\begin{proof}
$\langle V(t)\rangle$ is bounded for all $t>0$. From $(\ref{integene1})$ we know that $-\infty<\int_0^\infty\langle W\rangle_{\rho_{t}}dt<+\infty$. If $\mathcal{G}(W)\leq0$ by assumption, then the monotonic sequence $\langle W\rangle_{\rho_{t}}$ is bounded from below and hence will converge to a limit. The limit is exactly $0$ since the integral $\int_0^\infty\langle W\rangle_{\rho_{t}}dt$ is bounded.
\end{proof}
\section{Stability within the Invariant Set}\label{sec7}
We have used multiple Lyapunov conditions in Theorem~$6$ and Theorem~$7$. Similarly, we can use additional Lyapunov conditions to further engineer the dynamics of the trajectories within the invariant set. For example, we can make use of the Lyapunov operator $W=V^2$ to drive the system states to the zero solutions of $V$, where in general the system will converge only to the zero solutions of $\mathcal{G}(V)$ by LaSalle invariance principle.
As we have known from classical stochastic stability and quantum semigroup theory, the asymptotic dynamics of the trajectories are determined by the diffusion terms \cite{Khas} or the dissipation functional $\mathfrak{D}(\cdot)$ \cite{Frigerio78,Alberto1982,fagnola2003quantum}. As shown in the proof of Theorem~$3$, we can make explicit connection between the dissipation functional and the diffusion terms $j_t (\mathcal{B}(X)), j_t (\mathcal{C}(X))$ by calculating $\mathcal{G}(V^2)$.
\begin{thm}
Suppose $\mathcal{G}(V)\leq0$ for the Lyapunov operator $V$ of a finite-dimensional system. The state trajectory $\rho_t$ will converge to the set of zero solutions $Z_V=\{\rho:\langle V\rangle_\rho=0\}$ if $\langle[L^\dagger,V][V,L]\rangle_\rho>0$ for $\rho\notin Z_V$ and $[\mathcal G(V),V]=0$.
\end{thm}
\begin{proof}
Since $\mathcal{G}(V)\leq0$, $\lim_{t\rightarrow\infty}\langle V\rangle_{\rho_t}$ exists and $Z_V$ is an invariant set. We only need to prove that $\rho_t$ will exit the domain $\{\rho:\langle V\rangle_\rho\geq\epsilon\}$ for arbitrary $\epsilon>0$. Consider the positive operator $W(V)=V^2$. Similar to the derivations in Theorem~$3$, the generator for $W(V)$
can be calculated using the quantum It\={o} formula
\begin{equation*}
\mathcal{G}(W)=V\mathcal{G}(V)+\mathcal{G}(V)V+\mathfrak{D}(V)
\end{equation*}
with
\begin{equation*}
\mathfrak{D}(V)=\mathcal{B}(V)^2+ \mathcal{C}(V)^2+i\mathcal{B}(V)\mathcal{C}(V)-i\mathcal{C}(V)\mathcal{B}(V)=[L^\dagger,V][V,L].
\end{equation*}
For finite-dimensional system, any state trajectory is tight. Suppose the trajectory $\rho_t$ is restricted to a domain $\{\rho:\langle V\rangle_\rho\geq\epsilon\}$ for some $\epsilon>0$. Then by Theorem~$1$ there exists an invariant state $\rho_I$ which is the limit point of the tight sequence $\frac{1}{t}\int_0^{t}\rho_{t^{'}}dt^{'}$. Note that $\frac{1}{t}\int_0^{t}\rho_{t^{'}}dt^{'}$ is the mean of the sequence $\rho_t$, so $\rho_I$ is in the same domain $\{\rho:\langle V\rangle_\rho\geq\epsilon\}$ as $\rho_t$ which means $\rho_I\notin Z_V$.
Let the initial state be exactly the invariant state $\rho_I$. First we prove $\langle V\mathcal{G}(V)+\mathcal{G}(V)V\rangle_{\rho_I}=0$. Since $V$ is positive and $[\mathcal G(V),V]=0$, $V\mathcal{G}(V)$ and $\mathcal{G}(V)V$ are negative hermitian operators which make $\langle V\mathcal{G}(V)+\mathcal{G}(V)V\rangle_{\rho_I}\leq0$. Furthermore, we have $\langle V\mathcal{G}(V)+\mathcal{G}(V)V\rangle_{\rho_I}=\langle (V+\beta)\mathcal{G}(V)+\mathcal{G}(V)(V+\beta)\rangle_{\rho_I}$ due to the fact that $\langle\mathcal{G}(V)\rangle_{\rho_I}=0$. $V$ is bounded, so we can choose $\beta<0$ such that $V+\beta$ is negative. Given this $\beta$, we can conclude $\langle (V+\beta)\mathcal{G}(V)+\mathcal{G}(V)(V+\beta)\rangle_{\rho_I}\geq0$ which gives us $\langle V\mathcal{G}(V)+\mathcal{G}(V)V\rangle_{\rho_I}\geq0$. So $\langle V\mathcal{G}(V)+\mathcal{G}(V)V\rangle_{\rho_I}=0$. Next we have the following relation by integrating $\mathcal{G}(W)$
\begin{eqnarray}
\langle W(V)\rangle_{\rho_I}-{\langle W(V)\rangle_{\rho_I}}
&=&\int_0^t\langle V\mathcal{G}(V)+\mathcal{G}(V)V\nonumber +\mathfrak{D}(V)\rangle_{\rho_I}dt^{'}\\
&=& \int_0^t\langle \mathfrak{D}(V)\rangle_{\rho_I}dt^{'}.
\end{eqnarray}
The LHS of the equality is zero, however the RHS of the
equality is strictly positive, since $\langle \mathfrak{D}(V)\rangle_{\rho_I}=\langle[L^\dagger,V][V,L]\rangle_{\rho_I}>0$ by assumption. So we arrive at a contradiction. The contradiction shows that a trajectory $\rho_t$ cannot be confined to the domain $\{\rho:\langle V\rangle_\rho\geq\epsilon\}$. Hence $\rho_t$ will approach $Z_V$ asymptotically.
\end{proof}
\begin{coro}
Assume in a finite-dimensional system the Lyapunov operator $V$ has the decomposition $V=M^\dagger M$. If $M$ solves $M=[V,L]$ and then $\mathcal{G}(V)\leq0,[\mathcal G(V),V]=0$, the state trajectory $\rho_t$ will converge to the zero solutions of $V$.
\end{coro}
\begin{exam}\rm
Consider a qubit with $V=\frac{1}{2}(1+\sigma_z)$, or in matrix expression
$$
V=\left(
\begin{array}{cc}
1&0\\
0&0
\end{array}
\right).
$$
The decomposition is found to be $V=\sigma_+\sigma_-$ with
$$
\sigma_+=\left(
\begin{array}{cc}
0&1\\
0&0
\end{array}
\right),
\sigma_-=\left(
\begin{array}{cc}
0&0\\
1&0
\end{array}
\right).
$$
$\sigma_+^\dagger=\sigma_-=M$. The solution to $M=[V,L]$ is
$$
L=\left(
\begin{array}{cc}
a&0\\
1&b
\end{array}
\right)
$$
with $a$ and $b$ being arbitrary constants. With this $L$, the dissipation part $\mathfrak{L}(V)$ equals
$$
\left(
\begin{array}{cc}
-1&-\frac{b}{2}\\
-\frac{b}{2}&0
\end{array}
\right).
$$
Let $H=0$ and $b=0$, then $\mathcal{G}(V)=\mathfrak{L}(V)\leq0$. The system will converge to the ground state $|1\rangle\langle1|$.
\end{exam}
\section{Conclusion}\label{sec8}
Many theorems concerning asymptotic properties of quantum
Markov semigroups have the existence of a faithful invariant state as an
essential assumption. We have derived sufficient conditions to verify this
assumption. If these sufficient conditions hold, the unique and faithful
state is an equilibrium point which is also globally attractive. Our
approach makes use of the Lyapunov method complemented by additional algebraic
conditions. Our result exhibits some analogy with the classical
Foster-Lyapunov theory concerning the existence of invariant measures of
Markov processes. Beyond invariant states, we have introduced the quantum
invariance principle to characterize the set of limit states of the system
dynamics. More specifically, the system will asymptotically converge to the ground state of an operator $W$ if we are able to engineer the generator of a Lyapunov operator $V$. These invariance theorems are established via a Lyapunov
inequality between these two operators, which has potential to provide useful
tools for stability analysis of the non-commutative algebra associated with
general quantum coherent control systems. Moreover, the system can be
driven further to the ground state of $V$ within the invariant set if additional
conditions on the Lyapunov operator can be engineered. These results may also find essential applications in quantum information processing, since the outcomes of quantum computations can be encoded in the ground state of a particular operator \cite{Verstraete09}.
\section{Appendix}
In this appendix we introduce a constructive method to engineer a negative generator for the Lyapunov operator $V$.
First we will focus on engineering the dissipation part $\mathfrak{L}(V)=\frac{1}{2}(2L^{\dagger}VL-L^{\dagger}LV-VL^{\dagger}L)$ of the generator $\mathcal{G}(V)$ by assuming $[V,H]=0$. In a separable space we can decompose $L$ and $V$ as
$$L=\left(
\begin{array}{cc}
L_{00} & L_{01} \\
L_{10} & L_{11}
\end{array}
\right), V=\left(
\begin{array}{cc}
V_{00} & V_{01} \\
V_{10} & V_{11}
\end{array}
\right).$$$V$ is a positive hermitian operator, so we can always
make $V_{01}=V_{10}=0$ through spectral decomposition. The generator
$\mathcal{G}(V)$ is calculated to be
$$\mathcal{G}=\left(
\begin{array}{cc}
G_{00} & G_{01} \\
G_{10} & G_{11}
\end{array}
\right)$$with
\begin{eqnarray*}
G_{00}&=&(L_{00}^{\dagger}V_{00}L_{00}+L_{10}^{\dagger}V_{11}L_{10})-\frac{1}{2}\{L_{00}^{\dagger}L_{00}+L_{10}^{\dagger}L_{10},V_{00}\},\\
G_{01}&=&(L_{00}^{\dagger}V_{00}L_{01}+L_{10}^{\dagger}V_{11}L_{11})-\frac{1}{2}(L_{00}^{\dagger}L_{01}+L_{10}^{\dagger}L_{11})V_{11}-\frac{1}{2}V_{00}(L_{00}^{\dagger}L_{01}+L_{10}^{\dagger}L_{11}),\\
G_{10}&=&G_{01}^\dagger,\\
G_{11}&=&(L_{01}^{\dagger}V_{00}L_{01}+L_{11}^{\dagger}V_{11}L_{11})-\frac{1}{2}\{V_{11},L_{01}^{\dagger}L_{01}+L_{11}^{\dagger}L_{11}\}.
\end{eqnarray*}
Here we only present the calculations for two-level system. For higher dimensional systems, the blocks of $G,L$ and $H$ will be
matrices or operators. However, we can still do analysis for arbitrary dimensional systems by carefully engineering the two-dimensional subsystems and compensating the interactions between different subsystems. This approach is possible because of the
linearity of the generator $\mathcal{G}$. For example, $G_{00}$ can be divide into
the internal dynamics $2L_{00}^{\dagger}V_{00}L_{00}-\{L_{00}^{\dagger}L_{00},V_{00}\}$
and the interaction with other dimensions $2L_{10}^{\dagger}V_{11}L_{10}-\{L_{10}^{\dagger}L_{10},V_{00}\}$.
\subsection{$V_{00}=V_{11}$}
We can set $L_{00}=0$ and $L_{11}=0$ to make
$G_{01}=G_{10}=0$. However, due to the degeneracy of $V$, we also
have $G_{00}=G_{11}=0$. Therefore, $\mathcal{G}=0$ and the entire
two-dimensional space is irreducible. The off-diagonal elements of $L$ will not affect $\mathcal{G}$.
\subsection{$V_{00}{\neq}V_{11}$}
Without loss of generality we assume $V_{11}-V_{00}=v>0$. Again assuming $L_{00}=0$ and $L_{11}=0$, the
generator becomes
$$\mathcal{G}=\left(
\begin{array}{cc}
vL_{10}^{\dagger}L_{10}&0\\
0&-vL_{01}^{\dagger}L_{01}
\end{array}
\right).$$Now we can set $L_{10}=0,L_{01}=l\neq0$ so that
$G_{00}=0,G_{11}<0$. The coupling operator $L$ for engineering negative $\mathcal{G}(V)$ with non-degenerate spectrum could be
$$L=\left(
\begin{array}{cc}
0&l\\
0&0
\end{array}
\right).$$.
\subsection{$[V,H]\neq0$}
$[V,H]\neq0$ happens when $V$ is not representing the energy of the
system or additional Hamiltonian control $H_c$ is needed for
stabilization. Since $V_{01}=V_{10}=0$, the commutator $C=-i[V,H]$
can be calculated as
$$C=-i\left(
\begin{array}{cc}
[V_{00},H_{00}]&V_{00}H_{01}-H_{01}V_{11}\\
V_{11}H_{10}-H_{10}V_{00}&[V_{11},H_{11}]
\end{array}\right).$$
$C_{00}$ is the internal unitary dynamics within the subspace
$X_{00}$. For two-dimensional system, $V_{ij}$ and $H_{ij}$ are
complex numbers, which gives $C_{00}=C_{11}=0$. If
$V_{00}=V_{11}$, then $C_{01}=C_{10}=0$ and the unitary dynamics
induced by $H$ will not affect $\mathcal{G}$. If $V_{00}\neq V_{11}$,
we set $V_{11}-V_{00}=v>0$ and $L_{10}=0$. In this case, the
generator will still satisfy the relations $G_{00}=0$ and $G_{11}<0$
after adding $C$ to $\mathcal{G}$. However, $G_{01}$ cannot be made
vanish if the diagonal entries of $L$ are all zero. In fact, we have
\begin{equation*}
G_{01}=L_{00}^\dagger L_{01}(V_{00}-V_{11})-iH_{01}(V_{00}-V_{11}),
\end{equation*}
so $L_{00}^\dagger L_{01}=iH_{01}$ must be satisfied. If we choose
$L_{01}=l\neq0$, the coupling operator $L$ should be in the
following form
$$L=\left(
\begin{array}{cc}
-\frac{iH_{01}^*}{l^*}&l\\
0&L_{11}
\end{array}\right)
$$ to completely eliminate the influence of $H$.
\bibliographystyle{plain}
|
\section{Introduction}
Among the various approaches to an interpretation of quantum theory, one is to regard the
superposition principle, state vectors, and the Schr\"odinger equation as universally valid,
and to seek a solution to the ensuing many-worlds problem \cite{Everett1973}. An old
\cite{vNeumann1932} but still relevant \cite{Donald1999,Tegmark2014} conjecture is that the
solution should involve the physical functioning of an observer's consciousness.
The approach taken here falls in this category.
As to the merits of the superposition principle,
most concepts of quantum theory, as presented in textbooks, rely on the formalism of state
vectors. Even thermal systems, traditionally regarded as mixtures, can be
treated as pure state vectors \cite{Tasaki1998}.
The basic law of propagation of a particle takes an almost self-evident form
\cite{Feynman1965,Baym1978,Zee1991,Polley2001} when positions are restricted to a spatial
lattice, and superpositions are regarded as a logical possibility. For electromagnetic fields
to be incorporated, only the complex phases already present in the hopping amplitudes need to
be varied \cite{Wilson1974}. In comparison, Newton's laws are phenomenological.
A fundamental problem of a linear evolution equation emerges in application to a quantum system interacting with an observer. With a system in a superposition of properties $1,\ldots,B$, and an observer trying to determine the ``actual'' property, the Schr\"odinger equation implies a transition of the form
\begin{equation} \label{EqualAmplitudeSplitting}
\left(\sum_{n=1}^B |n\rangle \right)|\mathrm{ready}\rangle \longrightarrow
\sum_{n=1}^B \Big(|n\rangle|\mathrm{observed}\,n\rangle\Big)
\end{equation}
While experience suggests an observer should find himself
in one of the states $|\mathrm{observed}\,n\rangle$ after the measurement, the superposition state produced by the Schr\"odinger equation does not indicate which of the possible results has ``actually'' been obtained. The observer rather seems to have split into $B$ branches of himself. To date, no consensus exists as to whether a superposition of observer states like (\ref{EqualAmplitudeSplitting}) describes something physically real. The problem appears less dramatic when the state vector is converted into a density matrix, as in the theory of decoherence, but the ambiguity about the result of the measurement persists \cite{Schlosshauer2007}.
In Everett's many-worlds interpretation \cite{Everett1973} the superposition \emph{is} observer's real wavefunction. His inability to realise more than one of his branches is inferred from the (undisputed) impossibility of branches interacting with each other. While certain activities like talking in branch 1 about events in branch 2 can be ruled out in this way, a mere simultaneous awareness of branches is not covered by the argument
\cite{Penrose1997}. An attempt at resolving this problem was made by this author in
\cite{Polley2012}; in the present paper, that approach is simplified and generalised. The hypothesis is that an observer's awareness is extremal in one branch, in such a way that the sum of remaining branches can be regarded as a negligible contribution.
In the Copenhagen interpretation, the process of measurement is not described by
a linear equation for amplitudes in a superposition, but is described by state reduction
following Born's rule. Accordingly, superpositions evolve in a stochastic way. Superposition
(\ref{EqualAmplitudeSplitting}) would end up as $|n\rangle|\mathrm{observed}\,n\rangle$ with
probability $1/B$ for each case. The disturbing point here is that
probabilities occur at a fundamental level---no natural law supposedly exists that would
determine, in principle, the outcome of a single quantum measurement.
By contrast, an agreeable role for stochastics would be that of
an approximation to deterministic but uncontrollably complicated dynamics. In the model of this
paper, there will be a constant law of evolution (representing determinism) given by a unitary
operator which is a peculiar kind of random matrix (assumed to approximate some complicated
dynamics). Pseudo-random evolution results if the operator is applied repeatedly to an
initial state of an appropriate class.
Since the early days of quantum mechanics the idea has been pondered that state reduction may involve an observer's consciousness \cite{vNeumann1932}. While the observer remains conscious with every result of the measurement, there may be variations in the degree of consciousness. Of all the details around us that could in principle catch our attention, only a tiny fraction actually does so. If entire histories are considered, that fraction multiplies with every instant of time, so there is lots of room for outstanding extremes.
A formal description of the physical functioning of a \emph{real} observer, including both
Everettian branching and the influence of an irreducible stochastic process in the generation
of the branches, has been given by Donald \cite{Donald1999}. For a model of state reduction,
it may be an unnecessary complication to consider consciousness as versatile as that of a human
being.
Yet, as emphasised in \cite{Donald1999}, a lesson from real brain dynamics is that consciousness cannot be described statically, by assigning labels to mental states as in equation (\ref{EqualAmplitudeSplitting}), but that neural activity is required, like the switching between firing and resting states of a neuron. It makes a great difference for model building! Neurons thus come as subsystems with two dimensions at least, and fluctuations in the number of neurons appear vastly enhanced in the dimensions of Hilbert spaces involved.
The guiding idea of the present paper is that Everett's many worlds are not all
equivalent for an observer, but that his consciousness is \emph{physically} contained
almost exclusively in one world. The technical basis of this approch is a theorem of order
statistics \cite{Embrechts1997,David1981}, relating to statistical ensembles with power-law
distribution. The largest draw,
in that case, exceeds the second-largest by a quantity $\Delta$ of the order of the ensemble
size taken to some power. On an Everettian world tree, the ensemble size is huge, and
concentrated near the end of the branches. Thus the dominance of the extremal
draw is particularly pronounced, and it occurs near the end of a branch, thus singling it out.
If the draws are for numbers of neurons, the exceedance $\Delta$ exponentiates because the
dimensions of subsystems multiply. In fact, to see whether and how this statistical mechanism
might be relevant for state reduction was the main guide for the constructions below.
The model scenario is as follows. At an equidistant sequence of times, a quantum system is
observed, generating $B$ branches of itself, of a number of records, and of a corresponding
number of observer's neurons. The main interest is in the statistics of dimensions;
therefore, system states, records, and neurons are only distinguished by an index and
are not specified any further. The law of evolution, for a step of time, is given by a unitary
operator. A specific form of initial state must be assumed for the scenario to unfold. This
``objective'' part of the model dynamics is constructed in section \ref{secRS}.
As to an observer's consciousness, it should be memory-based \cite{Edelman1989};
``the history of a brain’s functioning is an essential part of its nature as an object on which
a mind supervenes'' \cite{Donald1997}. In the simplification of the model,
memories and the history of the brain's functioning are identified with the records.
The supervening mind is represented by neuronal activity drawing on memories, i.e., records.
The implementation of this drawing, by a (pseudo-)random cascade of neuronal activity,
serves two purposes: generating a draw from a power-law distribution, and composing in one
draw a conscious history---by ``recalling'' what happened at a certain time, then what happened
before and after, and before and after that again, and so on. This ``subjective'' part of the
model dynamics is constructed in section \ref{secObserversQuest}.
At several points in the construction of the evolution operator, random draws are made. As
mentioned already, these are supposed to approximate some complex deterministic law of
evolution, and they are made ``once and for all times''.
The evolution operator is the same at all times.
Some conclusions are given in section \ref{Conclusions}, and a technically convenient restriction of the dynamics is defined in appendix \ref{secAvoidingLoops}.
\section{Records and objective dynamics\label{secRS}}
\subsection{A paradigm: Griffiths' models of histories}
An elaborate version of the Copenhagen interpretation is the formalism of consistent histories \cite{Griffiths2002}. The notion of measurement on a quantum system is modified to that of a property (selected from an available spectrum) which the system has, irrespective of whether an observer is in place. Those properties a system has at times $t_0,t_1,\ldots, t_n$, while no property should be imagined for intermediate times. By a generalised Born's rule, probabilities are defined for sequences of the available properties (histories of the system) to occur. There are constraints, involving the time evolution operator, to be imposed on the sort of properties that can form a history. In order to demonstrate the constraints, Griffiths devises a number of toy models by immediately constructing operators of time-evolution, rather than Hamiltonians. For this purpose, the tensor-product structure of the so-called history Hilbert space turns out to be quite convenient. Moreover, models for simultaneously running processes in a time interval can be simply constructed as products of unitaries.
Although the intention in \cite{Griffiths2002} is to keep observers out of the theory, it certainly is a reasonable approximation for an observer model as well to consider awareness only at times $t_0,t_1,\ldots, t_n$ while leaving unspecified observer's state in between. In the model to be constructed, the times will be equidistant. The operator of evolution from an instant to the next will be the product $U_\mathrm{wit\,2}U_\mathrm{con}U_\mathrm{wit\,1}U_\mathrm{orb}U_\mathrm{age}$, their roles being to increase the ages of records, move the system along its (branching) orbit, create first half of witnessing records, run awareness cascade, and create second half of witnessing records. Each of these is a product of a large but finite number of unitary factors.
The tensor product structure of the space of states, which in Griffiths' formalism emerges from the notion of history Hilbert space, is a convenient element of model building independently of that notion, because it facilitates the construction of operators that manifestly commute. Below, the tensor product will describe a reservoir of potential records and their neuronal counterparts, each of which being treated as a subsystem. By construction of the evolution operator, every branch will consist of its own collection of subsystems. This makes the numbers of subsystems very large, raising the question of whether a ``multiverse'' in the classical sense is tacitly assumed for the model. It should therefore be noted that the subsystems can be mathematically identified with degrees of freedom of a conventional Hilbert space. For example, any unitary space of dimension $2^n$ can be rewritten as a tensor product by expressing the index of basis vectors in binary form and identifying
$$
|i_1,\ldots,i_n\rangle \equiv |i_1\rangle \cdots |i_n\rangle \qquad i_k=0,1
$$
For the counting of dimensions, both versions are equivalent.
In order to show that one branch nearly exhausts all dynamical dimension of awareness, what matters is the quantity of records and of neuronal response. The only quality of relevance is the date at which a record is created. For the model it therefore suffices to distinguish records by an unspecified index, and to make their age the only stored content.
\subsection{Structure of state vectors; dating of records\label{secAgeing}}
The subsystems of the model are ``objective'' records and ``subjective'' bits of
mental processing. For the records, two classes are assumed. Those in class ${\cal O}$ (orbits) induce further records in the course of evolution, giving rise to branching world lines. The branches emanating from a record $k\in{\cal O}$ are collected in a set ${\cal B}(k)$ of $B$ elements. They are envisioned as predetermined and (due to complex dynamics of real macroscopic systems) pseudorandom. They are defined here\footnote{A more technical specification, simplifying evaluation, is given in appendix \ref{secAvoidingLoops}.} by random draw, once and for all time:
\begin{equation} \label{DefB(k)}
{\cal B}(k) = \{j(k,1),\ldots,j(k,B)\} \mbox{ where } j(k,s) = \mbox{random draw from }
{\cal O}\backslash\{k\}
\end{equation}
The other class of objective records consists of mere witnesses, holding redundant information about records in ${\cal O}$:
\begin{equation} \label{DefW(k)}
k\in{\cal O}\mbox{ is witnessed by all records } l \in {\cal W}(k)
\end{equation}
All ${\cal W}(k)$ are assumed to have the same macroscopic number $W$ of elements.
Observer's neurons are associated with witnessing records, not immediately with orbital
records. The model neurons are distinguished by an index and are not specified any further.
However, it could make sense to address the same anatomical neuron at different times by
different indices. The multiple degrees of freedom would then be provided by the metabolic
environment.
Denoting by ${\cal R}_k$ and ${\cal N}_k$ the state-vector space of a record and an observer's ``neuron'', respectively, the model Hilbert space is
$$
{\cal H} = \bigotimes_{k\in{\cal O}}\left({\cal R}_k \otimes
\bigotimes_{l\in{\cal W}(k)} \left( {\cal R}_l\otimes {\cal N}_l \right) \right)
$$
The information to be stored in a recording subsystem is whether anything is recorded at all,
and if so, since how many steps of evolution. The basis states of a record of the orbital kind thus are
\begin{equation} \label{DefRecordBasisOrbit}
|\mathrm{blank}\rangle, ~ |\mathrm{age}~m\rangle \quad m\in{\bf Z}
~~~~ \mbox{ for each index in } {\cal O}
\end{equation}
Observer's mental processing of a record is modelled by transitions between the firing and resting state of a ``neuron''. The basis states of a witnessing record and its mental counterpart are
\begin{equation} \label{DefRecordBasisWitness}
\left\{ \begin{array}{l} |\mathrm{blank}\rangle \\
|\mathrm{age}~m\rangle \quad m\in{\bf Z} \end{array}\right\} \otimes
\left\{ \begin{array}{c} |\mathrm{rest}\rangle \\ |\mathrm{fire}\rangle \end{array}\right\}
~~ \mbox{for each index in } \bigcup_{k\in{\cal O}}{\cal W}(k)
\end{equation}
The ages of records could be limited to an observer's lifetime, as it was done in \cite{Polley2012}, but the infinite dimension implicit in (\ref{DefRecordBasisOrbit}) and (\ref{DefRecordBasisWitness}) is harmless with finite products of unitary operators, and avoids an unnecessary degree of subjectivity in the model.
For all recording subsystems, an ageing operation is defined:
\begin{equation} \label{DefUage}
U_\mathrm{age} = \prod_{k\in{\cal O}} U_\mathrm{age}(k)
\prod_{i\in{\cal W}(k)}U_\mathrm{age}(i)
\end{equation}
where for record number $r$
$$
\begin{array}{l}
U_\mathrm{age}(r)|\mathrm{blank}\rangle_r = |\mathrm{blank}\rangle_r \\
U_\mathrm{age}(r)|m\rangle_r = |m+1\rangle_r
\end{array}
$$
\subsection{Preferred initial state\label{secInitialState}}
A real observer's identity derives from a single DNA molecule. For a model of an observer's
history, this is taken here as justification for considering exclusively the evolution from an
initial state in which one orbital record $k_0$ and its witnesses ${\cal W}(k_0)$ are in the
zero-age state while all other records are in their blank states. The choice of a zero-age
record determines observer's entire history.
It is the only ``seed'' for all ensuing pseudo-random processes of evolution.
\subsection{Orbital branching\label{secOrbit}}
The idea is that observer's history branches at every step of evolution, like in a quantum
measurement. A new branch is described by an index of a new record, and is not specified any
further. Imagining a new quantity being measured at every step would seem to be consistent with
this scenario.
Deutsch \cite{Deutsch1999} showed, for systems with sufficiently many degrees of
freedom, that Born's rule for superpositions with coefficients more general than in equation
(\ref{EqualAmplitudeSplitting}) can be reduced to the
equal-amplitude case, providing the unitarity of any physical transformation is taken for
granted. In the model to be constructed, evolution will be unitary. Therefore,
invoking Deutsch's argument, only branching into equal-amplitude superpositions will be
considered.
Under the condition that record $k$ is older than zero, and that all records to
which the orbit possibly continues are blank, the orbit does continue as a superposition of
zero-age states of the records of address set ${\cal B}(k)$. Else, the identity operation is carried out. The corresponding evolution operator, specific to point $k$ on an orbit, is defined using the following basis of the \emph{partial tensor product} relating to the records of the
set ${\cal B}(k)$.
\begin{equation} \label{PartialBasis}
\begin{array}{l}
|\Psi_0\rangle = \prod_{l\in{\cal B}(k)} |\mathrm{~blank}\rangle_l \\[3mm]
|\Psi_l\rangle = \Big(|0\rangle\langle\mathrm{blank}|\Big)_l|\Psi_0\rangle
\qquad l\in{\cal B}(k)
\end{array}
\end{equation}
That is, one basis vector has all records of ${\cal B}(k)$ in the blank
state, while the remaining have one record promoted to the zero-age state.
The subspace orthogonal to $|\Psi_0\rangle$, \ldots, $|\Psi_B\rangle$
is spanned by product vectors with more than one record in a zero-age state or with records
in higher-age states.
The idea of equal-amplitude branching from point $k$ is that $|\Psi_0\rangle$ should evolve
into a superpositon of $|\Psi_1\rangle$ to $|\Psi_B\rangle$; in the basis (\ref{PartialBasis}),
\begin{equation} \label{SplittingShorthand}
\left(\begin{array}{c} 1 \\ 0 \\ \vdots \\ 0 \end{array} \right) \longrightarrow
\frac1{\sqrt B} \left(\begin{array}{c} 0 \\ 1 \\ \vdots \\ 1 \end{array} \right)
\end{equation}
A convenient way of completing this to define a unitary operator is to use a Fourier basis in
$B$ dimensions,
$$
F_m = \frac1{\sqrt B}\left(\begin{array}{c} \alpha_m^0 \\ \alpha_m^1 \\ \vdots \\
\alpha_m^{B-1} \end{array} \right) \qquad \alpha_m = \exp\frac{2\pi i m}{B}
\qquad m = 0,\ldots B-1
$$
Relating to the basis (\ref{PartialBasis}), and using the $F_m$ as $B$-dimensional column vectors, a branching operation can be defined using the $(B+1)\times(B+1)$ matrix
$$
S = \left(\begin{array}{ccccc} 0 & 1 & 0 & \cdots & 0 \\
F_0 & 0 & F_1 & \cdots & F_{B-1} \end{array} \right)
$$
whose columns form an orthonormal set.
Conditioning on $\mathrm{age} = 1$ of record $k$, the factor of orbital evolution triggered by
this record is then given by
\begin{equation} \label{DefUorb(k)}
U_\mathrm{orb}(k) = 1 + \left( \sum_{n,n'=0}^B |\Psi_n\rangle
(S_{nn'}-\delta_{nn'}) \langle\Psi_{n'}|\right)_{{\cal B}(k)}
|\mathrm{age}~1\rangle_k \langle \mathrm{age}~1|_k
\end{equation}
The bracket reduces to zero, in particular, when branching from $k$ has occurred previously in the evolution, so that, by subsequent ageing of non-blank records,
any zero-age components of records have been promoted to higher-age components;
cf.\ (\ref{PartialBasis}). In this way, repeated branching from the same point as well as loops of evolution are avoided. Technically, however, an additional means of avoiding loops (appendix \ref{secAvoidingLoops}) facilitates the evaluation of evolution.
The global operator of orbital evolution is
\begin{equation} \label{DefUorb}
U_\mathrm{orb} = \prod_{k\in{\cal O}} U_\mathrm{orb}(k)
\end{equation}
\subsection{Witnessing}
The basis states of the witnessing records are defined in (\ref{DefRecordBasisWitness}).
In order to represent (\ref{DefW(k)}) by an evolution operator, orbital records of zero age
are assumed to induce a change of witnessing records from blank to zero-age. The relevant part of the operation is, in self-explaining notation,
\begin{equation} \label{UwitSimpl}
\left( \prod_{l\in{\cal W}(k)} \Big(|0\rangle \langle\mathrm{blank}|\Big)_l\right)
\Big( |0\rangle \langle 0|\Big)_k
\end{equation}
For the sake of unitarity, however, this needs to be complemented by further operations,
although these will never become effective in the evolution of an initial state as defined in section \ref{secInitialState} and as age-promoted by the operators of section \ref{secAgeing}.
A complemented version of the operator above would be
$$
1 + \left( -1 + \prod_{l\in{\cal W}(k)} \left[|0\rangle \langle\mathrm{blank}|
+ |\mathrm{blank}\rangle \langle 0| + \sum_{m\neq 0}|m\rangle\langle m|\right]_l
\right) \Big( |0\rangle \langle 0|\Big)_k
$$
However, witnessing records just created can be read and processed by an observer within the
same step of evolution. It will be essential for the functioning of a stochastic mechanism below (section \ref{secObserversQuest}) that half of the witnessing records are created before the observer might immediately address them, while the other half is created thereafter. For this purpose, let the addresses of witnessing records be split into subsets of equal size,
$$
{\cal W}(k) = {\cal W}_1(k) \cup {\cal W}_2(k) ~~~~~~
{\cal W}_1(k) \cap {\cal W}_2(k) = \emptyset
$$
The two corresponding witness-generating evolution operators are
\begin{equation} \label{DefUwit}
U_\mathrm{wit\,1} = \prod_{k\in{\cal O}} U_\mathrm{wit\,1}(k) ~~~~~~~~~~~~~~~~~~
U_\mathrm{wit\,2} = \prod_{k\in{\cal O}} U_\mathrm{wit\,2}(k)
\end{equation}
where \vspace*{-4mm}
$$
~~~~~~~~ \begin{array}{l}
U_\mathrm{wit\,1,2}(k) ~ = ~ 1 ~ + \\[1mm] \displaystyle
\left( -1 + \!\!\! \prod_{l\in{\cal W}_{1,2}(k)} \left[|0\rangle \langle\mathrm{blank}|
+ |\mathrm{blank}\rangle \langle 0| + \sum_{m \neq 0}|m\rangle\langle m|\right]_l
\right) \Big( |0\rangle \langle 0|\Big)_k
\end{array}
$$
\section{Observer's mental programme\label{secObserversQuest}}
Consider an observer who is constantly trying to assemble his records into a coherent temporal
sequence. One way of doing this would be to see what happened at the middle $a/2$ of his life
at age $a$, by seeking an appropriate record; then what happend a quarter before and after, at
$a/4$ and $3a/4$, then at multiples of $a/8$, and so forth.
This defines a cascade of increasing temporal resolution. For
``coherence'', connection by a logical AND is required. It would fit in with the spirit (not
with the technical detail) of Tononi's concept of consciousness as Integrated Information
\cite{Tononi2008}:
``Phenomenologically, every experience is an integrated whole, one that means what it means by
virtue of being one, \ldots''. Moreover, memory
becomes a fundamental constituent of consciousness \cite{Edelman1989} in this way.
\subsection{Generating the power-law statistics\label{secGenPowStat}}
The idea for generating power-law statistics, within one step of evolution, is as follows.
While the AND condition is satisfied, the cascade of records addressed grows like $2^l$ where
$l$ is the generation number. By the scanning procedure to be constructed, records of non-zero
ages will be found with probability 1 whenever addressed, but records of zero age (being
created within the same step and representing the ``present'') will only be found with
probability 1/2. The entire cascade is stopped when the quest for a coherent picture of past
and present fails, for which the probability is 1/2 in every generation. This is a standard
mechanism for generating power law statistics \cite{SimkinRoychowdhury2006}, here with exponent
$-1$ for the cumulative distribution function since a number greater than $n=2^{l+1}-1$ (sum of
generations) is obtained with probability $\frac12(n+1)^{-1}$.
\subsection{Recursive construction of awareness cascade}
The unitary operator to be constructed in this section will go through all possible cascades
of records of ages $a/2$, multiples of $a/4$, of $a/8$ and so on, looking for a randomly
chosen witnessing record $i$ in every ${\cal W}(k)$ of the cascade. It should be noted again
that all random draws are made once and for all times. Parameter $a$ will be an eigenvalue of
observer's age operator, defined in section \ref{secObserversAge}.
We begin by constructing the $l$th generation of the cascade. Consider a set $g$ of addresses
given by pairs $(k,i)$ with
\begin{equation} \label{ikjDef}
\left. \begin{array}{rcl}
k(j) & = & \mbox{label of an orbital record} \\
& & \mbox{restricted by } k(j) < k(j') \mbox{ for } j < j'
\\[3mm]
i(j) & = & \mbox{random draw from ${\cal W}(k(j))$}
\end{array} \right\} \mbox{ for }j = 1,\ldots,2^l
\end{equation}
The ordering of the $k$ is for technical convenience; permutations are taken into account when
assigning ages, as below. Denote the collection of all possible sets of the above form by
$$
{\cal G}(l) = \big\{ \mbox{all possible $g$ of the form (\ref{ikjDef})} \big\}
$$
The random draws are understood to be \emph{independent for different} $g$.
The elementary projection on which the scanning operation is based is
\begin{equation} \label{PaikDefinition}
\Big( |m\rangle \langle m|\Big)_i = \mbox{projection on age $m$ of record $i$}
\end{equation}
Below, fractional ages are converted to integers by the ceiling function $\lceil~\rceil$.
To enable the scanning of all combinations of ages and records, let us define
\begin{equation} \label{alphaDef}
\alpha = \mbox{sequence consisting of ages }\lceil 2^{-l}(j-1)a \rceil,
j = 1,\ldots,2^l, \mbox{ reordered}
\end{equation}
This distinguishes permutations of different ages, but not of equal ages.
Denote the collection of all sequences of the form (\ref{alphaDef}) by
$$
{\cal A}(l) = \big\{ \mbox{all possible $\alpha$ of the form (\ref{alphaDef})} \big\}
$$
The projection operator testing whether the records given by $g$ have ages as given by
$\alpha$ is
\begin{equation} \label{DefP(g,a)}
P(g,\alpha) =
\left[\prod_{j=1}^{2^l} \Big( |\alpha(j)\rangle \langle \alpha(j)|\Big)_{i(j)}\right]
\left[\prod_{i \notin g} \Big( |\mbox{blank}\rangle \langle \mbox{blank}|\Big)_i\right]
\end{equation}
where $i(j)$ in the first bracket denotes elements of $g$. These projectors are mutually
orthogonal,
\begin{equation} \label{OrthogonalP(g,a)}
P(g,\alpha) P(g',\alpha') = 0 ~~ \mbox{ if } g\neq g' \mbox{ or } \alpha\neq \alpha'
\end{equation}
To show this, consider $g \neq g'$. Let $i$ be an index in $g$ but not in $g'$. Then in
$P(g,\alpha)$ we have a projector $( |m\rangle \langle m|)_i$ while in $P(g',\alpha')$ we have
$( |\mbox{blank}\rangle \langle \mbox{blank}|)_i$ instead. The product of these two is zero
already. Secondly, consider the case of $g=g'$ and $\alpha\neq\alpha'$.
Let $j_0$ be an index for which $\alpha(j_0)\neq \alpha'(j_0)$. Now the projectors are
orthogonal because they project on different ages for record $i(j_0)$.
The idea for the modelling of observer's neuronal reaction is as follows.
In the subspace where the test by $P(g,\alpha)$ is positive (all $p$ give 1) the observer
notices it by some neural activity, and the next generation of the scanning process takes
place. In the subspace where the test is negative (some $p$ give 0) nothing happens; evolution
reduces to an identity operation. The neural activity is modelled by
2-dimensional rotations $\sigma_i$ in the counterparts ${\cal N}_i$ of the records.
In the subspace where $g$ tests positive, the collective rotations are, with obvious
assignment to the tensorial factors,
$$
\sigma(g) = \prod_{i \in g} \sigma_i
$$
The awareness cascade, running within a step of evolution, is conveniently constructed
recursively, downward from higher to lower resolutions of time. This is enabled by the fact
that after many generations the finite contents of the address sets ${\cal W}(k)$ will be
exhausted. So there is a maximum $L$ for the generation number $l$, determined by the other
parameters of the model. The counting of the generations will be upward here as usual,
beginning with $l=1$ at age $a/2$. The recursion is initialised by
\begin{equation} \label{UawaInitial}
U_\mathrm{awa}(L+1,a) = 1
\end{equation}
and proceeds by
\begin{equation} \label{UawaRecursion}
U_\mathrm{awa}(l,a) = 1 + \sum_{g\in{\cal G}(l)} \sum_{\alpha\in{\cal A}(l)}
\Big( -1 + U_\mathrm{awa}(l+1,a) \sigma(g) \Big) P(g,\alpha)
\end{equation}
The awareness operator for the completed cascade is $ U_\mathrm{awa}(1,a)$. Defining it by
recursion is only a convenient way of representing the algebraic structure. In application
to a state vector, projections of the various generations automatically occur in the natural
order, $l=1,\ldots,L$.
In order to show that (\ref{UawaRecursion}) indeed defines a unitary operator, first note
that $P(g,\alpha)$ only consists of projections on the ages of witnessing records, so that
basis states of the form (\ref{DefRecordBasisWitness}) are eigenstates of the projectors.
All age projectors commute among themselves, and commute with the $\sigma$
operations because these do not act on records. It follows, starting from
(\ref{UawaInitial}) and going through (\ref{UawaRecursion}), that the projectors commute
with the $U_\mathrm{awa}$ of all generations. Using (\ref{OrthogonalP(g,a)}), unitarity in the
form $U_\mathrm{awa}(l,a)^\dag U_\mathrm{awa}(l,a) = 1$ can then be shown by
straightforward algebra.
\subsection{Observer's age and conscious history\label{secObserversAge}}
Parameter $a$ of the preceding section is identified here as an eigenvalue of observer's
age operator. It suffices to assign an age to any tensor product of basis vectors as defined in
(\ref{DefRecordBasisWitness}). Assigning 0 to the ``blank'' state here, any basis state of
record $i$ has an age value $a_i$. The age operator is defined by
$$
A \prod_i |a_i\rangle = ( \max a_i) \prod_i |a_i\rangle
$$
Observer's lifetime can be taken into account by restricting neuronal activity to ages
$ a \leq T$, by including an operator factor $\Theta(T-A)$.
Thus, the final expression for the evolution operator of observer's consciousness is
\begin{equation} \label{DefUconsc}
U_\mathrm{con} = 1 + \Big(-1 + U_\mathrm{awa}(1,A)\Big)\Theta(T-A)
\end{equation}
The complete evolution operator is a product of the factors constructed above.
In a new step of evolution, ages of all records are increased by one unit.
Next, orbital records develop. The creation of witnessing records and their processing by the
observer (operations that do not commute) are assumed to be intertwined in such a way that
unitarity is manifestly preserved. The full evolution operator of the model is
\begin{equation} \label{DefUmodel}
U = U_\mathrm{wit\,2} ~ U_\mathrm{con} ~ U_\mathrm{wit\,1} ~ U_\mathrm{orb} ~
U_\mathrm{age}
\end{equation}
\subsection{Verifying the Scenario\label{ScenarioRecovered}}
\subsubsection{Structure of branches}
We start out from a product state as specified in section \ref{secInitialState} and repeatedly apply the evolution operator (\ref{DefUmodel}). Clearly, since product states form a basis, we can always write the resulting states as superpositions of products; however, we wish to show that only a superposition of special products emerges, which will be regarded as the ``branches'' of the wave function. After $a$ applications of the evolution operator $U$, the properties of those product states are as follows.
\begin{enumerate}
\item In each branch there is precisely one orbital record of age $0$.
\item For each of the ages $0,\ldots,a$, there is one set ${\cal W}(k)$ of
witnessing records in the corresponding eigenstates of age; all other witnesses
are blank.
\item Neuronal states and record states factorise (do not entangle).
\end{enumerate}
The initial state has these properties with $a=0$ by definition. Let us assume then that $a$
applications of $U$ have produced a superposition of product states with properties 1-3. When
$U$ is applied once more, it suffices by linearity to consider the action on any of the product
states. The first action is to increase by 1 the ages of all non-blank records.
There is now for each of the ages $1,\ldots,a+1$ precisely one set ${\cal W}(k)$ of
witnessing records in the corresponding eigenstates of age, while all other witnesses
are blank; witnesses of age $0$ are missing so far.
Also, a single orbital record of age 1 is generated from the previous one of age 0; let its
address be $k_1$. Now applying $U_\mathrm{orb}$, as defined in (\ref{DefUorb}), only the factor
with $k=k_1$ can have an effect since the projection on age 1 gives zero for all other $k$.
Nontrivial action of $U_\mathrm{orb}(k_1)$ requires all records in ${\cal B}(k_1)$ to be blank,
which would not be true if the system had been on any of those points before. Invoking the loop-
avoiding specification of ${\cal B}(k_1)$, as given in appendix \ref{secAvoidingLoops}, we can
regard this condition as satisfied within observer's lifetime. The action of
$U_\mathrm{orb}(k_1)$ then consists in creating a new superposition of products, with a single
zero-age orbital record in each of them, as expressed in (\ref{SplittingShorthand}). Property 1
holds in each of these products. For the remaining operations of $U$ it suffices by linearity
again to apply them only on the product states just created by $U_\mathrm{orb}(k_1)$.
Let $k_0$ be the single zero-age orbital record in one of them. Then, of all
witness-generating factors of (\ref{DefUwit}), only $U_\mathrm{wit\,1}(k_0)$ and
$U_\mathrm{wit\,2}(k_0)$ act nontrivially, due to the conditioning on zero-age.
The records of the sets ${\cal W}_1(k_0)$ and ${\cal W}_2(k_0)$ are in blank states before this
action, because orbital point $k_0$ was not visited before, so the simplified expression
(\ref{UwitSimpl}) applies, and the records of ${\cal W}_1(k_0)$ are transformed from blank to
zero-age states. Thus, $U_\mathrm{wit\,1}(k_0)$ generates the first half of zero-age witnessing
records that were missing so far from the full range of ages.
Next comes the action of the action of $U_\mathrm{con}$, defined in (\ref{DefUconsc}). It is
the only factor of evolution which could affect property 3. It
consists in making certain neuronal factors rotate if certain projections on the ages of
records are nonzero, and no action else. All records are in definite ages or
blank, so the projections of $U_\mathrm{con}$ preserve the product form of the state vector.
The neuronal rotations preserve the product form by construction. Hence, property 3 continues
to hold. Finally, the second half of zero-age witnessing records is generated by
$U_\mathrm{wit\,2}(k_0)$, so property 2 holds as well after $a+1$ applications of the
evolution operator.
\subsubsection{Awareness cascades}
To evaluate the awareness cascades encoded in $U_\mathrm{con}$, defined in
(\ref{DefUconsc}), assume that observer's age is $a<T$ so that $U_\mathrm{awa}(1,a)$
applies. The projection operators $P(g,\alpha)$ of (\ref{UawaRecursion}), using
(\ref{ikjDef}) and (\ref{alphaDef}) for $l=1$, test for randomly chosen records
in ${\cal W}(k(0))$ and ${\cal W}(k(1))$, with ages $0$ and $\lceil a/2\rceil$ or the
permutation of these, while the pair of $k(0)$ and $k(1)$ is ordered. By property 2 above,
the product state (or branch) being considered has records of one set ${\cal W}(k_0)$ at age
$0$ and of one set ${\cal W}(k_1)$ at age $\lceil a/2\rceil$. Hence,
the only successful projection $P(g,\alpha)$ can be for $g=(k_0,k_1)$ and
$\alpha=(0,\lceil a/2\rceil)$ or for $g=(k_1,k_0)$ and $\alpha=(\lceil a/2\rceil,0)$,
depending on which of the addresses $k_0$ or $k_1$ is smaller. Only one term, at most,
contributes to the sum over $g$ and $\alpha$ in (\ref{UawaRecursion}).
The test for $k_1$ with age $\lceil a/2\rceil$ will be positive, since all
witnesses of ages $1$ to $a$ have been created during the preceding steps of evolution.
However, witnesses of age $0$ are created half before the action of $U_\mathrm{con}$ and half
thereafter. If the randomly chosen record from ${\cal W}(k_0)$ is contained in the first half,
${\cal W}_1(k_0)$, it is created by $U_\mathrm{wit\,1}$ before the action of $U_\mathrm{con}$,
so the test will result in $P=1$ in equation (\ref{UawaRecursion}); if it is created by
$U_\mathrm{wit\,2}$ instead, the test will result in $P=0$. In the latter case,
$U_\mathrm{awa}(1,a)$ reduces to the identity operation.
In the case of $P=1$, the neurons associated with the witnesses for $k_0$ or $k_1$
become active through the $\sigma$ factor, and the second generation of the cascade comes into
action through $U_\mathrm{awa}(2,a)$.
The argument repeats: As a consequence of property 2, at most one combination $(g,\alpha)$
contributes to the sum (\ref{UawaRecursion}) for $U_\mathrm{awa}(2,a)$, namely that combination
in which the records collected in $g$ are tested for the ages they actually have on the branch
considered. The zero-age orbital record $k_0$ and its witnesses are the same for all
generations of the cascade, but for each $l$ a new randomly chosen witness is tested.
Projection $P(g,\alpha)$ reduces to $1$ if that witness happens to be created by
$U_\mathrm{wit\,1}$, while it reduces to $0$ if it is created by $U_\mathrm{wit\,2}$. The
probability for the cascade to continue is $1/2$ in every generation. Witnesses of higher age
always test positive, as they have been created in the preceding steps of evolution.
\subsubsection{Statistics of dimensions of conscious subspaces\label{DimensionStatistics}}
If the observer lives to age $T$, the number of orbital points on his world-tree is
$$
N = \frac{B^{T+1}-1}{B-1}
$$
This is also the number of statistically independent awareness cascades, as we now show.
By definition (\ref{ikjDef}), a new series of random draws is made for every
sequence $g$, a selection of orbital addresses. This definition does not a priori relate to a
specific time, but its occurrence in the projector $P(g,\alpha)$ of the evolution operator,
defined in (\ref{DefP(g,a)}), combines it with a sequence of observer's ages.
We intend to show the following: If projections with the same sequence, $g_1=g_2$, give
positive results for two points on observer's world-tree, those points must be equal.
Let $a_1$ and $a_2$ be observer's ages at the two points, and let $k_1$ and $k_2$ be the
orbital points of zero age, the ``present'' points. By (\ref{DefP(g,a)}) and (\ref{alphaDef}),
the present point is always contained in $g$, so $k_1\in g_1$ in particular. This here implies
$k_1\in g_2$. Since $P(g_2,\alpha_2)$ is assumed to test positive, $k_1$ must be an orbital
record on the branch leading to $k_2$, so it either coincides with $k_2$ or is a record of age
greater than zero. In the latter case, it must have been the zero-age record at an earlier
age of the observer. Hence, $a_1<a_2$, unless $k_1=k_2$. Exchanging 1 and 2 in the
argument, we find $a_1>a_2$ unless $k_1=k_2$. This implies $k_1=k_2$ and $a_1=a_2$.
For the number of neurons activated in a cascade (section \ref{secGenPowStat}) the
probability distribution is a power law characterised by exponent $-1$. Hence, by a theorem of
order statistics \cite{Embrechts1997}, the largest number of neurons activated
exceeds the second-largest by a quantity of order $N$. More precisely, using notation of
\cite{Embrechts1997} corollary 4.2.13, given an ensemble of size $N$ of random draws with the
power-law distribution, we have for the difference between the largest draw $X_{1,N}$ and the
second-largest $X_{2,N}$
\begin{equation} \label{FrechetSeparation}
X_{1,N}-X_{2,N} = N \, Y \qquad \mbox{$Y$ = random variable independent of $N$}
\end{equation}
The distribution of $Y$ is non-singular.
The dimension of the active neuronal subspace in the extremal branch is $2^{X_{1,N}}$.
The total dimension of active neuronal subspaces in all other branches is bounded by
$2^{X_{2,N}}N$. For the latter to be larger than the former, the probability is
$$
P(2^{NY}<N) = P(Y<N^{-1}\log_2N) = \mbox{negligible}
$$
By a comfortable margin, an observer can expect to find his world-line well-defined, providing
it is indeed the \emph{dimension} of awareness that matters, rather than the number of neurons.
Two arguments in favour of the dimension are at hand. The simplest is Fermi's Golden Rule,
although it relies on state reduction and so goes beyond the framework of the model; any
transition probabilities into observer's subspace of awareness would be proportional to the
dimension of the subspace.
The other argument uses a change of basis in the union of conscious subspaces of \emph{all}
branches. Let $|1\rangle,\ldots,|N\rangle$ be a basis for the extremal branch, and
$|N+1\rangle,\ldots,|N+M\rangle$ a basis for the remaining branches. We know that $M$ is tiny
in comparison to $N$. Now consider instead a Fourier basis, which consists of superpositions
of all $|1\rangle,\ldots,|N+M\rangle$ with equal amplitudes but different phases. In each of
the new basis vectors, properties of the non-extremal branches only occur in a tiny
component. The situation is now similar to that of an electron bound to a proton on earth.
It resides by $10^{-10^{18}}$ of its wavefunction behind the moon, but we still regard it as
an electron on earth.
\subsection{Analysing states in a time-local basis\label{secLocalBasis}}
In order to analyse the properties of a state vector, it must be represented in a particular basis, such as the eigenbasis of an observable. For some applications, like representing dynamics in the Heisenberg picture, the basis may conveniently be chosen time-dependent.
As to the state vectors of observer's neurons, we have so far used a global basis which applies to all branches, and in which the evolution operator is constant. In this way, observer's entire experience emerges in a single step of evolution near the end of his lifetime. Observer's mental reactions thus appear highly non-local.
However, when analysed in a time-dependent basis adapted to the evolution in one particular branch, observer's reactions occur simultaneously with the creation of the records, while the operator of evolution appears to change in a random way after each observation. This is trivial mathematically, but not physically.
Consider a section of evolution of a neuronal subsystem,
$$
\left(\begin{array}{c} x_1 \\ x_2 \end{array}\right) \stackrel{1}{\longrightarrow}
\left(\begin{array}{c} x_1 \\ x_2 \end{array}\right) \stackrel{\sigma}{\longrightarrow}
\left(\begin{array}{c} y_1 \\ y_2 \end{array}\right)
$$
where entries relate to some initially chosen basis, and where $\sigma$ is a unitary $2\times 2$ matrix. Only in the second step something appears to happen here. Changing the basis for the second state vector such that
$$
\left(\begin{array}{c} x_1 \\ x_2 \end{array}\right)_\mathrm{old~basis}
= \sigma^{-1}\left(\begin{array}{c} y_1 \\ y_2 \end{array}\right)_\mathrm{new~basis}
$$
the section of evolution takes the form
$$
\left(\begin{array}{c} x_1 \\ x_2 \end{array}\right) \stackrel{\sigma}{\longrightarrow}
\left(\begin{array}{c} y_1 \\ y_2 \end{array}\right) \stackrel{1}{\longrightarrow}
\left(\begin{array}{c} y_1 \\ y_2 \end{array}\right)
$$
The step of evolution where change appears can be shifted to any position in the
sequence. Due to the tensor-product structure of the model's branches, the argument applies
separately to all neurons involved. It is thus \emph{possible} to choose a basis in which
observer's neuronal reactions appear local, but it is a different choice in each branch, and
the reason for the choice cannot be found in the state of records at the given time. In this
sense, the choice appears to be intrinsically random.
\section{Conclusions\label{Conclusions}}
It has been demonstrated for a toy model of a quantum system with conscious observer,
that a unitary evolution operator, repeatedly applied to an appropriate initial state,
can accomplish two things: gather information about an observer's world-tree,
and perform a random draw on the world-tree so as to single out a world-line of extreme
awareness. A theorem, known from order statistics, about the dominance of the extreme in a
power-law ensemble plays a central role.
What the model \emph{avoids} to do is giving up fundamental linearity, and
introducing fundamental stochastics. The framework is state vectors and unitary evolution
under a constant law. Yet certain vectors evolve pseudo-stochastically.
The role of time and causality in the model is precarious, inevitably so under
the working hypothesis that a world-line should be determined by an extremal
draw on a world-tree.
As was shown in section \ref{secLocalBasis}, the model reproduces the usual
scenario of alternating Schr\"odinger-type and Born-type evolution when
represented in an appropriate basis. Observer's mental reactions then appear at the same
instant as the generation of records, but the choice of the basis appears indeterminate at
that instant. The model resolves that indeterminacy by omitting any erosion of
``witnessing records'', keeping them readable throughout observer's lifetime.
Can this be true for more realistic ``witnessing records'', or is this the point where
attempts at realistic modifications of the model must fail?
Non-universality of time could be an argument in favour of the optimistic alternative.
From Special Relativity Theory,
time is known to be observer-dependent, but only with negligible effects if
observers move at low speed. On this basis, time is treated as universal in nonrelativistic
quantum mechanics (likewise in the preceding sections of this paper).
But quantum mechanics provides its own path to special relativity, in the sense that it
enables pre-relativistic derivations of the Dirac equation \cite{Zee1991,Polley2001};
it might also provide its own version of observer-dependent time. The existence of two
modes of evolution, Schr\"odinger and Born, might be an indication of it.
Having recovered the stochastic appearance of measurements in section \ref{secLocalBasis}
by referring to a specific basis, we may have some freedom in reinterpreting the
evolution \emph{operator} as something more general. Since logics is always part of a natural
law, and conceptually more general, could the role of the operator be to generate a logical
structure of which time evolution is only a representative in a particular basis?
In elementary cases like those described by a Dirac equation, the law of motion is close to
mere logics of nearest neighbours \cite{Polley2001}, so ``space-time'' might indeed reduce
to ``space-logics''. For quantum systems with great complexity, like an
elementary system coupled to a conscious observer, logical implications might depend on many
conditions, and could be halted as long as some conditions were not met by the state vector.
In different ways, ``halted'' evolutions are also considered in other scenarios
of quantum measurement. Stapp \cite{Stapp1993} proposed an interaction between mind and matter
based on the quantum Zeno effect; it would keep observer's attention focussed on one outcome in
a measurement, but a side effect would be the halting of processes in matter under observation.
With consistent histories \cite{Griffiths2002}, there is a copy of Hilbert space assigned to
each of the times $t_1,\ldots,t_n$ of measurement. In the present model, an analogue of such
``history Hilbert spaces'' may be seen in the subspaces defined by the written states of
records at a time $t_k$.
The modelling of
consciousness by a coherent, logically conjunctive neuronal activity appears to be in
the \emph{spirit} of Integrated Information \cite{Tononi2008}. As a \emph{measure} of
consciousness, however, section \ref{DimensionStatistics} of the present paper
suggests to take the total dimension of the subspace of neuronal activity, which is very
different from the entropy-based measure suggested in \cite{Tononi2008}. By taking logarithms
of dimensions, an elementary relation like that of one subspace covering
the union of many other subspaces becomes almost invisible.
If the model scenario could indeed be extended to more realistic systems and
observers, it would suggest an easier intuitive look on state vectors, and on the persistent
problem of ``the Now'' \cite{Mermin2013}. Intuitively, states of superposition have always been
associated with potentialities for a quantum system, but the need for an actuality
seemed to make it an insufficient characterisation. The model scenario suggests to identify
actuality with that potentiality which involves an extremal degree of awareness.
It is generated in one step of logical evolution, so an observer's impression of his entire
experience as one shifting moment would seem less surprising.
|
\section{Description of the problem and main theorems}
The aim of this manuscript is to analyze nonlinear diffusion problems when
the diffusion coefficients in certain directions are going towards zero. We
consider a general nonlinear elliptic singularly perturbed problem which can
be considered as a generalization to some class of integro-differential
problem (see \cite{7}), let us begin by describing the linear part of the
problem as given in \cite{1} and \cite{7}. For $\Omega =\omega _{1}\times
\omega _{2}$ a bounded cylindrical domain of
\mathbb{R}
^{N}$ ($N\geq 2$) where $\omega _{1},\omega _{2}$ are Lipschitz domains of
\mathbb{R}
^{p}$ and
\mathbb{R}
^{N-p}$ respectively, we denote by $x=(x_{1},...,x_{N})=(X_{1},X_{2})$ the
points in
\mathbb{R}
^{N}$ wher
\begin{equation*}
X_{1}=(x_{1},...,x_{p})\text{ }\in \omega _{1}\text{and
X_{2}=(x_{p+1},...,x_{N})\in \omega _{2},
\end{equation*}
i.e. we split the coordinates into two parts. With this notation we se
\begin{equation*}
\nabla =(\partial _{x_{1}},...,\partial _{x_{N}})^{T}=\binom{\nabla _{X_{1}
}{\nabla _{X_{2}}},
\end{equation*}
where
\begin{equation*}
\nabla _{X_{1}}=(\partial _{x_{1}},...,\partial _{x_{p}})^{T}\text{ and
\nabla _{X_{2}}=(\partial _{x_{p+1}},...,\partial _{x_{N}})^{T}
\end{equation*}
To make it simple we use this abuse of notation
\begin{equation*}
\nabla _{X_{i}}u\in L^{2}(\Omega )\text{ instead of }\nabla _{X_{i}}u\in
\left[ L^{2}(\Omega )\right] ^{p;N-p}\text{ for a function }u
\end{equation*}
Let $A=(a_{ij}(x))$ be a $N\times N$ symmetric matrix which satisfies the
ellipticity assumptio
\begin{equation*}
\exists \lambda >0:A\xi \cdot \xi \geq \lambda \left\vert \xi \right\vert
^{2}\text{ }\forall \xi \in R^{N}\text{ for a.e }x\in \Omega ,
\end{equation*}
and
\begin{equation}
a_{ij}(x)\in L^{\infty }(\Omega ),\forall i,j=1,2,....,N,\text{ } \label{1}
\end{equation}
where $"\cdot "$ is the canonical scalar product on $R^{N}$. We decompose $A$
into four block
\begin{equation*}
A=\left(
\begin{array}{cc}
A_{11} & A_{12} \\
A_{21} & A_{22
\end{array
\right) ,
\end{equation*}
where $A_{11}$, $A_{22}$ are respectively $p\times p$ and $(N-p)\times (N-p)$
matrices. For $0<\epsilon \leq 1$ we se
\begin{equation*}
A_{\epsilon }=\left(
\begin{array}{cc}
\epsilon ^{2}A_{11} & \epsilon A_{12} \\
\epsilon A_{21} & A_{22
\end{array
\right) ,
\end{equation*
\rule{0pt}{0pt}
then we have therefore, for a.e. $x\in \Omega $ and every $\xi \in R^{N}
\begin{eqnarray}
A_{\epsilon }\xi \cdot \xi &\geq &\lambda \left( \epsilon ^{2}\left\vert
\overline{\xi _{1}}\right\vert ^{2}+\left\vert \overline{\xi _{2}
\right\vert ^{2}\right) \geq \lambda \epsilon ^{2}\left\vert \xi \right\vert
^{2}\text{ }\forall \xi \in R^{N}\text{ }, \label{3} \\
\text{and }A_{22}\overline{\xi _{2}}\cdot \overline{\xi _{2}} &\geq &\lambda
\left\vert \overline{\xi _{2}}\right\vert ^{2}, \notag
\end{eqnarray}
where we have se
\begin{equation*}
\xi =\binom{\overline{\xi _{1}}}{\overline{\xi _{2}}},
\end{equation*}
with
\begin{equation*}
\overline{\xi _{1}}=(\xi _{1},.....,\xi _{p})^{T}\text{ and }\overline{\xi
_{2}}=(\xi _{p+1},.....,\xi _{N})^{T}
\end{equation*}
And finally let $B:L^{2}(\Omega )\rightarrow L^{2}(\Omega )$ be a nonlinear
locally-Liptchitz operator i.e, for every bounded set $E\subset L^{2}(\Omega
)$ there exists $K_{E}\geq 0$ such that
\begin{equation}
\forall u,v\in E:\left\Vert B(u)-B(v)\right\Vert _{L^{2}(\Omega )}\leq
K_{E}\left\Vert u-v\right\Vert _{L^{2}(\Omega )}, \label{2}
\end{equation}
,and $B$ satisfies the growth conditio
\begin{equation}
\exists r>2,\text{ }M\geq 0,\text{ }\forall u\in L^{2}(\Omega ):\left\Vert
B(u)\right\Vert _{L^{r}(\Omega )}\leq M\left( 1+\left\Vert u\right\Vert
_{L^{2}(\Omega )}\right) \text{, } \label{4}
\end{equation}
We define the spac
\begin{equation*}
V=\left\{ u\in L^{2}(\Omega ):\nabla _{X_{1}}u\in L^{2}(\Omega )\right\}
\end{equation*}
Moreover we suppose that for every $E\subset V$ bounded in $L^{2}(\Omega )$
we hav
\begin{equation}
\overline{conv}\left\{ B\left( E\right) \right\} \subset V, \label{5}
\end{equation}
where $\overline{conv}\left\{ B\left( E\right) \right\} $ is the closed
convex hull of $B\left( E\right) $ in $L^{2}(\Omega ).$This last condition
is the most crucial, it will be used in the proof of the interior estimates
and the convergence theorem.
For $\beta >M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}$ we
consider the problem
\begin{equation}
\left\{
\begin{array}{cc}
\dint\limits_{\Omega }A_{\epsilon }\nabla u_{\epsilon }.\nabla \varphi
dx+\beta \dint\limits_{\Omega }u_{\epsilon }\varphi dx\text{ } &
=\dint\limits_{\Omega }B(u_{\epsilon })\varphi dx\text{, \ }\forall \varphi
\in \mathcal{D}(\Omega ) \\
u_{\epsilon }\in H_{0}^{1}(\Omega )\text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }
&
\end{array
\right. \label{6}
\end{equation}
The existence of $u_{\epsilon }$ will be proved in the next section, Now,
passing to the limit $\epsilon \rightarrow 0$ formally in (\ref{6}) we
obtain the limit problem
\begin{equation}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}u_{0}.\nabla _{X_{2}}\varphi
dx+\beta \dint\limits_{\Omega }u_{0}\varphi dx=\dint\limits_{\Omega
}B(u_{0})\varphi dx\text{, \ }\forall \varphi \in \mathcal{D}(\Omega )
\label{10}
\end{equation}
Our goal is to prove that $u_{0}$ exists and it satisfies (\ref{10}), and
give a sense to the formal convergence $u_{\epsilon }\leadsto u_{0}$,
actually we would like to obtain convergence in $L^{2}(\Omega ).$ We refer
to \cite{1} for more details about the linear theory of problem (\ref{6}).
However the nonlinear theory is poorly known, a monotone problem has been
solved in \cite{2} (using monotonicity argument), and also a case where $B$
is represented by an integral operator has been studied in \cite{7} (in the
last section of this paper, we shall give an application to
integro-differential problems). Generally, in singular perturbation problems
for PDEs, a simple analysis of the problem gives only weak convergences, and
often it is difficult to prove strong convergence, the principal hardness is
the passage to the limit in the nonlinear term. In this article we expose a
resolution method based on the use of several approximated problems
involving regularization with compact operators and truncations. Let us give
the main results.
\begin{theorem}
(Existence and $L^{r}$-regularity of solutions) Assume (\ref{1}), (\ref{3}),
(\ref{4}), and that $B$ is continuous on $L^{2}(\Omega )$ ( not necessarily
locally-Lipschitz) then ($\ref{6}$) has at least a solution $u_{\epsilon
}\in H_{0}^{1}(\Omega ).$ Moreover, if $u_{\epsilon }\in H_{0}^{1}(\Omega )$
is a solution to (\ref{6}) then $\left\Vert u_{\epsilon }\right\Vert
_{L^{r}(\Omega )}\leq \frac{M}{\beta -M\left\vert \Omega \right\vert ^{\frac
1}{2}-\frac{1}{r}}}$ for every $\epsilon >0.$
\end{theorem}
For the convergence theorem and the interior estimates we need the following
assumption
\begin{equation}
\text{ }\partial _{k}A_{22}\text{, }\partial _{i}a_{ij}\text{, }\partial
_{j}a_{ij}\in L^{\infty }(\Omega )\text{ \ }k=1,...,p,\text{\ \ \ }i=1,...,p
\text{ \ \ }j=p+1,...,N \label{34}
\end{equation}
\begin{theorem}
(Interior estimates) Assume (\ref{1}), (\ref{3}), (\ref{2}), (\ref{4}), (\re
{5}), (\ref{34})$.$ Let $(u_{\epsilon })\subset H_{0}^{1}(\Omega )$ be a
sequence of solutions to (\ref{6}) then for every open set$\ \Omega ^{\prime
}\subset \subset \Omega $ (i.e $\overline{\Omega ^{\prime }}\subset \Omega
) there exists $C_{\Omega ^{\prime }}\geq 0$ (independent of $\epsilon $)
such tha
\begin{equation*}
\forall \epsilon :\left\Vert u_{\epsilon }\right\Vert _{H^{1}(\Omega
^{\prime })}\leq C_{\Omega ^{\prime }}
\end{equation*}
\end{theorem}
\begin{theorem}
( The convergence theorem) Assume (\ref{1}), (\ref{3}), (\ref{2}), (\ref{4
), (\ref{5}), (\ref{34})$.$ Let $(u_{\epsilon })\subset H_{0}^{1}(\Omega )$
be a sequence of solutions to (\ref{6}) then there exists a subsequence
(u_{\epsilon _{k}})$ and $u_{0}\in H_{loc}^{1}(\Omega )\cap L^{2}(\Omega )$
such that : $\nabla _{X_{2}}u_{0}\in L^{2}(\Omega )$ and
\begin{equation*}
u_{\epsilon _{k}}\rightarrow u_{0},\nabla _{X_{2}}u_{\epsilon
_{k}}\rightarrow \nabla _{X_{2}}u_{0}\text{ in }L^{2}(\Omega )\text{
strongly as }\epsilon _{k}\rightarrow 0
\end{equation*}
and for a.e $X_{1}$ we have $u_{0}(X_{1},.)\in H_{0}^{1}(\omega _{2}),$and
\begin{eqnarray}
&&\dint\limits_{\omega _{2}}A_{22}\nabla _{X_{2}}u_{0}(X_{1},.)\cdot \nabla
_{X_{2}}\varphi dX_{2}+\beta \dint\limits_{\omega _{2}}u_{0}(X_{1},.)\varphi
dX_{2} \notag \\
&=&\dint\limits_{\omega _{2}}B(u_{0})(X_{1},.)\varphi dX_{2},\text{ \
\forall \varphi \in \mathcal{D}(\omega _{2}) \label{11}
\end{eqnarray}
\end{theorem}
\begin{corollary}
If problem (\ref{11}) has a unique solution ( in the sense of \textbf
theorem\ 3}) then the convergences given in the previous theorem hold for
the whole sequence $(u_{\epsilon }).$
\end{corollary}
\begin{proof}
The proof is direct, let $(u_{\epsilon })$ be a sequence of solutions to
\ref{6}) and suppose that $u_{\epsilon }$ does not converge to $u_{0}$ (as
\epsilon \rightarrow 0$) then there exists a subsequence $(u_{\epsilon _{k}})
$ and $\delta >0$ such that $\forall \epsilon _{k},$ $\left\Vert u_{\epsilon
_{k}}-u_{0}\right\Vert _{L^{2}(\Omega )}>$ $\delta $ or $\left\Vert \nabla
_{X_{2}}(u_{\epsilon _{k}}-u_{0})\right\Vert _{L^{2}(\Omega )}>$ $\delta $.
By \textbf{theorem 3} one can extract a subsequence of $(u_{\epsilon _{k}})$
which converges to some $u_{1}$ in the sense of \textbf{theorem 3}, assume
that (\ref{11}) has a unique solution then $u_{1}=u_{0}.$and this
contradicts the previous inequalities.
\end{proof}
In the case of non-uniqueness we can reformulate the convergences ,given in
the previous theorem, using $\epsilon -nets$ like in \cite{7}. Let us recall
the definition of $\epsilon -nets$ (\cite{7})
\begin{definition}
Let $(X,d)$ be metric space, $Y,Y^{\prime }$ two subsets of \ $X,$ then we
say that $Y$ is an $\epsilon -net$ of $Y^{\prime },$ if \ for every $x\in
Y^{\prime }$ there exists an $a\in Y$ such tha
\begin{equation*}
d(x,a)<\epsilon
\end{equation*}
\end{definition}
We define the following space introduced in \cite{7}
\begin{equation*}
W=\left\{ u\in L^{2}(\Omega ):\nabla _{X_{2}}u\in L^{2}(\Omega ),\text{ and
for a.e }X_{1}\text{, }u(X_{1},.)\in H_{0}^{1}(\omega _{2})\text{ }\right\} ,
\end{equation*}
equipped with the Hilbertian norm (see \cite{7}
\begin{equation*}
\left\Vert u\right\Vert _{W}^{2}=\left\Vert u\right\Vert _{L^{2}(\Omega
)}^{2}+\left\Vert \nabla _{X_{2}}u\right\Vert _{L^{2}(\Omega )}^{2}
\end{equation*}
Now we can give \textbf{Theorem 3 }in the following form
\begin{theorem}
Under assumptions of theorem 3 then $\Xi $ ,the set of solutions of (\ref{11
) in $W,$ is non empty and we have $\Xi \cap H_{loc}^{1}(\Omega )\neq
\varnothing $, and moreover for every $\eta >0,$ there exists $\epsilon
_{0}>0$ such that $\Xi $ is an $\eta -net$ of $\Xi _{\epsilon _{0}}$ in $W$
wher
\begin{equation*}
\Xi _{\epsilon _{0}}=\left\{ u_{\epsilon }\text{ solution to (\ref{6}) for
0<\epsilon <\epsilon _{0}\right\}
\end{equation*}
\end{theorem}
\begin{proof}
\textbf{Theorem 1} and\textbf{\ 3} ensure that $\Xi \cap H_{loc}^{1}(\Omega
)\neq \varnothing .$ For the $\eta -net$ convergence, let us reasoning by
contradiction, then there exists $\eta >0$ and a sequence $\epsilon
_{k}\rightarrow 0$ such that $\Xi $ is not an $\eta -net$ of $\Xi _{\epsilon
_{k}}$ in $W$ for every $k$ ( remark that $\Xi _{\epsilon _{k}}\neq
\varnothing $ by \textbf{Theorem 1)} in other words there exists a sequence
(u_{\epsilon _{k}^{\prime }})$ with $\epsilon _{k}^{\prime }\rightarrow 0$
such that for every $u_{0}\in \Xi $ we have $\left\Vert u_{\epsilon
_{k}^{\prime }}-u_{0}\right\Vert _{W}\geq \eta $, according to theorem 3
there exists a subsequence of $(u_{\epsilon _{k}^{\prime }})$ which
converges to some $u_{0}\in \Xi $ in$W$ and this contradicts the previous
inequality.
\end{proof}
\section{Existence and $L^{r}-regularity$\ for the solutions and weak
convergences}
\subsection{\textbf{Existence and }$L^{r}-regularity$}
In this subsection we prove \textbf{Theorem 1, }we start by the following
result on the $L^{r}$- regularity for the solutions
\begin{proposition}
Assume (\ref{1}), (\ref{3}), (\ref{4} ) then if $u_{\epsilon }\in
H_{0}^{1}(\Omega )$ is a solution to (\ref{6}) then $u_{\epsilon }\in
L^{r}(\Omega )$ and $\left\Vert u_{\epsilon }\right\Vert _{L^{r}(\Omega
)}\leq \frac{M}{\beta -M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}
r}}}$ for every $\epsilon >0$
\end{proposition}
\begin{proof}
We will proceed as in \cite{3}. Let $u_{\epsilon }\in H_{0}^{1}(\Omega )$ be
a solution to (\ref{6}), given $g\in \mathcal{D}(\Omega )$ and let
w_{\epsilon }\in H_{0}^{1}(\Omega )$ be the unique solution to the linear
problem
\begin{equation}
\dint\limits_{\Omega }A_{\epsilon }\nabla w_{\epsilon }\cdot \nabla \varphi
dx+\beta \dint\limits_{\Omega }w_{\epsilon }\varphi dx\text{
=\dint\limits_{\Omega }g\varphi dx\text{, \ }\forall \varphi \in \mathcal{D
(\Omega ), \label{7}
\end{equation}
the existence of $w_{\epsilon }$ follows by the Lax-Milgram theorem (thanks
to assumptions (\ref{1}), (\ref{3})).
Take $u_{\epsilon }$ as a test function and using the symmetry of
A_{\epsilon }$ we get
\begin{eqnarray*}
\dint\limits_{\Omega }u_{\epsilon }gdx &=&\dint\limits_{\Omega }A_{\epsilon
}\nabla w_{\epsilon }\cdot \nabla u_{\epsilon }dx+\beta \dint\limits_{\Omega
}w_{\epsilon }u_{\epsilon }dx \\
&=&\dint\limits_{\Omega }A_{\epsilon }\nabla u_{\epsilon }\cdot \nabla
w_{\epsilon }dx+\beta \dint\limits_{\Omega }w_{\epsilon }u_{\epsilon }dx \\
&=&\dint\limits_{\Omega }B(u_{\epsilon })w_{\epsilon }dx.
\end{eqnarray*}
Given $s$ such that $\frac{1}{r}+\frac{1}{s}=1$, then by (\ref{4}) we obtain
\begin{equation}
\left\vert \dint\limits_{\Omega }u_{\epsilon }gdx\right\vert \leq
M(1+\left\Vert u_{\epsilon }\right\Vert _{L^{2}(\Omega )})\left\Vert
w_{\epsilon }\right\Vert _{L^{s}(\Omega )} \label{8}
\end{equation}
Now we have to estimate $\left\Vert w_{\epsilon }\right\Vert _{L^{s}(\Omega
)}.$ Let $\rho \in C^{1}
\mathbb{R}
\mathbb{R}
)$, such that $\rho (0)=0$ and $\rho ^{\prime }\geq 0$ and $\rho ^{\prime
}\in L^{\infty }$ then $\rho (w_{\epsilon })\in H_{0}^{1}(\Omega ),$ take
\rho (w_{\epsilon })$ as a test function in (\ref{7}) we ge
\begin{equation*}
\dint\limits_{\Omega }\rho ^{\prime }(w_{\epsilon })A_{\epsilon }\nabla
w_{\epsilon }\cdot \nabla w_{\epsilon }dx+\beta \dint\limits_{\Omega
}w_{\epsilon }\rho (w_{\epsilon })dx\text{ }=\dint\limits_{\Omega }g\rho
(w_{\epsilon })dx\text{.}
\end{equation*}
Now, using ellipticity assumption (\ref{3}) we deriv
\begin{equation*}
\lambda \left( \dint\limits_{\Omega }\rho ^{\prime }(w_{\epsilon
})\left\vert \epsilon \nabla _{X_{1}}w_{\epsilon }\right\vert
{{}^2
dx+\dint\limits_{\Omega }\rho ^{\prime }(w_{\epsilon })\left\vert \nabla
_{X_{2}}w_{\epsilon }\right\vert
{{}^2
dx\right) +\beta \dint\limits_{\Omega }w_{\epsilon }\rho (w_{\epsilon })d
\text{ }\leq \dint\limits_{\Omega }g\rho (w_{\epsilon })dx
\end{equation*}
Thus
\begin{equation*}
\beta \dint\limits_{\Omega }w_{\epsilon }\rho (w_{\epsilon })dx\text{ }\leq
\dint\limits_{\Omega }g\rho (w_{\epsilon })dx
\end{equation*}
Assume that $\forall x\in
\mathbb{R}
:$ $\left\vert \rho (x)\right\vert \leq \left\vert x\right\vert ^{\frac{1}
r-1}},$ so that $\left\vert \rho (x)\right\vert ^{r}\leq \left\vert
x\right\vert \left\vert \rho (x)\right\vert =x\rho (x)$ then, we obtai
\begin{eqnarray*}
\beta \dint\limits_{\Omega }w_{\epsilon }\rho (w_{\epsilon })dx\text{ }
&\leq &\left\Vert g\right\Vert _{L^{s}(\Omega )}\left( \dint\limits_{\Omega
}\left\vert \rho (w_{\epsilon })\right\vert ^{r}dx\right) ^{\frac{1}{r}} \\
&\leq &\left\Vert g\right\Vert _{L^{s}(\Omega )}\left( \dint\limits_{\Omega
}w_{\epsilon }\rho (w_{\epsilon })dx\text{ }\right) ^{\frac{1}{r}},
\end{eqnarray*}
then
\begin{equation*}
\beta \left( \dint\limits_{\Omega }w_{\epsilon }\rho (w_{\epsilon })\right)
^{\frac{1}{s}}\leq \left\Vert g\right\Vert _{L^{s}(\Omega )}
\end{equation*}
Now, for $\delta >0$ taking $\rho (x)=x(x^{2}+\delta )^{\frac{s-2}{2}}$ we
show easily that $\rho $ satisfies the above assumptions, so we obtain
\begin{equation*}
\beta \left( \dint\limits_{\Omega }w_{\epsilon }^{2}(w_{\epsilon
}^{2}+\delta )^{\frac{s-2}{2}}\right) ^{\frac{1}{s}}\leq \left\Vert
g\right\Vert _{L^{s}(\Omega )},
\end{equation*}
let $\delta \rightarrow 0$ by Fatou's lemma we get
\begin{equation*}
\beta \left\Vert w_{\epsilon }\right\Vert _{L^{s}(\Omega )}\leq \left\Vert
g\right\Vert _{L^{s}(\Omega )}
\end{equation*}
Finally by (\ref{8}) we get
\begin{equation}
\left\vert \dint\limits_{\Omega }u_{\epsilon }gdx\right\vert \leq \frac
M(1+\left\Vert u_{\epsilon }\right\Vert _{L^{2}(\Omega )})}{\beta
\left\Vert g\right\Vert _{L^{s}(\Omega )} \notag
\end{equation}
By density we can take $g\in L^{s}(\Omega )$ and therefore by duality we get
\begin{equation*}
\left\Vert u_{\epsilon }\right\Vert _{L^{r}(\Omega )}\leq \frac
M(1+\left\Vert u_{\epsilon }\right\Vert _{L^{2}(\Omega )})}{\beta },
\end{equation*}
hence by Holder's inequality we obtai
\begin{equation*}
\left\Vert u_{\epsilon }\right\Vert _{L^{r}(\Omega )}\leq \frac{M}{\beta }
\frac{M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}{\beta
\left\Vert u_{\epsilon }\right\Vert _{L^{r}(\Omega )},
\end{equation*}
the
\begin{equation*}
\left\Vert u_{\epsilon }\right\Vert _{L^{r}(\Omega )}\leq \frac{M}{\beta
-M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}
\end{equation*}
\end{proof}
Now, it remains to prove the existence of $u_{\epsilon }$\textbf{, }the
proof is based on the Schauder fixed point theorem. Let $v\in L^{2}(\Omega )$
and $v_{\epsilon }\in H_{0}^{1}(\Omega )$ be the unique solution to the
linearized problem
\begin{equation}
\dint\limits_{\Omega }A_{\epsilon }\nabla v_{\epsilon }\cdot \nabla \varphi
dx+\beta \dint\limits_{\Omega }v_{\epsilon }\varphi dx\text{
=\dint\limits_{\Omega }B(v)\varphi dx\text{, \ }\forall \varphi \in \mathcal
D}(\Omega ) \label{9}
\end{equation}
The existence of $v_{\epsilon }$ follows by the Lax-Milgram theorem ( thanks
to assumptions (\ref{1}), (\ref{3})). Let $\Gamma :L^{2}(\Omega )\rightarrow
L^{2}(\Omega )$ be the mapping defined by $\Gamma (v)=v_{\epsilon }$.We
prove that $\Gamma $ is continuous, fix $v\in L^{2}(\Omega )$ and let
v_{n}\rightarrow v$ in $L^{2}(\Omega )$, we note $v_{\epsilon }^{n}=\Gamma
(v_{n})$ then we hav
\begin{equation*}
\dint\limits_{\Omega }A_{\epsilon }\nabla (v_{\epsilon }^{n}-v_{\epsilon
})\cdot \nabla \varphi dx+\beta \dint\limits_{\Omega }(v_{\epsilon
}^{n}-v_{\epsilon })\varphi dx\text{ }=\dint\limits_{\Omega }\left(
B(v_{n})-B(v)\right) \varphi dx\text{, \ }\forall \varphi \in \mathcal{D
(\Omega )\text{ }
\end{equation*}
Take $(v_{\epsilon }^{n}-v_{\epsilon })$ as a test function, estimating
using ellipticity assumption (\ref{3}) and Holder's inequality we ge
\begin{equation*}
\beta \left\Vert v_{\epsilon }^{n}-v_{\epsilon }\right\Vert _{L^{2}(\Omega
)}\leq \left\Vert B(v_{n})-B(v)\right\Vert _{L^{2}(\Omega )}.
\end{equation*}
Passing to the limit as $n\rightarrow \infty $ and assume that $B$ is
continuous,\ then the continuity of $\Gamma $ follows. Now, we define the set
\begin{equation*}
S=\left\{ v\in H_{0}^{1}(\Omega ):\left\Vert \nabla v\right\Vert
_{L^{2}(\Omega )}\leq \frac{\sqrt{\beta }}{\epsilon \sqrt{2\lambda }}\left(
\frac{M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}{\beta
-M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}\right) \text{
and }\left\Vert v\right\Vert _{L^{2}(\Omega )}\leq \frac{M\left\vert \Omega
\right\vert ^{\frac{1}{2}-\frac{1}{r}}}{\beta -M\left\vert \Omega
\right\vert ^{\frac{1}{2}-\frac{1}{r}}}\right\}
\end{equation*
It is clear that $S$ is a convex bounded set in $H_{0}^{1}(\Omega )$ and it
is closed in $L^{2}(\Omega )$, then $S$ is compact in $L^{2}(\Omega )$
(thanks to the compact Sobolev embedding\ $H_{0}^{1}(\Omega )\hookrightarrow
L^{2}(\Omega )$). Let us check that $S$ is stable by $\Gamma .$ For $v\in S,$
taking $\varphi =v_{\epsilon }$ in (\ref{9}) and estimating using
ellipticity assumption (\ref{3}) and H\"{o}lder's inequality we get
\begin{equation*}
\lambda \epsilon ^{2}\left\Vert \nabla v_{\epsilon }\right\Vert
_{L^{2}(\Omega )}^{2}+\beta \left\Vert v_{\epsilon }\right\Vert
_{L^{2}(\Omega )}^{2}\leq \left\Vert B(v)\right\Vert _{L^{2}(\Omega
)}\left\Vert v_{\epsilon }\right\Vert _{L^{2}(\Omega )},
\end{equation*}
then by Young's inequality we derive
\begin{equation*}
\lambda \epsilon ^{2}\left\Vert \nabla v_{\epsilon }\right\Vert
_{L^{2}(\Omega )}^{2}+\beta \left\Vert v_{\epsilon }\right\Vert
_{L^{2}(\Omega )}^{2}\leq \frac{1}{2\beta }\left\Vert B(v)\right\Vert
_{L^{2}(\Omega )}^{2}+\frac{\beta }{2}\left\Vert v_{\epsilon }\right\Vert
_{L^{2}(\Omega )}^{2},
\end{equation*
and (\ref{4}) give
\begin{eqnarray*}
\lambda \epsilon ^{2}\left\Vert \nabla v_{\epsilon }\right\Vert
_{L^{2}(\Omega )}^{2}+\frac{\beta }{2}\left\Vert v_{\epsilon }\right\Vert
_{L^{2}(\Omega )}^{2} &\leq &\frac{\left\vert \Omega \right\vert ^{1-\frac{
}{r}}(M+M\left\Vert v\right\Vert _{L^{2}(\Omega )})^{2}}{2\beta } \\
&\leq &\frac{\left\vert \Omega \right\vert ^{1-\frac{2}{r}}\left( M+\frac
M^{2}\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}{\beta
-M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}\right) ^{2}}
2\beta }\leq \frac{\beta }{2}\left( \frac{M\left\vert \Omega \right\vert ^
\frac{1}{2}-\frac{1}{r}}}{\beta -M\left\vert \Omega \right\vert ^{\frac{1}{2
-\frac{1}{r}}}\right) ^{2},
\end{eqnarray*}
hence
\begin{equation*}
\left\{
\begin{array}{c}
\left\Vert v_{\epsilon }\right\Vert _{L^{2}(\Omega )}\leq \frac{M\left\vert
\Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}{\beta -M\left\vert \Omega
\right\vert ^{\frac{1}{2}-\frac{1}{r}}} \\
\left\Vert \nabla v_{\epsilon }\right\Vert _{L^{2}(\Omega )}\leq \frac{\sqrt
\beta }}{\epsilon \sqrt{2\lambda }}\left( \frac{M\left\vert \Omega
\right\vert ^{\frac{1}{2}-\frac{1}{r}}}{\beta -M\left\vert \Omega
\right\vert ^{\frac{1}{2}-\frac{1}{r}}}\right
\end{array
\right.
\end{equation*}
And therefore $v_{\epsilon }=\Gamma (v)\in S.$Whence, there exists at least
a fixed point $u_{\epsilon }\in S$ for $\Gamma ,$ in other words
u_{\epsilon }$ is a solution to (\ref{6}).
\subsection{\textbf{Weak convergences as }$\protect\epsilon \rightarrow 0$}
Throughout this article we use the notations $\rightharpoonup $ ,
\rightarrow $ for weak and strong convergences of sequences respectively.
Assume (\ref{1}), (\ref{3}), (\ref{4} ) and let $(u_{\epsilon })$ be a
sequence of solutions to (\ref{6})\textbf{, }$.$ We begin by a simple
analysis of the problem, considering problem (\ref{6}) and taking $\varphi
=u_{\epsilon }\in H_{0}^{1}(\Omega )$, by ellipticity assumption (\ref{3})
we get
\begin{equation*}
\lambda \left( \dint\limits_{\Omega }\left\vert \epsilon \nabla
_{X_{1}}u_{\epsilon }\right\vert
{{}^2
dx+\dint\limits_{\Omega }\left\vert \nabla _{X_{2}}u_{\epsilon }\right\vert
{{}^2
dx\right) +\beta \dint\limits_{\Omega }u_{\epsilon }^{2}dx\leq
\dint\limits_{\Omega }B(u_{\epsilon })u_{\epsilon }dx\text{ ,\ \ }
\end{equation*
and H\"{o}lder's inequality gives
\begin{equation*}
\lambda \epsilon ^{2}\left\Vert \nabla _{X_{1}}u_{\epsilon }\right\Vert
_{L^{2}(\Omega )}^{2}+\lambda \left\Vert \nabla _{X_{2}}u_{\epsilon
}\right\Vert _{L^{2}(\Omega )}^{2}+\beta \left\Vert u_{\epsilon }\right\Vert
_{L^{2}(\Omega )}^{2}\leq \left\Vert B(u_{\epsilon })\right\Vert
_{L^{2}(\Omega )}\left\Vert u_{\epsilon }\right\Vert _{L^{2}(\Omega )},
\end{equation*
and therefore (\ref{4}) and \textbf{Proposition 1 }give\textbf{\
\begin{equation*}
\lambda \epsilon ^{2}\left\Vert \nabla _{X_{1}}u_{\epsilon }\right\Vert
_{L^{2}(\Omega )}^{2}+\lambda \left\Vert \nabla _{X_{2}}u_{\epsilon
}\right\Vert _{L^{2}(\Omega )}^{2}+\beta \left\Vert u_{\epsilon }\right\Vert
_{L^{2}(\Omega )}^{2}\leq \frac{M^{2}\left\vert \Omega \right\vert ^{1-\frac
2}{r}}}{\beta -M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}
\left( 1+\frac{M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}
\beta -M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}\right)
\end{equation*
Whenc
\begin{equation}
\left\{
\begin{array}{c}
\left\Vert \epsilon \nabla _{X_{1}}u_{\epsilon }\right\Vert _{L^{2}(\Omega
)}\leq \frac{C}{\sqrt{\lambda }} \\
\left\Vert \nabla _{X_{2}}u_{\epsilon }\right\Vert _{L^{2}(\Omega )}\leq
\frac{C}{\sqrt{\lambda }} \\
\left\Vert u_{\epsilon }\right\Vert _{L^{2}(\Omega )}\leq \frac{M\left\vert
\Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}{\beta -M\left\vert \Omega
\right\vert ^{\frac{1}{2}-\frac{1}{r}}
\end{array
\right. \label{26}
\end{equation
$\ \ $,where $C^{2}=\frac{M^{2}\left\vert \Omega \right\vert ^{1-\frac{2}{r}
}{\beta -M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}\left( 1
\frac{M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}{\beta
-M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}\right) $. Remark
that the gradient of $u_{\epsilon }$ is not bounded uniformly in
H_{0}^{1}(\Omega )$ so we cannot obtain strong convergence (using Sobolev
embedding for example) in $L^{2}(\Omega )$, however there exists a
subsequence $(u_{\epsilon _{k}})$ and $u_{0}\in L^{2}(\Omega )$ such that:
u_{\epsilon _{k}}\rightharpoonup u_{0}$ , $\nabla _{X_{2}}u_{\epsilon
_{k}}\rightharpoonup \nabla _{X_{2}}u_{0}$ and $\epsilon _{k}\nabla
_{X_{1}}u_{\epsilon _{k}}\rightharpoonup 0$ \ weakly in $L^{2}(\Omega )$ (we
used weak compacity in $L^{2}(\Omega )$, and the continuity of the operator
of derivation on $\mathcal{D}^{\prime }(\Omega )$). The function $u_{0}$
constructed before represents a good candidate for solution to the limit
problems (\ref{10}),(\ref{11}).
\begin{corollary}
We have $u_{0}\in L^{r}(\Omega ).$
\end{corollary}
\begin{proof}
Since $(u_{\epsilon _{k}})$ is bounded in $L^{r}(\Omega )$ then one can
extract a subsequence noted always $(u_{\epsilon _{k}})$ which converges
weakly to some $u_{1}\in L^{r}(\Omega )$ and therefore $u_{\epsilon
_{k}}\rightharpoonup u_{1}$ in $\mathcal{D}^{\prime }(\Omega )$, so
u_{1}=u_{0}$
\end{proof}
\section{\protect\bigskip\ Interior estimates and $H_{loc}^{1}-regularity$}
For every $g\in V$ consider the linear problem (\ref{7}), then one can prove
the
\begin{theorem}
Assume (\ref{1}), (\ref{3}), (\ref{34}) then for every $\ \Omega ^{\prime
}\subset \subset \Omega $ (i.e $\overline{\Omega ^{\prime }}\subset \Omega
) there exists $C_{\Omega ^{\prime },g}\geq 0$ independent of $\epsilon $
such tha
\begin{equation}
\forall \epsilon :\left\Vert v_{\epsilon }\right\Vert _{H^{1}(\Omega
^{\prime })}\leq C_{\Omega ^{\prime },g} \label{23}
\end{equation}
\end{theorem}
\begin{proof}
The proof is the same as in \cite{1} (see the rate estimations theorem in
\cite{1}), remark that the additional term $\beta v_{\epsilon }$ is
uniformly bounded in $L^{2}(\Omega ).$
\end{proof}
To obtain interior estimates for the nonlinear problem we use the well known
Banach-Steinhaus's theorem
\begin{theorem}
\bigskip (see \cite{5}) Let $Y$ and $Z$ be two separated topological vector
spaces, and let $(\mathcal{A}_{\epsilon })$ be a family of continuous linear
mappings from $Y\rightarrow Z$ , $G$ is convex compact set in $Y$. Suppose
that for each $x\in G$ the orbit $\left\{ \mathcal{A}_{\epsilon }(x)\right\}
_{\epsilon }$ is bounded in $Z,$ then $(\mathcal{A}_{\epsilon })$ is
uniformly bounded on $G,$ i.e. there exists a bounded \ $F$ set in $Z$ such
that $\forall \epsilon ,$ $\mathcal{A}_{\epsilon }(G)\subset F.$
\end{theorem}
Now, we are ready to prove \textbf{Theorem 2. }Let $(\Omega _{j})_{j\in
\mathbb{N}
}$ ,($\forall j:\overline{\Omega _{j}}\subset \Omega _{j+1}$) be an open
covering of $\Omega $, so we can define a family $(p_{j})_{j}$ of seminorms
on $H_{loc}^{1}(\Omega )$ by{}{}\newline
\begin{equation*}
p_{j}(u)=\left\Vert u\right\Vert _{H^{1}(\Omega _{j})}\text{ for every }u\in
H_{loc}^{1}(\Omega )
\end{equation*}
Set $Z=(H_{loc}^{1}(\Omega ),(p_{j})_{j}),$ we can check easily that $Z$ is
a separated locally convex topological vector space where the topology is
generated by the family of seminorms $(p_{j})_{j}$, we also set
Y=L^{2}(\Omega )$. We define a family $(\mathcal{A}_{\epsilon })_{\epsilon }$
of linear mappings from $Y$ to $Z$ \ by $\mathcal{A}_{\epsilon
}(g)=v_{\epsilon }$ where $v_{\epsilon }$ is the unique solution to (\ref{7
) (existence and uniqueness follows by Lax-Milgram, thanks to (\ref{1}),
\ref{3})). $\forall \epsilon $, $\mathcal{A}_{\epsilon }:Y\rightarrow Z$ is
continuous (we can check easily that $\mathcal{A}_{\epsilon }:Y\rightarrow
H^{1}(\Omega )$ and the injection $H^{1}(\Omega )\hookrightarrow Z$ are
continuous). We note $Z_{w}$, $Y_{w}$ the spaces $Z$ and $Y$ equipped with
the weak topology, then for every $\epsilon ,$ $\mathcal{A}_{\epsilon
}:Y_{w}\rightarrow Z_{w}$ is still continuous. Let $E=\left\{ u\in
V:\left\Vert u\right\Vert _{L^{2}(\Omega )}\leq \frac{M\left\vert \Omega
\right\vert ^{\frac{1}{2}-\frac{1}{r}}}{\beta -M\left\vert \Omega
\right\vert ^{\frac{1}{2}-\frac{1}{r}}}\right\} $ and assume (\ref{5}) then
G=\overline{conv}\left\{ B\left( E\right) \right\} \subset V$, it is clear
that $G$ is bounded in $Y$ thus $G$ is compact in $Y_{w}.$ Recall that a set
is bounded in a locally convex topological space if and only if the
seminorms that generate the topology are bounded on this set, suppose (\re
{34}) then according to (\ref{23}) we have, for $g\in G$, $\left\{ \mathcal{
}_{\epsilon }(g)\right\} _{\epsilon }$ is bounded in $Z$, and therefore
\left\{ \mathcal{A}_{\epsilon }(g)\right\} _{\epsilon }$ is bounded in
Z_{w} $ so by \textbf{Theorem 6} there exists a bounded set $F$ in $Z_{w}$
(also note that $F$ is also bounded in $Z$) such that $\forall \epsilon ,$
\mathcal{A}_{\epsilon }(G)\subset F$. Now let $(u_{\epsilon })$ be a
sequence of solutions to (\ref{6}), and assume in addition (\ref{2}) and
\ref{4}) then (\ref{26}) gives\ $(u_{\epsilon })_{\epsilon }\subset E$
whence $(B(u_{\epsilon }))_{\epsilon }\subset G$, and therefore $\mathcal{A
_{\epsilon }(B(u_{\epsilon }))\subset F$ for every $\epsilon $, in other
words we have
\begin{equation*}
\forall j\text{ },\text{ }\exists C_{j}\geq 0\text{ such that }\forall
\epsilon :p_{j}(\mathcal{A}_{\epsilon }(B(u_{\epsilon })))\leq C_{j},
\end{equation*}
where $C_{j}$ is independent of $\epsilon $, and therefor
\begin{equation*}
\forall \epsilon ,\forall j,\left\Vert u_{\epsilon }\right\Vert
_{H^{1}(\Omega _{j})\text{ }}\leq C_{j}
\end{equation*}
Now, given $\Omega ^{\prime }\subset \subset \Omega $ then there exists $j$
such that $\Omega ^{\prime }\subset \Omega _{j}$ thu
\begin{equation}
\forall \epsilon ,\left\Vert u_{\epsilon }\right\Vert _{H^{1}(\Omega
^{\prime })\text{ }}\leq C_{j} \label{24}
\end{equation}
\begin{corollary}
Let $(u_{\epsilon })\subset H_{0}^{1}(\Omega )$ be a sequence of solutions
to (\ref{6}) such that $u_{\epsilon }\rightharpoonup $ $u_{0}$ in
L^{2}(\Omega )$ weakly, then under assumptions of \textbf{Theorem 2 }we
have, $u_{0}\in H_{loc}^{1}(\Omega )$
\end{corollary}
\begin{proof}
take $\Omega ^{\prime }\subset \subset \Omega $ an open set, and $\psi \in
\mathcal{D}(\Omega ^{\prime }),$ $1\leq i\leq N$ then by (\ref{24}) we hav
\begin{equation*}
\left\vert \dint\limits_{\Omega ^{\prime }}u_{\epsilon }\partial _{i}\psi
dx\right\vert =\left\vert \dint\limits_{\Omega ^{\prime }}\partial
_{i}u_{\epsilon }\psi dx\right\vert \leq C_{\Omega ^{\prime }}\left\Vert
\psi \right\Vert _{
{{}^2
(\Omega ^{\prime })}
\end{equation*}
Let $\epsilon \rightarrow 0$ and using the week convergence $u_{\epsilon
}\rightharpoonup u_{0}$ we get:
\begin{equation*}
\left\vert \dint\limits_{\Omega ^{\prime }}u_{0}\partial _{i}\psi
dx\right\vert \leq C_{\Omega ^{\prime }}\left\Vert \psi \right\Vert _{
{{}^2
(\Omega ^{\prime })}\text{ }
\end{equation*}
Hence, $u_{0}\in H_{loc}^{1}(\Omega ).$
\end{proof}
\section{Strong convergence and proof of theorem 3}
Let us begin by some useful propositions
\begin{proposition}
Let $(g_{n})$ be a sequence in $H_{0}^{1}(\Omega )$ and $g\in L^{2}(\Omega )$
such that $\nabla _{X_{2}}g\in L^{2}(\Omega )$ and $\nabla
_{X_{2}}g_{n}\rightarrow \nabla _{X_{2}}g$ in $L^{2}(\Omega )$, then we have:
$g_{n}\rightarrow g$ in $L^{2}(\Omega )$ and for a.e. $X_{1}$ $g(X_{1},.)\in
H_{0}^{1}(\omega _{2})$
\end{proposition}
\begin{proof}
We have for a.e $X_{1}:$ $\nabla _{X_{2}}g_{n}(X_{1},.)\rightarrow \nabla
_{X_{2}}g$ $(X_{1},.)$\ in $L^{2}(\omega _{2})$ (up to a subsequence), and
since for a.e $X_{1}$ and for every $n$ we have $g_{n}(X_{1},.)\in
H_{0}^{1}(\omega _{2})$ then we have for a.e. $X_{1},$ $g(X_{1},.)\in
H_{0}^{1}(\omega _{2}).$And finally the convergence $g_{n}\rightarrow g$ in
L^{2}(\Omega )$ follows by Poincar\'{e}'s inequality $\dint\limits_{\Omega
}\left\vert g_{n}-g\right\vert ^{2}\leq C$ $\dint\limits_{\Omega }\left\vert
\nabla _{X_{2}}(g_{n}-g)\right\vert ^{2}$
\end{proof}
\begin{proposition}
\bigskip Let $f,v\in L^{2}(\Omega )$ such that$\ \nabla _{X_{2}}v\in
L^{2}(\Omega )$ an
\begin{equation*}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}v\cdot \nabla _{X_{2}}\varphi
dx+\beta \dint\limits_{\Omega }v\varphi dx=\dint\limits_{\Omega }f\varphi d
\text{, \ }\forall \varphi \in \mathcal{D}(\Omega ),
\end{equation*}
then we have for a.e $X_{1}$
\begin{equation*}
\dint\limits_{\omega _{2}}A_{22}\nabla _{X_{2}}v(X_{1},.)\cdot \nabla
_{X_{2}}\varphi dX_{2}+\beta \dint\limits_{\omega _{2}}v(X_{1},.)\varphi
dX_{2}=\dint\limits_{\omega _{2}}f(X_{1},.)\varphi dX_{2},\text{ \ }\forall
\varphi \in \mathcal{D}(\omega _{2})\text{\ }
\end{equation*}
Moreover, if for a.e $X_{1}$ we have $v(X_{1},.)\in H_{0}^{1}(\omega _{2})$
then $v$ is the unique function which satisfies the previous equalities
\end{proposition}
\begin{proof}
Same arguments as in \cite{1}.
\end{proof}
\subsection{\textbf{The cut-off problem:}}
Let $\phi \in \mathcal{D}(\Omega )$, and let $(u_{\epsilon })\subset
H_{0}^{1}(\Omega )$ be a sequence of solutions to (\ref{6}) such that
u_{\epsilon }$ converges weakly in $L^{2}(\Omega )$ to some $u_{0}\in
L^{2}(\Omega )$. we define $w_{\epsilon }\in H_{0}^{1}(\Omega )$ to be the
unique solution to the cut-off problem (under assumptions (\ref{1}), (\ref{3
) existence and uniqueness of $w_{\epsilon }$ follows from the Lax-Milgram
theorem)
\begin{equation}
\dint\limits_{\Omega }A_{\epsilon }\nabla w_{\epsilon }\cdot \nabla \varphi
dx+\beta \dint\limits_{\Omega }w_{\epsilon }\varphi dx=\dint\limits_{\Omega
}B(\phi u_{\epsilon })\varphi dx\text{, \ }\forall \varphi \in \mathcal{D
(\Omega ) \label{12}
\end{equation}
The following Lemma is fundamental in this paper
\begin{lemma}
\bigskip Assume (\ref{1}), (\ref{3}), (\ref{2}),(\ref{4}), (\ref{5}), (\re
{34}) then there exists $w_{0}\in $ $W$ such that $w_{\epsilon }\rightarrow $
$w_{0}$ in $W$ strongly and
\begin{equation*}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}\cdot \nabla _{X_{2}}\varphi
dx+\beta \dint\limits_{\Omega }w_{0}\varphi dx=\dint\limits_{\Omega }B(\phi
u_{0})\varphi dx\text{, \ }\forall \varphi \in \mathcal{D}(\Omega ),
\end{equation*
\begin{multline*}
\dint\limits_{\omega _{2}}A_{22}\nabla _{X_{2}}w_{0}(X_{1},.)\cdot \nabla
_{X_{2}}\varphi dX_{2}+\beta \dint\limits_{\omega _{2}}w_{0}(X_{1},.)\varphi
dX_{2} \\
=\dint\limits_{\omega _{2}}B(\phi u_{0})(X_{1},.)\varphi dX_{2}\text{, \
\forall \varphi \in \mathcal{D}(\omega _{2}),
\end{multline*}
and $w_{0}$ is the unique function which satisfies the two previous weak
formulations.
\end{lemma}
Admit this\textbf{\ }lemma\textbf{\ }for the moment then we have the
following
\begin{proposition}
Assume (\ref{1}), (\ref{3}), (\ref{2}),(\ref{4}), (\ref{5}), (\ref{34}), let
$(u_{\epsilon })$ be a sequence of solutions to ($\ref{6}$) such that
u_{\epsilon }\rightharpoonup u_{0}$ weakly in $L^{2}(\Omega )$, then we have
$u_{\epsilon }\rightarrow u_{0}$ in $W$ strongly and
\begin{equation*}
\dint\limits_{\omega _{2}}A_{22}\nabla _{X_{2}}u_{0}(X_{1},.)\cdot \nabla
_{X_{2}}\varphi dX_{2}+\beta \dint\limits_{\omega _{2}}u_{0}(X_{1},.)\varphi
dX_{2}\text{ }=\dint\limits_{\omega _{2}}B(u_{0})(X_{1},.)\varphi dX_{2}
\text{ \ }\forall \varphi \in \mathcal{D}(\omega _{2})
\end{equation*}
\end{proposition}
\subsection{\protect\bigskip Proof of Proposition 4}
\subsubsection{\textbf{Approximation by truncations}}
Let $(u_{\epsilon })$ be a sequence in $H_{0}^{1}(\Omega )$ of solutions to
\ref{6}), assume (\ref{1}), (\ref{3}) and define $w_{\epsilon }^{n}\in
H_{0}^{1}(\Omega )$ the unique solution ( by Lax-Milgram theorem) to the
problem
\begin{equation}
\dint\limits_{\Omega }A_{\epsilon }\nabla w_{\epsilon }^{n}\cdot \nabla
\varphi +\beta \dint\limits_{\Omega }w_{\epsilon }^{n}\varphi
=\dint\limits_{\Omega }B(\phi _{n}u_{\epsilon })\varphi \text{, \ }\forall
\varphi \in \mathcal{D}(\Omega ), \label{13}
\end{equation}
where $(\phi _{n})$ is a sequence in $\mathcal{D}(\Omega )$ which converges
to $1$ in $L^{\frac{2r}{r-2}}(\Omega ).$
\begin{proposition}
Suppose (\ref{1}), (\ref{3}), (\ref{2}), (\ref{4}) then we have
\begin{equation*}
\left\Vert \nabla _{X_{2}}w_{\epsilon }^{n}-\nabla _{X_{2}}u_{\epsilon
}\right\Vert _{L^{2}(\Omega )}\rightarrow 0
\end{equation*
as $n\rightarrow \infty $ uniformly on $\epsilon $
\end{proposition}
\begin{proof}
\bigskip Subtracting (\ref{6}) from (\ref{13}) and taking $\varphi =$
(w_{\epsilon }^{n}-u_{\epsilon })\in H_{0}^{1}(\Omega )$ we get
\begin{multline*}
\dint\limits_{\Omega }A_{\epsilon }\nabla (w_{\epsilon }^{n}-u_{\epsilon
})\cdot \nabla (w_{\epsilon }^{n}-u_{\epsilon })dx+\beta
\dint\limits_{\Omega }(w_{\epsilon }^{n}-u_{\epsilon })^{2}dx \\
=\dint\limits_{\Omega }\left( B(\phi _{n}u_{\epsilon })-B(u_{\epsilon
})\right) (w_{\epsilon }^{n}-u_{\epsilon })dx
\end{multline*
By (\ref{3}) and H\"{o}lder's inequality we deriv
\begin{equation*}
\lambda \left\Vert \nabla _{X_{2}}(w_{\epsilon }^{n}-u_{\epsilon
})\right\Vert _{L^{2}(\Omega )}^{2}\leq \left\Vert \left( B(\phi
_{n}u_{\epsilon })-B(u_{\epsilon })\right) \right\Vert _{L^{2}}\left\Vert
w_{\epsilon }^{n}-u_{\epsilon }\right\Vert _{L^{2}},
\end{equation*}
and \textbf{Proposition 1 }gives $\left\Vert u_{\epsilon }\right\Vert
_{L^{2}(\Omega )}\leq \frac{M\left\vert \Omega \right\vert ^{\frac{1}{2}
\frac{1}{r}}}{\beta -M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r
}},$ $\left\Vert \phi _{n}u_{\epsilon }\right\Vert _{L^{2}(\Omega )}\leq
\frac{M}{\beta -M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}
\left\Vert \phi _{n}\right\Vert _{L^{\frac{2r}{r-2}}(\Omega )}$, we note $K$
the Lipschitz coefficient of $B$ associated with the bounded set
\begin{equation*}
\left\{ u\in L^{2}(\Omega ):\left\Vert u\right\Vert _{L^{2}}\leq \underset{n
{\sup }(\frac{M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}
\beta -M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}},\frac{M}
\beta -M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}\left\Vert
\phi _{n}\right\Vert _{L^{\frac{2r}{r-2}}})<\infty \right\} ,
\end{equation*
$,$
whence (\ref{2}) and H\"{o}lder's inequality give
\begin{equation*}
\left\Vert \nabla _{X_{2}}(w_{\epsilon }^{n}-u_{\epsilon })\right\Vert
_{L^{2}(\Omega )}^{2}\leq \frac{K}{\lambda }\left\Vert \phi
_{n}-1\right\Vert _{L^{\frac{2r}{r-2}}}\left\Vert u_{\epsilon }\right\Vert
_{L^{r}}\left\Vert w_{\epsilon }^{n}-u_{\epsilon }\right\Vert _{L^{2}}
\end{equation*}
And finally by \textbf{Proposition 1} and Poincar\'{e}'s inequality in the
X_{2}$ direction we get
\begin{equation*}
\left\Vert \nabla _{X_{2}}(w_{\epsilon }^{n}-u_{\epsilon })\right\Vert
_{L^{2}(\Omega )}\leq \frac{C^{\prime }KM}{\lambda (\beta -M\left\vert
\Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}})}\left\Vert \phi
_{n}-1\right\Vert _{L^{\frac{2r}{r-2}}}
\end{equation*
\textbf{\ }
Whence $\left\Vert \nabla _{X_{2}}(w_{\epsilon }^{n}-u_{\epsilon
})\right\Vert _{L^{2}(\Omega )}\rightarrow 0$ as $n\rightarrow \infty $
uniformly in $\epsilon $
\end{proof}
\subsubsection{\textbf{The convergence}}
Fix $n$ ,under assumptions of \textbf{Proposition 4} then it follows by
\textbf{Lemma 1} that there exists $w_{0}^{n}\in W$ such that
\begin{equation}
w_{\epsilon }^{n}\rightarrow w_{0}^{n}\text{ strongly in }W \label{22}
\end{equation}
\ and $w_{0}^{n}$ is the unique function in $W$ which satisfies
\begin{equation}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}^{n}\cdot \nabla
_{X_{2}}\varphi dx+\beta \dint\limits_{\Omega }w_{0}^{n}\varphi
dx=\dint\limits_{\Omega }B(\phi _{n}u_{0})\varphi dx\text{, \ }\forall
\varphi \in \mathcal{D}(\Omega ), \label{14}
\end{equation}
and for a.e $X_{1}$ we have
\begin{eqnarray}
&&\dint\limits_{\omega _{2}}A_{22}\nabla _{X_{2}}w_{0}^{n}(X_{1},.)\cdot
\nabla _{X_{2}}\varphi dX_{2}+\beta \dint\limits_{\omega
_{2}}w_{0}^{n}(X_{1},.)\varphi dX_{2} \label{15} \\
&=&\dint\limits_{\omega _{2}}B(\phi _{n}u_{0})(X_{1},.)\varphi dX_{2},\text{
\ }\forall \varphi \in \mathcal{D}(\omega _{2}) \notag
\end{eqnarray}
For a.e $X_{1}$ taking $\varphi =w_{0}^{n}(X_{1},.)\in H_{0}^{1}(\omega
_{2}) $ in (\ref{15}), by ellipticity assumption (\ref{3}), H\"{o}lder's
inequality we obtain
\begin{equation*}
\lambda \dint\limits_{\omega _{2}}\left\vert \nabla
_{X_{2}}w_{0}^{n}(X_{1},.)\right\vert ^{2}dX_{2}\leq \left\Vert B(\phi
_{n}u_{0})(X_{1},.)\right\Vert _{L^{2}(\omega _{2})}\left\Vert
w_{0}^{n}(X_{1},.)\right\Vert _{L^{2}(\omega _{2})},
\end{equation*}
and Poincar\'{e}'s inequality in the $X_{2}$ direction gives
\begin{eqnarray*}
\left\Vert \nabla _{X_{2}}w_{0}^{n}(X_{1},.)\right\Vert _{L^{2}(\omega
_{2})} &\leq &\frac{C^{\prime }}{\lambda }\left\Vert B(\phi
_{n}u_{0})(X_{1},.)\right\Vert _{L^{2}(\omega _{2})} \\
\left\Vert w_{0}^{n}(X_{1},.)\right\Vert _{L^{2}(\omega _{2})} &\leq &\frac
C^{\prime 2}}{\lambda }\left\Vert B(\phi _{n}u_{0})(X_{1},.)\right\Vert
_{L^{2}(\omega _{2})}
\end{eqnarray*}
integrating over $\omega _{1}$ yield
\begin{eqnarray*}
\left\Vert \nabla _{X_{2}}w_{0}^{n}\right\Vert _{L^{2}(\Omega )} &\leq
\frac{C^{\prime }}{\lambda }\left\Vert B(\phi _{n}u_{0})\right\Vert
_{L^{2}(\Omega )}, \\
\left\Vert w_{0}^{n}\right\Vert _{L^{2}(\Omega )} &\leq &\frac{C^{\prime 2}}
\lambda }\left\Vert B(\phi _{n}u_{0})\right\Vert _{L^{2}(\Omega )},
\end{eqnarray*}
and by (\ref{4}) and H\"{o}lder's inequality (remark that $u_{0}\in
L^{r}(\Omega )$ since $(u_{\epsilon })$ is bounded in $L^{r}(\Omega )$ and
u_{\epsilon }\rightharpoonup u_{0}$ in $L^{2}(\Omega )$) we obtain
\begin{eqnarray*}
\left\Vert \nabla _{X_{2}}w_{0}^{n}\right\Vert _{L^{2}(\Omega )} &\leq
\frac{C\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}M\left(
+\left\Vert \phi _{n}\right\Vert _{L^{\frac{2r}{r-2}}}\left\Vert
u_{0}\right\Vert _{L^{r}}\right) }{\lambda }, \\
\left\Vert w_{0}^{n}\right\Vert _{L^{2}(\Omega )} &\leq &\frac
C^{2}\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}M\left(
+\left\Vert \phi _{n}\right\Vert _{L^{\frac{2r}{r-2}}}\left\Vert
u_{0}\right\Vert _{L^{r}}\right) }{\lambda },
\end{eqnarray*}
(we note that The the right hand sides of the previous inequality is
uniformly bounded). Using weak compacity in $L^{2}(\Omega )$, one can
extract a subsequence noted always $(w_{0}^{n})$ which converges weakly to
some $w_{0}\in L^{2}(\Omega )$ and such that $\nabla
_{X_{2}}w_{0}^{n}\rightharpoonup \nabla _{X_{2}}w_{0}$ weakly. Now, passing
to the limit as $n\rightarrow \infty $ in (\ref{14}) and using
\begin{equation}
\left\Vert B(\phi _{n}u_{0})-B(u_{0})\right\Vert _{L^{2}(\Omega )}\leq
K\left\Vert \phi _{n}-1\right\Vert _{L^{\frac{2r}{r-2}}}\left\Vert
u_{0}\right\Vert _{L^{r}(\Omega )} \label{38}
\end{equation
we ge
\begin{equation}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}\cdot \nabla _{X_{2}}\varphi
dx+\beta \dint\limits_{\Omega }w_{0}\varphi dx=\dint\limits_{\Omega
}B(u_{0})\varphi dx\text{, \ }\forall \varphi \in \mathcal{D}(\Omega )
\label{16}
\end{equation
Now we will prove that $\nabla _{X_{2}}w_{0}^{n}\rightarrow \nabla
_{X_{2}}w_{0}$ in $L^{2}(\Omega )$ strongly, using ellipticity assumption
\ref{3}) we obtain
\begin{eqnarray}
&&\lambda \left\Vert \nabla _{X_{2}}(w_{0}^{n}-w_{0})\right\Vert
_{L^{2}(\Omega )}^{2}\leq \label{17} \\
&&\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}(w_{0}^{n}-w_{0})\cdot \nabla
_{X_{2}}(w_{0}^{n}-w_{0})dx+\beta \left\Vert w_{0}^{n}-w_{0}\right\Vert
_{L^{2}(\Omega )}^{2} \notag \\
&\leq &\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}^{n}\cdot \nabla
_{X_{2}}w_{0}^{n}dx-\dint\limits_{\Omega }A_{22}\nabla
_{X_{2}}w_{0}^{n}\cdot \nabla _{X_{2}}w_{0}dx-\dint\limits_{\Omega
}A_{22}\nabla _{X_{2}}w_{0}\cdot \nabla _{X_{2}}w_{0}^{n}dx \notag \\
&&+\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}\cdot \nabla
_{X_{2}}w_{0}dx+\beta \left\Vert w_{0}^{n}-w_{0}\right\Vert _{L^{2}(\Omega
)}^{2} \notag
\end{eqnarray
Taking $\varphi =w_{\epsilon }^{n}$ $\in H_{0}^{1}(\Omega )$ in (\ref{14})
and (\ref{16}) and letting $\epsilon \rightarrow 0$ we get (thanks to (\re
{22}))
\begin{equation}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}^{n}\cdot \nabla
_{X_{2}}w_{0}^{n}dx+\beta \dint\limits_{\Omega }\left\vert
w_{0}^{n}\right\vert ^{2}dx=\dint\limits_{\Omega }B(\phi
_{n}u_{0})w_{0}^{n}dx, \label{18}
\end{equation
an
\begin{equation}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}\cdot \nabla
_{X_{2}}w_{0}^{n}dx+\beta \dint\limits_{\Omega
}w_{0}w_{0}^{n}dx=\dint\limits_{\Omega }B(u_{0})w_{0}^{n}dx \label{19}
\end{equation
Replacing (\ref{18}) and (\ref{19}) in (\ref{17}) we get
\begin{eqnarray}
&&\lambda \left\Vert \nabla _{X_{2}}(w_{0}^{n}-w_{0})\right\Vert
_{L^{2}(\Omega )}^{2} \label{20} \\
&\leq &\dint\limits_{\Omega }B(\phi
_{n}u_{0})w_{0}^{n}dx-\dint\limits_{\Omega
}B(u_{0})w_{0}^{n}dx-\dint\limits_{\Omega }A_{22}\nabla
_{X_{2}}w_{0}^{n}\cdot \nabla _{X_{2}}w_{0}dx \notag \\
&&+\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}\cdot \nabla
_{X_{2}}w_{0}dx+\beta \dint\limits_{\Omega }\left\vert w_{0}\right\vert
^{2}dx-\beta \dint\limits_{\Omega }w_{0}w_{0}^{n}dx \notag
\end{eqnarray}
We have $B(\phi _{n}u_{0})\rightarrow B(u_{0})$ in $L^{2}(\Omega )$ and
since $w_{0}^{n}\rightharpoonup w_{0}$ in $L^{2}(\Omega )$ the
\begin{equation*}
\dint\limits_{\Omega }B(\phi _{n}u_{0})w_{0}^{n}dx\rightarrow
\dint\limits_{\Omega }B(u_{0})w_{0}dx
\end{equation*}
And since $\nabla _{X_{2}}w_{0}^{n}\rightharpoonup \nabla _{X_{2}}w_{0}$ in
L^{2}(\Omega )$ then $A_{22}\nabla _{X_{2}}w_{0}^{n}\rightharpoonup
A_{22}\nabla _{X_{2}}w_{0}$ in $L^{2}(\Omega )$ (since $A_{22}\in L^{\infty
}(\Omega )$). Now, let $n\rightarrow \infty $ in (\ref{20}) we ge
\begin{equation}
\left\Vert \nabla _{X_{2}}(w_{0}^{n}-w_{0})\right\Vert _{L^{2}(\Omega
)}\rightarrow 0 \label{21}
\end{equation}
Thanks to the uniform convergence proved in \textbf{proposition 5}, (\ref{21
) and (\ref{22}), we show by the triangular inequality that $\nabla
_{X_{2}}u_{\epsilon }\rightarrow \nabla _{X_{2}}w_{0}$ in $L^{2}(\Omega )$.
Now, we must check that $w_{0}=u_{0}$, according to \textbf{Proposition 2,
we have for a.e $X_{1},$\ $w_{0}(X_{1},.)\in H_{0}^{1}(\omega _{2})$ and
u_{\epsilon }\rightarrow w_{0}$ in $L^{2}(\Omega )$ , and therefore
w_{0}=u_{0}$. By (\ref{16}), we obtai
\begin{equation*}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}u_{0}\cdot \nabla _{X_{2}}\varphi
dx+\beta \dint\limits_{\Omega }u_{0}\varphi dx=\dint\limits_{\Omega
}B(u_{0})\varphi dx\text{, \ }\forall \varphi \in \mathcal{D}(\Omega ),
\end{equation*}
and we finish the proof of \textbf{proposition 4} by using\textbf{\
proposition 3. }Finally, if $(u_{\epsilon })$ is a sequence of solutions to
\ref{6}) then there exists a subsequence $(u_{\epsilon _{k}})$ which
converges to some $u_{0}$ in $L^{2}(\Omega )$ weakly ( see subsection 2.2),
whence \textbf{Theorem3 }follows from \textbf{Proposition 4}. Now, it
remains to prove \textbf{Lemma 1} which will be the subject of the next
section.
\section{Proof of Lemma 1}
Before starting , let us give some tools. For $n\in
\mathbb{N}
^{\ast }$ we note $\Delta _{n}=(I-n^{-1}\Delta )^{-1}$ the resolvent of the
Dirichlet Laplacian on $L^{2}(\Omega )$, this is a compact operator as well
known. Given $f\in L^{2}(\Omega )$ and we note $U_{n}=(I-n^{-1}\Delta )^{-1}f
$ \ , $U_{n}$ is the unique weak solution to the singularly perturbed
problem:
\begin{equation*}
-\frac{1}{n}\Delta U_{n}+U_{n}=f,
\end{equation*}
we have the
\begin{theorem}
(see \cite{6}): If $f$ $\in H_{0}^{1}(\Omega )$ then : $\left\Vert
U_{n}-f\right\Vert _{L^{2}(\Omega )}\leq C_{\Omega }n^{-\frac{1}{4
}\left\Vert f\right\Vert _{H^{1}(\Omega )}$
\end{theorem}
The following lemma will be used in the approximation
\begin{lemma}
For any functions $g\in H_{loc}^{1}(\Omega )\cap L^{2}(\Omega )$ , $\phi \in
\mathcal{D}(\Omega )$ we have : $\phi g\in H_{0}^{1}(\Omega )$ and moreover
there exists $\Omega ^{\prime }\subset \subset \Omega :$ $\left\Vert \phi
g\right\Vert _{H^{1}(\Omega )}\leq C_{\phi }\left\Vert g\right\Vert
_{H^{1}(\Omega ^{\prime })}$
\end{lemma}
\begin{proof}
the proof is direct.
\end{proof}
\subsection{Approximation of the cut-off problem by \textbf{regularization}}
Let $(u_{\epsilon })\mathbf{\subset }H_{0}^{1}(\Omega )$ be a sequence of
solution to (\ref{6}) such that $u_{\epsilon }\rightharpoonup u_{0}\in
L^{2}(\Omega )$ weakly, assume (\ref{1}), (\ref{3}), (\ref{2}), (\ref{4}),
\ref{5}), (\ref{34}). For $\phi \in \mathcal{D}(\Omega )$ fixed we note
w_{\epsilon }^{n}\in H_{0}^{1}(\Omega )$ the unique solution to the
following regularized problem (thanks to assumptions (\ref{1}), (\ref{3})
and Lax-Milgram theorem)
\begin{equation}
\dint\limits_{\Omega }A_{\epsilon }\nabla w_{\epsilon }^{n}\cdot \nabla
\varphi dx+\beta \dint\limits_{\Omega }w_{\epsilon }^{n}\varphi
dx=\dint\limits_{\Omega }B(\Delta _{n}(\phi u_{\epsilon }))\varphi dx,\text{
\ }\forall \varphi \in \mathcal{D}(\Omega ) \label{25}
\end{equation}
\begin{proposition}
As $n\rightarrow \infty $ we have $:
\begin{equation*}
\nabla _{X_{2}}w_{\epsilon }^{n}\rightarrow \nabla _{X_{2}}w_{\epsilon
\text{ in }L^{2}(\Omega )\text{ uniformly in }\epsilon \text{ ,}
\end{equation*}
where $w_{\epsilon }$ is the solution to the cut-off problem (\ref{12})
\end{proposition}
\begin{proof}
Subtracting (\ref{12}) from (\ref{25}) and taking $\varphi =(w_{\epsilon
}^{n}-w_{\epsilon })\in H_{0}^{1}(\Omega )$ yield
\begin{multline*}
\dint\limits_{\Omega }A_{\epsilon }\nabla (w_{\epsilon }^{n}-w_{\epsilon
})\cdot \nabla (w_{\epsilon }^{n}-w_{\epsilon })dx+\beta
\dint\limits_{\Omega }(w_{\epsilon }^{n}-w_{\epsilon })^{2}dx \\
=\dint\limits_{\Omega }\left\{ B(\Delta _{n}(\phi u_{\epsilon }))-B((\phi
u_{\epsilon }))\right\} (w_{\epsilon }^{n}-w_{\epsilon })dx
\end{multline*}
Remark that $(\phi u_{\epsilon })_{\epsilon }$ is bounded in $L^{2}(\Omega )$
(\textbf{Proposition 1}) and it is clear that $(\Delta _{n}(\phi u_{\epsilon
}))_{n,\epsilon }$ is bounded in $L^{2}(\Omega )$ ($\left\Vert \Delta
_{n}(\phi u_{\epsilon })\right\Vert _{L^{2}(\Omega )}\leq \left\Vert \phi
u_{\epsilon }\right\Vert _{L^{2}(\Omega )}$), then by ellipticity assumption
(\ref{3}) and the local Lipschitzness of $B$ (\ref{2}) we get
\begin{multline*}
\lambda \dint\limits_{\Omega }\left\vert \nabla _{X_{2}}(w_{\epsilon
}^{n}-w_{\epsilon })\right\vert ^{2}dx\leq \dint\limits_{\Omega }\left\{
B(\Delta _{n}(\phi u_{\epsilon }))-B(\phi u_{\epsilon })\right\}
(w_{\epsilon }^{n}-w_{\epsilon })dx \\
\leq K^{\prime }\left( \dint\limits_{\Omega }\left\vert \Delta _{n}(\phi
u_{\epsilon })-\phi u_{\epsilon }\right\vert ^{2}dx\right) ^{\frac{1}{2
}\left\Vert w_{\epsilon }^{n}-w_{\epsilon }\right\Vert _{L^{2}(\Omega )}
\end{multline*}
Hence, Poincar\'{e}'s inequality gives
\begin{equation*}
\text{ }\left\Vert \nabla _{X_{2}}(w_{\epsilon }^{n}-w_{\epsilon
})\right\Vert _{L^{2}(\Omega )}\leq \frac{CK^{\prime }}{\lambda }\left(
\dint\limits_{\Omega }\left\vert \Delta _{n}(\phi u_{\epsilon })-\phi
u_{\epsilon }\right\vert ^{2}dx\right) ^{\frac{1}{2}},
\end{equation*}
and by \textbf{Theorem 7} we get
\begin{equation*}
\left\Vert \nabla _{X_{2}}(w_{\epsilon }^{n}-w_{\epsilon })\right\Vert
_{L^{2}(\Omega )}\leq \frac{C^{\prime }K}{\lambda }n^{-\frac{1}{4
}\left\Vert \phi u_{\epsilon }\right\Vert _{H^{1}(\Omega )},
\end{equation*}
and \textbf{Lemma 2 }gives
\begin{equation*}
\left\Vert \nabla _{X_{2}}(w_{\epsilon }^{n}-w_{\epsilon })\right\Vert
_{L^{2}(\Omega )}\leq \frac{C^{\prime }K}{\lambda }C_{\phi }n^{-\frac{1}{4
}\left\Vert u_{\epsilon }\right\Vert _{H^{1}(\Omega ^{\prime })}
\end{equation*}
So finally by \textbf{Theorem 2} we ge
\begin{equation*}
\left\Vert \nabla _{X_{2}}(w_{\epsilon }^{n}-w_{\epsilon })\right\Vert
_{L^{2}(\Omega )}\leq C"n^{-\frac{1}{4}}
\end{equation*}
where $C"\geq 0$ is independent on $\epsilon $ and $n$
\end{proof}
\subsection{\protect\bigskip \textbf{The convergence}}
\subsubsection{\textbf{Passage to the limit as }$\protect\epsilon
\rightarrow 0$}
Let $n\in
\mathbb{N}
^{\ast }$ fixed, taking $\varphi =w_{\epsilon }^{n}\in H_{0}^{1}(\Omega )$
in (\ref{25})$,$and estimating using ellipticity assumption (\ref{3}) and
\ref{4}) and \textbf{Proposition 1}(as in subsection 2.2) then one can
extract a subsequence $(w_{\epsilon _{k}(n)}^{n})_{k}$ which converges (as
\epsilon _{k}(n)\rightarrow 0$) to some $w_{0}^{n}$ in the following
sense\smallskip
\begin{eqnarray}
w_{\epsilon _{k}(n)}^{n} &\rightharpoonup &w_{0}^{n}\text{ ,}\nabla
_{X_{2}}w_{\epsilon _{k}(n)}^{n}\rightharpoonup \nabla _{X_{2}}w_{0}^{n
\text{ and }\epsilon _{k}(n)\nabla _{X_{1}}w_{\epsilon
_{k}(n)}^{n}\rightharpoonup 0 \label{27} \\
&&\text{ in }L^{2}(\Omega )\text{ } \notag
\end{eqnarray
Now passing the limit (as $\epsilon _{k}(n)\rightarrow 0$) in (\ref{25}) we
get\
\begin{equation*}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}^{n}\cdot \nabla
_{X_{2}}\varphi dx+\beta \dint\limits_{\Omega }w_{0}^{n}\varphi dx=\underset
\epsilon _{k}(n)\rightarrow 0}{\lim }\dint\limits_{\Omega }B(\Delta
_{n}(\phi u_{\epsilon _{k}(n)}))\varphi dx,\text{ \ }\forall \varphi \in
\mathcal{D}(\Omega )\smallskip
\end{equation*
Since $u_{\epsilon _{k}(n)}\rightharpoonup u_{0}$ weakly in $L^{2}(\Omega )$
then $\phi u_{\epsilon _{k}(n)}\rightharpoonup \phi u_{0}$ weakly in
L^{2}(\Omega )$ so by compacity of $\Delta _{n}$ we get $\Delta _{n}(\phi
u_{\epsilon _{k}(n)})\rightarrow \Delta _{n}(\phi u_{0})$ in $L^{2}(\Omega )$
strongly. And therefore, the continuity of $B$ gives $B(\Delta _{n}(\phi
u_{\epsilon _{k}(n)}))\rightarrow B(\Delta _{n}(\phi u_{0}))$ in
L^{2}(\Omega )$ strongly, hence the previous equality become
\begin{equation}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}^{n}\cdot \nabla
_{X_{2}}\varphi dx+\beta \dint\limits_{\Omega }w_{0}^{n}\varphi
dx=\dint\limits_{\Omega }B(\Delta _{n}(\phi u_{0)}))\varphi dx,\text{ \
\forall \varphi \in \mathcal{D}(\Omega ) \label{28}
\end{equation
Take $\varphi =w_{\epsilon _{k}(n)}^{n}\in H_{0}^{1}(\Omega )$ in (\ref{28})
and let $\epsilon _{k}(n)\rightarrow 0$ we deriv
\begin{equation}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}^{n}\cdot \nabla
_{X_{2}}w_{0}^{n}dx+\beta \dint\limits_{\Omega }\left\vert
w_{0}^{n}\right\vert ^{2}dx=\dint\limits_{\Omega }B(\Delta _{n}(\phi
u_{0)}))w_{0}^{n}dx\ \ \label{29}
\end{equation}
Now, we prove strong convergences for the whole sequence (as $\epsilon
\rightarrow 0$)
\begin{proposition}
As $\epsilon \rightarrow 0$ we have $\nabla _{X_{2}}w_{\epsilon
}^{n}\rightarrow \nabla _{X_{2}}w_{0}^{n}$ strongly in $L^{2}(\Omega )$
\end{proposition}
\begin{proof}
Computin
\begin{multline*}
I_{\epsilon _{k}(n)}^{n}=\dint\limits_{\Omega }A_{\epsilon }\binom{\nabla
_{X_{1}}w_{\epsilon _{k}(n)}^{n}}{\nabla _{X_{2}}(w_{\epsilon
_{k}(n)}^{n}-w_{0}^{n})}\cdot \binom{\nabla _{X_{1}}w_{\epsilon _{k}(n)}^{n
}{\nabla _{X_{2}}(w_{\epsilon _{k}(n)}^{n}-w_{0}^{n})}dx+\beta
\dint\limits_{\Omega }\left\vert w_{\epsilon
_{k}(n)}^{n}-w_{0}^{n}\right\vert ^{2}dx \\
=\dint\limits_{\Omega }B(\Delta _{n}(\phi u_{\epsilon _{k}(n)}))\
w_{\epsilon _{k}(n)}^{n}dx-\dint\limits_{\Omega }\epsilon
_{k}(n)A_{12}\nabla _{X_{2}}w_{0}^{n}\cdot \nabla _{X_{1}}w_{\epsilon
_{k}(n)}^{n}dx-2\beta \dint\limits_{\Omega }w_{\epsilon
_{k}(n)}^{n}w_{0}^{n}dx \\
-\dint\limits_{\Omega }\epsilon _{k}(n)A_{21}\nabla _{X_{1}}w_{\epsilon
_{k}(n)}^{n}\cdot \nabla _{X_{2}}w_{0}^{n})dx-\dint\limits_{\Omega
}A_{22}\nabla _{X_{2}}w_{0}^{n}\cdot \nabla _{X_{2}}w_{\epsilon
_{k}(n)}^{n}dx \\
-\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{\epsilon _{k}(n)}^{n}\cdot
\nabla _{X_{2}}w_{0}^{n})dx+\dint\limits_{\Omega }B(\Delta _{n}(\phi
u_{0)}))w_{0}^{n}dx
\end{multline*}
Let $\epsilon _{k}(n)\rightarrow 0$ and using (\ref{27}) we ge
\begin{multline*}
\underset{\epsilon _{k}(n)\rightarrow 0}{\lim }I_{\epsilon _{k}(n)}^{n}
\underset{\epsilon _{k}(n)\rightarrow 0}{\lim }\left( \dint\limits_{\Omega
}B(\Delta _{n}(\phi u_{\epsilon _{k}(n)}))\ w_{\epsilon
_{k}(n)}^{n}dx+\dint\limits_{\Omega }B(\Delta _{n}(\phi
u_{0)}))w_{0}^{n}dx\right) -2\beta \dint\limits_{\Omega }\left\vert
w_{0}^{n}\right\vert ^{2}dx \\
-2\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}^{n}\cdot \nabla
_{X_{2}}w_{0}^{n}dx
\end{multline*}
Since $B(\Delta _{n}(\phi u_{\epsilon _{k}(n)}))\rightarrow B(\Delta
_{n}(\phi u_{0}))$ in $L^{2}(\Omega )$ strongly and $w_{\epsilon
_{k}(n)}^{n}\rightharpoonup w_{0}^{n}$ weakly the
\begin{equation*}
\dint\limits_{\Omega }B(\Delta _{n}(\phi u_{\epsilon _{k}(n)}))\ w_{\epsilon
_{k}(n)}^{n}\rightarrow \dint\limits_{\Omega }B(\Delta _{n}(\phi u_{0}))\
w_{0}^{n}
\end{equation*}
Whence by (\ref{29}) we get $\underset{\epsilon _{i}(n)\rightarrow 0}{\lim
I_{\epsilon _{i}(n)}^{n}=0$. Now using ellipticity assumption (\ref{3}) we
deriv
\begin{equation*}
\lambda \epsilon _{i}(n)^{2}\dint\limits_{\Omega }\left\vert \nabla
_{X_{1}}w_{\epsilon _{i}(n)}^{n}\right\vert ^{2}+\lambda
\dint\limits_{\Omega }\left\vert \nabla _{X_{2}}(w_{\epsilon
_{i}(n)}^{n}-w_{0}^{n})\right\vert ^{2}\leq I_{\epsilon _{i}(n)}^{n}\text{,}
\end{equation*}
and therefore we ge
\begin{equation*}
\left\Vert \nabla _{X_{2}}(w_{\epsilon _{i}(n)}^{n}-w_{0}^{n}\right\Vert
_{L^{2}(\Omega )}\rightarrow 0\text{ as }\epsilon _{i}(n)\rightarrow 0,
\end{equation*}
According to \textbf{Proposition 2 }we have for a.e $X_{1},$
w_{0}^{n}(X_{1},.)\in H_{0}^{1}(\omega _{2})$ and
\begin{eqnarray*}
\left\Vert \nabla _{X_{2}}(w_{\epsilon _{i}(n)}^{n}-w_{0}^{n})\right\Vert
_{L^{2}(\Omega )} &\rightarrow &0,\text{ }\left\Vert w_{\epsilon
_{i}(n)}^{n}-w_{0}^{n}\right\Vert _{L^{2}(\Omega )}\rightarrow 0\text{ } \\
\text{as }\epsilon _{i}(n) &\rightarrow &0,
\end{eqnarray*}
By (\ref{28}) and \textbf{Proposition 3 }we show that for every $n$ fixed,
w_{0}^{n}$ is the unique function which satisfies for a.e $X_{1}$
\begin{multline*}
\dint\limits_{\omega _{2}}A_{22}\nabla _{X_{2}}w_{0}^{n}(X_{1},.)\cdot
\nabla _{X_{2}}\varphi dX_{2}+\beta \dint\limits_{\omega
_{2}}w_{0}^{n}(X_{1},.)\varphi \ dX_{2} \\
=\dint\limits_{\omega _{2}}B(\Delta _{n}(\phi u_{0)}))(X_{1},.)\varphi
dX_{2},\text{ \ }\forall \varphi \in \mathcal{D}(\omega _{2})
\end{multline*}
Since the union of zero measure sets is a zero measure set then we have for
a.e $X_{1\text{ }}$ and $\forall n\in
\mathbb{N}
^{\ast }
\begin{eqnarray}
&&\dint\limits_{\omega _{2}}A_{22}\nabla _{X_{2}}w_{0}^{n}(X_{1},.)\cdot
\nabla _{X_{2}}\varphi dX_{2}+\beta \dint\limits_{\omega
_{2}}w_{0}^{n}(X_{1},.)\varphi dX_{2}\ \label{30} \\
&=&\dint\limits_{\omega _{2}}B(\Delta _{n}(\phi u_{0)}))(X_{1},.)\varphi
dX_{2},\text{ \ }\forall \varphi \in \mathcal{D}(\omega _{2}) \notag
\end{eqnarray}
And finally, the uniqueness of $w_{0}^{n}$ implies that the whole sequence
(w_{\epsilon }^{n})$ converges i.e $\forall n\in
\mathbb{N}
^{\ast }:
\begin{equation*}
\left\Vert \nabla _{X_{2}}(w_{\epsilon }^{n}-w_{0}^{n})\right\Vert
_{L^{2}(\Omega )}\rightarrow 0,\text{ and }\left\Vert w_{\epsilon
}^{n}-w_{0}^{n}\right\Vert _{L^{2}(\Omega )}\rightarrow 0\text{ as }\epsilon
\rightarrow 0
\end{equation*}
\end{proof}
\subsubsection{\textbf{Passage to the limit }$n\rightarrow \infty $}
For a.e $X_{1}$ and $\forall n\in
\mathbb{N}
^{\ast }$ taking $\varphi =w_{0}^{n}(X_{1},.)\in H_{0}^{1}(\omega _{2})$ in
\ref{30}), using ellipticity assumption (\ref{3}) and H\"{o}lder's
inequality we get
\begin{equation*}
\lambda \dint\limits_{\omega _{2}}\left\vert \nabla
_{X_{2}}w_{0}^{n}(X_{1},.)\right\vert ^{2}dX_{2}\leq \left\Vert B(\Delta
_{n}(\phi u_{0}))(X_{1},.)\right\Vert _{L^{2}(\omega _{2})}\left\Vert
w_{0}^{n}(X_{1},.)\right\Vert _{L^{2}(\omega _{2})}
\end{equation*}
and Poincar\'{e}'s inequality in the $X_{2}$ direction gives
\begin{eqnarray*}
\left\Vert \nabla _{X_{2}}w_{0}^{n}(X_{1},.)\right\Vert _{L^{2}(\omega
_{2})} &\leq &\frac{C^{\prime }}{\lambda }\left\Vert B(\Delta _{n}(\phi
u_{0}))(X_{1},.)\right\Vert _{L^{2}(\omega _{2})} \\
\left\Vert w_{0}^{n}(X_{1},.)\right\Vert _{L^{2}(\omega _{2})} &\leq &\frac
C^{\prime 2}}{\lambda }\left\Vert B(\Delta _{n}(\phi
u_{0}))(X_{1},.)\right\Vert _{L^{2}(\omega _{2})}
\end{eqnarray*}
integrating over $\omega _{1}$ yield
\begin{eqnarray*}
\left\Vert \nabla _{X_{2}}w_{0}^{n}\right\Vert _{L^{2}(\Omega )} &\leq
\frac{C^{\prime }}{\lambda }\left\Vert B(\Delta _{n}(\phi u_{0}))\right\Vert
_{L^{2}(\Omega )}, \\
\left\Vert w_{0}^{n}\right\Vert _{L^{2}(\Omega )} &\leq &\frac{C^{\prime 2}}
\lambda }\left\Vert B(\Delta _{n}(\phi u_{0}))\right\Vert _{L^{2}(\Omega )},
\end{eqnarray*}
and by (\ref{4}) and Holder's inequality we obtain
\begin{eqnarray*}
\left\Vert \nabla _{X_{2}}w_{0}^{n}\right\Vert _{L^{2}(\Omega )} &\leq
\frac{C^{\prime }\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r
}M\left( +\left\Vert \phi \right\Vert _{\infty }\left\Vert u_{0}\right\Vert
_{L^{2}}\right) }{\lambda }, \\
\left\Vert w_{0}^{n}\right\Vert _{L^{2}(\Omega )} &\leq &\frac{C^{^{\prime
}2}\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}M\left(
+\left\Vert \phi \right\Vert _{\infty }\left\Vert u_{0}\right\Vert
_{L^{2}}\right) }{\lambda },
\end{eqnarray*}
(we used the inequality $\left\Vert \Delta _{n}(\phi u_{0})\right\Vert
_{L^{2}(\Omega )}\leq \left\Vert \phi u_{0}\right\Vert _{L^{2}(\Omega )}$
and the notation $\left\Vert \phi \right\Vert _{\infty }=\underset{x\in
\Omega }{\sup }\left\vert \phi (x)\right\vert $)
Whence, it follows by weak compacity that there exists $w_{0}\in
L^{2}(\Omega )$ and a subsequence noted always $(w_{0}^{n})$ such tha
\begin{equation*}
\nabla _{X_{2}}w_{0}^{n}\rightharpoonup \nabla _{X_{2}}w_{0}\text{ and
w_{0}^{n}\rightharpoonup w_{0}\text{ in }L^{2}(\Omega )\text{ }
\end{equation*}
Remark that $\phi u_{0}\in H_{0}^{1}(\Omega )$ by \textbf{Lemma2}, then by
\textbf{Theorem 7 }$\Delta _{n}(\phi u_{0})\rightarrow \phi u_{0}$ in
L^{2}(\Omega )$ and therefore, continuity of $B$ gives $B(\Delta _{n}(\phi
u_{0)}))\rightarrow B(\phi u_{0})$ in $L^{2}(\Omega )$. Now, let
n\rightarrow \infty $ in (\ref{28}) yields
\begin{equation}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}\cdot \nabla _{X_{2}}\varphi
dx+\beta \dint\limits_{\Omega }w_{0}\varphi dx=\dint\limits_{\Omega }B(\phi
u_{0})\varphi dx,\text{ \ }\forall \varphi \in \mathcal{D}(\Omega )
\label{31}
\end{equation}
Take $\varphi =w_{\epsilon }^{n}\in H_{0}^{1}(\Omega )$ in (\ref{31}) and
let $\epsilon \rightarrow 0$ we obtain (by \textbf{Proposition 7)
\begin{equation*}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}\cdot \nabla
_{X_{2}}w_{0}^{n}dx+\beta \dint\limits_{\Omega
}w_{0}w_{0}^{n}dx=\dint\limits_{\Omega }B(\phi u_{0})w_{0}^{n}dx,
\end{equation*}
and as $n\rightarrow \infty $ we deriv
\begin{equation}
\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}\cdot \nabla
_{X_{2}}w_{0}dx+\beta \dint\limits_{\Omega }\left\vert w_{0}\right\vert
^{2}dx=\dint\limits_{\Omega }B(\phi u_{0})w_{0}dx \label{32}
\end{equation}
Now, we prove the strong convergences of $w_{0}^{n}$ and $\nabla
_{X_{2}}w_{0}^{n}$, by ellipticity assumption (\ref{3}), (\ref{29}) and (\re
{32}) we get
\begin{multline*}
\lambda \dint\limits_{\Omega }\left\vert \nabla
_{X_{2}}(w_{0}^{n}-w_{0})\right\vert ^{2}dx+\beta \dint\limits_{\Omega
}\left\vert w_{0}^{n}-w_{0}\right\vert ^{2}dx \\
\leq \dint\limits_{\Omega }A_{22}\nabla _{X_{2}}(w_{0}^{n}-w_{0})\cdot
\nabla _{X_{2}}(w_{0}^{n}-w_{0})dx+\beta \dint\limits_{\Omega }\left\vert
w_{0}^{n}-w_{0}\right\vert ^{2}dx \\
=\dint\limits_{\Omega }B(\Delta _{n}(\phi u_{0)}))w_{0}^{n}dx\
-\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}\cdot \nabla
_{X_{2}}w_{0}^{n}dx-2\beta \dint\limits_{\Omega }w_{0}^{n}w_{0}dx \\
-\dint\limits_{\Omega }A_{22}\nabla _{X_{2}}w_{0}^{n}\cdot \nabla
_{X_{2}}w_{0}dx+\dint\limits_{\Omega }B(\phi u_{0})w_{0}dx
\end{multline*
Since $B(\Delta _{n}(\phi u_{0)})\rightarrow B(\phi u_{0})$ in $L^{2}(\Omega
)$ and $w_{0}^{n}\rightharpoonup $ $w_{0}$ in $L^{2}(\Omega )$ the
\begin{equation*}
\dint\limits_{\Omega }B(\Delta _{n}(\phi u_{0)}))w_{0}^{n}\rightarrow
\dint\limits_{\Omega }B(\phi u_{0})w_{0}
\end{equation*}
Let $n\rightarrow \infty $ in the previous inequality we get
\begin{equation}
\nabla _{X_{2}}w_{0}^{n}\rightarrow \nabla _{X_{2}}w_{0}\text{ in
L^{2}(\Omega ) \label{33}
\end{equation
Finally by (\ref{33}), \textbf{Proposition 6 }and \textbf{Proposition 7 }and
the triangular inequality we get $\nabla _{X_{2}}w_{\epsilon }\rightarrow
\nabla _{X_{2}}w_{0}$ in $L^{2}(\Omega )$ and therefore (\ref{32}), \textbf
Proposition 2 }and\textbf{\ 3 }complete the proof.
\begin{remark}
In addition to convergences given in \textbf{Theorem 3} we also have
\epsilon _{k}u_{\epsilon _{k}}\rightarrow 0$ in $L^{2}(\Omega )$ strongly,
indeed ellipticity assumption gives
\begin{multline*}
\lambda \epsilon _{k}{}^{2}\dint\limits_{\Omega }\left\vert \nabla
_{X_{1}}u_{\epsilon _{k}}\right\vert ^{2}+\lambda \dint\limits_{\Omega
}\left\vert \nabla _{X_{2}}(u_{\epsilon _{k}}-u_{0})\right\vert ^{2} \\
\leq \dint\limits_{\Omega }A_{\epsilon }\binom{\nabla _{X_{1}}u_{\epsilon
_{k}}}{\nabla _{X_{2}}(u_{\epsilon _{k}}-u_{0})}\cdot \binom{\nabla
_{X_{1}}u_{\epsilon _{k}}}{\nabla _{X_{2}}(u_{\epsilon _{k}}-u_{0})}dx+\beta
\dint\limits_{\Omega }\left\vert u_{\epsilon _{k}}-u_{0}\right\vert ^{2}dx,
\end{multline*}
\end{remark}
and we can prove easily that the right-hand side of this inequality
converges to $0.$
\section{Some Applications}
\subsection{\textbf{A regularity result and rate of convergence}}
In this subsection we make some additional assumptions, suppose that for
every $u\in L^{2}(\Omega ),$
\begin{equation}
\nabla _{X_{1}}B(u)\in L^{2}(\Omega ), \label{43}
\end{equation}
and for every $\rho \in \mathcal{D(\omega }_{1}\mathcal{)}$ and $u,v$ $\in
L^{2}(\Omega )$ we have
\begin{equation}
\left\Vert \rho B(u)-\rho B(v)\right\Vert _{L^{2}}\leq \left\Vert B(\rho
u)-B(\rho v)\right\Vert \label{44}
\end{equation}
Remark that\textbf{\ Theorem 3} of section 1 gives only $H_{loc}^{1}-$
regularity for $u_{0}$, however we have the following
\begin{proposition}
Under assumptions of \textbf{Theorem 3 }and (\ref{43}) we have $u_{0}\in
H^{1}(\Omega ),$
\end{proposition}
\begin{proof}
We will proceed as in \cite{1}, let $\omega _{1}^{\prime }\subset \subset
\omega _{1}$, for $0<h<d(\omega _{1}^{\prime },\omega _{1}),$ $X_{1}\in
\omega _{1}^{\prime }$ we set $\tau
_{h}^{i}u_{0}(X_{1},X_{2})=u_{0}(X_{1}+he_{i},X_{2})$ $i=1,...,p.$ From (\re
{11}) we hav
\begin{multline*}
\dint\limits_{\omega _{2}}\tau _{h}^{i}A_{22}\nabla _{X_{2}}(\tau
_{h}^{i}u_{0}-u_{0})\nabla _{X_{2}}\varphi dX_{2}+\dint\limits_{\omega
_{2}}(\tau _{h}^{i}A_{22}-A_{22})\nabla _{X_{2}}u_{0}\nabla _{X_{2}}\varphi
dX_{2} \\
+\beta \dint\limits_{\omega _{2}}(\tau _{h}^{i}u_{0}-u_{0})\varphi
dX_{2}=\dint\limits_{\omega _{2}}\left\{ \tau
_{h}^{i}B(u_{0})-B(u_{0})\right\} \varphi dX_{2}\text{ }
\end{multline*}
Taking $\varphi =\frac{\tau _{h}^{i}u_{0}-u_{0}}{h^{2}}$ as a test function,
using ellipticity assumption (\ref{3}) and H\"{o}lder's inequality we derive
\begin{multline*}
\lambda \left\Vert \nabla _{X_{2}}\left( \frac{\tau _{h}^{i}u_{0}-u_{0}}{h
\right) \right\Vert _{L^{2}(\omega _{2})}^{2}\leq \\
\left\Vert \left( \frac{\tau _{h}^{i}A_{22}-A_{22}}{h}\right) \nabla
_{X_{2}}u_{0}\right\Vert _{L^{2}(\omega _{2})}\left\Vert \nabla
_{X_{2}}\left( \frac{\tau _{h}^{i}u_{0}-u_{0}}{h}\right) \right\Vert
_{L^{2}(\omega _{2})} \\
+\left\Vert \left( \frac{\tau _{h}^{i}B(u_{0})-B(u_{0})}{h}\right)
\right\Vert _{L^{2}(\omega _{2})}\left\Vert \left( \frac{\tau
_{h}^{i}u_{0}-u_{0}}{h}\right) \right\Vert _{L^{2}(\omega _{2})}
\end{multline*}
Using Poincar\'{e}'s inequality we deduc
\begin{equation*}
\left\Vert \frac{\tau _{h}^{i}u_{0}-u_{0}}{h}\right\Vert _{L^{2}(\omega
_{2})}\leq \frac{C}{\lambda }\left\{
\begin{array}{c}
\left\Vert \frac{\tau _{h}^{i}A_{22}-A_{22}}{h}\right\Vert _{L^{\infty
}(\omega _{2})}\left\Vert \nabla _{X_{2}}u_{0}\right\Vert _{L^{2}(\omega
_{2})} \\
+\left\Vert \frac{\tau _{h}^{i}B(u_{0})-B(u_{0})}{h}\right\Vert
_{L^{2}(\omega _{2})
\end{array
\right\}
\end{equation*}
Using regularity assumption (\ref{34}) and integrating over $\omega
_{1}^{\prime }$ (we use only the assumption $\partial _{k}A_{22}\in
L^{2}(\Omega )$) we deduce
\begin{equation*}
\left\Vert \frac{\tau _{h}^{i}u_{0}-u_{0}}{h}\right\Vert _{L(\omega
_{1}^{\prime }\times \omega _{2})}\leq C^{\prime }+\left\Vert \left( \frac
\tau _{h}^{i}B(u_{0})-B(u_{0})}{h}\right) \right\Vert _{L^{2}(\omega
_{1}^{\prime }\times \omega _{2})}
\end{equation*}
Thanks to regularity of $B(u_{0})$ in the $X_{1}$ direction (assumption (\re
{43})) we get
\begin{equation*}
\left\Vert \frac{\tau _{h}^{i}u_{0}-u_{0}}{h}\right\Vert _{L^{2}(\omega
_{1}^{\prime }\times \omega _{2})}\leq C^{\prime \prime },
\end{equation*}
where $C^{\prime \prime }$ is independent on $h$ , whence $\nabla
_{X_{1}}u_{0}\in L^{2}(\Omega )$ and the proof is finished.
\end{proof}
Now, we give a result on the rate of convergence
\begin{proposition}
Under assumptions of \textbf{Theorem 3} and \ (\ref{43}), (\ref{44}), for
\beta >\max (K,\beta _{0})$ (where $\beta _{0}>M\left\vert \Omega
\right\vert ^{\frac{1}{2}-\frac{1}{r}}$ (fixed), and $K$ is the Lipschitz
constant of $B$ associated with the bounded set $\left\{ \left\Vert
u\right\Vert _{L^{2}}\leq \frac{M\left\vert \Omega \right\vert ^{\frac{1}{2}
\frac{1}{r}}}{\beta _{0}-M\left\vert \Omega \right\vert ^{\frac{1}{2}-\frac{
}{r}}}\right\} $), we have $u_{\epsilon }\rightarrow u_{0}$ in $W$ \ and
\begin{equation*}
\left\Vert u_{\epsilon }-u_{0}\right\Vert _{L^{2}(\omega _{1}^{\prime
}\times \omega _{2})};\text{ }\left\Vert \nabla _{X_{1}}(u_{\epsilon
}-u_{0})\right\Vert _{L^{2}(\omega _{1}^{\prime }\times \omega _{2})}\leq
C^{\prime }\epsilon
\end{equation*}
where $C\geq 0$ is independent of $\epsilon .$
\end{proposition}
\begin{proof}
To make calculus easier we suppose that $A_{12},A_{21}=0$, $A_{11},A_{22}=I$
. According to \textbf{Theorem 3} the set of solutions to (\ref{11}) is non
empty, and we show easily that (\ref{11}) has a unique solution (thanks to
assumption $\beta >\max (K,\beta _{0})$), consequently \textbf{Corollary 1}
implies $u_{\epsilon }\rightarrow u_{0}$ in $W$.
From (\ref{6}) and (\ref{11}) we hav
\begin{equation*}
\epsilon ^{2}\dint\limits_{\Omega }\nabla _{X_{1}}u_{\epsilon }\nabla
_{X_{1}}\varphi dx+\dint\limits_{\Omega }\nabla _{X_{2}}(u_{\epsilon
}-u_{0})\nabla _{X_{2}}\varphi dx+\beta \dint\limits_{\Omega }(u_{\epsilon
}-u_{0})\varphi dx=\dint\limits_{\Omega }(B(u_{\epsilon })-B(u_{0}))\varphi
dx\text{ }
\end{equation*
Given $\omega _{1}^{\prime }\subset \subset $ $\omega _{1}^{^{\prime \prime
}}\subset \subset $ $\omega _{1}$, and let $\rho $ be a cut-off function
with $Supp(\rho )\subset \omega _{1}^{^{\prime \prime }}$ and $\rho =1$ on
\omega _{1}^{\prime }$(we can choose $0\leq \rho \leq 1$). We introduce the
test function used by M.Chipot and S.Guesmia in \cite{1}, $\varphi =$ $\rho
^{2}(u_{\epsilon }-u_{0})\in H_{0}^{1}(\Omega )$ ( thanks to the previous
proposition). Testing with $\varphi $ we obtain
\begin{multline*}
\epsilon ^{2}\dint\limits_{\Omega }\nabla _{X_{1}}u_{\epsilon }\nabla
_{X_{1}}\rho ^{2}(u_{\epsilon }-u_{0})dx \\
+\dint\limits_{\Omega }\nabla _{X_{2}}(u_{\epsilon }-u_{0})\nabla
_{X_{2}}\rho ^{2}(u_{\epsilon }-u_{0})dx+\beta \dint\limits_{\Omega }\rho
^{2}(u_{\epsilon }-u_{0})^{2}dx \\
=\dint\limits_{\Omega }(B(u_{\epsilon })-B(u_{0}))\rho ^{2}(u_{\epsilon
}-u_{0})dx\text{ }
\end{multline*}
we deduce
\begin{multline*}
\epsilon ^{2}\dint\limits_{\Omega }\left\vert \rho \nabla
_{X_{1}}(u_{\epsilon }-u_{0})\right\vert ^{2}dx+\dint\limits_{\Omega
}\left\vert \rho \nabla _{X_{2}}(u_{\epsilon }-u_{0})\right\vert ^{2}dx \\
+\beta \dint\limits_{\Omega }\rho ^{2}(u_{\epsilon }-u_{0})^{2}dx=-\epsilon
^{2}\dint\limits_{\Omega }\rho ^{2}\nabla _{X_{1}}u_{0}\nabla
_{X_{1}}(u_{\epsilon }-u_{0})dx-2\epsilon ^{2}\dint\limits_{\Omega
}(u_{\epsilon }-u_{0})\rho \nabla _{X_{1}}\rho \nabla _{X_{1}}u_{0}dx \\
-2\epsilon ^{2}\dint\limits_{\Omega }\rho (u_{\epsilon }-u_{0})\nabla
_{X_{1}}(u_{\epsilon }-u_{0})\nabla _{X_{1}}\rho dx+\dint\limits_{\Omega
}(B(u_{\epsilon })-B(u_{0}))\rho ^{2}(u_{\epsilon }-u_{0})dx
\end{multline*}
Using H\"{o}lder's inequality for the first three term in the right-hand
side, and assumptions (\ref{44}), (\ref{2}) and H\"{o}lder's inequality for
the last one, we obtai
\begin{multline*}
\epsilon ^{2}\left\Vert \rho \nabla _{X_{1}}(u_{\epsilon }-u_{0})\right\Vert
_{L^{2}(\omega _{1}^{\prime \prime }\times \omega _{2})}^{2}+\left\Vert \rho
\nabla _{X_{2}}(u_{\epsilon }-u_{0})\right\Vert _{L^{2}(\omega _{1}^{\prime
\prime }\times \omega _{2})}^{2}+ \\
\beta \left\Vert \rho (u_{\epsilon }-u_{0})\right\Vert _{L^{2}(\omega
_{1}^{\prime \prime }\times \omega _{2})}^{2}\leq \epsilon ^{2}\left\Vert
\rho \nabla _{X_{1}}u_{0}\right\Vert _{L^{2}(\omega _{1}^{\prime \prime
}\times \omega _{2})}\left\Vert \rho \nabla _{X_{1}}(u_{\epsilon
}-u_{0})\right\Vert _{L^{2}(\omega _{1}^{\prime \prime }\times \omega _{2})}
\\
+2\epsilon ^{2}\left\Vert (u_{\epsilon }-u_{0})\nabla _{X_{1}}\rho
\right\Vert _{L^{2}(\omega _{1}^{\prime \prime }\times \omega
_{2})}\left\Vert \rho \nabla _{X_{1}}u_{0}\right\Vert _{L^{2}(\omega
_{1}^{\prime \prime }\times \omega _{2})} \\
+\epsilon ^{2}\left\Vert (u_{\epsilon }-u_{0})\nabla _{X_{1}}\rho
\right\Vert _{L^{2}(\omega _{1}^{\prime \prime }\times \omega
_{2})}\left\Vert \rho (u_{\epsilon }-u_{0})\right\Vert _{L^{2}(\omega
_{1}^{\prime \prime }\times \omega _{2})} \\
+K\left\Vert \rho (u_{\epsilon }-u_{0})\right\Vert _{L^{2}(\omega
_{1}^{\prime \prime }\times \omega _{2})}^{2},
\end{multline*}
(thanks to \textbf{Proposition 1, }we remark that $\left\Vert \rho
u_{\epsilon }\right\Vert _{L^{2}}$, $\left\Vert \rho u_{0}\right\Vert
_{L^{2}}\in \left\{ \left\Vert u\right\Vert _{L^{2}}\leq \frac{M\left\vert
\Omega \right\vert ^{\frac{1}{2}-\frac{1}{r}}}{\beta _{0}-M\left\vert \Omega
\right\vert ^{\frac{1}{2}-\frac{1}{r}}}\right\} $\textbf{)}$.$
Using Young's inequality for ,the first term in the right-hand side of the
previous inequality, and boundedness of $(u_{\epsilon })$ for the rest, we
deduce
\begin{multline*}
\frac{\epsilon ^{2}}{2}\left\Vert \rho \nabla _{X_{1}}(u_{\epsilon
}-u_{0})\right\Vert _{L^{2}(\omega _{1}^{\prime \prime }\times \omega
_{2})}^{2}+\left\Vert \rho \nabla _{X_{2}}(u_{\epsilon }-u_{0})\right\Vert
_{L^{2}(\omega _{1}^{\prime \prime }\times \omega _{2})}^{2} \\
+(\beta -K)\left\Vert \rho (u_{\epsilon }-u_{0})\right\Vert _{L^{2}(\omega
_{1}^{\prime \prime }\times \omega _{2})}^{2}\leq C\epsilon ^{2}
\end{multline*}
whence
\begin{equation*}
\left\Vert u_{\epsilon }-u_{0}\right\Vert _{L^{2}(\omega _{1}^{\prime
}\times \omega _{2})};\text{ }\left\Vert \nabla _{X_{2}}(u_{\epsilon
}-u_{0})\right\Vert _{L^{2}(\omega _{1}^{\prime }\times \omega _{2})}\leq
C^{\prime }\epsilon ,
\end{equation*}
where $C^{\prime }$ is independent of $\epsilon .$
\end{proof}
\subsection{Application to integro-differential problem}
In this section we provide some concrete examples. In \cite{7} M. Chipot and
S. Guesmia studied problem (\ref{6}) with the following integral operator
\begin{equation}
B(u)=a\left( \dint\limits_{\omega _{1}}h(X_{1},X_{1}^{\prime
},X_{2})u(X_{1}^{\prime },X_{2})dX_{1}^{\prime }\right) \label{39}
\end{equation}
To prove the convergence theorem\ the authors based their arguments on the
compacity of the operator $u\rightarrow \dint\limits_{\omega
_{1}}h(X_{1},X_{1}^{\prime },X_{2})u(X_{1}^{\prime },X_{2})dX_{1}^{\prime }
. Indeed, for a sequence $u_{n}\rightharpoonup u_{0}$ in $L^{2}(\Omega )$ we
have $\dint\limits_{\omega _{1}}h(X_{1},X_{1}^{\prime
},X_{2})u_{n}(X_{1}^{\prime },X_{2})dX_{1}^{\prime }$ $\rightarrow
\dint\limits_{\omega _{1}}h(X_{1},X_{1}^{\prime },X_{2})u_{0}(X_{1}^{\prime
},X_{2})dX_{1}^{\prime }$ in $L^{2}(\Omega )$ (by compacity) and we use the
continuity of $a$ and Lebesgue's theorem (under additional assumption on $a
) to get $a\left( \dint\limits_{\omega _{1}}h(X_{1},X_{1}^{\prime
},X_{2})u_{n}(X_{1}^{\prime },X_{2})dX_{1}^{\prime }\right) $ $\rightarrow
a\left( \dint\limits_{\omega _{1}}h(X_{1},X_{1}^{\prime
},X_{2})u_{0}(X_{1}^{\prime },X_{2})dX_{1}^{\prime }\right) $ in
L^{2}(\Omega ).$
We can give another operator\ based on the aforementioned one
\begin{equation}
B(u)=\dint\limits_{\omega _{1}}h(X_{1},X_{1}^{\prime
},X_{2})a(u(X_{1}^{\prime },X_{2}))dX_{1}^{\prime }, \label{40}
\end{equation}
For $a
\mathbb{R}
\rightarrow
\mathbb{R}
$ we note a Liptchitz function i.e there exists $K\geq 0$ such tha
\begin{equation}
\forall x,y\in
\mathbb{R}
:\left\vert a(x)-a(y)\right\vert \leq K\left\vert x-y\right\vert \label{37}
\end{equation}
In addition, we suppose that $a$ satisfies the growth condition
\begin{equation}
\exists q\in \left[ 0,1\right[ ,\text{ }M\geq 0,\text{ }\forall x\in
\mathbb{R}
:\left\vert a(x)\right\vert \leq M(1+\left\vert x\right\vert ^{q}),
\label{35}
\end{equation}
and we suppose that
\begin{equation}
h\in L^{\infty }(\omega _{1}\times \Omega ),\text{ }\nabla _{X_{1}}h\in L^
\frac{2}{1-q}}(\omega _{1}\times \Omega ) \label{36}
\end{equation}
\begin{theorem}
Consider problem (\ref{6}) with $B$ given by (\ref{39}) or (\ref{40}).
Assume (\ref{1}), (\ref{3}), (\ref{34}), (\ref{37}), (\ref{35}) , (\ref{36})
and for $\beta $ suitably chosen, then we have the affirmations of \textbf
theorems 1 , 2 }and\textbf{\ 3 }of section 1 and those of \textbf
propositions 8, 9}
\end{theorem}
\begin{proof}
Take $B$ as in (\ref{40}) the proof of this theorem amounts to prove that
assumptions (\ref{2}), (\ref{4}), (\ref{5}), (\ref{43}) and (\ref{44}) hold.
(\ref{2}) follows directly from (\ref{37}) and (\ref{36}), Now assume (\re
{35}), (\ref{36}) then we can check easily that (\ref{4}) holds with $r
\frac{2}{q}.$ It remains to prove that (\ref{5}) holds. For every $u\in V$ (
we can also take $u\in L^{2}(\Omega )$)$,$ and $\varphi \in \mathcal{D
(\Omega )$ we have for $1\leq k\leq p$
\begin{multline*}
I(\varphi )=\left\vert \dint\limits_{\Omega }\left( \dint\limits_{\omega
_{1}}h(X_{1},X_{1}^{\prime },X_{2})a(u(X_{1}^{\prime },X_{2}))dX_{1}^{\prime
}\right) \partial _{k}\varphi (X_{1},X_{2})dX_{1}dX_{2}\right\vert \\
=\left\vert \dint\limits_{\omega _{1}}\left( \dint\limits_{\Omega
}h(X_{1},X_{1}^{\prime },X_{2})\partial _{k}\varphi
(X_{1},X_{2})a(u(X_{1}^{\prime },X_{2}))dX_{1}dX_{2}\right) dX_{1}^{\prime
}\right\vert \\
\leq \dint\limits_{\omega _{1}}\left\vert \left( \dint\limits_{\Omega
}h(X_{1},X_{1}^{\prime },X_{2})\partial _{k}\varphi
(X_{1},X_{2})a(u(X_{1}^{\prime },X_{2}))dX_{1}dX_{2}\right) \right\vert
dX_{1}^{\prime }
\end{multline*}
Since $\partial _{k}h\in L^{\frac{2r}{r-2}}(\omega _{1}\times \Omega )$ it
follows that for a.e $X_{1}^{\prime }\in \omega _{1}:$ $\partial _{k}\left[
a(u(X_{1}^{\prime },.))h(.,X_{1}^{\prime },.)\right] \in L^{\frac{2r}{r-2
}(\Omega )$, integrating by part we get
\begin{eqnarray*}
I(\varphi ) &\leq &\dint\limits_{\omega _{1}}\left\vert \left(
\dint\limits_{\Omega }\partial _{k}h(X_{1},X_{1}^{\prime },X_{2})\varphi
(X_{1},X_{2})a(u(X_{1}^{\prime },X_{2}))dX_{1}dX_{2}\right) \right\vert
dX_{1}^{\prime } \\
&\leq &\left\Vert a(u)\right\Vert _{L^{r}}\left\vert \omega _{1}\right\vert
^{\frac{1}{2}}\left\Vert \partial _{k}h\right\Vert _{L^{\frac{2r}{r-2
}}\left\Vert \varphi \right\Vert _{L^{2}(\Omega )} \\
&\leq &M^{\prime }(1+\left\Vert u\right\Vert _{L^{2}})\left\Vert \varphi
\right\Vert _{L^{2}(\Omega )}
\end{eqnarray*}
And therefore $\partial _{k}B(u)\in L^{2}(\Omega )$, whence (\ref{43}) holds
and we hav
\begin{equation*}
\left\Vert \nabla _{X_{1}}B(u)\right\Vert _{L^{2}}\leq M^{\prime \prime
}(1+\left\Vert u\right\Vert _{L^{2}}),
\end{equation*
then for every $L^{2}-$bounded set $E\subset V$ we hav
\begin{equation}
\left\Vert \nabla _{X_{1}}B(u)\right\Vert _{L^{2}}\leq M^{\prime \prime
\prime },\text{ }u\in E. \label{41}
\end{equation
\ Now, given a sequence $(U_{n})$ in $conv(B(E))$ which converges strongly
to some $U_{0}$ in $L^{2}(\Omega )$, by (\ref{41}) and the convexity of the
norm we show that $(\nabla _{X_{1}}U_{n})_{n}$ is bounded in $L^{2}(\Omega )
, hence one can extract a subsequence $(U_{n})$ such that $(\nabla
_{X_{1}}U_{n})$ converges weakly to some $c_{0}$ in $L^{2}(\Omega )$, thanks
to the continuity of derivation on $\mathcal{D}^{\prime }(\Omega )$ which
gives $c_{0}=\nabla _{X_{1}}U_{0}$ and therefore, $U_{0}\in V$ , whence (\re
{5}) follows. Finally, one can check easily that (\ref{44}) holds. Same
arguments when $B$ is given by (\ref{39})
\end{proof}
\subsection{\protect\bigskip A generalization}
Consider (\ref{39}) wit
\begin{equation}
h\in L^{\infty }(\Omega ),l\in L^{\infty }(\omega _{1}),\nabla _{X_{1}}l\in
L^{2}(\omega _{1}), \label{45}
\end{equation}
the operator $u\rightarrow a\left( l(X_{1})\dint\limits_{\omega
_{1}}h(X_{1}^{\prime },X_{2})u(X_{1}^{\prime },X_{2})dX_{1}^{\prime }\right)
$ belongs to a class of operators defined b
\begin{equation}
B(u)=a\left( lP(u)\right) , \label{46}
\end{equation}
where $P:L^{2}(\Omega )\rightarrow L^{2}(\omega _{2})$ is a linear bounded
operator (an orthogonal projector for example). The method used by M. Chipot
and S. Guesmia is not applicable here, in fact the linear operator $P$ is
not necessarily compact, for $u_{n}\rightharpoonup u_{0}$ we only have
P(u_{n})\rightharpoonup P(u_{0})$ weakly and therefore every subsequence
(a\left( lP(u_{n})\right) )$ is not necessarily convergent in $L^{2}(\Omega )
$ strongly. However we have the following.
\begin{theorem}
Consider problem (\ref{6}) with $B$ given by (\ref{46}). Assume (\ref{1}),
\ref{3}), (\ref{34}), (\ref{37}), (\ref{35}) and (\ref{45}), then for $\beta
$ suitably chosen, we have affirmations of \textbf{Theorems 1 , 2 }an
\textbf{\ 3} of section 1 and moreover we have $u_{0}\in H^{1}(\Omega )$
\end{theorem}
\begin{proof}
The proof of this theorem amounts to prove that assumptions (\ref{2}), (\re
{4}), (\ref{5}) and (\ref{43}) hold. Since $P$ is Lipschitz then (\ref{2})
follows by (\ref{37}). We also can prove (\ref{4}) using (\ref{35}) with $r
\frac{2}{q}.$It remains to check that (\ref{5}), (\ref{43}) hold, for every
u\in V$ (we can take $u\in L^{2}(\Omega )$) we have $\nabla
_{X_{1}}a(lP(u))\in L^{2}(\Omega )$ and $\nabla _{X_{1}}a(lPu)=a^{\prime
}(lP(u))P(u)\nabla _{X_{1}}l.$ We can show easily that $\nabla
_{X_{1}}a(lP(E))$ is bounded for any $L^{2}-$bounded set $E\subset V$ and we
finish the proof as in \textbf{Theorem 8}.
\end{proof}
|
\section{INTRODUCTION}
\subsection{Motivation}
The advent of personalized medicine together with rapid progress in techniques of genetics, next generation sequencing for example, the use of biomarkers, together with analytic techniques in bioinformatics have brought a renewed focus on the problems of model--based prediction.
The related but different question concerning goodness of fit for any model has been given rather greater attention, at least in the survival literature.
$R^2$ measures quantify the predictive capacity of a model, and this may be high even when the model assumptions are seriously violated, whereas goodness--of--fit measures focus on the model assumptions and aim to examine how well these are supported by the data themselves. Although a number of authors have carefully outlined that distinction it is true that some confusion still remains. \newline
Our motivation stems from a study of $1504$ breast cancer patients treated at the Institut Curie in Paris, France. Subsequent to initial treatment, patients were followed for a period of fifteen years. Among several study objective relating to this cohort was the aim to construct descriptive survival models that could provide a deeper understanding to prognosis after initial treatment. The problem is inherently a multi--factorial one. Combined effects of prognostic factors as well as conditional effets are a central concern. By conditional, we mean the impact of various risk factors on survival after having taken account of the impact of known or suspected risk factors. For instance, it can be of interest to try to quantify the added prognostic information of a more or less complex construction of biomarkers after having already accounted for known clinical risk factors. Finally, the effect of several of these risk factors can change with time and useful prognostic indices should reflect such time dependencies.
\subsection{Background}
Goodness--of--fit procedures can be directed at more than one aspect of any model. We may wish to consider overall fit of the model, i.e., how well the model when taken as a whole is supported by the observations or we may wish to focus on some particular feature of the model and how well it holds up in practice. For example, we may be interested in checking the working assumptions regarding treatment differences in presence of other covariates when the model fit of these covariates is of only indirect concern.
The goodness of fit of a model can be evaluated by using tests or graphical methods. In this paper, we focus on graphical methods that can not only indicate departures from working assumptions but can also, of themselves, suggest remedies.
The first graphical method for checking the proportional hazards assumption was proposed by \citet{Kay1977} who suggested to plot an estimate of the conditional cumulative hazard $\Lambda(t \vert Z)$ over time. When $Z$ is a categorical covariate, typically representing treatment groups, the plot should result in parallel curves under proportional hazards. \citet{Andersen1982b} extended this approach to continuous covariates by discretizing them.
Other graphical methods based on residuals can be sorted in two categories, depending on whether the residuals are cumulated or not. Amongst non-cumulative methods, a large class of martingale residuals described by \citet{Barlow1988} can be used by plotting their members over time. The \citet{Schoenfeld1982} residuals, weighted Schoenfeld residuals introduced by \citet{Lin1993} and the residuals of \citet{Kay1977} arise as special members of this class. \citet{Grambsch1994} suggested plotting standardized residuals over time to detect the validity of proportional hazards assumption and, in case of rejection, have an indication on the shape of the time-varying effect. This is the most commonly used approach and is implemented in the programming languages R and Splus.
More recently, \citet{Sasieni2003} proposed the use of martingale difference residuals. The latter method requires care in interpretation since several plots corresponding to several time points have to be considered. All of these non--cumulative residual methods presented so far make use of a smoothing function to average the residual points. As pointed out by \citet{Lin1993}, the result can be sensitive to the choice of the smoothing techniques.
To overcome this problem, several authors proposed the use of cumulative martingale residuals, such as \citet{Arjas1988},
\citet{Therneau1990} and \citet{Lin1993}. The method of \citet{Therneau1990} is based on the score process of \citet{Wei1984}.
Under the proportional hazards assumption, this process converges weakly to a Brownian bridge and a test of the supremum of a Brownian bridge can be performed. \citet{Lin1993} showed that Wei's score process can be asymptotically approximated by a gaussian process with a data--based variance--covariance matrix. Therefore, the comparison between the observed score process and a large numbers of simulated outcomes of the limiting gaussian process can give an indication of the validity of the proportional hazards assumption. In practice, the interpretation of such a plot is not always clear.
More details about goodness--of--fit methods can be found in \citet{Klein2003}, \citet{Therneau2000} and more recently \citet{Martinussen2005}.\\
Unlike the case of linear regression, if the multivariate proportional hazards model holds, the sub--models will no longer be simultaneously valid. Therefore, the evaluation of the goodness of fit of the multivariate model by evaluating the fit of the univariate sub-models will not suffice. However, in absence of tools for checking the overall validity of the model, most of the existing methods for checking the fit of one covariate assume proportional hazards for the other covariates, which is an erroneous assumption (\citealt{Scheike2004}). Besides, the validity of the results of such methods depends on the covariance between covariates. To adress this issue, \citet{Scheike2004} considered a non--proportional hazards model and developed estimation procedures and tests of the goodness of fit for one covariate with the possibility for the others not to have a constant regression effect. Their simulation work indicates the good performance of their method when compared to several existing and commonly used methods
when the proportional hazards assumption is not met and/or in the presence of correlated covariates. Their test statistic depends on the estimation of the regression parameter requiring an involved algorithm relying on kernel estimation. The shape of the resulting estimator of the regression parameter is not an explicit and smooth function of time. The expression of the asymptotic distribution is unavailable for their statistic.
The goodness--of--fit evaluation procedure presented in this article is a graphical method which does not require any estimation and is simple to understand. Our method is also based on the general framework of a non--proportional hazards model and is adapted to multivariate settings with correlated covariates.
\\
Measures of predictive ability, on the other hand, - we will focus specifically on $R^2$ type measures - are used to examine several different questions. Typical questions may be, how well does some set of biological markers perform, in a predictive sense, when compared to some other set. How much added predictive information is contained in a biomarker when added to already known clinical prognostic factors such as stage and grade. When all known factors are included in a model, how much of the variability is accounted for so that, in consequence, how much variability remains to be explained, either by physical or possibly genetic attributes. Finally, how does the relaxing of certain model assumptions - one example would be stratification rather than inclusion in the linear component of a proportional hazards model - impact prediction. This last observation draws attention to the fact that, although different techniques with a different purpose, the aims of goodness--of--fit procedures and predictive measures can to some degree overlap.
In the context of survival analysis, in particular when using the Cox proportional hazards model, several authors have proposed different measures of predictive ability.
A recent and exhaustive literature review on the predictive accuracy measures can be found in \citet{Choodari2012}.
No consensus has yet been established regarding the most suitable measure to use in practice (\citealt{Muller2008}, \citealt{Hielscher2010}, \citealt{Choodari2012}). \newline
It is not clear in what way, or in what sense, an improvement in predictability implies an improvement in goodness of fit. In fact it is not difficult to come up with counter examples and the notion itself is not very precise. The converse is however correct, and, in this work, we prove in a theorem that an improvement in fit of a proportional hazards model results in an improvement in predictability. This theorem underlies the purpose of this article which is
to investigate ways to improve goodness--of--fit for proportional hazards type models and to see how this impacts the resulting predictive power of the model. We work with goodness--of--fit procedures and measures of predictive ability that are closely related, having as their basis the residuals from the non--proportional hazards model. The goodness of fit is evaluated with a version of the score process introduced by O'Quigley (\citeyear{OQuigley2003}, \citeyear{OQuigley2008} chap. 8) which is extended here to the multivariate setting. We obtain the exact expression of the limiting distribution of the process. The predictive accuracy measure is the $R^2$ coefficient described by \citet{OQuigley1994} but is also extended to the multivariate non--proportional hazards situation.
This leads to easily assessed visual techniques and provides a complete and unified approach to the testing, fit and quantification of predictive effects. Several examples illustrate the ideas.
In the next section we describe the non--proportional hazards model and use it to derive stochastic processes of particular relevance to the problem we are studying. In Section 3, we present the main result that indicates why improvements in fit will result in improvements in predictive capability and how to proceed in practice. Section 4 summarizes simulations that provide additional support to our intuition and an application to a real dataset is provided. Before that, we recall the main notation.
\subsection{Notation}
The random variables of interest are the failure times $T_i$, the
censoring times $C_i$ and the vector of dimension $p$ of possibly time-dependent covariates
$\mathbf Z_i=(Z_i^1,\dots,Z_i^p)$, $i=1,...,n.$ We view these as a random sample from
the distribution of $T$, $C$ and $\mathbf Z=(Z^1,\dots,Z^p)$ which
have support on some finite interval. To emphasis the time-dependence, with a slight abuse of notation, we refer to any time-dependent quantity $A$ as $A(t)$, $A$ being either random or deterministic.
The time-dependent covariate $\mathbf Z(t) $ is assumed to be a predictable stochastic process which admits a moment of order 4.
For each subject $i$, the observed time is $X_i=\min(T_i, C_i)$, and the observed indicator of failure is
$\delta_i= I(T_i\leq C_i)$, where $I$ is the indicator function. The at--risk indicator $Y_i(t)$ is
defined as $Y_i(t)= I(X_i\geq t).$ The counting process $N_i(t)$
is defined as $N_i(t)=I(T_i\leq t, T_i\leq C_i)$ and we also
define $\bar{N}(t)=\sum_{i=1}^n N_i(t)$.
It is of
notational convenience to define ${\cal Z }(t)= \sum_{i=1}^n
\mathbf Z_i(t)I(X_i=t,\delta_i=1)$, in words a $\mathbb R^p$-valued function equal
to zero except at the observed failures where it assumes
the covariate value of the subject that fails.
In addition, $\|\mathbf{a}\|=\underset{i=1,\dots,p}{\max}|a_i| $ denotes the maximum norm of the vector $\mathbf{a}=(a_1,\dots,a_p) \in \mathbb R^p$. For a $p\times p$ matrix $\mathbf{A}$ with element $(i,j)$ denoted $A_{i,j}$, $i,j=1,\dots,p$, $\|\mathbf{A}\|=\underset{i,j=1,\dots,p}{\max}|A_{i,j}| $ denotes the maximum
norm of $\bf A$. Let ${\bf A}^T$ (respectively ${\bf a}^T$) denote the transpose of the matrix $\mathbf{A}$ (resp. vector $a$). The product $\mathbf{a}^{\otimes 2}=\mathbf{aa}^T$ is the matrix with element $[\mathbf{a}^{\otimes 2}]_{i,j}=a_ia_j$. Denote $\det(\mathbf{A})$ the determinant of the matrix $\bf A$. The space $D[0,1]^p=D[0,1]\times \dots \times D[0,1]$ is equipped with the Skorokhod product topology.
\section{MODEL-BASED EMPIRICAL PROCESSES}
Consider the non--proportional hazards model defined by
\begin{equation}
\label{nonph}
\lambda\left\{t\mid\mathbf Z(t)\right\}=\lambda_0(t)\exp\left\{\boldsymbol \beta(t)^T \mathbf Z(t)\right\},
\end{equation}
where
$\lambda(t|\cdot)$ is the conditional hazard function,
$\lambda_0(t)$ is a baseline hazard, $\boldsymbol \beta(t)$ is the time--dependent
regression effect and has dimension $p$ and $\boldsymbol \beta(t)^T \mathbf Z(t)$ is the usual inner product between $\boldsymbol \beta(t)$ and $\mathbf Z(t)$. This model has been considered previously by several authors \citep{Murphy1991,Hastie1990,Zucker1990,Cai2003,Winnett2003,Scheike2004}.
With covariates constant over time,
the above model becomes the proportional hazards model \citep{Cox1972} under the restriction that $\boldsymbol \beta(t)=\boldsymbol \beta$.
When we take the risk sets to be fixed and known and conditional on a failure at time $t$, the probability that the failure concerns individual $i$ is
\begin{equation}
\label{2.1} \pi_i(\boldsymbol \beta(t),t) = { Y_i(t) \exp\{
\boldsymbol \beta(t)^T \mathbf Z_i(t) \} }/\sum_{j=1}^n Y_j(t)\exp\{ \boldsymbol \beta(t)^T \mathbf Z_j(t) \}
,\quad i=1,\dots,n.
\end{equation}
The expectation and variance with respect to the probabilities $\{\pi_i(\beta(t),t)\}_{i=1,\dots,n}$ are respectively a vector $\mathbf E_{\boldsymbol \beta(t)}\left(Z\vert t\right)$ of dimension $p$ and a $p\times p$ matrix $\mathbf V_{\boldsymbol \beta(t)}\left(Z\vert t\right)$ such that,
\begin{align*}
\mathbf E_{\boldsymbol \beta(t)}\left(Z\vert t\right) &= \sum_{i=1}^n Z_i(t) \pi_i( \boldsymbol \beta(t),t), \\
\mathbf V_{\boldsymbol \beta(t)}\left(Z\vert t\right) &= \sum_{i=1}^n Z_i(t)^{\otimes 2} \pi_i( \boldsymbol \beta(t),t) -\mathbf E_{\boldsymbol \beta(t)}\left(Z\vert t\right)^{\otimes 2}.
\end{align*}
These quantities correspond to the conditional moments of the process $\mathcal{Z}(t)$ for a fixed $t$, given the risk sets.
The conditional variance-covariance matrix $\mathbf V_{\boldsymbol \beta(t)}\left(Z\vert t\right)$ is symmetric and positive definite. Thus, there exists an orthogonal matrix $\mathbf P_{\boldsymbol \beta(t)}(t)$ and a diagonal matrix $\mathbf D_{\boldsymbol \beta(t)}(t)$ such that
\begin{equation*}
\mathbf V_{\boldsymbol \beta(t)}\left(Z\vert t\right)=\mathbf P_{\boldsymbol \beta(t)}(t)\mathbf D_{\boldsymbol \beta(t)}(t)\mathbf P_{\boldsymbol \beta(t)}(t)^T.
\end{equation*}
This leads us to define the symmetric matrix $\mathbf V_{\boldsymbol \beta(t)}\left(Z\vert t\right)^{x}$ by
\begin{equation*}
\mathbf V_{\boldsymbol \beta(t)}\left(Z\vert t\right)^{x}=\mathbf P_{\boldsymbol \beta(t)}(t)\left(\mathbf D_{\boldsymbol \beta(t)}(t)\right)^{x}\,\mathbf P_{\boldsymbol \beta(t)}(t)^T,\quad x\in\{-1/2,1/2\}.
\end{equation*}
Denote \begin{equation}
\mathbf{r}_{\boldsymbol \beta(t)}(t)={\cal Z}(t)-\mathbf E_{\boldsymbol \beta(t)}(Z|t), \label{schoenfeldres}
\end{equation}
the residuals of the non--proportional hazards model (\ref{nonph}) with parameter $\boldsymbol \beta(t)$ evaluated at time $t$. In the case of the proportional hazards model, these residuals reduce to the well-known Schoenfeld residuals (\citealt{Schoenfeld1982}). Assume the case of a unique covariate ($p=1$) resulting in a univariate regression coefficient $\beta(t)$, a univariate conditional expectation $E_{\beta(t)}(Z\vert t)$ and a univariate residual $r_{\beta(t)}(t)$. Consider the partial scores
\begin{equation}
\label{gof1} U(\beta(t),t) = \int_0^t r_{\beta(s)}(s
d\bar N(s).
\end{equation}
With a constant regression effect $\beta$, these correspond to the partial scores of \citet{Wei1984}.
Wei was interested in goodness of fit for the two group problem and based a test on $\sup_t |
U(\hat\beta,t)|$, large values indicating departures away from
proportional hazards in the direction of non--proportional
hazards. Considerable exploration of this idea, and substantial
generalization via the use of martingale--based residuals, has
been carried out by
\citet{Lin1993,Lin1996} who showed that a wide choice of functions, potentially describing different
kinds of departures from the model could be used. Apart from the
two--group case, limiting distributions are complicated and
usually approximated via simulation. Furthermore, \citet{Lin1993} pointed out that extensions of their methodology to the multivariate case or to the integration of time--dependent covariates are not straightforward.
In order to overcome these difficulties, we follow the construction developed by \citet{Khmaladze1981}, working directly with the increments of the process
rather than the process itself. We are then able to derive related processes
for which the limiting distributions are available analytically. To be more specific, when working with the ranks of the failure times and standardizing each increment of the process with a particular value rather than applying the same standardization for the whole score process, the limiting distribution of the multivariate process can be anticipated analytically and time-dependent covariates can be directly taken into account.
\subsection{Time Scale}
Let $k_n=\#\{i : i=1,\dots,n,\ \delta_i=1, \, \det\left (\mathbf V_{\boldsymbol \beta(X_i)}(Z|X_i)\right)>0\}$, where $\#A$ denotes the cardinality of the set $A$, denote the number of observed failures such that the conditional variances assessed at the event-times are positive--definite matrices.
In our setting, a null conditional variance at any time implies null conditional variances at later times. We assume that the number of failures $k_n$ increases without bound as $n$ increases without bound. By virtue of the fact that in Equation (\ref{nonph}), $\lambda_0(t)$ is unspecified, a monotonically increasing transformation of the times leaves inference for the regression parameter of the proportional hazards model unchanged. Therefore, \citet{Chauvel2014} considered the transformed times $\phi_n(X_i)$ such that
\begin{equation}\label{transfo_phi}
\phi_n(X_i)= \dfrac{ \bar N(X_i) }{k_n}\left(1+ (1-\delta_i) \dfrac{\# \left\lbrace j:j=1,\dots,n, X_j<X_i,\ \bar N(X_j)=\bar N(X_i) \right\rbrace}{ \#\left\lbrace j:j=1,\dots,n, \bar N(X_j)=\bar N(X_i)\right\rbrace}\right).
\end{equation}
Recall that the counting process $\{\bar N(t)\}_{t \in \mathcal{T} }$ presents a unit jump at each observed failure time. On the new scale, the times in the set $\{0,1/k_n,2/k_n,\dots,1\}$ correspond to failure times, the $i$th ordered failure time, denoted $t_i$, is such that $t_i=i/k_n$. The set $\{0,1/k_n,2/k_n,\dots,1\}$ is included in but not equal to the set of images of all failure times.
Censoring times can assume any value as long as they keep their original locations between adjacent failure times. For simplicity in Formula (\ref{transfo_phi}), we take these times to be spread uniformly between adjacent failure times, maintaining the original ranking. The time $t_0$ on this scale corresponds to the $100\times t_0$th percentile of failure in the sample. For instance, at time $t_0=0.5$, half of the failures are observed. The inverse transformation of $\phi_n$ can be easily obtained and would enable us to interpret the results on the original time scale.
On this transformed time scale, we can define the at-risk indicator $ Y_i^*(t) $ by $ Y_i^*(t)=I(\displaystyle \phi_n(X_i)\geq t)$ and the individual counting process $ N_i^*(t) =I(\displaystyle \phi_n(X_i) \leq t, \delta_i=1)$, for individual $i=1,\dots,n$. In what follows, we only work with the standardized time scale, so that the process $\mathcal{Z}$, the expectation $\mathbf E_{\boldsymbol \beta(t)}(Z|t)$ and the variance $\mathbf V_{\boldsymbol \beta(t)}(Z|t)$, of which extensions are straightforward, are defined for $0\leq t\leq 1$.
Define the counting process associated with the transformed times which has unit jumps at failure--times on the new scale by $$ \bar N^* (t)=\sum_{i=1}^n I( \phi_n(X_i) \leq t, \delta_i=1),\quad 0\leq t\leq 1.$$
On the new time scale, the
partial scores (\ref{gof1}) can be re-expressed as
\begin{equation*}
\mathbf U(\beta(t),t) =\int_0^{t} \mathbf r_{\boldsymbol \beta(s)}(s)
d\bar N^*(s)=\sum_{i=1}^{\lfloor k_n t \rfloor}
\mathbf r_{\boldsymbol \beta(t_i)}(t_i)
, \quad \quad 0\leq t \leq 1,
\end{equation*}
where the $i$th element of the vector $\int_0^t\mathbf{a}(s)d\bar N^*(s)$ is $\int_0^t\mathbf{a}_i(s)d\bar N^*(s)$ for any $\mathbb R^p$-valued $\mathbf{a}(t)=(a_1(t),\dots,a_p(t))$, $i=1,\dots,p$ and
$\lfloor x \rfloor$ gives the largest integer less than or equal to $x$.
\subsection{Multivariate Standardized Score Process}
\label{sec:process}
Before defining the standardized score process, let us give the assumptions needed in the sequel.
Let $t \in [0,1]$, $\g(t)$ be a regression function, not necessarily equals to $\boldsymbol \beta(t)$ and
\begin{align*}
S^{(0)}(\g(t),t) = n^{-1} \sum^n_{i=1}
Y_i(t)&e^{\g(t) Z_i(t)}, \quad \quad
\mathbf{S}^{(1)}(\g(t),t) = n^{-1} \sum^n_{i=1}
Y_i(t)Z_i(t)e^{\g(t) Z_i(t)}, \\
\mathbf{S}^{(2)}(\g(t),t) &= n^{-1} \sum^n_{i=1}
Y_i(t)Z_i(t)^{\otimes 2}e^{\g(t) Z_i(t)}.
\end{align*}
Using these notations, we have the equalities
$\mathbf E_{\g(t)}(Z\vert t)=\mathbf{S}^{(1)}(\g(t),t)/S^{(0)}(\g(t),t)$
and
$\mathbf V_{\g(t)}(Z \vert t)=\mathbf{S}^{(2)}(\g(t),t)/S^{(0)}(\g(t),t) -\mathbf E_{\g(t)}(Z\vert t)^{\otimes 2}$. Notice that the Jacobian matrix of $\mathbf E_{\g(t)}(Z\vert t)$ is the variance--covariance matrix $\mathbf V_{\g(t)}(Z \vert t)$. Consider that the following assumptions, similar to those of \citet{Andersen1982} hold:
\begin{enumerate}[A.]
\item \label{point1_multiv} (Asymptotic stability). There exists a neighbourhood $\mathcal{B}$ of $\boldsymbol \beta(t)$ and vector and matrix functions $\mathbf{s}^{(r)}(\g(t),t)$, $r=0,1,2$, defined for $t\in[0,1]$ and $\g(t)\in \mathcal{B} $ such that $\bf 0$ and $\boldsymbol \beta(t)$ are in the interior of $\mathcal{B}$, for all $t \in [0,1]$ and
\begin{eqnarray*}
\sqrt n \sup_{t\in[0,1],\g(t)\in\mathcal{B}} \left\| \mathbf{S}^{(r)}(\g(t),t)-\mathbf{s}^{(r)}(\g(t),t)\right\| \underset{n \rightarrow \infty}{\overset{ \mathbb{P}}{\longrightarrow}} 0 .
\end{eqnarray*}
\item (Asymptotic regularity). \label{point2_multiv} All functions defined in assumption \ref{point1_multiv}. are uniformly continuous in $t\in [0,1]$. In addition, for $r=0,1,2$, $s^{(r)}(\g(t),t)$ are continuous functions of $\g(t) \in \mathcal{B}$, bounded on $\mathcal{B}\times [0,1]$ and $s^{(0)}(\g(t),t)$ is bounded away from $0$.
\item (Homoscedasticity). \label{homosced_multiv}
There exists a symmetric and positive definite matrix $\boldsymbol \Sigma$ and a series of positive constants $(M_n)_n$ converging to $0$ as $n$ goes to infinity such that
\begin{eqnarray*}
\sup_{t\in[0,1],\g(t)\in\mathcal{B}} \left\| \left.\dfrac{\partial}{\partial \boldsymbol \beta } \mathbf V_{\boldsymbol \beta}(Z \vert t)\right\vert_{\boldsymbol \beta=\g(t)} \right\| &\leq& M_n \quad a.s.,
\\
\sup_{t\in[0,1],\g(t)\in\mathcal{B}} \left\| \mathbf V_{\g(t)}(Z \vert t) - \boldsymbol \Sigma \right\| &\underset{n \rightarrow \infty}{\overset{\bf L^1}{\longrightarrow}}& 0.
\end{eqnarray*}
\end{enumerate}
By analogy with the empirical quantities, we denote $e(\g(t),t)=s^{(1)}(\g(t),t)/s^{(0)}({\bf 0},t)$ and
$v(\g(t),t)=s^{(2)}(\g(t),t)/s^{(0)}({\bf 0},t)-e(\g(t),t)^{\otimes 2}$.
The two first conditions are classical and introduced by \citet{Andersen1982} for using counting process and martingales theory, such as Lenglart's inequality or Rebolledo's theorem. Although we use a different approach here, that we believe is simpler to comprehend, the same assumptions are made.
Notice that $\mathbf V_{\g(t)}(Z \vert t)$ is, by definition, the sample-based variance of $Z$ given $T=t$ under the model with parameter $\g(t)$. Thus, condition \ref{homosced_multiv}. of homoscedasticity means that the asymptotic variance does not depend on time. This condition is implicitly used in the context of the proportional hazards regression, for instance when estimating the variance of the parameter $\boldsymbol \beta$ or when applying the log-rank test. The contribution to the global variance is the same at each failure time, by the use of an unweighted sum of each term.
This stability of variance has also been noticed by several authors, for example \citet{Grambsch1994}.
From the previous section, under the non--proportional hazards model (\ref{nonph}), the increments of the process $ \sum_{j=1}^{\lfloor k_n t \rfloor} \mathcal{Z}(t_j) $ at $t=t_i$ have mean ${\mathbf E}_{\boldsymbol \beta(t_i)}(Z|t_i)$ and
variance-covariance matrix ${\mathbf V}_{\boldsymbol \beta(t_i)} \left(Z\vert t_i\right).$ The increments of the process are
independent, either by design in view of the conditional model, or by the arguments of \citet{Cox1975}. Thus only the
existence of the variance is necessary to carry out
appropriate standardization and to be able to appeal to the
functional central limit theorem. This leads us to define a standardized version of the multivariate score process:
\begin{definition}
The multivariate standardized score process evaluated at parameter $\boldsymbol \beta_0$ and at failure time $t\in\{ 0,1/k_,2/k_n\dots,1\}$ is
\begin{equation*}
\mathbf U^*(\boldsymbol \beta_0,t)=\dfrac{1}{\sqrt{k_n}} \int_0^{t} \mathbf V_{\boldsymbol \beta_0} \left(Z\vert s\right)^{-1/2}\mathbf r_{\boldsymbol \beta_0}(s)
d\bar N^*(s)
=\dfrac{1}{\sqrt{k_n}}\sum_{i=1}^j {\mathbf V}_{\boldsymbol \beta_0} \left(Z\vert t_i\right)^{-1/2} \mathbf{r}_{\boldsymbol \beta_0}(t_i).
\end{equation*}
\end{definition}
The $\mathbf U^*$ process is only defined on $k_n$ equispaced points of the interval
$[0,1]$ but we extend our definition to the whole interval via
linear interpolation so that, for $u$ in the interval $[t_j,t_{j+1}[$, we write
$$\label{brown2}
\mathbf U^*\left(\boldsymbol \beta_0,
u\right) = \mathbf U^*\left(\boldsymbol \beta_0, t_j\right) +
\left\{uk_n-j\right\} \left\{\mathbf U^*\left(\boldsymbol \beta_0,
t_{j+1}\right)- \mathbf U^*\left(\boldsymbol \beta_0,t_j\right)
\right\}.
$$
The following theorem gives the asymptotic behaviour of $\mathbf U^*(\boldsymbol \beta_0,\cdot)$:
\begin{theorem}
\label{thNPH_multi}
Under the non--proportional hazards model of parameter $\boldsymbol \beta(t)$, we have the following convergence in distribution:
\begin{equation}
\quad \mathbf U^*(\boldsymbol \beta_0,\cdot)-\sqrt{k_n} \mathbf C_n\overset{\cal D}{\underset{n\rightarrow \infty}{\longrightarrow}} \mathbf W_p,
\label{derivebrownien_multiv}
\end{equation}
where $\mathbf W_p$ is a standard Brownian motion of dimension $p$ and, for all $t \in [0,1]$,
\begin{equation*}
\mathbf C_n(t)=\frac{1}{k_n}\sum_{i=1}^{\lfloor t k_n \rfloor} {\mathbf V}_{\boldsymbol \beta_0} \left(Z\vert t_i\right)^{-1/2}\left\{
{\mathbf E}_{\boldsymbol \beta_0} \left(Z\vert t_i\right)
-
{\mathbf E}_{\boldsymbol \beta(t_i)} \left(Z\vert t_i\right)
\right\}.
\end{equation*}
In addition, we have the convergence of probability
\begin{equation}
\label{CV_derive_multi}
\sup_{t \in [0,1]} \left\| \mathbf C_n(t)- \boldsymbol{ \Sigma}^{1/2}\int_0^t\left\{\boldsymbol \beta(s)-\boldsymbol \beta_0\right\}ds \right\| \overset{P}{\underset{n\rightarrow \infty}{\longrightarrow}} 0,
\end{equation}
where $\int_0^ta(s)ds\!=\!\left(\!\int_0^ta_1(s)ds,\dots,\int_0^ta_p(s)ds\!\right)$ for any $\mathbb R^p$--valued function $a\!=\!(a_1,\dots,a_p)$.
\end{theorem}
The proof is given in Appendix \ref{appendix_assproof} and is based on the multivariate functional central limit theorem of \citet{Helland1982}.
The second term of formula (\ref{derivebrownien_multiv}) increases without bound as the sample size goes to infinity. In practical situations, when the model generating the observations is based on $\boldsymbol \beta(t)$, Theorem \ref{thNPH_multi} in addition to Slutsky's lemma indicate that $\mathbf U^*(\boldsymbol \beta_0,\cdot)$ will look like a multivariate Brownian motion with an added drift term:
\begin{corollary}
\label{corNPH}
Under the model (\ref{nonph}) with parameter $\boldsymbol \beta(t)$, we have, for all $ \boldsymbol \beta_0$,
\begin{equation*}
\mathbf U^*(\boldsymbol \beta_0,\cdot) - \sqrt{k_n}\, \boldsymbol{ \Sigma}^{1/2} IB \overset{\cal D}{\underset{n\rightarrow \infty}{\longrightarrow}}
\mathbf W_p ,
\end{equation*}
where $IB(t)=\int_0^t\left\{\boldsymbol \beta(s)-\boldsymbol \beta_0\right\}ds,\ 0 \leq t \leq 1$.
In addition, $\hat{\boldsymbol{ \Sigma}}=k_n^{-1}\sum_{i=1}^{k_n}\mathbf V_{\boldsymbol \beta_0}(Z\vert t_i)$ is a consistent estimator of $\boldsymbol{ \Sigma}$, and
\begin{equation}
\label{derive_beta}
\hat{ \boldsymbol{ \Sigma}}^{-1/2}\mathbf U^*(\beta_0,\cdot) -\sqrt{k_n}IB \overset{\cal D}{\underset{n\rightarrow \infty}{\longrightarrow}} \boldsymbol{\Sigma}^{-1/2}\mathbf W_p.
\end{equation}
\end{corollary}
In the sequel, the standardized score process is evaluated in $\boldsymbol \beta_0=\bf 0$, where $ \bf 0$ is the null vector of $\mathbb R^p$. As a consequence, the plot of $\hat{ \boldsymbol{ \Sigma}}^{-1/2}\mathbf U^*({\bf 0},t) $ against the time $t$ gives an indication on the shape of $\int_0^t\boldsymbol \beta(s)ds$ which is reflected by the shape of the drift of the process (equation (\ref{derive_beta})). In the univariate case ($p=1$) or in the multivariate case with independent covariates, the process $\mathbf U^*({\bf 0},\cdot)$ can be directly plotted over time, with no additional standardization since $\Sigma$ is the identity matrix. However, when dealing with correlated covariates, a global standardization is needed for isolating each effect $\beta_i(t)$ on each process $ \left[ \hat{ \boldsymbol{ \Sigma}}^{-1/2}\mathbf U^*(\beta_0,\cdot) \right]_i$, $i=1,\dots,p$. A linear drift corresponds to a constant over time regression effect. Our proposed method takes into account the covariances between all covariates and the goodness--of--fit of the overall model is directly evaluated instead of checking proportionality of hazards for one covariate at a time.
Illustrations are given in the univariate case. Figure \ref{figure_beta0} represents a simulation of the process $U^*(0,t)$ over time $t$, under the model with a null regression parameter $\beta(t)=0$. Even under moderate to small sample size, the Brownian motion approximation appears accurate enough for reliable inference.
\begin{figure}[ht]
\begin{center}
\subfigure[$\beta=0$]{
\includegraphics[width=2.8in,height=2.5in]{score_standardise_beta0.pdf}
\label{figure_beta0}}
\subfigure[$\beta=0.5$]{
\includegraphics[width=2.8in,height=2.5in]{score_standardise_beta05.pdf}
\label{figure_beta0.5}}
\caption{Processes $U^*(0,t)$ based on data simulated from proportional hazards models of parameter $\beta$.}
\end{center}
\end{figure}
\begin{figure}[ht]
\begin{center}
\subfigure[$\beta(t)=I(t \leq 0.5)$]{
\includegraphics[width=2.8in,height=2.5in]{score_standardise_beta_1_0.pdf}
\label{figure_beta_1_0}}
\subfigure[$\beta(t)=I(t \leq 1/3)+0.5I(t>2/3)$]{
\includegraphics[width=2.8in,height=2.5in]{score_standardise_beta_1_0_05.pdf}
\label{figure_beta_1_0_05}}
\caption{Processes $U^*(0,t)$ based on data simulated from models (\ref{nonph}) of parameter $\beta(t)$.}
\end{center}
\end{figure}
Consider a proportional hazards model with $\beta(t)$ constant over time but not null. Corollary \ref{corNPH} suggests that a
good approximation for this process is a Brownian motion with a linear drift. An indication of the plausibility of this is shown in Figure \ref{figure_beta0.5}, where $\beta(t)$ is set to $0.5$.
Departures from the proportional hazards assumption can be of various forms. For instance, the effect can be constant and then decreasing after some time $\tau$, the effect can be piecewise constant over time or it can increase over time.
Corollary \ref{corNPH} implies that the shape of the drift of the process $U^*(0,\cdot)$ will reflect the shape of the cumulated regression coefficient.
As an illustration, Figure \ref{figure_beta_1_0} represents a simulated process under the non--proportional hazards model, with $\beta(t)$ piecewise constant. Before $t=0.5$, there is a linear trend corresponding to $\beta=1$ and for $t>0.5$, $\beta$ equals zero and the process $U^*(0,t)$ is constant in expectation over time. Figure \ref{figure_beta_1_0_05} represents a simulated standardized score process for a changepoint model with the regression parameter $\beta(t)=I(t \leq 1/3)+0.5I(t>2/3) $. The trend of the process can be separated into 3 straight lines reflecting the strength of the effect: the slope of the first part seems twice higher than the one of the last part and the slope of the second part is null.
The following proposition enables the construction of a confidence band for each process:
\begin{proposition} \label{confband}
Let $i=1,\dots,p$. Consider the hypothesis $H_{0,i}:\exists\, b_i,\,\beta_i(t)=b_i$ and its alternative $H_{1,i}:\nexists\, b_i,\,\beta_i(t)=b_i $.
Under the model (\ref{nonph}) of parameter $\boldsymbol \beta(t)=(\beta_1(t),\dots,\beta_p(t))$ not necessarily equals to $\boldsymbol \beta_0$ and $H_{0,i}$, we have, for all $a\geq 0$,
\begin{align} \label{CB_proc}
\lim_{n\rightarrow + \infty} &\text{P}\left( \left\| \hat{\Sigma}_{\cdot,i}^{-1/2} \right\|_2^{-1}\sup_{t\in[0,1]} \left\vert \left(\hat{ \boldsymbol{ \Sigma}}^{-1/2}\left\{\mathbf U^*(\boldsymbol \beta_0,t) -t\mathbf U^*(\boldsymbol \beta_0,1)\right\}\right)_i \right\vert \leq a\right) \nonumber \\
&=\text{P}\left( \sup_{t\in[0,1]} \left\vert B(t)\right\vert \leq a\right),
\end{align}
where $B$ is a Brownian bridge and $\left\| \hat{\Sigma}_{\cdot,i}^{-1/2} \right\|_2=\left(\sum_{j=1}^p\left(\hat{\Sigma}_{j,i}^{-1/2}\right)^2 \right)^{1/2}$. Therefore, by denoting $a(\alpha)$ the quantile of order $\alpha$ of the Kolmogorov distribution, we have
\begin{equation*}
\lim_{n\rightarrow + \infty }\text{P}\left(\forall t \in[0,1],\ \left[\hat{\Sigma}^{-1/2}\mathbf U^*(\beta_0,t)\right]_i \in IC_i(\alpha)\right) =1- \alpha,
\end{equation*}
with
\begin{equation*}
IC_i(\alpha)=\left[t\left[\hat{\Sigma}^{-1/2}\mathbf U^*(\boldsymbol \beta_0,1)\right]_i - \left\| \hat{\Sigma}_{\cdot,i}^{-1/2} \right\|_2 a(\alpha) ;t\left[\hat{\Sigma}^{-1/2}\mathbf U^*(\boldsymbol \beta_0,1)\right]_i -\left\| \hat{\Sigma}_{\cdot,i}^{-1/2} \right\|_2a(\alpha)\right].
\end{equation*}
\end{proposition}
The proof can be found in Appendix \ref{appendix_confband}. If the $i$th element of the process $ \hat{\Sigma}^{-1/2}\mathbf U^*(\beta_0,t)$ leaves the confidence band $IC(\alpha)_i$, we reject the hypothesis that the effect $\beta_i(t)$ is constant over time with an asymptotic level of $\alpha$. However, when testing simultaneously several hypotheses $H_{0,i}$ of constant effects for different covariates, the global type I error is inflated. This means that one process could leave its confidence band whereas the corresponding effect is constant over time with a level higher than $\alpha$. This does not seem to be a problem since the plot of the confidence interval is just one of the tools we use to select the variables respecting the proportional hazards assumption. We do not base a definitive conclusion regarding this assumption on this confidence band only, and the non--detection of a constant effect will be corrected with the other steps of the selection variable method we propose in this article.
Of course, corrections for multiple testings could be applied.
Whether effects are of a proportional hazards or a non--proportional hazards form, essentially, all of the information concerning the regression effect $\boldsymbol \beta(t)$ is captured in the process $\mathbf U^*(\mathbf{0},\cdot)$. The process allows the data to speak for themselves, not unlike a scatterplot in linear regression, in which trends and non-linearity may be apparent, since we evaluate the process at $\boldsymbol \beta_0= \bf 0$. No parameter has to be estimated and expectations and variance-covariance matrices are the usual sequential empirical quantities.
The process, based on the residuals of the non--proportional hazards model, is a useful tool in the evaluation of its goodness--of--fit. These residuals can also be used in the construction of a predictive accuracy measure of the model.
\FloatBarrier
\section{INTERPLAY OF FIT AND PREDICTION}
\subsection{$R^2$ Coefficient as a Measure of Predictive Ability}
For any random variables $X$ and $Y$ having second moments, the formula
\begin{equation}
\label{VarDecomp}
\text{Var}(Y)=E(\text{Var}(Y\vert X))+\text{Var}(E(Y \vert X)),
\end{equation}
leads to the natural definition of explained variation as the ratio of the variance of the expected values of the response variable under the model given the explanatory variables to the marginal variance of the response variable. In light of the Chebyshev inequality, we see that explained variation directly quantifies predictive strength.
In the non--proportional hazards model (\ref{nonph}) with one covariate $Z(t)$ ($p=1$), the explained variation makes use of the variance decomposition given in equation (\ref{VarDecomp}) in which $Y$ is replaced by $Z(t)$ and $X$ by $T$, leading to the definition:
\begin{definition} In the univariate non--proportional hazards model (\ref{nonph}), the explained variation, expressed as a function of the time-dependent regression coefficient $\beta(t)$, is defined by
\begin{equation*}
\Omega^2\left(\beta(t)\right)=\dfrac{{\rm Var}(E(Z|T))}{\rm{Var}(Z)} =1-\dfrac{E(\rm{Var}(Z|T))}{\rm{Var}(Z)}.
\end{equation*}
\end{definition}
In the multivariate non--proportional hazards model (\ref{nonph}), individual $i$ is characterized by its real-valued prognostic index $\eta_i(t)=\boldsymbol \beta(t)^T\mathbf Z_i(t)$, being a realization of $\eta(t)=\boldsymbol \beta(t)^T\mathbf Z(t)$. Therefore it is equivalent to evaluate the quality of prediction of the model via $\mathbf Z$ or $\eta$. We adopt the latter possibility.
\begin{definition} The explained variation of the non--proportional hazards model (\ref{nonph}) with multiple covariates can be defined by a function of the time-dependent regression coefficient $\boldsymbol \beta(t)$ by
\begin{equation*}
\Omega^2\left(\eta(t)\right)=\dfrac{{\rm Var}(E(\eta|T))}{\rm{Var}(\eta)} =1-\dfrac{E(\rm{Var}(\eta|T))}{\rm{Var}(\eta)}, \quad \eta(t)=\boldsymbol \beta(t)^T\mathbf Z(t).
\end{equation*}
\end{definition}
Some properties of $ \Omega^2$ can be found in \citet[chap. 13]{OQuigley2008} or in \citet[chap. 27]{OQuigley2012}. In these book chapters, $\Omega^2$ is a function of a constant regression parameter $\boldsymbol \beta$, corresponding to the proportional hazards model. Extension to a time-dependent regression parameter is straightforward. The explained variation coefficient remains constant when applying a monotonically increasing transformation on time. Thus, we work on the standardized time scale as described in Section \ref{sec:process}. \\
\noindent The explained variation is a population parameter that needs to be estimated. Several estimators have been proposed in the literature (\citealt{Choodari2012}). Our goal here is not to present an exhaustive review of these estimators. We focus on the $R^2$ coefficient introduced by \citet{OQuigley1994} since it is built with the same residuals as the standardized score process. We recall its definition by extending it to the non--proportional hazards case.
Let us define the expectation over time of the expected squared discrepancy between the covariate or prognostic index evaluated with parameter $\boldsymbol \alpha}\def\g{\boldsymbol \gamma_2(t)$ and their expected value under the non--proportional hazards model (\ref{nonph}) of parameter $\boldsymbol \boldsymbol \alpha}\def\g{\boldsymbol \gamma_1(t)$, not necessarily equals to $\boldsymbol \alpha}\def\g{\boldsymbol \gamma_2(t)$, of dimension $p$
\begin{align*}
Q\big(F&,\boldsymbol \alpha}\def\g{\boldsymbol \gamma_1 (t),\boldsymbol \alpha}\def\g{\boldsymbol \gamma_2(t) \big)\\
&= \left\lbrace
\begin{array}{ll}
\displaystyle{\int_0^1} {E}_{\boldsymbol \alpha}\def\g{\boldsymbol \gamma_1(t)}\left(\left.\left\{\boldsymbol \alpha}\def\g{\boldsymbol \gamma_2(t) ^T{Z}(t)-
{E}_{\boldsymbol \alpha}\def\g{\boldsymbol \gamma_1(t)}\left( \boldsymbol \alpha}\def\g{\boldsymbol \gamma_2(t) ^T{Z} \mid T=t\right)
\right\}^2\right\vert T=t\right)dF(t) & \text{if } p>1 \nonumber\\~\\
\displaystyle{\int_0^1} \ {E}_{\boldsymbol \alpha}\def\g{\boldsymbol \gamma_1(t)}\left(\left.\left\{{Z}(t)-
{E}_{\boldsymbol \alpha}\def\g{\boldsymbol \gamma_1(t)}\left({Z} \mid T=t\right) \right\}^2\right\vert T=t\right)dF(t) & \text{if } p=1,
\end{array}
\right.
\end{align*}
where $F$ is the cumulative distribution function of $T$. Then $\Omega^2(\boldsymbol \beta(t))$ can be expressed as
\begin{equation}
\Omega^2(\boldsymbol \beta(t))=1-\dfrac{Q(F,\boldsymbol \beta(t),\boldsymbol \beta(t))}{Q(F,{\bf 0},\boldsymbol \beta(t))}.
\end{equation}
Let us denote $\hat F$ the estimator of the cumulative distribution function of $T$ such that $\hat F(t)={k_n}^{-1} \bar N^*(t)$. $ \hat F$ corresponds to the usual empirical cumulative distribution function of $T$ in the uncensored case. Then, $Q\left(F,\boldsymbol \alpha}\def\g{\boldsymbol \gamma_1(t),\boldsymbol \alpha}\def\g{\boldsymbol \gamma_2(t)\right)$
can be estimated by
\begin{align*}
\hat{Q}(\hat{F},&\boldsymbol \alpha}\def\g{\boldsymbol \gamma_1(t),\boldsymbol \alpha}\def\g{\boldsymbol \gamma_2(t))\\
&=\left\{
\begin{array}{ll}
{\displaystyle \int_0^1} \left\{\boldsymbol \alpha}\def\g{\boldsymbol \gamma_2(s)^Tr_{\boldsymbol \alpha}\def\g{\boldsymbol \gamma_1(s)}(s)\right\}^2d\hat{F}(s)=
\dfrac{1}{k_n}\,\overset{k_n}{\underset{i=1}{\sum}}\left\{\boldsymbol \alpha}\def\g{\boldsymbol \gamma_2(t_i)^Tr_{\boldsymbol \alpha}\def\g{\boldsymbol \gamma_1(t_i)}(t_i)\right\}^2& \text{ if } p>1 \\~\\
{\displaystyle \int_0^1} \left\{r_{\boldsymbol \alpha}\def\g{\boldsymbol \gamma_1(s)}(s)\right\}^2d\hat{F}(s)=\dfrac{1}{k_n}\,\overset{k_n}{\underset{i=1}{\sum}}\left\{r_{\boldsymbol \alpha}\def\g{\boldsymbol \gamma_1(t_i)}(t_i)\right\}^2& \text{ if } p=1.
\end{array}
\right.
\end{align*}
The $R^2$ coefficient can then be defined by $R^2=R^2(\hat \boldsymbol \beta(t))$, where, for all vector $\boldsymbol \alpha}\def\g{\boldsymbol \gamma(t)$ of dimension $p$,
\begin{equation}
\label{R2def}
R^2(\boldsymbol \alpha}\def\g{\boldsymbol \gamma(t))=1-\dfrac{\hat Q( \hat F,\boldsymbol \alpha}\def\g{\boldsymbol \gamma(t),\boldsymbol \alpha}\def\g{\boldsymbol \gamma(t)) }{\hat Q( \hat F,{\bf 0},\boldsymbol \alpha}\def\g{\boldsymbol \gamma(t)) }
= \left\lbrace
\begin{array}{ll}
1-\dfrac{\overset{k_n}{\underset{i=1}{\sum}}\left\{\boldsymbol \alpha}\def\g{\boldsymbol \gamma(t_i)^Tr_{\boldsymbol \alpha}\def\g{\boldsymbol \gamma(t_i)}(t_i)\right\}^2}{\overset{k_n}{\underset{i=1}{\sum}} \left\{\boldsymbol \alpha}\def\g{\boldsymbol \gamma(t_i)^Tr_{\bf 0}(t_i)\right\}^2}
& \text{ if } p>1 \\
1-\dfrac{ \overset{k_n}{\underset{i=1}{\sum}} r_{\boldsymbol \alpha}\def\g{\boldsymbol \gamma(t_i)}(t_i)
^2 }
{\overset{k_n}{\underset{i=1}{\sum}} r_{\bf 0}(t_i)^2 } & \text{ if } p=1.
\end{array}
\right.
\end{equation}
The explained variation coefficient $\Omega^2(\boldsymbol \beta(t))$ can be estimated by $R^2(\hat{\boldsymbol \beta}(t)),$ where $\hat{\boldsymbol \beta}(t)$ is a consistent estimator of the true value of the regression coefficient $\boldsymbol \beta(t)$. The following theorem will be useful to evaluate the goodness of fit of the model (\ref{nonph}).
\begin{theorem}
\label{theoremR2}
Under the non--proportional hazards model (\ref{nonph}) of parameter $\boldsymbol \beta(t)$,
we have the following convergence
$$\vert R^2(\boldsymbol \beta(t))-R^2(\hat \boldsymbol \beta(t)) \vert\overset{a.s.}{\underset{n\rightarrow + \infty}{\longrightarrow}} 0,$$
and, if $p=1$, with probability one,
$$\underset{{b(t)}}{\arg\max}\ \lim_{n\rightarrow + \infty} R^2(b(t))=\boldsymbol \beta(t).$$
\end{theorem}
The proof is given in appendix \ref{appendix_R2}. The theorem states that if $\boldsymbol \beta(t)$ is the true regression coefficient, the maximum of $R^2$ is well approximated by $R^2(\hat \boldsymbol \beta(t))$ for a large enough sample size. This result has been shown for one covariate in the model, and in the case of multiple covariates, we conjecture an analogous result for the one--dimensional prognostic index.
The predictive ability measure and the standardized score process are built with the same ingredients.
The standardized score process enables to check the fit of the model, whereas the $R^2$ coefficient is a measure of the predictive ability of the
model. Although different, these two aspects of the model are related; their construction with the same quantities seems then to be quite natural.
\subsection{Using the $R^2$ Coefficient to Improve the Fit}
Using the results of Theorem \ref{thNPH_multi} and its corollary, the standardized score process can be used to determine the shape of the temporal regression effect. No other tools such as smoothing, the projection on a basis of functions or kernel estimation are needed
\citep{Cai2003,Hastie1990,Scheike2004}. For instance, as shown in Figure \ref{figure_beta_1_0}, a constant effect until time $\tau$ followed by a null effect is easily detectable, especially with moderate and larger sample sizes. Assume that the time-dependent regression parameter can be expressed as $\boldsymbol \beta(t)=(\beta_1(t),\dots,\beta_p(t))$, where $\beta_j(t)=\beta_{0,j} \, B_j(t)$ ($j=1,\dots,p$) with $\boldsymbol \beta_0=(\beta_{0,1},\dots,\beta_{0,p})\in \mathbb R^p$ an unknown regression parameter and $\mathbf{B}=(B_1,\dots,B_p)$ a known $\mathbb R^p$-valued function of the time. Thus, $\widehat{ \boldsymbol \beta}(t)= (\widehat{ \beta_1}(t),\dots,\widehat{ \beta_p}(t))$ where $\widehat{ \beta_j}(t)=\widehat{ \beta_{0,j}}\, B_j(t)$ ($j=1,\dots,p$) with $\widehat{ \boldsymbol \beta_0}$ the maximum likelihood estimator of $\boldsymbol \beta_0$, obtained via classical maximization of the partial likelihood \citep{Cox1972}. The function $B(t)$ can determined graphically using the standardized score process (see the examples of Section \ref{section_applications}). In addition, the confidence bands defined in Proposition \ref{confband} can help to evaluate the plausibility of a constant effect $\beta_j(t)$ over time resulting in a constant function $B_j$, for each $j=1,\dots,p$.
When dealing with non--proportional hazards, the investigator needs an instrument other than one that is focused solely on fit. This can be provided by the $R^2$ coefficient that not only indicates predictive strength but will tend to a maximum value when the correct form of $B(t)$ is chosen (Theorem \ref{theoremR2}).
When different competing models provide plausible forms for $B(t)$, the one maximizing the $R^2$ coefficient would be considered the best. Using this procedure, we obtain a non--proportional hazards model with a good fit and a maximal predictive ability.
The predictive ability measure is maximized on the set $\cal B$ of the temporal regression effects selected by the investigator. Formally,
let $\mathcal{B}=\{\boldsymbol \beta_1(t),\dots,\boldsymbol \beta_m(t)\}$ be a set of
$m$ functions from $[0,1]$ to $\mathbb R^p$.
The selected regression function $\boldsymbol \beta^*(t)$ is such that
$$\boldsymbol \beta^*(t)=\arg\max_{b(t) \in \mathcal{B}}R^2\left(b(t)\right) .$$
The following theorem gives an equivalence between this maximization problem when $n \rightarrow \infty$ and a problem of minimization of $L^2$ norms.
\begin{theorem}
\label{theoremR2L2}
Let $p=1$. Under the non--proportional hazards model (\ref{nonph}) with regression parameter $\beta(t)$ not necessarily in $\cal B$, asymptotically, $ \beta^*(t)=\underset{\alpha(t) \in \mathcal{B}}{\arg\max}\ \underset{n\rightarrow \infty}{\lim}R^2\left(\alpha(t)\right)$ is the solution of
$$ \beta^*(t)=\arg\min_{\alpha(t) \in \mathcal{B}} \left\| \beta(t)-\alpha(t)\right\|_{2,W},$$
where $ \left\| a(t)\right\|_{2,W}=\left(\int_0^1a(t)^2v(c(t),t)^2dt\right)^{1/2}$ is a weighted $L^2$ norm of the function $a(t)$ from $[0,1]$ to $\mathbb R$, with $c(t)$ lying between $ 0$ and $a(t)$.\newline
\end{theorem}
Proof can be found in Appendix \ref{proofR2L2}. In other words, for large enough sample sizes, selecting the regression coefficient by maximizing the $R^2$ coefficient is the same as selecting the closest temporal regression function to the true coefficient in the $L^2$ norm sense.
A model is chosen to fit a dataset because of either a good fit or a good predictive capacity. Several models could present one of these aspects or both of them, not only the "true" model.
Priority is given to the goodness of fit, with the selection of possible time-dependent coefficients, and in a second phase, the predictive capacity is considered.
We have chosen to work with the $R^2$ coefficient but notice that other predictive ability measures verifying Theorem \ref{theoremR2} might be considered.
When the trend of the process is a concave function, the effect disminishes over time, whereas in presence of a convex function, the effect increases. In order to obtain the largest possible $R^2$, we could create a temporal effect matching more and more closely the observed trend of the process, e.g. piecewise constant effects with multiple changepoints. In general, this would result in an overfit. In this case, the interpretation of the coefficient is not clear. A tradeoff has to be established between a high predictive ability and the simplicity of the coefficient, especially regarding its interpretation. This parallels linear regression where the estimated explained variation is positively biased
and this bias increases with the dimension of the model. Some balance needs to be struck between the goal of improved prediction and the dangers of over optimistic predictions as a result of over fitting.
\section{SOME SIMULATED EXAMPLES}
\label{section_applications}
The simulations are performed with a moderate sample size set to $n=200$ subjects and $\lambda_0(t)=1$. All cases presented here are uncensored. The effect of an independent censoring mechanism on the process is the same as a reduction in the sample size.
\subsection{Univariate cases}
In both considered cases, the covariate follows a Bernoulli distribution of parameter $0.5$.
First, we consider the proportional hazards situation by setting $\beta(t)=1.5$.
\begin{figure}[h]
\label{fig_PH}
\center
\includegraphics[width=2.5in,height=2.5in]{ameliorationfit_PH_15_200indiv_IC.pdf}
\caption{Standardized score process $U^*(0,\cdot)$ (solid line) and confidence bands (dotted lines)
on a simulated dataset with constant regression coefficient $\beta(t)=1.5$.}
\end{figure}
The standardized score process $U^*(0,\cdot)$ (solid line) and its confidence bands under proportional hazards assumption (dotted lines) are plotted over time in Figure \thefigure. A drift is observed, the effect is not null. The drift seems linear and the process stays between the confidence bands:
the hypothesis of a proportional hazards model seems reasonable. The usual maximum partial likelihood estimator is estimated at $1.59$ which gives an $R^2$ of $0.35$.
\FloatBarrier
\begin{figure}[h]
\label{fig_3(1-t)2_a}
\center
\includegraphics[width=2.5in,height=2.5in]{ameliorationfit_3_1-t_2_200indiv_IC.pdf}
\caption{
Standardized score process $U^*(0,\cdot)$ (solid line), confidence bands (dashed lines)
and a fitted changepoint model (dotted lines) on a simulated dataset with $\beta(t)=3(1-t)^2$.}
\end{figure}
The next case deals with a smooth decreasing effect. We simulate a dataset with $\beta(t)=3(1-t)^2$. The resulting standardized score process $U^*(0,\cdot)$ (solid line) is plotted over time in Figure \thefigure $ $ with its confidence bands under proportional hazards assumption (dotted lines). The process leaves the confidence bands which indicates that the proportional hazards assumption does not hold. The concavity of the trend gives an indication regarding the decrease of the effect. Amongst other possibilities, the effect could be linear, of a quadratic shape or a piecewise constant function of the time. In the latter case, the trend appears linear up to time $t=0.5$ corresponding to a constant coefficient. Then, the drift changes to a lower constant value, corresponding to a coefficient $\tilde \beta(t)=\beta_0\{I(t\leq 0.5)+C.I(t>0.5)\},$ where $\beta_0$ and $C$ are unknown. $C$ is the value by which the coefficient is multiplied in the second part of the study. In Figure \thefigure
, using linear regression, two straight dotted lines have been fitted to the process, before and after the changepoint time $t=0.5$. The ratio of the second slope over the first one is the value $C=0.16$. Various models with decreasing effect $\beta(t)=\beta_0 B(t)$ have been selected, their $R^2$ coefficients and $\hat \beta_0$ the maximum partial likelihood estimator of $\beta_0$ have been evaluated in Table \ref{table_3(1-t)2}. The lowest $R^2$ coefficient corresponds to the proportional hazards model and the largest $R^2$ coefficient of Table \ref{table_3(1-t)2} is the one associated with the model of regression coefficient $\beta(t)=\beta_0(1-t)^2$, with an estimation of $\beta_0$ equals to $0.37$. Using our procedure, the regression coefficient used to create the dataset has been selected.
\begin{table}[h]
\center
\begin{tabularx}{15cm}{|X||X|X|X|X|X|}
\hline $\beta(t)$ & $\beta_0$ & $\beta_0(1-t)$ & $\beta_0(1-t)^2$ & $\beta_0(1-t^2)$ & $\tilde \beta(t)$ \rule[-7pt]{0pt}{12pt} \\
\hline
\hline $\hat{\beta_0}$ & 1.06 & 2.45 & 3.73 & 1.77 & 1.83 \rule[-7pt]{0pt}{12pt} \\
\hline $R^2$ & 0.25 & 0.36 & 0.37 & 0.34 & 0.34 \rule[-7pt]{0pt}{12pt}\\
\hline
\end{tabularx}
\caption{
Maximum partial likelihood estimators $\hat \beta(t)$ and $R^2$ coefficients
on a simulated dataset with $\beta(t)=3(1-t)^2$.}
\label{table_3(1-t)2}
\end{table}
\subsection{Multivariate case}
\begin{figure}[h]
\center
\subfigure[Regression effect $\beta_1(t)$]{
\includegraphics[width=2.5in,height=2.5in]{ameliorationfit_NPH_multiv_200indiv_1_IC.pdf}
\label{fig_multiv_a}
}
\subfigure[Regression effect $\beta_2(t)$]{
\includegraphics[width=2.5in,height=2.5in]{ameliorationfit_NPH_multiv_200indiv_2_IC.pdf}
\label{fig_multiv_b}}
\caption{Standardized score process $\hat \boldsymbol \Sigma^{-1} \mathbf U^ *(0,\cdot)$ (solid line), confidence bands (dashed lines)
and a fitted changepoint model (dotted lines) on a simulated dataset with $\beta_1(t)=I(t\leq 0.5)$ and $\beta_2(t)=-1$.
}
\end{figure}
We simulate two standard normal covariates $Z^1$ and $Z^2$ with covariance equals to $0.5$. We set $\beta_1(t)=I(t\leq 0.5)$ and $\beta_2(t)=1$. Each component of the bivariate process $\hat \boldsymbol \Sigma^{-1/2}\mathbf U^*(\mathbf{0},\cdot)$ (solid lines) is plotted over time on Figures \ref{fig_multiv_a} and \ref{fig_multiv_b} with the confidence bands (dotted lines).
Clearly, the proportional hazards assumption is rejected for covariate $Z^1$ since the process leaves the confidence band. The shape of the process indicates a piecewise constant regression coefficient, with a changepoint at time $t=0.6$. As in the univariate case, two straight (dashed) lines have been fitted to the process, one before $t=0.6$ and one after. The ratio of the slopes is $-0.12$ which makes us consider the regression coefficient $\beta_1(t)=\beta_1\,B_{0.6}(t)$ where $B_{0.6}(t)=I(t\leq 0.6)-0.12\,I(t\geq 0.6)$. Other piecewise constant regression coefficients $\beta(t)=\beta_1\,B_{t_0}(t)$ have been considered with changepoints at times $t_0\in\{0.45,0.5,\dots,0.7\}$. For each time $t_0$, the ratio of slopes has been evaluated to determine the value which multiplies the coefficient in the second part on the study. The second covariate $Z^2$, however, seems to have a constant regression coefficient since the process stands between the confidence bands and
has a linear trend (Figure \ref{fig_multiv_b}). Therefore, we consider only the regression coefficient $\beta_2(t)=\beta_2$.
\begin{table}[h]
\label{table_multiv}
\center
\begin{tabular}{|c||c|c|c|c|c|c|c|c|}
\hline $\beta_1(t)$ & $ \beta_1$ &$ \beta_1\,B_{0.45}(t)$ &$ \beta_1\,B_{0.5}(t)$ &$ \beta_1\,B_{0.55}(t)$ & $ \beta_1\,B_{0.6}(t)$ & $ \beta_1\,B_{0.65}(t)$ &$ \beta_1\,B_{0.7}(t)$ \rule[-7pt]{0pt}{12pt} \\
\hline $\hat{\beta}_1$ & 0.45 & 0.93 & 0.96 & 0.89 & 0.95 & 0.86 & 0.72\rule[-7pt]{0pt}{12pt}\\
\hline $\hat \beta_2$ & -0.73& -0.72& -0.73& -0.74& -0.79& -0.80& -0.77\rule[-7pt]{0pt}{12pt} \\
\hline $R^2$ & 0.24 & 0.35 & 0.37 & 0.35 & 0.39 & 0.37 & 0.32 \rule[-7pt]{0pt}{12pt}\\
\hline
\end{tabular}
\caption{Maximum partial likelihood estimators $\hat \beta(t)$ and $R^2$ coefficients
on a simulated dataset with $\beta_1(t)=I(t\leq 0.5)$ and $\beta_2(t)=-1$.}
\end{table}
Estimation results are given in Table \thetable
. The proportional hazards model gives an $R^2$ of 0.24. The maximal $R^2$ is obtained when considering $\beta_1(t)=\beta_1\,B_{0.6}(t)$, with an increase of $60\%$ compared to the proportional hazards model. Therefore, we choose the model with $\beta_1(t)=\beta_1\,B_{0.6}(t)$ and $\beta_2(t)=\beta_2$.
\section{CLINICAL STUDY IN BREAST CANCER}
\begin{figure}[h]
\label{fig_breastcanc}
\center
\subfigure[Tumor size]{ \label{fig_breastcanctaille}
\includegraphics[width=2.5in,height=2.5in]{breastcancer_Taille60_IC.pdf}}
\subfigure[Progesterone receptor]{ \label{fig_breastcancrec}
\includegraphics[width=2.5in,height=2.5in]{breastcancer_Receptor_IC.pdf}}
\subfigure[Grading]{ \label{fig_breastcangrade}
\includegraphics[width=2.5in,height=2.5in]{breastcancer_Grade2_IC.pdf}}
\caption{Standardized score process $\hat \boldsymbol \Sigma^{-1} \mathbf U^ *(0,\cdot)$ (solid lines) and its confidence bands (dashed lines) on the breastcancer dataset for tumor size, progesterone receptor and grading.}
\end{figure}
We return to the motivating example of the $1504$ patients suffering from breast cancer. These patients were followed over a period of 15 years at the Institut Curie in Paris, France. Several studies were based on these data. One sub-study considered the predictive effects of the prognostic factors; progesterone receptor status, the tumor size over $60$ mm and the grading over 2. The multivariate standardized score process $\hat \boldsymbol \Sigma^{-1/2}\mathbf U^*(\mathbf{0},\cdot)$ and its confidence band are plotted over time in Figure \thefigure
. In Figure \ref{fig_breastcanctaille}, we illustrate the process corresponding to the tumor size effect. Clearly, the effect seems non--constant with slope gradually diminishing with time. So much so that the process ends up drifting beyond the limits of the 95$\%$ confidence band. A slightly more refined model providing a much better fit allows for a change in effect at time point $t=0.2$. As in our simulated examples, two straight lines have been fitted to the curve before and after $t=0.2$, leading us to consider the regression effect $\beta_{size}(t)=\beta_0(I(t\leq 0.2) +0.24 I(t\geq 0.2))$. From Table 3
we can quantify the predictive improvement of Model 2 (constant effects for hormone receptor status and grade, time dependent effects for tumor size) versus Model 1 (all 3 prognostic factors constant) by a greater than $30\%$ increase in the size of $R^2$, from $0.29$ to $0.39$.
Figure \ref{fig_breastcancrec} represents the process for the effect of the progesterone receptor over time. Again there is some evidence of a changing slope, although much weaker than for tumor size and, indeed, the process remains within the limits of the confidence bands. We considered various potential regression effects: a changepoint model with a cut at time $t=0.5,$ $\beta_{rec0}(t)=\beta_0(I(t\leq 0.5) +0.39 I(t\geq 0.5))$ and several smooth parameters $\beta_{rec1}(t)=\beta_0(1-t)$, $\beta_{rec2}(t)=\beta_0(1-t)^2$, $\beta_{rec3}(t)=\beta_0(1-t^2)$ and $\beta_{rec4}(t)=\beta_0\log(t)$.
Figure \ref{fig_breastcangrade} represents the process for the grading effect. There is a clear impression of the steepness of the negative slope attenuating with time. The process reaches the limits of the confidence bands but does not go beyond them. The simpler model, i.e., proportional hazards effects implying a linear slope, may be good enough although, in a model building context, it is also worth considering one with time dependent effects. Specifically, we chose to also look at a model with piecewise constant coefficients $\beta_{gra}(t)=\beta_0(t)(I(t\leq 0.4) +0.69 I(t\geq 0.4))$.
All of these several combinations, alongside models with constant effects, were looked at. For each combination, the regression effects have been estimated by maximizing the partial likelihood and the $R^2$ coefficient has been evaluated.
\begin{table}[h]
\label{table_breastcanc}
\center
\begin{tabular}{|c|c|c|c|}
\hline
Tumor size & Receptor & Grading & $R^2$ \\
\hline 0.84 & 1.03 & -0.68 & 0.29 \\
\hline $1.77(I(t\leq 0.2) +0.24 I(t\geq 0.2))$ & 1.03 & -0.66 & 0.39 \\
\hline 0.85 & $-1.02\log(t)$ & -0.67 & 0.39 \\
\hline $1.74 (I(t\leq 0.2) +0.24 I(t\geq 0.2))$ & $-1.02\log(t)$ & -0.66 &0.51 \\
\hline $1.72 (I(t\leq 0.2) +0.24 I(t\geq 0.2))$ & $-1.02\log(t)$ & $-0.82I(t\leq 0.4) +0.69 I(t\geq 0.4)$ & 0.52 \\
\hline
\end{tabular}
\caption{Maximum partial likelihood estimators and $R^2$ coefficients
on the breast cancer dataset.}
\end{table}
Partial results are given in Table \thetable
. The proportional hazards model gives an $R^2$ coefficient of $0.29$. As mentioned above, a more involved model allowing for the effect of tumor size to assume a simple time dependency results in a big jump in observed predictability of an order greater than 30\%.The highest $R^2$ is obtained with changepoints for tumor size and grading covariates, with a function of $\log(t)$ for the effect of progesterone receptor. The predictive accuracy of this model has increased by $80\%$ compared to the predictive accuracy of the corresponding proportional hazards model. This gives a strong indication that, as far as prediction is concerned, significant improvement can be consequent on allowing time dependency. On the other hand, allowing for time dependency grade, having already accounted for the joint effects of tumor size and receptor status, results in an increase in $R^2$ from 0.51 to 0.52. Such a negligible increase dose not justify the added complexity of the model so that, provided the other two risk factors are included, it makes sense to restrict the effects of grade to be constant.
\FloatBarrier
\section{DISCUSSION}
The related and complementary techniques of goodness of fit and predictive ability provide a coherent way to construct models. Intuitively, models constructed in this way ought provide a better predictive performance. This intuition is correct and is supported by the theoretical results of this paper. Our preference is to appeal to techniques based on the Schoenfeld residual processes for proportional and non--proportional hazards models since these processes provide the basis for both of these techniques. A large number of competing approaches appears possible since there is a large body of literature on goodness of fit procedures and a large body on predictive measures. Combinations of these could provide tools analogous to those described here. However, in order to make analogous claims to ours concerning predictive performance for some particular combination, we would require equivalent theorems to those presented in Sections 2 and 3.
We might consider that the first step away from a proportional hazards model is a similar model but with a changepoint. Before the changepoint we have one particular proportional hazards model whereas, after the changepoint, we have a model with a different value of $\beta .$ The methods described here would enable us to estimate the changepoint itself as well as the values of $\beta,$ before and after the changepoint. Extending this to more than a single changepoint is, at least in theory, straightforward. This suggests one possible systematic way of model construction. Another extension that would be worth considering is the estimation of the process drift with non--parametric estimation techniques in order to estimate the cumulative regression effect $\int_0^t\boldsymbol \beta(s)ds$.
\FloatBarrier
|
\chapter{Min-Sum-based decoders running on noisy hardware}\label{chap:ms_noisy_hw}
\section{Introduction}
In traditional models of communication or storage systems with error correction coding, it is assumed that the operations of an error correction encoder and decoder are deterministic and that the randomness exists only in the transmission or storage channel. However, with the advent of nanoelectronics, the reliability of the forthcoming circuits and computation devices is becoming questionable.
It is then becoming crucial to design and analyze error correcting decoders able to provide reliable error correction even if they are made of unreliable components.
Except the pioneered works by Taylor and Kuznetsov on reliable memories \cite{taylor1968reliablestorage, taylor1968reliablecomputing, kuznetsov1973information}, later generalized in \cite{chilappagari2006analysis, vasic2007information} to the case of hard-decision decoders, this new paradigm of noisy decoders has merely not been addressed until recently in the coding literature. However, over the last years, the study of error correcting decoders, especially Low-Density Parity-Check (LDPC) decoders, running on noisy hardware attracted more and more interest in the coding community. In \cite{winstead2009probabilistic} and \cite{tang2012ldpc} hardware redundancy is used to develop fault-compensation techniques, able to protect the decoder against the errors induced by the noisy components of the circuit. In \cite{hussien2011class}, a class of modified Turbo and LDPC decoders has been proposed, able to deal with the noise induced by the failures of a low-power buffering memory that stores the input soft bits of the decoder. Very recently, the characterization of the effect of noisy processing on message-passing iterative LDPC decoders has been proposed. In \cite{varshney2011performance}, the concentration and convergence properties were proved for the asymptotic performance of noisy message-passing decoders, and density evolution equations were derived for the noisy Gallager-A and Belief-Propagation decoders. In \cite{yazdi2012probabilistic, yazdi2012optimal, yazdi2013gallager}, the authors investigated the asymptotic behavior of the noisy Gallager-B decoder defined over binary and non-binary alphabets. The Min-Sum decoding under unreliable message storage has been investigated in \cite{balatsoukas2014characterization, balatsoukas2014density}. However, all these papers deal with very simple error models, which emulate the noisy implementation of the decoder, by passing each of the exchanged messages through a noisy channel.
In this work we focus on the Min-Sum decoder, which is widely implemented in real communication systems. In order to emulate the noisy implementation of the decoder, probabilistic error models are proposed for its arithmetic components (adders and comparators). The proposed probabilistic components are used to build the noisy finite-precision decoders. We further analyze the asymptotic performance of the noisy Min-Sum decoder, and provide useful regions and target-BER-thresholds \cite{varshney2011performance} for a wide range of parameters of the proposed error models.
We also highlight a wide variety of more or less conventional behaviors and reveal the existence of a specific threshold phenomenon, which is referred to as {\em functional threshold}. Finally, the asymptotic results are also corroborated through finite length simulations.
The remainder of the paper is organized as follows. Section~\ref{sec:ldpc_codes_and_ms} gives a brief introduction to LDPC codes and iterative decoding. Section~\ref{sec:error_injection} presents the error models for the arithmetic components. The density evolution equations and asymptotic analysis of the noisy finite-precision Min-Sum decoding are presented in Section~\ref{sec:density_evolution} and Section~\ref{sec:asymptotic-analysis} respectively. Section~\ref{sec:finite_length_perf} provides the finite-length performance and Section~\ref{sec:noisy_ms_conslusion} concludes the paper.
\section{LDPC Codes and the Min-Sum Algorithm}\label{sec:ldpc_codes_and_ms}
\subsection{LDPC Codes}
LDPC codes \cite{gallager1963low} are linear block codes defined by sparse parity-check matrices. They can be advantageously represented by bipartite (Tanner) graphs \cite{tanner1981recursive} and decoded by message-passing (MP) iterative algorithms. The Tanner graph of an LDPC code is a bipartite graph ${\cal H}$, whose adjacency matrix is the parity-check matrix $H$ of the code. Accordingly, ${\cal H}$ contains two types of nodes:
\begin{itemize}
\item {\em variable-nodes}, corresponding to the columns of $H$, or equivalently to the codeword bits, and
\item {\em check-nodes}, corresponding to the rows of $H$, or equivalently to the parity equations the codeword bits are checked by.
\end{itemize}
We consider an LDPC code defined by a Tanner graph ${\cal H}$, with $N$ variable-nodes and $M$ check-nodes. Variable-nodes are denoted by $n \in \{1, 2, ... , N\}$, and check-nodes by $m \in \{1, 2, ... , M\}$. We denote by ${\cal H}(n)$ and ${\cal H}(m)$ the {\em set of neighbor nodes} of the variable-node $n$ and of the check-node $m$, respectively. The number of elements of ${\cal H}(n)$ $\left( \mbox{or } {\cal H}(m)\right)$ is referred to as the {\em node-degree}.
The Tanner graph representation allows reformulating the {\em probabilistic decoding} initially proposed by Gallager \cite{gallager1963low} in terms of Belief-Propagation\footnote{Also referred to as Sum-Product (SP)} (BP): an MP algorithm proposed by J. Pearl in 1982 \cite{pearl1982bp} to perform Bayesian inference on trees, but also successfully used on general graphical models \cite{pearl1988probabilistic}. The BP decoding is known to be optimal on cycle-free graphs (in the sense that it outputs the maximum a posteriori estimates of the coded bits), but can also be successfully applied to decode linear codes defined by graphs with cycles, which is actually the case of all practical codes. However, in practical applications, the BP algorithm might be disadvantaged by its computational complexity and
its sensitivity to the channel noise density estimation (inaccurate estimation of the channel noisy density may cause significant degradation of the BP performance).
\subsection{Min-Sum Decoding}\label{subsec:ms-decoding}
One way to deal with complexity and numerical instability issues is to simplify the computation of messages exchanged within the BP decoding. The most complex step of the BP decoding is the computation of check-to-variable messages, which makes use of computationally intensive hyperbolic tangent functions. The Min-Sum (MS) algorithm is aimed at reducing the computational complexity of the BP, by using max-log approximations of the parity-check to coded-bit messages \cite{fossorier1999reduced, chung2000construction, eleftheriou2001reduced}. The only operations required by the MS decoding are additions, comparisons, and sign ($\pm 1$) products, which solves the complexity and numerical instability problems. The performance of the MS decoding is also known to be independent of the channel noise density estimation, for most of the usual channel models.
For the sake of simplicity, we only consider transmissions over {\em binary-input} memoryless noisy channels, and assume that the channel input alphabet is $\{-1, +1\}$, with the usual convention that $+1$ corresponds to the $0$-bit, and $-1$ corresponds to the $1$-bit.
We further consider a codeword $\vect{x} = (x_1,\dots,x_N) \in\{-1, +1\}^N$ and denote by $\vect{y} = (y_1,\dots,y_N)$ the received word. The following notation will be used throughout the paper, with respect to message passing decoders:
\begin{itemize}
\item $\gamma_n$ is the log-likelihood ratio (LLR) value of $x_n$ according to the received $y_n$ value; it is also referred to as the {\em a priori information} of the decoder concerning the variable-node $n$;
\item $\tilde{\gamma}_n$ is the {\em a posteriori information} (LLR value) of the decoder concerning the variable-node $n$;
\item $\alpha_{m,n}$ is the variable-to-check message sent from variable-node $n$ to check-node $m$;
\item $\beta_{m,n}$ is the check-to-variable message sent from check-node $m$ to variable-node $n$.
\end{itemize}
The (infinite precision) MS decoding is described in Algorithm~\ref{alg:ms}. It consists of an initialization step (in which variable-to-check messages are initialized according to the a priori information of the decoder), followed by an iteration loop, where each iteration comprises three main steps as follows:
\begin{itemize}
\item {\bf CN-processing} (check-node processing step): computes the check-to-variable messages $\beta_{m,n}$;
\item {\bf VN-processing} (variable-node processing step): computes the variable-to-check messages $\alpha_{m,n}$;
\item {\bf AP-update} (a posteriori information update step): computes the a posteriori information $\tilde{\gamma}_n$.
\end{itemize}
Moreover, each iteration also comprises a {\em hard decision} step, in which each transmitted bit is estimated according to the sign of the a posteriori information, and a {\em syndrome check} step, in which the syndrome of the estimated word is computed. The MS decoding stops when whether the syndrome is $+1$ (the estimated word is a codeword) or a maximum number of iterations is reached.
\input{algo_bin_ms}
The a priori information (LLR) of the decoder is defined by
$\displaystyle\gamma_n = \log\frac{\Pr(x_n=+1\mid y_n)}{\Pr(x_n=-1\mid y_n)}$, and
for the two following channel models (predominantly used in this work), it can be computed as follows:
\begin{itemize}
\item For the Binary Symmetric Channel (BSC), $\vect{y}\in \{-1,+1\}^N$ is obtained by flipping each entry of $\vect{x}$ with some probability $\varepsilon$, referred to as the channel's crossover probability. Consequently:
\begin{equation}
\gamma_n = \log\left(\frac{1-\varepsilon}{\varepsilon}\right)y_n
\end{equation}
\item For the Binary-Input Additive White Gaussian Noise (BI-AWGN) channel, $\vect{y}\in \mathbb{R}^N$ is obtained by $y_n = x_n + z_n$, where $z_n$ is the white Gaussian noise with variance $\sigma^2$. It follows that:
\begin{equation}
\gamma_n = \frac{2}{\sigma^2}y_n
\end{equation}
\end{itemize}
\noindent{\bf Remark:} It can be easily seen that if the a priori information vector $\vect{\gamma}=(\gamma_1,\dots,\gamma_N)$ is multiplied by a constant value, this value will factor out from all the processing steps in Algorithm~\ref{alg:ms} (throughout the decoding iterations), and therefore it will not affect in any way the decoding process. It follows that for both the BSC and BI-AWGN channel models, one can simply define the a priori information of the decoder by $\gamma_n = y_n$, $\forall n = 1,\dots, N$.
\section{Error Injection and Probabilistic Models for Noisy Computing}\label{sec:error_injection}
\subsection{Noisy Message-Passing decoders}
The model for noisy MP decoders proposed in \cite{varshney2011performance} incorporates two different sources of noise: {\em computation noise} due to noisy logic in the processing units, and {\em message-passing noise} due to noisy wires (or noisy memories) used to exchange messages between neighbor nodes.
\begin{itemize}
\item The computation noise is modeled as a random variable, which the variable-node or the check-node processing depends on.
Put differently, an outgoing message from a (variable or check) node depends not only on the incoming messages to that node (including the a priori information for the variable-node processing), but also on the realization of a random variable which is assumed to be independent of the incoming messages.
\item The message-passing noise is simply modeled as a noisy channel. Hence, transmitting a message over a noisy wire is emulated by passing that message through the corresponding noisy channel.
\end{itemize}
However, in \cite{varshney2011performance} it has been noted that {\em ``there is no essential loss of generality by combining computation noise and message-passing noise into a single form of noise''} (see also \cite[Lemma 3.1]{dobrushin1977lower}). Consequently, the approach adopted has been to merge noisy computation into message-passing noise, and to emulate noisy decoders by passing the exchanged messages through different noisy channel models. Thus, the noisy Gallager-A decoder has been emulated by passing the exchanged messages over independent and identical BSC wires, while the noisy BP decoder has been emulated by corrupting the exchanged messages with bounded and symmetrically distributed additive noise ({\em e.g.} uniform noise or truncated Gaussian noise).
\medskip The approach we follow in this work differs from the one in \cite{varshney2011performance} in that the computation noise is modeled at the lower level of {\em arithmetic and logic operations} that compose the variable-node and check-node processing units. This finer-grained noise modeling is aimed at determining the level of noise that can be tolerated in each type of operation. As the main focus of this work is on computation noise, we shall consider that messages are exchanged between neighbor nodes through error-free wires (or memories). However, we note that {\em this work can readily be extended to include different error models for the message-passing noise} (as defined in \cite{varshney2011performance}). Alternatively, we may assume that the message-passing noise is merged into the computation noise, in the sense that adding noise in wires would modify the probabilistic model of the noisy logic or arithmetic operations.
\subsection{Error Injection Models}
We only consider the case of finite-precision operations, meaning that the inputs (operands) and the output of the operator are assumed to be bounded integer numbers. We simulate a noisy operator by injecting errors into the output of the noiseless one. In the following, ${\cal V} \subset {\mathbb Z}$ denotes a finite set consisting of all the possible outputs of the noiseless operator.
\begin{defi}
An {\em error injection model} on ${\cal V}$, denoted by $({\cal E}, p_{{\cal E}}, \imath \mid {\cal V})$, is given by:
\begin{itemize}
\item A finite {\em error set} ${\cal E}\subset {\mathbb Z}$ together with a probability mass function $p_{{\cal E}}:{\cal E} \rightarrow [0, 1]$, referred to as the {\em error distribution};
\item A function $\imath : {\cal V} \times {\cal E} \rightarrow {\cal V}$, referred to as the {\em error injection function}.
\end{itemize}
\end{defi}
For a given set of inputs, the output of the noisy operator is the random variable defined by $\imath(v, e)$, where $v\in{\cal V}$ is the corresponding output of the noiseless operator, and $e$ is drawn randomly from ${\cal E}$ according to the probability distribution $p_{{\cal E}}$.
The error injection probability is defined by
\begin{equation}\label{eq:err-injection-proba}
p_0 = \frac{1}{|{\cal V}|} \sum_{v}\sum_{e} \bar{\delta}_{\imath(v,e)}^v p_{\cal E}(e),
\end{equation}
where $\bar{\delta}_{\imath(v,e)}^v = 0$ if $v = \imath(v,e)$, and $\bar{\delta}_{\imath(v,e)}^v = 1 $ if $v \neq \imath(v,e)$. In other word, $p_0 = \Pr(v \neq \imath(v, e))$, assuming that $v$ is drawn uniformly from ${\cal V}$ and $e$ is drawn from ${\cal E}$ according to $p_{{\cal E}}$.
The above definition makes some implicit assumptions which are discussed below.
\begin{itemize}
\item The set of possible outputs of the noisy operator is the same as the set of possible outputs of the noiseless operator (${\cal V}$). This is justified by the fact that, in most common cases, ${\cal V}$ is the set of all (signed or unsigned) integers that can be represented by a given number of bits. Thus, error injection will usually alter the bit values, but not the number of bits.
\item The injected error does not depend on the output value of the noiseless operator and, consequently, neither on the given set of inputs. In other words, the injected error is independent on the data processed by the noiseless operator. The validity of this assumption does actually depend on the size of the circuit implementing the operator. Indeed, this assumption tends to hold fairly well for large circuits \cite{i-RISC-D2.1}, but becomes more tenuous as the circuit size decreases.
\end{itemize}
Obviously, it would be possible to define more general error injection models, in which the injected error would depend on the data (currently and/or previously) processed by the operator. Such an error injection model would certainly be more realistic, but it would also make it very difficult to analytically characterize the behavior of noisy MP decoders. As a side effect, the decoding error probability would be dependent on the transmitted codeword, which would prevent the use of the {\em density evolution} technique for the analysis of the asymptotic decoding performance (since the density evolution technique relies on the all-zero codeword assumption).
\medskip However, the fact that the error injection model is data independent does not guarantee that the decoding error probability is independent of the transmitted codeword. In order for this to happen, the error injection model must also satisfy a {\em symmetry condition} that can be stated as follows.
\begin{defi}
An error injection model $({\cal E}, p_{{\cal E}}, \imath \mid {\cal V})$ is said to be {\em symmetric} if ${\cal V}$ is symmetric around the origin (meaning that $v\in{\cal V} \Leftrightarrow -v\in{\cal V}$, but $0$ does not necessarily belong to ${\cal V}$), and
the following equality holds
\begin{equation}
\sum_{\{e \mid\, \imath(v,e) = w\}} p_{\cal E}(e) \ \ = \sum_{\{e \mid\, \imath(-v,e) = -w\}} p_{\cal E}(e), \ \ \forall v, w\in {\cal V}
\end{equation}
\end{defi}
The meaning of the symmetry condition is as follows. Let $V$ be a random variable on ${\cal V}$. Let $\phi^{(\imath)}_{V}$ and $\phi^{(\imath)}_{-V}$ denote the probability mass functions of the random variables obtained by injecting errors in the output of $V$ and $-V$, respectively. Then the above symmetry condition is satisfied if and only if for any $V$ the following equality holds
\begin{equation}
\phi^{(\imath)}_{V}(w) = \phi^{(\imath)}_{-V}(-w), \ \forall w\in{\cal V}
\end{equation}
A particular case in which the symmetry condition is fulfilled is when $\imath(-v, e) = -\imath(v, e)$, for all $v\in {\cal V}$ and $e\in {\cal E}$. In this case, the error injection model is said to be {\em highly symmetric}.
\medskip Messages exchanged within message-passing decoders are generally in {\em belief-format}, meaning that the sign of the message indicates the bit estimate and the magnitude of the message the confidence level. As a consequence, errors occurring on the sign of the exchanged messages are expected to be more harmful than those occurring on their magnitude.
This motivates the following definition, which will be used in the following section (see also the discussion in Section~\ref{subsec:unit-sign-prez}).
\begin{defi}
An error injection model $({\cal E}, p_{{\cal E}}, \imath \mid {\cal V})$ is said to be {\em sign-preserving} if for any $v\in{\cal V}$ and $e\in{\cal E}$, $v$ and $\imath(v, e)$ are either both non-negative ($\geq 0$) or both non-positive ($\leq 0$).
\end{defi}
\subsection{Bitwise-XOR Error Injection}\label{subsec:bitwise-xor-injection}
We focus now on the two main symmetric error injection models that will be used in this work. Both models are based on a bitwise {\sc xor} operation between the noiseless output $v$ and the error $e$. The two models differ in the definition of the error set ${\cal E}$, which is chosen such that the bitwise {\sc xor} operation may or may not affect the sign of the noiseless output. In the first case, the bitwise {\sc xor} error injection model is said to be {\em full-depth}, while in the second it is said to be {\em sign-preserving}. These error injection models are rigorously defined below.
\medskip In the following, we fix $\theta \geq 2$ and set ${\cal V} = \{-\Theta, \dots, -1, 0, +1, \dots, +\Theta\}$, where $\Theta = 2^{\theta-1}-1 \geq 1$.\break We also fix a {\em signed number binary representation}, which can be any of the {\em sign-magnitude}, {\em one's complement}, or {\em two's complement} representation. There are exactly $2^\theta$ signed numbers that can be represented by $\theta$ bits in any of the above formats, one of which does not belong to ${\cal V}$ (note that ${\cal V}$ contains only $2\Theta + 1 = 2^\theta-1$ elements for symmetry reasons!). We denote this element by $\zeta$. Hence:
\begin{itemize}
\item In sign-magnitude format, $\zeta = -0$, with binary representation $10\cdots0$;
\item In one's complement format, $\zeta = -0$, with binary representation $11\cdots1$;
\item In two's complement format, $\zeta = -(\Theta+1)$, with binary representation $10\cdots0$.
\end{itemize}
For any $u,v\in {\cal V}$, we denote by $u\wedge v$ the bitwise {\sc xor} operation between $u$ and $v$. From the above discussion, it follows that $u\wedge v \in {\cal V}\cup \{\zeta\}$.
\subsubsection{Full-depth error injection} For this error model, the error set is ${\cal E} = {\cal V}$.
The error injection probability is denoted by $p_0$, and all the possible error values $e\neq 0$ are assumed to occur with the same probability (for symmetry reasons). It follows that the error distribution function is given by $p_{\cal E}(0) = 1-p_0$ and $p_{\cal E}(e) = \frac{p_0}{2\Theta}$, $\forall e\neq 0$. Finally, the error injection function is defined by:
\begin{equation}
\imath(v, e) = \left\{\begin{array}{cl}
v\wedge e, & \mbox{if } v\wedge e \in {\cal V}\\
e, & \mbox{if } v\wedge e = \zeta\end{array}\right.
\end{equation}
\subsubsection{Sign-preserving error injection} For this error model, the error set is ${\cal E} = \{0, +1,\dots, +\Theta\}$.
The error injection probability is denoted by $p_0$, and all the possible error values $e\neq 0$ are assumed to occur with the same probability (for symmetry reasons). It follows that the error distribution function is given by $p_{\cal E}(0) = 1-p_0$ and $p_{\cal E}(e) = \frac{p_0}{\Theta}$, $\forall e\neq 0$. Finally, the error injection function is defined by:
\begin{equation}
\imath(v, e) = \left\{\begin{array}{cl}
v\wedge e, & \mbox{if } v \neq 0 \mbox{ and } v\wedge e \in {\cal V}\\
\pm e, & \mbox{if } v = 0 \\
0, & \mbox{if } v\wedge e = \zeta\end{array}\right.
\end{equation}
In the above definition, $\imath(0, e)$ is randomly set to either $-e$ or $+e$, with equal probability (this is due once again to symmetry reasons). Note also that the last two conditions, namely $v = 0$ and $v\wedge e = \zeta$, cannot hold simultaneously (since $e\neq \zeta$).
Finally, we note that both of the above models are highly symmetric, if one of the sign-magnitude or the one's complement representation is used. In case that the two's complement representation is used, they are both symmetric, but not highly symmetric.
An example of sign-preserving bitwise-{\sc xor} error injection is given in Table~\ref{tab:sign-preserve-bitwise-xor}. The number of bits is $\theta = 5$ and two's complement binary representation is used. The sign bit of the error is not displayed, as it is equal to zero for any $e\in {\cal E}$. The positions of $1$'s in the binary representation of $e$ correspond to the positions of the erroneous bits in the noisy output.
\begin{table}[t!]
\begin{center}
\caption{Example of sign-preserving bitwise-xor error injection}
\label{tab:sign-preserve-bitwise-xor}
\begin{tabular}{|@{\,}r@{\,}|c@{\,}||*{5}{@{\,}m{12.5mm}@{\,}|}}
\hline
& integer & \multicolumn{5}{c|}{$2$'s complement binary representation}\\
\hline
noiseless output: $v$ & $-11$ &\ccol{$1$}&\ccol{$0$}&\ccol{$1$}&\ccol{$0$}&\ccol{$1$}\\
\hline
error: $e$ & $6$ & &\ccol{$0$}&\ccol{$1$}&\ccol{$1$}&\ccol{$0$}\\
\hline
noisy output: $\imath(v, e)$ & $-13$ & \ccol{$1$} & \ccol{$0$} & \ccol{$\textcolor{red}0$} & \ccol{$\textcolor{red}1$} & \ccol{$1$} \\
\hline
\multicolumn{2}{|@{\,}r@{\,}||@{\,}}{bit position} & \ccol{$\theta\!=\!5$} & \ccol{$4$} & \ccol{$3$} & \ccol{$2$} & \ccol{$1$} \\
\hline
\end{tabular}
\end{center}
\end{table}
\medskip\noindent{\bf Remark:} It is also possible to define a {\em variable depth error injection} model, in which errors are injected in only the $\lambda$ least significant bits, with $\lambda\leq \theta$. Hence, $\lambda = \theta$ corresponds to the above full-depth model, while $\lambda = \theta-1$ corresponds to the sign-preserving model. However, for $\lambda < \theta-1$ such a model is {\bf not} symmetric, if the the two's complement representation is used.
\subsection{Output-Switching Error Injection}\label{subsec:output-swithcing}
A particular case is represented by error injection on binary output. Assuming that ${\cal V} = \{0, 1\}$,
the {\em bit-flipping} error injection model is defined as follows. The error set is ${\cal E} = \{0, 1\}$, with error distribution function given by $p_{\cal E}(0) = 1-p_0$ and $p_{\cal E}(1) = p_0$, where $p_0$ is the error injection probability, and the error injection function is given by $\imath(v, e) = v\wedge e$. Put differently, the error injection model flips the value of a bit in ${\cal V}$ with probability $p_0$.
Clearly, the above error injection model can be applied on any set ${\cal V}$ with two elements, by switching one value to another with probability $p_0$. In this case, we shall refer to this error injection model as {\em output-switching}, rather than bit-flipping.
Moreover, if one takes ${\cal V} = \{-1, +1\}$ (with the usual $0{,}1$ to $\pm 1$ conversion), it can be easily verified that this error injection model is highly symmetric.
\subsection{Probabilistic models for noisy adders, comparators and XOR-gates}\label{subsec:proba-models-operators}
In this section we describe the probabilistic models for noisy adders, comparators and xor-gates, built upon the above error injection models. These probabilistic models will be used in the next section, in order to emulate the noisy implementation of the quantized (finite-precision) MS decoder.
\subsubsection{Noisy adder model} We consider a $\theta$-bit adder, with $\theta \geq 2$. The inputs and the output of the adder are assumed to be in ${\cal V} = \{-\Theta, \dots, -1, 0, +1, \dots, +\Theta\}$, where $\Theta = 2^{\theta-1}-1$.
We denote by $\mathbf{s}_{\cal V} : \mathbb{Z}\rightarrow {\cal V}$, the $\theta$-bit saturation map, defined by:
\begin{equation}\label{eq:saturation-map}
\mathbf{s}_{\cal V}(v) = \left\{\begin{array}{ll}
-\Theta, & \mbox{if } v < -\Theta \\
\ \ v, & \mbox{if } v \in {\cal V}\\
+\Theta, & \mbox{if } v > +\Theta \end{array}\right.
\end{equation}
For inputs $(x, y)\in{\cal V}$, the output of the noiseless adder is defined as $s_{\cal V}(x + y)$. Hence, for a given error injection model $({\cal E}, p_{{\cal E}}, \imath \mid {\cal V})$, the output of the noisy adder is given by:
\begin{equation}\label{eq:noisy-add-def}
\mathbf{a}_{\mbox{\scriptsize pr}}(x, y) = \imath\left(\mathbf{s}_{\cal V} (x+y), e\right),
\end{equation}
where $e$ is drawn randomly from ${\cal E}$ according to the probability distribution $p_{{\cal E}}$. The {\em error probability of the noisy adder}, assuming uniformly distributed inputs, is equal to the error injection probability (parameter $p_0$ defined in (\ref{eq:err-injection-proba})), and will be denoted in the sequel by $p_{\text{\it a}}$.
\subsubsection{Noisy comparator model} Let $\mathbf{lt}$ denote the noiseless {\em less than} operator, defined by $\mathbf{lt}(x, y) = 1$ if $x < y$, and $\mathbf{lt}(x, y) = 0$ otherwise.
The {\em noisy less than} operator, denoted by $\mathbf{lt}_{\mbox{\scriptsize pr}}$, is defined by injecting errors on the output of the noiseless one, according to the bit-flipping model defined in Section~\ref{subsec:output-swithcing}. In other words, the output of the noiseless $\mathbf{lt}$ operator is flipped with some probability value, which will be denoted in the sequel by $p_{\text{\it c}}$.
Finally, the {\em noisy minimum} operator is defined by:
\begin{equation}\label{eq:noisy-comp-def}
\mathbf{m}_{\mbox{\scriptsize pr}}(x, y) = \left\{\begin{array}{rl}
x, & \mbox{if } \mathbf{lt}_{\mbox{\scriptsize pr}}(x, y) = 1\\
y, & \mbox{if } \mathbf{lt}_{\mbox{\scriptsize pr}}(x, y) = 0
\end{array}\right.
\end{equation}
\subsubsection{Noisy XOR model} The noisy {\sc xor} operator, denoted by $\mathbf{x}_{\mbox{\scriptsize pr}}$ is defined by flipping the output of the noiseless operator with some probability value, which will be denoted in the sequel by $p_{\text{\it x}}$ (according to the bit-flipping error injection model in Section~\ref{subsec:output-swithcing}). It follows that:
\begin{equation}\label{eq:noisy-xor-def}
\mathbf{x}_{\mbox{\scriptsize pr}}(x, y) = \left\{\begin{array}{rl}
x\wedge y, & \mbox{with probability } 1-p_{\text{\it x}}\\
\overline{x\wedge y}, & \mbox{with probability } p_{\text{\it x}}
\end{array}\right.
\end{equation}
\medskip\noindent{\bf Assumption:} We further assume that the inputs and the output of the {\sc xor} operator may take values in either $\{0, 1\}$ or $\{-1, +1\}$ (using the usual $0{,}1$ to $\pm 1$ conversion). This assumption will be implicitly made throughout the paper.
\medskip\noindent{\bf Remark:} As a general rule, we shall refer to a noisy operator according to its underlying error injection model. For instance, a sign-preserving (resp. full-depth or sign-preserving bitwise-{\sc xor}ed) noisy adder, is a noisy adder whose underlying error injection model is sign-preserving (resp. one of the bitwise-{\sc xor} error injection models defined in Section~\ref{subsec:bitwise-xor-injection}). We shall also say that a noisy operator is {\em (highly) symmetric} if its underlying error injection model is so.
\subsection{Nested Operators}
As it can be observed from Algorithm~\ref{alg:ms}, several arithmetic/logic operations must be nested\footnote{For instance, $(d_n-1)$ additions -- where $d_n$ denotes the degree of the variable-node $n$ -- are required in order to compute each $\alpha_{m,n}$ message. Similarly, each $\beta_{m,n}$ message requires $(d_m-2)$ {\sc xor} operations and $(d_m-2)$ comparisons.} in order to compute the exchanged messages. Since all these operations (additions, comparisons, {\sc xor}) are commutative, the way they are nested does not have any impact on the infinite-precision MS decoding. However, this is no longer true for finite-precision decoding, especially in case of noisy operations. Therefore, one needs an assumption about how operators extend from two to more inputs.
\medskip Our assumption is the following. For $n \geq 2$ inputs, we randomly pick any two inputs and apply the operator on this pair. Then we replace the pair by the obtained output, and repeat the above procedure until there is only one output (and no more inputs) left.
The formal definition goes as follows. Let $\Omega\subset {\mathbb Z}$ and $\omega: \Omega\times \Omega\rightarrow \Omega$ be a noiseless or noisy operator with two operands. Let $\{x_i\}_{i=1:n} \subset \Omega$ be an unordered set of $n$ operands. We define:
$$\omega\left(\{x_i\}_{i=1:n}\right) = \omega(\cdots(\omega(x_{\pi(1)}, x_{\pi(2)}), \cdots), x_{\pi(n)}),$$
where $\pi$ is a random permutation of $1,\dots,n$.
\section{Noisy Min-Sum Decoding}\label{sec:noisy_ms_decoding}
\subsection{Finite-Precision Min-Sum Decoder}\label{subsec:finite-prec-dec-notation}
We consider a finite-precision MS decoder, in which the a priori information ($\gamma_n$) and the exchanged messages ($\alpha_{m,n}$ and $\beta_{m,n}$) are quantized on $q$ bits. The a posteriori information ($\tilde{\gamma}_n$) is quantized on $\tilde{q}$ bits, with $\tilde{q}>q$ (usually $\tilde{q}=q+1$, or $\tilde{q}=q+2$). We further denote:
\begin{itemize}
\item ${\cal M} = \{-Q,\dots,-1,0,+1,\dots,Q\}$, where $Q=2^{q-1}-1$, the alphabet of both the a priori information and the exchanged messages;
\item $\widetilde{\cal M} = \{-\widetilde{Q},\dots,-1,0,+1,\dots,\widetilde{Q}\}$, where $\widetilde{Q}=2^{\tilde{q}-1}-1$, the alphabet of the a posteriori information;
\item $\mathbf{q} : {\cal Y} \rightarrow {\cal M}$, a quantization map, where ${\cal Y}$ denotes the channel output alphabet;
\item $\mathbf{s}_{\cal M} : \mathbb{Z}\rightarrow {\cal M}$, the $q$-bit saturation map (defined in a similar manner as in (\ref{eq:saturation-map}));
\item $\mathbf{s}_{\widetilde{\cal M}} : \mathbb{Z}\rightarrow \widetilde{\cal M}$, the $\tilde{q}$-bit saturation map
\end{itemize}
\medskip\noindent{\bf Remark:} The quantization map $\mathbf{q}$ determines the $q$-bit quantization of the decoder soft input. Since $\mathbf{q}$ is defined on the channel input ({\em i.e.} $y_n$ values), it must also encompass the computation of the corresponding LLR values, whenever is necessary (see also the Remark at the end of Section~\ref{subsec:ms-decoding}).
Saturation maps $\mathbf{s}_{\cal M}$ and $\mathbf{s}_{\widetilde{\cal M}}$ define the finite-precision saturation of the exchanged messages and of the a posteriori information, respectively.
\subsection{Noisy Min-Sum Decoder}
The noisy (finite-precision) MS decoding is presented in Algorithm~\ref{alg:noisy_ms}.
We assume that $\tilde{q}$-bit adders are used to compute both $\alpha_{m,n}$ messages in the {\bf VN-processing} step, and $\tilde{\gamma}_n$ values in the {\bf AP-update} processing step. This is usually the case in practical implementations\footnote{In practical implementation, the $\tilde{\gamma}_n$ is computed first, and then $\alpha_{m,n}$ is obtained from $\tilde{\gamma}_n$ by subtracting the incoming $\beta_{m,n}$ message}, and allows us to use the same type of adder in both processing steps. This assumption explains as well the $q$-bit saturation of $\alpha_{m,n}$ messages in the {\bf VN-processing} step. Note also that the saturation of $\tilde{\gamma}_n$ values is actually done within the adder (see Equation~(\ref{eq:noisy-add-def})).
Finally, we note that the {\em hard decision} and the {\em syndrome check} steps in Algorithm~\ref{alg:noisy_ms} are assumed to be {\em noiseless}. We note however that the syndrome check step is optional, and if missing, the decoder stops when the maximum number of iterations is reached.
\input{algo_noisy_ms}
\subsection{Sign-Preserving Properties}\label{subsec:unit-sign-prez}
Let $\mathbf{U}$ denote any of the VN-processing or CN-processing units of the noiseless MS decoder. We denote by $\mathbf{U}_{\text{pr}}$ the corresponding unit of the noisy MS decoder. We say that $\mathbf{U}_{\text{pr}}$ is {\em sign-preserving} if for any incoming messages and any noise realization, the outgoing message is of the same sign as the message obtained when the same incoming messages are supplied to $\mathbf{U}$.
\medskip Clearly, $\text{CN}_{\text{pr}}$ is sign-preserving if and only if the {\sc xor}-operator is noiseless ($p_{\text{\it x}} = 0$). In case that the noisy {\sc xor}-operator severely degrades the decoder performance, it is possible to increase its reliability by using classical fault-tolerant techniques (as for instance modular redundancy, or multi-voltage design by increasing the supply voltage of the corresponding {\sc xor}-gate). The price to pay, when compared to the size or the energy consumption of the whole circuit, would be reasonable.
\medskip Concerning the VN-processing, it is worth noting that the $\text{VN}_{\text{pr}}$ is {\bf not} sign-preserving, even if the noisy adder is. This is due to the fact that multiple adders must be nested in order to complete the VN-processing. However, a sign-preserving adder might have several benefits. First, the error probability of the sign of variable-node messages would be lowered, which would certainly help the decoder. Second, if the noisy adder is sign-preserving and all the variable-node incoming messages have the same sign, then the $\text{VN}_{\text{pr}}$ does preserve the sign of the outgoing message. Put differently, in case that all the incoming messages agree on the same hard decision, the noisy VN-processing may change the confidence level, but cannot change the decision. This may be particularly useful, especially during the last decoding iterations.
\medskip Finally, the motivation behind the sign-preserving noisy adder model is to investigate its possible benefits on the decoder performance. If the benefits are worth it ({\em e.g.} one can ensure a target performance of the decoder), the sign-bit of the adder could be protected by using classical fault-tolerant solutions.
\section{Density Evolution}\label{sec:density_evolution}
\subsection{Concentration and Convergence Properties}
First, we note that our definition of {\em symmetry} is slightly more general than the one used in \cite{varshney2011performance}. Indeed, even if all the error injection models used within the noisy MS decoder are {\em symmetric}, the noisy MS decoder does not necessarily verify the \textsl{\textbf{symmetry}} property from \cite{varshney2011performance}. However, this property is verified in case of {\em highly symmetric} fault injection\footnote{According to the probabilistic models introduced in Section~\ref{subsec:proba-models-operators}, the noisy comparator and the noisy {\sc xor}-operator are highly symmetric, but the noisy adder does not necessarily be so!}. Nevertheless, the concentration and convergence properties proved in \cite{varshney2011performance} for \textsl{\textbf{symmetric}} noisy message-passing decoders, can easily be generalized to our definition of {\em symmetry}.
We summarize below the most important results; the proof relies essentially on the same arguments as in \cite{varshney2011performance}. We consider an {\em ensemble} of LDPC codes, with length $N$ and fixed degree distribution polynomials \cite{richardson2001capacity}. We choose a random code $\mathbf{C}$ from this ensemble and assume that a random codeword $\vect{x}\in\{-1, +1\}^N$ is sent over a binary-input memoryless symmetric channel. We fix some number of decoding iterations $\ell > 0$, and denote by $E^{(\ell)}_{\mathbf{C}}(\vect{x})$ the expected fraction of incorrect {\em messages}\footnote{Here, {\em ``messages''} may have any one of the three following meanings: ``variable-node messages'', or ``check-node messages'', or ``a posteriori information values''.} at iteration $\ell$.
\begin{theo}
Assume that all the error injection models used within the MS decoder are symmetric. Then, the following properties hold:
\begin{enumerate}
\item{}[{\em Conditional Independence of Error}] For any decoding iteration $\ell > 0$, the
expected fraction of incorrect messages $E^{(\ell)}_{\mathbf{C}}(\vect{x})$ does not depend on $\vect{x}$. Therefore, we may define $E^{(\ell)}_{\mathbf{C}} := E^{(\ell)}_{\mathbf{C}}(\vect{x})$.
\item{} [{\em Cycle-Free Case}] If the graph of $\mathbf{C}$ contains no cycles of length $2\ell$ or less, the expected fraction of incorrect messages $E^{(\ell)}_{\mathbf{C}}$ does not depend on the code $\mathbf{C}$ or the code-length $N$, but only on the degree distribution polynomials; in this case, it will be further denoted by $E^{(\ell)}_{\infty}(\vect{x})$.
\item{} [{\em Concentration Around the Cycle-Free Case}] For any $\delta > 0$, the probability that $E^{(\ell)}_{\mathbf{C}}$ lies outside the interval $\left(E^{(\ell)}_{\infty}(\vect{x})-\delta, E^{(\ell)}_{\infty}(\vect{x})+\delta\right)$ converges to zero exponentially fast in $N$.
\end{enumerate}
\end{theo}
\subsection{Density Evolution Equations}
In this section we derive density evolution equations for the noisy finite-precision MS decoding for a regular $(d_v,d_c)$ LDPC code. The study can be easily generalized to irregular LDPC codes, simply by averaging according to the degree distribution polynomials.
The objective of the density evolution technique is to recursively compute the probability mass functions of exchanged messages, through the iterative decoding process. This is done under the independence assumption of exchanged messages, holding in the asymptotic limit of the code length, in which case the decoding performance converges to the cycle-free case. Due to the symmetry of the decoder, the analysis can be further simplified by assuming that the all-zero codeword is transmitted through the channel. We note that our analysis applies to any memoryless symmetric channel.
\medskip Let $\ell > 0$ denote the decoding iteration. Superscript ${(\ell)}$ will be used to indicate the messages and the a posteriori information computed at iteration $\ell$. To indicate the value of a message on a randomly selected edge, we drop the variable and check node indexes from the notation (and we proceed in a similar manner for the a priori and a posteriori information). The corresponding probability mass functions are denoted as follows.
$$\begin{array}{r@{\ }c@{\ }ll}
C(z) & = & \Pr(\gamma = z), & \forall z\in{\cal M} \\
\widetilde{C}^{(\ell)}(\tilde{z})& = &\Pr(\tilde{\gamma}^{(\ell)}= \tilde{z}), &\forall \tilde{z}\in{\cal \widetilde{M}}\\
A^{(\ell)}(z)& = &\Pr\left(\alpha^{(\ell)} = z\right),& \forall z\in {\cal M}\\
B^{(\ell)}(z)& = &\Pr\left(\beta^{(\ell)} = z\right), &\forall z\in {\cal M}
\end{array}$$
\subsubsection{Expression of the input probability mass function $C$} \label{subsubsec:express-C}
The probability mass function $C$ depends only on the channel and the quantization map $\mathbf{q} : {\cal Y} \rightarrow {\cal M}$, where ${\cal Y}$ denotes the channel output alphabet (Section \ref{subsec:finite-prec-dec-notation}). We also note that for $\ell = 0$, we have $A^{(0)} = C$.
We give below the expression of $C$ for the BSC and the BI-AWGN channel models (see Section~\ref{subsec:ms-decoding}). For the BSC, the channel output alphabet is ${\cal Y} = \{-1, +1\}$, while for the BI-AWGN channel, ${\cal Y} = {\mathbb R}$.
Let $\mu$ be a positive number, such that $\mu \leq Q$. The quantization map $\mathbf{q}_{\mu}$ is defined as follows:
\begin{equation} \label{eq:quant-map}
\mathbf{q}_{\mu} : {\cal Y} \rightarrow {\cal M}, \ \ \mathbf{q}_{\mu}(y) = \mathbf{s}_{\cal M}([\mu{\cdot}y]),
\end{equation}
where $[\mu{\cdot}y]$ denotes the nearest integer to $\mu{\cdot}y$, and $\mathbf{s}_{\cal M}$ is the saturation map (Section~\ref{subsec:finite-prec-dec-notation}). For the BSC, we will further assume that $\mu$ is an integer. It follows that $\mathbf{q}_{\mu}(y) = \mu{\cdot}y$, $\forall y \in {\cal Y} = \{-1, +1\}$.
Considering the all-zero ($+1$) codeword assumption, the probability mass function $C$ can be computed as follows.
\begin{itemize}
\item For the BSC with crossover probability $\varepsilon$:
\begin{equation}
C(z) = \left\{\begin{array}{cl}
1-\varepsilon, & \mbox{if } z = \mu \\
\varepsilon, & \mbox{if } z = -\mu \\
0, & \mbox{otherwise}
\end{array}\right.
\end{equation}
\item For the BI-AWGN channel with noise variance $\sigma^2$:
\begin{equation}\label{eq:awgn_c}
C(z) = \left\{\begin{array}{cl}
1-q\left(\frac{-Q+0.5-\mu}{\mu\sigma}\right), & \mbox{if } z = -Q \\
q\left(\frac{z-0.5-\mu}{\mu\sigma}\right) - q\left(\frac{z+0.5-\mu}{\mu\sigma}\right),
& \mbox{if } -Q < z < +Q \\
q\left(\frac{Q-0.5-\mu}{\mu\sigma}\right), & \mbox{if } z = +Q
\end{array}\right.
\end{equation}
where $\displaystyle q(x) = \frac{1}{\sqrt{2\pi}}\int_{x}^{+\infty}\mbox{exp}\left(-\frac{u^2}{2}\right) du$ is the tail probability of the standard normal distribution (also known as the {\em $Q$-function}).
\end{itemize}
\subsubsection{Expression of $B^{(\ell)}$ as a function of $A^{(\ell-1)}$} \
In the sequel, we make the convention that $ \Pr({\mbox{\rm sgn}}(0) = 1)= \Pr({\mbox{\rm sgn}}(0) = -1)=1/2$.
The following notation will be used:
\begin{itemize}
\item $A_{[x,y]} = \displaystyle\sum_{z=x}^y A(z)$, for $x\leq y\in{\cal M}$
\item $A_{[0^{+},y]} = \displaystyle\frac{1}{2}A(0) + \sum_{z=1}^y A(z)$, for $y\in{\cal M}, \ y > 0$
\item $A_{[x, 0^{-}]} = \displaystyle\frac{1}{2}A(0) + \sum_{z=x}^{-1} A(z)$, for $x\in{\cal M}, \ x < 0$
\end{itemize}
For the sake of simplicity, we drop the iteration index, thus $B := B^{(\ell)}$ and $A := A^{(\ell-1)}$. We proceed by recursion on $i=2,\dots,d_c-1$, where $d_c$ denotes the check-node degree.
\noindent Let $\beta_1 := \alpha_1$, and for $i=2,\dots,d_c-1$ define:
$$\beta_i = {\color{red}\mathbf{x}_{\mbox{\scriptsize pr}}}({\mbox{\rm sgn}}(\beta_{i-1}),{\mbox{\rm sgn}}(\alpha_i)){\color{red}\mathbf{m}_{\mbox{\scriptsize pr}}}(|\beta_{i-1}|, |\alpha_i|)$$
\noindent Let also $B_{i-1}$ and $B_i$ denote the probability mass functions of $\beta_{i-1}$ and $\beta_i$, respectively (hence, $B_1 = A$).
\medskip \noindent First of all, for $z = 0$, we have:\\
$B_i(0) = \Pr(\beta_i = 0) = A(0)B_{i-1}(0) + \left[ B_{i-1}(0)(1-A(0)) + A(0)(1-B_{i-1}(0)) \right](1-p_c)$.
\medskip \noindent For $z\neq 0$, we proceed in several steps as follows:
\medskip\noindent\begin{minipage}{.495\linewidth}
For $z > 0$:
\medskip\resizebox{\linewidth}{!}{$\begin{array}{@{}r@{\ }c@{\ }l}
F^\prime_i(z) &\stackrel{\mbox{\tiny def}}{=}& \Pr(\beta_i \geq z \mid p_{\text{\it x}}=0)\\ & = &
\left[{B_{i-1}}_{[0^+,z-1]}A_{[z,Q-1]} + A_{[0^+,z-1]}{B_{i-1}}_{[z,Q-1]}\right]p_c \\
& + & \left[{B_{i-1}}_{[1-z,0^-]}A_{[-Q,-z]} + A_{[1-z,0^-]}{B_{i-1}}_{[-Q,-z]}\right]p_c \\
& + & {B_{i-1}}_{[z,Q-1]}A_{[z,Q-1]} + {B_{i-1}}_{[-Q,-z]}A_{[-Q,-z]}
\end{array}$}
\medskip$\begin{array}{@{}r@{\ }c@{\ }l}
F_i(z) & \stackrel{\mbox{\tiny def}}{=}& \Pr(\beta_i \geq z) \\
& = & (1-p_{\text{\it x}}).F^\prime_i(z)+p_{\text{\it x}}.G^\prime_i(-z)\\
B_i(z) & = & \Pr(\beta_i = z) = F_i(z) - F_i(z+1)
\end{array}$
\end{minipage}\hfill\begin{minipage}{.495\linewidth}
For $z < 0$:
\medskip\resizebox{\linewidth}{!}{$\begin{array}{@{}r@{\ }c@{\ }l}
G^\prime_i(z) &\stackrel{\mbox{\tiny def}}{=}& \Pr(\beta_i \leq z \mid p_{\text{\it x}}=0) \\ & = &
\left[{B_{i-1}}_{[0^+,-z-1]}A_{[-Q,z]} + A_{[0^+,-z-1]}{B_{i-1}}_{[-Q,z]}\right]p_c \\
&+ & \left[{B_{i-1}}_{[-z,Q-1]}A_{[z+1,0^{-}]} + A_{[-z,Q-1]}{B_{i-1}}_{[z+1,0^{-}]}\right]p_c \\
& + & {B_{i-1}}_{[-z,Q-1]}A_{[-Q,z]} + A_{[-z,Q-1]}{B_{i-1}}_{[-Q,z]}
\end{array}$}
\medskip$\begin{array}{@{}r@{\ }c@{\ }l}
G_i(z) & \stackrel{\mbox{\tiny def}}{=}&\Pr(\beta_i \geq z)\\
& =&(1-p_{\text{\it x}}).G^\prime_i(z)+p_{\text{\it x}}.F^\prime_i(-z)\\
B_i(z) & = & \Pr(\beta_i = z) = G_i(z) - G_i(z+1)
\end{array}$
\end{minipage}
\medskip \noindent Finally, we have that $B = B_{d_c-1}$.
\subsubsection{Expression of $A^{(\ell)}$ as a function of $B^{(\ell)}$ and $C$}
We derive at the same time the {\em expression of $\widetilde{C}^{(\ell)}$ as a function of $B^{(\ell)}$ and $C$}.
\medskip For simplicity, we drop the iteration index, so $A:= A^{(\ell)}$, $B:=B^{(\ell)}$, and $\widetilde{C}:=\widetilde{C}^{(\ell)}$. We denote by $\left({\cal E}, p_{\cal E}, \imath \mid \widetilde{\cal M}\right)$ the error injection model used to define the noisy adder.
We decompose each noisy addition into three steps (noiseless infinite-precision addition, saturation, and error injection), and proceed by recursion on $i=0,1,\dots,d_v$, where $d_v$ denotes the variable-node degree:
\begin{itemize}
\item For $i = 0$:\\
$\begin{array}{r@{\ }c@{\ }lr@{\ }c@{\ }l}
\Omega_0 & \stackrel{\mbox{\tiny def}}{=} &
\gamma\in {\cal M}\subseteq \widetilde{\cal M}, \hspace*{6.3mm}&
\widetilde{C}_{0}(\tilde{z}) & \stackrel{\mbox{\tiny def}}{=} &
\Pr(\Omega_0 = \tilde{z})
= \left\{\begin{array}{cl} C(\tilde{z}), &\mbox{if } \tilde{z}\in{\cal M}\\
0, & \mbox{if } \tilde{z}\in\widetilde{\cal M}\setminus{\cal M}
\end{array}\right.
\end{array}$
\item For $i = 1,\dots, d_v$:\\
$\begin{array}{r@{\ }c@{\ }lr@{\ }c@{\ }l}
\omega_i & \stackrel{\mbox{\tiny def}}{=} &
\Omega_{i-1} + \beta_{m_i,n} \in \mathbb{Z}, &
c_{i}(w) & \stackrel{\mbox{\tiny def}}{=} &
\Pr(\omega_i = w) = \sum_u \widetilde{C}_{i-1}(u) B(w-u), \forall w \in \mathbb{Z} \\[3mm]
%
\tilde{\omega}_i & \stackrel{\mbox{\tiny def}}{=} &
\mathbf{s}_{\widetilde{\cal M}}(\omega_i) \in \widetilde{\cal M}, &
\tilde{c}_{i}(\tilde{w}) & \stackrel{\mbox{\tiny def}}{=} &
\Pr(\tilde{\omega}_i = \tilde{w}) =
\renewcommand{\arraystretch}{1.2}
\left\{\begin{array}{l@{\ }l}
c_i(\tilde{w}), & \mbox{if } \tilde{w} \in \widetilde{\cal M} \setminus\{\pm \widetilde{Q}\} \\
\sum_{w \leq -\widetilde{Q}} c_i(w), & \mbox{if } \tilde{w} = -\widetilde{Q}\\
\sum_{w \geq +\widetilde{Q}} c_i(w), & \mbox{if } \tilde{w} = +\widetilde{Q}
\end{array}\right.\\[10mm]
%
\Omega_i & \stackrel{\mbox{\tiny def}}{=} &
\imath(\tilde{\omega}_i, e) \in \widetilde{\cal M}, &
\widetilde{C}_{i}(\tilde{z}) & \stackrel{\mbox{\tiny def}}{=} &
\Pr(\Omega_i = \tilde{z}) =
\sum_{\tilde{\omega}} \sum_{e} {\delta}_{\imath(\tilde{\omega}, e)}^{\tilde{z}}
p_{\cal E}(e) \tilde{c}_{i}(\tilde{\omega}),
\forall \tilde{z} \in \widetilde{\cal M} \\
& & & \multicolumn{3}{l}{\mbox{where } {\delta}_{x}^y = 1 \mbox{ if } x = y,
\mbox{ and } {\delta}_{x}^y = 0 \mbox{ if } x \neq y. }
\end{array}$
\end{itemize}
\renewcommand{\arraystretch}{1}
\medskip\noindent
Note that in the definition of $\Omega_i$ above, $e$ denotes an error drown from the error set ${\cal E}$ according to the error probability distribution $p_{\cal E}$.
\medskip\noindent Finally, we have:
\begin{itemize}
\item ${A} = \mathbf{s}_{\cal M}\left( \widetilde{C}_{d_v-1}\right)$
\item $\widetilde{C} = \widetilde{C}_{d_v}$
\end{itemize}
In the first equation above, applying the saturation operator $\mathbf{s}_{\cal M}$ on the probability mass function $\widetilde{C}_{d_v-1}$ means that all the probability weights corresponding to values $\tilde{w}$ outside ${\cal M}$ must be accumulated to the probability of the corresponding boundary value of ${\cal M}$ (that is, either $-Q$ or $+Q$, according to whether $\tilde{w}< -Q$ or $\tilde{w}< +Q$).
\medskip\noindent {\bf Remark:} If the noisy adder is defined by one of the bitwise-{\sc xor} error injection models (Section \ref{subsec:bitwise-xor-injection}), then the third equation from the above recursion (expression of $\widetilde{C}_{i}$ as a function of $\tilde{c}_{i}$) may be rewritten as follows:
\begin{itemize}
\item Sign-preserving bitwise-{\sc xor}ed noisy adder
\renewcommand{\arraystretch}{1.5}
\begin{equation}\label{eq:ct-sign-preserve}
\widetilde{C}_{i}(\tilde{z}) = \left\{\begin{array}{ll}
\displaystyle (1-p_{\text{\it a}})\tilde{c}_{i}(\tilde{z}) +
\frac{1}{\widetilde{Q}}p_{\text{\it a}}\left( \tilde{c}_{i\,{[\leq\, 0^{-}]}}
- \tilde{c}_{i}(z) \right), & \mbox{if } \tilde{z} < 0 \\
\displaystyle (1-p_{\text{\it a}})\tilde{c}_{i}(0) +
\frac{1}{\widetilde{Q}}p_{\text{\it a}}\left(1 - \tilde{c}_{i}(0) \right), &
\mbox{if } \tilde{z} = 0 \\
\displaystyle (1-p_{\text{\it a}})\tilde{c}_{i}(\tilde{z}) +
\frac{1}{\widetilde{Q}}p_{\text{\it a}}\left( \tilde{c}_{i\,{[\geq\, 0^{+}]}}
- \tilde{c}_{i}(z) \right), & \mbox{if } \tilde{z} > 0
\end{array}\right.
\end{equation}
where $\tilde{c}_{i\,{[\leq\, 0^{-}]}} = \sum_{\tilde{\omega} < 0} \tilde{c}_{i}(\tilde{\omega}) + \frac{1}{2}\tilde{c}_{i}(0)$, and $\tilde{c}_{i\,{[\geq\, 0^{+}]}} = \frac{1}{2}\tilde{c}_{i}(0) + \sum_{\tilde{\omega} > 0} \tilde{c}_{i}(\tilde{\omega})$.
\renewcommand{\arraystretch}{1}
\item Full-depth bitwise-{\sc xor}ed noisy adder
\begin{equation}\label{eq:ct-full-depth}
\widetilde{C}_{i}(\tilde{z}) = (1-p_{\text{\it a}})\tilde{c}_{i}(\tilde{z}) + \frac{1}{2\widetilde{Q}}p_{\text{\it a}}
\left(1 - \tilde{c}_{i}(\tilde{z})\right)
\end{equation}
\end{itemize}
\medskip Finally, we note that the density evolution equations for the noiseless finite-precision MS decoder can be obtained by setting $p_{\text{\it a}} = p_{\text{\it c}} = p_{\text{\it x}}=0$.
\subsection{Error Probability and Useful and Target-BER regions} \label{subsec:err_proba_and_useful_reg}
\subsubsection{Decoding Error Probability}
The error probability at decoding iteration $\ell$, is defined by:
\begin{equation}\label{eq:p_iter}
P_e^{(\ell)} = \displaystyle\sum_{\tilde{z}=-\widetilde{Q}}^{-1}\widetilde{C}^{(\ell)}(\tilde{z})+\frac{\widetilde{C}^{(\ell)}(0)}{2}
\end{equation}
\begin{prop}\label{prop:lower-bounds}
The error probability at decoding iteration $\ell$ is lower-bounded as follows:
\begin{itemize}
\item[(a)] For the sign-preserving bitwise-{\sc xor}ed noisy adder: $P_e^{(\ell)} \geq \displaystyle \frac{1}{2\widetilde{Q}}p_{\text{\it a}}$.
\item[(b)] For the full-depth bitwise-{\sc xor}ed noisy adder: $P_e^{(\ell)} \geq \displaystyle \frac{1}{2}p_{\text{\it a}} + \frac{1}{4\widetilde{Q}}p_{\text{\it a}}$.
\end{itemize}
\end{prop}
\noindent{\em Proof.} {\em (a)} Using $\widetilde{C} = \widetilde{C}_{d_v}$ and equations (\ref{eq:p_iter}) and (\ref{eq:ct-sign-preserve}), it follows that
$P_e^{(\ell)} = (1-p_{\text{\it a}}) \tilde{c}_{d_v\,{[\leq\, 0^{-}]}} + \frac{1}{2\widetilde{Q}}p_{\text{\it a}}\left(1-2\tilde{c}_{d_v\,{[\leq\, 0^{-}]}}\right) +
p_{\text{\it a}}\left(\tilde{c}_{d_v\,{[\leq\, 0^{-}]}} - \frac{1}{2}\tilde{c}_{d_v}(0)\right) \geq (1-p_{\text{\it a}}) \tilde{c}_{d_v\,{[\leq\, 0^{-}]}} + \frac{1}{2\widetilde{Q}}p_{\text{\it a}}\left(1-2\tilde{c}_{d_v\,{[\leq\, 0^{-}]}}\right) \geq \frac{1}{2\widetilde{Q}}p_{\text{\it a}}$, since the function
$(1-p_{\text{\it a}}) x + \frac{1}{2\widetilde{Q}}p_{\text{\it a}}\left(1-2x\right)$ is an increasing function of $x\in[0,1]$.
\noindent {\em (b)} Equations (\ref{eq:p_iter}) and (\ref{eq:ct-full-depth}) imply that $P_e^{(\ell)} = \frac{1}{2}p_{\text{\it a}} + (1-p_{\text{\it a}})\tilde{c}_{d_v\,{[\leq\, 0^{-}]}} + \frac{1}{4\widetilde{Q}}p_{\text{\it a}}\left(1 - 2\tilde{c}_{d_v\,{[\leq\, 0^{-}]}}\right)\geq \frac{1}{2}p_{\text{\it a}} + \frac{1}{4\widetilde{Q}}p_{\text{\it a}}$\\
$\mbox{ }$ \hfill$\square$
\medskip Note that the above lower bounds are actually inferred from the error injection in the {\em last (the $d_v$-th) addition} performed when computing the a posteriori information value. Therefore, these lower bounds are not expected to be tight. However, if the channel error probability is small enough, the sign-preserving lower bound proves to be tight in the asymptotic limit of $\ell$ (this will be discussed in more details in Section~\ref{sec:asymptotic-analysis}). Note also that by protecting the sign of the noisy adder, the bound is lowered by a factor of roughly $\widetilde{Q}$, which represents an exponential improvement with respect to the number of bits of the adder.
\medskip In the asymptotic limit of the code-length, $P_e^{(\ell)}$ gives the probability of the hard bit estimates being in error at decoding iteration $\ell$. For the (noiseless, infinite-precision) BP decoder, the error probability is usually a decreasing function of $\ell$. This is no longer true for the noiseless, infinite-precision MS decoder, for which the error probability may increase with $\ell$. However, both decoders exhibit a {\em threshold phenomenon}, separating the region where error probability goes to zero (as the number of decoding iterations goes to infinity), from that where it is bounded above zero \cite{richardson2001capacity}.
Things get more complicated for the noisy (finite-precision) MS decoder. First, the error probability have a more unpredictable behavior. It does not always converge and it may become periodic\footnote{In fact, for both BSC and BI-AWGN channels, the only cases we observed, in which the sequence $\left(P_e^{(\ell)}\right)_{\ell > 0}$ does not converge, are those cases in which this sequence becomes periodic for $\ell$ large enough.} when the number of iterations goes to infinity. Second, the error probability is always bounded above zero (Proposition~\ref{prop:lower-bounds}), since there is a non-zero probability of fault injection at any decoding iteration. Hence, a decoding threshold, similar to the noiseless case, cannot longer be defined.
\medskip
Following \cite{varshney2011performance}, we define below the notions of useful decoder and target error rate threshold. We consider a channel model depending on a channel parameter $\chi$, such that the channel is degraded by increasing $\chi$ (for example, the crossover probability for the BSC, or the noise variance for the BI-AWGN channel). We will use subscript $\chi$ to indicate a quantity that depends on $\chi$. Hence, in order to account that $P_{e}^{(\ell)}$ depends also on the value of the channel parameter, it will be denoted in the following by $P_{e,\chi}^{(\ell)}$.
\subsubsection{Useful Region}
The first step is to evaluate the channel and hardware parameters yielding a final probability of error (in the asymptotic limit of the number of iterations) less than the {\em input error probability}. The latter probability is given by
$P_{e,\chi}^{(0)} = \sum_{z=-Q}^{-1} C(z) + \frac{1}{2} C(0)$, where $C$ is the probability mass function of the quantized a priori information of the decoder (see Section~\ref{subsubsec:express-C}).
Following \cite{varshney2011performance}, the decoder is said to be {\em useful} if $\left(P_{e,\chi}^{(\ell)}\right)_{\ell > 0}$ is convergent, and:
\begin{equation}
P_{e,\chi}^{(\infty)} \stackrel{\text{def}}{=} \lim_{\ell \rightarrow \infty} P_{e,\chi}^{(\ell)} < P_{e,\chi}^{(0)}
\end{equation}
The ensemble of the parameters that satisfy this condition constitutes the {\em useful region} of the decoder.
\subsubsection{Target Error Rate Threshold}
For noiseless-decoders, the decoding threshold is defined as the supremum channel noise, such that the error probability converges to zero as the number of decoding iterations goes to infinity. However, for noisy decoders this error probability does not converge to zero, and an alternative definition of the decoding threshold has been introduced in \cite{varshney2011performance}. Accordingly, for a target bit-error rate $\eta$, the $\eta$-threshold is defined\footnote{In \cite{varshney2011performance}, the $\eta$-threshold is defined by $\chi^{\ast} (\eta)=\sup \left\{\chi \mid P_{e,\chi}^{(\infty)} \mbox{ exists and } P_{e,\chi}^{(\infty)} < \eta \right\}$, and consequently, there might exist a channel parameter value $\chi' < \chi^{\ast} (\eta)$, for which $P_{e,\chi'}^{(\infty)}$ does not exist. In order to avoid this happening, our definition is slightly different from the one in \cite{varshney2011performance}.} by:
\begin{equation}\label{eq:eta_threshold}
\chi^{\ast} (\eta)=\sup \left\{\chi \mid P_{e,\chi'}^{(\infty)} \mbox{ exists and } P_{e,\chi'}^{(\infty)} < \eta, \ \forall \chi'\in[0, \chi] \right\}
\end{equation}
\subsection{Functional Threshold}\label{subsec:funct_threshold}
Although the $\eta$-threshold definition allows determining the maximum channel noise for which the bit error probability can be reduced below a target value, there is not significant change in the behavior of the decoder when the channel noise parameter $\lambda$ increases beyond the value of $\chi^{\ast} (\eta)$.
In this section, a new threshold definition is introduced in order to identify the channel and hardware parameters yielding to a sharp change in the decoder behavior, similar to the change that occurs around the threshold of the noiseless decoder.
This threshold will be referred to as the {\em functional threshold}.
The aim is to detect a sharp increase (e.g. discontinuity) in the error probability of the noisy decoder, when $\lambda$ goes beyond this functional threshold value. The threshold definition we propose make use of the Lipschitz constant of the function $\chi \mapsto P_{e}^{(\infty)}(\chi)$ in order to detect a sharp change of $P_{e}^{(\infty)}(\chi)$ with respect to $\chi$. The definition of the Lipschitz constant is first restated for the sake of clarity.
\begin{defi}
Let $f:I \rightarrow \mathbb{R}$ be a function defined on an interval $I\subseteq \mathbb{R}$.
The {\em Lipschitz constant} of $f$ in $I$ is defined as
\begin{equation}
L(f, I) = \sup_{x\neq y \in I} \frac{|f(x)-f(y)|}{|x-y|} \in \mathbb{R}_{+} \cup \{+\infty\}
\end{equation}
For $a\in I$ and $\delta > 0$, let $I_a(\delta) = I \cap (a-\delta, a+\delta)$.
The {\em (local) Lipschitz constant} of $f$ in $a\in I$ is defined by:
\begin{equation}
L(f, a) = \inf_{\delta > 0} L(f, I_a(\delta)) \in \mathbb{R}_{+} \cup \{+\infty\}
\end{equation}
\end{defi}
Note that if $a$ is a discontinuity point of $f$, then $L(f, a) = +\infty$.
On the opposite, if $f$ is differentiable in $a$, then the Lipschitz constant in $a$ corresponds to the absolute value of the derivative.
Furthermore, if $L(f, I) < +\infty$, then $f$ is uniformly continuous on $I$ and almost everywhere differentiable.
In this case, $f$ is said to be {\em Lipschitz continuous} on $I$.
The functional threshold is then defined as follows.
\begin{defi}\label{def:ft}
For given hardware parameters and a channel parameter $\chi$, the decoder is said to be {\em functional} if
\begin{description}
\item[$(a)$] The function $x \mapstoP_{e}^{(\infty)}(x)$ is defined on $[0,\chi]$
\item[$(b)$] $P_{e}^{(\infty)}$ is Lipschitz continuous on $[0, \chi]$
\item[$(c)$] $L\left(P_{e}^{(\infty)}, x\right)$ is an increasing function of $x\in [0, \chi]$
\end{description}
Then, the functional threshold $\bar{\chi}$ is defined as:
\begin{equation}
\bar{\chi} = \sup \{ \chi \mid \mbox{conditions } (a), (b) \mbox{ and } (c) \mbox{ are satisfied}\}
\end{equation}
\end{defi}
The use of the Lipschitz constant allows a rigorous definition of the functional threshold, while avoiding the use of the derivative (which would require $P_{e}^{(\infty)}(\lambda)$ to be a piecewise differentiable function of $\lambda$). As it will be further illustrated in Section~\ref{sec:asymptotic-analysis}, the functional threshold corresponds to a transition between two modes.
The first mode corresponds to the channel parameters leading to a low level of error probability, \emph{i.e.}, for which the decoder can correct most of the errors from the channel.
In the second mode, the channel parameters lead to a much higher error probability level.
If $L\left(P_{e}^{(\infty)}, \bar{\chi}\right) = +\infty$, then $\bar{\chi}$ is a discontinuity point of $P_{e}^{(\infty)}$ and the transition between the two levels is sharp.
If $L\left(P_{e}^{(\infty)}, \bar{\chi}\right) < +\infty$, then $\bar{\chi}$ is an inflection point of $P_{e}^{(\infty)}$ and the transition is smooth.
With the Lipschitz constant, one can characterize the transition in both cases.
However, the second case corresponds to a degenerated one, in which the hardware noise is too high and leads to a non-standard asymptotic behavior of the decoder.
That is why a set of admissible hardware noise parameters is defined as follows.
\begin{defi}
The set of \emph{admissible hardware parameters} is the set of hardware noise parameters $(p_a,p_c,p_x)$ for which $L\left(P_{e}^{(\infty)}, \bar{\chi}\right) = +\infty$.
\end{defi}
In the following, as each threshold definition helps at illustrating different effects, one or the other definition will be used, depending on the context.
\section{Asymptotic analysis of the noisy Min-sum decoder}\label{sec:asymptotic-analysis}
In this section, the density evolution equations derived previously are used to analyze the asymptotic performance ({\em i.e.} in the asymptotic limit of both the code length and number of iterations) of the noisy MS decoder.
Unless specified otherwise, the following parameters are used throughout this section:
\medskip\noindent{\bf Code parameters:}
\begin{itemize}
\item We consider the ensemble of regular LDPC codes with variable-node degree $d_v=3$ and check-node degree $d_c=6$
\end{itemize}
\noindent{\bf Quantization parameters:}
\begin{itemize}
\item The a priori information and exchanged messages are quantized on $q = 4$ bits; hence, $Q = 7$ and ${\cal M} = \{-7,\dots, +7\}$.
\item The a posteriori information is quantized on $\tilde{q} = 5$ bits; hence, $\widetilde{Q} = 15$ and $\widetilde{\cal M} = \{-15,\dots, +15\}$.
\end{itemize}
\medskip\noindent We analyze the decoding performance depending on:
\begin{itemize}
\item The quantization map $\mathbf{q}_{\mu} : {\cal Y} \rightarrow {\cal M}$, defined in Equation~(\ref{eq:quant-map}). The factor $\mu$ will be referred to as the {\em channel-output scale factor}, or simply the {\em channel scale factor}.
\item The parameters of the noisy adder, comparator, and {\sc xor}-operator, defined respectively in Equations (\ref{eq:noisy-add-def}), (\ref{eq:noisy-comp-def}), and (\ref{eq:noisy-xor-def}).
\end{itemize}
\subsection{Numerical results for the BSC}\label{subsec:num_results_bsc}
For the BSC, the channel output alphabet is ${\cal Y} = \{-1, +1\}$ and the quantization map is defined by $\mathbf{q}_{\mu}(-1) = -\mu$ and $\mathbf{q}_{\mu}(+1) = +\mu$, with $\mu\in\{1,\dots,Q\}$.
The infinite-precision MS decoder (Algorithm~\ref{alg:ms}), is known to be independent of the scale factor $\mu$. This is because $\mu$ factors out from all the processing steps in Algorithm~\ref{alg:ms}, and therefore does not affect in any way the decoding process. This is no longer true for the finite precision decoder (due to saturation effects), and we will show in this section that, even in the noiseless case, the scale factor $\mu$ can significantly impact the performance of the finite precision MS decoder.
We start by analyzing the performance of the MS decoder with quantization map $\mathbf{q}_{1}$, and then we will analyze its performance with an optimized quantization map $\mathbf{q}_{\mu}$.
\subsubsection{Min-Sum decoder with quantization map $\mathbf{q}_{1}$}\label{subsec:ms_q1}
The case $\mu=1$ leads to an ``unconventional'' behavior,
as in some particular cases the noise introduced by the device can help the MS decoder to escape from fixed points attractors, and may actually result in an increased correction capacity with respect to the noiseless decoder. This behavior will be discussed in more details in this section.
We start with the noiseless decoder case. Figure~\ref{fig:pinf_noiselessMS} shows the asymptotic error probability $P_{e}^{(\infty)}$ as a function of $p_0$. It can be seen that $P_{e}^{(\infty)}$ decreases slightly with $p_0$, until $p_0$ reaches a threshold value $p_{\text{th}} = 0.039$, where $P_{e}^{(\infty)}$ drops to zero. This is the {\em classical} threshold phenomenon mentioned in Section~\ref{subsec:err_proba_and_useful_reg}: for $p_0 > p_{\text{th}}$, the decoding error probability is bounded far above zero ($P_{e}^{(\infty)} > 0.31$), while for $p_0 < p_{\text{th}}$, one has $P_{e}^{(\infty)} = 0$.
\begin{figure}[!thb]
\centering
\includegraphics[width=90mm]{./fig_noisy_ms/pinf_noiselessMS}
\caption{Asymptotic error probability $P_{e}^{(\infty)}$ of the noiseless MS decoder as a function of $p_0$}
\label{fig:pinf_noiselessMS}
\end{figure}
\begin{figure}[!thb]
\centering
\subfigure[$P_{e}^{(\ell)}$ plotted in linear scale]{\includegraphics[width=80mm]{./fig_noisy_ms/piter_noisy_adder}\label{subfig:piter_noisy_adder_lin}}
\subfigure[$P_{e}^{(\ell)}$ plotted in logarithmic scale]{\includegraphics[width=80mm]{./fig_noisy_ms/piter_noisy_adder_log}\label{subfig:piter_noisy_adder_log}}
\caption{Effect of the noisy adder on the asymptotic performance of the MS decoder ($p_0 = 0.06$)}
\label{fig:piter_noisy_adder}
\end{figure}
Now, we consider a $p_0$ value slightly greater than the threshold of the noiseless decoder, and investigate the effect of the noisy adder on the decoder performance. Let us fix $p_0 = 0.06$.
Figure~\ref{subfig:piter_noisy_adder_lin} shows the decoding error probability at iteration $\ell$, for different parameters $p_a \in\{10^{-30}, 10^{-15}, 10^{-5}\}$ of the noisy adder. For each $p_a$ value, there are two superimposed curves, corresponding to the full-depth (``fd'', solid curve) and sign-preserving (``sp'', dashed curve) error models of the noisy adder.
\noindent The error probability of the noiseless decoder is also plotted (solid black curve): it can be seen that it increases rapidly from the initial value $P_{e}^{(0)} = p_0$ and closely approaches the limit value $P_{e}^{(\infty)} = 0.323$ after a few number of iterations.
When the adder is noisy, the error probability increases during the first decoding iterations, and behaves similarly as in the noiseless case. It may approach the limit value from the noiseless case, but starts decreasing after some number of decoding iterations. However, it remains bounded above zero, according to the lower bounds from Proposition~\ref{prop:lower-bounds}. This can be seen in Figure~\ref{subfig:piter_noisy_adder_log}, where $P_{e}^{(\ell)}$ plotted in logarithmic scale. The asymptotic values $P_{e}^{(\infty)}$ and the corresponding lower-bounds values from Proposition~\ref{prop:lower-bounds} are shown in Table~\ref{tab:pinf_noisy_ader}. It can be seen that these bounds are tight, especially in the sign-preserving case.
\begin{table}[!htb]
\caption{Asymptotic error probability of the MS decoding with noisy adder ($p_0 = 0.06$)}
\label{tab:pinf_noisy_ader}
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
\multicolumn{2}{|r|}{$p_a$} & $10^{-30}$ & $10^{-15}$ & $10^{-5}$ \\
\hline\hline
full & $P_{e}^{(\infty)}$ & $8.500\times 10^{-31}$ & $8.500\times 10^{-16}$ & $8.507\times 10^{-6}$ \\
\cline{2-5}
depth & lower-bound & $5.167\times 10^{-31}$ & $5.167\times 10^{-16}$ & $5.167\times 10^{-6}$ \\
\hline\hline
sign & $P_{e}^{(\infty)}$ & $3.333\times 10^{-32}$ & $3.333\times 10^{-17}$ & $3.333\times 10^{-7}$ \\
\cline{2-5}
preserving & lower-bound & $3.333\times 10^{-32}$ & $3.333\times 10^{-17}$ & $3.333\times 10^{-7}$ \\
\hline
\end{tabular}
\end{table}
\begin{figure}[!htb]
\centering
\subfigure[Iteration $\ell = 0$]{\includegraphics[width=.33\linewidth]{./fig_noisy_ms/ct_noiselessMS_i0}\label{subfig:ct_noiselessMS_i0}}%
\subfigure[Iteration $\ell = 5$]{\includegraphics[width=.33\linewidth]{./fig_noisy_ms/ct_noiselessMS_i5}\label{subfig:ct_noiselessMS_i5}}%
\subfigure[Iteration $\ell = 20$]{\includegraphics[width=.33\linewidth]{./fig_noisy_ms/ct_noiselessMS_i20}\label{subfig:ct_noiselessMS_i20}}%
\caption{Probability mass function of the a posteriori information $\widetilde{C}^{(\ell)}$ (noiseless MS decoder)}
\label{fig:ct_noiselessMS}
\end{figure}
\begin{figure}[!htb]
\centering
\subfigure[Iteration $\ell = 20$]{\includegraphics[width=.33\linewidth]{./fig_noisy_ms/ct_noisy_adder_i20}\label{subfig:ct_noisy_adder_i20}}%
\subfigure[Iteration $\ell = 23$]{\includegraphics[width=.33\linewidth]{./fig_noisy_ms/ct_noisy_adder_i23}\label{subfig:ct_noisy_adder_i23}}%
\subfigure[Iteration $\ell = 30$]{\includegraphics[width=.33\linewidth]{./fig_noisy_ms/ct_noisy_adder_i30}\label{subfig:ct_noisy_adder_i30}}%
\caption{Probability mass function of the a posteriori information $\widetilde{C}^{(\ell)}$ (MS decoder with full-depth noisy adder, $p_a = 10^{-15}$)}
\label{fig:ct_noisy_adder}
\end{figure}
\medskip The above behavior of the MS decoder is explained by the fact that the noise present in the adder helps the MS decoder to escape from fixed points attractors. Figure~\ref{fig:ct_noiselessMS} illustrates the evolution of the probability mass function $\widetilde{C}^{(\ell)}$ for the noiseless decoder. At iteration $\ell = 0$, $\widetilde{C}^{(0)}$ is supported in $\pm 1$, with $\widetilde{C}^{(0)}(-1) = p_0$ and $\widetilde{C}^{(0)}(+1) = 1-p_0$. It evolves during the iterative decoding, and reaches a fixed point of the density evolution for $\ell = 20$. Note that since all variable-nodes are of degree $d_v = 3$, it can be easily seen that, for $\ell \geq 1$, $\widetilde{C}^{(\ell)}$ is supported only on even values. These ``gaps'' in the probability mass function seem lead to favorable conditions for the occurrence of density-evolution fixed-points.
Figure~\ref{fig:ct_noisy_adder} illustrates the evolution of the probability mass function $\widetilde{C}^{(\ell)}$ when the full-depth noisy adder with $p_a = 10^{-15}$ is used within the MS decoder. At iteration $\ell = 20$, $\widetilde{C}^{(\ell)}$ is virtually the same as in the noiseless case. However, the noisy adder allows the decoder to escape from this fixed-point, as it can be seen for iterations $\ell = 23$ and $\ell = 30$. For $\ell > 30$, the $\widetilde{C}^{(\ell)}$ moves further on the right, until the corresponding error probability $P_e^{(\ell)}$ reaches the limit value $P_e^{(\infty)} = 8.5\times 10^{-16}$.
It is worth noting that neither the noisy comparator nor the {\sc xor}-operator can help the decoder to escape from fixed-point distributions, as they do not allow ``filling the gaps'' in the support of $\widetilde{C}^{(\ell)}$.
\medskip We focus now on the useful region of the noisy MS decoder. We assume that only the adder is noisy, while the comparator and the {\sc xor}-operator are noiseless.
\begin{figure}[!thb]
\centering
\includegraphics[width=90mm]{./fig_noisy_ms/ureg_add_sp_mu1_pt}
\vspace*{-3mm}\caption{Useful and non-convergence regions of the MS decoder with sign-preserving noisy adder}
\label{fig:ureg_add_sp_mu1_pt}
\end{figure}
\begin{figure}[!thb]
\centering
\subfigure[Point $A(p_0 = 0.03, p_a = 0.027)$]{\includegraphics[width=.4\linewidth]{./fig_noisy_ms/piter_noisy_adder_ptA}\label{subfig:piter_noisy_adder_ptA}}\ \
\subfigure[Point $B(p_0 = 0.03, p_a = 0.03)$]{\includegraphics[width=.4\linewidth]{./fig_noisy_ms/piter_noisy_adder_ptB}\label{subfig:piter_noisy_adder_ptB}}%
\subfigure[Point $C(p_0 = 0.03, p_a = 0.039)$]{\includegraphics[width=.4\linewidth]{./fig_noisy_ms/piter_noisy_adder_ptC}\label{subfig:piter_noisy_adder_ptC}}\ \
\subfigure[Point $D(p_0 = 0.03, p_a = 0.042)$]{\includegraphics[width=.4\linewidth]{./fig_noisy_ms/piter_noisy_adder_ptD}\label{subfig:piter_noisy_adder_ptD}}%
\vspace*{-2mm}\caption{Decoding error probability $P_{e}^{(\ell)}$ of the noisy MS decoder, for $p_0 = 0.03$ and sign-preserving noisy adder with various $p_a$ values
\label{fig:piter_noisy_adder_pts}
\end{figure}
The useful region for the sign-preserving noisy adder model is shown in Figure~\ref{fig:ureg_add_sp_mu1_pt}. The
useful region is shaded in gray and delimited by either a solid black curve or a dashed red curve. Although one would expect that $P_e^{(\infty)} = p_0$ on the border of the useful region, this equality only holds on the solid black border. On the dashed red border, one has $P_e^{(\infty)} < p_0$. The reason why the useful region does not extend beyond the dashed red border is that for points located on the other side of this border the sequence $(P_e^{(\ell)})_{\ell > 0}$ is periodic, and hence it does not converge! The region shaded in brown in Figure~\ref{fig:ureg_add_sp_mu1_pt} is the {\em non-convergence region} of the decoder. Note that the non-convergence region gradually narrows in the upper part, and there is a small portion of the useful region delimited by the non-convergence region on the left and the black border on the right. Finally, we note that points with $p_a = 0$ (noiseless decoder) and $p_0 > 0.039$ (threshold of the noiseless decoder) -- represented by the solid red line superimposed on the vertical axis in Figure~\ref{fig:ureg_add_sp_mu1_pt} -- are excluded from the useful region. Indeed, for such points $P_e^{(\infty)} > p_0$; however, for $p_a$ greater than but close to zero, we have $P_e^{(\infty)}\approx \frac{p_a}{2\widetilde{Q}}$ (see Figure \ref{fig:piter_noisy_adder} and related discussion).
We exemplify the decoder behavior on four points located on one side and the other of the left and right boundaries of the non-convergence region. These points are indicated in Figure~\ref{fig:ureg_add_sp_mu1_pt} by $A, B, C$, and $D$. For all the four points $p_0 = 0.03$, while $p_a = 0.027, 0.03, 0.039$, and $0.042$, respectively.
The error probability $(P_e^{(\ell)})_{\ell > 0}$ is plotted for each one of these points in Figure~\ref{fig:piter_noisy_adder_pts}. The point $A$ belongs to the useful region, and it can be seen from Figure~\ref{subfig:piter_noisy_adder_ptA} that $(P_e^{(\ell)})_{\ell > 0}$ converges to $P_e^{(\infty)} = 9.11\times 10^{-4} < p_0$. For the point $B$, located just on the other side of the dashed red border of the useful region, $(P_e^{(\ell)})_{\ell > 0}$ exhibits a periodic behavior (although we only plotted the first $500$ iterations, we verified the periodic behavior on the first $5\times 10^4$ iterations). Crossing the non-convergence region from left to the right, the amplitude between the inferior and superior limits of $(P_e^{(\ell)})_{\ell > 0}$ decreases (point C), until it reaches again a convergent behavior (point D). Note that $D$ is outside the useful region, as $(P_e^{(\ell)})_{\ell > 0}$ converges to $P_e^{(\infty)} = 0.0605 > p_0$.
The non-convergence region gradually narrows in the upper part, and for $0\leq p_a < 0.01$ it takes the form of a {\em discontinuity line}: $P_e^{(\infty)}$ takes values close to $10^{-4}$ just below this line, and values greater than $0.05$ above this line.
Note that points $(p_a, p_0)$ with $p_0 < \frac{p_a}{2\widetilde{Q}} = \frac{p_a}{30}$ cannot belong to the useful region, since from Proposition~\ref{prop:lower-bounds} we have $P_e^{(\infty)}\geq \frac{p_a}{2\widetilde{Q}} > p_0$. Moreover, we note that the bottom border of the useful region (solid black curve) is virtually identical to, but slightly above, the line defined by $p_0 = \frac{p_a}{2\widetilde{Q}}$.
\subsubsection{Optimization of the quantization map}
In this section we show that the decoder performance can be significantly improved by using an appropriate choice of the channel scale factor $\mu$. Figure~\ref{fig:csf_optimisation} shows the threshold values for the noiseless and several noisy decoders with channel scale factors $\mu\in\{1,2,\dots,7\}$. For the noisy decoders, the threshold values are computed for a target error probability $\eta = 10^{-5}$ (see Equation~(\ref{eq:eta_threshold})).
\begin{figure}[!bht]
\centering
\includegraphics[width=90mm]{./fig_noisy_ms/csf_optimisation}
\caption[Threshold values of noiseless and noisy MS decoders with various channel scale factors]{Threshold values of noiseless and noisy MS decoders with various channel scale factors (for noisy decoders, threshold values correspond to a target error probability $\eta = 10^{-5}$)}
\label{fig:csf_optimisation}
\end{figure}
The solid black curve in Figure~\ref{fig:csf_optimisation} correspond to the noiseless decoder. The solid red curve and the dotted blue curve correspond to the MS decoder with sign-preserving noisy adder and full-depth noisy adder, respectively. The adder error probability is $p_a = 10^{-4}$ for the sign-preserving noisy adder, and $p_a = 10^{-5}$ for the full-depth adder\footnote{Note that according to Proposition~\ref{prop:lower-bounds}, a necessary condition to achieve a target error probability $P_{e}^{(\infty)} \leq \eta = 10^{-5}$ is $p_a\leq 2\widetilde{Q}\eta = 3\times 10^{-4}$ for the signed-preserving adder, and $p_a\leq 2\eta\frac{2\widetilde{Q}+1}{2\widetilde{Q}} = 2.07\times 10^{-5}$ for the full-depth adder.}. The two curves are superimposed for $1\leq \mu \leq 6$, and differ only for $\mu=7$. The corresponding threshold values are equal to those obtained in the noiseless case for $\mu\in\{2, 4, 6\}$. For $\mu\in\{1, 3, 5\}$, the MS decoders with noisy-adders exhibit better thresholds than the noiseless decoder. This is due to the fact that the
messages alphabet ${\cal M}$ is underused by the noiseless decoder, since all the exchanged messages are necessarily odd (recall that all variable-nodes are of degree $d_v=3$). For the MS decoders with noisy adders, the noise present in the adders leads to a more efficient use of the messages alphabet, which allows the decoder
to escape from fixed-point attractors and hence results in better thresholds (Section \ref{subsec:ms_q1}).
Figure~\ref{fig:csf_optimisation} also shows a curve corresponding to the MS decoder with a noisy comparator having $p_{\text{\it c}} = 0.005$, and two curves for the MS decoder with noisy {\sc xor}-operators, having respectively $p_{\text{\it x}} = 2\times 10^{-4}$ and $p_{\text{\it x}} = 3\times 10^{-4}$.
Concerning the noisy {\sc xor}-operator, it can be seen that the threshold values corresponding to $p_{\text{\it x}} = 2\times 10^{-4}$ are very close to those obtained in the noiseless case, except for $\mu = 7$ (the same holds for values $p_{\text{\it x}} < 2\times 10^{-4}$). However, a significant degradation of the threshold can be observed when slightly increasing the {\sc xor} error probability to $p_{\text{\it x}} = 3\times 10^{-4}$. Moreover, although not shown in the figure, it is worth mentioning that for $p_{\text{\it x}} \geq 5\times 10^{-4}$, the target error probability $\eta = 10^{-5}$ can no longer be reached (thus, all threshold values are equal to zero).
Finally, we note that except for the noisy {\sc xor}-operator with $p_{\text{\it x}} = 3\times 10^{-4}$, the best choice of the channel scale factor is $\mu=6$. For the noisy {\sc xor}-operator with $p_{\text{\it x}} = 3\times 10^{-4}$, the best choice of the channel scale factor is $\mu=3$. This is rather surprising, as in this case the messages alphabet is underused by the decoder: all the exchanged messages are odd, and the fact that the {\sc xor}-operator is noisy does not change their parity.
\medskip\noindent{\bf Assumption:} In the following sections, we will investigate the impact of the noisy adder, comparator and {\sc xor}-operator on the MS decoder performance, assuming that the channel scale factor is $\mu=6$.
\subsubsection{Study of the impact of the noisy adder (quantization map $\mathbf{q}_6$)}
In order to evaluate the impact of the noisy adder on the MS decoder performance, the useful region and the $\eta$-threshold regions have been computed, assuming that only the adders within the VN-processing step are noisy ($p_a > 0$), while the CN-processing step is noiseless ($p_{\text{\it x}} = p_c = 0$). This regions are represented in Figure~\ref{fig:bsc_45_eta_regions_add}, for both sign-preserving and full-depth noisy adder models.
The useful region is delimited by the solid black curve. The vertical lines delimit the $\eta$-threshold regions, for $\eta = 10^{-3}, 10^{-4}, 10^{-5}, 10^{-6}$ (from right to the left).
Note that unlike the case $\mu = 1$ (Section~\ref{subsec:ms_q1}), there is no non-convergence region when the channel scale factor is set to $\mu=6$. Hence, the border of the useful region corresponds to points $(p_a, p_0)$ for which $P_{e}^{(\infty)} = p_0$. However, it can be observed that there is still a {\em discontinuity line} (dashed red curve) inside the useful region. This discontinuity line does not hide a periodic (non-convergent) behavior, but it is due to the occurrence of an {\em early plateau phenomenon} in the convergence of $(P_{e}^{(\ell)})_\ell$. This phenomenon is illustrated in Figure~\ref{fig:bsc_45_plateau}, where the error probability $(P_{e}^{(\ell)})_\ell$ is plotted as a function of the iteration number $\ell$, for the two points A and B from Figure~\ref{subfig:bsc_45_eta_regions_add_sp}. For point A, it can be observed that the error probability $P_{e}^{(\ell)}$ reaches a first plateau for $\ell \approx 50$, then drops to $3.33\times 10^{-6}$ for $\ell \geq 250$. For point B, $P_{e}^{(\ell)}$ behaves in a similar manner during the first iterations, but it does not decrease below the plateau value as $\ell$ goes to infinity. Although we have no analytic proof of this fact, it was numerically verified for $\ell \leq 5\times 10^{5}$.
\begin{figure}[!thb]
\centering
\subfigure[Sign-preserving noisy adder]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/bsc_45_eta_regions_add_sp}\label{subfig:bsc_45_eta_regions_add_sp}}%
\subfigure[Full-depth noisy adder]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/bsc_45_eta_regions_add_fd}\label{subfig:bsc_45_eta_regions_add_fd}}%
\caption{Useful and $\eta$-threshold regions of the MS decoder with noisy adder}
\label{fig:bsc_45_eta_regions_add}
\end{figure}
\begin{figure}[!thb]
\centering
\subfigure[Point $A(p_0 = 0.0770, p_a = 10^{-4})$]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/bsc_45_plateau_ptA}\label{subfig:bsc_45_plateau_ptA}}%
\subfigure[Point $B(p_0 = 0.0772, p_a = 10^{-4})$]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/bsc_45_plateau_ptB}\label{subfig:bsc_45_plateau_ptB}}%
\caption[Illustration of the early plateau phenomenon]{Illustration of the early plateau phenomenon (points A and B from Figure~\ref{subfig:bsc_45_eta_regions_add_sp})}
\label{fig:bsc_45_plateau}
\end{figure}
\begin{figure}[!htb]
\centering
\subfigure[$p_a = 0$ (noiseless decoder )]{\includegraphics[width=.32\linewidth]{./fig_noisy_ms/pinf_noiselessMS_mu6}\label{subfig:pinf_noiselessMS_mu6}}%
\subfigure[$p_a = 10^{-4}$]{\includegraphics[width=.33\linewidth]{./fig_noisy_ms/pinf_noisy_adder_mu6}\label{subfig:pinf_noisy_adder_mu6}}%
\subfigure[$p_a = 0.05$]{\includegraphics[width=.32\linewidth]{./fig_noisy_ms/pinf_noisy_adder_mu6_sec}\label{subfig:pinf_noisy_adder_mu6_sec}}%
\caption{Asymptotic error probability $P_e^{(\infty)}$ as a function of $p_0$; noiseless and noisy MS decoder with sign-preserving noisy adder}
\label{fig:pinf_mu6}
\end{figure}
In Figure~\ref{fig:pinf_mu6}, we plotted the asymptotic error probability $P_e^{(\infty)}$ as a function of $p_0$, for the noiseless decoder ($p_a = 0$), and for the sign-preserving noisy adder with error probability values $p_a = 10^{-4}$ and $p_a = 0.05$. In each plot we have also represented two points $p_0^{(\text{U})}$ and $p_0^{(\text{FT})}$, corresponding respectively to the values of $p_0$ on the upper-border of the useful region, and on the discontinuity line. Hence, $p_0^{(\text{FT})}$ coincides with the classical threshold of the MS decoder in the noiseless case, and it is equal to the functional threshold defined in Section~\ref{subsec:funct_threshold} in case of noisy decoders. In the following, the
sub-region of the useful region located below the discontinuity line will be referred to as the \textbf{\em functional region}. Within this region, if the adder error probability is small enough, it can be observed that:
\noindent\parbox[t]{6mm}{$(a)$}\parbox[t]{\linewidth-6mm}{For the sign-preserving adder: $P_{e}^{(\infty)} \approx \frac{p_a}{30}$, for $p_a \lessapprox 3\times 10^{-2}$, which corresponds to the value given by the lower-bound ($\frac{1}{2\widetilde{Q}}p_a = \frac{1}{30} p_a$) from Proposition~\ref{prop:lower-bounds}.}
\noindent\parbox[t]{6mm}{$(b)$}\parbox[t]{\linewidth-6mm}{For the full-depth adder: $P_{e}^{(\infty)} \approx 1.17 p_a$, for $p_a \lessapprox 10^{-3}$, which is about twice higher than the value given by the lower-bound ($\frac{1}{2}p_{\text{\it a}} + \frac{1}{4\widetilde{Q}}p_{\text{\it a}} = 0.52 p_a$) from Proposition~\ref{prop:lower-bounds}.}
Finally, we note that by protecting the sign of the noisy adder, the useful region is expanded by a factor of roughly $2\widetilde{Q}$, representing an exponential improvement with respect to the number of bits of the adder (see also the discussion following the proof of Proposition ~\ref{prop:lower-bounds}).
\subsubsection{Study of the impact of the noisy XOR-operator (quantization map $\mathbf{q}_6$)}
The useful region and the $\eta$-threshold regions of the decoder, assuming that only the {\sc xor}-operator used within the CN-processing step is noisy, are plotted in Fig. \ref{fig:bsc_45_eta_regions_xor}. Similar to the noisy-adder case, a discontinuity (functional threshold) line can be observed inside the useful region, which delimits the {\em functional region} of the decoder.
Comparing the $\eta$-threshold regions from Figure~\ref{fig:bsc_45_eta_regions_add} and Figure~\ref{fig:bsc_45_eta_regions_xor}, it can be observed that in order to achieve a target error probability $P_e^{(\infty)} \leq 10^{-6}$, the error probability parameters of the noisy adder and of the noisy {\sc xor}-operator must satisfy:
\begin{itemize}
\item $p_{\text{\it a}} < 1.17 \times 10^{-6}$, for the full-depth noisy-adder;
\item $p_{\text{\it a}} < 3 \times 10^{-5}$, for the sign-preserving noisy-adder;
\item $p_{\text{\it x}} < 7 \times 10^{-5}$, for the noisy {\sc xor}-operator. \\ (moreover, values of $p_{\text{\it x}}$ up to $1.4 \times 10^{-4}$ are tolerable if $p_0$ is sufficiently small)
\end{itemize}
The most stringent requirement concerns the error probability of the full-depth noisy-adder, thus we may consider that it has the most negative impact on the decoder performance. On the other hand, the less stringent requirement concerns the error probability of the noisy {\sc xor}-operator.
Finally, it is worth noting that in practical cases the value of $p_{\text{\it x}}$ should be significantly lower than the value of $p_a$ (given the high number of elementary gates contained in the adder). Moreover, since the {\sc xor}-operators used to compute the signs of CN messages represent only a small part of the decoder, this part of the circuit could be made reliable by using classical fault-tolerant methods, with a limited impact on the overall decoder design.
\begin{figure}[!htb]
\centering
\includegraphics[width=.5\linewidth]{./fig_noisy_ms/bsc_45_eta_regions_xor}
\caption{Useful and $\eta$-threshold regions of the MS decoder with noisy {\sc xor}-operator}
\label{fig:bsc_45_eta_regions_xor}
\end{figure}
\subsubsection{Study of the impact of the noisy comparator (quantization map $\mathbf{q}_6$)}
This section investigates the case when comparators used within the CN-processing step are noisy ($p_{\text{\it c}} > 0$), but $p_{\text{\it a}} = p_{\text{\it x}} = 0$. Contrary to the previous cases, this case exhibits a ``classical'' threshold phenomenon, similar to the noiseless case: for a given $p_{\text{\it c}} > 0$, there exists a $p_0$-threshold value, denoted by $p_0^{\text{(TH)}}$, such that $P_e^{(\infty)} = 0$ for any $p_0 < p_0^{\text{(TH)}}$.
\begin{figure}[!thb]
\centering
\includegraphics[width=.5\linewidth]{./fig_noisy_ms/bsc_45_eta_regions_comp_2}
\caption{Useful region and threshold curve of the MS decoder with noisy comparator}
\label{fig:bsc_45_eta_regions_comp}
\end{figure}
The threshold value $ p_0^{\text{(TH)}}$ is plotted as a function of $p_{\text{\it c}}$ in Figure~\ref{fig:bsc_45_eta_regions_comp}. The functional region of the decoder is located below the threshold curve, and $P_e^{(\infty)} = 0$ for any point within this region. In particular, it can be seen that $P_e^{(\infty)} = 0$ for any $p_0 \lessapprox 0.039$ and any $p_c > 0$.
Although such a threshold phenomenon might seem surprising for a noisy decoder, it can be easily explained.
The idea behind is that in this case the crossover probability of the channel is small enough, so that in the CN-processing step only the sign of check-to-variable messages is important, but not their amplitudes. In other words a decoder that only computes (reliably) the signs of check-node messages and randomly chooses their amplitudes, would be able to perfectly decode the received word.
Finally, we note that the useful region of the decoder extends slightly above the threshold curve: for $p_c$ close to $0$, there exists a small region above the threshold curve, within which $0 < P_e^{(\infty)} <p_0$.
\subsection{Numerical results for the BI-AWGN channel}
For the BI-AWGN, the channel output is given by $y = x + z$, where $x\in\{\pm 1\}$ is the channel input and $z$ is the additive white Gaussian noise with variance $\sigma^2$. Threshold values and useful regions of the decoder will be described in terms of Signal to Noise Ratio (SNR), defined by $\text{SNR} = -10\log_{10}(\sigma^2)$.
\medskip For a given channel scale factor $\mu$, the quantization map $\mathbf{q}_{\mu}$ is defined by $\mathbf{q}_{\mu}(y) = \mathbf{s}_{\cal M}([\mu{\cdot}y])$, where $[\mu{\cdot}y]$ denotes the nearest integer to $\mu{\cdot}y$, and $\mathbf{s}_{\cal M}$ is the saturation map (see also Equation~(\ref{eq:quant-map})).
Similar to the BSC case, the choice of the channel scale factor $\mu$ may significantly impact the decoder performance. Hence, we start first by optimizing the channel scale factor value, and then we investigate the impact of the different noisy components on the decoder performance.
\medskip \noindent{\bf Remark:} For the BI-AWGN channel we denote by $p_0 \stackrel{\text{def}}{=} P_e^{(0)}$ the error probability at iteration $0$, which is, by definition, the probability of the {\em a priori information} $\gamma = \mathbf{q}_{\mu}(y)$ being in error. Hence, $p_0 = \sum_{z=-Q}^{-1}C(z)+\frac{1}{2}C(0)$. Using Equation~(\ref{eq:awgn_c}) it follows that:
\begin{equation}
p_0 = 1 - \frac{1}{2}\left[q\left(\frac{-0.5-\mu}{\mu\sigma}\right) + q\left(\frac{0.5-\mu}{\mu\sigma}\right) \right]
\end{equation}
\subsubsection{Optimization of the quantization map}
The goal of this section is to provide an optimal choice of the channel scale factor $\mu$. Figure~\ref{fig:awgn_csf_optimisation} shows the threshold SNR values for the noiseless and several noisy decoders for channel scale factors $\mu$ varying within the interval $[1,7]$. For the noisy decoders, the threshold values are computed for a target error probability $\eta = 10^{-5}$ (see Equation~(\ref{eq:eta_threshold})).
\begin{figure}[!bht]
\centering
\includegraphics[width=90mm]{./fig_noisy_ms/awgn_csf_optimisation}
\caption[Threshold SNR values of noiseless and noisy decoders with various channel scale factors]{Threshold SNR values of noiseless and noisy decoders with various channel scale factors (for noisy decoders, threshold values correspond to a target error probability $\eta = 10^{-5}$)}
\label{fig:awgn_csf_optimisation}
\end{figure}
The solid black curve in Figure~\ref{fig:awgn_csf_optimisation} correspond to the noiseless decoder. The dashed red curve and the dotted blue curve correspond to the MS decoder with sign-preserving noisy adder and full-depth noisy adder, respectively. The adder error probability is $p_a = 2\times 10^{-4}$ for the sign-preserving noisy adder, and $p_a = 10^{-5}$ for the full-depth adder\footnote{Note that according to Proposition~\ref{prop:lower-bounds}, a necessary condition to achieve a target error probability $P_{e}^{(\infty)} \leq \eta = 10^{-5}$ is $p_a\leq 2\widetilde{Q}\eta = 3\times 10^{-4}$ for the signed-preserving adder, and $p_a\leq 2\eta\frac{2\widetilde{Q}+1}{2\widetilde{Q}} = 2.07\times 10^{-5}$ for the full-depth adder.}. These three curves are virtually indistinguishable.
Figure~\ref{fig:awgn_csf_optimisation} also shows two curves corresponding respectively to the MS decoder with a noisy {\sc xor}-operator ($p_{\text{\it x}} = 2\times 10^{-4}$) and to the MS decoder with a noisy comparator ($p_{\text{\it c}} = 0.005$). Finally, we note that in all cases the best choice of the channel scale factor is $\mu\approx 5.5$.
\medskip \noindent{\bf Assumption:} In the following sections, we will investigate the impact of the noisy adder, comparator and {\sc xor}-operator on the MS decoder performance, assuming that the channel scale factor is $\mu=5.5$.
\subsubsection{Study of the impact of the noisy adder}
Useful and $\eta$-regions of the MS decoder with noisy adders are represented in Figure~\ref{fig:awgn_45_eta_regions_add}, for both sign-preserving and full-depth noisy adder models.
The useful region is delimited by the solid black curve, while vertical lines delimit the $\eta$-threshold regions, for $\eta = 10^{-3}, 10^{-4}, 10^{-5}, 10^{-6}$ (from right to the left). The {\em functional threshold} of the decoder is also displayed by a red dashed curve.
\begin{figure}[!b]
\centering
\subfigure[Sign-preserving noisy adder]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/awgn_45_eta_regions_add_sp}\label{subfig:awgn_45_eta_regions_add_sp}}%
\subfigure[Full-depth noisy adder]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/awgn_45_eta_regions_add_fd}\label{subfig:awgn_45_eta_regions_add_fd}}%
\caption{Useful and $\eta$-threshold regions of the MS decoder with noisy adder ({\sc bi-awgn})}
\label{fig:awgn_45_eta_regions_add}
\end{figure}
\begin{figure}[!thb]
\centering
\subfigure[sign-preserving noisy adder, $p_a = 10^{-4}$]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/awgn_pinf_noisy_adder_sp}\label{subfig:awgn_pinf_noisy_adder_sp}}%
\subfigure[full-depth noisy adder, $p_a = 10^{-4}$]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/awgn_pinf_noisy_adder_fd}\label{subfig:awgn_pinf_noisy_adder_fd}}%
\caption{Asymptotic error probability $P_e^{(\infty)}$ of the MS decoder with noisy-adder as a function of the SNR}
\label{fig:awgn_pinf_noisy_adder}
\end{figure}
Figure~\ref{fig:awgn_pinf_noisy_adder} shows the input and output error probabilities of the decoder ($p_0$ and $P_e^{(\infty)}$) as functions of the SNR value, for the sign-preserving and full-depth noisy adder models with $p_a = 10^{-4}$. The two intersection points between the two curves correspond to the points on the lower and upper borders of the useful region in Figure~\ref{fig:awgn_45_eta_regions_add}, for $p_a = 10^{-4}$. The discontinuity point of the $P_e^{(\infty)}$ curve corresponds to the functional threshold value in Figure~\ref{fig:awgn_45_eta_regions_add}, for $p_a = 10^{-4}$.
\subsubsection{Study of the impact of the noisy XOR-operator and noisy comparator}
The useful region and the $\eta$-threshold regions of the MS decoder, assuming that only the {\sc xor}-operator used within the CN-processing step is noisy, are plotted in Fig. \ref{fig:awgn_45_eta_regions_xor}. The {\em functional threshold} of the decoder is also displayed by a red dashed curve.
The case of a noisy comparator is illustrated in Figure~\ref{fig:awgn_45_eta_regions_comp}. Similar to the BSC channel, this case exhibits a ``classical'' threshold phenomenon: for any SNR value above the functional threshold curve, one has $P_e^{(\infty)} = 0$.
\begin{figure}[!htb]
\centering
\parbox{.49\linewidth}{%
\includegraphics[width=\linewidth]{./fig_noisy_ms/awgn_45_eta_regions_xor}
\caption{Useful and $\eta$-threshold regions of the MS decoder with noisy {\sc xor}-operator ({\sc bi-awgn})}
\label{fig:awgn_45_eta_regions_xor}}\hfill
\parbox{.49\linewidth}{%
\includegraphics[width=\linewidth]{./fig_noisy_ms/awgn_45_eta_regions_comp}
\caption{Useful region and threshold curve of the MS decoder with noisy comparator ({\sc bi-awgn})}
\label{fig:awgn_45_eta_regions_comp}}%
\end{figure}
\endinput
\section{Introduction}\label{sec:introduction} %
\input{asymptotic_ms_noisy_hw}
\input{finite_length_ms_noisy_hw}
\section{Conclusion}\label{sec:noisy_ms_conslusion}
This paper investigated the asymptotic and finite length behavior of the noisy MS decoder. We demonstrated the impact of the channel scale factor on the decoder performance, both for the noiseless and for the noisy decoder. We also highlighted the fact that an inappropriate choice the channel scale factor may lead to an {\em unconventional} behavior, in the sense that the noise introduce by the device may actually result in an increased correction capacity with respect to the noiseless decoder. We analyzed the asymptotic performance of the noisy MS decoder in terms of useful regions and target-BER thresholds, and further revealed the existence of a different threshold phenomenon, which was referred to as functional threshold. Finally, we also corroborated the asymptotic analysis through finite-length simulations, and highlighted the excellent performance of the noisy SCMS decoder, which provides virtually the same performance as the noiseless decoder, for a wide range of values of the hardware noise parameters.
\bibliographystyle{IEEEtran}
\section{Finite Length Performance of Min-Sum based decoders}\label{sec:finite_length_perf}
The goal of this section is twofold:
\begin{description}
\item[(1)] To corroborate the asymptotic analysis through finite-length simulations;
\item[(2)] To investigate ways of increasing the robustness of the MS decoder to hardware noise.
\end{description}
\noindent {\bf Assumption:} Unless otherwise stated, the $(3, 6)$-regular LDPC code with length $N = 1008$ bits from \cite{mackaywebsite} will be used for finite length simulations throughout this section.
\subsection{Practical implementation and early stopping criterion}
First of all, we note that the practical implementation of the noisy MS decoder differs slightly from the one presented in Algorithm~\ref{alg:noisy_ms}:
\begin{itemize}
\item The order of the {\bf VN-processing} and {\bf AP-update} steps is inverted;
\item The variable-to-check node messages are computed by subtracting the incoming check-to-variable message from the corresponding a posteriori information value:
\fbox{\begin{minipage}{.98\linewidth}
{\bf for all} $n=1,\dots,N$ {\bf do} \hfill $\vartriangleright$ {\bf AP-update}
\hspace*{5mm} $\tilde{\gamma}_n = {\color{red}\mathbf{a}_{\mbox{\scriptsize pr}}}\left(\{\gamma_n\} \cup
\{\beta_{m,n}\}_{{m} \in{\cal H}(n)}\right)\text{;}$\\
{\bf for all} $n=1,\dots,N$ and $m\in{\cal H}(n)$ {\bf do} \hfill $\vartriangleright$ {\bf VN-processing}
\hspace*{5mm}$\begin{array}{@{}r@{\ }c@{\ }l}
\alpha_{m,n} & = & {\color{red}\mathbf{a}_{\mbox{\scriptsize pr}}}\left(\tilde{\gamma}_n, -\beta_{m,n}\right)\text{;} \\
\alpha_{m,n} & = & \mathbf{s}_{\cal M}\left(\alpha_{m,n} \right)\text{;}
\end{array}$
\end{minipage}}
\end{itemize}
For floating-point noiseless decoders, the two ways of computing the variable-to-check messages are completely equivalent. However, this equivalence does not hold anymore for finite-precision (noisy or noiseless) decoders, because of saturation effects and, in case of noisy decoders, of probabilistic computations. We note that the practical implementation might result in a degradation of the decoder performance compared to the ``Density-Evolution like'' implementation (Algorithm~\ref{alg:noisy_ms}), since each variable-to-check node message {\em encompasses} $d_v+1$ additions ($d_v$ additions to compute $\tilde{\gamma}_n $ and one subtraction).
Finally, it is worth noting that the density-evolution analysis cannot be applied to the practical implementation, due to the fact that in the VN-processing step, the computation of variable-to-check messages $\alpha_{m,n} = {\color{red}\mathbf{a}_{\mbox{\scriptsize pr}}}\left(\{\gamma_n\}, -\beta_{m,n}\right)$ involves two correlated variables, namely $\gamma_n$ and $\beta_{m,n}$.
\subsubsection{Early stopping criterion (syndrome check)}
As described in Algorithm~\ref{alg:noisy_ms}, each decoding iteration also comprises a {\em hard decision} step, in which each transmitted bit is estimated according to the sign of the a posteriori information, and a {\em syndrome check} step, in which the syndrome of the estimated word is computed.
Both steps are assumed to be {\em noiseless}, and the syndrome check step acts as an {\em early stopping criterion}: the decoder stops when whether the syndrome is $+1$ (the estimated word is a codeword) or a maximum number of iterations is reached.
We note however that the syndrome check step is optional and, if missing, the decoder stops when the maximum number of iterations is reached.
\medskip \noindent {\bf Remark:} The reason why we stress the difference between the MS decoder with and without the syndrome check step is because, as we will see shortly, the {\em noiseless} early stopping criterion may significantly improve the bit error rate performance of the {\em noisy} decoder in the error floor region.
\medskip \noindent {\bf Assumptions:}
\begin{itemize}
\item Unless otherwise stated, the MS decoder is assumed to implement the {\em noiseless} stopping criterion (syndrome check step).
\item The maximum number of decoding iterations is fixed to 100 throughout this section.
\end{itemize}
\subsection{Corroboration of the asymptotic analysis through finite-length simulations}
We start by analyzing the finite-length decoder performance over the BSC channel. Figure~\ref{fig:bsc_flen_mu1_mu6} shows the bit error rate (BER) performance of the finite-precision MS decoder (both noiseless and noisy) with various channel scale factors. For comparison purposes, we also included the BER performance of the Belief-Propagation decoder (solid black curve, no markers) and of the infinite-precision MS decoder (dashed blue curve, no markers).
It can be observed that the worst performance is achieved by the infinite-precision MS decoder~(!) and the finite-precision noiseless MS decoder with channel scale factor $\mu = 1$ (both curves are virtually indistinguishable). The BER performance of the latter improves significantly when using a sign-preserving noisy adder with error probability $p_a = 0.001$ (dashed red curve with empty circles).
For a channel scale factor $\mu = 6$, both noiseless and noisy decoders have almost the same performance (solid and dashed green curves, with triangular markers). Remarkably, the achieved BER is very close to the one achieved by the Belief-Propagation decoder!
These results corroborate the asymptotic analysis from Section~\ref{subsec:num_results_bsc} concerning the channel scale factor optimization.
\begin{figure}[!thb]
\centering
\includegraphics[width=90mm]{./fig_noisy_ms/bsc_flen_mu1_mu6}
\caption{BER performance of noiseless and noisy MS decoders with various channel scale factors}
\label{fig:bsc_flen_mu1_mu6}
\end{figure}
\subsubsection{Error floor performance}
Surprisingly, the BER curves of the noisy decoders from Figure~\ref{fig:bsc_flen_mu1_mu6} do not show any error floor down to $10^{-7}$. However, according to Proposition~\ref{prop:lower-bounds}, the decoding error probability should be lower-bounded by $P_e^{(\ell)} \geq \frac{1}{2\widetilde{Q}}p_a = 3.33 \times 10^{-5}$ (see also
the $\eta$-threshold regions in Figure~\ref{subfig:bsc_45_eta_regions_add_sp}).
The fact that the observed decoding error probability may decrease below the above lower-bound is due to the early stopping criterion (syndrome check step) implemented within the MS decoder. Indeed, as we observed in the previous section, the above lower-bound is tight, when $\ell$ (the iteration number) is sufficiently large. Therefore, as the iteration number increases, the expected number of erroneous bits gets closer and closer to $\frac{1}{2\widetilde{Q}}p_aN = 0.034$, and the probability of not having any erroneous bit within one iteration approaches $\left(1 - \frac{1}{2\widetilde{Q}}p_a\right)^N = 0.967$. As the decoder performs more and more iterations, it will eventually reach an error free iteration. The absence of errors is at once detected by the noiseless syndrome check step, and the decoder stops.
\begin{figure}[!thb]
\centering
\parbox{.49\linewidth}{%
\includegraphics[width=\linewidth]{./fig_noisy_ms/bsc_ber_variousN}
\caption[BER performance with and without early stopping criterion]{BER performance with and without early stopping criterion (MS decoder with sign-preserving noisy adder, $p_a = 0.001$)}
\label{fig:bsc_ber_variousN}}\hfill
\parbox{.49\linewidth}{%
\includegraphics[width=\linewidth]{./fig_noisy_ms/bsc_ave_iter_nb_variousN}
\caption[Average number of decoding iterations with early stopping criterion]{Average number of decoding iterations with early stopping criterion (MS decoder with sign-preserving noisy adder, $p_a = 0.001$)}
\label{fig:bsc_ave_iter_nb_variousN}}%
\end{figure}
\medskip To illustrate this behavior, we plotted the Figure~\ref{fig:bsc_ber_variousN} the BER performance of the noisy MS decoder, with and without early stopping criterion. The noisy MS decoder comprises a sign-preserving noisy adder with $p_a = 0.001$, while the comparator and the {\sc xor}-operator are assumed to be noiseless ($p_{\text{\it c}} = p_{\text{\it x}} = 0$). Two codes are simulated, the first with length $N = 1008$ bits, and the second with length $N = 10000$ bits.
In case that the noiseless early stopping criterion is implemented (solid curves), it can be seen that none of the BER curves show any error floor down to $10^{-8}$. However, if the early stopping criterion is not implemented (dashed curves), corresponding BER curves exhibit an error floor at $\approx 3.33\times 10^{-5}$, as predicted by Proposition~\ref{prop:lower-bounds}.
In Figure~\ref{fig:bsc_ave_iter_nb_variousN} we plotted the average number of decoding iterations in case that the early stopping criterion is implemented. It can be seen that the average number of decoding iterations decreases with the channel crossover probability $p_0$, or equivalently, with the achieved bit error rate.
However, for a fixed BER -- say BER $= 10^{-6}$, achieved either at $p_0 \approx 0.04$ for the code with $N = 1008$, or at $p_0 \approx 0.063$ for the code with $N = 10000$ -- the average number of iterations is about $8$ for the first code and about $21$ for the second.
Note that in case the early stopping criterion is not implemented, both codes have nearly the same performance
for the above $p_0$ values. Thus, when the early stopping criterion is implemented, the decoder needs to perform more iterations to eventually reach an error free iteration when $N = 10000$, which explains the increased average number of decoding iterations.
\subsubsection{Further results on the finite-length performance}
In this section we investigate the finite-length performance when all the MS components (adder, comparator, and {\sc xor}-operator) are noisy. In order to reduce the number of simulations, we assume that $p_a = p_c \geq p_x$. Concerning the noisy adder, we evaluate the BER performance for both the sign-preserving and the full-depth error models. Simulation results are presented in Figure~\ref{fig:ber_perf_noisy_adder}. The error probability of the {\sc xor}-operator is $p_{\text{\it x}} = 0.0001$ in sub-figures~\ref{subfig:bsc_ber_add_sp_px0001} and~\ref{subfig:bsc_ber_add_fd_px0001}, and $p_{\text{\it x}} = 0.001$ in sub-figures~\ref{subfig:bsc_ber_add_sp_px001} and~\ref{subfig:bsc_ber_add_fd_px001}. The noisy adder is sign-preserving in sub-figures~\ref{subfig:bsc_ber_add_sp_px0001} and~\ref{subfig:bsc_ber_add_sp_px001}, and full-depth in sub-figures~\ref{subfig:bsc_ber_add_fd_px0001} and~\ref{subfig:bsc_ber_add_fd_px001}.
In case the noisy-adder is sign-preserving, it can be seen that the MS decoder can provide reliable error protection for all the noise parameters that have been simulated. Of course, depending on the error probability parameters of the noisy components, there is a more or less important degradation of the achieved BER with respect to the noiseless case. But in all cases the noisy decoder can achieve a BER less than $10^{-7}$. This is no longer true for the full-depth noisy adder: it can be seen that for $p_{\text{\it c}} = p_{\text{\it a}} \geq 0.005$, the noisy decoder cannot achieve bit error rates below $10^{-2}$.
\clearpage
\begin{figure}[!thb]
\centering
\subfigure[sign-preserving noisy adder, $p_{\text{\it c}} = p_{\text{\it a}}$, $p_{\text{\it x}} = 0.0001$]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/bsc_ber_add_sp_px0001}\label{subfig:bsc_ber_add_sp_px0001}}%
\subfigure[full-depth noisy adder, $p_{\text{\it c}} = p_{\text{\it a}}$, $p_{\text{\it x}} = 0.0001$]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/bsc_ber_add_fd_px0001}\label{subfig:bsc_ber_add_fd_px0001}}%
\subfigure[sign-preserving noisy adder, $p_{\text{\it c}} = p_{\text{\it a}}$, $p_{\text{\it x}} = 0.001$]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/bsc_ber_add_sp_px001}\label{subfig:bsc_ber_add_sp_px001}}%
\subfigure[full-depth noisy adder, $p_{\text{\it c}} = p_{\text{\it a}}$, $p_{\text{\it x}} = 0.001$]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/bsc_ber_add_fd_px001}\label{subfig:bsc_ber_add_fd_px001}}%
\caption{BER performance of the noisy MS decoder with various noise parameters}
\label{fig:ber_perf_noisy_adder}
\end{figure}
\subsection{Noisy Self-Corrected Min-Sum decoder}
In this section we investigate the finite-length performance of the Self-Corrected Min-Sum (SCMS) decoder \cite{savin2008self}. The objective is to determine if a correction circuit ``plugged into'' the noisy MS decoder can improve the robustness of the decoder to hardware noise.
The specificity of the SCMS decoder is to {\em erase} ({\em i.e.} set to zero) any variable-to-check message that
changes its sign between two consecutive iterations. However, in order to avoid erasures propagation, a message cannot be erased if it has also been erased at the previous iteration. Hence, the SCMS decoder performs the same computations as the noisy MS, except that the {\bf VN processing} step further includes a {\em correction step}, as follows\footnote{Superscript $(\ell)$ used to denote the iteration number}:
\medskip\noindent\fbox{\begin{minipage}{.98\linewidth}
{\bf for all} $n=1,\dots,N$ and $m\in{\cal H}(n)$ {\bf do} \hfill $\vartriangleright$ {\bf VN-processing}
\vspace*{2mm}\hspace*{5mm}$\begin{array}{@{}r@{\ }c@{\ }l}
\alpha_{m,n}^{(\ell)} & = & \mathbf{s}_{\cal M}\left({\color{red}\mathbf{a}_{\mbox{\scriptsize pr}}}\left(\tilde{\gamma}_n^{(\ell)}, -\beta_{m,n}^{(\ell)}\right)\right)\text{;}
\end{array}$
\vspace*{2mm}\hspace*{5mm}$\text{\bf if } \
{\mbox{\rm sgn}}\left(\alpha_{m,n}^{(\ell)}\right) \neq {\mbox{\rm sgn}}\left(\alpha_{m,n}^{(\ell-1)}\right)
\text{ and } \alpha_{m,n}^{(\ell-1)} \neq 0 \\
\hspace*{15mm} \alpha_{m,n}^{(\ell)} = 0\,; \\
\hspace*{5mm}\text{\bf end}$
\end{minipage}}
\bigskip The body enclosed between the {\bf if} condition and the matching {\bf end} is referred to as the {\em correction step}. In practical implementations, one needs to store the signs of the variable-to-check node messages and to keep a record of messages that have been erased by the self-correction step. We use the following notation:
\begin{itemize}
\item $s_{m,n}^{(\ell)} = \text{sgn}\left(\alpha_{m,n}^{(\ell)}\right)$, the sign of the message $\alpha_{m,n}^{(\ell)}$;
\item $e_{m,n}^{(\ell)}\in\{0,1\}$, with $e_{m,n}^{(\ell)} = 1$ if and only if the corresponding variable-to-check message has been erased at iteration $\ell$; for $\ell = 0$, these values are all initialized as zero.
\item $\text{\sc scu}(s_1, s_2, e) \stackrel{\text{def}}{=} (s_1 \oplus s_2) \otimes (1 \oplus e)$, for any $s_1, s_2, e\in\{0, 1\}$, where $\oplus$ denotes the {\sc xor} operation (sum modulo $2$) and $\otimes$ denotes the {\sc and} operation (product). Clearly $\text{\sc scu}(s_1, s_2, e) = 1$ if and only if $s_1\neq s_2$ and $e = 0$.
\end{itemize}
\noindent Therefore, the VN-processing step of the SCMS decoder can be rewritten as follows:
\medskip\noindent\fbox{\begin{minipage}{.98\linewidth}
{\bf for all} $n=1,\dots,N$ and $m\in{\cal H}(n)$ {\bf do} \hfill $\vartriangleright$ {\bf VN-processing}
\vspace*{2mm}\hspace*{5mm}$\begin{array}{@{}r@{\ }c@{\ }l}
\alpha_{m,n}^{(\ell)} & = & \mathbf{s}_{\cal M}\left({\color{red}\mathbf{a}_{\mbox{\scriptsize pr}}}\left(\tilde{\gamma}_n^{(\ell)}, -\beta_{m,n}^{(\ell)}\right)\right)\text{;} \hbox{\protect\raisebox{0mm}[0mm][4mm]{}}\\
e_{m,n}^{(\ell)} & = & \text{\sc scu}\left(s_{m,n}^{(\ell)}, s_{m,n}^{(\ell-1)},
e_{m,n}^{(\ell-1)}\right)\text{;}
\end{array}$
\vspace*{2mm}\hspace*{5mm} $\text{\bf if } \ e_{m,n}^{(\ell)} = 1 \ {\bf then }\ \
\alpha_{m,n}^{(\ell)} = 0\,; \ \ \text{\bf end}$
\end{minipage}}
\bigskip This reformulation of the VN-processing step allows defining a {\em noisy self-correction} step, by injecting errors in the output of the {\sc scu} operator. The noisy {\sc scu} operator with error probability $p_{\text{scu}}$ is defined by:
\begin{equation}\label{eq:noisy-scu-def}
\text{\sc scu}_{\mbox{\scriptsize pr}}(s_1, s_2, e) = \left\{\begin{array}{cl}
\text{\sc scu}(s_1, s_2, e), & \mbox{with probability } 1-p_{\text{scu}}\\
1 - \text{\sc scu}(s_1, s_2, e), & \mbox{with probability } p_{\text{scu}}
\end{array}\right.
\end{equation}
This error model captures the effect of the noisy logic or of the noisy storage of $s_{m,n}$ and $e_{m,n}$ values on the {\sc scu} operator. The SCMS decoder with noisy self-correction step is detailed in Algorithm~\ref{alg:noisy_scms}.
\begin{figure}[!b]
\centering
\vspace*{-5mm}
\subfigure[BSC channel ($\mu = 6$)]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/bsc_ber_ms_scms} \label{subfig:bsc_ber_ms_scms}}%
\subfigure[BI-AWGN channel ($\mu = 5.5$)]{\includegraphics[width=.49\linewidth]{./fig_noisy_ms/awgn_ber_ms_scms}
\label{subfig:awgn_ber_ms_scms}}
\vspace*{-3mm}\caption{BER performance comparison between noisy MS and noisy SCMS decoders}
\label{fig:noisy_ms_vs_scms}
\end{figure}
\input{algo_noisy_scms}
\smallskip The finite length performance of the noisy SCMS decoder is presented in Figure~\ref{fig:noisy_ms_vs_scms}, for both BSC and BI-AWGN channels. For comparison purposes, Figure~\ref{fig:noisy_ms_vs_scms} also shows the performance of the noisy MS decoder. The parameters of the different noisy components are as follows:
\medskip \noindent {\bf [P1]} sign-preserving adder with $p_{\text{\it a}} = 0.01$, $p_{\text{\it c}} = 0.01$, $p_{\text{\it x}} = p_{\text{scu}} = 0.001$ (red curves, diamond markers);
\smallskip \noindent {\bf [P2]} full-depth adder with $p_{\text{\it a}} = 0.001$, $p_{\text{\it c}} = 0.001$, $p_{\text{\it x}} = p_{\text{scu}} = 0.001$ (blue curves, circle markers).
Solid and dashed curves correspond respectively to the MS and SCMS performance. While the hardware noise alters the performance of the MS decoder, it can be seen that the noisy SCMS decoder exhibits very good performance, very close to that of the noiseless decoder. Therefore, one can think of the self-correction circuit
as a {\em noisy patch} applied to the noisy MS decoder, in order to improve its robustness to hardware noise. The robustness of the SCMS decoder to hardware noise is explained by the fact that it has an intrinsic capability to detect unreliable messages, and discards them from the iterative decoding process \cite{savin2008self}.
|
\section{Introduction}
We are living in an era enriched with many experimental breakthroughs and
results especially in the area of astro-particle physics and cosmology.
The most recent one is the identification of a weak line at $E\sim 3.5
~\rm{keV}$ in the X-ray spectra of the Andromeda galaxy and many other galaxy
clusters including the Perseus galaxy cluster, observed by XMM-Newton X-ray
Space observatory\cite{Bulbul, Boyarsky}. The observed flux and the best fit
energy peak are at
\begin{eqnarray}
\Phi_{\gamma}&=&4\pm 0.8\times 10^{-6}
~\rm{photons~cm^{-2} sec^{-1}},\nonumber \\
E_{\gamma}&=&3.57\pm 0.02 ~\rm{keV}.
\end{eqnarray}
Since atomic transitions in thermal plasma cannot account for this energy,
therefore the concept of a dark matter, providing the possible explanation
regarding the appearance of this photon line becomes extremely important.
This result can be explained by a sterile neutrino \cite{Ishida, Abazajian,
Modak, Cline, Barry, Robinson}, axion or axion like warm dark matter
\cite{Higaki,Jaeckel,Lee,Conlon}, axino \cite{Kong,Choi,Liew}, excited dark
matter \cite{Okada,Okada-1}, gravitino \cite{Bomark,Demidov}
and keV scale LSP \cite{Kolda} as decaying dark matter. Other interesting
scenarios with an annihilating scalar dark matter \cite{Dudas}, decaying
Majoron \cite{Sinha} and a keV scale dark gaugino \cite{Kang} have also
been considered in this context. In this work we consider sterile neutrino in a
$U(1)_{R^-}$lepton number model, which could provide a possible explanation for
the emergence of the photon line. The observed flux and the peak of the
energy readily translates to an active-sterile mixing in the range $2.2
\times 10^{-11}<\sin^2 2\theta_{14} < 2 \times 10^{-10}$ and the mass of the
sterile neutrino dark matter $M^R_N = 7.06 \pm 0.05$ keV \cite{Boyarsky}.
On the other hand, in high energy collider frontier two CERN based
experiments ATLAS and CMS have confirmed the existence of a neutral
elementary scalar boson of nature, with mass around 125 GeV
\cite{Aad,Chatrchyan}. Nevertheless, more analysis is required to confirm it as
the Standard Model (SM) Higgs boson. In order to explain the mass of this
scalar boson in a natural way, to address the question of nonzero neutrino mass
and mixing and to provide a candidate for dark matter, many beyond standard
model (BSM) theories have been pursued for quite some time and supersymmetry
remains one of the most celebrated ones as of now. However, supersymmetric
particle searches by ATLAS and CMS experiments for pp collision at center of
mass energy 7 and 8 TeV, have observed no significant excess
\cite{Chatrchyan-1, Aad-1} over the standard model background. This has
put very stringent lower limits on the superpartner masses.
In the light of this present situation, $U(1)_{R^-}$symmetric models with Dirac
gauginos are well motivated because they can relax the strong bounds on the
superpartner masses, explain the 125 GeV Higgs boson mass, provide non-zero
neutrino mass at the tree as well as at the one loop level and can also
accommodate a suitable dark matter candidate. Various aspects of different
R-symmetric models have been studied and can be found in the literature
\cite{Fayet,Hall-1,Hall-2, Nelson,Fox-1,Chacko,Choi-S,Kribs,Benakli,Kumar,Fox,
Benakli-2,Benakli-1,Davies,Davies-1,Gregoire,Greg,Bertuzzo,Goodsell,Riva,Kumar-1,Claudia,
Chakraborty,Dudas-1,Beauchesne,Benakli:2014cia}. In this work, we study a
particular $U(1)_{R^-}$symmetric model where we have identified the R-charges
with lepton numbers in such a way that the lepton numbers of the standard model
fermions correspond to the negative of their R-charges \cite{Kumar-1,Claudia}.
The role of the down-type Higgs is played by the sneutrino since its vacuum
expectation value (vev) is not constrained by the Majorana mass of the neutrino.
The minimal extension of this model by adding a single right handed neutrino
superfield also gives rise to very interesting phenomenological consequences
\cite{Chakraborty}. It generates a tree level Dirac mass for one of the
neutrinos in the R-symmetry preserving scenario. If R-symmetry is broken because
of the presence of a non zero gravitino mass, then for small neutrino Yukawa
coupling, $f\sim\mathcal O(10^{-4})$, the extended neutralino-neutrino mass
matrix provides a sterile neutrino state accompanied by an active neutrino
state. Here we identify the sterile neutrino as the warm dark matter in our
model.
The presence of R-symmetry inhibits gauginos to acquire a Majorana mass.
However, gauginos can have Dirac masses and to introduce the Dirac gaugino
mass, one must consider a singlet chiral superfield $\hat S$, a triplet
$\hat T$ and an octet $\hat O$ living in the adjoint representation of
$U(1)_Y$, $SU(2)_L$ and $SU(3)_C$ respectively. The Dirac gaugino masses
are also coined as `supersoft' mass terms since they do not contribute
to any logarithmic corrections to the scalar masses. The presence of Dirac
gluino also helps to relax the bound on squark masses compared to MSSM and
in addition flavor and CP violating constraints are suppressed in this class
of models \cite{Kribs}.
The plan of the paper is as follows. At first we describe the model in section
II, with appropriate R-charge assignments. In section III we discuss very
briefly, the scalar sector of the model and point out the extra contributions
to the Higgs boson mass, which can arise both at the tree level as well as
at the one loop level. Section IV addresses the issue of R-symmetry breaking
and tree level Majorana masses of the sterile and one of the active neutrinos.
Next in section V the essential features of the sterile neutrino as a keV
warm dark matter candidate are discussed and its production mechanism and the
dominant decay modes relevant to our model are highlighted. In section VI we
briefly present a discussion related to the cosmology of the gravitino in
this model with a few GeV mass and finally, in section VII, we summarise
our results.
\section{$U(1)_{R^-}$lepton number model with a right handed neutrino superfield}
We study a $U(1)_{R^-}$lepton number model, where in addition to the standard
superfields of the MSSM - $\hat H_u$, $\hat H_d$, $\hat Q_i$, $\hat {U_i}^c$,
$\hat {D_i}^c$, $\hat L_i$, $\hat {E_i}^c$, this model includes a right handed
neutrino superfield and a pair of vector-like $SU(2)_L$ doublet superfields
$\hat R_u$ and $\hat R_d$, with opposite hypercharge \cite{Chakraborty}. These two
doublets carry non zero R-charges (The R-charge assignments are given in Table I)
and therefore, to avoid spontaneous R-breaking and the emergence of R-axions, they
do not acquire any non-zero vev and would remain inert. R symmetry
prohibits soft supersymmetry breaking terms like Majorana gaugino masses and
trilinear scalar couplings. However, gauginos can acquire Dirac masses as
mentioned in the introduction. The implications of adding a right-handed
neutrino superfield $\hat N^c$ is discussed later in detail. We would like
to reiterate that the R-charge assignments are such that the lepton number of
the SM fermions are negative of their corresponding R-charges.
\begin{table}[h!]
\begin{center}
\begin{tabular}{|c|ccccccccccccc|}
\hline
\rule{0mm}{5mm}
& $\hat Q_{i}$ & $\hat U_{i}^{c}$ & $\hat D_{i}^{c}$ & $\hat L_{i}$
& $\hat E_{i}^{c}$ & $\hat H_{u}$ & $\hat H_{d}$ & $\hat R_{u}$
& $\hat R_{d}$ & $\hat S$ & $\hat T$ & $\hat O$ & $\hat N^{c}$
\\[0.3em]
\hline
\rule{0mm}{5mm}
$U(1)_{R}$ & 1 & 1 & 1 & 0 & 2 & 0 & 0 & 2 & 2 & 0 & 0 & 0 & 2 \\
[0.3em]
\hline
\end{tabular}
\end{center}
\vspace{-10pt}
\caption{$U(1)_{R}$ charge assignments of the chiral superfields.}
\label{R-charges}
\vspace{-15pt}
\end{table}
The generic superpotential, carrying R-charge of 2 units is
\begin{eqnarray}
W&=&y^{u}_{ij}\hat H_{u}\hat Q_{i}\hat U^{c}_{j}
+\mu_{u}\hat H_{u}\hat R_{d}+f_{i}\hat L_{i}\hat H_{u}\hat N^{c}
+\lambda_{S}\hat S\hat H_{u}\hat R_{d}
+2\lambda_{T}
\hat H_{u}\hat T\hat R_{d}-M_{R}\hat N^{c}\hat S+
\mu_{d}\hat R_{u}\hat H_{d}\nonumber \\
&+&\lambda^\prime_{S}\hat S\hat R_{u}\hat H_{d}
+\frac{1}{2}\lambda_{ijk}
\hat L_{i}\hat L_{j}\hat E^{c}_{k}
+\lambda^{\prime}_{ijk}\hat L_{i}\hat Q_{j}\hat D^{c}_{k}
+2\lambda^\prime_{T}\hat R_{u}\hat T\hat H_{d}
+y^{d}_{ij}\hat H_{d}
\hat Q_{i}\hat D^{c}_{j}+y^{l}_{ij}\hat H_{d}
\hat L_{i}\hat E^{c}_{j} \nonumber \\
&+& \lambda_N {\hat N}^c {\hat H}_u {\hat H}_d.
\label{superpotential}
\end{eqnarray}
Note that a subset ($\lambda$, $\lambda^{\prime}$) of standard R-parity violating
operators are present in the superpotential although the model is $U(1)_R$
conserving (i.e. lepton number conserving). In a somewhat simplistic approach
we have omitted the terms $\hat N^c \hat S\hat S$ and $\hat N^c$ from the
superpotential.
In a realistic model one should also include supersymmetry breaking terms,
such as the gaugino and scalar mass terms. The Dirac gaugino `supersoft' mass
terms are constructed from a spurion superfield $W^{\prime}_{\alpha} =
\lambda_{\alpha}+\theta_{\alpha}D^{\prime}$, if supersymmetry breaking is of
the D-type. The Lagrangian containing the Dirac gaugino masses are
\cite{Benakli-2,Benakli-1}
\begin{eqnarray}
{\cal L}^{\rm Dirac}_{\rm gaugino} &=& \int d^2 \theta
\dfrac{W^\prime_\alpha}{\Lambda}[\sqrt{2} \kappa_1 ~W_{1 \alpha} {\hat S}
+ 2\sqrt{2} \kappa_2 ~{\rm tr}(W_{2\alpha} {\hat T})
+ 2\sqrt{2} \kappa_3 ~{\rm tr}(W_{3\alpha} {\hat O})] + h.c.
\label{dirac-gaugino}
\end{eqnarray}
\vspace {2mm}
This D-term breaking generates Dirac mass for the gauginos, proportional
to $k_i\frac{<D^{\prime}>}{\Lambda}$, where $\Lambda$ denotes the scale of
SUSY breaking mediation. In a similar manner the $U(1)_R$ conserving soft
supersymmetry breaking terms in the scalar sector are generated by the spurion
superfield $\hat X$, defined as $\hat X=x+\theta^2 F_X$. The non-zero
vev of the F-term generates the scalar soft terms as
\begin{eqnarray}
V_{soft}&=& m^{2}_{H_{u}} H_{u}^{\dagger}H_{u}+m^{2}_{R_{u}}
R_{u}^{\dagger}R_{u}+ m^{2}_{H_{d}}H_{d}^{\dagger}H_{d}
+m^{2}_{R_{d}}R_{d}^{\dagger}R_{d}
+m^{2}_{\tilde L_{i}}\tilde L_{i}^{\dagger}\tilde L_{i}\nonumber \\
&+&m^2_{{\tilde R}_i}{{\tilde l}^\dagger_{Ri} {\tilde l}_{Ri}} +
M_{N}^{2}\tilde N^{c\dagger}\tilde N^{c}
+m_{S}^{2} S^{\dagger}S+2m_{T}^{2} {\rm tr}(T^{\dagger}T)
+2 m_O^2 {\rm tr}(O^\dagger O) \nonumber \\
&+& (B\mu H_u H_d + {\rm h.c.})- (b\mu_L^i H_u {\tilde L}_i + {\rm h.c.})
+(t_{S}S+{\rm h.c.})\nonumber \\
&+&\frac{1}{2} b_{S}(S^{2}+{\rm h.c.})
+b_{T} ({\rm tr}(TT) + {\rm h.c.})+B_O({\rm tr}(OO) + {\rm h.c.}).
\label{soft-scalar-terms}
\end{eqnarray}
The presence of the bilinear term $b\mu_L^i H_u\tilde L_i$ in the soft
supersymmetry breaking potential implies all the three left handed sneutrinos
can acquire a non zero vevs ($v_i$). To simplify, we perform a basis rotation as
$\hat L_i=\frac{v_i}{v_a}\hat L_a+e_{ib}\hat L_b$ by which only one of the
sneutrinos acquire a non zero vev ($v_a$) and we choose it to be the electron
sneutrino ($a=1(e)$). We also choose the neutrino Yukawa coupling ($f$) in such a
manner that only $\hat L_a$ couples with $\hat N^c$, the right-handed neutrino
superfield \cite{Chakraborty}. Finally, we choose a very large $\mu_d$ such that
the superfields $\hat H_d$ and $\hat R_u$ gets decoupled, which also implies
that the left handed electron type sneutrino now plays the role of a down type
Higgs field. We would like to emphasise that the model is lepton number
conserving and therefore, the sneutrino vev is not constrained from the Majorana
mass of the neutrinos. This is clearly different from the standard R-parity
violating scenario.
In the mass eigenstate basis (primed superfields) of the down-type quarks and the
charged leptons\footnote{Note that the mass of the charged lepton of flavor $a$
can come from R-symmetry preserving supersymmetry breaking operators \cite{Kumar-1}.}
the superpotential takes the following form \cite{Chakraborty}
\begin{eqnarray}
W&=&y_{ij}^{u}\hat H_{u}\hat Q_{i}\hat U_{j}^{c}+\mu_{u}\hat
H_{u}\hat R_{d}+f\hat L_{a}\hat H_{u}\hat N^{c}+
\lambda_{S}\hat S\hat H_{u}\hat R_{d}
+2\lambda_{T}\hat H_{u}
\hat T\hat R_{d}-M_{R}\hat N^{c}\hat S + W^{\prime},\nonumber \\
\label{final-superpotential}
\end{eqnarray}
and
\begin{eqnarray}
W^{\prime}&=&\sum_{b=2,3} f^l_b {\hat L^{\prime}}_a {\hat L^\prime}_b
{\hat E^{\prime c}}_b + \sum_{k=1,2,3} f^d_k {\hat L^\prime}_a
{\hat Q^\prime}_k {\hat D^{\prime c}}_k
+ \sum_{k=1,2,3} \dfrac{1}{2}
{\tilde \lambda}_{23k}{\hat L^\prime}_2 {\hat L^\prime}_3
{\hat E^{\prime c}}_k \nonumber \\
&+& \sum_{j,k=1,2,3;b=2,3}{\tilde \lambda}^\prime_{bjk}
{\hat L^\prime}_b {\hat Q^\prime}_j {\hat D^{\prime c}}_k.
\label{W-diag}
\end{eqnarray}
In our subsequent analysis we stay in this mass basis
but remove the prime from the fields and make the replacement $\tilde\lambda$,
$\tilde\lambda^{\prime}\rightarrow \lambda$, $\lambda^{\prime}$.
The soft supersymmetry breaking but $U(1)_R$ preserving terms in
the rotated basis are
\begin{eqnarray}
V_{soft}&=& m^{2}_{H_{u}}H_{u}^{\dagger}H_{u}+m^{2}_{R_{d}}
R_{d}^{\dagger}R_{d}+m^{2}_{\tilde L_{a}} \tilde L_{a}^{\dagger}
\tilde L_{a}
+\sum_{b=2,3} m^{2}_{\tilde L_{b}} \tilde L_{b}^{\dagger}
{\tilde L_{b}}+M_{N}^{2}{\tilde N}^{c\dagger} {\tilde N}^{c}
+m^2_{{\tilde R}_i}{{\tilde l}^\dagger_{Ri} {\tilde l}_{Ri}} \nonumber \\
&+&m_{S}^{2} S^{\dagger}S+2m_{T}^{2} {\rm tr}(T^{\dagger}T)
+2m_O^2 {\rm tr}(O^\dagger O)
- (b\mu_L H_u {\tilde L}_a + {\rm h.c.})
+(t_{S}S+{\rm h.c.}) \nonumber \\
&+&\frac{1}{2} b_{S}(S^{2}+{\rm h.c.})
+b_{T} ({\rm tr}(TT) + {\rm h.c.}) +B_O({\rm tr}(OO) + {\rm h.c.}).
\label{final-softsusy-terms}
\end{eqnarray}
In the R-symmetric case, the lightest eigenvalue of the neutralino mass matrix,
written in the basis ($\tilde b^0$, $\tilde w^0$, $\tilde R_d^0$, $N^c$) and
($\tilde S$, $\tilde T^0$, $\tilde H_u^0$, $\nu_e$), provides a tree level
Dirac neutrino mass, which can be written as \cite{Chakraborty}
\begin{eqnarray}
m_{\nu_e}^D&=&\frac{v^3 \sin\beta f g \lambda_T}{\sqrt 2\gamma
M_1^D M_2^D}(M_2^D-M_1^D),
\label{Dirac-neutrino}
\end{eqnarray}
where $M_1^D$, $M_2^D$ stands for Dirac bino and wino masses respectively,
$\gamma = \mu_u + \lambda_S v_S + \lambda_T v_T$, $g$ is the $SU(2)$ gauge
coupling, $\tan\beta = \frac{v_u}{v_a}$, $v\equiv \sqrt{v_u^2+v_a^2} =
\frac{\sqrt 2 M_W}{g}$.
To obtain this particular form in eq.~(\ref{Dirac-neutrino}) we have assumed
certain relations involving the parameters and they are
\begin{eqnarray}
\lambda_T &=& \tan\theta_W\lambda_S, \nonumber \\
M_R &=& \frac{\sqrt 2 f M_1^D \tan\beta}{g\tan\theta_W}.
\label{MR-lambda}
\end{eqnarray}
Therefore, with appropriate choice of parameters one can easily obtain a small
tree level Dirac neutrino mass $\sim 0.1$ eV.
\section{Scalar sector}
In this section we shall mention very briefly about the scalar sector of this
particular model. For a detailed discussion we refer the reader to
\cite{Chakraborty}. The lightest CP even scalar mass matrix, in the basis of
($H_u$, $\tilde \nu$, $S$, $T$), provide the CP even Higgs boson. It is
remarkable that the neutrino Yukawa coupling $f$ renders a tree level correction
to the lightest Higgs boson mass, which we calculate as
\begin{eqnarray}
M_h^2 \leq M_z^2 \cos^2 2\beta + f^2 v^2 \sin^2 2\beta.
\end{eqnarray}
For $f\sim\mathcal O(1)$ and for small $\tan\beta$, the tree level
\footnote{In this paper we shall not explore such a possibility and
concentrate on the region of parameter space where $f\sim \mathcal O(10^{-4})$,
which produces a keV sterile neutrino state.}
Higgs boson mass can satisfy the present observed value, close to $125$ GeV
\cite{Chakraborty}. It is also pertinent to mention that the singlet and the
triplet fields provide very important loop corrections to the Higgs boson mass.
These contributions can be sizable if the singlet and the triplet couplings
$\lambda_S$ and $\lambda_T$ are large. The dominant radiative corrections to
the quartic potential can be written as \cite{Belanger}, $\frac{1}{2}\delta
\lambda_u(|H_u|^2)^2$, $\frac{1}{2}\delta\lambda_\nu(|\tilde\nu_a|^2)^2$ and
$\frac{1}{2}\delta\lambda_3|H_u^0|^2|\tilde\nu_a|^2$, where
\begin{eqnarray}
\delta\lambda_{u}&=& \frac{3 y_{t}^{4}}{16\pi^{2}}
\ln \left(\frac{m_{\tilde t_{1}}m_{\tilde t_{2}}}{m_{t}^2}\right)
+\frac{5\lambda_{T}^{4}}{16\pi^{2}}\ln\left(\frac{m_{T}^{2}}
{v^{2}}\right)
+\frac{\lambda_{S}^{4}}{16\pi^{2}}\ln\left(\frac{m_{S}^2}{v^{2}}\right)
-\frac{1}{16\pi^{2}}\frac{\lambda_{S}^{2}\lambda_{T}^{2}}
{m_{T}^{2}-m_{S}^{2}}\nonumber \\
&&\Big(m_{T}^{2}\Big\{\ln\left(\frac{m_{T}^{2}}{v^{2}}\right)-1\Big\}
-m_{S}^{2}\Big\{\ln \left(\frac{m_{S}^{2}}{v^{2}}\right)-1\Big\}\Big),\nonumber \\
\end{eqnarray}
\begin{eqnarray}
\delta\lambda_{\nu}&=& \frac{3 y_{b}^{4}}{16\pi^{2}}
\ln \left(\frac{m_{\tilde b_{1}}m_{\tilde b_{2}}}{m_{b}^2}\right)
+\frac{5\lambda_{T}^{4}}{16\pi^{2}}\ln\left(\frac{m_{T}^{2}}
{v^{2}}\right)
+\frac{\lambda_{S}^{4}}{16\pi^{2}}
\ln\left(\frac{m_{S}^2}{v^{2}}\right)
- \frac{1}{16\pi^{2}}\frac{\lambda_{S}^{2}\lambda_{T}^{2}}
{m_{T}^{2}-m_{S}^{2}}\nonumber \\
&&\Big(m_{T}^{2}\Big\{\ln\left(\frac{m_{T}^{2}}{v^{2}}\right)-1\Big\}
-m_{S}^{2}\Big\{\ln \left(\frac{m_{S}^{2}}{v^{2}}\right)-1\Big\}\Big), \nonumber \\
\end{eqnarray}
and finally,
\begin{eqnarray}
\delta\lambda_{3}&=& \frac{5 \lambda_{T}^{4}}{32\pi^{2}}
\ln (\frac{m_{T}^{2}}{v^{2}})
+\frac{1}{32\pi^{2}}\lambda_{S}^{4}
\ln\left(\frac{m_{S}^{2}}{v^{2}}\right)
+\frac{1}{32\pi^{2}}\frac{\lambda_{S}^{2}\lambda_{T}^{2}}{m_{T}^{2}
-m_{S}^{2}}\nonumber \\
&&\Big(m_{T}^{2}\Big\{\ln \left(\frac{m_{T}^{2}}{v^{2}}\right)-1\Big\}
-m_{S}^{2}\Big\{\ln\left(\frac{m_{S}^{2}}{v^{2}}\right)-1\Big\}\Big).\nonumber \\
\end{eqnarray}
Therefore, for large $\lambda_S$, $\lambda_T\sim\mathcal O(1)$, a
125 GeV Higgs boson mass can easily be accommodated in this model even in
the presence of a light stop mass and negligible left-right mixing.
\section{R-symmetry breaking}
Until now we have constrained ourselves in the R-symmetry preserving
scenario. Although the R-symmetric case in this regard is interesting and
should be explored in much more detail but in our work we pursue the path,
where R-symmetry is broken. Recent cosmological observations point towards
a vanishingly small vacuum energy or cosmological constant associated with
our universe. Spontaneously broken supergravity theory in a hidden sector
requires a non zero value of the superpotential in vacuum in order to have
this small vacuum energy. As the superpotential carries R-charge of two units
($R[W]=2$), therefore R-symmetry is broken when the superpotential acquires
a non zero vev $\langle W \rangle$. Furthermore, a non zero gravitino
mass also requires a non zero $\langle W \rangle$, thereby one can consider
the gravitino mass as the order parameter of R-symmetry breaking.
The breaking of R-symmetry has to be communicated to the visible sector and
in this context we confine ourselves to the case of anomaly mediation,
which plays the role of the messenger of R-symmetry breaking \cite{Kumar-1,
Chakraborty}. Such a scenario generates very small ($\sim$ a few MeV) Majorana
gaugino masses and trilinear scalar couplings, $M_i\sim \frac{g_i^2}
{16\pi^2}m_{3/2}$ and $A_{u/d}=\frac{\hat\beta_{h_{u/d}} v_{u/d}}{16\pi^2
m_{u/d}} m_{3/2}$ \cite{Guidice,Ghosh}, as long as the gravitino mass is in
the range of a few GeV.
In the R-breaking case, the neutralino mass matrix written in the
basis \\
$(\tilde b^{0}, \tilde S, \tilde w^{0}, \tilde T, \tilde R_{d}^{0},
\tilde H_{u}^{0}, N^{c}, \nu_{e})$, is given by
\begin{eqnarray}
M_{\chi}^{M}=\left(
\begin{array}{cccccccc}
M_{1} & M_{1}^{D} & 0 & 0 & 0 & \frac{g^{\prime}v_{u}}{\sqrt 2} & 0
& -\frac{g^{\prime}v_{a}}{\sqrt 2}\\
M_{1}^{D} & 0 & 0 & 0 & \lambda_{S}v_{u} & 0 & M_{R} & 0\\
0 & 0 & M_{2} & M_{2}^{D} & 0 & -\frac{g v_{u}}{\sqrt 2} & 0 &
\frac{g v_{a}}{\sqrt 2}\\
0 & 0 & M_{2}^{D} & 0 & \lambda_{T}v_{u} & 0 & 0 & 0 \\
0 & \lambda_{S}v_{u} & 0 & \lambda_{T}v_{u} & 0 & \mu_{u}+
\lambda_{S}v_{S}+\lambda_{T}v_{T} & 0 & 0\\
\frac{g^{\prime}v_{u}}{\sqrt 2} & 0 & -\frac{gv_{u}}{\sqrt 2}& 0 &
\mu_{u}+\lambda_{S}v_{S}+\lambda_{T}v_{T} & 0 & -fv_{a} & 0\\
0 & M_{R} & 0 & 0 & 0 & -fv_{a} & 0 & -fv_{u} \\
-\frac{g^{\prime}v_{a}}{\sqrt 2}& 0 & \frac{g v_{a}}{\sqrt 2} & 0
& 0 & 0 & -fv_{u} & 0
\end{array} \right). \nonumber \\
\label{majorana-neutralino}
\end{eqnarray}
An approximate expression for the tree level Majorana neutrino mass
is given by \cite{Chakraborty}
\begin{eqnarray}
\label{majorana-mass-tree-level}
(m_\nu)_{\rm Tree} &\simeq&-v^{2}\frac{\left[g \lambda_{T} v^{2}
(M_{2}^{D}-M_{1}^{D})\sin\beta\right]^{2}}{\left[M_{1}\alpha^{2}
+M_{2}\delta^{2}\right]},
\end{eqnarray}
where
\begin{eqnarray}
\alpha&=&\frac{2 M_{1}^{D} M_{2}^{D}\gamma\tan\beta}
{g\tan\theta_{w}}
+\sqrt 2 v^{2}\lambda_{S}\tan\beta(M_{1}^{D}\sin^{2}\beta+
M_{2}^{D}\cos^{2}\beta),\nonumber \\
\delta&=&\sqrt2 M_{1}^{D}v^{2}\lambda_{T}\tan\beta,
\end{eqnarray}
and $\gamma$ has been defined earlier.
Note that the neutrino Yukawa coupling $f$ does not arise in this expression
because of our choice in eq.~(\ref{MR-lambda}). Therefore, it is obvious from
eq.~(\ref{majorana-mass-tree-level}) that in order to obtain a small tree level
Majorana neutrino mass, we either require a small $\lambda_T$ or nearly
degenerate Dirac gaugino masses\footnote{A detailed discussion on how to
fit the light neutrino masses and mixing in this model can be found in
\cite{Chakraborty}.}. In this work we are interested in the sterile
neutrino which might play the role of keV dark matter. From the $8\times 8$
neutralino mass matrix, the sterile neutrino mass can be approximated as
\begin{eqnarray}
M_N^R\simeq M_1\frac{2 f^2 \tan^2 \beta}{g^{\prime 2}}.
\label{sterile-mass}
\end{eqnarray}
For a wide range of parameters the active-sterile mixing can also be
estimated as
\begin{eqnarray}
\theta_{14}^2\simeq \frac{(m_{\nu})_{\rm Tree}}{M_N^R}.
\label{mixing}
\end{eqnarray}
\begin{figure}[htb]
\begin{center}
\includegraphics[height=3.5in,width=3.5in]{tanb_f.eps}
\caption{\label{fig:tanb_f}
The contour in the black thick line represents a sterile neutrino mass of
7 keV. Contours in red (dotted) and blue (dashed) colours show active-sterile
mixing $2.2\times 10^{-11}$ and $2\times 10^{-10}$ respectively.}
\end{center}
\end{figure}
In figure~(\ref{fig:tanb_f}), we show in the ($f - \tan\beta$) plane the contour
of sterile neutrino mass fixed at 7.06 keV and also two different contours of
$\sin^2 2\theta_{14}$, fixed at the lower and upper limit at $2.2\times 10^{-11}$
and $2\times 10^{-10}$ respectively. We have chosen the gravitino mass, $m_{3/2}$
to be 10 GeV and $M_1^D = 900 ~{\rm GeV}$, keeping a degeneracy between the Dirac
gaugino masses, $\epsilon \equiv (M_2^D -M_1^D) = 10^{-4} ~{\rm GeV}$. We have
also fixed $\mu_u = 750 ~{\rm GeV}$, $\lambda_S = 1.1$, $v_S = -0.1 ~{\rm GeV}$
and $v_{T} = 0.1 ~{\rm GeV}$.
The sterile neutrino mass contour can be easily explained by looking at
eq.~(\ref{sterile-mass}). Similarly from eq.~(\ref{majorana-mass-tree-level}),
eq.~(\ref{sterile-mass}) and eq.~(\ref{mixing}), it is straightforward to show
that $\sin^2 2\theta_{14}$ goes as $\frac{1}{1+\tan^2 \beta}$. This means that
for smaller $\tan\beta$ one would expect larger mixing angle for fixed values
of other parameters. This is also evident from figure~\ref{fig:tanb_f}. Furthermore,
for larger Dirac gaugino masses, the active neutrino mass gets reduced
(see eq.~(\ref{majorana-mass-tree-level})), which also implies a reduction
in the active-sterile mixing.
\begin{figure}[htb]
\begin{center}
\includegraphics[height=3.5in,width=3.5in]{M1D_tanb.eps}
\caption{\label{fig:M1D_tanb}
Showing the lower and upper limits of $\tan\beta$ from X-ray analysis
as a function of $M_2^D$ for $\mu_u$ = 700 GeV, $m_{3/2}$ = 10 GeV and
$\epsilon = 10^{-4}$ GeV.}
\end{center}
\end{figure}
Looking at figure~\ref{fig:tanb_f}, we observe that the largest value of the
active-sterile mixing, required to explain the observed photon line flux at
an energy $E \approx$ 3.5 keV, corresponds to the minimum value of $\tan\beta$.
In fact, for this particular case shown in figure~\ref{fig:tanb_f}, $(\tan\beta)_
{\rm min} \approx 11.3$. Similarly the smallest active-sterile mixing
($\sin^2 2\theta_{14} = 2.2 \times 10^{-11}$) provides the maximum allowed
value of $\tan\beta$, which in this case turns out to be $(\tan\beta)_{\rm max}
\approx 33$. In order to obtain an analytical relationship between the lower
limit of $\tan\beta$ and $M_2^D$, we can solve for $\tan\beta$ using
eq.~(\ref{mixing}), with $\sin^2 2\theta_{14}=2\times 10^{-10}$ and
$M_N^R = 7.06 ~{\rm keV}$. This gives rise to
\begin{eqnarray}
(\tan^2\beta)_{\rm min} &=& \dfrac{4 v^2 \{g \lambda_T v^2 (M^D_2 - M^D_1) \}^2}
{(1.4 \times 10^{-15} ~{\rm GeV}) [M_1 {\alpha^\prime}^2 + M_2 {\delta^\prime}^2]} -1,
\nonumber \\
\label{eq:tanb_low_high}
\end{eqnarray}
where
\begin{eqnarray}
\alpha^\prime &\simeq& \dfrac{2 (M_2^D)^2 \mu_u}{g^\prime} + \sqrt{2} v^2 \lambda_s M^D_2,
\nonumber \\
\delta^\prime &\simeq& \sqrt{2} M^D_2 v^2 \lambda_T.
\end{eqnarray}
In a similar way an analytical expression for the upper limit of $\tan\beta$
can also be derived.
Figure~\ref{fig:M1D_tanb} shows the lower and upper limits of $\tan\beta$ as
a function of $M_2^D$, for $\mu_u$ = 700 GeV, $m_{3/2}$ = 10 GeV and
$\epsilon = 10^{-4}$ GeV. We have fixed $\lambda_S$ at the previously mentioned
value. The horizontal grey line shows the upper limit on $\tan\beta$ arising
from the contribution of the leptonic Yukawa coupling, $f_{\tau} \equiv
\lambda_{133}$ to the ratio $R_\tau \equiv \Gamma(\tau \rightarrow e
{\bar \nu}_e \nu_\tau)/\Gamma(\tau \rightarrow \mu {\bar \nu}_\mu \nu_\tau)$. The
resulting constraint is $f_{\tau}<0.07\left(\frac{m_{{\tilde \tau}_R}}
{100~{\rm GeV}}\right)$ \cite{Kumar-1} and considering stau mass, close to
$280 ~{\rm GeV}$, translates into an upper limit on $\tan\beta \approx 19$.
For higher stau mass this upper limit on $\tan\beta$ gets relaxed.
The blue dashed line shows the lower bound on $\tan\beta$, as a function
of $M_2^D$, arising from the precision measurements of the deviations in
the couplings of the Z boson to charged leptons \cite{Kumar-1}.
We infer from the above discussions, that in a large region of the parameter
space, the lower limit on $\tan\beta$, satisfying the estimated mass and mixing
of the sterile neutrino dark matter particle coming from the recent observation
of an X-ray line signal at energy 3.5 keV is stronger than the lower limit on
$\tan\beta$ coming from the electroweak precision measurements. On the other
hand, the upper limit on $\tan\beta$ coming from the X-ray observations becomes
stronger than the upper limit arising from the $\tau$ Yukawa coupling contribution
to $R_\tau$ only for higher values of $M^D_2$ as shown in figure~\ref{fig:M1D_tanb}
for specific choices of $\mu_u$ and $m_{3/2}$. Combining these lower and upper
limits on $\tan\beta$ from X-ray observations and measurement of $R_\tau$, we
can find a range of $M^D_2$ that is allowed. For smaller values of $\mu_u$ and
$m_{3/2}$, the upper and lower limits of $M^D_2$ shift to higher values
(see eq.~(\ref{eq:tanb_low_high})).
We also observe from figure~\ref{fig:tanb_f} that the allowed values of $f$ is of
the order of $10^{-4}$. Such a small value of $f$, implies negligible extra
contribution to the tree level Higgs boson mass. Therefore, to elevate the Higgs
boson mass to $125$ GeV, we have to rely on the loop corrections. Sizable
radiative corrections are obtained if $\lambda_S$, $\lambda_T$ are large
($\mathcal O(1)$) and this would imply nearly degenerate Dirac gaugino masses
($\epsilon \sim 10^{-4}$ GeV) in order to have the active-sterile mixing
$\sin^2 2\theta_{14} \sim 10^{-11}$ and a tree level active neutrino mass
$\lsim$ 0.05 eV.
The other case, which can relax this strong degeneracy between Dirac gaugino
masses, corresponds to the case of small $\lambda_S$, $\lambda_T\sim\mathcal
O(10^{-4})$, which implies multi-TeV stop to fit the Higgs boson mass. Therefore,
this model provides a very interesting possibility where we can connect the
Higgs sector with the neutrino sector (both active as well as sterile neutrino).
\section{Right handed neutrino as a keV warm dark matter}
To accommodate sterile neutrino as a warm dark matter candidate, it
is very important to make sure that the active sterile mixing is
very small~\cite{Abazajian-1,Shaposhnikov,Shaposhnikov-1,Biermann,Dahle,
Roy} and within the valid range of different X-ray experiments.
A rough bound on the active-sterile mixing can be parametrised as
\cite{Boyarsky-1}
\begin{eqnarray}
\theta_{14}^{2}\leq 1.8\times 10^{-5}\Big(\frac{1 \rm{keV}}{M_N^R}\Big)^5.
\end{eqnarray}
Along with the strict bound coming from different X-ray experiments,
the keV sterile neutrino must produce the correct relic density $\Omega_N
h^2\sim 0.1$, in order to identify itself with the warm dark matter.
An approximate formula for the relic density of sterile neutrinos, produced
in the early universe with negligible lepton asymmetry via non-resonant
oscillations with active neutrinos, known as the Dodelson-Widrow (DW) mechanism
\cite{Dodelson} can be written as \cite{Fuller}
\begin{eqnarray}
\Omega_N h^2\sim 0.3 \Big(\frac{\sin^2 2\theta_{14}}{10^{-10}}\Big)
\Big(\frac{M_N^R}{100 ~\rm{keV}}\Big)^2,
\end{eqnarray}
where $\Omega_N$ is the ratio of the sterile neutrino density to the
critical density of the Universe and $h=0.673$.
Different experimental observations have also put lower limits on the mass
of the keV warm dark matter. A very robust bound for fermionic dark matter
particles comes from Pauli exclusion principle. By claiming the maximal
(Fermi) velocity of the degenerate fermionic gas in the dwarf spheroidal
galaxies is less compared to the escape velocity, translates into a lower
bound on the sterile neutrino dark matter mass, i.e $M_N^R>0.41$ keV
\cite{Tremaine-Gunn,Boyarsky-mass-bound}. Model dependent bounds on the
mass of the warm dark matter are much more stringent and obtained from
analysing Lyman-$\alpha$ experiment \cite{Seljak, Lesgourgues}.
\begin{figure}[htb]
\begin{center}
\includegraphics[height=3.5in,width=3.5in]{final_light.eps}
\caption{\label{fig:light_stop}
The red (grey) points in the mass-mixing plane are obtained by scanning the parameter
space as mentioned in the text. The yellow (light) region is ruled out from the
Tremaine Gunn bound \cite{Tremaine-Gunn,Boyarsky-mass-bound}. Cosmic X-ray
background (CXB) rules out the region in red stripes \cite{X-ray-bound}.
Constraints from M31, observed by Chandra rules out the region in grey
\cite{Yuksel-Watson}. The blue region is ruled out from the diffuse X-ray background
observations \cite{X-ray-bound}. XMM-Newton observations from Coma and Virgo
clusters rule out the region in green. The light blue line represents the 100
\% relic density of the sterile neutrino dark matter, produced via DW mechanism.
The light blue region above this line leads to over abundance of the sterile
neutrino warm dark matter. Finally, the black star represents the central value of
the mass and active-sterile mixing, from the 3.5 keV X-ray line observation.}
\end{center}
\end{figure}
In figure~\ref{fig:light_stop} we present a scatter plot by scanning the parameter
space of our model and also show the compatibility of those points with the
current experimental findings. The red circles are the points obtained by varying
the parameters as $500~\rm{GeV}<M_1^D<1.2~\rm{TeV}$, $10^{-5}<f<10^{-3}$,
$2.7<\tan\beta<17$, $400~\rm{GeV}<m_{\tilde t_1,\tilde t_2}<1.2 ~{\rm TeV}$, keeping
$\epsilon \equiv (M_2^D-M_1^D)\sim 10^{-4}~\rm{GeV}$. $\mu_u$ and $\lambda_S$ are
fixed at 750 GeV and 1.1 respectively ($\lambda_T = \lambda_S \tan\theta_W \sim 0.6$).
All these points respect a Higgs boson mass in between 124.4 GeV and 126.2 GeV
avoiding any tachyonic scalar states.
Similar plot can also be generated where $\lambda_T\sim 10^{-5}$. Therefore,
to fit the Higgs boson mass in that case, one requires $m_{\tilde t} > 5~\rm{TeV}$.
However, the degeneracy between $M_1^D$ and $M_2^D$ is somewhat lifted where
$\epsilon \gsim 1~\rm{GeV}$.
The horizontal yellow band in figure\ref{fig:light_stop} is ruled out by the Tremaine
Gunn bound, which implies $M_N^R<0.4~\rm{keV}$ \cite{Tremaine-Gunn,
Boyarsky-mass-bound}. The blue region is excluded by taking into consideration
the diffuse X-ray background \cite{X-ray-bound}. Cluster X-ray bound rules out
a region in the mass-mixing plane by taking into consideration XMM-Newton
observations from the Coma and Virgo clusters \cite{Cluster-x-ray}. Constraints
from the cosmic X-ray background (CXB) rules out the region in red stripes
\cite{X-ray-bound}. Chandra observation of M31 \cite{Yuksel-Watson} rules
out the region in grey. The light blue line corresponds to the correct relic density
provided by the sterile neutrino warm dark matter via DW mechanism. The light
blue region above this line marked as DW is ruled out because of the over abundance of
sterile neutrino dark matter. The horizontal and vertical lines show the region
in the mass and mixing plane consistent with the observed 3.5 keV X-ray line with more
than 3$\sigma$ significance. The black star corresponds to the best fit point.
It is clearly evident from this figure that such
a small mixing is completely in conflict with the DW production mechanism of
sterile neutrinos. However, resonant production of sterile neutrinos in the
presence of a lepton asymmetry in primordial plasma can be very important and
produce correct relic abundance of the keV sterile neutrinos \cite{Fuller,Shi}.
Recent studies have shown that a cosmological lepton asymmetry $L \sim
\mathcal O(10^{-3})$ is capable of producing correct relic density of 0.119
\cite{Abazajian}. It was shown in \cite{Volkas, Shi-1, Foot-Volkas, Foot, Kishimoto,
Dolgov,Dolgov-1,Maalampi}
that active-sterile neutrino oscillations can themselves create a cosmological
lepton number of this magnitude, assuming that the number of sterile neutrinos
is negligible to start with. Such a possibility can be easily conceived in
our model to generate a large lepton asymmetry.
Let us note in passing that sterile neutrino production in non-standard
cosmology with low reheating temperature ($\sim$ a few MeV) has also been
discussed in the literature \cite{Pascoli,Yaguna,Osoba}. If the universe has
undergone inflation and was never reheated to a temperature above a few MeV
then the relic abundance of the sterile neutrinos can be written as
\begin{eqnarray}
\Omega_N h^2 = 10^{-7} d_{\alpha}\Big(\frac{\sin^2 2\theta_{14}}{10^{-10}}\Big)
\Big(\frac{M_N^R}{10 ~\rm{keV}}\Big)\Big(\frac{T_R}{5 ~{\rm MeV}}\Big)^3,
\end{eqnarray}
where $d_{\alpha}=1.13$, assuming that the sterile neutrino couples only with
$\nu_e$ as in our case. It is obvious from the above expression that for allowed
values of $\sin^2 2\theta_{14}$ and $M_N^R$ (from the recent X-ray observation)
this production mechanism will give rise to severe under abundance of sterile
neutrinos.
In our model sterile neutrinos can also be produced non-thermally via the decay
of heavier scalar particles. However, a quantitative estimate of the relic
density requires a thorough investigation and we postpone the discussion of
this method of production for a future work \cite{Chakraborty-1}.
\begin{figure}[htb]
\begin{center}
\includegraphics[height=3.5in,width=3.5in]{sterile_lifetime.eps}
\caption{\label{fig:sterile_lifetime}
The plot shows the sterile neutrino lifetime as a function of
the sterile neutrino mass. The black vertical line represents
the 7 keV mass of the sterile neutrino. The red points represent the
total life time of the sterile neutrino.}
\end{center}
\end{figure}
\subsection{Sterile neutrino decay}
The most dominant decay mode of the sterile neutrino is $N\rightarrow 3\nu$.
The corresponding decay rate for this process is given by \cite{Abazajian-1}
\begin{eqnarray}
\Gamma_{3\nu}=8.7\times 10^{-31} {~\rm{sec}}^{-1}
\Big(\frac{\sin^2 2\theta_{14}}{10^{-10}}\Big)
\Big(\frac{M_N^R}{1 ~\rm{keV}}\Big)^5.
\end{eqnarray}
The principal radiative decay mode of the sterile neutrino which
is of concern here is $N\rightarrow \nu\gamma$ and the decay width
is
\begin{eqnarray}
\Gamma_{\nu\gamma} = 1.38 \times 10^{-32} ~{\rm{sec}}^{-1}
\Big(\frac{\sin^2 2\theta_{14}}{10^{-10}}\Big)
\Big(\frac{M_N^R}{1 ~\rm{keV}}\Big)^5.
\label{sterile_decay}
\end{eqnarray}
This decay produces a monochromatic photon line at $E_{\gamma}=
\frac{M_N^R}{2}$.
From figure~(\ref{fig:sterile_lifetime}) we can see that the lifetime of
the sterile neutrino is much larger than the age of the universe.
\section{Gravitino cosmology}
As mentioned earlier, the gravitino mass is the order parameter of R-breaking.
If the mass is around a few GeV, it can be a candidate for cold dark matter
\cite{Covi}. In our scenario, the gravitino is an unstable particle and decays
to an active/sterile neutrino and a monochromatic photon. The tree level decay mode
into an active neutrino final state $\tilde G\rightarrow \gamma\nu_e$
is suppressed by the very small mixing $U_{\tilde b\nu_e}$ ($\sim 10^{-7}$)
between the bino and active neutrino $\nu_e$ \cite{Takayama}. Interestingly, in
our model the most dominant decay mode of gravitino is into a photon and a sterile
neutrino ($\tilde G\rightarrow N\gamma$) and the decay width is given as
\begin{eqnarray}
\Gamma_{\tilde G\rightarrow N\gamma}\sim \frac{|U_{\tilde b N}|^{2}m_{3/2}^{3}}
{32\pi M_{P}^{2}},
\end{eqnarray}
where $U_{\tilde b N}$ is the bino sterile neutrino mixing angle. Because of the
presence of the term $M_R {\hat N}^c {\hat S}$ in the superpotential and the
bino Dirac mass term in the Lagrangian, the tree level bino sterile neutrino
mixing is not strongly suppressed ($\sim 10^{-2}$).
For the sake of completeness, let us mention that at the one loop level the
decay $\tilde G\rightarrow \gamma\nu_e$ occurs \cite{Masiero,Lola,Lola-1,
Ghosh-Zhang} via trilinear R-parity violating coupling $\lambda_{133}^{\prime}$
which we have identified with the bottom Yukawa coupling. We have checked that
this process is also suppressed compared to the tree level decay $\tilde
G\rightarrow N\gamma$. The one-loop contribution to the decay $\tilde G\rightarrow
N\gamma$ is negligible because of small active-sterile mixing.
Taking into account the most dominant decay mode of the gravitino in the sterile
neutrino plus photon final state, for a 10 GeV gravitino mass, the lifetime is close
to $10^{15}$ sec. Therefore, to satisfy the experimental constraints coming from the
diffuse photon background, one has to consider a scenario where the gravitino density
is very much diluted. In order to provide a quantitative analysis we note that
for a gravitino of mass 10 GeV the limit on the diffuse photon flux is around
$6.89\times 10^{-7}\rm{GeV}\rm {cm^{-2}}\rm{sec^{-1}}$ \cite{Yuksel}. This can
be translated into a bound on the gravitino relic density and we find
\begin{eqnarray}
\Omega_{3/2}h^2< 4.34\times 10^{-13} \left(\frac{10^{-2}}{U_{\tilde bN}}\right)^2,
\end{eqnarray}
for a 10 GeV gravitino. Note that this bound depends strongly on the mass of
the gravitino and will get relaxed for a smaller gravitino mass.
To satisfy such a strong bound on the gravitino relic density,
one must account for a very low reheating temperature. If the reheating
temperature is above the SUSY scale, the gravitino relic density would be
too large \cite{Buchmuller}. Therefore, the reheating temperature must lie
much below the SUSY threshold.
Following \cite{Gregoire}, we see that if the reheating temperature is below
the SUSY threshold, the gravitinos are produced by thermal scattering with
neutrinos and bottom quarks. Using the results of \cite{Gregoire} for
production of gravitinos, we obtain an upper bound on the reheating temperature
for a 10 GeV gravitino as
\begin{eqnarray}
T_{R}&<&127 \left(\frac{v_a}{30\rm{GeV}}\right)^{2/7}
\left(\frac{m_{\tilde b}}{500\rm{GeV}}\right)^{4/7}
\left(\frac{10^{-2}}{U_{\tilde b N}}\right)^{2/7}\rm{GeV}. \nonumber \\
\end{eqnarray}
Such a low reheating temperature might have important implications
in the context of different baryogenesis and leptogenesis scenarios.
\section{Conclusion}
Recent observation of a weak X-ray line around $E_{\gamma}=3.5~\rm{keV}$
by XMM-Newton telescope coming from Andromeda galaxy and various
galaxy clusters have been studied in the light of a $U(1)_{R^-}$lepton
number model, with a single right handed neutrino. We have shown explicitly
that a sterile neutrino of mass about 7 keV and with appropriate
active-sterile mixing can easily be obtained in our model. We briefly
mention different production mechanisms of the sterile neutrino.
Allowed ranges of the mass and mixing helped us to put bounds on
$\tan\beta$ as a function of the Dirac wino mass $M_2^D$. Combining these
bounds with the limits coming from the measurements of the $\tau$ Yukawa
coupling contribution to the ratio $R_\tau \equiv \Gamma(\tau \rightarrow
e {\bar \nu}_e \nu_\tau)/\Gamma(\tau \rightarrow \mu {\bar \nu}_\mu \nu_\tau)$,
one obtains strong upper and lower bounds on $M^D_2$.
In addition, we have also discussed the Higgs sector briefly and pointed out
different possibilities to have a Higgs boson mass around 125 GeV. Finally,
gravitino is the LSP in our model with a mass about a few GeV and gravitino
mass is the order parameter of R-symmetry breaking. The gravitino can decay
into a photon plus active or sterile neutrino. Therefore, we have also
presented a short discussion on the cosmological implications of the gravitino.
We have taken into account the most robust constraint coming from the diffuse
photon background, which readily puts a very stringent bound on the gravitino
relic density. This eventually imposes an upper limit ($\lsim$ 130 GeV) on the
reheating temperature of the universe.
\begin{acknowledgments}
We thank Kevork N. Abazajian and George G Raffelt for helpful discussions.
S.C would also like to thank the Council of Scientific and Industrial Research
(CSIR), Government of India for financial support obtained as a Senior Research
Fellow.
\end{acknowledgments}
|
\section*{Methods}
The single crystals of BaFe$_2$(As$_{1-x}$P$_x$)$_2$ were grown by the self-flux method \cite{Kasahara10} and characterised by several techniques as reported previously \cite{Shibauchi14}. The observation of the quantum oscillations in this series of crystals and the sharp superconducting transition indicate the very high quality of our pristine crystals. We used several crystals from three batches, which exhibit slightly different $T_{\rm c0}$ values of 28, 29, and 30\,K.
Electron irradiation experiments were performed on SIRIUS platform operated by LSI at Ecole Polytechnique, composed by Pelltron type NEC accelerator and closed cycle cryocooler maintaining sample immersed in liquid hydrogen at 20-22\,K during irradiation. The low-temperature environment is important to prevent defect migration and agglomeration. Partial annealing of introduced defects occurs upon warming to room temperature and sample transfer \cite{Prozorov14}.
The resistivity measurements were performed at LSI, Ecole Polytechnique by the van der Pauw method with four contacts on corners of crystals to minimize the possible effect of the unirradiated area due to contacts. The penetration depth measurements by using 13\,MHz TDOs \cite{Hashimoto10,Hashimoto12} were performed at Kyoto University before and after irradiation.
|
\section{Introduction}
\label{sec:intro}
At an instantaneous luminosity of $10^{34}$\percms, typical of that expected at the Large Hadron Collider (LHC), with the proton bunches crossing at intervals of 25\unit{ns},
the Compact Muon Solenoid (CMS) tracker is expected to be traversed by about 1000 charged particles at each bunch crossing, produced by
an average of more than twenty proton--proton (pp) interactions. These multiple interactions
are known as \textit{pileup}, to which prior or later bunch crossings can also contribute because of the
finite time resolution of the detector.
Reconstructing tracks in such a high-occupancy environment is immensely challenging. It is difficult to
attain high track-finding efficiency, while keeping the fraction of \textit{fake} tracks small. Fake tracks
are falsely reconstructed tracks that may be formed from a combination of unrelated hits or from a genuine
particle trajectory that is badly reconstructed through the inclusion of spurious hits.
In addition, the tracking software must run sufficiently fast to be used not only for offline event
reconstruction (of ${\approx}10^9$ events per year), but also for the CMS High-Level Trigger (HLT),
which processes events at rates of up to 100~kHz.
The scientific goals of CMS \cite{Bayatian:2006zz,Ball:2007zza} place demanding requirements on the
performance of the tracking system. Searches for high-mass dilepton resonances, for example, require
good momentum resolution for transverse momenta $\pt$ of up to 1\TeV. At the same time,
efficient reconstruction of tracks with very low \pt of order 100\MeV is needed for studies of hadron
production rates and to obtain optimum jet energy resolution with particle-flow techniques
\cite{CMS_PAS_PFT-09-001}. In addition, it is essential to resolve nearby tracks, such as those
from 3-prong $\tau$-lepton decays. Furthermore, excellent impact parameter resolution is needed for
a precise measurement of the positions of primary pp interaction vertices as well as for identifying b-quark jets
\cite{Chatrchyan:2012jua}.
While the CMS tracker \cite{:2008zzk} was designed with the above requirements in mind, the track-finding
algorithms must fully exploit its capabilities, so as to deliver the desired performance. The goal of this
paper is to describe the algorithms used to achieve this and show the level of performance attained.
The focus here is purely on pp collisions, with heavy ion collisions being beyond the scope of this document.
Section~\ref{sec:cmsTracker} introduces the CMS tracker; and Section~\ref{sec:localReco} describes the
reconstruction of the \textit{hits} created by charged particles crossing the tracker's sensitive layers.
The algorithms used to reconstruct tracks from these hits are explained in Section~\ref{sec:trackReco}; and
the performance obtained in terms of track-finding efficiency, proportion of fake tracks and track parameter resolution is presented in
Section~\ref{sec:trackPerformance}. Primary vertices from pp collisions are distributed over a
luminous region known as the beam spot. Reconstruction of the beam spot
and of the primary vertex positions is described in Section~\ref{sec:beamSpotAndPV}. This is intimately connected with tracking,
since on the one hand, the beam spot and primary vertices are found using reconstructed tracks, and on the other hand, an approximate knowledge
of their positions is needed before track finding can begin.
All results shown in this paper are based on
pp collision data collected or events simulated at a centre-of-mass energy of $\sqrt{s}=7$\TeV in 2011.
The simulated events include a full simulation of the CMS detector response based on \GEANTfour
\cite{Agostinelli:2002hh}. All events are reconstructed using software from the same period.
The track-reconstruction algorithms have been steadily evolving since then, but still have a similar
design now.
The CMS detector \cite{:2008zzk} was commissioned initially using cosmic ray muons and subsequently
using data from the first LHC running period.
Results obtained using cosmic rays in 2008 \cite{craft_paper} are extensively documented
in several publications pertaining to the pixel detector~\cite{craft_pixel_paper}, strip
detector~\cite{craft_strip_paper}, tracker alignment~\cite{craft_alignment_paper}, and magnetic
field~\cite{craft_bfield_paper}, and are of particular relevance to the present paper.
Results from the commissioning of the tracker using pp collisions in 2010 are presented in
\cite{TRK10_001}.
\section{The CMS tracker \label{sec:det}}
\label{sec:cmsTracker}
The CMS collaboration uses a right-handed coordinate system, with the origin at the
centre of the detector, the $x$-axis pointing to the centre of the
LHC ring, the $y$-axis pointing up (perpendicular to the plane of the LHC ring), and with
the $z$-axis along the anticlockwise-beam direction. The polar angle
$\theta$ is defined relative to the positive $z$-axis and the azimuthal
angle $\phi$ is defined relative to the $x$-axis in the $x$-$y$ plane.
Particle pseudorapidity $\eta$ is defined as $-\ln[\tan(\theta/2)]$.
The CMS tracker \cite{:2008zzk} occupies a cylindrical volume 5.8\ensuremath{\,\text{m}}\xspace in length and 2.5\ensuremath{\,\text{m}}\xspace in diameter,
with its axis closely aligned to the LHC beam line. The tracker is
immersed in a co-axial magnetic field of 3.8\unit{T} provided by the CMS solenoid.
A schematic drawing of the CMS tracker is shown in
Fig.~\ref{fig:TrackerLayout}.
The tracker comprises a large silicon strip tracker with a small silicon pixel tracker inside it. In the
central pseudorapidity region, the pixel tracker consists of three co-axial barrel layers at radii between
4.4\cm and 10.2\cm and the strip tracker consists of ten co-axial barrel layers extending outwards to a
radius of 110\cm. Both subdetectors are completed by endcaps on either side of the barrel,
each consisting of two disks in the pixel tracker, and three small plus nine large disks in the
strip tracker. The endcaps extend the acceptance of the tracker up to a pseudorapidity of $\abs{\eta}<2.5$.
\begin{figure}[hbtp]
\centering
\includegraphics[width=1.00\textwidth]{figs_2011/cmsTracker/TrackerLayoutNew.pdf}
\caption{Schematic cross section through the CMS tracker in the $r$-$z$ plane. In this view,
the tracker is symmetric about the horizontal line $r=0$, so only the top half is shown here.
The centre of the tracker, corresponding to the approximate position of the pp collision point,
is indicated by a star. Green dashed lines help the reader understand which modules belong to each of
the named tracker subsystems. Strip tracker modules that provide 2-D hits are shown by thin, black
lines, while those permitting the reconstruction of hit positions in 3-D are shown by thick, blue lines.
The latter actually each consist of two back-to-back strip modules, in which one module is rotated
through a `stereo' angle. The pixel modules, shown by the red lines, also provide 3-D hits. Within a given
layer, each module is shifted slightly in $r$ or $z$ with respect to its neighbouring modules,
which allows them to overlap, thereby avoiding gaps in the acceptance.}
\label{fig:TrackerLayout}
\end{figure}
The pixel detector consists of cylindrical barrel layers at radii of 4.4, 7.3
and 10.2\cm, and two pairs of endcap disks at $z= {\pm}34.5$ and $\pm$46.5\cm.
It provides three-dimensional (3-D) position measurements of the hits
arising from the interaction of charged particles with its sensors.
The hit position resolution is approximately 10\mum in the transverse coordinate and 20--40\mum in the
longitudinal coordinate, while the third coordinate is given by the sensor plane position.
In total, its 1440 modules cover an area of about 1\ensuremath{\,\text{m}^{2}}\xspace and have 66 million pixels.
The strip tracker has 15\,148 silicon modules, which in total cover
an active area of about 198\ensuremath{\,\text{m}^{2}}\xspace and have 9.3 million strips.
It is composed of four
subsystems. The Tracker Inner Barrel (TIB) and Disks (TID)
cover $r < 55$\cm and $\abs{z} < 118$\cm, and are composed of four barrel layers,
supplemented by three disks at each end.
These provide position measurements in $r\phi$ with a resolution of approximately 13--38\mum.
The Tracker Outer Barrel (TOB) covers $r > 55$\cm and $\abs{z} < 118$\cm and consists
of six barrel layers providing position measurements in $r\phi$ with a resolution of approximately 18--47\mum.
The Tracker EndCaps (TEC) cover the region $124 < \abs{z} < 282$\cm. Each TEC
is composed of nine disks, each containing up to seven concentric rings of silicon strip modules, yielding
a range of resolutions similar to that of the TOB.
To refer to the individual layers/disks within a subsystem, we use a numbering convention
whereby the barrel layer number increases with its radius and the endcap disk number increases
with its $\abs{z}$-coordinate. When referring to individual rings within an endcap disk, the ring
number increases with the radius of the ring.
The modules of the pixel detector use silicon of 285\mum thickness, and achieve resolutions that are
roughly the same in $r\phi$ as in $z$, because of the chosen pixel cell size of $100\times 150$\ensuremath{\,\mu\text{m}^{2}}\xspace
in $r\phi\times z$.
The modules in the TIB, TID and inner four TEC rings use silicon that is 320\mum thick, while those in
the TOB and the outer three TEC rings use silicon of 500\mum thickness.
In the barrel, the silicon strips usually run parallel to
the beam axis and have a pitch (\ie, the distance between neighbouring strips) that varies from 80\mum in the inner TIB layers to 183\mum in the inner
TOB layers. The endcap disks use wedge-shaped sensors with radial strips, whose pitch varies from
81\mum at small radii to 205\mum at large radii.
The modules in the innermost two layers of both the TIB and the TOB,
as well as the modules in rings 1 and 2 of the TID, and 1, 2 and 5 of the TEC, carry a second
strip detector module, which is mounted back-to-back to the first and rotated in the plane
of the module by a `stereo' angle of 100\ensuremath{\,\text{mrad}}\xspace. The hits from these two modules, known as `$r\phi$' and
`stereo hits', can be combined into \textit{matched hits} that provide
a measurement of the second coordinate ($z$ in the barrel and $r$ on the disks). The achieved
single-point resolution of this measurement is an order of magnitude worse than in $r\phi$.
The principal characteristics of the tracker are summarized in Table~\ref{tab:TrackerGeom}.
Figure~\ref{fig:MaterialBudget} shows the material budget of the
CMS tracker, both in units of radiation lengths and nuclear interaction lengths, as estimated
from simulation. The simulation describes the tracker material budget with an
accuracy better than 10\% \cite{CMS_PAS_TRK-10-003}, as was established by measuring the distribution of
reconstructed nuclear interactions and photon conversions in the tracker.
\begin{table}[htbp]
\centering
\topcaption{\label{tab:TrackerGeom} A summary of the principal characteristics of the various tracker
subsystems. The number of disks corresponds to that in a single endcap. The location specifies
the region in $r$ ($z$) occupied by each barrel (endcap) subsystem.}
\begin{tabular}{llll}
\hline
Tracker subsystem & Layers & Pitch & Location \\
\hline
Pixel tracker barrel & 3 cylindrical & $100\times 150$\ensuremath{\,\mu\text{m}^{2}}\xspace & $4.4 < r < 10.2$\cm \\
Strip tracker inner barrel (TIB) & 4 cylindrical & 80--120\mum & $20 < r < 55$\cm \\
Strip tracker outer barrel (TOB) & 6 cylindrical & 122--183\mum & $55 < r < 116$\cm \\
\hline
Pixel tracker endcap & 2 disks & $100\times 150$\ensuremath{\,\mu\text{m}^{2}}\xspace & $34.5 < \abs{z} < 46.5$\cm \\
Strip tracker inner disks (TID) & 3 disks & 100--141\mum & $58 < \abs{z} < 124$\cm \\
Strip tracker endcap (TEC) & 9 disks & 97--184\mum & $124 < \abs{z} < 282$\cm \\
\hline
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/cmsTracker/MaterialBudget_RadLengths.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/cmsTracker/MaterialBudget_InteractionLengths.pdf}
\caption{Total thickness $t$ of the tracker material traversed by a particle produced at the
nominal interaction point, as a function of pseudorapidity $\eta$, expressed in units of radiation
length $X_0$ (left) and nuclear interaction length $\lambda_I$ (right). The contribution to the total
material budget of each of the subsystems that comprise the CMS tracker is shown, together
with contributions from the beam pipe and from the support tube that surrounds the tracker.
\label{fig:MaterialBudget}}
\end{figure}
\section{Reconstruction of hits in the pixel and strip tracker}
\label{sec:localReco}
The first step of the reconstruction process is referred to as \textit{local reconstruction}. It consists of the clustering of \textit{zero-suppressed} signals above specified thresholds in
pixel and strip channels into hits,
and then estimating the cluster positions and their uncertainties defined in a local orthogonal coordinate system $(u,v)$ in
the plane of each sensor.
A pixel sensor consists of $100\times 150\mum^2$ pixels with the $u$-axis oriented parallel to the shorter pixel edge.
In the strip sensors, the $u$-axis is chosen perpendicular to the central strip in each sensor (which in the TEC is not parallel to the
other strips in the same sensor).
\subsection{Hit reconstruction in the pixel detector}
\label{sec:localPixelReco}
In the data acquisition system of the pixel detector~\cite{Kotlinski:2006jz}, zero-suppression is performed in the readout chips of the sensors~\cite{Kastli:2005jj},
with adjustable thresholds for each pixel.
This pixel readout threshold is set to a single-pixel threshold corresponding to an equivalent charge of 3200 electrons.
Offline, pixel clusters are formed from adjacent pixels, including both side-by-side and corner-by-corner
adjacent cells. Each cluster must have a minimum charge equivalent to 4000 electrons. For comparison, a minimum ionizing particle deposits usually around 21000 electrons.
Miscalibration of residual charge caused by pixel-to-pixel differences of the charge injection capacitors, which are used to calibrate the pixel gain,
are extracted from laboratory measurements and included in the Monte Carlo (MC) simulation.
Two algorithms are used to determine the position of pixel clusters.
A fast algorithm (described in Section~\ref{sec:firstPassPixel}) is used during track seeding and pattern recognition, and a more precise algorithm (Section~\ref{sec:templatesPixel}), based on cluster shapes, is used in the final track fit.
\subsubsection{First-pass hit reconstruction}
\label{sec:firstPassPixel}
The position of a pixel cluster along the transverse ($u$) and longitudinal ($v$) directions on the sensor is obtained
as follows. The procedure is described only for the case of the $u$ coordinate, but is identical for
the $v$ coordinate.
The cluster is projected onto the $u$-axis by summing the charge collected in pixels with the same
$u$-coordinate \cite{Cucciarelli:687475}. The result is referred to as a \textit{projected cluster}.
For projected clusters that are only one pixel large, the $u$-position is given by the
centre of that pixel, corrected for the Lorentz drift of the collected charge in the CMS magnetic field.
For larger projected clusters, the hit position $u_{\text{hit}}$ is determined using the relative charge in the
two pixels at each end of the projected cluster:
\begin{linenomath}
\begin{equation}
\label{eqn:pixelhit}
u_{\text{hit}} = u_{\text{geom}} + \frac{Q^u_{\text{last}} - Q^u_{\text{first}}}{2(Q^u_{\text{last}} + Q^u_{\text{first}})}
\left|W^u - W^u_{\text{inner}}\right| - \frac{L_u}{2},
\end{equation}
\end{linenomath}
where $Q_{\text{first}}$ and $Q_{\text{last}}$ are the charges collected in the first and last pixel of the
projected cluster, respectively; $u_{\text{geom}}$ is the position of the geometrical centre of the
projected cluster; and the parameter $L_u/2 = D\tan \Theta^u_L/2$ is the Lorentz shift along the $u$-axis, where $\Theta^u_L$ is the
Lorentz angle in this direction, and $D$ is the sensor thickness.
For the pixel barrel, the Lorentz shift is approximately $59\mum$.
The parameter $W^u_{\mathrm{inner}}$ is the geometrical width of the projected cluster, excluding its first and last
pixels. It is zero if the width of the projected cluster is less than three pixels.
The \textit{charge width} $W^u$ is defined as the width expected for the deposited charge, as estimated from the angle of the track with respect to the sensor, and equals
\begin{linenomath}
\begin{equation}
W^u = D \left| \tan \left ( \alpha^u - \pi/2 \right ) + \tan\Theta^u_L \right|,
\end{equation}
\end{linenomath}
where the angle $\alpha^u$ is the impact angle of the track relative to the plane of the sensor, measured
after projecting the track into the plane perpendicular to the $v$-axis. If no track is available,
$\alpha^u$ is calculated assuming that the particle producing the hit moved in a straight line from the
centre of the CMS detector.
The motivation for Eq.~(\ref{eqn:pixelhit}) is that the charge deposited by the traversing particle is expected to only partially cover the two pixels at each end of the projected cluster.
The quantity $W^u - W^u_{\text{inner}}$, which is expected to have a value between zero and twice the pixel pitch, (a modified version of Eq.~(\ref{eqn:pixelhit}) is used for any hits
that do not meet this
expectation), provides an estimate
of the total extension of charge into these two outermost pixels, while the relative
charge deposited in these two pixels provides a way to deduce how this total distance is shared between them.
The distance that the charge extends into each of the two pixels can thereby be deduced.
This gives the position of the two edges of the charge distribution, and the mean value of these edges, corrected for the Lorentz drift, equals the position of the cluster.
\subsubsection{Template-based hit reconstruction}
\label{sec:templatesPixel}
The high level of radiation exposure of the pixel detector can affect significantly the collection of charge
by the pixels during the detector's useful life.
This degrades particularly the performance of the standard hit reconstruction algorithm, sketched in the
previous section, as this algorithm only uses the end pixels of projected clusters when determining
hit positions.
The reconstructed positions of hits can be biased by up to $50\mum$ in highly irradiated sensors, and the
hit position resolution can be severely degraded.
In the template-based reconstruction algorithm, the observed distribution of the cluster charge is compared to expected projected distributions,
called \textit{templates}, to estimate the positions of hits~\cite{CMS_NOTE_2007_033}.
The templates are generated based on a large number of simulated particles traversing
pixel modules, which are modelled using the detailed \textsc{Pixelav}\xspace simulation~\cite{Swartz:2003ch,Chiochia:2004qh, Swartz:2005vp}.
Since the \textsc{Pixelav}\xspace program can describe the behaviour of irradiated sensors, new templates can be generated over the life of the detector to maintain the
performance of the hit reconstruction.
To allow the template-based algorithm to be applied to tracks crossing the silicon at various angles,
different sets of templates are generated for several ranges of the angle between the particle trajectory and
the sensor.
Working in each dimension independently, each pixel is subdivided into nine bins along the $u$ (or $v$) axis, where
each bin has a width of one-eighth of the size of a pixel and the end bins are centred on the pixel boundaries.
The $u$ (or $v$) coordinate of the point of interception of the particle trajectory and the pixel (defined as the
position at which the track crosses the plane that lies halfway between the front and back faces of the sensor)
is used to assign the interception point to one of the nine bins, $j$, indicating its location within the pixel.
The charge profile of the cluster produced by each particle is projected into an array that is 13
pixels long along the $u$ axis (or 23 pixels long along the $v$ axis) and centred on the intercepted pixel.
The resulting charge in each element $i$ of this array is recorded.
Only clusters with a charge below some specified angle-dependent maximum, determined from simulation, are used, as the charge distributions can be distorted by the significant ionization caused by energetic delta rays.
This procedure provides an accurate determination of the projected cluster distributions, determined by effects of geometry, charge drift, trapping, and charge induction.
In each dimension, the mean charge $S_{i,j}$ in bin $(i,j)$, averaged over all the particles, is then
determined.
In addition, the \rms charge distributions for the two projected pixels at the two ends
of the cluster are extracted, as are the charge in the projected pixel that has the
highest charge within the cluster, and the cluster charge, both averaged over all tracks.
The charge distribution of a reconstructed cluster, projected onto either the $u$ or $v$ axis,
can be described in terms of a charge $P_i$ in each pixel $i$ of the cluster. This can be compared to the
expected charge distributions $S_{i,j}$ stored in the templates, so as to determine the bin $j$
where the particle is likely to have crossed the sensor, and hence the best estimate of the
reconstructed hit position. This is accomplished by minimizing a $\chi^2$
function for several or all of the bins:
\begin{align}
\chi^2 (j) & = \sum_i \left ( \frac {P_i - N_j S_{i,j} }
{\Delta P_i } \right )^2,
\label{eq:fulleq}
\\
\intertext{with}
N_j & = \sum_i \frac {P_i}{ \left (\Delta P_i \right )^2 }
\Big / \sum_i \frac {S_{i,j}}{ \left (\Delta P_i \right )^2 }.
\end{align}
In this expression, $\Delta P_i$ is the expected \rms of a charge $P_i$ from the \textsc{Pixelav}\xspace
simulation and $N_j$ represents a normalization factor between the observed cluster charge and the
template.
While a sum over all the template bins yields an absolute minimum, different strategies can be used to optimize the performance of the algorithm as a function of allowed CPU time.
As described in Section~\ref{sec:TrackFit}, this $\chi^2$ is also used to reject outliers
during track fitting, in particular pixel hits on a track that are incompatible with
the distribution expected for the reconstructed track angle.
A simplified estimate of the position of a hit is performed for cluster projections consisting of a single pixel by correcting the
position of the hit for bias from
Lorentz drift and possible radiation damage.
The bias is defined by the average residual of all single-pixel clusters, as detailed below..
For cluster projections consisting of multiple pixels, the estimate of the hit position is further
refined. The charge template expected for a track crossing the pixel at an arbitrary position $r$, near
the best $j$ bin is approximated by the expression
$(1-r)S_{i,j-1} + r S_{i,j+1}$. Substituting this expression in place of $S_{i,j}$ in Eq.~(\ref{eq:fulleq}), and
minimizing $\chi^2$ with respect to $r$, yields an improved estimate of the hit position.
Finally, the above-mentioned hit reconstruction algorithm is
applied to the same \textsc{Pixelav}\xspace MC samples originally used to generate the templates.
Since the true hit position is known, any bias in the reconstructed hit position can be determined
and accounted for when the algorithm is run on collision data.
In addition, the \rms of the difference
between the reconstructed and true hit position is used to define the uncertainty in
the position of a reconstructed hit.
\subsection{Hit reconstruction in the strip detector}
\label{sec:striphit}
The data acquisition system of the strip detector \cite{ttdr} runs algorithms on off-detector electronics
(namely, on the modules of the \textit{front-end driver} (FED)~\cite{FED}) to subtract pedestals (the baseline signal
level when no particle is present) and common mode noise (event-by-event
fluctuations in the baseline within each tracker readout chip), and to perform zero-suppression.
Zero-suppression accepts a strip if its charge
exceeds the expected channel noise by at least a factor of five, or if both the strip and one of its neighbours have
a charge exceeding twice the channel noise.
As a result, information for only a small fraction of the channels in any given event is retained for offline
storage.
Offline, clusters are seeded by any channel passing zero-suppression that has a charge at least a factor of three greater than the corresponding
channel noise~\cite{Bayatian:2006zz}. Neighbouring strips are added to each seed, if their strip charge
is more than twice the strip noise. A cluster is kept if its total charge is a factor five larger than the cluster noise,
defined as $\sigma_{\text{cluster}} = \sqrt{\scriptstyle{ \sum_i \sigma_i^2}}$, where $\sigma_i$ is the noise for strip $i$,
and the sum runs over all the strips in the cluster.
The position of the hit corresponding to each cluster is determined from the charge-weighted average of
its strip positions, corrected by approximately 10\mum (20\mum) in the TIB (TOB)
to account for the Lorentz drift. One additional correction is made to compensate
for the fact that charge generated near the back-plane of the sensitive volume of the thicker silicon
sensors is inefficiently collected. This inefficiency shifts the cluster barycentre along the direction
perpendicular to the sensor
plane by approximately 10\mum in the 500\mum thick silicon, while its effect is negligible in the
320\mum thick silicon.
The inefficient charge collection from the sensor backplane is caused by the narrow time window during
which the APV25 readout chip~\cite{French:2001xb} integrates the collected charge,
and whose purpose is to reduce background from out-of-time hits.
The uncertainty in the hit position is usually parametrized as a function of the expected
width of the cluster obtained from the track angle (\ie, the `charge width' defined in
Section~\ref{sec:firstPassPixel}). However, in rare cases, when the observed width of a cluster exceeds the expected width by at least a factor of 3.5, and is incompatible with it, the uncertainty in the position is then
set to the `binary resolution', namely, the width of the cluster divided by $\sqrt{12}$. This broadening of the cluster is caused by capacitive coupling between the strips or energetic delta rays.
\subsection{Hit efficiency}
\label{s:hit_efficiency}
The hit efficiency is the probability to find a cluster in a given silicon sensor that has been traversed by a charged particle.
In the pixel detector, the efficiency is measured using isolated tracks originating from the primary vertex.
The \pt is required to be $>$1\GeV,
and the tracks are required to be reconstructed with a minimum of 11~hits measured in the strip detector.
Hits from the pixel layer under study are not removed when the tracks are reconstructed. To minimize any ensuing bias,
all tracks are required to have hits in the other two pixel layers, ensuring thereby
that they would be found even without using the studied layer.
A restrictive selection is set on the impact parameter to reduce false tracks and tracks from secondary interactions.
To avoid inactive regions and to allow for residual misalignment, track trajectories passing near the edges
of the sensors or their readout chips are excluded. Specifically, they must not pass within
0.6\mm (1.0--1.5\mm) of a sensor edge in the pixel endcap (barrel) or within 0.6\mm of the edge of a pixel
readout chip.
The efficiency is determined from the fraction of tracks to which either a hit is associated in the layer under study, or if it is found within $500\mum$
of the predicted position of the track. Given the high track density, only tracks that have no additional trajectories within 5\mm are considered so as to reduce
false track-to-cluster association.
The average efficiency for reconstructing hits is $>$99\%, as shown in Fig.~\ref{fig:hitEffPixel}(left),
when excluding the 2.4\% of the pixel modules known to be defective.
The hit efficiency depends on the instantaneous luminosity and on the trigger rate, as shown in Fig.~\ref{fig:hitEffPixel}(right).
The systematic uncertainty in these measurements is estimated to be 0.2\%.
Several sources of loss have been identified.
First, the limited size of the internal buffer of the readout chips cause a dynamic inefficiency that increases with the instantaneous luminosity and with the trigger rate.
Single-event upsets temporarily cause loss of information at a negligible rate of approximately two readout chips per hour.
Finally, readout errors signalled by the FED modules depend on the rate of beam induced background.
The efficiency in the strip tracker is measured using tracks that have a minimum of eight hits in the pixel and strip detectors. Where two hits are found in one of the closely-spaced double layers, which consist
of $r\phi$ and stereo modules, both hits are counted separately.
The efficiency in any given layer is determined using only the subset of tracks that have at least one hit in
subsequent layers, further away from the beam spot.
This requirement ensures that the particle traverses the layer under study, but
also means that the efficiency cannot be measured in the outermost layers of the TOB
(layer~6) and the TEC (layer~9).
To avoid inactive regions and to take account of any residual misalignment, tracks that cross a module
within five standard deviations from the sensor's edges, based on the uncertainty in the extrapolated track trajectory,
are excluded from consideration.
The efficiency is determined from the fraction of traversing tracks with a hit anywhere within the non-excluded region of a traversed module.
In the strip tracker, 2.3\% of the modules are excluded because of short circuits of the high voltage,
communication problems with the front-end electronics, or other faults.
Once the defective modules are excluded from the measurement, the overall hit efficiency is 99.8\%, as shown in Fig.~\ref{fig:hitEffStrip}.
This number is compatible with the 0.2\% fraction of defective channels observed during the construction of the strip tracker.
All defective components of the tracker are taken into account, both in the MC simulation of the detector and in the reconstruction of tracks.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\linewidth]{figs_2011/localReco/Raw_Efficiency_vs_Layers_2011.pdf}
\includegraphics[width=0.48\linewidth]{figs_2011/localReco/Raw_Efficiency_vs_InstLumi_2011.pdf}
\caption{The average hit efficiency for layers or disks in the pixel detector excluding defective modules (left), and
the average hit efficiency as a function of instantaneous luminosity (right). The peak luminosity ranged from 1 to $4\,\mathrm{nb^{-1}s^{-1}}$ during the data taking.}
\label{fig:hitEffPixel}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.5\linewidth]{figs_2011/localReco/StripHitEffTDRstyle.pdf}
\caption{Average hit efficiency for layers or disks in the strip tracker.
The black squares show the hit efficiency in all modules, and the red dots for modules included in the readout.}
\label{fig:hitEffStrip}
\end{figure}
\subsection{Hit resolution}
The hit resolution in the pixel and strip barrel sensors has been studied by measuring residuals, defined by the difference between the measured and the expected hit
position as predicted by the fitted track. Each trajectory is refitted excluding the
hit under study in order to minimize biases of the procedure.
The resolution of the pixel detector is measured from the RMS width of the hit residual distribution in the
middle of the three barrel layers, using only tracks with $\pt > 12$\GeV, for which multiple scattering
between the layers does not affect the measurement. The expected hit position in the middle layer, as
determined from the track trajectory, has an uncertainty that is dominated by the resolution of the hits
assigned to the track in the first and third barrel layers. Assuming that the three barrel layers all have
the same hit resolution $\sigma_\text{hit}$ and because they are approximately equally spaced in radius from
the z-axis of CMS, then this uncertainty is given by $\sigma_\text{hit}/\sqrt{2}$. Adding this in quadrature
with the uncertainty $\sigma_\text{hit}$ in the measured position of the hit in the middle layer, demonstrates
that the RMS width of the residual distribution is given by $\sigma_\text{hit}\sqrt{3/2}$. The measured hit
resolution $\sigma_\text{hit}$ in the $r\phi$ coordinate, as derived using this formula, is 9.4\mum.
The resolution in the longitudinal direction is shown in Fig.~\ref{fig:pixResZ}, and found to agree within
1\mum with MC simulation.
The longitudinal resolution depends on the angle of the track relative to the sensor.
For longer clusters, sharing of charge among pixels improves the resolution, with optimal resolution reached for interception angles of $\pm 30^\circ$.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\linewidth]{figs_2011/localReco/zres-vs-dip.pdf}
\caption{Resolution in the longitudinal ($z$) coordinate of hits in the barrel section of the pixel detector, shown
as a function of the incident angle of the track, which is defined as $90^\circ-\theta$, and equals the angle of the
track relative to the normal to the plane of the sensor. Data are compared with MC simulation for tracks with $\pt>12$\GeV.
}
\label{fig:pixResZ}
\end{figure}
Because of multiple scattering, the uncertainty in track position in the strip detector is usually much larger than the inherent resolution;
consequently, individual residuals of hits are not sensitive to the resolution.
However, the difference in a track's residuals for two closely spaced modules can be
measured with much greater precision.
Any offset in a track's position caused by multiple scattering will
be largely common to both modules.
A technique based on tracks passing through overlapping modules from the same tracker layer is
employed to compare the difference in residuals for the two measurements in the overlapping modules~\cite{tifPaper}.
The difference in hit positions ($\Delta x_\text{hit}$) can be compared to the difference in
predicted positions ($\Delta x_\text{pred}$) derived from the track trajectory, and their difference, fitted to a Gaussian function, provides a hit resolution convoluted with the uncertainty from
the trajectory propagation. The bias from translational misalignment between modules affects only
the mean of the Gaussian distribution, and not its \rms width.
As the two overlapping modules are expected to have the
same resolution, the resolution of a single sensor is determined by dividing this \rms width by $\sqrt{2}$.
Only tracks of \textit{high purity} (defined in Section~\ref{sec:TrackSelection})
are used for the above-described study. To reduce the uncertainty from multiple
Coulomb scattering, the track momenta are required to be $>$10\GeV.
The $\chi^2$ probability of the track fit is required to be $>$0.1\%,
and the tracks are required to be reconstructed using a minimum of six hits in the strip detector.
Tracks in the overlapping barrel modules are analysed only when the residual rotational
misalignment is less than 5\mum.
Remaining uncertainties from multiple scattering and rotational
misalignment for the overlapping modules are included as systematic
uncertainties of the measurement.
Sensor resolution depends strongly on the size of the cluster and on the pitch of the sensor.
The resolutions for the strip detector are shown in Table~\ref{tab:sstHitRes}, where they are compared to the predictions from MC simulation.
The resolution varies not only as a function of the cluster width, but also as a function of pseudorapidity, as the energy deposited by a charged particle
in the silicon depends on the angle at which it crosses the sensor plane. The resolution is worse in simulation than in data, implying the need for additional tuning
of the MC simulation. The results in the table are valid only for tracks with momenta $>$10\GeV. At lower momenta, the simulations indicate
that the resolution in hit position improves, but this is not important for tracking performance, as
the resolution of the track parameters for low-momentum tracks is dominated by the multiple scattering and by not the hit resolution.
\begin{table}[bt]
\centering
\topcaption{\label{tab:sstHitRes}
A comparison of hit resolution in the barrel strip detector as measured in data with the
corresponding prediction from simulation, for track momenta $>$10\GeV.
The resolution is given as function of both the barrel layer and the width
of the cluster in strips. Since the resolution is observed to vary with $\phi$ and $\eta$,
a range of resolution values is quoted in each case.}
\begin{tabular}{ccccccc}
\hline
Sensor & Pitch & & \multicolumn{4}{c}{Resolution [$\mu \mathrm{m}$] vs. width of cluster [strips]}\\ \cline{4-7}
layer & $(\mu \mathrm{m})$ & & width=1 & =2 & =3 & =4\\
\hline
\multirow{2}{*}{TIB 1--2} & \multirow{2}{*}{ 80} &
Data & 11.7--19.1 & 10.9--17.9 & 10.1--18.1 & \\
&& MC & 14.5--20.5 & 15.0--19.8 & 14.0--20.6 & \\
\hline
\multirow{2}{*}{TIB 3--4} & \multirow{2}{*}{ 120} &
Data & 20.9--29.5 & 21.8--28.8 & 20.8--29.2 & \\
&& MC & 26.8--30.4 & 27.6--30.8 & 27.9--32.5 & \\
\hline
\multirow{2}{*}{TOB 1--4} & \multirow{2}{*}{ 183} &
Data & & 23.4--40.0 & 32.3--42.3 & 16.9--28.5 \\
&& MC & & 42.5--50.5 & 43.0--48.6 & 18.8--35.2 \\
\hline
\multirow{2}{*}{TOB 5--6} & \multirow{2}{*}{ 122} &
Data & & & 18.4--26.6 & 11.8--19.4 \\
&& MC & & & 26.1--29.5 & 17.8--21.6 \\
\hline
\end{tabular}
\end{table}
\section{Track reconstruction}
\label{sec:trackReco}
Track reconstruction refers to the process of using the hits, obtained from the local reconstruction
described in Section~\ref{sec:localReco}, to obtain estimates for the momentum and position
parameters of the charged particles responsible for the hits (tracks). As part of this process, a
translation between the local coordinate system of the hits and the global coordinate system of the
track is necessary. This translation takes into account discrepancies between the assumed and
actual location and surface deformation of detector elements as found through the alignment
process~\cite{Chatrchyan:2014wfa}. In addition, the uncertainty in the detector element location is
added to the intrinsic uncertainty in the local hit position.
Reconstructing the trajectories of charged particles is a computationally challenging
task. An overview of the difficulties and solutions can be found in review
articles~\cite{Regler:1996yw,Mankel:2004yv,Strandlie:2010zz}.
The tracking software at CMS is commonly referred to as the Combinatorial Track Finder (CTF),
which is an adaptation of the combinatorial Kalman filter~\cite{Billoir:1989mh,Billoir:1990we,Mankel:1997dy}, which in turn
is an extension of the Kalman filter~\cite{Fruhwirth:1987fm} to allow pattern recognition and
track fitting to occur in the same framework.
The collection of reconstructed tracks is produced by multiple passes (iterations) of the CTF track
reconstruction sequence, in a process called \textit{iterative tracking}.
The basic idea of iterative tracking is that the initial iterations search for tracks that
are easiest to find (\eg, of relatively large \pt, and produced near the interaction region).
After each iteration, hits associated with tracks are removed, thereby reducing the
combinatorial complexity, and simplifying subsequent iterations in a search for more difficult
classes of tracks (\eg, low-\pt, or greatly displaced tracks).
The presented results reflect the status of the software in use from May through August, 2011, which is
applied in a series of six iterations of the track reconstruction algorithm. Later versions of the
software retain the same basic structure but with different iterations and tuned values for the
configurable parameters to adapt to the higher pileup conditions. Iteration 0, the
source of most reconstructed tracks, is designed for prompt tracks (originating near the pp interaction point)
with $\pt>0.8$\GeV that
have three pixel hits. Iteration 1 is used to recover prompt tracks that have only two pixel hits.
Iteration 2 is configured to find low-\pt prompt tracks. Iterations 3--5
are intended to find tracks that originate outside the beam spot (luminous region of the pp
collisions) and to recover tracks not found in
the previous iterations. At the beginning of each iteration, hits associated with high-purity
tracks (defined in Section~\ref{sec:TrackSelection}) found in previous iterations are excluded
from consideration (masked).
Each iteration proceeds in four steps:
\begin{itemize}
\item{Seed generation provides initial track candidates found using only a few (2 or 3) hits.
A seed defines the initial estimate of the trajectory parameters and their uncertainties.}
\item{Track finding is based on a Kalman filter. It extrapolates
the seed trajectories along the expected flight path of a charged particle, searching for additional hits
that can be assigned to the track candidate.}
\item{The track-fitting module is used to provide the best possible estimate of the parameters
of each trajectory by means of a Kalman filter and smoother.}
\item{Track selection sets quality flags, and discards tracks that fail certain specified criteria.}
\end{itemize}
The main differences between the six iterations lie in the configuration of the seed generation and the final
track selection.
\subsection{Seed generation}
\label{sec:SeedGeneration}
The seeds define the starting trajectory parameters and associated uncertainties of potential
tracks. In the quasi-uniform magnetic field of the tracker, charged particles follow helical paths and
therefore five parameters are needed to define a
trajectory. Extraction of these five parameters requires either three 3-D hits, or two 3-D hits and a
constraint on the origin of the trajectory based on the assumption that the particle originated near the
beam spot. (A `3-D hit' is defined to be any hit that provides a 3-D position measurement).
To limit the number of hit combinations, seeds are required to satisfy
certain weak restrictions, for
example, on their minimum \pt and their consistency with originating from the pp
interaction region.
In principle, it is possible to construct seeds in the outermost regions of the tracker, where the track density is
smallest, and then construct track candidates by searching inwards from the seeds for additional hits
at smaller distances from the beam-line. However, there are several reasons why an alternative approach, of
constructing seeds in the inner part of the tracker and building the track candidates outwards, has been chosen
instead.
First, although the track density is much higher in the inner region of the tracker, the high granularity
of the pixel detector ensures that the channel occupancy (fraction of channels that are hit) of the inner pixel layer is much lower than
that of the outer strip layer. This can be seen in Fig.~\ref{fig:TrackerOccupancy},
which shows the mean channel occupancy in strip and pixel sensors
in data collected with a `zero-bias' trigger, (which took events from randomly selected non-empty LHC bunch
crossings). This data had a mean of about nine pp interactions per bunch crossing.
The channel occupancy
is 0.002--0.02\% in the pixel detector and 0.1--0.8\% in the strip detector.
Second, the pixel
layers produce 3-D spatial measurements, which provide more constraints and
better estimates of trajectory parameters. Finally, generating seeds in the inner tracker leads to
a higher efficiency for reconstructing tracks. Although most high-\pt muons traverse the
entire tracker, a significant fraction of the produced pions interact inelastically in the
tracker~(Fig.~\ref{fig:PionSurvival}). In addition, many electrons lose a significant
fraction of their energy to bremsstrahlung radiation in the tracker. Therefore, to ensure high
efficiency, track finding begins with trajectory seeds created in the inner region of the tracker.
This also facilitates reconstruction of low-momentum tracks that are deflected by the strong magnetic field
before reaching the outer part of the tracker.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.9\textwidth]{figs_2011/cmsTracker/occupancy_map_blueyellow_LE_CMSlabel.pdf}
\caption{Channel occupancy (labelled by the scale on the right) for CMS silicon detectors in events taken
with unbiased triggers with an average of nine pp interactions per beam crossing, displayed as a function of $\eta$,
$r$, and $z$.}
\label{fig:TrackerOccupancy}
\end{figure}
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.6\textwidth]{figs_2011/cmsTracker/PionSurvivalProbability.pdf}
\caption {Fraction of pions produced with $\abs{\eta}<2.5$ that do not undergo a nuclear interaction in the
tracker volume, as a function of the number of traversed layers.}
\label{fig:PionSurvival}
\end{figure}
Seed generation requires information on the position of the centre of the reconstructed beam spot,
obtained prior to track finding using the method described in Section~\ref{sec:beamspot}. It
also requires the locations of primary vertices in the event, including
those from pileup events.
This information is obtained by running a very fast track and vertex reconstruction algorithm,
described in Section~\ref{sec:pixeltrackvertex}, that uses only hits from the pixel detector.
The tracks and primary vertices found with this algorithm are known as \textit{pixel tracks} and
\textit{pixel vertices}, respectively.
The seed generation algorithm is controlled by two main sets of parameters: \textit{seeding layers} and
\textit{tracking regions}. The seeding layers are pairs or triplets of detector layers in which hits are searched
for. The tracking regions specify the limits on the acceptable track parameters, including the
minimum \pt, and the maximum transverse and longitudinal distances of closest approach to the assumed production point of the particle, taken to be located either at the centre of the reconstructed beam spot or at a
pixel vertex.
If the seeding layers correspond to pairs of detector layers, then seeds are constructed using one hit in each layer.
A hit pair is accepted as a seed if the corresponding track parameters are
consistent with the requirements of the tracking region. If the seeding layers correspond to triplets of
detector layers, then, after pairs of hits are found in the two inner layers of each triplet, a
search is performed in the outer detector layer for another hit. If the track parameters derived from the
three hits are compatible with the tracking region requirements, the seed is accepted. It is
also possible to check if the hits associated with the seed have the expected charge distribution from the
track parameters: a particle that enters the detector at a grazing angle will have a larger
cluster size than a particle that enters the detector at a normal angle. Requiring the reconstructed
charge distribution to match the expected charge distribution can remove many fake seeds.
In simulated $\ttbar$ events at $\sqrt{s} = 7$\TeV, more than 85\% of the charged particles produced
within the geometrical acceptance of the tracker $(\abs{\eta}<2.5)$ cross three pixel layers and
can therefore be reconstructed starting from trajectory seeds
obtained from triplets of pixel hits. Nevertheless, other trajectory seeds are also needed, partially to
compensate for inefficiencies in the pixel detector (from gaps in coverage, non-functioning modules, and saturation
of the readout), and partially to reconstruct particles not produced directly at the pp
collision point (decay products of strange hadrons, electrons from photon conversions, and particles from
nuclear interactions). To improve the speed and quality of the seeding algorithm, only 3-D space
points are used, either from a pixel hit or a \textit{matched} strip hit. Matched strip hits are obtained
from the closely-spaced double strip layers, which are composed of two sensors mounted back-to-back,
one providing an $r\phi$ view and one providing a stereo view (rotated by 100\ensuremath{\,\text{mrad}}\xspace relative to the other,
in the plane of the sensor). The `$r\phi$' and `stereo hits' in such a layer are combined into a
matched hit, which provides a 3-D position measurement. Table~\ref{tab:IterativeSeeds} shows the seeding requirements for
each of the six tracking iterations. The seeding layers listed in this table are defined as follows:
\begin{itemize}
\item{Pixel triplets are seeds produced from three pixel hits.
These seeds are used to find most of the tracks corresponding to promptly produced charged
particles. The three precise 3-D space points provide seeds of high quality and with well-measured
starting trajectories. A mild constraint on the compatibility of these trajectories with the centre of the beam spot is employed, to remove seeds inconsistent with promptly produced particles. Also, the
charge distribution of each pixel hit is
required to be compatible with that expected for the crossing angle of the seed trajectory and the
corresponding sensor.}
\item{Mixed pairs with vertex constraint are seeds that use two hits and a third
space-point given by the location of a pixel vertex. If more than one pixel vertex is found in
an event, which often happens because of pileup, all are considered in turn. The pixel vertices are required to pass
quality criteria; the most important is that a vertex must contain at least four
pixel tracks. The two hits used for these seeds can be provided by the pixel tracker, or by the two inner
rings of the three inner TEC layers, where the TEC layers are used to increase coverage in the
very forward regions.}
\item{Mixed triplets are seeds produced from three hits formed from a combination of pixel
hits and matched strip hits. Each triplet contains between one and three pixel hits and $< 3$
strip hits. This iteration is implemented for finding displaced tracks and prompt tracks that do not
have three hits in the pixel detector.
The beam spot related constraint is less restrictive, providing higher efficiency for finding
tracks arising from decays of hadrons containing s, c, or b quarks, photon conversions, and
nuclear interactions.}
\item{Strip pairs are seeds constructed using two matched hits from the strip detector.
Iteration 4 uses the two inner TIB layers and rings 1--2 of the TID/TEC, which are the same
strip layers used in Iteration 3. In Iteration 5, hits from the two inner TOB layers and ring 5
of the TEC are used for seeds. These two iterations have even weaker constraints on the
compatibility of the seed trajectory with the centre of the beam spot than has
Iteration 3, and they do not require pixel hits. These iterations are therefore useful for
finding tracks produced outside of the pixel detector volume or tracks that do not leave hits in the
pixel detector.}
\end{itemize}
\begin{table}[htbp]
\centering
\topcaption{\label{tab:IterativeSeeds} The configuration of the track seeding for each of the six
iterative tracking steps. Shown are the layers used to seed the tracks, as well as the requirements on the minimum \pt and
the maximum transverse ($d_0$) and longitudinal ($z_0$) impact parameters relative to the centre of the beam spot.
The Gaussian standard deviation corresponding to the length of the beam spot along the $z$-direction is $\sigma$. The asterisk symbol
indicates that the longitudinal impact parameter is calculated relative to a pixel vertex
instead of to the centre of the beam spot.}
\begin{tabular}{cclll}
\hline
Iteration & Seeding layers & \multicolumn{1}{c}{\pt (\GeV)} & \multicolumn{1}{c}{$d_0$ (cm)} & \multicolumn{1}{c}{$\abs{z_0}$} \\
\hline
0 & Pixel triplets & $>$0.8 & $<$0.2 & ${<}3\sigma$ \\
1 & Mixed pairs with vertex& $>$0.6 & $<$0.2 & ${<}0.2\cm^*$ \\
2 & Pixel triplets & $>$0.075 & $<$0.2 & ${<} 3.3\sigma$ \\
3 & Mixed triplets & $>$0.35 & $<$1.2 & ${<}10$\cm \\
4 & TIB 1+2 \& TID/TEC ring 1+2 &$>$0.5 & $<$2.0 & ${<}10$\cm \\
5 & TOB 1+2 \& TEC ring 5 & $>$0.6 & $<$5.0 & ${<}30$\cm \\
\hline
\end{tabular}
\end{table}
\subsection{Track finding}
\label{sec:TrackFinding}
The track-finding module of the CTF algorithm is based on the Kalman filter
method~\cite{Fruhwirth:1987fm,Billoir:1989mh,Billoir:1990we,Mankel:1997dy}.
The filter
begins with a coarse estimate of the track parameters provided by the trajectory seed, and then
builds track candidates by adding hits from successive detector layers, updating the
parameters at each layer. The information needed at each layer includes the location and
uncertainty of the detected hits, as well as the amount of material crossed, which is used to estimate
the effects of multiple Coulomb scattering and energy loss. The track finding is implemented in
the four steps listed below.
The first step (navigation) uses the parameters of the track candidate, evaluated at the current layer,
to determine which adjacent layers of the detector can be intersected through an extrapolation of the
trajectory, taking into account the current uncertainty in that trajectory.
The navigation service can be configured to propagate along or opposite to the momentum vector, and
uses a fast \textit{analytical propagator} to find the intercepted layers. The analytical propagator assumes a
uniform magnetic field, and does not include effects of multiple Coulomb scattering or energy
loss. With these assumptions, the track trajectory is a perfect helix, and the propagator can therefore
extrapolate the trajectory from one layer to the next using rapid analytical calculations.
In the barrel, the cylindrical geometry makes navigation particularly easy, since the extrapolated
trajectory can only intercept the layer adjacent to the current one. In the endcap and
barrel-endcap transition regions, navigation is more complex, as the crossing from one layer
does not uniquely define the next one.
The second step involves a search for compatible silicon modules in the layers returned by the
navigation step. A module is considered compatible with the trajectory if the position at which
the trajectory intercepts the module surface is no more than some given number (currently three) of
standard deviations outside the module boundary. The propagation of the trajectory parameters, and of the
corresponding uncertainties, to the sensor surface involves mathematical operations and routines that are
generally quite time-consuming~\cite{Strandlie:927379}. Hence, the code responsible for searching for
compatible modules has been optimized to limit the number of sensors that are
considered, while preserving an efficiency of $>$99\% in finding the relevant
sensors. A complication is that the design of the CMS tracker is such that sensors
often slightly overlap their neighbours,
meaning that a particle can cross two sensors in the same layer. This possibility is accommodated by
dividing the compatible modules in each layer into groups of mutually exclusive modules,
defined such that if a particle passes through one member of a group, it is not physically possible for
it to pass through a second member of the same group. Any two modules that have some overlap are not mutually
exclusive, and are therefore assigned to different groups. This feature is used in the third and fourth
steps of the track finding, described next.
The third step forms groups of hits, each of which is defined by the collection of all the hits from one of the
module groups. A configurable parameter provides the possibility of adding a ghost hit
to represent the possibility that the particle failed to produce a hit in the module group, for example,
as a result of module inefficiency. The hit positions and uncertainties are refined using the trajectory
direction on the sensor surface, to calculate more accurately the Lorentz drift of the ionization-charge
carriers inside the silicon bulk. A $\chi^2$ test is used to check which of the hits are
compatible with the extrapolated trajectory. The current (configurable) requirement is
$\chi^2<30$ for one degree of freedom (dof). The $\chi^2$ calculation takes into account both the hit and trajectory uncertainties.
In the endcap regions and the barrel-endcap transition regions, the extrapolation distances and the
amount of material traversed are generally greater,
with correspondingly larger uncertainties in the trajectory, and the probability
of finding spurious hits compatible with the track tends therefore to be greater.
The fourth and last step is to update the trajectories. From each of the original track candidates,
new track candidates are formed by adding exactly one of the compatible hits from each module grouping
(where this hit may be a ghost hit).
As the modules in a given group are mutually exclusive,
it would not be expected that a track would have more than one hit contributing from each group.
The trajectory parameters for each new candidate are then
updated at the location of the module surface, by combining the information from the added hits with the extrapolated
trajectory of the original track candidate.
For the above second, third, and fourth steps of the procedure, a more accurate \textit{material propagator}
is used when extrapolating the track trajectory, which includes the effect of the material in the tracker. This differs from the method of the
simple analytical propagator, in that it increases the uncertainty in the trajectory parameters according
to the predicted \rms scattering angle in the tracker material. It also adjusts the momentum of
the trajectory by the predicted mean energy loss of the Bethe--Bloch equation. Since all
detector material is assumed to be concentrated in the detector layers, the track propagates along a simple helix
between the layers, allowing the material propagator to extrapolate the track analytically. The ghost
hits include the effect of material without providing position information to the propagator.
All resulting track candidates found at each layer are then propagated to the next compatible
layers, and the procedure is repeated until a termination condition is satisfied. However, to avoid a rapid
increase in the number of candidates, only a limited number (default is 5) of the candidates
are retained at each step, with the best candidates chosen based on the normalized $\chi^2$ and a bonus given for
each valid hit, and a penalty for each ghost hit. The standard termination conditions are if a
track reaches the end of the tracker or contains too many missing hits (limit is $N_\text{lost}$),
or if its \pt drops below a user specified value.
The number of missing hits on a track is equal
to the number of ghost hits, except that hits not found due to attributable known detector
conditions, for example, if a detector module is turned off, are not counted. The building of a
trajectory can also be terminated when the uncertainty in its parameters falls below a given
threshold or the number of hits is above a threshold; these kinds of termination conditions tend to
be used only in the high-level trigger (HLT), where the required accuracy on track parameters
is often reached after 5 or 6 hits are added to the track candidate, and the continuation of the
track building would correspond to a waste of CPU time.
When the search for hits in the outward direction reveals a minimum number of valid hits
($N_\text{rebuild}$), an inwards search is initiated for
additional hits. Otherwise, the track candidate remains as formed.
The inwards search starts by taking all of the hits assigned to the track, excluding those belonging
to the track seed, and using them to fit the track trajectory. In case this exclusion of the seeding hits leaves
fewer than $N_\text{rebuild}$ hits to fit, some of the seeding hits are also used
(taking first the outer contributions) so as to obtain at least $N_\text{rebuild}$ hits.
Then, as in the outward track building, the
trajectory is propagated inwards through the seeding layers and then further, until the inner
edge of the tracker is reached or too many ghost hits are found. There are three reasons for
this inward search. First, additional hits can be found in the seeding layers (for example,
from overlapping sensors). Second, hits can be found in layers closer to the interaction region
than the seeding layers. Third, when strip layers are used in seeding, matched hits are used to
increase computational speed and reduce the combinations of hits available for seeding. However,
some $r\phi$ or stereo hits are not part of any matched hit. While these hits are
not available during seeding, they can be found during the inward track building process.
The effect of the inward search is an increase in the mean number of hits per track by 0.15,
(\ie, a 1\% increase relative to a total of $\approx$14 hits), which translates to a better
signal-to-background ratio, impact parameter resolution, and \pt resolution, with maximum improvements of
2\%, 1\%, and 0.5\%, respectively.
The track of a single charged particle can be reconstructed more than once, either starting from
different seeds, or when a given seed develops into more than one track candidate. To remedy this
feature, a trajectory cleaner is applied after all the track candidates in a given iteration
have been found. The trajectory cleaner calculates the fraction of shared hits between two track
candidates:
$f_\text{shared} = \frac{N^\text{hits}_\text{shared}}{\min(N^\text{hits}_1,N^\text{hits}_2)}$ where $N^\text{hits}_1$
and $N^\text{hits}_2$ are, respectively, the number of hits used in forming the first (second) track candidate. If this fraction
exceeds the (configurable) value of 19\% (determined empirically), the trajectory cleaner removes the track with the
fewest hits; if both tracks have the same number of hits, the track with the largest
$\chi^2$ value is discarded. The procedure is repeated iteratively on all pairs of
track candidates. The same algorithm is applied when tracks from the six iterations are
combined into a single track collection.
The requirements applied during the track-finding stage are shown in Table~\ref{tab:IterativeFinding}
for each tracking iteration. In addition to the requirement on $N_\text{lost}$,
the completed track candidates must also pass requirements on the minimum number of hits $(N_\text{hits})$
and minimum track \pt.
The minimum \pt requirements have very little effect, as they are
weaker than those applied to the seeds, given in Table~\ref{tab:IterativeSeeds}. Since the later iterations
do not have strong requirements that the tracks originate close to the centre of the beam spot, the probability of
random hits forming tracks increases, which leads to more fake tracks and greater usage of CPU time. To
compensate for this tendency, the criteria for the minimum number of hits, and maximum number of lost hits,
are tightened in the later iterations.
\begin{table}[htbp]
\centering
\topcaption{\label{tab:IterativeFinding} Selection requirements applied to track candidates
during the six iterative steps of track finding, the minimum \pt,
the minimum number of hits $N_\text{hits}$, and the maximum number of missing hits
$N_\text{lost}$. Also shown is the minimum number of hits needed to be found in the outward
track building step to trigger the inward track building step $N_\text{rebuild}$,
although candidates failing this requirement are not rejected.}
\begin{tabular}{ccccc}
\hline
Iteration & \pt (\GeVns) & $N_\text{hits}$ & $N_\text{lost}$ & $N_\text{rebuild}$ \\
\hline
0 & 0.3 & 3 & 1 & 5 \\
1 & 0.3 & 3 & 1 & 5 \\
2 & 0.1 & 3 & 1 & 5 \\
3 & 0.1 & 4 & 0 & 5 \\
4 & 0.1 & 7 & 0 & 5 \\
5 & 0.1 & 7 & 0 & 4 \\
\hline
\end{tabular}
\end{table}
\subsection{Track fitting}
\label{sec:TrackFit}
For each trajectory, the track-finding stage yields a collection of hits and an estimate of
the track parameters. However, the full information about the trajectory is only available at the final hit of the
trajectory (when all hits are known). Furthermore, the estimate can be biased by constraints, such as a beam spot constraint
applied to the trajectory during the seeding stage. The trajectory is therefore refitted using a
Kalman filter and smoother.
The Kalman filter is initialized at the location of the innermost hit, with the trajectory estimate obtained
by performing a Kalman filter fit to the innermost hits (typically four) on the track.
The corresponding covariance matrix is scaled up by a large factor (10 for the last iteration and 100 for the
other iterations) in order to
limit the bias. The fit then proceeds in an iterative way through the full list of hits, from the inside
outwards, updating the track trajectory estimate sequentially with each hit. For each
valid hit, the estimated hit position uncertainty is reevaluated
using the current values of the track parameters. In the case of pixel hits, the estimated hit
position is also reevaluated.
This first filter is followed by the smoothing stage, whereby a second filter is initialized with the
result of the first one (except for the covariance matrix, which is scaled by a large factor),
and is run backward towards the beam-line.
The track parameters at the surface associated with any of its hits, can then be obtained from the
weighted average of the track parameters of these two filters, evaluated on this same surface, as
one filter uses information from all the hits found before, and the other uses information from all the hits
found after the surface. This provides the optimal track parameters at any point, including the
innermost and outermost hit on the track, which are used to extrapolate the trajectory to the
interaction region and to the calorimeter and muon detectors, respectively. A configurable parameter
determines whether the silicon strip matched hits are used as is or split into their component
$r\phi$ and stereo hits. For the standard offline reconstruction, the split hits are used to improve
the track resolution, while for the HLT, the matched hits are used to improve speed.
To obtain the best precision, this filtering and smoothing procedure uses a \textit{Runge--Kutta propagator}
to extrapolate the trajectory from one hit to the next.
This not only takes into account the effect of material, but it also accommodates an inhomogeneous
magnetic field. The latter means that the particle may not move along a perfect helix, and its equations
of motion in the magnetic field must therefore be solved numerically. To do so, the Runge--Kutta
propagator divides the distance to be extrapolated into many small
steps. It extrapolates the track trajectory over each of these steps in turn, using
a well-known mathematical technique for solving first-order differential equations, called the fourth-order
Runge--Kutta method, so called because it is accurate to fourth order in the step size. The optimal step size is chosen
automatically, according to how non-linear the problem is. This automatic determination of step size
employs the method \cite{Cash:1990:VOR:79505.79507},
which is based on how well the fourth and fifth order Runge--Kutta predictions agree with
each other. Use of the Runge--Kutta propagator is most important in the region $\abs{\eta} > 1$, where
the magnetic field inhomogeneities are greatest. For example, in this region, tracks
fitted using the simple material propagator are biased by up to 1\% for particles with $\pt = 10\GeV$.
This bias is almost completely eliminated when using the Runge--Kutta propagator. To assure an accurate
extrapolation of the track trajectory, the Runge--Kutta propagator uses a detailed map of the magnetic
field, which was measured before LHC collisions to a precision of $<0.01\%$.
Estimates of the track trajectory at any other points, such as the point of closest approach to the beam-line,
can be obtained by extrapolating the trajectory evaluated at the nearest hit to that very point. This extrapolation
also uses the Runge--Kutta propagator.
After filtering and smoothing, a search is made for spurious hits (outliers), incorrectly
associated to the track. Such hits can be related to an otherwise well-defined track,
\eg, from $\delta$-rays, or unrelated, such as hits from nearby tracks or electronic noise.
Two methods are used to find outliers. One uses the
measured residual between a hit and the track to reject hits whose $\chi^2$ compatibility with the track
exceeds a
configurable threshold (20 for Iterations 0--4 and 30 for Iteration 5). While a $\chi^2$ requirement
of 30 on each hit is already applied during track finding, the outlier rejection criterion provides a more
powerful restriction as it uses information from the full fit~\cite{Fruhwirth:1987fm}. The other method
calculates a probability that a pixel hit is consistent with the track, taking into account the
charge distribution of the pixel hit, which generally comprises several pixel channels. This probability
corresponds to the $\chi^2$ defined in Eq.~(\ref{eq:fulleq}). After removing the outlier,
the track is again filtered and smoothed and another check for outliers is made. This continues
until no more outliers are found. In cases where removing an outlier results in two consecutive ghost
hits, the track is terminated and the remaining outer hits discarded (although not used, a configurable parameter is available
to allow the track fitting to continue). If a track is found to have less than three hits after outlier
rejection or for the track fitting to fail, the track is discarded (although not used, a configurable parameter
is available to return the original track).
The default value of 20 for the $\chi^2$ requirement is chosen to reject a significant fraction of
outliers, while removing few genuine hits. With this value, approximately 20\% of the spurious outliers
are removed from tracks reconstructed in high-density dijet events, whereas ${<} 0.2\%$ of
the good hits are removed.
\subsection{Track selection}
\label{sec:TrackSelection}
In a typical LHC event containing jets, the track-finding procedure described above yields a
significant fraction of fake tracks, where a fake track is defined as a reconstructed track not
associated with a charged
particle, as defined in Section~\ref{sec:trackPerformance}. The fake
rate (fraction of reconstructed tracks that are fake) can be reduced substantially through
quality requirements. Tracks are selected on the basis of the number of layers that have hits, whether
their fit yielded a good $\chi^2/\mathrm{dof}$, and how compatible they are with originating from
a primary interaction vertex. If several primary vertices are present in the event,
as often happens due to pileup, all are considered. To optimize the performance, several
requirements are imposed as a function of the track $\eta$ and \pt, and on the number of layers
$(N_\text{layers})$ with an
assigned hit (where a layer with both $r\phi$ and stereo strip modules is counted as a single layer). The selection criteria are as follows.
\begin{itemize}
\itemsep 0pt
\item A requirement on the minimum number of layers in which the track has at least one associated hit. This differs from
selections based on the number of hits on the track, because more than one hit in a given layer can be assigned to a track,
as in the case of layers with overlapping sensors or double-sided layers in which two sensors are mounted back-to-back.
\item A requirement on the minimum number of layers in which the track has an associated 3-D hit (\ie, in the pixel
tracker or matched hits in the strip tracker).
\item A requirement on the maximum number of layers intercepted by the track containing no assigned hits, not
counting those layers inside its innermost hit or outside its outermost hit, nor those layers where no hit was expected because
the module was known to be malfunctioning.
\item $\chi^2/\mathrm{dof} < \alpha_0 N_\text{layers}$.
\item $\abs{d_0^\mathrm{BS}}/\delta d_0 < \left( \alpha_3 N_\text{layers}\right)^\beta$.
\item $\abs{z_0^\mathrm{PV}}/\delta z_0 < \left( \alpha_4 N_\text{layers}\right)^\beta$.
\item $\abs{d_0^\mathrm{BS}}/\sigma_{d_0}(\pt) < \left( \alpha_1 N_\text{layers}\right)^\beta$.
\item $\abs{z_0^\mathrm{PV}}/\sigma_{z_0}(\pt,\eta) < \left( \alpha_2 N_\text{layers}\right)^\beta$.
\end{itemize}
The parameters $\alpha_i$ and $\beta$ are configurable constants. The track's impact parameters
are $d_0^\mathrm{BS}$ and $z_0^\mathrm{PV}$, where $d_0^\mathrm{BS}$ is the distance from the centre of the
beam spot in the plane transverse to the beam-line and $z_0^\mathrm{PV}$ is the distance along the beam-line from the closest
pixel vertex. These
pixel vertices, described in Section~\ref{sec:pixeltrackvertex}, are required to have at least three
pixel tracks and if no pixel vertices meet this requirement, then $z_0^\mathrm{PV}$ is required to be within 3$\sigma$ of
the $z$-position of the centre of the beam spot, where $\sigma$ is the Gaussian standard deviation corresponding to the length of the beam spot in the $z$-direction.
The above selection criteria include requirements on the transverse $\abs{d_0^\mathrm{BS}}/\delta d_0$ and
longitudinal $\abs{z_0^\mathrm{PV}}/\delta z_0$ impact parameter significances of the track, where
the impact parameter uncertainties, $\delta d_0$ and $\delta z_0$, are calculated from the covariance
matrix of the fitted track trajectory. A second pair of requirements is also imposed on these
significances, but calculated differently, with
the uncertainties in the impact parameters being parametrized in terms of \pt and polar angle of the track:
$\sigma(d_0) = \sigma(z_0\sin\theta) = a \oplus \frac{b}{\pt}$, where $\oplus$ represents the
sum in quadrature and $a$ and $b$ are
parameters. Their nominal values are $a=30\mum$ and $b=10\mum\GeV$, but
$b$ increases to $100\mum\GeV$ for the
\textit{loose} and \textit{tight} selection criteria used (and defined below) in Iterations 0 and 1.
The fraction of fake tracks decreases roughly exponentially as a function of the
number of layers in which the track has associated hits: $dN_{\rm fake}/dN_\text{layers} \sim \exp(-\omega N_\text{layers}) $, with
$\omega$ in the range 0.9--1.3 depending on the \pt of the track. As a consequence, weaker selection criteria can be applied for
tracks having many hit layers, which is the reason for the chosen selection criteria.
For tracks with hits in at least 10 layers, the selection requirements on $\chi^2$ and impact parameters
are found to reject no tracks. However, the criteria become far more stringent for tracks with relatively
few hit layers.
The above quality criteria were initially optimized as a function of track \pt and $N_\text{layers}$,
so as to maximize the quality $Q(\rho) = {s}/{\sqrt{s+\rho b}}$, where $s$ is the
number of selected genuine (non-fake) tracks, $b$ is the number of selected fake tracks and $\rho\simeq 10$
inflates the importance of the fake tracks to achieve low fake rates (below 1\% for \PYTHIA QCD events with
$\hat{p}_{\mathrm{T}}$ of the two outgoing partons in the range 170--230\GeV). As data taking conditions have evolved,
the parameters have been adjusted to maintain high efficiency and low fake rate.
The track selection criteria for each iteration are given in Table~\ref{tab:TrackSelection}.
The \textit{loose} criteria denote the minimum requirements for a track to be kept in the general track collection.
The \textit{tight} and \textit{high-purity} criteria provide progressively more stringent requirements, which reduce the
efficiency and fake rate.
In general,
high-purity tracks are used for scientific analysis, although in cases where efficiency is essential and purity is not a major
concern, the loose tracks can be used. The criteria for the initial tracking iterations emphasise compatibility
with originating from a primary vertex as a means of assuring quality, while the criteria used for the later iterations
rely on other measures of track quality such as fit $\chi^2$ and the number of hits,
ensuring thereby that they are still useful for selecting displaced tracks. This matches the seeding and
track-finding requirements shown in Tables~\ref{tab:IterativeSeeds}--\ref{tab:IterativeFinding}, and is aligned with the
goals for the six iterations.
After the track selection is complete, the tracks found by each of the six iterations are
merged into a single collection.
\begin{table}[htbp]
\centering
\topcaption{\label{tab:TrackSelection} Parameter values used in selecting tracks reconstructed by each
of the six iterative tracking steps. The first table shows the three requirements on the number
of layers that contain hits assigned to tracks and the parameter $\alpha_0$
that controls selection criteria based on $\chi^2/\mathrm{dof}$. The second table shows the parameters $\alpha_i$ and $\beta$
that define compatibility of impact parameters with the interaction point. Each parameter has
three entries, corresponding to the loose (L), tight (T), and high-purity (H) selection requirements.
Iterations 2 and 3 use two paths that emphasise track quality (Trk) or primary-vertex compatibility (Vtx).
A track produced by these iterations is retained if it passes either of these criteria.
}
\begin{tabular}{c|ccc|ccc|ccc|ccc}
\hline
\multicolumn{1}{c|}{\multirow{2}{*}{Iteration}} & \multicolumn{3}{c|}{Min layers} & \multicolumn{3}{c|}{Min 3-D layers} & \multicolumn{3}{c|}{Max lost layers} & \multicolumn{3}{c}{$\alpha_0$} \\\cline{2-13}
& \multicolumn{1}{c}{~~L~~} &\multicolumn{1}{c}{~~T~~} &\multicolumn{1}{c|}{~~H~~} & \multicolumn{1}{c}{~~L~~} &\multicolumn{1}{c}{~~T~~} &\multicolumn{1}{c}{~~H~~}
& \multicolumn{1}{c}{~~L~~} &\multicolumn{1}{c}{~~T~~} &\multicolumn{1}{c|}{~~H~~} & \multicolumn{1}{c}{~~L~~} &\multicolumn{1}{c}{~~T~~} &\multicolumn{1}{c}{~~H~~}
\\
\hline
0 \& 1 & 0 & 3 & 4 & 0 & 3 & 4 & $\infty$ & 2 & 2 & 2.0 & 0.9 & 0.9 \\
2 Trk & 4 & 5 & 5 & 0 & 3 & 3 & $\infty$ & 1 & 1 & 0.9 & 0.7 & 0.5 \\
2 Vtx & 3 & 3 & 3 & 0 & 3 & 3 & $\infty$ & 1 & 1 & 2.0 & 0.9 & 0.9 \\
3 Trk & 4 & 5 & 5 & 2 & 3 & 4 & 1 & 1 & 1 & 0.9 & 0.7 & 0.5 \\
3 Vtx & 3 & 3 & 3 & 2 & 3 & 3 & 1 & 1 & 1 & 2.0 & 0.9 & 0.9 \\
4 & 5 & 5 & 6 & 3 & 3 & 3 & 1 & 0 & 0 & 0.6 & 0.4 & 0.3 \\
5 & 6 & 6 & 6 & 2 & 2 & 2 & 1 & 0 & 0 & 0.6 & 0.35 & 0.25 \\
\hline
\end{tabular}
\vspace{10pt}
\begin{tabular}{@{~}c@{~}@{~}c@{~}|ccc|ccc|ccc|ccc}
\hline
\multicolumn{1}{c}{\multirow{2}{*}{Iteration}} & \multicolumn{1}{c|}{\multirow{2}{*}{$\beta$}} & \multicolumn{3}{c|}{$\alpha_1$} & \multicolumn{3}{c|}{$\alpha_2$} & \multicolumn{3}{c|}{$\alpha_3$} & \multicolumn{3}{c}{$\alpha_4$} \\\cline{3-14}
& & \multicolumn{1}{c}{~~L~~} &\multicolumn{1}{c}{~~T~~} &\multicolumn{1}{c|}{~~H~~} & \multicolumn{1}{c}{~~L~~} &\multicolumn{1}{c}{~~T~~} &\multicolumn{1}{c|}{~~H~~}
& \multicolumn{1}{c}{~~L~~} &\multicolumn{1}{c}{~~T~~} &\multicolumn{1}{c|}{~~H~~} & \multicolumn{1}{c}{~~L~~} &\multicolumn{1}{c}{~~T~~} &\multicolumn{1}{c}{~~H~~} \\
\hline
0 \& 1 & 4 & 0.55 & 0.30 & 0.30 & 0.65 & 0.35 & 0.35 & 0.55 & 0.40 & 0.40 & 0.45 & 0.40 & 0.40 \\
2 Trk & 4 & 1.50 & 1.00 & 0.90 & 1.50 & 1.00 & 0.90 & 1.50 & 1.00 & 0.90 & 1.50 & 1.00 & 0.90 \\
2 Vtx & 3 & 1.20 & 0.95 & 0.85 & 1.20 & 0.90 & 0.80 & 1.30 & 1.00 & 0.90 & 1.30 & 1.00 & 0.90 \\
3 Trk & 4 & 1.80 & 1.10 & 1.00 & 1.80 & 1.10 & 1.00 & 1.80 & 1.10 & 1.00 & 1.80 & 1.10 & 1.00 \\
3 Vtx & 3 & 1.20 & 1.00 & 0.90 & 1.20 & 1.00 & 0.90 & 1.30 & 1.10 & 1.00 & 1.30 & 1.10 & 1.00 \\
4 & 3 & 1.50 & 1.20 & 1.00 & 1.50 & 1.20 & 1.00 & 1.50 & 1.20 & 1.00 & 1.50 & 1.20 & 1.00 \\
5 & 3 & 1.80 & 1.30 & 1.20 & 1.50 & 1.20 & 1.10 & 1.80 & 1.30 & 1.20 & 1.50 & 1.20 & 1.10 \\
\hline
\end{tabular}
\end{table}
\subsection{Specialized tracking}
The track reconstruction described above produces the main track collection used by the CMS collaboration.
However, variants of this software are also used for more specialized purposes, as described in
this section.
\subsubsection{Electron track reconstruction}
\label{sec:GSF}
Electrons, being charged particles, can be reconstructed through the standard track reconstruction.
However, as electrons lose energy primarily through bremsstrahlung, rather than ionization, large energy
losses are common.
For example, about 35\% of electrons radiate more than 70\% of their initial energy before reaching
the electromagnetic calorimeter (ECAL)
that surrounds the tracker. The energy loss
distribution is highly non-Gaussian, and therefore the standard Kalman filter, which is optimal when all
variables have Gaussian uncertainties, is not appropriate. As a result, the
efficiency and resolution of the standard tracking are not particularly good for electrons and therefore
electron candidates are reconstructed using a combination of two techniques that make use of information,
not only from the tracker, but also from the ECAL. As this is a subject beyond the scope of this paper,
only a brief description of these methods is given.
The first method \cite{Baffioni:2006cd} starts by searching for clusters
of energy in the ECAL. The curvature of electrons in the strong CMS magnetic field means that
bremsstrahlung photons emitted by the electrons will, in general, strike the ECAL at $\eta$ values similar
to that of the electron, but at different azimuthal coordinates $(\phi)$. To recover this radiated energy,
ECAL \textit{superclusters} are formed, by merging clusters of similar $\eta$ over some range of $\phi$. The
knowledge of the energy and position of each supercluster, and the assumption that the electron originated near the centre of the beam spot,
constrains the trajectory of the electron through the tracker (aside from a two-fold
ambiguity introduced by its unknown charge). Tracker seeds compatible with this trajectory are sought
in the pixel tracker (and also in the TEC to improve efficiency in the forward region).
The second method \cite{CMS_PAS_PFT-10-003}
takes the standard track collection (excluding tracks found by Iteration 5, as described in Table~\ref{tab:IterativeSeeds})
and attempts to identify a subset of these tracks that are
compatible with being electrons. Electrons that suffer only little bremsstrahlung loss can be identified by searching
for tracks extrapolated to the ECAL that pass close to an ECAL cluster. Electrons that suffer large
bremsstrahlung loss can be identified by the fact that the fitted track will often have poor $\chi^2$ or few associated hits.
The track seeds originally used to generate these electron-like tracks are retained.
The seed collections obtained by using these two methods are merged, and used to initiate electron track finding. This procedure
is similar to that used in standard tracking, except that the $\chi^2$ threshold, used by the Kalman filter to decide whether a
hit is compatible with a trajectory, is weakened from 30 to 2000. This is to accommodate tracks that deviate from their
expected trajectory because of bremsstrahlung. In addition, the penalties assigned to track candidates for passing through
a tracker layer without being assigned a hit are adjusted. This is necessary because bremsstrahlung photons can
convert into $\Pep\Pem$ pairs with the track-finding algorithm
incorrectly forming a track by combining hits from the primary electron with one of the conversion electrons.
To obtain the best parameter estimates, the final track fit is performed using a
modified version of the Kalman filter, called the Gaussian Sum Filter (GSF) \cite{Adam:2003kg}.
In essence, the fractional energy loss of an electron, as it traverses material of a given thickness, is expected
to have a distribution described by the Bethe--Heitler formula. This distribution is non-Gaussian,
making it unsuitable for use in a conventional Kalman filter algorithm. The GSF technique solves this by
approximating the Bethe--Heitler energy-loss distribution as the sum of several Gaussian functions, whose means, widths, and
relative amplitudes are chosen so as to optimize this approximation.
The parameters of these Gaussian energy-loss functions are determined only once.
Each track trajectory is also represented by a mixture of several `trajectory components', where each trajectory component has
helix parameters with Gaussian uncertainties, and a `weight' corresponding to the probability that it correctly describes
the true path of the particle. Initially, a track trajectory is described by only a single such trajectory component, derived from
the track seed.
When propagating a trajectory component through a layer of material in the tracker, the
estimated mean energy of the trajectory component is reduced and its uncertainty increased, according to the
mean and width of each Gaussian component of the energy-loss distribution applied independently, in turn, to the original trajectory component.
Thus after passing through the a layer of material, each original trajectory component gives rise to several new trajectory components,
each one obtained using one of the Gaussian energy-loss functions. The weight of each new trajectory component is given by
the product of the weight of the original trajectory component and the weight of the corresponding Gaussian component of the energy-loss distribution.
To avoid an exponential explosion in the number of
trajectory components being followed, as the track candidate is propagated through successive tracker layers, the less probable
trajectory components are dropped or merged (by grouping together similar trajectory components), so as to limit their number to 12.
Each trajectory component will also be updated by the Kalman filter if an additional hit is assigned to it when passing through a layer.
When this happens, the weight of the trajectory component is further adjusted according to its compatibility with the hit.
The GSF fit provides estimates of the track parameters, whose uncertainties are described not by
a single Gaussian distribution, but instead by the sum of several Gaussian distributions, each corresponding to the uncertainty
on one of the trajectory components that make up the track.
For each parameter, the mode of this distribution is used as it
is found to provide the best estimates of the parameters.
The performance of the GSF electron tracking has been studied both with simulations \cite{Adam:2003kg}
and with data \cite{CMS_PAS_EGM-10-004}, with good agreement observed between the two.
\subsubsection{Track reconstruction in the high-level trigger}
The CMS high-level trigger (HLT) \cite{LHCC_2007_021} uses a processor farm running C++ software to achieve
large reductions in data rate. The HLT filters events selected at rates of up to 100\unit{kHz} using the
{Level-1} (hardware) trigger. Whereas {Level-1} uses information only from the CMS calorimeters and
muon detectors, the HLT is also able to capture information from the tracker, thereby adding the powerful
tool of track reconstruction to the HLT\@. Some examples of how this improves the HLT performance are
listed below.
\begin{itemize}
\item
Requiring muon candidates that are reconstructed in the muon detectors to be confirmed
through the presence of a corresponding track in the tracker greatly reduces the false reconstruction rate and substantially
improves momentum resolution.
\item
Energy clusters found in the electromagnetic calorimeters can be identified as electrons or
photons through the presence of a track of appropriate momentum pointing to the cluster.
\item
The background rejection rate for lepton triggers can be enhanced by requiring leptons to be isolated. One method
of doing this is to use a veto on the presence of (too many) tracks in a cone around the lepton direction.
\item
Triggering on jets produced by b~quarks can be done by counting the number of tracks in a jet that have
transverse impact parameters statistically incompatible with the track originating from the
beam-line.
\item Triggers on $\tau$ decays $\tau \to \ell \nu_\ell \nu_\tau$, where $\ell =\Pe$ or $\mu$, can
be extended to $\tau \to h \nu_\tau$ decays, where $h$ represents one or more charged hadrons, by
reconstructing a narrow, isolated jet using tracks in combination with calorimeter information.
\end{itemize}
The HLT uses track reconstruction software that is identical to that used for offline reconstruction, but
it must run much faster. This is achieved by using a modified configuration of the track reconstruction.
Tracks can be reconstructed from triplets of hits found using only the pixel tracker,
as documented in Sections~\ref{sec:SeedGeneration} and \ref{sec:pixeltrackvertex}. This is extremely fast, and can be used with great effect
in the reconstruction of the primary-vertex position in the HLT, described in Section~\ref{sec:pixeltrackvertex}.
Tracks can also be reconstructed in the HLT using hits from both the pixel and strip detectors. Such
tracks have superior momentum resolution and a lower probability of being fake. However, this requires much more CPU time than just reconstructing pixel tracks, since the strip tracker does not provide the precise 3-D hits of the pixel tracker, and suffers from a higher hit occupancy. This can be mitigated using
some or all of the following techniques (the details vary significantly, depending on the type of
trigger).
\begin{itemize}
\item
Rather than trying to reconstruct all tracks in the event, \textit{regional} track reconstruction can be
performed instead, where the software is used to reconstruct tracks lying within a specified $\eta$-$\phi$ region
around some object of interest (which might be a muon, electron, or jet candidate reconstructed using the
calorimeters or muon detectors). This saves CPU time, and is accomplished by using
regional seeding. This method differs from the track seeding described in Section~\ref{sec:SeedGeneration},
in that it only forms seeds from combinations of hits that are consistent with a track heading
into the desired $\eta$-$\phi$ region. Another important ingredient of regional tracking concerns
the extraction of hits. As discussed in Section~\ref{sec:striphit}, hits are reconstructed after
unpacking the original data blocks produced by the FED readout boards. Significant time is saved
by unpacking only the data from those FED units that read out tracker modules within the
region of interest~\cite{Wingham:2008zz}. This is not used in the offline reconstruction as the
track reconstruction searches the entire $\eta$-$\phi$ region and therefore needs all hits.
\item
Further gains in speed can be made by performing just a single iteration in the iterative tracking,
such that only seeds made from pairs of pixel hits are considered, where these hits are compatible with a
track originating within a few millimetres of a primary pixel vertex. Furthermore, the HLT uses a higher
\pt requirement when forming the seeds (usually $>$1\GeV) than is used for offline reconstruction.
These stringent requirements on track impact parameter and \pt reduce the number of seeds, and thereby the amount of
time spent building track candidates.
\item
Track finding can differ from that described in Section~\ref{sec:TrackFinding}, in that it can rely on
partial track reconstruction. With this technique, the building of each track candidate is stopped
once a specific condition is met, for example, a given minimum number of hits (typically eight), or a certain
precision requirement on the track parameters. As a consequence,
the hits in the outermost layers of the tracker tend not to be used. While such partially
reconstructed tracks will have slightly poorer momentum resolution and higher fake rates than fully reconstructed
tracks, they also take less CPU time to construct.
\item
Other changes in the tracking configuration can further enhance the speed of reconstruction. For example, when building
track candidates from a given seed, the offline track reconstruction retains at most the best five
partially reconstructed candidates for extrapolating to the next layer. Changing this configurable
parameter to retain fewer candidates can save CPU time.
\end{itemize}
Pixel tracking and other aspects of track reconstruction absorb about 20\% of the total HLT CPU
time. This is kept low by performing track reconstruction only when necessary, and only after other
requirements have been satisfied, so as to reduce the rate at which tracking must be performed.
Track reconstruction is employed in a variety of ways to satisfy different needs in the HLT\@.
Examples of track reconstruction at the HLT include seeds originating in the muon detector, tracking
in a specific $\eta$-$\phi$ region defined by a jet, and searching for tracks over the full
detector. Even the most comprehensive (and slowest) track reconstruction configuration at the HLT
is more than ten times faster than the offline reconstruction of tracks in events representative of
data taken in 2011 (\ttbar + 10 pileup events).
\section{Track reconstruction performance}
\label{sec:trackPerformance}
In this section, the performance of the CTF tracking algorithm is evaluated
in terms of tracking efficiency and fake rate, track parameter resolutions, and the CPU time
required for processing collision events. Two different categories of
simulated samples are used: isolated particles and pp collision events. Comparing the results helps
one understand both the performance of the tracking for isolated particles and to what extent it is
degraded in a high hit occupancy environment.
Simulated events offer the possibility of detailed studies of track reconstruction, such as the way
characteristics of the tracker and the design of the track reconstruction algorithms influence its
performance over a wide range of particle momenta and rapidities, and how much its performance
depends on the type of charged particle being reconstructed, and on whether this particle is isolated or
not. The performance in simulation can be compared with that in data in certain regions of phase space
to verify that the results from simulation are realistic. The CMS collaboration demonstrated previously that its
simulation describes the momentum resolution of muons from \JPsi decay to an accuracy of better than
5\% \cite{CMS_PAS_TRK-10-004}; and does similarly well in describing the dimuon mass resolution of muons from Z boson
decay \cite{Chatrchyan:2013mxa}. The transverse and longitudinal impact parameters of tracks reconstructed in typical
multijet events agree in data and simulation to better than 10\% \cite{CMS_PAS_TRK-10-005}. The CMS collaboration also showed that the
tracking efficiency for particles from \JPsi and charmed hadron decays is simulated with a precision
better than 5\% \cite{CMS_PAS_TRK-10-002}. A similar comparison for the higher-momentum muons from Z boson decay will
be presented in Section~\ref{subSec:TagAndProbe} of the present work.
The isolated particle samples that are used here consist of simple events with just a single generated muon, pion or electron,
although secondary particles may also be present due to interactions with the detector material.
The single particles are generated with a flat distribution in pseudorapidity inside the tracker acceptance $\abs{\eta} < 2.5$.
Their transverse momenta are either fixed to
1, 10 or 100\GeV, or are generated according to a flat distribution in $\ln(\pt)$.
The former set of particles with fixed momenta is used for studying the tracking performance
as a function of $\eta$, while the latter is used to quantify the performance as a function of \pt.
For pp collisions, simulated inclusive \ttbar events are used,
either with or without superimposed pileup events.
The average number of pileup collisions per LHC bunch crossing depends on the instantaneous
luminosity of the machine and on the period of data-taking over which the luminosity is averaged.
For the sake of simplicity, the number of pileup interactions superimposed on each
simulated \ttbar event is randomly generated from a Poisson distribution with mean equal to 8.
This amount of pileup corresponds roughly to what was delivered by the LHC, when averaged over the whole 2011 running period.
The \ttbar events, and also the minimum-bias events used for the pileup, are generated with
the \PYTHIA6 program~\cite{Pythia6}.
Simulated particles are paired to reconstructed tracks
for evaluating tracking efficiency, fake rate, and other quantities discussed in this section.
A simulated particle is \emph{associated} with a
reconstructed track if at least 75\% of the hits assigned to the reconstructed track originate from the simulated particle. The association of simulated hits with reconstructed hits is possible because
the simulation software records the particles responsible for the signal in each channel of the tracker.
Strip and pixel response to electronic noise is also recorded.
Reconstructed tracks that are not associated with a simulated particle are referred to as \textit{fake} tracks.
Results for tracking efficiencies and for fake rates are presented in Section~\ref{subSec:EfficiencyFakeRate}.
While the latter is evaluated only using simulated samples,
the former is also measured in data as described in Section~\ref{subSec:TagAndProbe}.
The resolution obtained for track parameters is discussed in Section~\ref{subSec:TrackParResolution}. Unless
indicated otherwise, all results pertaining to the performance are obtained using the set of `high-purity' tracks
defined in Section~\ref{sec:TrackSelection}. Finally, Section~\ref{sec:CPU} provides estimates of
the CPU time required for different components of track reconstruction.
\subsection{Tracking efficiency and fake rate}
\label{subSec:EfficiencyFakeRate}
For simulated samples, the tracking efficiency is defined as the fraction of simulated charged particles
that can be associated with corresponding reconstructed tracks,
where the association criterion is the one described at the beginning of this section.
This definition of efficiency depends not only on the quality of the track-finding
algorithm, but also upon the intrinsic properties of the tracker, such as its geometrical acceptance
and material content.
Using the same association criterion as used for the efficiency,
the fake rate is defined as the fraction of reconstructed tracks that are not
associated with any simulated particle.
This quantity represents the probability that a reconstructed
track is either a combination of unrelated hits or a genuine
particle trajectory that is badly reconstructed through the inclusion of spurious hits. The efficiency and
fake rate presented in this section are given as a function of \pt and $\eta$ of the
simulated particle and reconstructed track, respectively.
The efficiency is obtained for simulated particles generated
within $\abs{\eta}<2.5$,
with a production
point $<$3\cm and $<$30\cm from the centre of the beam spot for $r$ and $\abs{z}$, respectively. These criteria
select fairly prompt particles. We also require $\pt > 0.9$\GeV, for the study of efficiency
as a function of $\eta$, or $\pt > 0.1$\GeV for studying efficiency over the entire \pt spectrum.
Since the `high-purity' requirement described in Section~\ref{sec:TrackSelection} is the default track selection
for the majority of analyses in CMS, unless otherwise stated efficiency and fake rate are measured and presented here
using only the subset of reconstructed tracks that are identified as `high-purity'.
\subsubsection{Results from simulation of isolated particles}
\label{sec:PerfEffAndFakesSingleMC}
This section presents the performance of the CTF tracking software in reconstructing
trajectories of particles in events containing just a single muon, a pion or an electron.
Muons are reconstructed better than any other charged particle in the tracker, as they mainly interact with the silicon detector
through ionization of the medium and, unlike electrons, their energy loss through bremsstrahlung is negligible.
Muons therefore tend to cross the entire volume of the tracking system,
producing detectable hits in several sensitive layers of the apparatus. Finally, muon trajectories are
altered almost exclusively by Coulomb scattering and energy loss, whose effects are straightforward to
include within the formalism of Kalman filter.
For isolated muons with $1 < \pt < 100\GeV$, the tracking efficiency is
$>$99\% over the full $\eta$-range of tracker acceptance, and does
not depend on \pt~(Fig.~\ref{fig:SingleTrackEfficiencyMC}, top). The fake rate is completely
negligible.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/efficiencyVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/efficiencyVsPt.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/pi/efficiencyVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/pi/efficiencyVsPt.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/efficiencyVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/efficiencyVsPt.pdf}
\caption {
Track reconstruction \textit{efficiencies} for \textit{single, isolated muons} (top), \textit{pions} (middle) and \textit{electrons} (bottom)
passing the \textit{high-purity} quality requirements.
Results are shown as a function of $\eta$ (left), for $\pt = 1$, 10, and 100\GeV. They are
also shown as a function of \pt (right), for the barrel, transition, and endcap regions, which are defined by the
$\eta$ intervals of 0--0.9, 0.9--1.4 and 1.4--2.5, respectively.
}
\label{fig:SingleTrackEfficiencyMC}
\end{figure}
Charged pions, as muons, undergo multiple scattering and energy loss through ionization as they cross the tracker volume.
However, like all hadrons, pions are also subject to elastic and inelastic nuclear interactions.
The elastic nuclear interactions
introduce long tails in the distribution of the scattering angle, well beyond expectations from Coulomb
scattering. The current implementation of the track-finding algorithm assumes a track trajectory modelled by the
material propagator described in Section~\ref{sec:TrackFinding}. This takes into account Coulomb scattering, but neglects
elastic nuclear interactions. As a result, the formation of a track can be interrupted if a hadron undergoes a large-angle
elastic nuclear scattering.
Hence, a hadron can be reconstructed as a single track with fewer hits, or as two separate tracks, or it may not be found at all.
A loss of hits also degrades the precision with which the parameters of the trajectory can be estimated
(Section~\ref{subSec:TrackParResolution}).
Inelastic nuclear interactions are the main source of tracking inefficiency for hadrons, particularly in those regions of the
tracker with large material content.
Depending on $\eta$, up to 20\% of the simulated pions are not reconstructed~(Fig.~\ref{fig:SingleTrackEfficiencyMC}, middle).
This effect is most significant for hadrons with $\pt\lesssim 700\MeV$, because of the larger
cross sections for nuclear interactions at low energies~\cite{PDG2012}.
The tracking efficiency is also affected, along with the fake rate (Fig.~\ref{fig:SingleTrackFakeRateMC}, top), by
the secondary particles produced in inelastic processes.
This is because the products of nuclear interactions are often emitted with trajectories approximately
aligned to that of the traversing pion, particularly for large pion momenta.
As a result, it is common for the trajectory builder to combine hits
of the incoming pion with those of a secondary particle into a single track.
The degradation in efficiency and the increase in fake rate are correlated, as expected, and the loss in performance
is greatest for highest momentum pions. In general, the merging of separate
trajectories during reconstruction is more common in the region of the barrel to endcap transition and in the endcap
regions of the tracker, as these regions contain large amounts of material. In the transition region, the proportion of
fake tracks is also high because the distances between successive hits on each track are longer,
particularly when passing from a hit in the barrel to a hit in an endcap detector.
These longer distances result in correspondingly larger uncertainties in the track trajectory extrapolation that is
performed during track building. This makes it more probable that spurious hits, such as those from secondary particles,
will be incorrectly assigned to the track. Although the extrapolation uncertainties would be equally large
for muons, the fake rate remains very small for muons, as they rarely produce secondary particles.
While the fake rate is generally $<$2--3\% for tracks reconstructed in the sample of single pions with $\pt = 1$ or 10\GeV,
in a sample of single pions with a \pt of 100\GeV, the fake rate peaks at $\approx$15\% for $\abs{\eta} \approx 1.3$.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/pi/fakerateVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/pi/fakerateVsPt.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/fakerateVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/fakerateVsPt.pdf}
\caption {
Tracking \textit{fake rate} for \textit{single, isolated pions} (top) and \textit{electrons} (bottom)
passing \textit{high-purity} quality requirements.
Results are shown, as a function of the reconstructed $\eta$ (left),
for generated $\pt = 1$, 10, and 100\GeV. Results are
also shown as a function of the reconstructed \pt (right), for the barrel, transition, and endcap regions,
which are defined by the
$\eta$ intervals of 0--0.9, 0.9--1.4 and 1.4--2.5, respectively. The results for \pt are obtained
using particles generated with a flat distribution in $\ln\pt$. NB The measured fake rate depends
strongly on the \pt distribution of the generated particles, since, for example, if no particles are
generated in a given \pt range, most tracks reconstructed in that range must necessarily be fake.
The generated particles used to make the plots of fake rate versus $\eta$ have a different \pt spectrum
to those used to make the plots of fake rate versus \pt, therefore the measured fake rates in these
two sets of plots are not directly comparable.
}
\label{fig:SingleTrackFakeRateMC}
\end{figure}
Electrons lose a large fraction of their energy
via bremsstrahlung radiation before they reach the outer
layers of the silicon tracker. Such radiation has an impact on the reconstruction of electrons, similar to that of
inelastic nuclear interactions on the reconstruction of charged hadrons.
First, if an electron loses most
of its energy before reaching the outer layer of the tracker, the number of hits assigned to the track
can be reduced significantly. Second, if a radiated photon converts to an electron-positron pair or
induces an electromagnetic shower,
the track finder can assign a mixture of hits from the primary electron and
from the secondary particles to a single track. This reduces tracking efficiency, increases fake rate, and
is the principal source of misidentification of charge for electrons. The efficiency and fake rate of the CTF algorithm for reconstructing electrons are shown in
Fig.~\ref{fig:SingleTrackEfficiencyMC} (bottom) and Fig.~\ref{fig:SingleTrackFakeRateMC} (bottom), respectively.
In the barrel, the efficiency for electrons exceeds 90\% for $\pt > 0.4$\GeV, and the fake rate is very small.
However, the performance is significantly worse in the endcap and barrel-endcap transition regions, because of the larger
amount of material and the correspondingly greater chance of an electron to produce an electromagnetic shower within the tracker
volume. The fake rate is particularly high in the sample of electrons with $\pt = 100$\GeV, since any secondary particle they
produce will tend to be emitted tangentially to the direction of the original electron, with the consequence that
the tracking algorithm tends to reconstruct the primary and the secondary particle as a single track.
It is important to note that, in practice, CMS achieves considerably better performance for electron reconstruction,
by using the dedicated GSF algorithm~\cite{CMS_PAS_EGM-10-004}, described in Section~\ref{sec:GSF}, rather
than the standard CTF algorithm.
\subsubsection{Results from simulated pp collision events}
\label{sec:PerfEffAndFakesCollisionMC}
This section presents the performance of the CTF tracking software for reconstructing
trajectories of non-isolated charged particles generated in simulated LHC
collisions. Compared to the results shown in the previous section for isolated
particles, the tracking performance discussed in this section is affected by an additional important feature
of LHC events: the large number of hits produced in the tracker at each LHC bunch crossing.
These hits originate from the hundreds of primary particles and their interactions selected by the CMS triggers.
Their number is increased by the combined effects of low-energy particles spiralling in the CMS magnetic
field (``loopers''), and particles produced in temporally overlapping pileup collisions.
In the following, we give examples of the kind of difficulties encountered by
the tracking algorithm during the reconstruction of these events.
\begin{itemize}
\item Many particles can be emitted within highly collimated jets and the hits they produce are closer
to each other than the typical uncertainty in the position of extrapolated trajectories at the sensors.
In such situations, the trajectory builder is
unable to assign unambiguously the hits to the corresponding trajectories. For example, hits
corresponding to two distinct charged particles can be mixed into one or two reconstructed tracks that
do not describe accurately either of the trajectories of the two particles.
\item Trajectories of nearby particles can be separated sufficiently in the outer layers of the
tracker so as to be correctly identified by the track-finding module. Nevertheless, their hits
in the innermost layers can be so near to each other that the reconstruction algorithm often
assigns incorrectly the hits to the relevant trajectories. Particularly
in the innermost pixel layer particles can be so close to each other that their
ionization signals can merge into a single cluster. In this
case, even if the individual trajectories are reconstructed, and their momenta are
well measured, the resolutions in their impact parameter are degraded by the formation of this merged cluster.
\item Many of the low-\pt particles from the underlying event of the hard collision, or from the other pileup
collisions, have such a low transverse momentum that they cannot escape the volume of the tracker, but
instead spiral in the magnetic field, producing many hits in the detector, increasing the complexity of the track-finding
task. Even when these circulating particles are not close to each other at their production vertex, their
large number of hits increases the probability of having uncorrelated hits
accepted as legitimate trajectories, and thereby generate reconstructed fake track.
\end{itemize}
Since most charged particles produced in LHC collisions
are hadrons, all the sources of inefficiency discussed for single, isolated pions similarly affect
the \ttbar results. The efficiency for reconstructing charged particles in \ttbar events, which is shown in
Fig.~\ref{fig:TTbarEfficiencyFakeRateAllvsHP} (top), closely resembles the reconstruction efficiency for
isolated pions shown in Fig.~\ref{fig:SingleTrackEfficiencyMC} (middle). The similarity between the two
indicates that tracking efficiency is not strongly degraded by particle multiplicity in
typical \ttbar events.
The tracking efficiency as a function of the \pt is approximately constant
for $1 < \pt < 80\GeV$, but, at small \pt, the efficiency decreases quickly
(Fig.~\ref{fig:TTbarEfficiencyFakeRateAllvsHP}, top-right) for several reasons.
\begin{itemize}
\item The pion-nucleus cross section increases rapidly for pions of energies below 0.7\GeV.
\item Track selection criteria~(see Section~\ref{sec:TrackSelection}) are much more stringent for trajectories
of small momentum, as they correspond to the main source of fake tracks.
\item When estimating the \rms scattering angle and mean energy loss in the detector material, the trajectory
propagator assumes that all particles have a pion mass, since the pion is the most
common particle produced in LHC collisions. While this assumption is good for
relativistic particles, it breaks down at low energies when particle masses become more important.
\end{itemize}
For the considered \ttbar events, charged particles with \pt larger than 80\GeV are mostly produced inside the core
of collimated jets. The inability of the trajectory builder to cope fully with regions of the tracker characterised
by extremely high-density of particles is reflected in the drop in tracking efficiency for
large \pt values.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/TTbarGeneralTracksVsHP/efficiencyVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/TTbarGeneralTracksVsHP/efficiencyVsPt.pdf} \\
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/TTbarGeneralTracksVsHP/fakerateVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/TTbarGeneralTracksVsHP/fakerateVsPt.pdf} \\
\caption {
Tracking \textit{efficiency} (top) and \textit{fake rate} (bottom) for simulated \textit{$t\bar t$ events} that include
superimposed pileup collisions.
The number of pileup interactions superimposed on each simulated event is generated randomly from a Poisson distribution with mean
value of 8.
Plots are for all reconstructed tracks, and also for the subset of tracks passing \textit{high-purity} requirements.
The efficiency and fake rate plots are plotted for $\abs{\eta}<2.5$, and the efficiency
for charged particles refers to those generated less than 3\cm (30\cm) from the centre of the beam spot in $r$ ($z$) directions.
The efficiency as a function of $\eta$ is for generated particles with $\pt>0.9$\GeV.
\label{fig:TTbarEfficiencyFakeRateAllvsHP}}
\end{figure}
The fake rate, shown in Fig.~\ref{fig:TTbarEfficiencyFakeRateAllvsHP} (bottom), has a similar
dependence on $\eta$ as that observed for isolated pions in Fig.~\ref{fig:SingleTrackFakeRateMC}.
However, the fake rate has a very different dependence on \pt for the two cases. In the pp
collisions, the fake rate increases for \pt
values $<$1\GeV. This is because the smaller the \pt of an initial trajectory seed, the larger the search
windows that must be used (because of multiple scattering) when searching for additional hits to form
the corresponding track
candidates. This increases the probability to assign wrong hits to a track.
The fake rate also increases at large \pt, as was the case for the single-pion samples,
partially because of the production of secondary particles in nuclear interactions, and partially
because comparatively few high-\pt particles are produced in pp collisions.
The distributions in efficiency and fake rate in Fig.~\ref{fig:TTbarEfficiencyFakeRateAllvsHP} are generated for two sets of
reconstructed tracks: all the tracks produced using the default tracking software, and only those tracks that pass
the \textit{high-purity} requirements. For a 1--2\% reduction in efficiency, the quality requirement reduces the fake rate over
the entire \pt range by more than a factor of two.
Figure~\ref{fig:TTbarEfficiencyFakeRateWithVsWithoutPU} shows the efficiency and fake rate plots for \ttbar events
simulated either with or without superimposed pileup interactions. After applying the quality requirement,
the presence of pileup significantly degrades the efficiency and fake rate only for tracks with a $\pt < 1\GeV$.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/TTbarPuVsNoPu/efficiencyVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/TTbarPuVsNoPu/efficiencyVsPt.pdf} \\
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/TTbarPuVsNoPu/fakerateVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/TTbarPuVsNoPu/fakerateVsPt.pdf} \\
\caption {
Tracking \textit{efficiency} (top) and \textit{fake rate} (bottom) for \textit{$t\bar t$ events} simulated with and without superimposed pileup
collisions.
The number of pileup interactions superimposed on each simulated event is generated randomly from a Poisson distribution with mean
value of 8.
Plots are produced for the subset of tracks passing the \textit{high-purity} quality requirements.
The efficiency and fake rate plots cover $\abs{\eta}<2.5$. The efficiency results are for charged particles produced less
than 3\cm (30\cm) from the centre of the beam spot in $r$ ($z$) directions.
The efficiency as a function of $\eta$ is for generated particles with $\pt>0.9$\GeV.
\label{fig:TTbarEfficiencyFakeRateWithVsWithoutPU}}
\end{figure}
The CMS tracker is capable of reconstructing highly displaced tracks, such as pions from \PKz\xspace decay, or particles
produced in nuclear interactions and photon conversions. This is very useful
for studies of B~physics, photon reconstruction, and for improving energy resolution for particle-flow reconstruction \cite{CMS_PAS_PFT-09-001}.
This capability also makes it possible to search for signatures of new phenomena, such as new long-lived particles
that decay with displaced tracks. Reconstruction of displaced tracks is carried out in
Iterations~3--5 of the 6-step iterative tracking scheme described in Section~\ref{sec:trackReco}. Charged
particles originating outside the pixel detector can also be reconstructed.
The efficiency for reconstructing this kind of charged particle
as a function of the radius of its point of production is shown in Fig.~\ref{fig:TTbarEffVsR} for \ttbar
events.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.6\textwidth]{figs_2011/trackPerformance/MC/TTbarEfficiencyVsRadius.pdf}
\caption{Cumulative contributions to the overall tracking performance from the six iterations in track reconstruction.
The tracking efficiency for simulated \ttbar events is shown as a function of
transverse distance ($r$) from the beam axis to the production point of each particle, for tracks with
$\pt > 0.9$\GeV and $\abs{\eta}<2.5$, transverse (longitudinal) impact parameter $<$60 (30)\cm.
The reconstructed tracks are required to pass the \textit{high-purity} quality requirements.}
\label{fig:TTbarEffVsR}
\end{figure}
\subsubsection{Efficiency estimated from data}
\label{subSec:TagAndProbe}
A ``tag-and-probe'' method~\cite{CMS:2011aa,Chatrchyan:2012xi} allows an extraction of muon tracking
efficiency directly from
decays of known resonances.
For example, $\cPZ\to\Pgmp\Pgmm$ candidates are reconstructed using pairs of oppositely charged muons
identified in the muon chambers. Each $\cPZ$ candidate must have
one \textit{tag} muon, meaning that it is reconstructed in both the tracker and muon chambers, and one
\textit{probe} muon, meaning that it is reconstructed just in the muon chambers, with no requirement
on the tracker.
The invariant mass of each $\mu^+\mu^-$ candidate is required to be within the 50--130\GeV range, around the
91\GeV mass of the Z boson~\cite{PDG2012}.
For both data and simulated events, the tracking efficiency can be estimated as the fraction of the probe muons in $\cPZ\to\Pgmp\Pgmm$
events that can be associated with
a track reconstructed in the tracker. A correction must be made for the fact that some of the probe muons
are not genuine. This correction is obtained by fitting the dilepton mass spectrum in order to
subtract the non-resonant background, since only genuine dimuons will contribute to the resonant peak.
This must be done separately for $\Pgmp\Pgmm$ candidates in which the probe is associated (or not) with
a track in the tracker.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.4\textwidth]{figs_2011/trackPerformance/Data/effEta.pdf}
\includegraphics[width=0.4\textwidth]{figs_2011/trackPerformance/Data/effVtx.pdf}
\caption{Tracking efficiency measured with a tag-and-probe technique,
for muons from $\cPZ$ decays, as a function of
the muon $\eta$ (left) and the number of reconstructed primary vertices in the event (right)
for data (black dots) and simulation (blue bands). }
\label{f:eff_meas}
\end{figure}
The results of fits using the tag-and-probe method are shown for data and simulation in Fig.~\ref{f:eff_meas}
as a function of the $\eta$ of the probe, as well as the number of reconstructed primary vertices in the event.
The measured tracking efficiency is $>$99\% in both data and simulation.
The data displays a $\lesssim$0.3\% drop in tracking efficiency with increasing pileup, which is not reproduced in
the simulation. This may originate from the dynamic (pileup dependent) inefficiency of the pixel detector,
discussed in Section~\ref{s:hit_efficiency}, which is not modelled in the simulation.
The structure in the tracking efficiency when shown as a function of
$\eta$ is caused by inactive modules
and residual misalignment of the tracker. As the figure shows, these detector conditions are well reproduced in simulation.
\subsection{Resolution in the track parameters}
\label{subSec:TrackParResolution}
In the context of the reconstruction software of CMS,
the five parameters used to describe a track are: $d_0$, $z_0$, $\phi$, $\cot \theta$, and the
\pt of the track, defined at the point of closest approach of the
track to the assumed beam axis. This point is called the \textit{impact point},
with global coordinates ($x_0$, $y_0$, $z_0$). Thus $d_0$ and $z_0$ define the
coordinates of the impact point in the radial and $z$ directions ($d_0 = -y_0 \cos \phi+
x_0 \sin \phi$). The azimuthal and polar angles of the momentum vector of the track are denoted by $\phi$ and $\theta$, respectively.
The resolution in the parameters is studied using simulated events, and estimated from \textit{track residuals},
which are defined as the differences between the reconstructed track parameters and the
corresponding parameters of the generated particles. For each of the five track parameters, the
resolution is plotted as a function of the $\eta$ or \pt of the simulated charged particle.
In every bin of $\eta$ or \pt, the distribution in track residuals defines the resolution as the
half-width of the interval that satisfies both of the following requirements.
\begin{itemize}
\item The width contains 68\% of all entries (including underflows and overflows) in the distribution of
the residuals.
\item The interval is centred on the most probable value (mode) of the residuals, where this value is
taken from the peak of a double-tailed Crystal Ball function~\cite{Oreglia:1980cs} fitted to the residuals.
The function must provide different parametrizations of the tails on the left and right sides of the
residuals distribution as, especially for electrons, the distribution can be very asymmetric.
\end{itemize}
For all resolution plots, we also provide a second measure of the resolution, defined such that the
interval contains 90\% of the track residuals. This quantifies the impact of the extreme values, whereas
the resolutions for the 68\% intervals represent the core of the distribution.
\subsubsection{Results from simulation of isolated particles}
\label{sec:PerfResolutionsMCsingle}
Muons do not undergo strong interactions, and
therefore they tend to traverse the entire volume of the tracker, so
the hits on their trajectories provide a long lever arm for reconstruction.
Figure~\ref{fig:resolutionsVsEtaMuMC} shows the dependence on $\eta$ of the resolution for the
five track parameters, for isolated muons with $\pt = 1,$ 10, and 100\GeV.
The same resolutions, but as a function of \pt, are shown in Fig.~\ref{fig:resolutionsVsPtMuMC}.
The resolutions in both the impact parameters and the angular parameters generally deteriorate
for larger values of $\abs{\eta}$ because the extrapolation from the innermost
hit to the beam axis, where the parameters are calculated, becomes larger.
The resolutions in the transverse and longitudinal impact parameters $d_0$ and $z_0$ are shown in
the first two plots of Figs.~\ref{fig:resolutionsVsEtaMuMC} and \ref{fig:resolutionsVsPtMuMC}.
At high momentum, the impact parameter resolution
is dominated by the position resolution of the innermost hit in the pixel detector, while at
lower momenta, the resolution is degraded progressively by multiple scattering. The
improvement in $z_0$ resolution as $\abs{\eta}$ increases to 0.4 can be attributed to the
beneficial effect of charge sharing in the estimation of position of pixel
clusters~(see Fig.~\ref{fig:pixResZ}); in the barrel, as the crossing angle for the tracks in
the pixel layers increases, the clusters broaden, distributing thereby the signal over more than one
pixel, and improving the resolution in position.
The resolutions in the $\phi$ and $\cot \theta$ parameters, shown in the middle two
panels of Figs.~\ref{fig:resolutionsVsEtaMuMC} and \ref{fig:resolutionsVsPtMuMC},
have distributions in resolutions similar to those found for $d_0$ and $z_0$, respectively, for likewise reasons.
However, as the contribution of the strip tracker to the
measurement of $\phi$ and $\theta$ is important, the influence of charge sharing in the pixel
tracker is smaller.
As a function of $\eta$, the resolutions in the four track parameters $d_0$, $z_0$, $\phi$, and $\theta$,
are not exactly symmetric around $\eta = 0$. This effect is not caused by the tracker geometry, but is rather due to the noisy and dead channels
of pixel and strip modules, whose defective components are taken into account in simulation to reproduce the condition of the real detector.
The resolution in \pt is shown in the bottom panel of
Figs.~\ref{fig:resolutionsVsEtaMuMC} and \ref{fig:resolutionsVsPtMuMC}. At high transverse momentum (100\GeV),
the resolution is $\approx$2--3\% up to $\abs{\eta}=1.6$, but deteriorates at
higher $\abs{\eta}$ values, because of the shorter lever arm of these tracks in the $x$-$y$ plane of the tracker.
The degradation at $\abs{\eta} \approx 1.0$ and beyond is due to the gap between the barrel and
the endcap disks~(Fig.~\ref{fig:TrackerLayout}), and due to the inferior hit resolution of the last hits of
the track measured in TEC ring~7 compared to the hit resolution in TOB layers~5 and
6~(Table~\ref{tab:sstHitRes}).
At a transverse momentum of 100\GeV, the material in the tracker accounts for between 20 and
30\% of the transverse momentum resolution; at lower momenta, the resolution is dominated by
multiple scattering and its value reflects the amount of material traversed by the track.
The relative precision in \pt is measured to be best for tracks with $\pt \approx 3\GeV$.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/resolutionD0VsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/resolutionDzVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/resolutionPhiVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/resolutionThetaVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/resolutionPtVsEta.pdf}
\caption {
Resolution, \textit{as a function of pseudorapidity}, in the five track parameters for \textit{single, isolated muons} with
$\pt = 1$, 10, and 100\GeV. From top to bottom and left
to right: transverse and longitudinal impact parameters, $\phi$, $\cot \theta$
and transverse momentum.
For each bin in $\eta$, the solid (open) symbols
correspond to the half-width for 68\% (90\%) intervals centered on the
mode of the distribution in residuals, as described in the text.
}
\label{fig:resolutionsVsEtaMuMC}
\end{figure}
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/resolutionD0VsPt.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/resolutionDzVsPt.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/resolutionPhiVsPt.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/resolutionThetaVsPt.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/mu/resolutionPtVsPt.pdf}
\caption {
Resolution, \textit{as a function of \pt}, in the five track parameters for \textit{single, isolated muons} in
the barrel, transition, and endcap regions, defined by
$\eta$ intervals of 0--0.9, 0.9--1.4 and 1.4--2.5, respectively.
From top to bottom and left to right: transverse and longitudinal impact parameters, $\phi$, $\cot \theta$
and \pt.
For each bin in \pt, the solid (open) symbols
correspond to the half-width for 68\% (90\%) intervals centered on the
mode of the distribution in residuals, as described in the text.
}
\label{fig:resolutionsVsPtMuMC}
\end{figure}
Charged pions that do not undergo nuclear interactions behave similarly to muons, as they
are subjected to the same multiple scattering effects and to the same mechanism of energy loss through
ionization. The trajectories of this subset of pions are reconstructed using the CTF algorithm with a
precision that is close to that achieved for muons, and therefore these trajectories populate
the core of the distributions of residuals. The five plots in
Fig.~\ref{fig:resolutionsVsEtaPiMC} show resolutions in the five track parameters as a function
of $\eta$. As expected, the
results are very close to those observed for muons in Fig.~\ref{fig:resolutionsVsEtaMuMC}.
However, the resolutions obtained
for the 90\% interval have a somewhat different pattern for muons than for pion tracks crossing
the barrel-endcap transition region of the tracker. The residuals are generally larger for pions,
as they can interact inelastically, and thereby fail to reach the outer layers of the tracking system.
Their trajectories are measured therefore using smaller lever arms, with degraded resolutions.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/pi/resolutionD0VsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/pi/resolutionDzVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/pi/resolutionPhiVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/pi/resolutionThetaVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/pi/resolutionPtVsEta.pdf}
\caption {
Resolution, \textit{as a function of pseudorapidity}, in the five track parameters for \textit{single, isolated pions} with
transverse momenta of 1, 10, and 100\GeV. From top to bottom and left
to right: transverse and longitudinal impact parameters, $\phi$, $\cot \theta$
and \pt.
For each bin in $\eta$, the solid (open) symbols
correspond to the half-width for 68\% (90\%) intervals centered on the
mode of the distribution in residuals, as described in the text.
}
\label{fig:resolutionsVsEtaPiMC}
\end{figure}
Three of the track parameters ($d_0$, $\phi$ and \pt) for electrons have very asymmetric residual distributions
because of bremsstrahlung, and we therefore alter the definition of their resolution. The distribution in track residuals
is split into two regions, separated at the mode of the distribution. Only one of these two regions contains long, non-Gaussian tails due to bremsstrahlung and the resolution is now redefined using only the distribution in this region. It is given by the width of an interval that starts at the
mode of the distribution and is wide enough to include 68\% of the entries in the region. A similar definition
of the resolution corresponding to the width of a 90\% probability interval is used to quantify the size of the non-Gaussian
tails. Note that if the distribution of residuals had been symmetric, then the results obtained with
these new definitions of the resolution would be identical to those that would have been obtained with
the original definitions from the beginning of Section~\ref{subSec:TrackParResolution}.
The other two parameters ($\cot \theta$ and $z_0$) are less affected by bremsstrahlung, and we therefore continue
to use the same definition of resolution as for muons and pions.
In Fig.~\ref{fig:resolutionsVsEtaEleMCyesBrem}, we show the resolutions in the
$d_0$, $\phi$, and \pt track parameters as a function of $\eta$ for single, isolated electrons
for simulated \pt values of 10 and 100\GeV. These resolutions
are calculated for using both the standard CTF algorithm as well as using the GSF
algorithm, described in Section~\ref{sec:GSF}. However, the GSF requirements described in
Section~\ref{sec:GSF} for consistency of tracks with energy depositions in the ECAL were not applied, as they are
beyond the scope of this discussion. Because the GSF algorithm handles bremsstrahlung in a better way both the 68\% and the
90\% resolutions are significantly improved relative to those obtained with CTF. Similar effects can also be observed
for the resolution in the $\cot \theta$ and $z_0$ parameters, as shown in Fig.~\ref{fig:resolutionsVsEtaEleMCnoBrem}.
Figure~\ref{fig:biasVsEtaEleMC} shows the bias in the reconstructed \pt of electrons as a
function of $\eta$. The bias is defined by the mode of the distribution of residuals.
An alternative definition, based on the mean value of residuals is also shown.
The momenta are systematically underestimated by the CTF algorithm for electrons outside the barrel region. However, the bias
is almost completely recovered using the GSF algorithm except for electrons with $\abs{\eta} > 2.0$, where it is affected
more severely by the large amount of material in the pixel endcaps.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test3/resolutionD0VsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test6/resolutionD0VsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test3/resolutionPhiVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test6/resolutionPhiVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test3/resolutionPtVsEta.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test6/resolutionPtVsEta.pdf}
\caption {
Resolution, \textit{as a function of pseudorapidity},
in the $d_0$, $\phi$ and \pt track parameters for \textit{single, isolated electrons} with
$\pt = 10$ and 100\GeV. For each bin in $\eta$, the solid (open) symbols
correspond to the width of the 68\% (90\%) intervals having its origin on the
mode of the distribution in residuals, as described in the text.
Only the half of the residuals distribution
that does contain the non-Gaussian tail due to bremsstrahlung is considered in the interval calculation. The left (right) plots are
of electrons reconstructed with the CTF (GSF) algorithm.
}
\label{fig:resolutionsVsEtaEleMCyesBrem}
\end{figure}
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.39\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test1/resolutionThetaVsEta.pdf}
\includegraphics[width=0.39\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test4/resolutionThetaVsEta.pdf}
\includegraphics[width=0.39\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test1/resolutionDzVsEta.pdf}
\includegraphics[width=0.39\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test4/resolutionDzVsEta.pdf}
\caption {
Resolution, \textit{as a function of pseudorapidity},
in the $\cot \theta$ and $z_0$ track parameters for \textit{single, isolated electrons} with
$\pt = 10$ and 100\GeV.
For each bin in $\eta$, the solid (open) symbols
correspond to the half-width of the 68\% (90\%) intervals centered on the
mode of the distribution in residuals, as described in the text.
The left (right) plots are of electrons reconstructed with the CTF (GSF) algorithm.
}
\label{fig:resolutionsVsEtaEleMCnoBrem}
\end{figure}
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.39\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test1/biasPtVsEta.pdf}
\includegraphics[width=0.39\textwidth]{figs_2011/trackPerformance/MC/SingleParticles/el/splitTwoRanges/test4/biasPtVsEta.pdf}
\caption {
\textit{Bias, as a function of pseudorapidity},
on the \pt track parameter for \textit{single, isolated electrons} with
$\pt = 10$ and 100\GeV. For each bin in $\eta$, the solid (open) symbols
correspond to the mode (mean) of the distribution in residuals. The left (right) plots are
of electrons reconstructed with the CTF (GSF) algorithm.
}
\label{fig:biasVsEtaEleMC}
\end{figure}
\clearpage
\subsubsection{Results from simulated pp collision events}
\label{sec:PerfResolutionsMCcollision}
The resolutions for tracks in \ttbar events, with superimposed pileup interactions, are shown as
a function of track \pt in Fig.~\ref{fig:resolutionVsPtTTbar}. For the five track parameters,
the functional dependence is very similar to that observed for single particles
(Fig.~\ref{fig:resolutionsVsPtMuMC}), except for \pt beyond 20--30\GeV and $\eta$
corresponding to the regions outside the tracker barrel.
The impact of pileup on these resolutions is generally negligible.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/ttbarHpWithPuSplitByEta/resolutionD0VsPt.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/ttbarHpWithPuSplitByEta/resolutionDzVsPt.pdf} \\
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/ttbarHpWithPuSplitByEta/resolutionPhiVsPt.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/ttbarHpWithPuSplitByEta/resolutionThetaVsPt.pdf} \\
\includegraphics[width=0.45\textwidth]{figs_2011/trackPerformance/MC/ttbarHpWithPuSplitByEta/resolutionPtVsPt.pdf}
\caption {
Resolution, \textit{as a function of \pt}, in the five track parameters for charged particles in
simulated \textit{$t\bar t$ events} with pileup.
The number of pileup interactions superimposed to each simulated event is generated randomly from a Poisson distribution
with a mean value of 8.
From top to bottom and left
to right: transverse and longitudinal impact parameters, $\phi$, $\cot \theta$,
and \pt.
For each bin in $\pt$, the solid (open) symbols
correspond to the half-width of the 68\% (90\%) intervals centered on the
mode of the distribution in residuals, as described in the text.
}
\label{fig:resolutionVsPtTTbar}
\end{figure}
\subsection{CPU execution time}
\label{sec:CPU}
Track reconstruction is, by far, the most computationally challenging part of CMS data reconstruction: for
processing pp events with pileup, it requires almost as much CPU time as all the other reconstruction modules together.
Furthermore, as the number of pileup events increases, the number of tracks increases in proportion, but
the number of hit combinations that can be assembled into seeds and track candidates increases much more quickly,
leading to a far more rapid increase in the required CPU time. The mean
CPU time per event for reconstructing tracks is shown in Table~\ref{tab:CPU}, separated into needs for
tracking iterations and for
computational steps (track seeding, finding, fitting, etc). The CPU times are given for \ttbar events, simulated
either without pileup or with an average of 8 pileup events.
As the table shows, the presence of pileup significantly increases the total required CPU time, for
example, by a factor 2.4 for Iteration~0 and a factor 8.6 for Iteration~1. Figure~\ref{fig:TracksByAlgo}
shows the number of tracks per event reconstructed in each iteration. The presence of pileup
increases the number of low-\pt tracks, and as these are mainly reconstructed in Iterations~1--3, pileup
has the biggest effect on these three iterations, increasing thereby both the number of tracks and
the use of CPU time.
\begin{table}[htbp]
\centering
\topcaption{\label{tab:CPU} The mean CPU time per event attributable to components of the track reconstruction algorithm.
The top table is divided according to the iteration and the bottom table according to the type of task. The ``other'' category
includes removing clusters between iterations, assignment of track quality, and merging of track collections. Results are given
for simulated \ttbar events without pileup and with an average of 8 pileup events. The times are obtained
from an unloaded machine containing an Intel Core i7~CPU~960 running at 3.20\unit{GHz}.}
\begin{tabular}{l|cc}
\hline
Task & Time for \ttbar (s) & Time for \ttbar + 8 pileup (s) \\
\hline
Pixel tracking \& vertexing & 0.01 & 0.03 \\
Iteration 0 & 0.17 & 0.40 \\
Iteration 1 & 0.13 & 1.12 \\
Iteration 2 & 0.10 & 0.67 \\
Iteration 3 & 0.08 & 0.59 \\
Iteration 4 & 0.11 & 0.48 \\
Iteration 5 & 0.07 & 0.19 \\
Merging of track collections & 0.02 & 0.10 \\
\hline
Total & 0.68 & 3.58 \\
\hline
\end{tabular}
\vspace{10pt}
\begin{tabular}{l|cc}
\hline
Task & Time for \ttbar (s) & Time for \ttbar + 8 pileup (s) \\
\hline
Pixel tracking \& vertexing & 0.01 & 0.03 \\
Seeding & 0.06 & 0.55 \\
Track finding & 0.43 & 2.29 \\
Track fitting & 0.14 & 0.56 \\
Other & 0.04 & 0.15 \\
\hline
Total & 0.68 & 3.58 \\
\hline
\end{tabular}
\end{table}
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.6\textwidth]{figs_2011/trackPerformance/MC/TTbarTracksByIteration.pdf}
\caption {
The number of additional tracks per event reconstructed after each individual iteration, for \ttbar events
generated without pileup and with an average of 8 pileup events.
The distributions include only tracks associated with a simulated charged particle.}
\label{fig:TracksByAlgo}
\end{figure}
\clearpage
\section{Beam spot and primary-vertex reconstruction and its performance}
\label{sec:beamSpotAndPV}
\subsection{Primary-vertex reconstruction}
\label{sec:pvtxreco}
The goal of primary-vertex reconstruction~\cite{CMS_NOTE_2006-032} is to measure the
location, and the associated uncertainty, of all proton-proton interaction vertices in each event,
including the `signal' vertex and any vertices from pileup collisions, using the available
reconstructed tracks. It consists of three steps: (i) selection of the tracks, (ii) clustering
of the tracks that appear to originate from the same interaction vertex, and (iii) fitting for
the position of each vertex using its associated tracks.
Track selection involves choosing tracks consistent with being produced promptly in the primary interaction region,
by imposing requirements on the maximum value of significance of the transverse impact parameter ($<$5)
relative to the centre of the beam spot (which is reconstructed as described in Section~\ref{sec:beamspot}),
the number of strip and pixel hits associated with a track ($\ge$2 pixel layers, pixel+strip $\ge$5 ),
and the normalized $\chi^2$ from a fit to the trajectory ($<$20).
To ensure high reconstruction efficiency, even for minimum-bias events,
there is no requirement on the \pt of the tracks.
The selected tracks are then clustered on the basis of their $z$-coordinates at their point of closest
approach to the centre of the beam spot. This clustering allows for the reconstruction of any number of proton-proton
interactions in the same LHC bunch crossing. The clustering algorithm must balance the
efficiency for resolving nearby vertices in cases of high pileup against the possibility of
accidentally splitting a single, genuine interaction vertex into more than one cluster of tracks.
A simple `gap clustering' algorithm was used in past reconstruction of the CMS data recorded in 2010~\cite{CMS_PAS_TRK-10-005},
with all tracks ordered according to the $z$-coordinate of their point of closest approach to the centre of the beam spot.
When any two neighbouring elements in this ordered set of coordinates had a gap exceeding a distance $z_\text{sep} = 2$\mm,
the gap was used for splitting the tracks on either side into separate vertices.
Interaction vertices separated by less than $z_\text{sep}$ were merged in this algorithm, making it
a poor choice for high-pileup LHC conditions.
Track clustering is therefore now performed using a \textit{deterministic annealing} (DA) algorithm
\cite{IEEE_DetAnnealing},
finding the global minimum for a problem with many degrees of freedom,
in a way that is analogous to that of a physical system approaching a state of minimal energy through
a series of gradual temperature reductions.
The $z$-coordinates of the points of closest approach of the tracks to the centre of the beam spot are referred to as
\ensuremath{z^T_i}, and their associated uncertainties as $\sigma_i^z$.
The tracks must be assigned to some unknown number of vertices at positions \ensuremath{z^V_k}.
`Hard' assignments, where a track is assigned to one and only one vertex,
can be represented by values of probability $p_{ik}$ that equal 1 if track $i$ is assigned
to vertex $k$, and 0 otherwise.
In the DA framework, assignments are `soft', meaning tracks can be associated with more than one vertex,
with probability $p_{ik}$ between $0$ and $1$ that
can be interpreted as the probability of the assignment of track $i$ to vertex $k$
in a large ensemble of possible assignments.
Postulating that a priori every possible configuration is equally likely, this
is analogous to calculations in statistical mechanics if the vertex $\chi^2$ represents the
role of the energy.
The most probable vertex positions at ``temperature'' $T$ follow from the
minimization of the analogue of the free energy in statistical mechanics,
\begin{linenomath}
\begin{equation}
F = -T \sum_i^{\text{\# tracks}} p_i \log \sum_k^{\text{\# vertices}} \rho_k \exp \left[-\frac{1}{T} \frac{(\ensuremath{z^T_i}-\ensuremath{z^V_k})^2}{{\sigma_i^z}^2}\right],
\label{eq:DAF}
\end{equation}
\end{linenomath}
relative to the positions of the vertices \ensuremath{z^V_k} with vertex weights $\rho_k$.
The sums run over the tracks $i$, and the set of vertices $k$
that reflect the temperature $T$.
Tracks enter with constant weights, $p_i$, reflecting their consistency with originating from
the beam spot.
The number of prototype vertices can be chosen to be arbitrarily large, but
after minimizing $F$ with respect to the $z_k$,
many of the prototype positions coincide. Then a finite number of effective vertices
emerge at distinct positions, independent of the number of prototypes.
It is computationally more efficient to use those effective vertices with weights $\rho_k$ that correspond to
the fraction of unweighted prototypes that coincide at position $z_k$.
The weights are variable, but the sum is always constrained to unity.
(This version of DA is called ``mass-constrained clustering'' in~\cite{IEEE_DetAnnealing}, because $\sum_k \rho_k = 1$.)
The assignment probabilities are given by
\begin{linenomath}
\begin{equation}
p_{ik} = \frac{ \rho_k \exp \left[-\frac{1}{T} \frac{\left(\ensuremath{z^T_i}-\ensuremath{z^V_k}\right)^2}{{\sigma_i^z}^2}\right] }{\sum_{k'} \rho_{k'} \exp \left[-\frac{1}{T} \frac{\left(\ensuremath{z^T_i}-\ensuremath{z^V_{k'}}\right)^2}{{\sigma_i^z}^2}\right] },
\label{eq:DApik}
\end{equation}
\end{linenomath}
where
the resolutions $\sigma_i^z$ are effectively scaled by $\sqrt{T}$.
At very high $T$, all $p_{ik}$ become equal, and all tracks become compatible with a
single vertex.
For $T\rightarrow 0$ every track becomes compatible with exactly one vertex, resulting in hard assignment.
The DA algorithm is initiated at a very high temperature with a single vertex.
$T$ is gradually decreased,
and $\partial{F}/\partial{\ensuremath{z^V_k}}=\partial{F}/\partial{\rho_k}=0$ is implemented iteratively
at each new temperature,
starting with the result of the previous step in temperature.
Because local minima are smeared out by the effective scaling of resolutions as a function of $T$,
this procedure traces the global minimum of $F$ from high to low temperature.
The number of vertices increases as the temperature falls, and rises each time the minimum of $F$ turns
into a saddle point at lower temperatures. This happens whenever $T$ falls below
the critical temperature of one of the vertices,
\begin{equation}
T_c^k=2\sum_i \frac{p_i p_{ik}}{{\sigma_i^z}^2} \left(\frac{\ensuremath{z^T_i}-\ensuremath{z^V_k}}{\sigma_i^z}\right)^2\Big/\sum_i \frac{p_i p_{ik}}{{\sigma_i^z}^2}.
\end{equation}
When this happens,
the vertex involved is then replaced by two nearby vertices before the temperature is decreased again.
The sum of the weights $\rho_k$ of the two resultant vertices is initially set equal to the weight of the parent.
The DA process thereby finds not only positions and assignments of tracks to vertices but also the
number of vertices.
The starting temperature for the whole process is chosen to be above the first critical temperature,
evaluated for $\rho_1=p_{i1}=1$. The temperature is decreased at every step by a
cooling factor of $0.6$.
The `annealing' is continued down to a minimum temperature of $T_{\text{min}}=4$, which
represents a compromise between the resolving power and the possibility of
incorrectly splitting true vertices.
Because of the inherently tentative assignment of tracks in the DA framework,
there is a possibility that tracks can be assigned to multiple vertices.
For the final assignment, the annealing is
continued down to $T=1$, but without more splitting.
As described, the DA algorithm is not robust against outliers,
such as secondary or mismeasured tracks.
Above $T_{\text{min}}$, outlier rejection competes with splitting, and
is therefore not used.
Below $T_{\text{min}}$, an outlier rejection term
$Z_i=\exp(-\mu^2/T)$
is added to the vertex sums in Eq.~(\ref{eq:DAF}), which acts as a cutoff
for the assignment probabilities in the denominator of Eq.~(\ref{eq:DApik}).
Tracks that are more than $\mu$ standard deviations away from the nearest vertex
are down-weighted, and the algorithm becomes a one-dimensional robust adaptive multi-vertex fit \cite{FruehwirtStrandlie99}.
The default value for the cutoff is $\mu=4$.
Outliers tend to create false vertices when other tracks, typically
worse in resolution, are available nearby.
Candidate vertices are therefore retained only if
at least two of their
tracks are incompatible with originating from other vertices.
The tracks assigned to the rejected candidate vertices are not removed but reassigned to other vertices through another minimization of $F$.
The outlier rejection term at this stage allows individual tracks to have low assignment probability
to all remaining vertex candidates.
A minimal probability of 0.5 is required for making the final assignment
when $T=1$ has been reached.
After identifying candidate vertices based on the DA clustering in $z$,
those candidates containing at least two tracks are then fitted using an \textit{adaptive vertex
fitter}~\cite{CMS_NOTE_2007-008} to compute the best estimate of vertex
parameters, including its $x$, $y$ and $z$ position and covariance matrix, as well as the
indicators for the success of the fit, such as the number of degrees of
freedom for the vertex, and weights of the tracks used in the vertex.
In the adaptive vertex fit, each track in the vertex is assigned a
weight between 0 and 1, which reflects the likelihood that it genuinely belongs to the vertex.
Tracks that are consistent with the position of the reconstructed vertex have
a weight close to 1, whereas tracks that lie more than a few standard deviations from the vertex have small
weights. The number of degrees of freedom in the fit is defined as
\begin{linenomath}
\begin{equation}
n_\mathrm{dof} = -3 + 2\sum_{i=1}^{\text{\# tracks}}{w_i},
\label{eq:ndofvtx}
\end{equation}
\end{linenomath}
where $w_i$ is the weight of the $i$th track, and the sum runs over all tracks associated with the vertex.
The value of $n_\mathrm{dof}$ is therefore strongly correlated with the number of
tracks compatible with arising from the interaction region. For this
reason,
$n_\mathrm{dof}$
can be also used to
select true proton-proton interactions.
\subsubsection{Primary-vertex resolution \label{subsec:pvtxresolution}}
The resolution in a reconstructed primary-vertex position depends strongly on the number of tracks
used to fit the vertex and the \pt of those tracks. In this
section, we introduce a `splitting method' for measuring the resolution
as a function of the number of tracks emanating from a vertex.
The tracks used in any given vertex are split equally into two sets. During the splitting
procedure, the tracks are first sorted in descending order of \pt,
and then grouped in pairs starting from the track with the largest \pt.
For each pair, tracks are assigned randomly to one or the other set of tracks.
This ensures that the two sets of tracks have, on average, the
same kinematic properties. These two sets of tracks
are then fitted independently with the adaptive vertex fitter.
To extract the resolution, the distributions in the difference of the fitted vertex positions for a
given number of tracks are fitted using a single Gaussian distribution,
whose fitted \rms width is then divided by $\sqrt{2}$, because the two measurements
of the vertex used in the difference have the same resolution.
The range of the fit is constrained to be within twice
the \rms of the distribution.
Results from a study of the primary-vertex resolution in $x$
and $z$ as a function of the number of tracks associated to the vertex,
using both minimum-bias and jet-enriched data samples,
are shown in Fig.~\ref{fig:pvtx_respt}.
The resolution in $y$ is almost identical to that in $x$, and is therefore omitted.
The minimum-bias sample is collected from a suite of triggers requiring, for example,
only a coincidence of signals from the Beam Scintillator Counters or minimal requirements on the hit or track multiplicity in the pixel detectors.
The jet-enriched samples are produced by requiring each event to have a reconstructed jet
with transverse energy $\et > 20$\GeV. The tracks in these events have significantly higher
mean \pt, resulting in higher resolution in the track impact parameter and consequently better vertex resolution.
For minimum-bias events, the resolutions in $x$ and $z$ are, respectively, less than
$20\mum$ and $25\mum$, for primary vertices reconstructed using at least 50 tracks.
The resolution is better for the jet-enriched sample across the full
range of the number of tracks used to fit the vertex,
approaching $10\mum$ in $x$ and $12\mum$ in $z$ for primary vertices using at least 50 tracks.
The primary-vertex resolution
for the minimum-bias data from pp collisions has also been compared with simulated minimum-bias events
(\PYTHIA6, Tune Z2~\cite{Field:2011iq}), and found to be in excellent agreement.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/ResX_MB_Jet.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/ResZ_MB_Jet.pdf}
\caption{Primary-vertex resolution
in $x$ (left)
and
in $z$ (right) as a function of the number of
tracks at the fitted vertex, for two kinds of events with different average track \pt values (see text).
\label{fig:pvtx_respt}}
\end{figure}
The difference between the measured vertex positions,
divided by the sum of the contributions to the uncertainty from the fit, taken in quadrature, is
referred to as the ``pull.''
The standard deviation of the Gaussian function fitted to the pull distribution is roughly
independent of the number of tracks at the vertex and
is found to be approximately 0.93 in data and 0.90 in simulation,
indicating that the position uncertainty from the fit to a vertex is slightly overestimated for both.
This is consistent with the slightly overestimated track uncertainties observed in MC studies.
\ifANnote{The mean of the pull distribution in $x$ and $z$ as a function of the number
of tracks in the vertex for minimum-bias data and simulation
is shown in Fig.~\ref{fig:pvtx_pulls}.}
\subsubsection{Efficiency of primary-vertex reconstruction \label{subsec:pvtxefficiency}}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/PrimVtx_efficiency.pdf}
\caption[Primary-vertex reconstruction efficiency as a function of the number of
tracks.]{Primary-vertex reconstruction efficiency as a function of the number of
tracks in a cluster, measured in minimum-bias data and in MC simulation. \label{fig:pvtx_efficiency}}
\end{figure}
Given an input set of reconstructed tracks, the primary-vertex reconstruction efficiency
is evaluated based on how often a vertex is reconstructed successfully and its
position found consistent with the true value. Neither
the tracking efficiency nor the probability to produce a minimal
number of charged particles in a minimum-bias interaction is considered
in the extraction of the efficiency for reconstruction of the vertex.
Just as in the measurement of the resolution, the efficiency for
primary-vertex reconstruction depends strongly on the
number of tracks in the cluster.
The same splitting method described in the previous section can be used
to also extract the reconstruction efficiency as a function
of the number of tracks in the vertex cluster.
In this implementation of the method, the tracks used at the vertex
are sorted first in descending order of \pt and then split into two
different sets, such that two-thirds (one-third) of the tracks are randomly assigned as
\textit{tag} (\textit{probe}) tracks. The asymmetric splitting is used to increase the
number of vertices with a small number of tracks, where the efficiency is
expected to be lowest. The sets of tag and probe tracks are then
clustered and fitted independently to extract the vertex
reconstruction efficiency. While each event is not entirely reclustered,
the contribution to the efficiency from such clustering is not neglected, as the
possibility still remains that a new cluster, using the reduced set of tracks following
splitting, will not be formed. The effect of pileup on the measurement of the vertexing
efficiency has been checked in simulation, and found to be small.
The efficiency is calculated based on the number of times the probe
vertex is reconstructed and matched
to the original vertex, given that the tag vertex is reconstructed and
matched to the original vertex. A tag or probe vertex is considered
to be matched to the original vertex if the tag or probe vertex position
in $z$ is within $5\sigma$ from the original vertex. The value of $\sigma$ here is
chosen to be the larger of the uncertainty in the fit to a vertex for the tag
or probe tracks and the uncertainty in the original vertex.
Figure~\ref{fig:pvtx_efficiency} shows the efficiency of the primary-vertex reconstruction
as a function of the number of tracks clustered in $z$.
The results are obtained using the splitting method described above,
applied to both minimum-bias data and to MC simulation, and show agreement between the
two samples. The primary-vertex efficiency is estimated to be close
to 100\% when more than two tracks are used to reconstruct the
vertex. The effect of pileup on the efficiency is checked using
simulated minimum-bias events, with and without added pileup, and the
loss of efficiency is found to be $ < 0.1$\% for the pileup with a mean value of 8.
\subsection{Track and vertex reconstruction with the pixel detector
\label{sec:pixeltrackvertex}}
CMS has an independent reconstruction of tracks and primary vertices based purely on pixel
hits. The pixel track reconstruction is extremely fast, because only three tracker layers are used, and
the low occupancy and high 3-D spatial resolution of the pixel detector make it ideally suited to track finding.
Such reconstructed pixel tracks and primary vertices can be found extremely fast, hence making them valuable tools for the HLT.
Pixel tracks are formed in the same fashion as the pixel triplets, described in
Section~\ref{sec:SeedGeneration}, requiring $\pt>0.9$\GeV.
Vertex finding using pixel tracks provides a simple
and efficient method for measuring the position of the primary vertex.
The clustering of tracks
is performed using a gap clustering algorithm, with vertex candidates having at least
two tracks fitted using an adaptive vertex fit, as described in Section~\ref{sec:pvtxreco}.
The great speed with which pixel tracks and pixel primary vertices can be reconstructed also makes
them a useful tool for many algorithms in the HLT. For example, counting
the number of pixel tracks near a lepton can help determine if the lepton is isolated.
Similarly, measuring the impact parameter of pixel tracks relative to their
vertex can be used to identify the displaced tracks expected from b-hadron decays.
\subsubsection{Tracking efficiency and fake rate for pixel tracks}
The reconstruction efficiency of pixel tracks is estimated by
comparing the reconstructed tracks with the particles generated in simulation.
Since pixel tracks have only three hits, it is required that all three hits must
be produced by the same simulated particle, for the pixel track and simulated
particle to be associated.
The efficiency for reconstructing a particle as a pixel track is defined
as the fraction of simulated particles that can be associated with a
reconstructed pixel track. The fake rate is defined as the fraction of reconstructed tracks that
are not associated with any simulated particle.
Plots on top left and top right in Fig.~\ref{fig:pixel_track_eff} show the
dependence of the measured pixel track efficiency
on the simulated track $\eta$ and \pt, for inclusive \ttbar events with
and without superimposed pileup (where the number of pileup interactions
is 8, as mentioned in Section~\ref{sec:trackPerformance}). The maximum efficiency for the
pixel tracks is $\sim$85\%. The $\sim$15\% inefficiency arises
mainly from the presence of defective pixel modules ($\sim$2.4\% of the
read out chips in the CMS pixel detector are inoperative) and geometric inefficiency.
The asymmetry between positive and negative $\eta$ reflects the non-uniform distribution
of the affected pixel modules.
In the top-right plot of Fig.~\ref{fig:pixel_track_eff}, the
efficiency drops at low \pt because of the $\pt>0.9$\GeV requirement on pixel tracks.
Figure~\ref{fig:pixel_track_eff} also shows that the addition of
pileup events leads to only a small loss in efficiency.
Plots at the bottom left and bottom right in Fig.~\ref{fig:pixel_track_eff}
show the fake rate as a function of $\eta$ and \pt, both with and
without the presence of pileup.
As observed for the full tracking algorithms in Section~\ref{subSec:EfficiencyFakeRate}, the fake rate increases significantly with $\abs{\eta}$
and \pt.
The effect of pileup is also clearly visible, as the fake
rate increases by 50\% with high pileup.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/efficiencyVsEta_140203.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/efficiencyVsPt_140203.pdf} \\
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/fakerateVsEta_140203.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/fakerateVsPt_140203.pdf} \\
\caption {
Pixel tracking \textit{efficiency} (top) and \textit{fake rate} (bottom) for \textit{$t\bar t$ events} simulated with and without superimposed pileup
collisions.
The number of pileup interactions superimposed on each simulated event is randomly generated from a Poisson distribution with mean
equal to 8.
The two plots of
efficiency and fake rate as a function of pseudorapidity are produced
applying a $\pt >0.9$\GeV selection.
\label{fig:pixel_track_eff}}
\end{figure}
\subsubsection{Resolution in the parameters of pixel tracks}
The resolutions in transverse and longitudinal impact parameters $d_0$ and $d_z$ can be
extracted from simulated events in the same way as in Section~\ref{subSec:TrackParResolution}.
Figure~\ref{fig:pixel_ip_resolution_vs_pt} shows
the resolutions for the five pixel track parameters as a function of
pixel track $\pt$ that includes the effect from pileup.
The distributions are similar in form, but somewhat poorer resolution than those shown for standard tracking
in Fig.~\ref{fig:resolutionVsPtTTbar}. The pixel track resolution in \pt degrades by
over 30\% for track $\pt>10$\GeV.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/resolutionD0VsPt_140203.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/resolutionDzVsPt_140203.pdf} \\
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/resolutionPhiVsPt_140203.pdf}
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/resolutionThetaVsPt_140203.pdf} \\
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/resolutionPtVsPt_140203.pdf}
\caption {
\textit{Resolution, as a function of \pt}, in the five track parameters for pixel tracks
in simulated \textit{\ttbar events} with pileup in the barrel, transition and endcap regions,
defined by the pseudorapidity intervals 0--0.9, 0.9--1.4 and 1.4--2.5, respectively. From top
to bottom and left to right: transverse and longitudinal impact parameters, $\varphi$,
$\cot \vartheta$ and transverse momentum. For each bin in $\pt$, the solid (open) symbol
corresponds to the half-width of the 68\% (90\%) interval centered on the most probable
value of the residuals distributions.
}
\label{fig:pixel_ip_resolution_vs_pt}
\end{figure}
\subsubsection{Position resolution for pixel based vertices}
\label{sec:pixelPV}
The position resolution for pixel vertices is extracted using the same method
used to measure primary-vertex
resolutions in Section~\ref{subsec:pvtxresolution} (split method).
Figure~\ref{fig:pixelpvtx_resolution_MBias_Jets_data} shows the measured
resolution as a function of the number of tracks in $x$ (left),
and $z$ (right), using both minimum-bias and a
jet-enriched data.
(The resolution in $y$ is almost identical to that in $x$, and hence it is omitted.)
The resolution is better for the jet-enriched sample, across the full
range of associated tracks used to reconstruct the pixel vertex. For example, with 50
tracks, the $x$ resolution of a pixel vertex is $30\mum$ for the minimum-bias
sample, compared to $25\mum$ for the jet-enriched sample. This is due to
the fact that tracks in the jet-enriched data have higher
mean \pt compared to those in the minimum-bias sample. As before,
the pixel vertex resolution in the minimum-bias data has also been compared with
that in simulated minimum-bias events and again found to be in good agreement.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.45\textwidth]{figs_2011/beamSpotAndPV/Resy_MBias_Jets_140507.pdf}
\includegraphics[width=0.48\textwidth]{figs_2011/beamSpotAndPV/Resz_MBias_Jets_140507.pdf}
\caption{Pixel vertex position resolutions in $x$ (left) and $z$ (right) as a function of the number of
tracks used in the fitted vertex, for minimum-bias and jet-enriched data.}
\label{fig:pixelpvtx_resolution_MBias_Jets_data}
\end{figure}
\subsection{Reconstruction of the LHC beam spot
\label{sec:beamspot}}
The beam spot represents a 3-D profile of the luminous
region, where the LHC beams collide in the CMS detector. The beam spot parameters are
determined from an average over many events, in contrast to the
event-by-event primary vertex that gives the precise position of a
single collision. Measurements of the centre and dependence of
the luminous region on $r$ and $z$ are important components of event reconstruction.
The position of the centre of the beam spot, corresponding to the centre of the luminous region,
is used, especially in the HLT,
(i) to estimate the position of the interaction point prior
to the reconstruction of the primary vertex; (ii) to provide an additional
constraint in the reconstruction of all the primary vertices of an event;
and (iii) to provide the primary interaction point in the full reconstruction of
low-multiplicity data.
\subsubsection{Determination of the position of the centre of the beam spot}
The position of the centre of the beam spot can be determined in two ways. The first method is through the
reconstruction of primary vertices (see Section~\ref{sec:pvtxreco}),
which map out the collisions as a function of $x$, $y$, and $z$, and therefore the shape of
the beam spot. The mean position in $x$, $y$, and $z$, and the size
of the luminous region can be
determined through a fit of a likelihood to the 3-D distribution of vertex positions.
The second method utilises a correlation between $d_{0}$ and $\phi$ that
appears when the centre of the beam spot is displaced relative to its expected position.
The $d_{0}$ for tracks coming from a primary vertex can
be parametrized as:
\begin{linenomath}
\begin{equation}
d_{0}(\phi, z_0) = x_\mathrm{BS} \sin\phi + \dd{x}{z} \sin\phi\, (z_0-z_\mathrm{BS})
- y_\mathrm{BS} \cos\phi - \dd{y}{z} \cos\phi\, (z_0-z_\mathrm{BS}) ,
\end{equation}
\end{linenomath}
where $x_\mathrm{BS}$ and $y_\mathrm{BS}$ are the $x$ and $y$ positions of the beam at
$z=z_\mathrm{BS}$ (the centre of the beam spot along the beam direction), and
$\ddinline{x}{z}$ and $\ddinline{y}{z}$ are the derivatives (slopes) of $x$ and $y$ relative to $z$.
The fit of the beam spot~\cite{CMS_NOTE_2007-021} uses an iterative
$\chi^2$ fit to utilise this correlation between $d_{0}$ and
$\phi$. With a sample of 1000 tracks, the position can be
determined with a statistical precision of approximately 5\mum.
The two methods have been checked against each other, and provide consistent results.
The precision of the $d_{0}$--$\phi$ fit is better in lower-multiplicity events, however
the width and length of the luminous region can not be obtained with the same algorithm.
Therefore, a combination of the two methods is used
to measure the beam spot used in the full reconstruction of each event.
The $d_{0}$--$\phi$ fit is used to determine the centre of the luminous region
in the transverse plane, ($x_\mathrm{BS}, y_\mathrm{BS}$), and the slopes, $\ddinline{x}{z}$ and $\ddinline{y}{z}$;
while $z_\mathrm{BS}$ and the \rms widths of the luminous region
$\sigma_x$, $\sigma_y$, and $\sigma_z$ are all determined from the
fit to the 3-D vertex distribution. The beam spot is determined in every
luminosity section (LS), corresponding to the events collected during
a period of 23\unit{seconds}.
When the results from all LS intervals of a run are available, they are combined to extract
the final beam spot values.
A weighted average is performed,
with a check implemented to assure that no significant shift occurred in the parameters that might indicate a movement of the beam spot.
To protect against slow drifts
of the beam, no more than 60 consecutive LS are combined at a time.
\begin{figure}[p]
\centering
\includegraphics[width=0.9\textwidth]{figs_2011/beamSpotAndPV/BeamSpot_Selected2011_x0.pdf}
\includegraphics[width=0.9\textwidth]{figs_2011/beamSpotAndPV/BeamSpot_Selected2011_y0.pdf}
\includegraphics[width=0.9\textwidth]{figs_2011/beamSpotAndPV/BeamSpot_Selected2011_z0.pdf}
\caption[Position of the centre of the beam spot as a function of time.]
{Fitted $x_\mathrm{BS}$ (top), $y_\mathrm{BS}$ (middle) and $z_\mathrm{BS}$ (bottom)
positions of the centre of the luminous region as a
function of time during early 2011 running. The $x_\mathrm{BS}$ and $y_\mathrm{BS}$ values are extracted from the
a fit to the $d_{0}$--$\phi$ distribution, and the value of $z_\mathrm{BS}$ is extracted from the fit to the primary-vertex distribution.
Each point represents one luminosity section of 23\unit{seconds}.
The error bars reflect the statistical uncertainty from the fit.
\label{fig:beamxyz_bylumi}}
\end{figure}
Figure~\ref{fig:beamxyz_bylumi} shows the fitted positions as a function of time
for LHC fills during early 2011.
The results demonstrate that, within a fill, the position
is quite stable, while occasionally there are larger shifts between fills.
\subsubsection{Determining the size of the beam spot}
The size of the luminous region is also determined with two methods.
The first one is based on
the reconstructed primary-vertex distribution, where the values of $\sigma$ are obtained through
the likelihood fit described above.
The second method, described below, which measures only the transverse size, is based on
event-by-event correlations between the transverse impact
parameters of two tracks originating from the same vertex.
The displacement of the interaction point within the interaction
region introduces a common displacement of trajectories
of all particles from the interaction. This shift of the trajectories
produces a correlation between the transverse impact parameters of
tracks relative to the nominal beam position. The strength of the
correlation reflects the transverse size of the beam.
The correlation between the transverse impact
parameters of two tracks from the vertex of one
interaction (labelled (1) and (2)) can be expressed by the expectation
value
\begin{linenomath}
\begin{equation}
\left< d_{0}^{(1)} \ d_{0}^{(2)} \right> =
\frac{\sigma_x^2+\sigma_y^2}{2} \cos(\phi_1 - \phi_2)
+ \frac{\sigma_y^2-\sigma_x^2}{2} \cos(\phi_1 + \phi_2),
\label{eq:ipcorr}
\end{equation}
\end{linenomath}
where $\phi_1$ and $\phi_2$ are the azimuthal angles of the tracks
measured at the point of closest approach to the beam. A particular
feature of this correlation is that its size is independent of
the resolutions in vertex positions and impact parameters, and therefore
corrections to remove contributions from the resolution are not required.
Assuming no correlation between $\phi_1 - \phi_2$ and $\phi_1 + \phi_2$, the
coefficients in Eq.~(\ref{eq:ipcorr}) can be obtained through the slopes
of straight lines fitted to the respective dependence of $\left<
d_{0}^{(1)} \ d_{0}^{(2)} \right>$.
Both methods have been used to extract $\sigma_x$ and $\sigma_y$ and
the results averaged over an LHC fill are found to be
consistent to 2-3\mum~\cite{CMS_PAS_TRK-10-005}.
Figure~\ref{fig:beamwidth_bylumi}
shows $\sigma_x$, $\sigma_y$, and $\sigma_z$ as a function of time for LHC fills in early
2011, obtained using the likelihood fit to the primary-vertex distribution.
The size of the beam grows with time during each fill, reflecting the growth of
the beam emittance.
The emittance growth has been directly observed with dedicated instrumentation
and correlated with real-time measurements of the beam size~\cite{White:2012zzc}.
\begin{figure}[p]
\centering
\includegraphics[width=0.9\textwidth]{figs_2011/beamSpotAndPV/BeamSpot_Selected2011_sigmaX.pdf}
\includegraphics[width=0.9\textwidth]{figs_2011/beamSpotAndPV/BeamSpot_Selected2011_sigmaY.pdf}
\includegraphics[width=0.9\textwidth]{figs_2011/beamSpotAndPV/BeamSpot_Selected2011_sigmaZ.pdf}
\label{fig:beamwidth_z}
\caption[Beam size as a function of time.]{Fitted widths
$\sigma_x$ (top) and $\sigma_y$ (middle),
and length $\sigma_z$ (bottom) of the luminous region as a
function of time during early 2011 running, extracted from the fit to
the distribution of reconstructed primary vertices.
Each point represents one luminosity section of 23~seconds.
The error bars reflect the statistical uncertainty from the fit.
\label{fig:beamwidth_bylumi}}
\end{figure}
\clearpage
\section{Summary and conclusions}
\label{sec:conclusions}
CMS has developed sophisticated tracking and vertexing software algorithms,
based on
the Kalman filter, the Gaussian sum filter, and the deterministic annealing filter, to reconstruct
the proton-proton collision data provided by the CMS silicon tracker.
The implementation of these algorithms has been optimized for computational efficiency,
required to keep up with the high data rates recorded using the CMS apparatus.
The flexibility of this software is evident from the fact that, with only few changes,
it has been adapted to provide the fast tracking needed for the CMS high-level trigger,
which processes events at rates of up to 100\unit{kHz}. Furthermore, a dedicated version of the software that
accommodates bremsstrahlung energy loss in the tracker material, is used to reconstruct electrons.
The tracking algorithms reconstruct tracks over the full pseudorapidity
range of the tracker $\abs{\eta} < 2.5$, finding charged particles with \pt as low as 0.1\GeV, or produced as
far as 60\cm from the beam line (such as pions from \PKz\xspace decay).
Promptly produced, isolated muons of $\pt > 0.9$\GeV are reconstructed with essentially 100\%
efficiency for $\abs{\eta} < 2.4$. In the central region ($\abs{\eta} < 1.4$), where the resolution is best,
muons of $\pt = 100$\GeV have resolutions of approximately 2.8\% in \pt, and
10 and 30\mum in transverse and longitudinal impact parameter, respectively.
For prompt, charged particles of $\pt > 0.9$\GeV in simulated \ttbar events, under typical 2011 LHC
pileup conditions, the average track-reconstruction efficiency is 94\% in the barrel region ($\abs{\eta} < 0.9$) of the
tracker and 85\% at higher pseudorapidity. Most of the inefficiency is caused by hadrons undergoing
nuclear interactions in the tracker material. In the same \pt range, the fraction of falsely reconstructed
tracks is at the few percent level.
In the central region, tracks with $1 < \pt < 10$\GeV have a resolution in \pt of approximately 1.5\%.
The resolution in their transverse (longitudinal) impact parameters improves from
90\mum (150\mum) at $\pt = 1$\GeV to 25\mum (45\mum) at $\pt = 10$\GeV. In this momentum range,
the resolution in the track parameters is dominated by multiple scattering.
Tracks are used to reconstruct the primary interaction vertices in each event. For vertices
with many tracks, characteristic of interesting events, the achieved vertex position resolution
is 10--12\mum in each of the three spatial dimensions.
When the LHC was first proposed, it was not at all certain that tracking of such high quality could be
achieved. To make this possible, the CMS collaboration elected to build the world's largest all-silicon
tracker, which would provide a relatively small number of high precision hit position measurements, and
immersed it in a powerful coaxial magnetic field. The collaboration then devoted many years
to the development and study of different tracking algorithms, before finally selecting the ones
described in this paper.
For example, it was thought initially that track finding should be seeded using hits in the outer
layers of the tracker, where the channel occupancy is relatively low. Only later was it broadly appreciated that
the pixel tracker is much better for this purpose, with its high granularity giving it
excellent resolution in three dimensions and an even lower channel occupancy, despite the high
track density.
The CMS track and primary-vertex reconstruction software has already achieved or surpassed the
performance levels predicted at the time that the tracker was originally designed~\cite{ttdr}.
Evolution and refinement of tracking and vertexing algorithms will continue in the future, in order to meet
the challenges of ever increasing LHC luminosity.
\section*{Acknowledgements}
\hyphenation{Bundes-ministerium Forschungs-gemeinschaft Forschungs-zentren} We congratulate our colleagues in the CERN accelerator departments for the excellent performance of the LHC and thank the technical and administrative staffs at CERN and at other CMS institutes for their contributions to the success of the CMS effort. In addition, we gratefully acknowledge the computing centres and personnel of the Worldwide LHC Computing Grid for delivering so effectively the computing infrastructure essential to our analyses. Finally, we acknowledge the enduring support for the construction and operation of the LHC and the CMS detector provided by the following funding agencies: the Austrian Federal Ministry of Science, Research and Economy and the Austrian Science Fund; the Belgian Fonds de la Recherche Scientifique, and Fonds voor Wetenschappelijk Onderzoek; the Brazilian Funding Agencies (CNPq, CAPES, FAPERJ, and FAPESP); the Bulgarian Ministry of Education and Science; CERN; the Chinese Academy of Sciences, Ministry of Science and Technology, and National Natural Science Foundation of China; the Colombian Funding Agency (COLCIENCIAS); the Croatian Ministry of Science, Education and Sport, and the Croatian Science Foundation; the Research Promotion Foundation, Cyprus; the Ministry of Education and Research, Estonian Research Council via IUT23-4 and IUT23-6 and European Regional Development Fund, Estonia; the Academy of Finland, Finnish Ministry of Education and Culture, and Helsinki Institute of Physics; the Institut National de Physique Nucl\'eaire et de Physique des Particules~/~CNRS, and Commissariat \`a l'\'Energie Atomique et aux \'Energies Alternatives~/~CEA, France; the Bundesministerium f\"ur Bildung und Forschung, Deutsche Forschungsgemeinschaft, and Helmholtz-Gemeinschaft Deutscher Forschungszentren, Germany; the General Secretariat for Research and Technology, Greece; the National Scientific Research Foundation, and National Innovation Office, Hungary; the Department of Atomic Energy and the Department of Science and Technology, India; the Institute for Studies in Theoretical Physics and Mathematics, Iran; the Science Foundation, Ireland; the Istituto Nazionale di Fisica Nucleare, Italy; the Korean Ministry of Education, Science and Technology and the World Class University program of NRF, Republic of Korea; the Lithuanian Academy of Sciences; the Ministry of Education, and University of Malaya (Malaysia); the Mexican Funding Agencies (CINVESTAV, CONACYT, SEP, and UASLP-FAI); the Ministry of Business, Innovation and Employment, New Zealand; the Pakistan Atomic Energy Commission; the Ministry of Science and Higher Education and the National Science Centre, Poland; the Funda\c{c}\~ao para a Ci\^encia e a Tecnologia, Portugal; JINR, Dubna; the Ministry of Education and Science of the Russian Federation, the Federal Agency of Atomic Energy of the Russian Federation, Russian Academy of Sciences, and the Russian Foundation for Basic Research; the Ministry of Education, Science and Technological Development of Serbia; the Secretar\'{\i}a de Estado de Investigaci\'on, Desarrollo e Innovaci\'on and Programa Consolider-Ingenio 2010, Spain; the Swiss Funding Agencies (ETH Board, ETH Zurich, PSI, SNF, UniZH, Canton Zurich, and SER); the Ministry of Science and Technology, Taipei; the Thailand Center of Excellence in Physics, the Institute for the Promotion of Teaching Science and Technology of Thailand, Special Task Force for Activating Research and the National Science and Technology Development Agency of Thailand; the Scientific and Technical Research Council of Turkey, and Turkish Atomic Energy Authority; the National Academy of Sciences of Ukraine, and State Fund for Fundamental Researches, Ukraine; the Science and Technology Facilities Council, UK; the US Department of Energy, and the US National Science Foundation.
Individuals have received support from the Marie-Curie programme and the European Research Council and EPLANET (European Union); the Leventis Foundation; the A. P. Sloan Foundation; the Alexander von Humboldt Foundation; the Belgian Federal Science Policy Office; the Fonds pour la Formation \`a la Recherche dans l'Industrie et dans l'Agriculture (FRIA-Belgium); the Agentschap voor Innovatie door Wetenschap en Technologie (IWT-Belgium); the Ministry of Education, Youth and Sports (MEYS) of the Czech Republic; the Council of Science and Industrial Research, India; the Compagnia di San Paolo (Torino); the HOMING PLUS programme of Foundation for Polish Science, cofinanced by EU, Regional Development Fund; and the Thalis and Aristeia programmes cofinanced by EU-ESF and the Greek NSRF.
|
\section{Introduction\label{secintro}}
This paper studies the isotropy representation of (generalized) real flag
manifolds associated to a noncompact real simple Lie algebra $\mathfrak{g}$.
Here we consider the case where $\mathfrak{g}$ is a split real form of a
complex simple Lie algebra.
A flag manifold of $\mathfrak{g}$ is a coset space $\mathbb{F}_{\Theta
}=G/P_{\Theta }$ where $G$ is any connected Lie group with Lie algebra
\mathfrak{g}$ and $P_{\Theta }\subset G$ is a parabolic subgroup. The Lie
algebra $\mathfrak{p}_{\Theta }$ of $P_{\Theta }$ is a parabolic subalgebra
which is the sum of the eigenspaces of the nonnegative eigenvalues of
\mathrm{ad}\left( H_{\Theta }\right) $ with $H_{\Theta }\in \mathfrak{g}$ a
suitable chosen element. If $K\subset G$ is a maximal compact subgroup and
K_{\Theta }=K\cap P_{\Theta }$ then $\mathbb{F}_{\Theta }=K/K_{\Theta }$ as
well.
The two presentations $\mathbb{F}_{\Theta }=G/P_{\Theta }$ and $\mathbb{F
_{\Theta }=K/K_{\Theta }$ yield the isotropy representations of $P_{\Theta }$
and $K_{\Theta }$ on the tangent space $T_{b_{\Theta }}\mathbb{F}_{\Theta }$
at the origin $b_{\Theta }$. The $K_{\Theta }$-representation is obtained by
restricting the $P_{\Theta }$-representation.
Our objective in this paper is to describe the isotropy representation of
K_{\Theta }$. This means that the invariant and irreducible subspaces of
T_{b_{\Theta }}\mathbb{F}_{\Theta }$ must be obtained as well as the
possible decompositions
\begin{equation}
T_{b_{\Theta }}\mathbb{F}_{\Theta }=V_{1}\oplus \cdots \oplus V_{k}
\label{fordecintro}
\end{equation
into $K_{\Theta }$-invariant irreducible components.
The description of the isotropy representatin of $K_{\Theta }$ is essential
to get $K$-invariant geometries on $\mathbb{F}_{\Theta }$. For example the
K $-invariant Riemannian metrics on $\mathbb{F}$ are given by the $K_{\Theta
}$-invariant inner products on $T_{b_{\Theta }}\mathbb{F}_{\Theta }$, which
in turn are direct sum of invariant inner products on the components of a
decomposition (\ref{fordecintro}).
Too look at the $K_{\Theta }$-representation we consider first the the
isotropy representation of $P_{\Theta }$. It is completely determined by the
restriction to its Levi component $Z_{\Theta }$, which is the centralizer in
$G$ of $H_{\Theta }$. The group $Z_{\Theta }$ is reductive, so that its
representation decomposes as a sum of invariant irreducible subspaces. This
decomposition is unique and coincide with the decomposition for the ensuing
representation of the Lie algebra $\mathfrak{z}_{\Theta }$ of $Z_{\Theta }$.
In fact, each $\mathfrak{z}_{\Theta }$-irreducible component is a sum of
root spaces (for a Cartan subalgebra) associated to different roots for
different componets. This implies uniqueness of the decomposition. For the
moment we write the $\mathfrak{z}_{\Theta }$-decomposition as
\begin{equation*}
T_{b_{\Theta }}\mathbb{F}_{\Theta }=W_{1}^{\mathfrak{z}}\oplus \cdots \oplus
W_{n}^{\mathfrak{z}}.
\end{equation*}
The subspaces $W_{i}^{\mathfrak{z}}$ are invariant by $K_{\Theta }$ since
K_{\Theta }\subset Z_{\Theta }$. Hence we are faced to the following
problems:
\begin{enumerate}
\item Find the $K_{\Theta }$-invariant irreducible subspaces inside each
W_{i}^{\mathfrak{z}}$. This includes the question of deciding whether
W_{i}^{\mathfrak{z}}$ is $K_{\Theta }$-irreducible.
\item Among the invariant subspaces of item (1), find those pairs $U_{1}$,
U_{2}$ such that the $K_{\Theta }$-representations on them are equivalent.
Given such a pair we get further invariant irreducible subspaces contained
in $U_{1}\oplus U_{2}$ as graphs of operators $T:U_{1}\rightarrow U_{2}$,
intertwining the representations on $U_{1}$ and $U_{2}$.
\end{enumerate}
The answers to these two questions give the full picture of the $K_{\Theta }
-invariant subspaces.
At this point it is worthwhile to compare the real flag manifolds with the
complex ones. In the complex case the above questions have trivial answers:
The subspaces $W_{1}^{\mathfrak{z}}$ are $K_{\Theta }$-irreducible and no
two of them are equivalent. This is due to the fact that in a complex Lie
group $K_{\Theta }$ is a compact real form of the semi-simple component
G\left( \Theta \right) $ of $Z_{\Theta }$, which is also a complex group. So
that the equivalence classes of $K_{\Theta }$-representations are in
bijection to the $G\left( \Theta \right) $-representations.
On the contrary for real flag manifolds new phenomena occur: There are
\mathfrak{z}_{\Theta }$-irreducible subspaces that are not $K_{\Theta }
-irreducible and there are equivalent $K_{\Theta }$-invariant irreducible
subspaces. Such equivalence gives rise to continuous sets of invariant
subspaces and to the nonuniqueness of the decompositions (\ref{fordecintro}).
The basic differences of the real case to the complex one is that $K$ is not
in general simple and $K_{\Theta }$ is not connected (if $\mathfrak{g}$ is a
split real form). When $K$ is not simple we get a supply of $K_{\Theta }
-invariant subspaces as tangent spaces to the orbits through the origin
b_{\Theta }$ of the simple components of $K$. In many cases these tangent
spaces decompose the $\mathfrak{z}_{\Theta }$-irreducible subspaces. The
fact that $K_{\Theta }$ is not connected requires a separate analysis of the
representations of its group of connected components, the so called $M
-group.
Now we describe the contents of the paper. Section \ref{secisotro} contains
generalities about isotropy representations.
The main technical part of the paper starts at Section \ref{secmequiv} where
we look at the representations of the discrete group $M$. This is the
centralizer in $K$ of the Cartan subalgebra $\mathfrak{a}$ and contains
information about the group of connected components of any $K_{\Theta }$.
Also $M=K_{\Theta }$ if $\mathbb{F}_{\Theta }$ is the maximal flag manifold.
The one dimensional root spaces $\mathfrak{g}_{\alpha }$ are $M$-invariant
thus defining representations of $M$. For the roots $\alpha $ and $\beta $
we put $\alpha \sim _{M}\beta $ if the representations of $M$ on $\mathfrak{
}_{\alpha }$ and $\mathfrak{g}_{\beta }$ are equivalent. The purpose of
Section \ref{secmequiv} is to find $M$-equivalence classes of roots. After
some preparations we proceed to a case by case analysis of the diagrams. For
each case the $M$-equivalence classes are described at the beginning of the
corresponding subsection. For the classical diagrams there are exceptions
since in low dimension the sizes of the classes tend to increase. The
detemination of the $M$-equivalence classes furnishes the complete picture
of the isotropy representation on the maximal flag manifolds. They will be
also a basic tool to detect inequivalent subrepresentations in the other
flag manifolds.
Section \ref{seclemaux} is preparatory. There we prove several lemmas to be
applied in the study of isotropy representations on the partial flag
manifolds. Some of these lemmas have independent interest, like Lemma \re
{lemtransimplelong} which ensures transitivity of the Weyl group on the set
of weights of a given representation. This fact is far from to be true for
general representations.
In Section \ref{secirreduc} we go through the isotropy representations of
the partial flag manifolds, again in a case by case analysis. For the
classical diagrams we use their standard realizations as algebras of
matrices: $A_{l}=\mathfrak{sl}\left( l+1,\mathbb{R}\right) $, $B_{l}
\mathfrak{so}\left( l,l+1\right) $, $C_{l}=\mathfrak{sp}\left( l,\mathbb{R
\right) $ and $D_{l}=\mathfrak{so}\left( l,l\right) $. These realizations
allow the use of nice expressions for the roots. The analysis of the
classical diagrams have the following pattern: First we describe the
\mathfrak{z}_{\Theta }$-irreducible components. Then we check their
K_{\Theta }$-irreducibility and finally we look at equivalence between
irreducible subspaces. The results are summarized at the end of each
corresponding subsection. Regarding to the exceptional diagrams, $G_{2}$ is
clear by its low dimensionality. For $E_{6}$, $E_{7}$ and $E_{8}$, it
follows easily by the general lemmas of Section \ref{seclemaux} that the
K_{\Theta }$-invariant subspaces are the $\mathfrak{z}_{\Theta }
-irreducible components. \ As to $F_{4}$ we refrain to make a detailed and
annoying description of the fifteen flag manifolds. Besides the maximal flag
manifold, where the picture is given by the $M$-equivalence classes, we just
look at a minimal flag manifold.
In conclusion we say that our initial motivation to study the isotropy
representation came from the attempt to understand the $K$-invariant
Riemannian metrics on the real flag manifolds. There is an extensive
literature on invariant Riemannian geometry on complex flag manifolds. See
for example Burstall-Rawnsley \cite{br}, Burstall-Salamon \cite{bs},
Negreiros \cite{neg}, San Martin-Negreiros \cite{smneg}, San Martin-Silva
\cite{smrit}, and Wang-Ziller \cite{wz}, and references therein. In a
complex flag manifold the isotropy representation has a unique decomposition
into invariant irreducible components, which makes the set of invariant
Riemannian metrics a finite dimensional manifold. Our results in this paper
show the existence of infinitely many decompositions on a real flag
manifold, pointing to a great richness of the invariant Riemannian geometry.
\section{Isotropy representation\label{secisotro}}
Let $\mathfrak{g}$ be a split real form of a complex simple Lie algebra,
\mathfrak{g}=\mathfrak{k}\oplus \mathfrak{s}$ be a Cartan decomposition and
\mathfrak{a}\subset \mathfrak{s}$ be a maximal abelian subalgebra. Denote by
$\Pi $ the associated set of roots and by
\begin{equation*}
\mathfrak{g}=\mathfrak{g}_{0}\oplus \sum_{\alpha \in \Pi }\mathfrak{g
_{\alpha }
\end{equation*
the associated root space decomposition. Denote by $G$ the group of inner
automorphisms of $\mathfrak{g}$, which is the subgroup of $\mathrm{Gl}
\mathfrak{g})$ generated by $\exp \mathrm{ad}(\mathfrak{g})$. Let $K$ be the
subgroup of $G$ generated by $\mathrm{ad}(\mathfrak{k})$. Fixing a set $\Pi
^{+}$ of positive roots let $\Sigma $ be the corresponding set of simple
roots. We denote by $\mathfrak{a}^{+}=\{H\in \mathfrak{a}:\forall \alpha \in
\Sigma $, $\alpha \left( H\right) >0\}$ the Weyl chamber associated to
\Sigma $.
A subset $\Theta \subset \Sigma $ defines the parabolic subalgebra of type
\Theta $ given by
\begin{equation*}
\mathfrak{p}_{\Theta }=\mathfrak{g}_{0}\oplus \sum_{\alpha \in \Pi ^{+}
\mathfrak{g}_{\alpha }\oplus \sum_{\alpha \in \langle \Theta \rangle ^{-}
\mathfrak{g}_{\alpha },
\end{equation*
where $\langle \Theta \rangle ^{-}$ is the set of negative roots generated
by $\Theta $. The standard parabolic subgroup $P_{\Theta }$ defined by
\Theta $ is the normalizer of $\mathfrak{p}_{\Theta }$ in $G$. The
associated flag manifold is defined by $\mathbb{F}_{\Theta }=G/P_{\Theta }$.
Since $K$ acts transitively on $\mathbb{F}_{\Theta }$, this flag manifold
can be given by $\mathbb{F}_{\Theta }=K/K_{\Theta }$, where $K_{\Theta
}=P_{\Theta }\cap K$.
When $\Theta =\emptyset $ we get the minimal parabolic subalgebra $\mathfrak
p}=\mathfrak{p}_{\emptyset }$. In this case the subscript is omited and the
maximal flag manifold is written $\mathbb{F}=G/P$. We have $\mathbb{F}=K/M$,
where $M=K_{\emptyset }$ is the centralizer of $\mathfrak{a}$ in $K$.
For an alternative description of the parabolic subalgebra write
\begin{equation*}
\mathfrak{a}_{\Theta }=\{H\in \mathfrak{a}:\forall \alpha \in \Theta ,\alpha
\left( H\right) =0\}
\end{equation*
for the anihilator of $\Theta $. Let $H_{\Theta }$ be characteristic for
\Theta $, that is $H_{\Theta }$ is in the \textquotedblleft partial
chamber\textquotedblright\ $\mathfrak{a}_{\Theta }\cap \mathrm{cl}\mathfrak{
}^{+}$ and satisfies
\begin{equation*}
\Theta =\{\alpha \in \Sigma :\alpha (H_{\Theta })=0\}.
\end{equation*
Then
\begin{equation*}
\mathfrak{p}_{\Theta }=\sum_{\lambda \geq 0}V_{\lambda }\left( H_{\Theta
}\right)
\end{equation*
where $V_{\lambda }\left( H_{\Theta }\right) =\sum_{\alpha \left( H_{\Theta
}\right) =\lambda }\mathfrak{g}_{\alpha }$ is the $\lambda $-eigenspace of
\mathrm{ad}\left( H_{\Theta }\right) $. Clearly any $H_{\Theta }$ satisfying
(\ref{forcaracteris}) yield the same $\mathfrak{p}_{\Theta }$, although the
eigenspaces $V_{\lambda }\left( H_{\Theta }\right) $ may change.
The centralizer of $H_{\Theta }$, $\mathfrak{z}_{\Theta }=\mathrm{cent}_
\mathfrak{g}}\left( H_{\Theta }\right) =\sum_{\alpha \left( H_{\Theta
}\right) =0}\mathfrak{g}_{\alpha }$ is the Levi component of $\mathfrak{p
_{\Theta }$. It is a reductive Lie algebra that decomposes as
\begin{equation*}
\mathfrak{z}_{\Theta }=\mathfrak{g}\left( \Theta \right) \oplus \mathfrak{a
_{\Theta }
\end{equation*
where the semi-simple component $\mathfrak{g}\left( \Theta \right) $ is the
subalgebra generated by $\mathfrak{g}_{\alpha }$, $\alpha \in \pm \Theta $.
Since $\mathfrak{g}$ is a split real form, it follows that $\mathfrak{g
\left( \Theta \right) $ is also a split real form, having Cartan subalgebra
the subspace $\mathfrak{a}\left( \Theta \right) $ spanned by $H_{\alpha }$,
\alpha \in \Theta $ (where $\alpha \left( \cdot \right) =\langle H_{\alpha
},\cdot \rangle $). Put $G\left( \Theta \right) =\langle \exp \mathfrak{g
\left( \Theta \right) \rangle $ for the connected subgroup with Lie algebra
\mathfrak{g}\left( \Theta \right) $.
With this notation we have that $K_{\Theta }=\mathrm{Cent}_{K}\left(
H_{\Theta }\right) $ is the centralizer of $H_{\Theta }$ in $K$ and its Lie
algebra $\mathfrak{k}_{\Theta }=\mathrm{Cent}_{\mathfrak{k}}\left( H_{\Theta
}\right) =\mathfrak{z}_{\Theta }\cap \mathfrak{k}$. Also, $\left( K_{\Theta
}\right) _{0}\subset G\left( \Theta \right) $.
The nilpotent subalgebra
\begin{equation*}
\mathfrak{n}_{\Theta }^{-}=\sum_{\alpha \in \Pi ^{-}\backslash \langle
\Theta \rangle ^{-}}\mathfrak{g}_{\alpha }=\sum_{\lambda <0}V_{\lambda
}\left( H_{\Theta }\right)
\end{equation*
complements $\mathfrak{p}_{\Theta }$ in $\mathfrak{g}$. Hence we identify
\mathfrak{n}_{\Theta }^{-}$ with the tangent space $T_{b_{\Theta }}\mathbb{F
_{\Theta }$ at the origin $b_{\Theta }$. Under this identification the
isotropy representations of $K_{\Theta }$ and $G\left( \Theta \right) $ are
just the adjoint representation, since $\mathfrak{n}_{\Theta }^{-}$ is
normalized by these groups. The same statement holds for the representations
of the Lie algebras $\mathfrak{k}_{\Theta }$, $\mathfrak{g}\left( \Theta
\right) $ and $\mathfrak{z}_{\Theta }$.
Since $\mathfrak{z}_{\Theta }$ is reductive its representation on $\mathfrak
n}_{\Theta }$ is a direct sum
\begin{equation*}
\mathfrak{n}_{\Theta }=\sum_{\sigma }V_{\Theta }^{\sigma }
\end{equation*
where the subspaces $V_{\Theta }^{\sigma }$ are $\mathfrak{z}_{\Theta }
-invariant and irreducible. Here we use $\sigma $ to distinguish the
different invariant subspaces.
\begin{proposicao}
\label{propcomproot}Each $\mathfrak{z}_{\Theta }$-invariant and irreducible
subspace $V_{\Theta }^{\sigma }$ is a direct sum of root spaces,
\begin{equation*}
V_{\Theta }^{\sigma }=\sum \mathfrak{g}_{\alpha }
\end{equation*
where the sum extended to a subset of roots $\Pi _{\Theta }^{\sigma }\subset
\Pi ^{-}\backslash \langle \Theta \rangle ^{-}$. Conversely if $\alpha \in
\Pi ^{-}\backslash \langle \Theta \rangle ^{-}$ then $\mathfrak{g}_{\alpha }$
is contained in a unique $\mathfrak{z}_{\Theta }$-component denoted by
V_{\Theta }\left( \alpha \right) $. We write $\Pi _{\Theta }\left( \alpha
\right) $ for the roots $\beta $ with $\mathfrak{g}_{\beta }\subset
V_{\Theta }\left( \alpha \right) $.
\end{proposicao}
\begin{profe}
This follows by a standard argument using the fact that $\mathfrak{a}\subset
\mathfrak{z}_{\Theta }$. In fact, if $V$ is a $\mathfrak{z}_{\Theta }
-invariant subspace and $X=\sum X_{\alpha }\in V$ then
\begin{equation*}
\mathrm{ad}\left( H\right) X=\sum \alpha \left( H\right) X_{\alpha }\in V
\end{equation*
if $H\in \mathfrak{a}$. By taking suitable values of $H\in \mathfrak{a}$ one
concludes that each component $X_{\alpha }\in V$, so that $\mathfrak{g
_{\alpha }\subset V$. The last statement follows directly from the fact tha
\mathfrak{n}_{\Theta }^{-}$ is the direct sum of the roots spaces as well as
the $\mathfrak{z}_{\Theta }$-components.
\end{profe}
The restriction to $\mathfrak{g}\left( \Theta \right) $ of the $\mathfrak{z
_{\Theta }$-representation on $V_{\Theta }^{\sigma }$ is also irreductible.
This is because $\mathfrak{z}_{\Theta }=\mathfrak{g}\left( \Theta \right)
\oplus \mathfrak{a}_{\Theta }$ with $\mathfrak{a}_{\Theta }$ the center of
\mathfrak{g}\left( \Theta \right) $, so that $\mathrm{ad}\left( H\right) $
is a scalar $\lambda \cdot \mathrm{id}$ in $V_{\Theta }^{\sigma }$ for any
H\in \mathfrak{a}_{\Theta }$. Hence, any $\mathfrak{g}\left( \Theta \right)
-invariant subspace $U\subset V_{\Theta }^{\sigma }$ is also $\mathfrak{z
_{\Theta }$-invariant, ensuring that $V_{\Theta }^{\sigma }$ is $\mathfrak{g
\left( \Theta \right) $-irreducible.
The weight spaces of the representation of $\mathfrak{g}\left( \Theta
\right) $, w.r.t. $\mathfrak{a}\left( \Theta \right) $, are root spaces of
\mathfrak{g}$, so that the weights of the representation are restrictions to
$\mathfrak{a}\left( \Theta \right) $ of some roots $\alpha \in \Pi
^{-}\backslash \langle \Theta \rangle ^{-}$. There is just one highest
weight, say $\mu _{\sigma }$, and two representations of $\mathfrak{g}\left(
\Theta \right) $ on $V_{\Theta }^{\sigma _{1}}$ and $V_{\Theta }^{\sigma
_{2}}$ are equivalent if and only if $\mu _{\sigma _{1}}=\mu _{\sigma _{2}}
. (We note that different representations of $\mathfrak{z}_{\Theta }$ on
V_{\Theta }^{\sigma _{1}}$ and $V_{\Theta }^{\sigma _{2}}$ cannot be
equivalent, even if the $\mathfrak{g}\left( \Theta \right) $-representations
are equivalent.)
The subspaces $V_{\Theta }^{\sigma }$ are also invariant and irreductible by
$G\left( \Theta \right) $, since by definition this group is connected.
Hence $V_{\Theta }^{\sigma }$ is invariant by the identity component $\left(
K_{\Theta }\right) _{0}$ of $K_{\Theta }$, because $\left( K_{\Theta
}\right) _{0}$ $\subset G\left( \Theta \right) $.\ As to $K_{\Theta }$ we
have $K_{\Theta }=M\cdot \left( K_{\Theta }\right) _{0}$ which ensures that
V_{\Theta }^{\sigma }$ is $K_{\Theta }$-invariant, because $M$ leaves
invariant each root space.
Our objective is to get the invariant irreducible subspaces of $\mathfrak{n
_{\Theta }^{-}$ by the $K_{\Theta }$ representation, which is equivalent to
the isotropy representation of the flag $\mathbb{F}_{\Theta }$.
In view of the above discussion we are reduced to the following questions:
\begin{enumerate}
\item Describe the irreducible components $V_{\Theta }^{\sigma }$ of the
\mathfrak{z}_{\Theta }$ representation.
\item Find the $K_{\Theta }$-invariant subspaces of each $V_{\Theta
}^{\sigma }$.
\item Find the pairs of irreducible subspaces having equivalent $K_{\Theta }
-representations.
\end{enumerate}
Finally we note that if $H_{\Theta }\in \mathfrak{a}_{\Theta }\cap \mathrm{c
}\mathfrak{a}^{+}$ then an eigenspace $V_{\lambda }\left( H_{\Theta }\right)
=\sum_{\alpha \left( H_{\Theta }\right) =\lambda }\mathfrak{g}_{\alpha }$ is
contained in $\mathfrak{n}_{\Theta }^{-}$ if $\lambda <0$ and is invariant
by $\mathfrak{z}_{\Theta }$. Hence $V_{\lambda }\left( H_{\Theta }\right) $
is the direct sum of some irreducible components $V_{\Theta }^{\sigma }$.
This remark will be used later to determine the irreducible components
V_{\Theta }^{\sigma }$. Actually, in some cases an eigenspace $V_{\lambda
}\left( H_{\Theta }\right) $ is irreducible and hence is itself a component.
\section{$M$-equivalence classes\label{secmequiv}}
Let \ $M=\mathrm{Cent}_{K}\left( \mathfrak{a}\right) $ be the centralizer of
$\mathfrak{a}$ in $K$. It is known that $M\subset K_{\Theta }=M(K_{\Theta
})_{0}$. Also, any root space $\mathfrak{g}_{\alpha }$ is $M$-invariant. In
this section we determine the pairs of root spaces $\mathfrak{g}_{\alpha }$,
$\mathfrak{g}_{\beta }$ having equivalent representations of $M$. This will
be used later to check equivalence or nonequivalence of $K_{\Theta }
-representations on invariant subspaces.
\begin{defi}
The roots $\alpha $ and $\beta $ are said to be $M$-equivalent (in symbols
\alpha \sim _{M}\beta $) if the representations of $M$ on $\mathfrak{g
_{\alpha }$ and $\mathfrak{g}_{\beta }$ are equivalent. We write $\left[
\alpha \right] _{M}$ for the $M$-equivalence class of the root $\alpha $.
\end{defi}
If $\mathfrak{g}$ is a split real form of a complex semi-simple Lie algebra
then $M$ is a discrete abelian subgroup equals to
\begin{equation*}
M=\{m_{\gamma }=\exp (\pi iH_{\gamma }^{\vee }):\gamma \in \Pi \}
\end{equation*
where $H_{\gamma }^{\vee }=\frac{2H_{\gamma }}{\langle \gamma ,\gamma
\rangle }$ is the co-root associated to $\gamma $ and $H_{\gamma }$ is
defined by $\gamma (H)=\langle H_{\gamma },H\rangle $, $H\in \mathfrak{a}$.
In the above formula the exponential $\exp (\pi iH_{\gamma }^{\vee })$ is in
the complex group $\mathrm{Aut}\left( \mathfrak{g}_{\mathbb{C}}\right) $,
where $\mathfrak{g}_{\mathbb{C}}$ is the complexification of $\mathfrak{g}$
(see \cite{knp}, Theorems 7.53 and 7.55).
The following statement gives a necessary and sufficient condition for the
M $-equivalence between the roots $\alpha $ and $\beta $.
\begin{proposicao}
\label{propmequivalent}The root $\alpha $ and $\beta $ are $M$-equivalent if
and only if, for every $\gamma \in \Pi $ we have
\begin{equation*}
\frac{2\langle \gamma ,\alpha \rangle }{\langle \gamma ,\gamma \rangle
\equiv \frac{2\langle \gamma ,\beta \rangle }{\langle \gamma ,\gamma \rangle
} \quad \mbox{mod}2.
\end{equation*}
\end{proposicao}
\begin{profe}
Take a root $\gamma $ and write as above $m_{\gamma }=\exp (\pi iH_{\gamma
}^{\vee })$. If $X\in \mathfrak{g}_{\alpha }$ and $Y\in \mathfrak{g}_{\beta
} $ then
\begin{equation*}
\mathrm{Ad}(m_{\gamma })X=e^{\pi i\alpha (H_{\gamma }^{\vee })}X\qquad
\mbox{and}\qquad \mathrm{Ad}(m_{\gamma })Y=e^{\pi i\beta (H_{\gamma }^{\vee
})}Y,
\end{equation*
by definition of $m_{\gamma }$. It follows that $\alpha \sim _{M}\beta $ if
and only if $e^{\pi i\alpha (H_{\gamma }^{\vee })}=e^{\pi i\beta (H_{\gamma
}^{\vee })}$, which is equivalent to
\begin{equation*}
\frac{2\langle \gamma ,\alpha \rangle }{\langle \gamma ,\gamma \rangle
\equiv \frac{2\langle \gamma ,\beta \rangle }{\langle \gamma ,\gamma \rangle
} \quad \mbox{mod}2
\end{equation*
as desired.
\end{profe}
As a corollary we get the following necessary condition.
\begin{corolario}
\label{cormequivalent}If $\alpha \sim _{M}\beta $ then $\langle \alpha
,\beta \rangle =0$.
\end{corolario}
\begin{profe}
Suppose that $\langle \alpha ,\beta \rangle \neq 0$. Then we have the
following possibilities for the Killing numbers:
\begin{enumerate}
\item $\alpha $ and $\beta $ have the same length and the angle between them
is $6
{{}^\circ
$ or $12
{{}^\circ
$. Then
\begin{equation*}
\frac{2\langle \alpha ,\alpha \rangle }{\langle \alpha ,\alpha \rangle
=2\qquad \mbox{and}\qquad \frac{2\langle \alpha ,\beta \rangle }{\langle
\alpha ,\alpha \rangle }=\pm 1
\end{equation*
showing that $\alpha $ and $\beta $ are not $M$-equivalent.
\item The angle between $\alpha $ and $\beta $ is $4
{{}^\circ
$ or $13
{{}^\circ
$. If $\alpha $ is the long root then
\begin{equation*}
\frac{2\langle \alpha ,\alpha \rangle }{\langle \alpha ,\alpha \rangle
=2\qquad \mbox{and}\qquad \frac{2\langle \alpha ,\beta \rangle }{\langle
\alpha ,\alpha \rangle }=\pm 1
\end{equation*
and $\alpha $ and $\beta $ are not $M$-equivalent. If otherwise $\beta $ is
the long root then we interchange the roles of $\alpha $ and $\beta $ to get
the same result.
\item The angle between $\alpha $ and $\beta $ is $3
{{}^\circ
$ or $15
{{}^\circ
$. Then
\begin{equation*}
\frac{2\langle \alpha ,\alpha \rangle }{\langle \alpha ,\alpha \rangle
=2\qquad \mbox{and}\qquad \frac{2\langle \alpha ,\beta \rangle }{\langle
\alpha ,\alpha \rangle }=\pm 1,\pm 3,
\end{equation*
concluding the proof.
\end{enumerate}
\end{profe}
In the sequel we apply the above proposition and its corollary to find for
each Dynkin diagram the classes of $M$-equivalence between roots. The
following simple remarks are used throughout with no further reference.
\begin{enumerate}
\item $\alpha \sim _{M}\left( -\alpha \right) $. In fact the Cartan
involution $\theta $ is an equivalence between the $M$-representations in
\mathfrak{g}_{\alpha }$ and $\mathfrak{g}_{-\alpha }$ because $\theta \left(
m\right) =m$ if $m\in M$. This implies that $\mathrm{Ad}\left( m\right)
\circ \theta =\theta \circ \mathrm{Ad}\left( m\right) $ if $m\in M$ and
since $\theta \left( \mathfrak{g}_{\alpha }\right) =\mathfrak{g}_{-\alpha }$
the equivalence follows. Hence we are reduced to check $M$-equivalence
between positive roots alone.
\item The criterion of Proposition \ref{propmequivalent} implies easily that
if $w\in \mathcal{W}$ then $\alpha \sim _{M}\beta $ if and only if $w\alpha
\sim _{M}w\beta $. Hence it will will be enough to get the $M$-equivalence
classes for just one element in each orbit of the Weyl group, that is, for
one root in the simply laced diagrams and for one long root and a short root
in diagrams with multiple edges.
\end{enumerate}
We proceed now to look at the $M$-equivalences for each Dynkin diagram.
\subsection{Diagram $A_{l}$, $l\geq 1$}
We use the standard realization of $A_{l}$ where the positive roots are
written as $\lambda _{i}-\lambda _{j}$, $1\leq i<j\leq l+1$. There are two
cases:
\subsubsection{$A_{l}$, $l\neq 3$}
The classes of \textbf{\ }$M$-equivalence on the positive roots are
singletons. (That is the $M$-representation on different root spaces are not
equivalent.)
Since the Weyl group is transitive on the set of roots it is enough to fix a
specific root $\alpha $ and check that any positive root $\beta \neq \alpha $
is not $M$-equivalent to $\alpha $.
Suppose that $l>3$ and take $\alpha =\lambda _{1}-\lambda _{2}$. The
positive roots orthogonal to $\alpha $ are $\lambda _{i}-\lambda _{j}$ with
3\leq i<j$. By Corollary \ref{cormequivalent} we are reduced to check that
these roots are not $M$-equivalent to $\alpha =\lambda _{1}-\lambda _{2}$.
There are the following cases for $3\leq i<j$:
\begin{enumerate}
\item If $j<l+1$ then $\langle \gamma ,\lambda _{i}-\lambda _{j}\rangle \neq
0$ but $\langle \gamma ,\lambda _{1}-\lambda _{2}\rangle =0$ where $\gamma
=\lambda _{j}-\lambda _{j+1}$. Hence
\begin{equation*}
\frac{2\langle \gamma ,\lambda _{1}-\lambda _{2}\rangle }{\langle \gamma
,\gamma \rangle }=0\qquad \mathrm{and}\qquad \frac{2\langle \gamma ,\lambda
_{i}-\lambda _{j}\rangle }{\langle \gamma ,\gamma \rangle }=\pm 1
\end{equation*
so that $\lambda _{1}-\lambda _{2}$ is not $M$-equivalent to $\lambda
_{i}-\lambda _{j}$, $3\leq i<j$.
\item If $i>3$ then $\langle \gamma ,\lambda _{i}\pm \lambda _{j}\rangle
\neq 0$ but $\langle \gamma ,\lambda _{1}-\lambda _{2}\rangle =0$ where
\gamma =\lambda _{i-1}-\lambda _{i}$, and again $\lambda _{1}-\lambda _{2}$
is not $M$-equivalent to $\lambda _{3}-\lambda _{j}$, $3<j$.
\item If $i=3$ and $j=l+1$ then $\lambda _{4}-\lambda _{l+1}$ is a root
orthogonal to $\lambda _{1}-\lambda _{2}$ but not orthogonal to $\lambda
_{3}-\lambda _{l+1}$.
\end{enumerate}
Finally if $l=1$ there is just one positive root. If $l=2$ then the positive
roots are not orthogonal to each other so by Corollary \ref{cormequivalent}
they are not $M$-equivalent.
\subsubsection{$A_{3}$}
The $M$-equivalence classes on the positive roots are $\{\lambda
_{1}-\lambda _{2},\lambda _{3}-\lambda _{4}\}$, $\{\lambda _{1}-\lambda
_{3},\lambda _{2}-\lambda _{4}\}$ and $\{\lambda _{1}-\lambda _{4},\lambda
_{2}-\lambda _{3}\}$.
In this case the unique root orthogonal to $\alpha =\lambda _{1}-\lambda
_{2} $ is $\lambda _{3}-\lambda _{4}$ and hence, by Corollary \re
{cormequivalent}, $\lambda _{3}-\lambda _{4}$ is the only candidate to be $M
-equivalent to $\lambda _{1}-\lambda _{2}$. To see that indeed $\lambda
_{3}-\lambda _{4}\sim _{M}\lambda _{1}-\lambda _{2}$ note that a root
\gamma =\lambda _{i}-\lambda _{j}$ with $\left( i,j\right) \neq \left(
1,2\right) $ or $\left( 3,4\right) $ is not orthogonal to $\lambda
_{1}-\lambda _{2}$ neither to $\lambda _{3}-\lambda _{4}$. So that the
Killing numbers $\frac{2\langle \gamma ,\lambda _{1}-\lambda _{2}\rangle }
\langle \gamma ,\gamma \rangle }$ and $\frac{2\langle \gamma ,\lambda
_{3}-\lambda _{4}\rangle }{\langle \gamma ,\gamma \rangle }$ are $\pm 1$,
that is, \ the condition of Proposition \ref{propmequivalent} is satisfied
showing that $\lambda _{3}-\lambda _{4}\sim _{M}\lambda _{1}-\lambda _{2}$.
By applying the Weyl group (permutation group) we see that the classes of $M
-equivalences are as stated.
\subsection{Diagram $B_{l}$, $l\geq 2$}
In the standard realization of $B_{l}=\mathfrak{so}\left( l,l+1\right) $ the
positive roots are written as $\lambda _{i}\pm \lambda _{j}$, $1\leq i<j\leq
l$ and $\lambda _{i}$, $1\leq i\leq l$. These are the long and short roots
respectively.
The $M$-equivalence classes depend on the rank $l$, according to the
following cases:
\subsubsection{$B_{l}$, $l\geq 5$}
The $M$-equivalence classes on the positive roots are $\{\lambda
_{i}-\lambda _{j},\lambda _{i}+\lambda _{j}\}$ and $\{\lambda _{i}\}$,
1\leq i<j\leq l$.
We find the equivalence classes of the long and short roots.
Take long root $\lambda _{1}-\lambda _{2}$. We must check $M$-equivalence
only for the roots orthogonal to it, namely $\lambda _{1}+\lambda _{2}$,
\lambda _{i}\pm \lambda _{j}$ and $\lambda _{i}$ with $3\leq i<j$. The roots
$\lambda _{i}$, $3\leq i$, are not $M$-equivalent to $\lambda _{1}-\lambda
_{2}$. In fact, $\lambda _{i}\pm \lambda _{i+1}$ is a root because $l\geq 5
. Now, $\langle \lambda _{i},\lambda _{i}\pm \lambda _{i+1}\rangle \neq 0$
and the Killing number $\langle \left( \lambda _{i}\pm \lambda _{i+1}\right)
^{\vee },\lambda _{i}\rangle =\pm 1$ because $\lambda _{i}$ is a short root.
Since $\langle \lambda _{i},\lambda _{1}-\lambda _{2}\rangle =0$ the
condition of Proposition \ref{propmequivalent} is violated by $\gamma
=\lambda _{i}\pm \lambda _{i+1}$. The same argument used in the $A_{l}$ case
show that $\lambda _{i}\pm \lambda _{j}$, $3\leq i<j$, is not $M$-equivalent
to $\lambda _{1}-\lambda _{2}$ (when $l\geq 5$). On the other hand $\lambda
_{1}-\lambda _{2}\sim _{M}\lambda _{1}+\lambda _{2}$ because for any root
\gamma $ it holds $\langle \gamma ,\lambda _{1}-\lambda _{2}\rangle =\pm
\langle \gamma ,\lambda _{1}+\lambda _{2}\rangle $. It follows that
\{\lambda _{1}-\lambda _{2},\lambda _{1}+\lambda _{2}\}$ is an $M
-equivalence class. To conclude this case we note that $w\in \mathcal{W}$
acts on $\lambda _{i}$ by a permutation followed by a change of sign, that
is, $w\lambda _{i}=\pm \lambda _{j}$, for some index $j$. Hence $\lambda
_{i}-\lambda _{j}\sim _{M}\lambda _{i}+\lambda _{j}$, $1\leq i<j\leq l$ and
the sets $\{\lambda _{i}-\lambda _{j},\lambda _{i}+\lambda _{j}\}$ are the
only $M$-equivalence classes containing a long root.
By the previous paragraph no long root is $M$-equivalent to the short root
\lambda _{i}$. Finaly two short roots $\lambda _{i}$ and $\lambda _{j}$,
i\neq j$, are not $M$-equivalent. For example $\gamma =\lambda _{i}+\lambda
_{k}$, $k\neq i,j$ satisfies $\langle \gamma ^{\vee },\lambda _{i}\rangle =1
$ while $\langle \gamma ^{\vee },\lambda _{j}\rangle =0$.
\subsubsection{$B_{4}$}
The $M$-equivalence classes on the positive roots are $\{\lambda
_{1}-\lambda _{2},\lambda _{1}+\lambda _{2},\lambda _{3}-\lambda
_{4},\lambda _{3}+\lambda _{4}\}$, $\{\lambda _{1}-\lambda _{3},\lambda
_{1}+\lambda _{3},\lambda _{2}-\lambda _{4},\lambda _{2}+\lambda _{4}\}$,
\{\lambda _{1}-\lambda _{4},\lambda _{1}+\lambda _{4},\lambda _{2}-\lambda
_{3},\lambda _{2}+\lambda _{3}\}$, and the short roots $\{\lambda _{i}\}$,
1\leq i\leq 4$.
The difference from the general case is that $\lambda _{3}-\lambda _{4}\sim
_{M}\lambda _{1}-\lambda _{2}$. In fact if $\lambda _{i}$ is a short root
then $\langle \lambda _{i}{}^{\vee },\lambda _{1}-\lambda _{2}\rangle $ and
\langle \lambda _{i}{}^{\vee },\lambda _{3}-\lambda _{4}\rangle $ equals $0$
or $2$. Also if $\gamma $ is a long root different from $\lambda
_{1}-\lambda _{2}$ and $\lambda _{3}-\lambda _{4}$ then $\langle \gamma
^{\vee },\lambda _{1}-\lambda _{2}\rangle $ and $\langle \gamma ^{\vee
},\lambda _{3}-\lambda _{4}\rangle $ equals $\pm 1$.
Again the same arguments show that a long root and a short root as well as
two short roots are not $M$-equivalent.
\subsubsection{$B_{3}$}
The $M$-equivalence classes on the positive roots are $\{\lambda
_{1}-\lambda _{2},\lambda _{1}+\lambda _{2},\lambda _{3}\}$, $\{\lambda
_{1}-\lambda _{3},\lambda _{1}+\lambda _{3},\lambda _{2}\}$ and $\{\lambda
_{2}-\lambda _{3},\lambda _{2}+\lambda _{3},\lambda _{1}\}$.
Here $\lambda _{3}\sim _{M}\lambda _{1}-\lambda _{2}$. The point is that if
\gamma \neq \lambda _{1}-\lambda _{2}$ is a long root then $\langle \gamma
^{\vee },\lambda _{1}-\lambda _{2}\rangle =\pm 1$ and since $\gamma $ cannot
be orthogonal to $\lambda _{3}$ we have $\langle \gamma ^{\vee },\lambda
_{3}\rangle =\pm 1$ as well. On the other hand if $\gamma $ is short then
\langle \gamma ^{\vee },\lambda _{1}-\lambda _{2}\rangle $ and $\langle
\gamma ^{\vee },\lambda _{3}\rangle $ are even.
\subsubsection{$B_{2}$}
The $M$-equivalence classes on the positive roots are the long roots
\{\lambda _{1}-\lambda _{2},\lambda _{1}+\lambda _{2}\}$ and the short roots
$\{\lambda _{1},\lambda _{2}\}$.
\subsection{Diagram $C_{l}$, $l\geq 3$}
In the standard realization of $C_{l}=\mathfrak{sp}\left( l,\mathbb{R
\right) $ the positive roots are written as $\lambda _{i}\pm \lambda _{j}$,
1\leq i<j\leq l$ and $2\lambda _{i}$, $1\leq i\leq l$. These are the short
and long roots respectively.
Here any two long roots $2\lambda _{i}$ and $2\lambda _{j}$ are $M
-equivalent. In fact, for any root $\gamma $ the Killing number $\langle
\gamma ^{\vee },2\lambda _{i}\rangle $ is even ($0$ or $\pm 2$). In fact, if
$\gamma $ is a short root then $\langle \gamma ^{\vee },2\lambda _{i}\rangle
$ is either $0$ (orthogonal roots) or $\pm 2$ (Killing number between a
short root and a long root). On the other hand two long roots are either
equal or orthogonal.
As in the previous diagrams the $M$-equivalence classes increase for small
ranks. For $C_{l}$ the exception is when $l=4$.
\subsubsection{$C_{l}$, $l\neq 4$}
The $M$-equivalence classes are $\{\lambda _{i}-\lambda _{j},\lambda
_{i}+\lambda _{j}\}$ and the set of long roots $\{2\lambda _{1},\ldots
,2\lambda _{l}\}$.
The roots orthogonal to the short root $\lambda _{1}-\lambda _{2}$ are
\lambda _{1}+\lambda _{2}$, $\lambda _{i}\pm \lambda _{j}$ and $2\lambda
_{i} $ with $3\leq i<j$. As in the $B_{l}$ case (with $l\geq 5$) the roots
\lambda _{i}\pm \lambda _{j}$, $3\leq i<j$, are not $M$-equivalent to
\lambda _{1}-\lambda _{2}$. On the other hand if $3\leq i$ then $\gamma
=\lambda _{1}-\lambda _{j}$ with $j\neq i$ violates the criterion of
Proposition \ref{propmequivalent} for $M$-equivalence between $\lambda
_{1}-\lambda _{2}$ and $2\lambda _{i}$. In fact, $\langle \gamma ^{\vee
},\lambda _{1}-\lambda _{2}\rangle =\pm 1$ (non orthogonal roots of same
length) and $\langle \gamma ^{\vee },2\lambda _{i}\rangle =0$.
Since the long roots are equivalente to each other it follows that $\lambda
_{1}-\lambda _{2}$ is not $M$-equivalent to any long root. Hence we get the
classes stated above.
These arguments remain true if $l=3$. (Diferently from $B_{3}$ in $C_{3}$
long roots are not $M$-equivalent to short roots.)
\subsubsection{$C_{4}$}
The $M$-equivalence classes are $\{\lambda _{1}-\lambda _{2},\lambda
_{1}+\lambda _{2},\lambda _{3}-\lambda _{4},\lambda _{3}+\lambda _{4}\}$,
\{\lambda _{1}-\lambda _{3},\lambda _{1}+\lambda _{3},\lambda _{2}-\lambda
_{4},\lambda _{2}+\lambda _{4}\}$, $\{\lambda _{1}-\lambda _{4},\lambda
_{1}+\lambda _{4},\lambda _{2}-\lambda _{3},\lambda _{2}+\lambda _{3}\}$ and
the long roots $\{2\lambda _{1},2\lambda _{2},2\lambda _{3},2\lambda _{4}\}$.
This is seen as in $B_{4}$ where $\lambda _{3}-\lambda _{4}\sim _{M}\lambda
_{1}-\lambda _{2}$.
\subsection{Diagram $D_{l}$, $l\geq 4$}
In the standard realization of $D_{l}=\mathfrak{so}\left( l,l\right) $ the
positive roots are written as $\lambda _{i}\pm \lambda _{j}$, $1\leq i<j\leq
l$.
\subsubsection{$D_{l}$, $l>4$}
The $M$-equivalence classes on the positive roots are $\{\lambda
_{i}-\lambda _{j},\lambda _{i}+\lambda _{j}\}$, $1\leq i<j\leq l$.
This is verified by arguments similar to the $B_{l}$ case, simplified by the
fact that the roots have the same length.
First the only root $M$-equivalent to $\lambda _{1}-\lambda _{2}$ is
\lambda _{1}+\lambda _{2}$. In fact, the roots orthogonal to $\lambda
_{1}-\lambda _{2}$ are $\lambda _{1}+\lambda _{2}$ and $\lambda _{i}\pm
\lambda _{j}$, $3\leq i<j$. A root $\lambda _{i}\pm \lambda _{j}$ with
3\leq i<j$ is not $M$-equivalent to $\lambda _{1}-\lambda _{2}$ by the
following reasons:
\begin{enumerate}
\item If $j<l$ and $\gamma =\lambda _{j}-\lambda _{j+1}$ then $\langle
\gamma ,\lambda _{1}-\lambda _{2}\rangle =0$ and $\langle \gamma ,\lambda
_{i}\pm \lambda _{j}\rangle \neq 0$ which implies that $\langle \gamma
^{\vee },\lambda _{i}\pm \lambda _{j}\rangle =\pm 1$. Thus by Proposition
\ref{propmequivalent} $\lambda _{1}-\lambda _{2}$ is not $M$-equivalent to
\lambda _{i}\pm \lambda _{j}$.
\item If $i>3$ and $\gamma =\lambda _{i-1}-\lambda _{i}$ then $\langle
\gamma ,\lambda _{1}-\lambda _{2}\rangle =0$ and $\langle \gamma ^{\vee
},\lambda _{i}\pm \lambda _{j}\rangle =\pm 1$.
\item Since $l>4$, $\lambda _{4}-\lambda _{l}$ is a root satisfying $\langle
\gamma ,\lambda _{1}-\lambda _{2}\rangle =0$ and $\langle \gamma ^{\vee
},\lambda _{3}\pm \lambda _{l}\rangle =\pm 1$.
\end{enumerate}
Finally $\lambda _{1}-\lambda _{2}\sim _{M}\lambda _{1}+\lambda _{2}$,
because $\langle \gamma ,\lambda _{1}-\lambda _{2}\rangle =0$ if and only if
$\langle \gamma ,\lambda _{1}+\lambda _{2}\rangle =0$ for any root $\gamma
. Also, if $\gamma $ is not orthogonal to both roots then the Killing
numbers are $\pm 1$, since the roots have the same length.
Since the Weyl group is transitive on the set of roots we get the
equivalence classes stated above.
\subsubsection{$D_{4}$}
The $M$-equivalence classes on the positive roots are $\{\lambda
_{1}-\lambda _{2},\lambda _{1}+\lambda _{2},\lambda _{3}-\lambda
_{4},\lambda _{3}+\lambda _{4}\}$, $\{\lambda _{1}-\lambda _{3},\lambda
_{1}+\lambda _{3},\lambda _{2}-\lambda _{4},\lambda _{2}+\lambda _{4}\}$ and
$\{\lambda _{1}-\lambda _{4},\lambda _{1}+\lambda _{4},\lambda _{2}-\lambda
_{3},\lambda _{2}+\lambda _{3}\}$.
In this case, appart from $\lambda _{1}+\lambda _{2}$ the roots $\lambda
_{3}-\lambda _{4}$ and $\lambda _{3}+\lambda _{4}$ are $M$-equivalent to
\lambda _{1}-\lambda _{2}$ (see the discussion for $B_{4}$). Hence an
application of the Weyl group yield the stated classes.
\subsection{Diagram $G_{2}$}
The $M$-equivalence classes on the positive roots are the pairs $\{\alpha
_{1},\alpha _{1}+2\alpha _{2}\}$, $\{\alpha _{1}+\alpha _{2},\alpha
_{1}+3\alpha _{2}\}$ and $\{\alpha _{2},2\alpha _{1}+3\alpha _{2}\}$ where
\alpha _{1}$ and $\alpha _{2}$ are the simple roots with $\alpha _{1}$ the
long one.
The reason is that these are the only pairs of positive roots orthogonal to
each other. Moreover if two roots are not orthogonal then their Killing are
odd ($\pm 1$ ou $\pm 3$).
\subsection{Diagrams $E_{6}$, $E_{7}$ and $E_{8}$}
For these diagrams the $M$-equivalence classes on the positive roots are
singletons.
Since these diagrams are simply-laced it is enough to find a positive root
which is not $M$-equivalent to any other positive root.
In any of the diagrams $E_{6}$, $E_{7}$ and $E_{8}$ we choose the highest
root $\mu $. To check that $\{\mu \}$ is an $M$-equivalence class we prove
the
\begin{itemize}
\item \textbf{Claim:} For every $\beta >0$ with $\langle \mu ,\beta \rangle
=0$ there exists $\gamma \neq \beta $ such that $\langle \mu ,\gamma \rangle
=0$ and $\langle \beta ,\gamma \rangle \neq 0$.
\end{itemize}
From the claim we get $\langle \gamma ^{\vee },\mu \rangle =0$ and $\langle
\gamma ^{\vee },\beta \rangle $ odd because the diagrams are simply laced.
Hence, by Proposition \ref{propmequivalent}, no $\beta $ orthogonal to $\mu $
is $M$-equivalent to $\mu $. By Corollary \ref{cormequivalent} we conclude
that $\{\mu \}$ is an $M$-equivalence class.
Now the roots orthogonal to the highest root $\mu $ have the following
simple description: Denote by $\Sigma =\{\alpha _{1},\ldots ,\alpha _{l}\}$
the simple system of roots, and let $\{\omega _{1},\ldots ,\omega _{l}\}$ be
the fundamental weights, defined by
\begin{equation*}
\langle \alpha _{i}^{\vee },\omega _{j}\rangle =\delta _{ij}.
\end{equation*
It is known that in the diagrams $E_{6}$, $E_{7}$ and $E_{8}$ the highest
root $\mu =\omega _{i}$ for some fundamental weight. (The formula for $\mu $
in terms of the fundamental weights can be read off from the affine Dynkin
diagrams. The extra root is precisely $-\mu $, see \cite{he}, Chapter X,
Table of Diagrams S(A).) Let $\alpha =b_{1}\alpha _{1}+\cdots +b_{l}\alpha
_{l}$, $b_{i}\geq 0$, be a positive root. Since $\mu =\omega _{i}$ we have
by definition
\begin{equation*}
\langle \alpha ,\mu \rangle =\frac{\langle \alpha _{i},\alpha _{i}\rangle }{
}a_{i}b_{i}.
\end{equation*
So that $\langle \alpha ,\mu \rangle =0$ if and only if $a_{i}b_{i}=0$.
Therefore the roots orthogonal to $\mu $ are those spanned by $\Sigma
\setminus \{\alpha _{i}\}$. This set of roots is a root system whose Dynkin
diagram is the subdiagram of $\Sigma $ given by $\Sigma \setminus \{\alpha
_{i}\}$. A glance at the affine Dynkin diagrams provides the diagrams
\Sigma \setminus \{\alpha _{i}\}$, namely,
\begin{itemize}
\item $\Sigma \setminus \{\alpha _{i}\}=A_{5}$ if $\Sigma =E_{6}$.
\item $\Sigma \setminus \{\alpha _{i}\}=D_{6}$ if $\Sigma =E_{7}$.
\item $\Sigma \setminus \{\alpha _{i}\}=E_{7}$ if $\Sigma =E_{8}$.
\end{itemize}
Now it is clear that in any of the root systems spanned by $\Sigma \setminus
\{\alpha _{i}\}$ ($A_{5}$, $D_{6}$ or $E_{7}$), the conclusion of the claim
holds, that is, given $\beta $ there exists $\gamma $ with $\langle \beta
,\gamma \rangle \neq 0$. This concludes the proof that the $M$-equivalence
classes on the positive roots are singletons.
\subsection{$F_{4}$}
The $24$ positive roots of
\begin{picture}(100,60)(0,0)
\put(0,27){$F_4$}
\put(20,30){\circle{6}}
\put(16,20){${\alpha}_1$}
\put(23,30){\line(1,0){20}}
\put(46,30){\circle{6}}
\put(42,20){${\alpha}_2$}
\put(49,31.2){\line(1,0){20}}
\put(49,28.8){\line(1,0){20}}
\put(72,30){\circle{6}}
\put(68,20){${\alpha}_3$}
\put(69,30){\line(-1,2){5}}
\put(69,30){\line(-1,-2){5}}
\put(75,30){\line(1,0){20}}
\put(98,30){\circle{6}}
\put(94,20){${\alpha}_4$}
\end{picture
\noinden
split into the following $M$-equivalence classes:
\begin{itemize}
\item $12$ singletons $\{\alpha \}$ with $\alpha $ running through the set
of short roots.
\item $3$ sets of long roots $\{2\alpha _{1}+3\alpha _{2}+4\alpha
_{3}+2\alpha _{4},\alpha _{2},\alpha _{2}+2\alpha _{3},\alpha _{2}+2\alpha
_{3}+2\alpha _{4}\}$, $\{\alpha _{1}+3\alpha _{2}+4\alpha _{3}+2\alpha
_{4},\alpha _{1}+\alpha _{2},\alpha _{1}+\alpha _{2}+2\alpha _{3},\alpha
_{1}+\alpha _{2}+2\alpha _{3}+2\alpha _{4}\}$ and
$\{\alpha _{1}+2\alpha _{2}+4\alpha _{3}+2\alpha _{4},\alpha _{1},\alpha
_{1}+2\alpha _{2}+2\alpha _{3},\alpha _{1}+2\alpha _{2}+2\alpha _{3}+2\alpha
_{4}\}$.
\end{itemize}
We let $\{\omega _{1},\omega _{2},\omega _{3},\omega _{4}\}$ be the
fundamental weights.
The fundamental weight $\omega _{4}$ is also the short\ positive roo
\begin{equation*}
\omega _{4}=\alpha _{1}+2\alpha _{2}+3\alpha _{3}+2\alpha _{4}.
\end{equation*
We look at its $M$-equivalence class by the same method of the $E_{l}$'s.
The set of roots orthogonal to the fundamental weight $\omega _{4}$ is
spanned by $\{\alpha _{1},\alpha _{2},\alpha _{3}\}$ which is a $B_{3}$
Dynkin diagram. Now if $\beta $ is a root of $B_{3}$ then there exists a
root $\gamma $ (in $B_{3}$) such that $\langle \gamma ^{\vee },\beta \rangle
$ is odd. It follows by Proposition \ref{propmequivalent} and its Corollary
\ref{cormequivalent} that $\{\omega _{4}\}$ is an $M$-equivalence class.
This gives the classes of the short roots.
As to the long roots we first recall that they form a $D_{4}$ root system
(see ????). Now if $\gamma $ is a short root and $\alpha $ a long root then
\langle \gamma ^{\vee },\alpha \rangle $ is even. Hence to check if $\alpha
\sim _{M}\beta $ for the two long roots $\alpha $ and $\beta $ it is enough
to test the condition of Proposition \ref{propmequivalent} when $\gamma $ is
also a long root. This means that two long roots are $M$-equivalent if and
only if they are equivalent as roots of $D_{4}$. Since no short root is $M
-equivalent to a long root we conclude that the classes of $D_{4}$ are also
M$-equivalence classes in $F_{4}$. These are the three sets with four
orthogonal roots each as stated. (To get these sets start with the highest
root $\omega _{1}=2\alpha _{1}+3\alpha _{2}+4\alpha _{3}+2\alpha _{4}$. Then
the first set is $\omega _{1}$ together with the long roots orthogonal to
it. The next two sets are obtained by applying first the reflection
r_{\alpha _{1}}$ and then $r_{\alpha _{2}}$.)
\ \
\section{Auxiliary lemmas\label{seclemaux}}
In this section we prove some lemmas to be used later in the determination
of the irreducible $K_{\Theta }$-invariant subspaces of $\mathfrak{n
_{\Theta }^{-}$. We choose once and for all a generator $E_{\alpha }\in
\mathfrak{g}_{\alpha }$ for each root space.
Recall that in Section \ref{secisotro} we denoted the irreducible components
for the adjoint representation of $\mathfrak{z}_{\Theta }$ on $\mathfrak{n
_{\Theta }^{-}$ by $V_{\Theta }^{\sigma }$. Write $\Pi _{\Theta }^{\sigma
}\subset \Pi ^{-}\backslash \langle \Theta \rangle ^{-}$ for the set of
roots such that
\begin{equation*}
V_{\Theta }^{\sigma }=\sum_{\alpha \in \Pi _{\Theta }^{\sigma }}\mathfrak{g
_{\alpha }.
\end{equation*
The subgroup $K_{\Theta }$ leave invariant $V_{\Theta }^{\sigma }$, hence
\Pi _{\Theta }^{\sigma }$ is $\mathcal{W}_{\Theta }$-invariant. \
Our first results give conditions ensuring that a \ $\mathfrak{z}_{\Theta }
-invariant subspace is $K_{\Theta }$-irreducible.
\begin{lema}
\label{leminterinvMclass}Let $V=\sum_{\alpha \in \Pi _{V}}\mathfrak{g
_{\alpha }$ be a $\mathfrak{z}_{\Theta }$-invariant subspace and suppose
that $W\subset V$ is a $K_{\Theta }$-invariant subspace. Take
X=\sum_{\alpha \in \Pi _{V}}a_{\alpha }E_{\alpha }\in W$ and let $\alpha $
be a root such that $a_{\alpha }\neq 0$. Define
\begin{equation*}
V_{[\alpha ]_{M}}=\sum_{\beta \in \left[ \alpha \right] _{M}\cap \Pi _{V}
\mathfrak{g}_{\beta }.
\end{equation*}
Then $W\cap V_{[\alpha ]_{M}}\neq \{0\}$.
\end{lema}
\begin{profe}
Let $c_{X,\alpha }$ be the cardinality of $\{\beta \notin \lbrack \alpha
]_{M}:a_{\beta }\neq 0\}$. If $c_{X,\alpha }=0$ we are done. Otherwise we
find $0\neq Y\in U$ with $c_{Y,\alpha }<c_{X,\alpha }$. In fact, if
c_{X,\alpha }>0$ then there are $\beta \notin \lbrack \alpha ]_{M}$ with
a_{\alpha },a_{\beta }\neq 0$. So that there exists $m\in M$ with
mE_{\alpha }=E_{\alpha }$ and $mE_{\beta }=-E_{\beta }$.
Now, $M\subset K_{\Theta }$, hence $Y=X+mX\in W$. Clearly $Y\neq 0$ and
since the $\beta $ component of $Y$ is zero we have $c_{Y,\alpha
}<c_{X,\alpha }$. Repeating this argument successively we arrive at $Z\in W$
such that $c_{Z,\alpha }=0$, concluding the proof.
\end{profe}
\begin{lema}
\label{lemirreducible}Take a subset $\Pi _{V}\subset \Pi ^{-}\backslash
\langle \Theta \rangle ^{-}$ such that the subspace $V=\sum_{\alpha \in \Pi
_{V}}\mathfrak{g}_{\alpha }$ is $\mathfrak{z}_{\Theta }$-invariant. Suppose
that
\begin{enumerate}
\item $\mathcal{W}_{\Theta }$ acts transitively on $\Pi _{V}$, and
\item two different roots in $\Pi _{V}$ are not $M$-equivalent.
\end{enumerate}
Then $V$ is $K_{\Theta }$-irreducible.
\end{lema}
\begin{profe}
Let $U\subset V$ be a nontrivial $K_{\Theta }$-invariant subspace. By
transitivity of $\mathcal{W}_{\Theta }\subset K_{\Theta }$ it is enough to
prove that $U$ contains a root space $\mathfrak{g}_{\alpha }$, $\alpha \in
\Pi _{V}$. But this follows by the previous lemma and the assumption that
different roots in $\Pi _{V}$ are not $M$-equivalent.
\end{profe}
As a complement of the above lemma we exhibit next general cases where
\mathcal{W}_{\Theta }$ acts transitively on sets of roots.
\begin{lema}
\label{lemtransimplelong}Let $\Pi _{\Theta }^{\sigma }\subset \Pi
^{-}\backslash \langle \Theta \rangle ^{-}$ be the set of roots
corresponding to an irreducible component $V_{\Theta }^{\sigma
}=\sum_{\alpha \in \Pi _{\Theta }^{\sigma }}\mathfrak{g}_{\alpha }$. In each
of the following cases $\mathcal{W}_{\Theta }$ acts transitively on $\Pi
_{\Theta }^{\sigma }$.
\begin{enumerate}
\item The Dynkin diagram of $\mathfrak{g}$ has only simple edges ($A_{l}$,
D_{l}$, $E_{6}$, $E_{7}$ and $E_{8}$).
\item For the diagrams $B_{l}$, $C_{l}$ and $F_{4}$ there are the cases:
\begin{enumerate}
\item The roots in $\Theta \subset \Sigma $ are long.
\item The roots in $\Pi _{\Theta }^{\sigma }$ are short.
\end{enumerate}
\end{enumerate}
\end{lema}
\begin{profe}
Let $\mu $ be the highest root of $\Pi _{\Theta }^{\sigma }$. By
representation theory we know that any other root $\beta \in \Pi _{\Theta
}^{\sigma }$ (weight of the representation) is given by
\begin{equation*}
\beta =\mu -\alpha _{1}-\cdots -\alpha _{k}
\end{equation*
with $\alpha _{i}\in \Theta $ and such that any partial difference $\mu
-\alpha _{1}-\cdots -\alpha _{i}$, $i\leq k$, is also a root. Fix $i\leq k$
put $\delta =\mu -\alpha _{1}-\cdots -\alpha _{i-1}$ and let $r_{\alpha
_{i}} $ be the reflection with respect to $\alpha _{i}$.
We claim that $\delta -\alpha _{i}=r_{\alpha _{i}}\left( \delta \right) $.
This follows by the Killing formula applied to the string of roots $\delta
+k\alpha _{i}$. There are the following cases:
\begin{enumerate}
\item In the simply laced diagrams of (1) the Killing number
\begin{equation*}
\langle \alpha _{i}^{\vee },\delta \rangle =\frac{2\langle \alpha
_{i},\delta \rangle }{\langle \alpha _{i},\alpha _{i}\rangle }
\end{equation*
is $0$ or $\pm 1$. Since $\delta -\alpha _{i}$ is a root we have $\langle
\alpha _{i}^{\vee },\delta \rangle =1$, and hence $r_{\alpha _{i}}\left(
\delta \right) =\delta -\alpha _{i}$.
\item If the roots in $\Theta $ are long as in (2a) then $\alpha _{i}$ is a
long root implying that $\langle \alpha _{i}^{\vee },\delta \rangle $ is $0$
or $\pm 1$. Again, the fact that $\delta -\alpha _{i}$ is a root implies
that $\langle \alpha _{i}^{\vee },\delta \rangle =1$, so that $r_{\alpha
_{i}}\left( \delta \right) =\delta -\alpha _{i}$.
\item If the roots in $\Pi _{\Theta }^{\sigma }$ are short in a double laced
diagram as in (2b) then $\delta $ and $\delta -\alpha _{i}$ are a short
roots. If $\alpha _{i}$ is a long root then $\delta $ and $\delta -\alpha
_{i}$ are the only roots of the form $\delta +k\alpha _{i}$, $k\in \mathbb{Z}
$. Hence by the Killing formula $\langle \alpha _{i}^{\vee },\delta \rangle
=1$, that is $r_{\alpha _{i}}\left( \delta \right) =\delta -\alpha _{i}$.
On the other hand if $\alpha _{i}$ is short then there are two possibilities
for the string of roots $\delta +k\alpha _{i}$: i) $\delta -\alpha _{i}$,
\delta $ and $\delta +\alpha _{i}$ are roots in which case $\langle \alpha
_{i},\delta \rangle =0$ and $\delta -\alpha _{i}$ and $\delta +\alpha _{i}$
are long roots; ii) $\delta -\alpha _{i}$ and $\delta $ are roots and
\langle \alpha _{i},\delta \rangle =1$. The first case is ruled out because
otherwise we would have the long roots $\delta -\alpha _{i},\delta +\alpha
_{i}\in \Pi _{\Theta }^{\sigma }$, contradicting the assumption. Therefore
\langle \alpha _{i},\delta \rangle =1$, that is, $r_{\alpha _{i}}\left(
\delta \right) =\delta -\alpha _{i}$.
\end{enumerate}
Since $r_{\alpha _{i}}\in \mathcal{W}_{\Theta }$, it follows by induction
that $\beta $ belongs to the $\mathcal{W}_{\Theta }$-oribt of $\mu $,
proving transitivity of $\mathcal{W}_{\Theta }$. \
\end{profe}
We turn now to the equivalence of irreducible representations.
\begin{lema}
\label{lemnotequivalent}Let $V_{\Theta }^{\sigma }$ and $V_{\Theta }^{\tau }$
be $\mathfrak{z}_{\Theta }$-irreducible components. Suppose that there
exists $\alpha \in \Pi _{\Theta }^{\sigma }$ which is not $M$-equivalent to
any $\beta \in \Pi _{\Theta }^{\tau }$. Then $V_{\Theta }^{\sigma }$ and
V_{\Theta }^{\tau }$ are not $K_{\Theta }$-equivalent.
\end{lema}
\begin{profe}
Suppose to the contrary that there exists an isomorphism $T:V_{\Theta
}^{\sigma }\rightarrow V_{\Theta }^{\tau }$ intertwining the $K_{\Theta }
-representations. In particular
\begin{equation*}
TmX=mTX,
\end{equation*
for all $m\in M\subset K_{\Theta }$ and $X\in V_{\Theta }^{\sigma }$.
Take $0\neq E_{\alpha }\in \mathfrak{g}_{\alpha }$. Then for every $m\in M$
we have $mE_{\alpha }=\allowbreak \varepsilon _{m}E_{\alpha }$ with
\varepsilon _{m}=\pm 1$. Write
\begin{equation*}
TE_{\alpha }=\sum_{\beta \in \Pi _{\Theta }^{\tau }}a_{\beta }E_{\beta }.
\end{equation*
Then for $m\in M$ we have
\begin{equation*}
\varepsilon _{m}\sum_{\beta \in \Pi _{\Theta }^{\tau }}a_{\beta }E_{\beta
}=\varepsilon _{m}TE_{\alpha }=mTE_{\alpha }=\sum_{\beta \in \Pi _{\Theta
}^{\tau }}a_{\beta }mE_{\beta }.
\end{equation*
Since $mE_{\beta }=\pm E_{\beta }$ and the set $E_{\beta }$ is linearly
independent, it follows that $mE_{\beta }=\varepsilon _{m}E_{\beta }$ if
a_{\beta }\neq 0$. For any such $\beta $ the representation of $M$ on
\mathfrak{g}_{\beta }$ is equivalent to the representation on $\mathfrak{g
_{\alpha }$. This contradicts the assumption that $\alpha $ is not $M
-equivalent to $\beta \in \Pi _{\Theta }^{\tau }$.
\end{profe}
The next statement gives a sufficient condition for equivalence.
\begin{proposicao}
\label{propequivalent}Let $V_{\Theta }^{\sigma }$ and $V_{\Theta }^{\tau }$
be $\mathfrak{z}_{\Theta }$-irreducible components. Suppose that there is a
bijection $\iota :\Pi _{\Theta }^{\sigma }\rightarrow \Pi _{\Theta }^{\tau }$
such that $\mathfrak{g}_{\alpha }$ and $\mathfrak{g}_{\iota \left( \alpha
\right) }$ are $M$-equivalent for every $\alpha \in \Pi _{\Theta }^{\sigma }
. Assume also that the linear map $T:V_{\Theta }^{\sigma }\rightarrow
V_{\Theta }^{\tau }$, given by $TE_{\alpha }=E_{\iota \left( \alpha \right)
} $, commutes with $\mathrm{ad}\left( X\right) $, $X\in \mathfrak{k}_{\Theta
}$ for every $\alpha \in \Pi _{\Theta }^{\sigma }$. Then $T$ is an
intertwining operator for the $K_{\Theta }$-representations on $V_{\Theta
}^{\sigma }$ and $V_{\Theta }^{\tau }$. Moreover the subspaces
\begin{equation*}
V_{[(x,y)]}=\left\{ xX+yTX:X\in V_{\Theta }^{\sigma }\right\} ,
\end{equation*
where $[(x,y)]\in \mathbb{RP}^{2}$, are the only $K_{\Theta }$-invariant
subspaces in $V_{\Theta }^{\sigma }\oplus V_{\Theta }^{\tau }$.
\end{proposicao}
\begin{profe}
The first assumption implies that $T$ intertwines the $M$-representations,
while the second assumption means that $T$ intertwines the representations
of $\left( K_{\Theta }\right) _{0}$. Since $K_{\Theta }=M\left( K_{\Theta
}\right) _{0}$, we conclude that $T$ is in fact an intertwining operator for
the $K_{\Theta }$-representations.
This implies that $V_{[(x,y)]}$ is a $K_{\Theta }$-invariant subspace in
V_{\Theta }^{\sigma }\oplus V_{\Theta }^{\tau }$ for any $[(x,y)]\in \mathbb
RP}^{2}$.
Now, if $V$ is a $K_{\Theta }$-invariant subspace in $V_{\Theta }^{\lambda
}\oplus V_{\Theta }^{\mu }$ different from $V_{\Theta }^{\sigma
}=V_{[(1,0)]} $ or $V_{\Theta }^{\tau }=V_{[(0,1)]}$, then there exist a
linear isomorphism $L:V_{\Theta }^{\sigma }\rightarrow V_{\Theta }^{\tau }$
such that
\begin{equation*}
V=\{X+LX:X\in V_{\Theta }^{\lambda }\}
\end{equation*
and
\begin{equation*}
LkX=kLX,
\end{equation*
for every $k\in K_{\Theta }$. Since $M$ is a subset of $K_{\Theta }$ and
since $mE_{\alpha }=\varepsilon _{m}E_{\alpha }$, we can argue as in the
proof of the previous Lemma to show that
\begin{equation*}
LE_{\alpha }=y_{i}E_{\iota \left( \alpha \right) }.
\end{equation*
Since $\mathcal{W}_{\Theta }$ acts transitively in the set of the directions
$\{\mathfrak{g}_{1}^{\lambda },\ldots ,\mathfrak{g}_{n_{\lambda }}^{\lambda
}\}$, we conclude that $y_{i}=y$, is independent of the index $i\in
\{1,\ldots ,n_{\lambda }\}$. Thus we have that $V=V_{[(1,y)]}$, concluding
the proof.
\end{profe}
The previous results are complemented by the following standard basic fact
in representation theory.
\begin{proposicao}
\label{propsumirredcomp}Let $V$ be the space of a finite dimensional
representation of a group $L$. Suppose that
\begin{equation*}
V=V_{1}\oplus \cdots \oplus V_{s}
\end{equation*
with $V_{i}$ invariant and irreducible. If the representations of $L$ on
different components $V_{i}$, $V_{j}$, $i\neq j$, are not equivalent then
the only $L$-invariant subspaces are sums of the components.
\end{proposicao}
\begin{profe}
(Sketch) If $\{0\}\neq W\subset V$ is an invariant subspace then the
projection $W_{i}$ to $V_{i}$ is invariant and hence either $\{0\}$ or
V_{i} $. Suppose that there are $i\neq j$ such that $W_{i}=V_{i}$ and
W_{j}=V_{j}$ and write $W_{ij}$ for the projection on $V_{i}\oplus V_{j}$.
Then $W_{ij}\cap V_{i}$ is $\{0\}$ or $V_{i}$. If $W_{ij}\cap V_{i}=V_{i}$
then $W_{ij}=V_{i}\oplus V_{j}$, which implies that $V_{i}\oplus
V_{j}\subset W$. Otherwise $W_{ij}\cap V_{i}=W_{ij}\cap V_{j}=\{0\}$. In
this case $W_{ij}$ is the graph of an isomorphism $V_{i}\rightarrow V_{j}$,
intertwining the representations on $V_{i}$ and $V_{j}$.
\end{profe}
Finally for several split simple Lie algebras the compact subalgebra
\mathfrak{k}$ is not simple. Via the next lemma we exploit this fact to get
K_{\Theta }$-invariant subspaces in $\mathfrak{n}_{\Theta }^{-}$.
\begin{lema}
\label{lemorbitnormal}Let $U\subset K$ be a normal subgroup and denote by
V\subset \mathfrak{n}_{\Theta }^{-}$ the tangent space to the $U$-orbit
U\cdot b_{\Theta }$ through the origin. Then $V$ is $K_{\Theta }$-invariant.
\end{lema}
\begin{profe}
The orbit $U\cdot b_{\Theta }$ is invariant by $K_{\Theta }$. In fact, if
u\cdot b_{\Theta }\in U\cdot b_{\Theta }$ and $k\in K_{\Theta }$ then
kuk^{-1}\in U$ so that $ku\cdot b_{\Theta }=kuk^{-1}\cdot kb_{\Theta
}=kuk^{-1}\cdot b_{\Theta }$ belongs to $U\cdot b_{\Theta }$. Hence its
tangent space at $b_{\Theta }$ is invariant by the isotropy representation.
\end{profe}
\section{Irreducible $K_{\Theta }$-invariant subspaces\label{secirreduc}}
Inn this section we describe \ the previous results to each diagram.
\subsection{Flags of $A_{l}=\mathfrak{sl}\left( l+1,\mathbb{R}\right) $}
As checked in Section \ref{secmequiv} no two different negative roots of
A_{l}$ are $M$-equivalent if $l\neq 3$. On the other hand by Lemma \re
{lemtransimplelong}, on any flag manifold of $A_{l}$, the subgroup $\mathcal
W}_{\Theta }$ acts transitively on each set of roots $\Pi _{\Theta }^{\sigma
}$ corresponding to an irreducible representation of $\mathfrak{z}_{\Theta }$
on $V_{\Theta }^{\sigma }$. Therefore by Lemma \ref{lemirreducible} we
conclude that $K_{\Theta }$ is irreducible on each $V_{\Theta }^{\sigma }$.
Looking again the $M$-equivalence classes we see that two different
irreducible subspaces are not $K_{\Theta }$-equivalent. Hence we get the
following description of the $K_{\Theta }$-invariant irreducible subspaces
in a flag manifold of $A_{l}$.
\begin{proposicao}
For any flag manifold $\mathbb{F}_{\Theta }$ of $A_{l}$, $l\neq 3$, the
K_{\Theta }$-invariant irreducible subspaces are the irreducible components
V_{\Theta }^{\sigma }$ for the $\mathfrak{z}_{\Theta }$ representation. Two
such representations are not $K_{\Theta }$-equivalent.
\end{proposicao}
The irreducible components $V_{\Theta }^{\sigma }$ are easily described in
terms of the matrices in $\mathfrak{sl}(n,\mathbb{R})$, $n=l+1$. In fact,
let
\begin{equation*}
H_{\Theta }=\mathrm{diag}\{a_{1},\ldots ,a_{n}\}\in \mathfrak{a}_{\Theta
}\cap \mathrm{cl}\mathfrak{a}^{+}\quad a_{1}\geq \cdots \geq a_{n}
\end{equation*
be characteristic for $\Theta $. The multiplicities of the eigenvalues of
H_{\Theta }$ determine the sizes of a block decomposition of the $n\times n$
matrices. With respect to this decomposition the matrices in $\mathfrak{z
_{\Theta }$ are block diagonal while a block outside the diagonal determines
a $\mathfrak{z}_{\Theta }$-irreducible component. These are also the
K_{\Theta }$-irreducible components.
Now we look at the case $l=3$. The matrix
\begin{equation*}
\left(
\begin{array}{llll}
\ast & a & b & c \\
a & \ast & c & b \\
b & c & \ast & a \\
c & b & a & \as
\end{array
\right)
\end{equation*
summarizes the $M$-equivalence classes of $\mathfrak{sl}\left( 4,\mathbb{R
\right) $, where root spaces represented by the same letter are $M
-equivalent (see Section \ref{secmequiv}).
This shows that in the maximal flag manifold $\mathbb{F}_{\Theta }$, $\Theta
=\emptyset $, the $K_{\Theta }=M$ invariant irreducible subspaces are the
one-dimensional subspaces of $\mathfrak{g}_{21}\oplus \mathfrak{g}_{43}$,
\mathfrak{g}_{31}\oplus \mathfrak{g}_{42}$ or $\mathfrak{g}_{32}\oplus
\mathfrak{g}_{41}$. The irreducible subspaces in the other flag manifolds
are easily obtained from this $M$-equivalence.
We discuss further the instructive case when $\mathbb{F}_{\Theta }=\mathrm{G
}_{2}\left( 4\right) $, the Grassmannian of two dimensional subspaces of
\mathbb{R}^{4}$. In this case $\mathfrak{n}_{\Theta }^{-}$ is the subalgebra
of matrices writen in $2\times 2$ blocks as
\begin{equation*}
X=\left(
\begin{array}{ll}
0 & 0 \\
B &
\end{array
\right) .
\end{equation*
The representations of $\mathfrak{z}_{\Theta }$ and $\mathfrak{g}\left(
\Theta \right) $ on $\mathfrak{n}_{\Theta }^{-}$ are irreducible. Here
K_{\Theta }=\mathrm{SO}\left( 2\right) \times \mathrm{SO}\left( 2\right) $
whose representation on $\mathfrak{n}_{\Theta }^{-}$ decomposes into two $2
-dimensional irreducible subspaces. This is due to the fact that $\mathfrak
so}\left( 4\right) =\mathfrak{so}\left( 3\right) _{1}\oplus \mathfrak{so
\left( 3\right) _{2}$ is a sum of two copies of $\mathfrak{so}\left(
3\right) $. The matrices in these components have the for
\begin{equation}
\mathfrak{so}\left( 3\right) _{1}:\left(
\begin{array}{cc}
A & -B^{T} \\
B &
\end{array
\right) \qquad \mathfrak{so}\left( 3\right) _{2}:\left(
\begin{array}{cc}
A & -B^{T} \\
B & -
\end{array
\right) \label{fordecomso4so3so3}
\end{equation
with $A+A^{T}=0$ where $B$ is symmetric with $\mathrm{tr}B=0$ for $\mathfrak
so}\left( 3\right) _{1}$ while
\begin{equation}
B=\left(
\begin{array}{cc}
a & -b \\
b &
\end{array
\right) \label{forBcomplex}
\end{equation
for $\mathfrak{so}\left( 3\right) _{2}$. Hence by Lemma \ref{lemorbitnormal
, the tangent spaces $V_{i}$ to orbits of $\mathrm{SO}\left( 3\right)
_{i}=\langle \exp \mathfrak{so}\left( 3\right) _{i}\rangle $, $i=1,2$, are
K_{\Theta }$-invariant. The subspace $V_{i}$, $i=1,2$, is given by the
matrices in $\mathfrak{n}_{\Theta }^{-}$ with $B$ as $\mathfrak{so}\left(
3\right) _{1}$ or $\mathfrak{so}\left( 3\right) _{2}$, respectively.
\subsection{Flags of $B_{l}=\mathfrak{sl}\left( l+1,l\right) $}
In the standard realization $\mathfrak{sl}\left( l+1,l\right) $ is the
algebra of matrice
\begin{equation*}
\left(
\begin{array}{ccc}
0 & a & b \\
-b^{T} & A & B \\
-a^{T} & C & -A^{T
\end{array
\right) \qquad B+B^{T}=C+C^{T}=0.
\end{equation*
In this case $\mathfrak{a}$ is the subalgebra of matrices
\begin{equation*}
\left(
\begin{array}{ccc}
0 & 0 & 0 \\
0 & \Lambda & 0 \\
0 & 0 & -\Lambd
\end{array
\right)
\end{equation*
with $\Lambda =\mathrm{diag}\{a_{1},\ldots ,a_{l}\}$. The set of roots are
i) the long ones $\pm \left( \lambda _{i}-\lambda _{j}\right) $ and $\pm
\left( \lambda _{i}+\lambda _{j}\right) $, $1\leq i<j\,\leq l$ and ii) the
short ones $\pm \lambda _{i}$, $1\leq i\leq l$. The set of simple roots is
\Sigma =\{\lambda _{1}-\lambda _{2},\ldots ,\lambda _{l-1}-\lambda
_{l},\lambda _{l}\}$, which we write also as $\Sigma =\{\alpha _{1},\ldots
,\alpha _{l}\}$, that is, $\alpha _{i}=\lambda _{i}-\lambda _{i+1}$ if $i<l$
and $\alpha _{l}=\lambda _{l}$.
The Weyl chamber $\mathfrak{a}^{+}\subset \mathfrak{a}$ is defined by the
inequalities
\begin{equation*}
a_{1}>a_{2}>\cdots >a_{l-1}>a_{l}>0,
\end{equation*
and a partial chamber $\mathfrak{a}_{\Theta }\cap \mathrm{cl}\mathfrak{a
^{+} $ is defined by a similar relations where some of the strict
inequalities are changed by equalities (e.g. if $\lambda _{i}-\lambda
_{j}\in \Theta $ then $a_{i}=a_{j}$). In particular a characteristic element
$H_{\Theta }$ for the subset $\Theta =\{\alpha \in \Sigma :\alpha \left(
H_{\Theta }\right) =0\}\subset \Sigma $ is defined by one of these relations.
The subalgebra $\mathfrak{k}$ is composed of the skew-symmetric matrices in
\mathfrak{sl}\left( l,l\right) $, that is
\begin{equation*}
\left(
\begin{array}{ccc}
0 & a & a \\
-a^{T} & A & B \\
-a^{T} & B &
\end{array
\right) \qquad A+A^{T}=B+B^{T}=0.
\end{equation*
It is isomorphic to $\mathfrak{so}\left( l+1\right) \oplus \mathfrak{so
\left( l\right) $. The isomorphism is provided by the decompositio
\begin{equation*}
\left(
\begin{array}{lll}
0 & a & a \\
-a^{T} & A & B \\
-a^{T} & B &
\end{array
\right) =\left(
\begin{array}{lll}
0 & a & a \\
-a^{T} & \left( A+B\right) /2 & \left( A+B\right) /2 \\
-a^{T} & \left( A+B\right) /2 & \left( A+B\right) /
\end{array
\right) +\left(
\begin{array}{lll}
0 & 0 & 0 \\
0 & \left( A-B\right) /2 & -\left( A-B\right) /2 \\
0 & -\left( A-B\right) /2 & \left( A-B\right) /
\end{array
\right) ,
\end{equation*
so that $\mathfrak{k}=\mathfrak{k}_{l+1}\oplus \mathfrak{k}_{l}\approx
\mathfrak{so}\left( l+1\right) \oplus \mathfrak{so}\left( l\right) $ where
the ideals are given by matrices as follow
\begin{equation}
\mathfrak{k}_{l+1}:\left(
\begin{array}{ccc}
0 & a & a \\
-a^{T} & A & A \\
-a^{T} & A &
\end{array
\right) \qquad \mathfrak{k}_{l}:\left(
\begin{array}{ccc}
0 & 0 & 0 \\
0 & A & -A \\
0 & -A &
\end{array
\right) . \label{forideaiskbele}
\end{equation
In both cases $A$ is skew-symmetric. We write $K_{l+1}=\langle \exp
\mathfrak{k}_{l+1}\rangle $ and $K_{l+1}=\langle \exp \mathfrak{k
_{l}\rangle $.
We start our analysis by describing the irreducible components $V_{\Theta
}^{\sigma }$ defined by the set of roots $\Pi _{\Theta }^{\sigma }$. For
this we separate the cases where $\lambda _{l}$ belongs or not to $\Theta $.
\begin{lema}
\label{lembelecompseml}Suppose that $\lambda _{l}\notin \Theta $ and let
\begin{equation*}
V_{\Theta }^{\sigma }=\sum_{\alpha \in \Pi _{\Theta }^{\sigma }}\mathfrak{g
_{\alpha }
\end{equation*
be an irreducible component. Then $\Pi _{\Theta }^{\sigma }$ contains only
short roots or long roots. These sets are described as follows:
\begin{enumerate}
\item \textbf{Short roots:} Take a simple root $\alpha _{i}\notin \Theta $.
Then there are two possibilities:
\begin{enumerate}
\item $\alpha _{i-1}\notin \Theta $. Then $\mathfrak{g}_{-\lambda _{i}}$ is
\mathfrak{z}_{\Theta }$-invariant and hence is an irreducible component.
\item $\alpha _{i-1}\in \Theta $. Let $j\left( i\right) <i$ be the smallest
index such that $\{\alpha _{j\left( i\right) },\ldots ,\alpha
_{i-1}\}\subset \Theta $. Then $\Pi _{\Theta }^{\sigma }=\{-\lambda
_{j\left( i\right) },\ldots ,-\lambda _{i-1}\}$ defines a $\mathfrak{z
_{\Theta }$-irreducible component.
\end{enumerate}
These sets form a disjoint union of the negative short roots $-\lambda _{i}
, $1\leq i\leq l$. (Note that this disjoint union completely determines
\Theta $.)
\item \textbf{Long roots:} A subset $\Pi _{\Theta }^{\sigma }$ contains only
roots of the type $\lambda _{i}-\lambda _{j}$ or of the type $-\lambda
_{i}-\lambda _{j}$.
\end{enumerate}
\end{lema}
\begin{profe}
To see the components corresponding to the short roots take an index $i$
with $\alpha _{i}\notin \Theta $. An easy check shows that the only simple
roots $\alpha $ such that $-\lambda _{i}+\alpha $ is a root are $\alpha
=\lambda _{i}-\lambda _{i+1}$ or $\alpha =\lambda _{l}$. By assumption these
simple roots are not in $\Theta $. This implies that $-\lambda _{i}$ is the
highest weight of an irreducible representation of $\mathfrak{g}\left(
\Theta \right) $. The weights of this representation are restrictions of
\mathfrak{a}\left( \Theta \right) $ of roots. They have the form $-\lambda
_{i}-\beta _{1}-\cdots -\beta _{k}$ with $\beta _{i}\in \Theta $. But these
successive diferences are roots only when $\beta _{1}=\lambda _{i-1}-\lambda
_{i}$, $\beta _{2}=\lambda _{i-2}-\lambda _{i-1}$, and so on, obtaining the
-\lambda _{i}$, $-\lambda _{i-1}$, \ldots , $-\lambda _{j\left( i\right) }$
with $j\left( i\right) $ as in (b). This concludes the case of the short
roots.
Now, take a long root, e.g. $\lambda _{i}-\lambda _{j}$. Then $\Pi _{\Theta
}\left( \lambda _{i}-\lambda _{j}\right) $ does not contain short roots that
were already exhausted. On the other hand, by assumption $\Theta $ is
contained in the set of roots of the type $\lambda _{r}-\lambda _{s}$. Since
this set is closed by sum we conclude that the roots in $\Pi _{\Theta
}\left( \lambda _{i}-\lambda _{j}\right) $ have the type $\lambda
_{r}-\lambda _{s}$. The same argument applies to $\Pi _{\Theta }\left(
-\lambda _{i}-\lambda _{j}\right) $.
\end{profe}
\begin{lema}
\label{lembelecompcoml}Suppose that $\lambda _{l}\in \Theta $ and let $i_{0}$
be the largest index such that $\lambda _{i_{0}}-\lambda _{i_{0}+1}\notin
\Theta $, that is, $\{\alpha _{i_{0}+1},\ldots ,\alpha _{l}=\lambda _{l}\}$
is the connected component of $\Theta $ containing $\lambda _{l}$. Then the
sets of roots defining the $\mathfrak{z}_{\Theta }$-irreducible components
are as follows:
\begin{enumerate}
\item \textbf{Components containing short roots}: If $i\leq i_{0}$ then $\Pi
_{\Theta }\left( -\lambda _{i}\right) $ contains $-\lambda _{i}+\lambda _{k}$
and $-\lambda _{i}-\lambda _{k}$ for all $k\geq i_{0}+1$. (The short roots
-\lambda _{i}$, $i>i_{0}$, belong to $\langle \Theta \rangle ^{-}$.)
Moreover the sets of short roots belonging to the same component are as in
Lemma \ref{lembelecompseml} (1), namely $\{-\lambda _{j\left( i\right)
},\ldots ,-\lambda _{i-1}\}$ where $\{\alpha _{j\left( i\right) },\ldots
,\alpha _{i-1}\}$ is a connected component of $\Theta $.
\item \textbf{Components containing only long roots}: If $i<j\leq i_{0}$
then $\Pi _{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) $ has only
roots $\lambda _{r}-\lambda _{s}$ and $\Pi _{\Theta }\left( -\lambda
_{i}-\lambda _{j}\right) $ has only roots $-\lambda _{r}-\lambda _{s}$.
\end{enumerate}
(These sets exhaust the roots because $-\lambda _{i},-\lambda _{i}\pm
\lambda _{j}\in \langle \Theta \rangle ^{-}$ if $i_{0}+1\leq i<j$.)
\end{lema}
\begin{profe}
By assumption $\Theta $ contains the subdiagram simple roots
B_{l-i_{0}}=\{\alpha _{i_{0}+1},\ldots ,\alpha _{l}\}$. This implies that
the roots $\pm \lambda _{k}\pm \lambda _{j}$ and $\pm \lambda _{k}$ belong
to $\langle \Theta \rangle $ if $i_{0}+1\leq k<j$. Take a short root
-\lambda _{i}$ with $i\leq i_{0}$, which is not in $\langle \Theta \rangle
^{-}$. For any root $\alpha \in \langle \Theta \rangle $ such that $-\lambda
_{i}+\alpha $ is a root we have $-\lambda _{i}+\alpha \in \Pi _{\Theta
}\left( -\lambda _{i}\right) $. If we take $\alpha =\pm \lambda _{k}$,
k\geq i_{0}+1$, we see that $-\lambda _{i}\pm \lambda _{k}\in \Pi _{\Theta
}\left( -\lambda _{i}\right) $, proving the first part of (1). By the same
argument of the proof Lemma \ref{lembelecompseml} we get the statement about
the short roots.
Now, a long root $-\lambda _{i}+\lambda _{j}$, $i<j\leq i_{0}$, is
orthogonal to every root in $B_{l-i_{0}}$. Hence the only way to get new
roots from $-\lambda _{i}+\lambda _{j}$ is by adding or subtracting roots in
$\Theta \setminus B_{l-i_{0}}$. These roots have the type $\lambda
_{r}-\lambda _{s}$, so that as in the proof of Lemma \ref{lembelecompseml}
we see that $\Pi _{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) $
contains only roots of the type $\lambda _{r}-\lambda _{s}$. The same
argument works for $-\lambda _{i}-\lambda _{j}$, $i<j\leq i_{0}$, showing
(2).
\end{profe}
The next step is to look at the $K_{\Theta }$-irreducibility of the
\mathfrak{z}_{\Theta }$-irreducible components. For this we use the $M
-equivalence classes so we are led to consider separetely different values
of $l$.
\begin{lema}
Take $B_{l}$ with $l\geq 5$.
\begin{enumerate}
\item Suppose that $\lambda _{l}\notin \Theta $. Then any component
V_{\Theta }^{\sigma }$ is $K_{\Theta }$-irreducible.
\item If $\lambda _{l}\in \Theta $ then $K_{\Theta }$ is irreducible in the
components $V_{\Theta }^{\sigma }$ such that $\Pi _{\Theta }^{\sigma }$
contains only long roots as in Lemma \ref{lembelecompcoml} (2).
\end{enumerate}
\end{lema}
\begin{profe}
We just piece together different facts proved previously. First if $\lambda
_{l}\notin \Theta $ then $\Theta $ contains only long roots. Hence by Lemma
\ref{lemtransimplelong} the subgroup $\mathcal{W}_{\Theta }$ acts
transitively on the sets $\Pi _{\Theta }^{\sigma }$ for any irreducible
component $V_{\Theta }^{\sigma }$. Now if $l\geq 5$ then the $M$-equivalence
classes are the short roots $\{-\lambda _{i}\}$ and $\{-\lambda _{i}+\lambda
_{j},-\lambda _{i}-\lambda _{j}\}$, $i<j$. Hence by Lemma \re
{lembelecompseml} the intersection of a $M$-equivalence class with a set
\Pi _{\Theta }^{\sigma }$ has just one root. Therefore, the assumptions of
Lemma \ref{lemirreducible} are satisfied, and we get the conclusion that any
$V_{\Theta }^{\sigma }$ is $K_{\Theta }$-irreducible, proving (1).
The proof of (2) is similar. Take a subset $\Pi _{\Theta }^{\sigma }$ as in
the statement. Again no two roots in $\Pi _{\Theta }^{\sigma }$ are $M
-equivalent. As to the transitive action of $\mathcal{W}_{\Theta }$ consider
the subset $B_{l-i_{0}}=\{\alpha _{i_{0}+1},\ldots ,\alpha _{l}\}$ defined
in the proof of Lemma \ref{lembelecompcoml}. Then $\mathcal{W}_{\Theta }$ is
the direct product $\mathcal{W}_{\Theta }=\mathcal{W}_{\Theta \setminus
B_{l-i_{0}}}\times \mathcal{W}_{B_{l-i_{0}}}$, and any $w\in \mathcal{W
_{B_{l-i_{0}}}$ is the identity in $\Pi _{\Theta }^{\sigma }$, because the
sets $\Pi _{\Theta }^{\sigma }$ and $B_{l-i_{0}}$ are orthogonal (see the
proof of Lemma \ref{lembelecompcoml}). Now $\mathcal{W}_{\Theta \setminus
B_{l-i_{0}}}$ acts transitively on $\Pi _{\Theta }^{\sigma }$ since $\Theta
\setminus B_{l-i_{0}}$ has only long roots (see the proof of Lemma \re
{lemtransimplelong}).
\end{profe}
It remains to analyze the components $V_{\Theta }\left( -\lambda _{i}\right)
$ containing short roots $-\lambda _{i}$ in case $\lambda _{l}\in \Theta $.
Contrary to the others these are not $K_{\Theta }$-irreducible. Let us write
them explicitly as follows: Let $i_{0}$ be, as in Lemma \ref{lembelecompcoml
, the largest index such that $\alpha _{i_{0}}=\lambda _{i_{0}}-\lambda
_{i_{0}+1}\notin \Theta $, so that $-\lambda _{i}\notin \langle \Theta
\rangle ^{-}$. For $i\leq i_{0}$ and $k>i_{0}$ put
\begin{equation*}
W_{\Theta }^{ik}=\mathfrak{g}_{-\lambda _{i}+\lambda _{k}}\oplus \mathfrak{g
_{-\lambda _{i}}\oplus \mathfrak{g}_{-\lambda _{i}-\lambda _{k}}\qquad
\mathrm{and}\qquad W_{\Theta }^{i}=\sum_{k\geq i_{0}+1}W_{\Theta }^{ik}.
\end{equation*
(Note that the last sum is not direct.)
By Lemma \ref{lembelecompcoml} (2) we have one irreducible component
V_{\Theta }\left( -\lambda _{i}\right) $ for each index $i\leq i_{0}$ such
that $\alpha _{i}=\lambda _{i}-\lambda _{i+1}\notin \Theta $. To write it in
terms of the subspaces $W_{\Theta }^{i}$ let $j\left( i\right) \leq i$ be
defined by
\begin{enumerate}
\item $j\left( i\right) =i$ if $\alpha _{i-1}\notin \Theta $, and
\item $j\left( i\right) $ is such that $\{\alpha _{j\left( i\right) },\ldots
,\alpha _{i-1}\}$ is a connected component of $\Theta $ if $\alpha _{i-1}\in
\Theta $.
\end{enumerate}
Then
\begin{equation*}
V_{\Theta }\left( -\lambda _{i}\right) =\sum_{k=j\left( i\right)
}^{i}W_{\Theta }^{k}.
\end{equation*}
Now for $i,j$ let
\begin{equation}
E_{ij}^{-}=\left(
\begin{array}{ccc}
0 & 0 & 0 \\
0 & E_{ij} & 0 \\
0 & 0 & -E_{ji
\end{array
\right) ~E_{ij}^{+}=\left(
\begin{array}{ccc}
0 & 0 & 0 \\
0 & 0 & 0 \\
0 & E_{ij}-E_{ji} &
\end{array
\right) ~E_{i}^{0}=\left(
\begin{array}{ccc}
0 & e_{i} & 0 \\
0 & 0 & 0 \\
-e_{i}^{T} & 0 &
\end{array
\right) \label{forbasesbele}
\end{equation
where $E_{ij}$ and $e_{i}$ are basic $l\times l$ and $1\times l$ matrices.
These matrices are generators of $\mathfrak{g}_{-\lambda _{i}+\lambda _{j}}
, $\mathfrak{g}_{-\lambda _{i}-\lambda _{j}}$ and $\mathfrak{g}_{-\lambda
_{i}}$, respectively. So that $\{E_{ik}^{-},E_{ik}^{+},E_{ik}^{0}\}$ is a
basis of $W_{\Theta }^{ik}$ and $\{E_{ik}^{-},E_{ik}^{+},E_{ik}^{0}:k\geq
i_{0+1}\}$ is a basis of $W_{\Theta }^{i}$.
Before proceeding we note that the subspace $W_{\Theta }^{ik}$ is invariant
and irreducible by adjoint representation of the subalgebra $\mathfrak{g
\left( \lambda _{k}\right) \approx \mathfrak{sl}\left( 2,\mathbb{R}\right) $
generated by $\mathfrak{g}_{\pm \lambda _{k}}$, thus defining an irreducible
representation of $\mathfrak{sl}\left( 2,\mathbb{R}\right) $. By
dimensionality this representation is equivalent to the adjoint
representation of $\mathfrak{sl}\left( 2,\mathbb{R}\right) $, which in turn
is not $\mathfrak{so}\left( 2\right) $-irreducible: It decomposes into the
subspaces of skew-symmetric ($1$-dimensional) and symmetric ($2
-dimensional) matrices. The equivalence $\mathfrak{g}\left( \lambda
_{k}\right) \approx \mathfrak{sl}\left( 2,\mathbb{R}\right) $ maps
\mathfrak{g}_{-\lambda _{i}+\lambda _{k}}$ and $\mathfrak{g}_{-\lambda
_{i}-\lambda _{k}}$ onto the upper and lower triangular matrices,
respectively and $\mathfrak{g}_{-\lambda _{i}}$ onto the diagonal matrices.
So we get
\begin{lema}
The representation of $\mathfrak{g}\left( \lambda _{k}\right) $ decomposes
W_{\Theta }^{ik}$ into two $\mathfrak{k}_{\{\lambda _{k}\}}$-invariant
subspaces, namel
\begin{equation*}
\left( W_{\Theta }^{ik}\right) _{1}=\mathrm{span}\{E_{ik}^{-}-E_{ik}^{+}\
\qquad \mathrm{and}\qquad \left( W_{\Theta }^{ik}\right) _{2}=\mathrm{span
\{E_{ik}^{-}+E_{ik}^{+},E_{ik}^{0}\}.
\end{equation*}
\end{lema}
Now we can decompose $V_{\Theta }\left( -\lambda _{i}\right) $, $i\leq i_{0}$
when $\lambda _{l}\in \Theta $ into $K_{\Theta }$-irreducible subspaces.
\begin{lema}
\label{lemirreducw}In $B_{l}$, $l\geq 5$, suppose $\lambda _{l}\in \Theta $
and let $i_{0}$ be the largest index such that $\lambda _{i_{0}}-\lambda
_{i_{0}+1}\notin \Theta $. If $i\leq i_{0}$ then $-\lambda _{i}\notin
\langle \Theta \rangle ^{-}$ and the $\mathfrak{z}_{\Theta }$-irreducible
component $V_{\Theta }\left( -\lambda _{i}\right) $ is the direct sum of the
following $K_{\Theta }$-irreducible and invariant subspaces
\begin{equation*}
\left( W_{\Theta }^{i}\right) _{1}=\sum_{j=j\left( i\right) }^{i}\sum_{k\geq
i_{0}+1}\left( W_{\Theta }^{jk}\right) _{1}\qquad \mathrm{and}\qquad \left(
W_{\Theta }^{i}\right) _{2}=\sum_{j=j\left( i\right) }^{i}\sum_{k\geq
i_{0}+1}\left( W_{\Theta }^{jk}\right) _{2}
\end{equation*
with $\dim \left( W_{\Theta }^{i}\right) _{1}=l-i_{0}+i-j\left( i\right) +1$
and $\dim \left( W_{\Theta }^{i}\right) _{2}=2\left( l-i_{0}+i-j\left(
i\right) +1\right) $.
Furthermore, $\left( W_{\Theta }^{i}\right) _{1}=V_{\Theta }\left( -\lambda
_{i}\right) \cap T_{b_{\Theta }}K_{l}\cdot b_{\Theta }$ and $\left(
W_{\Theta }^{i}\right) _{2}=V_{\Theta }\left( -\lambda _{i}\right) \cap
T_{b_{\Theta }}K_{l+1}\cdot b_{\Theta }$.
\end{lema}
\begin{profe}
The intersections in the last statement with the tangent space to the orbits
$K_{l}\cdot b_{\Theta }$ and $K_{l+1}\cdot b_{\Theta }$ are readily obtained
from the matrices in $\mathfrak{k}_{l}$ and $\mathfrak{k}_{l+1}$ given in
\ref{forideaiskbele}) and the definition of the subspaces $\left( W_{\Theta
}^{i}\right) _{1}$ and $\left( W_{\Theta }^{i}\right) _{2}$. It follows by
Lemma \ref{lemorbitnormal} that these subspaces are $K_{\Theta }$-invariant.
To check irreducibility consider $\left( W_{\Theta }^{i}\right) _{2}$ and
take a nonzero $K_{\Theta }$-invariant subspace $Z\subset \left( W_{\Theta
}^{i}\right) _{2}$.
We claim that there are $j\in \left[ j\left( i\right) ,i\right] $ and $k\geq
i_{0}+1$ such that $\left( W_{\Theta }^{jk}\right) _{2}\subset Z$.
By Lemma \ref{leminterinvMclass} we have a nontrivial intersection of $Z$
with a subspace $\sum_{\alpha }\mathfrak{g}_{\alpha }$, with the sum
extended to a $M$-equivalence class. Since we are assuming that $l\geq 5$,
the $M$-equivalence classes are $\{-\lambda _{s}+\lambda _{r},-\lambda
_{s}-\lambda _{r}\}$ and $\{-\lambda _{s}\}$. Hence either there exists
j\in \left[ j\left( i\right) ,i\right] $ such that $Z\cap \mathfrak{g
_{-\lambda _{j}}\neq \{0\}$ or there are $j\in \left[ j\left( i\right) ,
\right] $ and $k\geq i_{0}+1$ such that $Z\cap \left( \mathfrak{g}_{-\lambda
_{j}+\lambda _{k}}\oplus \mathfrak{g}_{-\lambda _{j}-\lambda _{k}}\right)
\neq \{0\}$. In both cases we have
\begin{equation*}
Z\cap \left( W_{\Theta }^{jk}\right) _{2}\neq \{0\}
\end{equation*
However, $\left( W_{\Theta }^{jk}\right) _{2}$ is invariant and irreducible
for $\mathfrak{k}_{\{\lambda _{k}\}}$. Since $k\geq i_{0}+1$ we have
\lambda _{k}\in \langle \Theta \rangle $ and $\mathfrak{k}_{\{\lambda
_{k}\}}\subset \mathfrak{k}_{\Theta }$. By $K_{\Theta }$-invariance of $Z$
we conclude that $\left( W_{\Theta }^{jk}\right) _{2}\subset Z$.
Now let $B_{l-i_{0}+1}$ be the connected component of $\Theta $ containing
\lambda _{l}$. Then its Weyl group $\mathcal{W}_{B_{l-i_{0}+1}}\subset
\mathcal{W}_{\Theta }$ acts transitively on the set of its short root. This
means that if $k_{1},k_{2}\geq i_{0}+1$ then there exists $w\in \mathcal{W
_{B_{l-i_{0}+1}}$ such that $w\lambda _{k_{1}}=\lambda _{k_{2}}$. Combining
this transitivity with the claim it follows that $\left( W_{\Theta
}^{jk}\right) _{2}\subset Z$ for every $k\geq i_{0}+1$. Consequently, there
exists $j\in \left[ j\left( i\right) ,i\right] $ such that $\sum_{k\geq
i_{0}+1}\left( W_{\Theta }^{jk}\right) _{2}\subset Z$.
To finish the proof we use the subgroup $\mathcal{W}_{\left[ j\left(
i\right) ,i-1\right] }$ of $\mathcal{W}_{\Theta }$ generated by the
reflections with respect to the roots in the connected component $\{\alpha
_{j\left( i\right) },\ldots ,\alpha _{i-1}\}\subset \Theta $. This subgroup
is the permutation group of $\{j\left( i\right) ,\ldots ,i-1,i\}$. Since
\sum_{k\geq i_{0}+1}\left( W_{\Theta }^{jk}\right) _{2}\subset Z$ for some
j\in \left[ j\left( i\right) ,i\right] $ we conclude $\sum_{k\geq
i_{0}+1}\left( W_{\Theta }^{sk}\right) _{2}\subset Z$ for all $s\in \left[
j\left( i\right) ,i\right] $, so that $\left( W_{\Theta }^{i}\right)
_{2}\subset Z$, showing irreducibility of $\left( W_{\Theta }^{i}\right)
_{2} $.
The proof for $\left( W_{\Theta }^{i}\right) _{1}$ is similar.
\end{profe}
Summarizing the above discussion we have the following $K_{\Theta }
-invariant irreducible subspaces for $B_{l}$, $l\geq 5$:
\begin{enumerate}
\item A $\mathfrak{z}_{\Theta }$-component $V_{\Theta }\left( -\lambda
_{i}\right) $ containing only short roots. These components occur only when
\lambda _{l}\notin \Theta $.
\item $\mathfrak{z}_{\Theta }$-components \ $V_{\Theta }\left( -\lambda
_{i}+\lambda _{j}\right) $ and $V_{\Theta }\left( -\lambda _{i}-\lambda
_{j}\right) $, $i<j$, containing only long roots. These subspaces occur in
both cases when $\lambda _{l}$ belongs or not to $\Theta $. When $\lambda
_{l}\in \Theta $ the indexes $i,j$ satisfy $i<j\leq i_{0}$ where $\{\alpha
_{i_{0}},\ldots ,\alpha _{l}=\lambda _{l}\}$ is the connected component of
\Theta $ containing $\lambda _{l}$.
\item The subspaces $\left( W_{\Theta }^{i}\right) _{1}$ and $\left(
W_{\Theta }^{i}\right) _{2}$ contained in a $\mathfrak{z}_{\Theta }
-component \ $V_{\Theta }\left( -\lambda _{i}\right) $ when $\lambda _{l}\in
\Theta $.
\end{enumerate}
These are not the only invariant irreducible subspaces of $K_{\Theta }$,
since among them some pairs $V_{1}\neq V_{2}$ are $K_{\Theta }$-equivalent,
enabling the existence of invariant subspaces inside $V_{1}\oplus V_{2}$.
Among these pairs we can discard the following by $M$-equivalence we discard
the following pairs: i) $V_{1}$ is a subspace in item (1) and $V_{2}$ in (1)
or (2); ii) $V_{1}$ is a subspace in (2) and $V_{2}$ in (3); iii) $V_{1}$ is
a subspace $\left( W_{\Theta }^{i}\right) _{1,2}$ and $V_{2}=\left(
W_{\Theta }^{j}\right) _{1,2}$ if $i\neq j$. Since (1) and (3) are subspaces
for different $\Theta $ and $\left( W_{\Theta }^{i}\right) _{1}$ and $\left(
W_{\Theta }^{i}\right) _{2}$ do not have the same dimension, it remains the
subspaces $V_{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) $ and
V_{\Theta }\left( -\lambda _{i}-\lambda _{j}\right) $ of (2). These are
indeed equivalent.
\begin{lema}
In $B_{l}$, $l\geq 5$, the subspaces $V_{\Theta }\left( -\lambda
_{i}+\lambda _{j}\right) $ and $V_{\Theta }\left( -\lambda _{i}-\lambda
_{j}\right) $ as in (2) above are $K_{\Theta }$-equivalent if both roots
-\lambda _{i}+\lambda _{j}$ and $-\lambda _{i}-\lambda _{j}$ do not belong
to $\langle \Theta \rangle ^{-}$.
\end{lema}
\begin{profe}
To prove equivalence we shall exhibit a proper $K_{\Theta }$-invariant
subspace $\{0\}\neq V\subset V_{\Theta }\left( -\lambda _{i}+\lambda
_{j}\right) \oplus V_{\Theta }\left( -\lambda _{i}-\lambda _{j}\right) $
different from the irreducible components $V_{\Theta }\left( -\lambda
_{i}+\lambda _{j}\right) $ and $V_{\Theta }\left( -\lambda _{i}-\lambda
_{j}\right) $. This implies equivalence by Proposition \ref{propsumirredcomp
).
The required subspace $V$ will be obtained from the tangent space at the
origin of the orbit of the normal subgroup $K_{l}$. By (\ref{forideaiskbele
) the matrices in the Lie algebra $\mathfrak{k}_{l}$ of $K_{l}$ are
\begin{equation*}
\left(
\begin{array}{ccc}
0 & 0 & 0 \\
0 & A & -A \\
0 & -A &
\end{array
\right) \quad A+A^{T}=0.
\end{equation*
Looking at these matrices we see that after identifying $T_{b_{\Theta }
\mathbb{F}_{\Theta }$ with $\mathfrak{n}_{\Theta }^{-}$ the tangent space
T_{b_{\Theta }}\left( K_{l}\cdot b_{\Theta }\right) $ is identified to the
subspace $T_{l}\subset \mathfrak{n}_{\Theta }^{-}$ spanned by $\mathrm{pr
\left( E_{rs}^{-}-E_{rs}^{+}\right) $, $r>s$, where $E_{rs}^{\pm }$ were
defined in (\ref{forbasesbele}) and $\mathrm{pr}:\mathfrak{n}^{-}\rightarrow
\mathfrak{n}_{\Theta }^{-}$ is the projection w.r.t. the root spaces
decomposition.
The tangent space $T_{b_{\Theta }}\left( K_{l}\cdot b_{\Theta }\right) $ is
invariant by the isotropy representation of $K_{\Theta }$, by Lemma \re
{lemorbitnormal}. Hence $T_{l}$ is invariant by the adjoint action of
K_{\Theta }$.
Now if $-\lambda _{r}+\lambda _{s}$, $-\lambda _{r}-\lambda _{s}\in \Pi
^{-}\setminus \langle \Theta \rangle ^{-}$ then $E_{rs}^{-}-E_{rs}^{+}
\mathrm{pr}\left( E_{rs}^{-}-E_{rs}^{+}\right) $. Hence the following
vectors form a basis of $T_{l}$:
\begin{enumerate}
\item $E_{rs}^{-}$ such that $-\lambda _{r}+\lambda _{s}\in \Pi _{\Theta
}\left( -\lambda _{i}+\lambda _{j}\right) $ and $-\lambda _{r}-\lambda
_{s}\notin \Pi _{\Theta }\left( -\lambda _{i}-\lambda _{j}\right) $.
\item $E_{rs}^{+}$ such that $-\lambda _{r}-\lambda _{s}\in \Pi _{\Theta
}\left( -\lambda _{i}-\lambda _{j}\right) $ and $-\lambda _{r}+\lambda
_{s}\notin \Pi _{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) $.
\item $E_{rs}^{-}-E_{rs}^{+}$ such that $-\lambda _{r}+\lambda _{s}\in \Pi
_{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) $ and $-\lambda
_{r}-\lambda _{s}\in \Pi _{\Theta }\left( -\lambda _{i}-\lambda _{j}\right)
.
\end{enumerate}
The third case is not empty (e.g. $\left( r,s\right) =\left( i,j\right) $
fall in this case), which means that $E_{rs}^{-}-E_{rs}^{+}\in T_{l}$ for
some pair $\left( r,s\right) $. For this pair $T_{l}\cap V_{\Theta }\left(
\lambda _{r}-\lambda _{s}\right) =T_{1}\cap V_{\Theta }\left( -\lambda
_{r}-\lambda _{s}\right) =\{0\}$, which shows that $T_{l}$ is proper and
different from $V_{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) $ and
V_{\Theta }\left( -\lambda _{i}-\lambda _{j}\right) $. By Proposition \re
{propsumirredcomp} it follows that these irreducible subspaces are
K_{\Theta }$-equivalent.
\end{profe}
In conclusion we have:
\begin{teorema}
Let $\mathbb{F}_{\Theta }$ be a flag manifold of $B_{l}=\mathfrak{so}\left(
l+1,l\right) $, $l\geq 5$. Then the $K_{\Theta }$-invariant irreducible
subspaces of $\mathfrak{n}_{\Theta }^{-}$ are in the following classes:
\begin{enumerate}
\item Isolated subspaces:
\begin{enumerate}
\item $V_{\Theta }\left( -\lambda _{i}\right) $ when $\lambda _{l}\notin
\Theta $. These subspaces contain root spaces of short roots only.
\item $V_{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) $, $i<j$, when
-\lambda _{i}-\lambda _{j}\in \langle \Theta \rangle ^{-}$. Any such pair
occur if $-\lambda _{l}\in \Theta $. Otherwise we have $i<j\leq i_{0}$,
where $\{\alpha _{i_{0}+1},\ldots ,\alpha _{l}=\lambda _{l}\}$ is the
connected component of $\Theta $ containing $\lambda _{l}$.
\item The same as (b) interchanging the roles of $-\lambda _{i}+\lambda _{j}$
and $-\lambda _{i}-\lambda _{j}$.
\item The subspace
\begin{equation*}
\left( W_{\Theta }^{i}\right) _{1}=\sum_{j=j\left( i\right) }^{i}\sum_{k\geq
i_{0}+1}\left( W_{\Theta }^{jk}\right) _{1}\qquad \mathrm{and}\qquad \left(
W_{\Theta }^{i}\right) _{2}=\sum_{j=j\left( i\right) }^{i}\sum_{k\geq
i_{0}+1}\left( W_{\Theta }^{jk}\right) _{2}
\end{equation*
defined in Lemma \ref{lemirreducw}. These subspaces decompose $V\left(
-\lambda _{i}\right) $ when $\lambda _{l}\in \Theta $.
\end{enumerate}
\item A continuum of invariant subspaces parametrized by $[(x,y)]\in \mathbb
R}P^{2}$ given by
\begin{equation*}
V_{[(x,y)]}^{ij}=\left\{ xX+yTX:X\in V_{\Theta }\left( -\lambda _{i}+\lambda
_{j}\right) \right\} \qquad i<j.
\end{equation*
The indexes $ij$ are as in (1.b), and here both $-\lambda _{i}+\lambda _{j}$
and $-\lambda _{i}-\lambda _{j}$ are not in $\langle \Theta \rangle ^{-}$.
\end{enumerate}
\end{teorema}
The low dimensional cases $l=2,3,4$ must be treated separetely because of
the difference in the $M$-equivalence classes.
For instance $B_{2}$ has three flag manifolds. The maximal one whose
irreducible components are detected by the $M$-equivalence classes
\{\lambda _{1}-\lambda _{2},\lambda _{1}+\lambda _{2}\}$ and $\{\lambda
_{1},\lambda _{2}\}$. Hence, there are two continuous families of $1
-dimensional irreducible subspaces. In the flag $\mathbb{F}_{\{\lambda
_{1}-\lambda _{2}\}}$ there are two $\mathfrak{z}_{\Theta }$-irreducible
components defined by the sets $\{-\lambda _{2},-\lambda _{1}\}$ and
\{-\lambda _{1}-\lambda _{2}\}$. Both are $K_{\Theta }$-irreducible and
clearly they are not equivalent. On the other hand the flag $\mathbb{F
_{\{\lambda _{2}\}}$ has just one $3$-dimensional $\mathfrak{z}_{\Theta }
-irreducible component which decomposes into a $1$-dimensional plus a $2
-dimensional irreducible subspaces of $K_{\Theta }$ (as happens to the
adjoint representation of $\mathfrak{sl}\left( 2,\mathbb{R}\right) $).
For $B_{3}$ and $B_{4}$ the compact subalgebra $\mathfrak{k}$ ($\mathfrak{so
\left( 3\right) \oplus \mathfrak{so}\left( 4\right) $ and $\mathfrak{so
\left( 4\right) \oplus \mathfrak{so}\left( 5\right) $, respectively) splits
once more because $\mathfrak{so}\left( 4\right) =\mathfrak{so}\left(
3\right) \oplus \mathfrak{so}\left( 3\right) $. By Lemma \ref{lemorbitnormal}
these simple components of $\mathfrak{k}$ can yield new $K_{\Theta }
-invariant subspaces. invariant subspaces. The example with $D_{4}$ below,
which has a similar splitting, illustrates this occurence of new invariant
subspaces. Another aspect that differs $B_{3}$ and $B_{4}$ from the general
case are the $M$-equivalence classes that have more elements. This can
introduce more $K_{\Theta }$-equivalence than the general case. The example
with $C_{4}$ below illustrates this fact.
\subsection{Flags of $C_{l}=\mathfrak{sp}\left( l,\mathbb{R}\right) $}
The symplectic Lie algebra $\mathfrak{sp}\left( l,\mathbb{R}\right) $ is
composed of the real $2l\times 2l$ matrice
\begin{equation*}
\left(
\begin{array}{cc}
A & B \\
C & -A^{T
\end{array
\right) \qquad B-B^{T}=C-C^{T}=0
\end{equation*
written in the basis $\{e_{1},\ldots ,e_{l},f_{1},\ldots ,f_{l}\}$. In this
case $\mathfrak{a}$ is the subalgebra of matrices
\begin{equation*}
\left(
\begin{array}{ll}
\Lambda & 0 \\
0 & -\Lambd
\end{array
\right)
\end{equation*
with $\Lambda =\mathrm{diag}\{a_{1},\ldots ,a_{l}\}$. The set of roots are
i) the long ones $\pm 2\lambda _{i}$, $1\leq i\leq l$ and ii) the short ones
$\pm \left( \lambda _{i}-\lambda _{j}\right) $ and $\pm \left( \lambda
_{i}+\lambda _{j}\right) $, $1\leq i<j\,\leq l$. The set of simple roots is
\Sigma =\{\lambda _{1}-\lambda _{2},\ldots ,\lambda _{l-1}-\lambda
_{l},2\lambda _{l}\}$, which we write also as $\Sigma =\{\alpha _{1},\ldots
,\alpha _{l}\}$, that is, $\alpha _{i}=\lambda _{i}-\lambda _{i+1}$ if $i<l$
and $\alpha _{l}=2\lambda _{l}$.
The Weyl chamber $\mathfrak{a}^{+}\subset \mathfrak{a}$ is defined by the
inequalities
\begin{equation*}
a_{1}>a_{2}>\cdots >a_{l-1}>a_{l}>0,
\end{equation*
and a partial chamber $\mathfrak{a}_{\Theta }\cap \mathrm{cl}\mathfrak{a
^{+} $ is defined by a similar relations where some of the strict
inequalities are changed by equalities (e.g. if $\lambda _{i}-\lambda
_{j}\in \Theta $ then $a_{i}=a_{j}$). In particular a characteristic element
$H_{\Theta }$ for the subset $\Theta =\{\alpha \in \Sigma :\alpha \left(
H_{\Theta }\right) =0\}\subset \Sigma $ is defined by one of these relations.
The subalgebra $\mathfrak{k}$ is composed of the skew-symmetric matrices in
\mathfrak{sp}\left( l,\mathbb{R}\right) $, that is
\begin{equation*}
\left(
\begin{array}{cc}
A & -B \\
B &
\end{array
\right) \qquad A+A^{T}=B-B^{T}=0.
\end{equation*
It is isomorphic to $\mathfrak{u}\left( l\right) =\mathfrak{su}\left(
l\right) \oplus \mathbb{R}$, where the isomorphism associates the above
matrix the complex matrix $A+iB$.
To describe the $\mathfrak{z}_{\Theta }$-irreducible components $V_{\Theta
}^{\sigma }$ defined by the set of roots $\Pi _{\Theta }^{\sigma }$ we
consider first the components $V_{\Theta }\left( -2\lambda _{i}\right) $
containing the long roots.
\begin{lema}
\label{lemcelelong}Let $i=1,\ldots ,l$ be an index such that $\alpha
_{i}\notin \Theta $.
\begin{enumerate}
\item If $i=1$ or $\alpha _{i-1}\notin \Theta $ then $V_{\Theta }\left(
-2\lambda _{i}\right) =\mathfrak{g}_{-2\lambda _{i}}$.
\item Otherwise let $j\left( i\right) <i$ be such that $\{\alpha _{j\left(
i\right) },\ldots ,\alpha _{i-1}\}$ is the connected component of $\Theta $
containing $\alpha _{i-1}$. Then
\begin{equation}
V_{\Theta }\left( -2\lambda _{i}\right) =\sum_{k,r=j\left( i\right) }^{i
\mathfrak{g}_{-\lambda _{k}-\lambda _{r}}. \label{forcelelong}
\end{equation}
\end{enumerate}
If $\alpha _{i}\in \Theta $ then either $-2\lambda _{i}\in \langle \Theta
\rangle $ if $2\lambda _{l}\in \Theta $ and $\alpha _{i}$ and $2\lambda _{l}
$ are in the same connected component of $\Theta $ or $-2\lambda _{i}\in \Pi
_{\Theta }\left( -2\lambda _{j}\right) $ where $j>i$ is the smallest index
such that $\alpha _{j}\notin \Theta $.
\end{lema}
\begin{profe}
Since $\alpha _{i}=\lambda _{i}-\lambda _{i+1}\notin \Theta $ the only way
that $-2\lambda _{i}\pm \alpha $ is a root with $\alpha \in \Theta $ is in
the string $-2\lambda _{i}-\left( \lambda _{i-1}-\lambda _{i}\right)
=-\lambda _{i-1}-\lambda _{i}$ and $-2\lambda _{i}-2\left( \lambda
_{i-1}-\lambda _{i}\right) =-2\lambda _{i-1}$. Hence if $\alpha _{i-1}\notin
\Theta $ (or $i=1$) no such sum occurs and $V_{\Theta }\left( -2\lambda
_{i}\right) =\mathfrak{g}_{-2\lambda _{i}}$.
On the other hand if $\alpha _{i-1}=\lambda _{i-1}-\lambda _{i}\in \Theta $
then the roots $-\lambda _{i-1}-\lambda _{i}=-2\lambda _{i}-\left( \lambda
_{i-1}-\lambda _{i}\right) $ and $-2\lambda _{i-1}=-2\lambda _{i}-2\left(
\lambda _{i-1}-\lambda _{i}\right) $ belong to $\Pi _{\Theta }\left(
-2\lambda _{i}\right) $. Proceeding successively it follows that $-\lambda
_{k}-\lambda _{k+1},-2\lambda _{k}\in \Pi _{\Theta }\left( -2\lambda
_{i}\right) $ if $k=j\left( i\right) ,\ldots ,i-1$. The roots $-\lambda
_{k}-\lambda _{r}$, $j\left( i\right) \leq k<r-1\leq i-1$, also belong to
\Pi _{\Theta }\left( -2\lambda _{i}\right) $, since $-\lambda _{k}-\lambda
_{r}=\left( -\lambda _{k}-\lambda _{k+1}\right) +\left( \lambda
_{k+1}-\lambda _{r}\right) $ and $\lambda _{k+1}-\lambda _{r}\in \langle
\Theta \rangle $. Hence $V_{\Theta }\left( -2\lambda _{i}\right) $ contains
the subspace $\sum_{k,r=j\left( i\right) }^{i}\mathfrak{g}_{-\lambda
_{k}-\lambda _{r}}$. This subspace is $\mathfrak{z}_{\Theta }$-invariant
because a root in a connected component of $\Theta $ different from
\{\alpha _{j\left( i\right) },\ldots ,\alpha _{i-1}\}$ is orthogonal to the
roots $-\lambda _{k}-\lambda _{r}$, $k,r=j\left( i\right) ,\ldots ,i$.
The last statement is a consequence of the expression for $V_{\Theta }\left(
-2\lambda _{i}\right) $ in (2).
\end{profe}
\vspace{12pt
\noinden
\textbf{Remark:} The first case of the above lemma is included in the second
case by taking $j\left( i\right) =i$.
To look at the representation of $K_{\Theta }$ on the subspace $V\left(
-2\lambda _{i}\right) $ of (\ref{forcelelong}) we make use of the following
geometric meaning: \ Let $\mathfrak{sp}\left( j\left( i\right) ,i\right) $
be the subalgebra generated by the root spaces $\mathfrak{g}_{\pm \left(
\lambda _{k}\pm \lambda _{r}\right) }$, $k,r=j\left( i\right) ,\ldots ,i$.
Its elements are symplectic matrices
\begin{equation*}
\left(
\begin{array}{cc}
A & -B \\
B &
\end{array
\right)
\end{equation*
with $A$, $B$ and $C$ having non zero entries only at the positions
k,r=j\left( i\right) ,\ldots ,i$, which shows that it is isomorphic to the
Lie algebra of symplectic matrices in the subspace spanned by $\{e_{j\left(
i\right) },\ldots ,e_{i},f_{j\left( i\right) },\ldots ,f_{i}\}$. Let
\mathrm{Sp}\left( j\left( i\right) ,i\right) =\langle \exp \mathfrak{sp
\left( j\left( i\right) ,i\right) \rangle $ be the connected subgroup with
Lie algebra $\mathfrak{sp}\left( j\left( i\right) ,i\right) $ and put
\mathrm{U}\left( j\left( i\right) ,i\right) =\mathrm{Sp}\left( j\left(
i\right) ,i\right) \cap K$ for its maximal compact subgroup, which is
isomorphic to the unitarian group $\mathrm{U}\left( i-j\left( i\right)
+1\right) $.
The inclusion $\{\alpha _{j\left( i\right) },\ldots ,\alpha _{i-1}\}\subset
\Theta $ shows that the root spaces $\mathfrak{g}_{\lambda _{k}-\lambda
_{r}} $, $k,r=j\left( i\right) ,\ldots ,i$, are contained in the isotropy
subalgebra at the origin $b_{\Theta }\in \mathbb{F}_{\Theta }$. From this it
is \ easily seen that the orbit $\mathrm{Sp}\left( j\left( i\right)
,i\right) \cdot b_{\Theta }=\mathrm{U}\left( j\left( i\right) ,i\right)
\cdot b_{\Theta }$ is a flag manifold of $\mathrm{Sp}\left( j\left( i\right)
,i\right) $ and identifies to the coset $\mathrm{U}\left( j\left( i\right)
,i\right) /\mathrm{SO}\left( j\left( i\right) ,i\right) $ where $\mathrm{SO
\left( j\left( i\right) ,i\right) $ is the subgroup isomorphic to $\mathrm{S
}\left( i-j\left( i\right) +1\right) $, whose Lie algebra is contained in
\sum_{k,r=j\left( i\right) }^{i}\mathfrak{g}_{\lambda _{k}-\lambda _{r}}$.
Further $V_{\Theta }\left( -2\lambda _{i}\right) =\sum_{k,r=j\left( i\right)
}^{i}\mathfrak{g}_{-\lambda _{k}-\lambda _{r}}$ is the tangent space at the
origin $b_{\Theta }\in \mathbb{F}_{\Theta }$ of the orbit $\mathrm{Sp}\left(
j\left( i\right) ,i\right) \cdot b_{\Theta }=\mathrm{U}\left( j\left(
i\right) ,i\right) \cdot b_{\Theta }$.
Now we can get the $K_{\Theta }$-irreducible components of $V_{\Theta
}\left( -2\lambda _{i}\right) $.
\begin{lema}
\label{lemceledecomp}The $\mathfrak{z}_{\Theta }$-irreducible subspace
V_{\Theta }\left( -2\lambda _{i}\right) =\sum_{k,r=j\left( i\right) }^{i
\mathfrak{g}_{-\lambda _{k}-\lambda _{r}}$ has two $K_{\Theta }$-irreducible
components if $j\left( i\right) <i$. They are given as follows:
\begin{enumerate}
\item The one dimensional subspace $V_{\Theta }\left( -2\lambda _{i}\right)
_{\mathrm{cent}}\subset \sum_{k=j\left( i\right) }^{i}\mathfrak{g
_{-2\lambda _{k}}$ spanned by the matrix
\begin{equation*}
\left(
\begin{array}{cc}
0 & 0 \\
I_{\left[ j\left( i\right) ,i\right] } &
\end{array
\right) \in \mathfrak{sp}\left( j\left( i\right) ,i\right) \subset \mathfrak
sp}\left( l,\mathbb{R}\right)
\end{equation*
where $I_{\left( j\left( i\right) ,i\right) }$ is the diagonal matrix with
1 $ in the positions $j\left( i\right) ,\ldots ,i$ and $0$ otherwise.
\item The subspace $V_{\Theta }\left( -2\lambda _{i}\right) _{\mathrm{su
\left( j\left( i\right) ,i\right) }$ given by the matrices
\begin{equation*}
\left(
\begin{array}{cc}
A & 0 \\
B & -A^{T
\end{array
\right) \in \mathfrak{sp}\left( j\left( i\right) ,i\right)
\end{equation*
with $A$ lower triangular and $\mathrm{tr}B=0$.
\end{enumerate}
\end{lema}
\begin{profe}
The compact group $\mathrm{U}\left( j\left( i\right) ,i\right) $ being
isomorphic to $\mathrm{U}\left( i-j\left( i\right) +1\right) $ is the
product of its center $Z_{\left( j\left( i\right) ,i\right) }$ by $\mathrm{S
}\left( j\left( i\right) ,i\right) $. The Lie algebra of the center is given
by matrices
\begin{equation*}
\left(
\begin{array}{cc}
0 & -B \\
B &
\end{array
\right)
\end{equation*
with $B\in \mathbb{R}\cdot I_{\left( \left( j\left( i\right) ,i\right)
\right) }$ (corresponding to the scalar matrices in $\mathfrak{u}\left(
i-j\left( i\right) +1\right) $). The Lie algebra $\mathfrak{su}\left( \left(
j\left( i\right) ,i\right) \right) $ of $\mathrm{SU}\left( j\left( i\right)
,i\right) $ is given by matrices
\begin{equation*}
\left(
\begin{array}{cc}
A & -B \\
B &
\end{array
\right) \in \mathfrak{sp}\left( j\left( i\right) ,i\right)
\end{equation*
with $A$ skew symmetric and $B$ symmetric with $\mathrm{tr}B=0$. It follows
that the tangent spaces to the orbits $Z_{\left( j\left( i\right) ,i\right)
}\cdot b_{\Theta }$ and $\mathrm{SU}\left( j\left( i\right) ,i\right) $ are
V_{\Theta }\left( -2\lambda _{i}\right) _{\mathrm{cent}}$ and $V_{\Theta
}\left( -2\lambda _{i}\right) _{\mathrm{su}\left( j\left( i\right) ,i\right)
}$, respectively.
Hence, by Lemma \ref{lemorbitnormal} these subspaces are invariant by the
isotropy representation of $\mathrm{SO}\left( j\left( i\right) ,i\right)
\mathrm{U}\left( j\left( i\right) ,i\right) \cap K_{\Theta }$. They are
K_{\Theta }$-invariant as well because the connected components of $\Theta $
besides $\{\alpha _{j\left( i\right) },\ldots ,\alpha _{i-1}\}$ are
orthogonal to $\Pi _{\Theta }\left( -2\lambda _{i}\right) $. Hence the
simple components of $K_{\Theta }$ different from $\mathrm{SO}\left( j\left(
i\right) ,i\right) $ act trivially on $V_{\Theta }\left( -2\lambda
_{i}\right) $.
Finally, both subspaces $V_{\Theta }\left( -2\lambda _{i}\right) _{\mathrm
cent}}$ and $V_{\Theta }\left( -2\lambda _{i}\right) _{\mathrm{su}\left(
j\left( i\right) ,i\right) }$ are irreducible. This is obvious for
V_{\Theta }\left( -2\lambda _{i}\right) _{\mathrm{cent}}$ which is one
dimensional. On the other hand the representation of $\mathrm{SO}\left(
j\left( i\right) ,i\right) $ on $V_{\Theta }\left( -2\lambda _{i}\right) $
is equivalent to the isotropy representation of the symmetric space $\mathrm
U}\left( j\left( i\right) ,i\right) /\mathrm{SO}\left( j\left( i\right)
,i\right) $, which is known to be irreducible.
\end{profe}
The $\mathfrak{z}_{\Theta }$-irreducible components described in Lemma \re
{lemcelelong} contain all the root spaces of the long roots not in $\langle
\Theta \rangle $. They include also the short roots $-\lambda _{i}-\lambda
_{j}$ such that $\lambda _{i}-\lambda _{j}\in \langle \Theta \rangle $. The
other $\mathfrak{z}_{\Theta }$-components are given as follows.
\begin{lema}
\label{lemceleshort}Suppose the root $\lambda _{i}-\lambda _{j}$, $i<j$,
does not belong to $\langle \Theta \rangle $.
\begin{enumerate}
\item If $2\lambda _{l}\notin \Theta $ then the set $\Pi _{\Theta }\left(
-\lambda _{i}+\lambda _{j}\right) $ corresponding to the $\mathfrak{z
_{\Theta }$-irreducible component $V_{\Theta }\left( -\lambda _{i}+\lambda
_{j}\right) $ contains only short roots of the type $-\lambda _{r}+\lambda
_{s}$. Similarly $\Pi _{\Theta }\left( -\lambda _{i}-\lambda _{j}\right) $
contains only roots of the type $-\lambda _{r}-\lambda _{s}$.
\item In case $2\lambda _{l}\in \Theta $ let $i_{0}$ be such that
C_{l-i_{0}+1}=\{\alpha _{i_{0}+1},\ldots ,\alpha _{l}=2\lambda _{l}\}$ is
the connected component of $\Theta $ containing $2\lambda _{l}$. Then
V_{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) =V_{\Theta \setminus
\{2\lambda _{l}\}}\left( -\lambda _{i}+\lambda _{j}\right) $ and $V_{\Theta
}\left( -\lambda _{i}-\lambda _{j}\right) =V_{\Theta \setminus \{2\lambda
_{l}\}}\left( -\lambda _{i}-\lambda _{j}\right) $ if $j\leq i_{0}$.
\item On the other hand if $j\geq i_{0}+1$ then
\begin{eqnarray}
V_{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) &=&V_{\Theta }\left(
-\lambda _{i}-\lambda _{j}\right) \label{forceledirectsum} \\
&=&V_{\Theta \setminus C_{l-i_{0}+1}}\left( -\lambda _{i}+\lambda
_{j}\right) \oplus V_{\Theta \setminus C_{l-i_{0}+1}}\left( -\lambda
_{i}-\lambda _{j}\right) . \notag
\end{eqnarray}
\end{enumerate}
Moreover these $\mathfrak{z}_{\Theta }$-irreducible components are
K_{\Theta }$-irreducible.
\end{lema}
\begin{profe}
The first statement is proved as Lemma \ref{lembelecompseml} for $B_{l}$.
The proof of Lemma \ref{lembelecompcoml} also works for the components in
(2) with $j\leq i_{0}$. The direct sum in (\ref{forceledirectsum}) \ comes
from the pair of roots $-\lambda _{i}+\lambda _{j}$, $\left( -\lambda
_{i}+\lambda _{j}\right) -2\lambda _{j}$ and the fact that $2\lambda _{j}\in
\langle \Theta \rangle ^{-}$ if $j\geq i_{0}+1$.
The $K_{\Theta }$-irreducibility of the subspaces in (1) is a consequence of
Lemma \ref{lemirreducible}. In fact no two roots in $\Pi _{\Theta }\left(
-\lambda _{i}+\lambda _{j}\right) $ or in $\Pi _{\Theta }\left( -\lambda
_{i}-\lambda _{j}\right) $ are $M$-equivalent and by Lemma \re
{lemtransimplelong} 2b, $\mathcal{W}_{\Theta }$ acts transitively on these
sets of short roots.
The same argument hold for the subspaces $V_{\Theta }\left( -\lambda
_{i}+\lambda _{j}\right) =V_{\Theta \setminus \{2\lambda _{l}\}}\left(
-\lambda _{i}+\lambda _{j}\right) $ and $V_{\Theta }\left( -\lambda
_{i}-\lambda _{j}\right) =V_{\Theta \setminus \{2\lambda _{l}\}}\left(
-\lambda _{i}-\lambda _{j}\right) $ when $j\leq i_{0}$, which are indeed
K_{\Theta \setminus \{2\lambda _{l}\}}$-irreducible.
Finally, in (2) if $j\geq i_{0}+1$ then $2\lambda _{j}\in \langle \Theta
\rangle $. From the equality $-\lambda _{i}-\lambda _{j}=-\lambda
_{i}+\lambda _{j}-2\lambda _{j}$ we see that $\mathfrak{g}_{-\lambda
_{i}+\lambda _{l}}\oplus \mathfrak{g}_{-\lambda _{i}-\lambda _{l}}$ is an
irreducible subspace for the three dimensional subalgebra $\mathfrak{g
\left( 2\lambda _{j}\right) $, isomorphic to $\mathfrak{sl}\left( 2,\mathbb{
}\right) $, generated by $\mathfrak{g}_{\pm 2\lambda _{j}}$. In this two
dimensional subspace the compact part $\mathfrak{k}\left( 2\lambda
_{j}\right) $ of $\mathfrak{g}\left( 2\lambda _{j}\right) $ is also
irreducible. Now let $\{0\}\neq U\subset V_{\Theta }\left( -\lambda
_{i}+\lambda _{j}\right) $ be a $K_{\Theta }$-invariant subspace. Since the
roots $-\lambda _{i}\pm \lambda _{j}\in \Pi _{\Theta }\left( -\lambda
_{i}+\lambda _{j}\right) =\Pi _{\Theta }\left( -\lambda _{i}-\lambda
_{j}\right) $ are $M$-equivalent, it follows by Lemma \ref{leminterinvMclass}
that $U\cap \left( \mathfrak{g}_{-\lambda _{i}+\lambda _{l}}\oplus \mathfrak
g}_{-\lambda _{i}-\lambda _{l}}\right) \neq \{0\}$. This subspace is
\mathfrak{k}\left( 2\lambda _{j}\right) $-invariant, hence $\mathfrak{g
_{-\lambda _{i}+\lambda _{l}}\oplus \mathfrak{g}_{-\lambda _{i}-\lambda
_{l}}\subset U$. Hence irrducibility of $V_{\Theta }\left( -\lambda
_{i}+\lambda _{j}\right) $ is a consequence of Lemma \ref{lemirreducible}
combined with the fact that $\mathcal{W}_{\Theta \setminus \{2\lambda
_{l}\}}\subset \mathcal{W}_{\Theta }$ acts transitively on the sets $\Pi
_{\Theta \setminus \{2\lambda _{l}\}}\left( -\lambda _{i}+\lambda
_{j}\right) $ and $\Pi _{\Theta \setminus \{2\lambda _{l}\}}\left( -\lambda
_{i}-\lambda _{j}\right) $.
\end{profe}
With this lemma we finish the description of the irreducible $K_{\Theta }
-components. Among them the only $K_{\Theta }$-equivalents are the following:
\begin{enumerate}
\item The one dimensional subspaces $V_{\Theta }\left( -2\lambda _{i}\right)
_{\mathrm{cent}}$ of Lemma \ref{lemceledecomp} (1). The representation of
K_{\Theta }$ on each one of them is trivial.
\item $V_{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) \approx V_{\Theta
}\left( -\lambda _{i}-\lambda _{j}\right) $ when $2\lambda _{l}\notin \Theta
$ as in Lemma \ref{lemceleshort} (1). This equivalence follows by
Proposition \ref{propequivalent}, since there is a bijection between $\Pi
_{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) $ and $\Pi _{\Theta
}\left( -\lambda _{i}-\lambda _{j}\right) $ mapping a root $-\lambda
_{r}+\lambda _{s}\in \Pi _{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) $
to the $M$-equivalent root $-\lambda _{r}-\lambda _{s}\in \Pi _{\Theta
}\left( -\lambda _{i}-\lambda _{j}\right) $.
\item The subspaces $V_{\Theta }\left( -\lambda _{i}+\lambda _{j}\right)
=V_{\Theta \setminus \{2\lambda _{l}\}}\left( -\lambda _{i}+\lambda
_{j}\right) $ and $V_{\Theta }\left( -\lambda _{i}-\lambda _{j}\right)
=V_{\Theta \setminus \{2\lambda _{l}\}}\left( -\lambda _{i}-\lambda
_{j}\right) $ with $j\leq i_{0}$ as in Lemma \ref{lemceleshort} (2).
\end{enumerate}
Any other pair of subspaces are not $K_{\Theta }$-equivalent because the
lack of $M$-equivalence in the corresponding sets of roots (cf. Lemma \re
{lemnotequivalent}).
Summarizing we get the $K_{\Theta }$-invariant subspaces for the flags of
C_{l}$, $l>4$.
\begin{teorema}
Let $\mathbb{F}_{\Theta }$ be a flag manifold of $C_{l}=\mathfrak{sp}\left(
l,\mathbb{R}\right) $, $l\geq 5$. Then the $K_{\Theta }$-invariant
irreducible subspaces of $\mathfrak{n}_{\Theta }^{-}$ are the following:
\begin{enumerate}
\item Continuous families:
\begin{enumerate}
\item One dimensional subspaces spanned by matrices
\begin{equation*}
\left(
\begin{array}{cc}
0 & 0 \\
\Lambda &
\end{array
\right)
\end{equation*
where $\Lambda $ is a diagonal matrix $a_{1}I_{\left[ j\left( i_{1}\right)
,i_{1}\right] }+\cdots +a_{k}I_{\left[ j\left( i_{k}\right) ,i_{k}\right] }$
where $\left[ j\left( i_{1}\right) ,i_{1}\right] $, \ldots ,$\left[ j\left(
i_{1}\right) ,i_{1}\right] $ are the connected components of $\Theta $ not
containing $2\lambda _{l}$, and $I_{\left[ j\left( i_{s}\right) ,i_{s}\right]
}$ is the identity matrix corresponing to these indexes.
\item The subspaces parametrized by $[(x,y)]\in \mathbb{R}P^{2}$ given by
\begin{equation*}
V_{[(x,y)]}^{ij}=\left\{ xX+yTX:X\in V_{\Theta }\left( -\lambda _{i}+\lambda
_{j}\right) \right\} \qquad i<j
\end{equation*
where $T:V_{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) \rightarrow
V_{\Theta }\left( -\lambda _{i}-\lambda _{j}\right) $ is an itertiwining
operator for the $K_{\Theta }$-representations. Here the indexes $ij$ are
arbitrary if $2\lambda _{l}\notin \Theta $. Otherwise $j\leq i_{0}$, where
\{\alpha _{i_{0}+1},\ldots ,\alpha _{l}=2\lambda _{l}\}$ is the component of
$\Theta $ containing $2\lambda _{l}$.
\end{enumerate}
\item Isolated subspaces:
\begin{enumerate}
\item The subspaces $V_{\Theta }\left( -2\lambda _{i}\right) _{\mathrm{su
\left( j\left( i\right) ,i\right) }$ of codimension $1$ contained in
V_{\Theta }\left( -2\lambda _{i}\right) $ as defined in Lemma \re
{lemceledecomp}.
\item The subspace
\begin{eqnarray*}
V_{\Theta }\left( -\lambda _{i}+\lambda _{j}\right) &=&V_{\Theta }\left(
-\lambda _{i}-\lambda _{j}\right) \\
&=&V_{\Theta \setminus C_{l-i_{0}+1}}\left( -\lambda _{i}+\lambda
_{j}\right) \oplus V_{\Theta \setminus C_{l-i_{0}+1}}\left( -\lambda
_{i}-\lambda _{j}\right) ,
\end{eqnarray*
when $2\lambda _{l}\in \Theta $ and $i<i_{0}+1\leq j$ where $\{\alpha
_{i_{0}+1},\ldots ,\alpha _{l}=2\lambda _{l}\}$ is the component of $\Theta $
containing $2\lambda _{l}$.
\end{enumerate}
\end{enumerate}
\end{teorema}
When $l=4$ the $M$-equivalence classes of the short roots increase to
\{\lambda _{1}-\lambda _{2},\lambda _{1}+\lambda _{2},\lambda _{3}-\lambda
_{4},\lambda _{3}+\lambda _{4}\}$, $\{\lambda _{1}-\lambda _{3},\lambda
_{1}+\lambda _{3},\lambda _{2}-\lambda _{4},\lambda _{2}+\lambda _{4}\}$,
\{\lambda _{1}-\lambda _{4},\lambda _{1}+\lambda _{4},\lambda _{2}-\lambda
_{3},\lambda _{2}+\lambda _{3}\}$ while the long roots are kept the same
\{2\lambda _{1},2\lambda _{2},2\lambda _{3},2\lambda _{4}\}$. Since there
are more $M$-equivalent pair of roots we can have more $K_{\Theta }
-equivalent subspaces than in the general case.
For example consider flag $\mathbb{F}_{\{\lambda _{2}-\lambda _{3}\}}$. By
the general result the subspaces $V_{\{\lambda _{2}-\lambda _{3}\}}\left(
-\lambda _{1}+\lambda _{2}\right) $ and $V_{\{\lambda _{2}-\lambda
_{3}\}}\left( -\lambda _{1}-\lambda _{2}\right) $ are $K_{\Theta }
-equivalent. Their corresponding roots are $\Pi _{\{\lambda _{2}-\lambda
_{3}\}}\left( -\lambda _{1}+\lambda _{2}\right) =\{-\lambda _{1}+\lambda
_{2},-\lambda _{1}+\lambda _{3}\}$ and $\Pi _{\{\lambda _{2}-\lambda
_{3}\}}\left( -\lambda _{1}-\lambda _{2}\right) =\{-\lambda _{1}-\lambda
_{2},-\lambda _{1}-\lambda _{3}\}$. For $l=4$ we have $\left( -\lambda
_{1}+\lambda _{2}\right) \sim _{M}\left( -\lambda _{3}+\lambda _{4}\right) $
and $\left( -\lambda _{1}+\lambda _{3}\right) \sim _{M}\left( -\lambda
_{2}+\lambda _{4}\right) $. Since the set of root for $V_{\{\lambda
_{2}-\lambda _{3}\}}\left( -\lambda _{3}+\lambda _{4}\right) $ is $\Pi
_{\{\lambda _{2}-\lambda _{3}\}}\left( -\lambda _{3}+\lambda _{4}\right)
=\{-\lambda _{3}+\lambda _{4},-\lambda _{2}+\lambda _{4}\}$ we conclude that
$V_{\{\lambda _{2}-\lambda _{3}\}}\left( -\lambda _{1}+\lambda _{2}\right) $
is also $K_{\Theta }$-equivalent to $V_{\{\lambda _{2}-\lambda _{3}\}}\left(
-\lambda _{3}+\lambda _{4}\right) $. The same way $V_{\{\lambda _{2}-\lambda
_{3}\}}\left( -\lambda _{1}-\lambda _{2}\right) $ and $V_{\{\lambda
_{2}-\lambda _{3}\}}\left( -\lambda _{3}-\lambda _{4}\right) $ are
K_{\Theta }$-equivalent. Thus we get new continuous families of invariant
subspaces that are not present in the general case.
\subsection{Flags of $D_{l}=\mathfrak{so}\left( l,l\right) $}
The Dynkin diagram $D_{l}$ has no multiple edges. Hence on any flag manifold
$\mathbb{F}_{\Theta }$ we have, by Lemma \ref{lemtransimplelong}, that
\mathcal{W}_{\Theta }$ acst transitively on each set of roots $\Pi _{\Theta
}^{\sigma }$ corresponding to an irreducible representation of $\mathfrak{z
_{\Theta }$ on $V_{\Theta }^{\sigma }\subset \mathfrak{n}_{\Theta }^{-}$.
This transitivity is one of the conditions of Lemma \ref{lemirreducible},
ensuring that the subspaces $V_{\Theta }^{\sigma }$ are $K_{\Theta }
-irreducible. To look at the condition involving the $M$-equivalence classes
we work, as in Section \ref{secmequiv}, with the standard realization of
D_{l}=\mathfrak{so}\left( l,l\right) $.
In this realization $\mathfrak{a}$ is the subalgebra of matrices
\begin{equation*}
\left(
\begin{array}{cc}
\Lambda & 0 \\
0 & -\Lambd
\end{array
\right)
\end{equation*
with $\Lambda =\mathrm{diag}\{a_{1},\ldots ,a_{l}\}$ and the set of simple
roots is $\Sigma =\{\lambda _{1}-\lambda _{2},\ldots ,\lambda _{l-1}-\lambda
_{l},\lambda _{l-1}+\lambda _{l}\}$.
The Weyl chamber $\mathfrak{a}^{+}\subset \mathfrak{a}$ is defined by the
inequalities
\begin{equation}
a_{1}>a_{2}>\cdots >a_{l-1}>a_{l}>-a_{l-1}. \label{forinechamber}
\end{equation
A partial chamber $\mathfrak{a}_{\Theta }\cap \mathrm{cl}\mathfrak{a}^{+}$
is defined by a similar relation where some of the inequalities are changed
by equalities. In particular a characteristic element $H_{\Theta }$ for the
subset $\Theta =\{\alpha \in \Sigma :\alpha \left( H_{\Theta }\right)
=0\}\subset \Sigma $ is defined by one of these relations.
The following statement is specific for $D_{l}$ and will be used soon to
check that $M$-equivalent root spaces are not contained in an irreducible
component.
\begin{lema}
Given a subset $\Theta \subset \Sigma $ there exists characteristic element
\begin{equation*}
H_{\Theta }=\left(
\begin{array}{cc}
\Lambda _{\Theta } & 0 \\
0 & -\Lambda _{\Theta
\end{array
\right) \in \mathfrak{a}_{\Theta }\cap \mathrm{cl}\mathfrak{a}^{+}
\end{equation*
with $\Lambda _{\Theta }=\mathrm{diag}\{a_{1},\ldots ,a_{l}\}$ such that
a_{i}\neq 0$, $i=1,\ldots ,l$.
\end{lema}
\begin{profe}
By the last two inequalities in (\ref{forinechamber}) we have $a_{l-1}\geq
-a_{l-1}$, that is, $a_{l-1}\geq 0$. Also, $a_{l-1}=0$ if and only if
a_{l-1}=a_{l}=-a_{l-1}$, that is, $a_{l-1}-a_{l}=a_{l-1}+a_{l}=0$, which
means that both roots $\lambda _{l-1}-\lambda _{l}$ and $\lambda
_{l-1}+\lambda _{l}$ belong to $\Theta $. This being so we consider the
possibilities:
\begin{enumerate}
\item $\{\lambda _{l-1}-\lambda _{l},\lambda _{l-1}+\lambda _{l}\}\subset
\Theta $. Let $i<l-1$ be the maximum such that $\lambda _{i}-\lambda
_{i+1}\notin \Theta $ (it is tacitly assumed that $\Theta \neq \Sigma $).
Then the conditions to define a characteristic element for $\Theta $ have
the for
\begin{equation*}
a_{1}\geq \cdots \geq a_{i}>a_{i+1}=a_{i+2}=\cdots =a_{l}.
\end{equation*
Thus we can choose a characteristic element having $a_{i+1}=a_{i+2}=\cdots
=a_{l}>0$, so that all the entries of $\Lambda _{\Theta }$ will be $>0$.
\item One of the roots $\lambda _{l-1}-\lambda _{l}$ or $\lambda
_{l-1}+\lambda _{l}$ does not belong to $\Theta $. In this case
a_{l-1}>0>-a_{l-1}$ for any $H_{\Theta }$ so that $a_{i}>0$ for any $i\leq
l-1$. Also the relations defining $\mathfrak{a}_{\Theta }\cap \mathrm{cl
\mathfrak{a}^{+}$ end with
\begin{equation*}
a_{l-1}>a_{l}>-a_{l-1}\quad \mathrm{or}\quad a_{l-1}\geq a_{l}>-a_{l-1}\quad
\mathrm{or}\quad a_{l-1}>a_{l}\geq -a_{l-1}.
\end{equation*
In each case we can choose $a_{l}\neq 0$ without violating the conditions.
\end{enumerate}
\end{profe}
From now on we distinguish the cases where $l>4$ and $l=4$.
If $l>4$ then $M$-equivalence classes in the positive roots are $\{\lambda
_{i}-\lambda _{j},\lambda _{i}+\lambda _{j}\}$, $i<j$, and $\{\lambda
_{i}-\lambda _{j},-\lambda _{i}-\lambda _{j}\}$, $i>j$, in the negative
roots. By the previous lemma we get easily that the corresponding $M
-equivalent root spaces are not contained in a single irreducible component.
\begin{lema}
\label{lemdelecompequiv}Let $V_{\Theta }^{\sigma }$ be an irreducible
component containing the root spaces $\mathfrak{g}_{\alpha }$ and $\mathfrak
g}_{\beta }$, $\alpha \neq \beta $. If $l>4$ then $\alpha $ and $\beta $ are
not $M$-equivalent.
\end{lema}
\begin{profe}
Take a characteristic element $H_{\Theta }$ with $a_{i}\neq 0$ as in the
previous lemma. Then $\left( \lambda _{i}-\lambda _{j}\right) \left(
H_{\Theta }\right) \neq -\left( \lambda _{i}+\lambda _{j}\right) \left(
H_{\Theta }\right) $ for otherwise $a_{i}-a_{j}=-a_{i}-a_{j}$, that is,
a_{i}=0$. The result follows, since $V_{\Theta }^{\sigma }$ is contained in
an eigenspace of $\mathrm{ad}\left( H_{\Theta }\right) $.
\end{profe}
Combining this lemma with Lemma \ref{lemtransimplelong} (about the
transitivity of $\mathcal{W}_{\Theta }$) we get at once $K_{\Theta }
-irreducibility of $V_{\Theta }^{\sigma }$, by Lemma \ref{lemirreducible}.
\begin{proposicao}
In any flag manifold $\mathbb{F}_{\Theta }$ of $D_{l}$, $l>4$, the
\mathfrak{z}_{\Theta }$-irreducible components $V_{\Theta }^{\sigma }$ are
also $K_{\Theta }$-irreducible.
\end{proposicao}
To get the full picture of the invariant subspaces we must find the pairs of
$\mathfrak{z}_{\Theta }$-irreducible components that are mutually $K_{\Theta
}$-equivalent. Our method to check $K_{\Theta }$-equivalence is via the
orbits on $\mathbb{F}_{\Theta }$ of the simple components of the maximal
compact subgroup $K$ of $G$.
To this purpose we need some further notation concerning the standard
realization of $D_{l}$. The Lie algebra $\mathfrak{so}\left( l,l\right) $ is
the algebra of $2l\times 2l$ matrices of the form
\begin{equation*}
\left(
\begin{array}{cc}
A & B \\
C & -A^{T
\end{array
\right) \quad B+B^{T}=C+C^{T}=0.
\end{equation*
We have that $\mathfrak{k}$ is the subalgebra of skew-symmetric matrices in
\mathfrak{so}\left( l,l\right) $, that is
\begin{equation*}
\left(
\begin{array}{cc}
A & B \\
B &
\end{array
\right) \quad A+A^{T}=B+B^{T}=0.
\end{equation*
$\mathfrak{k}$ is the direct sum of two copies of $\mathfrak{so}\left(
l\right) $. In fact, via the decomposition
\begin{equation*}
\left(
\begin{array}{cc}
A & B \\
B &
\end{array
\right) =\left(
\begin{array}{cc}
\left( A+B\right) /2 & \left( A+B\right) /2 \\
\left( A+B\right) /2 & \left( A+B\right) /
\end{array
\right) +\left(
\begin{array}{cc}
\left( A-B\right) /2 & -\left( A-B\right) /2 \\
-\left( A-B\right) /2 & \left( A-B\right) /
\end{array
\right)
\end{equation*
we get $\mathfrak{k}=\mathfrak{so}\left( l\right) _{1}\oplus \mathfrak{so
\left( l\right) _{2}$ wit
\begin{equation*}
\mathfrak{so}\left( l\right) _{1}:\left(
\begin{array}{ll}
A & A \\
A &
\end{array
\right) \qquad \mathfrak{so}\left( l\right) _{2}:\left(
\begin{array}{cc}
A & -A \\
-A &
\end{array
\right)
\end{equation*
where in both cases $A$ is skew-symmetric. We write $\mathrm{SO}\left(
l\right) _{i}=\langle \exp \mathfrak{so}\left( l\right) _{i}\rangle $,
i=1,2 $. \
As to the root spaces we write
\begin{equation}
E_{ij}^{-}=\left(
\begin{array}{cc}
E_{ij} & 0 \\
0 & -E_{ij}^{t
\end{array
\right) \qquad \mathrm{and}\qquad E_{ij}^{+}=\left(
\begin{array}{cc}
0 & 0 \\
E_{ij}-E_{ij}^{t} &
\end{array
\right) \label{forrootdele}
\end{equation
where $E_{ij}$ is a basic $l\times l$ matrix. Then $E_{ij}^{-}$ spans the
root space $\mathfrak{g}_{\lambda _{i}-\lambda _{j}}$ and $E_{ij}^{+}$ spans
$\mathfrak{g}_{-\lambda _{i}-\lambda _{j}}$.
We can return now to the question of $K_{\Theta }$-equivalence of the
\mathfrak{z}_{\Theta }$-components $V_{\Theta }^{\sigma }$.
For a root $\alpha \in \Pi ^{-}\setminus \langle \Theta \rangle ^{-}$ write
V_{\Theta }\left( \alpha \right) $ for the irreducible component containing
\mathfrak{g}_{\alpha }$ (cf. Proposition \ref{propcomproot}). By Lemma \re
{lemdelecompequiv} we have $V_{\Theta }\left( \alpha \right) \neq V_{\Theta
}\left( \beta \right) $ if $\alpha \sim _{M}\beta $ and $\alpha \neq \beta
. Moreover, by Lemma \ref{lemnotequivalent} a component $V_{\Theta }^{\sigma
}$ is not $K_{\Theta }$-equivalent to $V_{\Theta }\left( \alpha \right) $
unless there exists $\beta \sim _{M}\alpha $ such that $V_{\Theta }^{\sigma
}=V_{\Theta }\left( \beta \right) $.
Now by Section \ref{secmequiv} we have that if $l\neq 4$ then the $M
-equivalent classes of $D_{l}$ (on the negative roots) has exactly two
elements. If $\{\alpha ,\beta \}$ is a $M$-equivalence class with say
\alpha \in \Pi ^{-}\setminus \langle \Theta \rangle ^{-}$ and $\beta \in
\langle \Theta \rangle ^{-}$ then $V_{\Theta }\left( \alpha \right) $ is not
$K_{\Theta }$-equivalent to any other irreducible component $V_{\Theta
}^{\sigma }$. On the other hand if both $\alpha ,\beta \in \Pi ^{-}\setminus
\langle \Theta \rangle ^{-}$ we have $K_{\Theta }$-equivalence between
V_{\Theta }\left( \alpha \right) $ and $V_{\Theta }\left( \beta \right) $.
\begin{lema}
In $D_{l}$, $l>4$, let $\{\alpha ,\beta \}$ be a $M$-equivalence class
contained in $\Pi ^{-}\setminus \langle \Theta \rangle ^{-}$. Then the
K_{\Theta }$ representations on $V_{\Theta }\left( \alpha \right) $ and
V_{\Theta }\left( \beta \right) $ are equivalent.
\end{lema}
\begin{profe}
To prove equivalence we shall exhibit a $K_{\Theta }$-invariant subspace
\{0\}\neq V\subset V_{\Theta }\left( \alpha \right) \oplus V_{\Theta }\left(
\beta \right) $ which is different from the irreducible components
V_{\Theta }\left( \alpha \right) $ and $V_{\Theta }\left( \beta \right) $.
This will imply that the components are indeed $K_{\Theta }$-equivalent (see
Proposition \ref{propsumirredcomp}).
The required subspace $V$ will be obtained from the tangent space at the
origin of the orbit of one of the normal subgroups $\mathrm{SO}\left(
l\right) _{j}$, $j=1,2$.
Take for instance $\mathrm{SO}\left( l\right) _{1}$ whose Lie algebra
\mathfrak{so}\left( l\right) _{1}$ constitutes of the matrices
\begin{equation*}
\left(
\begin{array}{cc}
A & A \\
A &
\end{array
\right) \quad A+A^{T}=0.
\end{equation*
Looking at these matrices we see that after identifying $T_{b_{\Theta }
\mathbb{F}_{\Theta }$ with $\mathfrak{n}_{\Theta }^{-}$ the tangent space
T_{b_{\Theta }}\left( \mathrm{SO}\left( l\right) _{1}\cdot b_{\Theta
}\right) $ to the orbit $\mathrm{SO}\left( l\right) _{1}\cdot b_{\Theta }$
is identified to the subspace $W_{1}\subset \mathfrak{n}_{\Theta }^{-}$
spanned by $\mathrm{pr}\left( E_{rs}^{-}+E_{rs}^{+}\right) $, $r>s$, where
E_{rs}^{\pm }$ were defined in (\ref{forrootdele}) and $\mathrm{pr}
\mathfrak{n}^{-}\rightarrow \mathfrak{n}_{\Theta }^{-}$ is the projection
w.r.t. the root spaces decomposition.
The tangent space $T_{b_{\Theta }}\left( \mathrm{SO}\left( l\right)
_{1}\cdot b_{\Theta }\right) $ is invariant by the isotropy representation
of $K_{\Theta }$, by Lemma \ref{lemorbitnormal}. Hence $W_{1}$ is invariant
by the adjoint action of $K_{\Theta }$.
Now a $M$-equivalence class is given by $\{\lambda _{i}-\lambda
_{j},-\lambda _{i}-\lambda _{j}\}$, $i>j$, whose root spaces are spanned by
E_{ij}^{-}$ and $E_{ij}^{+}$, so that $E_{ij}^{-}\in V_{\Theta }\left(
\lambda _{i}-\lambda _{j}\right) $ and $E_{ij}^{+}\in V_{\Theta }\left(
-\lambda _{i}-\lambda _{j}\right) $. If both roots are in $\Pi ^{-}\setminus
\langle \Theta \rangle ^{-}$ we have
\begin{equation*}
E_{rs}^{-}+E_{rs}^{+}=\mathrm{pr}\left( E_{rs}^{-}+E_{rs}^{+}\right) \in
W_{1}\cap \left( V_{\Theta }\left( \lambda _{i}-\lambda _{j}\right) \oplus
V_{\Theta }\left( -\lambda _{i}-\lambda _{j}\right) \right) .
\end{equation*
Hence $W_{1}\cap \left( V_{\Theta }\left( \lambda _{i}-\lambda _{j}\right)
\oplus V_{\Theta }\left( -\lambda _{i}-\lambda _{j}\right) \right) \neq
\{0\} $. This is a $K_{\Theta }$-invariant subspace different from
V_{\Theta }\left( \lambda _{i}-\lambda _{j}\right) $ and $V_{\Theta }\left(
-\lambda _{i}-\lambda _{j}\right) $. It follows that the representation of
K_{\Theta }$ on the irreducible subspaces $V_{\Theta }\left( \lambda
_{i}-\lambda _{j}\right) $ and $V_{\Theta }\left( -\lambda _{i}-\lambda
_{j}\right) $ are equivalent by Proposition \ref{propsumirredcomp}.
\end{profe}
Summarizing we get the $K_{\Theta }$-invariant subspaces for the flags of
D_{l}$, $l>4$.
\begin{teorema}
In a flag $\mathbb{F}_{\Theta }$ of $D_{l}$, $l>4$, there are the following
two classes of $K_{\Theta }$-invariant subspaces in $\mathfrak{n}_{\Theta
}^{-}$.
\begin{enumerate}
\item The $\mathfrak{z}_{\Theta }$-irreducible component $V_{\Theta }\left(
\alpha \right) $, containing the root space $\mathfrak{g}_{\alpha }$ in case
$\alpha \in \Pi ^{-}\setminus \langle \Theta \rangle ^{-}$ is not $M
-equivalent to $\beta \in \Pi ^{-}\setminus \langle \Theta \rangle ^{-}$.
(These are isolated invariant subspaces.)
\item Let $\{\alpha ,\beta \}$ be a $M$-equivalence class contained in $\Pi
^{-}\setminus \langle \Theta \rangle ^{-}$. Then there is a continuum of
invariant subspaces parametrized by $[(x,y)]\in \mathbb{R}P^{1}$ given by
\begin{equation*}
V_{[(x,y)]}=\left\{ xX+yTX:X\in V_{\Theta }\left( \alpha \right) \right\}
\end{equation*
where $T:V_{\Theta }\left( \alpha \right) \rightarrow V_{\Theta }\left(
\beta \right) $ is an isomorphism intertwining the $K_{\Theta }
-representations.
\end{enumerate}
\end{teorema}
The case $l=4$ differs from the general one in two aspects, \ namely each $M
-equivalence class has now $4$ elements and the compact subalgebra
\mathfrak{k}=\mathfrak{so}\left( 4\right) \oplus \mathfrak{so}\left(
4\right) $ decomposes further into four copies of $\mathfrak{so}\left(
3\right) $. These simple components of $\mathfrak{k}$ yield new invariant
subspaces.
To see what can happen let us consider the example with $\Theta =\{\lambda
_{1}-\lambda _{2},\lambda _{2}-\lambda _{3},\lambda _{3}-\lambda _{4}\}$.
Then $\mathfrak{n}_{\Theta }^{-}$ is formed by matrices
\begin{equation*}
Y=\left(
\begin{array}{ll}
0 & 0 \\
X &
\end{array
\right)
\end{equation*
with $X$ a $4\times 4$ skew-symmetric matrix. Here $\mathfrak{k}$ is the Lie
algebra of matrices
\begin{equation*}
\left(
\begin{array}{cc}
\alpha & \beta \\
\beta & \alph
\end{array
\right) \qquad \alpha +\alpha ^{T}=\beta +\beta ^{T}=0
\end{equation*
which is isomorphic to $\mathfrak{so}\left( 4\right) \oplus \mathfrak{so
\left( 4\right) $ by
\begin{equation*}
\left(
\begin{array}{cc}
\alpha & \beta \\
\beta & \alph
\end{array
\right) \longmapsto \left(
\begin{array}{cc}
\alpha +\beta & 0 \\
0 & \alpha -\bet
\end{array
\right) .
\end{equation*
Now $\mathfrak{so}\left( 4\right) $ is the direct sum of two copies of
\mathfrak{so}\left( 3\right) $ as in (\ref{fordecomso4so3so3}) (see the case
$A_{3}$, above). Thus we can see that if we take $X$ in each one of the sets
of matrices in (\ref{fordecomso4so3so3}) we get subespaces $V_{1}$ and
V_{2} $ of $\mathfrak{n}_{\Theta }^{-}$ that are the tangent spaces to the
orbits of the simple components of $K$. Hence $\mathfrak{n}_{\Theta
}^{-}=V_{1}\oplus V_{2}$ is a decomposition into two $3$-dimensional
K_{\Theta }$-invariant subspaces. These two representations are equivalent,
since they are just the adjoint representation of $\mathfrak{so}\left(
3\right) $ on each component.
\subsection{Flags of $E_{6}$, $E_{7}$ and $E_{8}$}
For a flag manifold $\mathbb{F}_{\Theta }$ of one of these exceptional Lie
algebras the $K_{\Theta }$-invariant irreducible subspaces finite and
coincide with the invariant irreducible components $V_{\Theta }^{\sigma
}\subset \mathfrak{n}_{\Theta }^{-}$ for the representation of $\mathfrak{z
_{\Theta }$.
This is because the Dynkin diagrams are simply laced. Hence, by Lemma \re
{lemtransimplelong} it follows that $\mathcal{W}_{\Theta }$ acts
transitively on each set of roots $\Pi _{\Theta }^{\sigma }$ corresponding
to $V_{\Theta }^{\sigma }$. Also, as checked in Section \ref{secmequiv} the
classes of $M$-equivalence for these Lie algebras are singletons. Hence by
Lemma \ref{lemirreducible} we have $K_{\Theta }$-irreducibility of each
V_{\Theta }^{\sigma }$. Furthermore the representations of $K_{\Theta }$ on
different subspaces $V_{\Theta }^{\sigma _{1}}$ and $V_{\Theta }^{\sigma
_{2}}$ are not $M$-equivalent, as follows by combining Lemma \re
{lemnotequivalent} and the fact that the $M$-equivalence classes are
singletons.
\subsection{Flags of $G_{2}$}
Let $\alpha _{1}$ and $\alpha _{2}$ be the simple roots of $G_{2}$ with
\alpha _{1}$ the long one. There are three flag manifolds, $\mathbb{F
_{\emptyset }$, $\mathbb{F}_{\{\alpha _{1}\}}$ and $\mathbb{F}_{\{\alpha
_{2}\}}$. The irreducible components on them are easily obtained by direct
inspection of the positive roots. Recall that the $M$-equivalence classes on
the positive roots are $\{\alpha _{1},\alpha _{1}+2\alpha _{2}\}$, $\{\alpha
_{1}+\alpha _{2},\alpha _{1}+3\alpha _{2}\}$ and $\{\alpha _{2},2\alpha
_{1}+3\alpha _{2}\}$. They are listed below:
\begin{enumerate}
\item In $\mathbb{F}=\mathbb{F}_{\emptyset }$ there are three families of
\mathfrak{z}_{\Theta }$ and $K_{\Theta }$-irreducible subspaces,
parametrized by $\mathbb{R}P^{1}$, corresponding to the three $M
-equivalence classes on the negative roots.
\item For $\mathbb{F}_{\{\alpha _{1}\}}$ there are three $\mathfrak{z
_{\Theta }$-irreducible components corresponding to the sets of roots
\{\alpha _{2},\alpha _{1}+\alpha _{2}\}$, $\{\alpha _{1}+2\alpha _{2}\}$ and
$\{\alpha _{1}+3\alpha _{2},2\alpha _{1}+3\alpha _{2}\}$. They are
K_{\Theta }$-irreducible because the $2$-dimensional irreducible
representation of $\mathfrak{sl}\left( 2,\mathbb{R}\right) $ is $\mathfrak{s
}\left( 2\right) $-irreducible. By checking the $M$-equivalence classes we
see that the $2$-dimensional subspaces are equivalent. Hence, we have the
irreducible subspace $\mathfrak{g}_{-\alpha _{1}-2\alpha _{2}}$ and a family
of $2$-dimensional irreducible subspaces parametrized by $\mathbb{R}P^{1}$,
contained in $\mathfrak{g}_{-\alpha _{2}}\oplus \mathfrak{g}_{-\alpha
_{1}-\alpha _{2}}\mathfrak{g}_{-\alpha _{1}-3\alpha _{2}}\oplus \mathfrak{g
_{-2\alpha _{1}-3\alpha _{2}}$.
\item For $\mathbb{F}_{\{\alpha _{2}\}}$ the $\mathfrak{z}_{\Theta }
-irreducible components correspond to the sets of roots $\{\alpha
_{1},\alpha _{1}+\alpha _{2},\alpha _{1}+2\alpha _{2},\alpha _{1}+3\alpha
_{2}\}$ and $\{2\alpha _{1}+3\alpha _{2}\}$. The $4$-dimensional irreducible
representation of $\mathfrak{z}_{\Theta }\approx \mathfrak{sl}\left( 2
\mathbb{R}\right) $ decomposes into two $K_{\Theta }\approx \mathrm{SO
\left( 2\right) $-invariant irreducible $2$-dimensional inequivalent
representations. Hence there are three $K_{\Theta }$-invariant irreducible
subspaces. \
\end{enumerate}
\subsection{Flags of $F_{4}$}
Recall that the $M$-equivalence classes on the positive roots of $F_{4}$ are
given by
\begin{itemize}
\item $12$ singletons $\{\alpha \}$ with $\alpha $ running through the set
of short roots.
\item $3$ sets of long roots $\{2\alpha _{1}+3\alpha _{2}+4\alpha
_{3}+2\alpha _{4},\alpha _{2},\alpha _{2}+2\alpha _{3},\alpha _{2}+2\alpha
_{3}+2\alpha _{4}\}$, $\{\alpha _{1}+3\alpha _{2}+4\alpha _{3}+2\alpha
_{4},\alpha _{1}+\alpha _{2},\alpha _{1}+\alpha _{2}+2\alpha _{3},\alpha
_{1}+\alpha _{2}+2\alpha _{3}+2\alpha _{4}\}$ and
$\{\alpha _{1}+2\alpha _{2}+4\alpha _{3}+2\alpha _{4},\alpha _{1},\alpha
_{1}+2\alpha _{2}+2\alpha _{3},\alpha _{1}+2\alpha _{2}+2\alpha _{3}+2\alpha
_{4}\}$.
\end{itemize}
Hence in the maximal flag manifold the invariant subspaces are $\mathfrak{g
_{-\alpha }$, $\alpha $ short root, and the one dimensional subspaces
contained in $\mathfrak{g}_{-\alpha }\oplus \mathfrak{g}_{-\beta }\oplus
\mathfrak{g}_{-\gamma }\oplus \mathfrak{g}_{-\delta }$ with $\{\alpha ,\beta
,\gamma ,\delta \}$ a $M$-equivalence class of long roots.
We will not make an extensive analysis of the other $14$ flag manifolds but
look only at the specific flag manifold $\mathbb{F}_{\Theta }$ where $\Theta
=\{\alpha _{2},\alpha _{3},\alpha _{4}\}$ which $C_{3}$ subdiagram. In this
case the $\mathfrak{z}_{\Theta }$-representation decomposes into two
irreducible components $V_{\Theta }^{1}$ and $V_{\Theta }^{2}$ with $\dim
V_{\Theta }^{1}=14$ and $\dim V_{\Theta }^{2}=1$. The sets of roots $\Pi
_{\Theta }^{i}$ corresponding to $V_{\Theta }^{i}$ are those having
coefficient $-i$ ($i=1,2$) with respect to $\alpha _{1}$. Namely, $-\Pi
_{\Theta }^{2}=\{2\alpha _{1}+3\alpha _{2}+4\alpha _{3}+2\alpha _{4}\}$ (the
highest root) and $-\Pi _{\Theta }^{1}$ contains the remaining positive
roots outside $\langle \Theta \rangle $.
Clearly the $K_{\Theta }$-representation on $V_{\Theta }^{2}$ is
irreducible. The representation on $V_{\Theta }^{1}$ is irreducible as well.
In fact, $-\alpha _{1}$ is the highest weight for the $\mathfrak{z}_{\Theta
} $-representation. Since $\langle -\alpha _{1},\alpha _{2}^{\vee }\rangle
=1 $ and $\langle -\alpha _{1},\alpha _{3}^{\vee }\rangle =\langle -\alpha
_{1},\alpha _{4}^{\vee }\rangle =0$, it follows that this is a fundamental
weight of $\mathfrak{sp}\left( 3,\mathbb{R}\right) $, namely the weight
\lambda _{1}+\lambda _{2}+\lambda _{3}$, where $\lambda _{i}$ has the same
meaning as in Section \ref{secirreduc}. Hence $V_{\Theta }^{1}$ is the space
of a basic representation of $\mathfrak{sp}\left( 3,\mathbb{R}\right) $. It
is known that this basic representation is irreducible by the compact
subalgebra $\mathfrak{u}\left( 3\right) $. Hence $V_{\Theta }^{1}$ is
K_{\Theta }$-irreducible.
|
\section{Introduction}
\label{sec:introduction}
Let $\mathscr{H}_1$, $\mathscr{H}_2$ be separable Hilbert spaces. We
consider the following space:
\begin{equation}
\label{eq:2}
\mathscr{H}=\mathscr{H}_1\otimes \Gamma_s(\mathscr{H}_2)\; ;
\end{equation}
where $\Gamma_s(\mathscr{K})$ is the symmetric Fock space based on
$\mathscr{K}$ \citep[see][for mathematical presentations of Fock
spaces and second quantization]{MR0493420,MR3060648,MR0044378}. The
symmetric structure of the Fock space does not play a role in the
argument: in principle it is possible to formulate the same criterion
for anti-symmetric Fock spaces
$\mathscr{H}_1\otimes\Gamma_a(\mathscr{H}_2)$. We focus on symmetric
spaces, the corresponding antisymmetric results should be deduced
without effort.
We are interested in proving a criterion of essential self-adjointness
for densely defined operators of the form:
\begin{equation}
\label{eq:1}
H= H_{01}\otimes 1 + 1\otimes H_{02}+H_I\; ;
\end{equation}
with suitable assumptions on $H_{01}$, $H_{02}$ and $H_I$. Operators
based on these spaces and with such structure are crucial in physics,
to describe the quantum dynamics of interacting particles and fields.
Self-adjointness of operators in Fock spaces has been widely studied,
in particular in the context of Constructive Quantum Field Theory
\citep[e.g.][]{MR947959,MR0215585,MR0210416,MR0270671,MR0266533,Se}
and Quantum ElectroDynamics
\citep[e.g.][]{nelson:1190,MR1891842,MR2097788,BFS,BFS2,BFSS,MR797278,MR1809881,MR2436496}.
A variety of advanced tools has been utilized, for even ``simple''
systems present technical difficulties to overcome: many questions
still remain unsolved.
In some favourable situations, however, it is possible to take
advantage of the peculiar structure of the Fock space and prove
essential self-adjointness with almost no effort. The idea first
appeared in a paper by \citet{MR0266533}; and the author utilized it
in \citep{falconi-phd,Ammari:2014aa} for the Nelson model with cut
off: essential self-adjointness can be proved with less assumptions
than using the Kato-Rellich Theorem (and that becomes particularly
significative in dimension two), see
Section~\ref{sec:nels-type-hamilt}. Another remarkable application is
the Pauli-Fierz Hamiltonian describing particles coupled with a
radiation field. For general coupling constants, essential
self-adjointness has been first proved in a probabilistic setting,
using stochastic integration \citep{MR1773809,MR1891842}. In this
paper we prove the same result directly in
Section~\ref{sec:pauli-fierz-hamilt}, applying the criterion
formulated in Assumptions~\ref{ass:1}, \ref{ass:2} and
Theorem~\ref{thm:1}.
In the literature, self-adjointness of operators in Fock spaces has
been studied using various tools of functional analysis: the
Kato-Rellich and functional integration arguments mentioned above are
two examples, as well as the Nelson commutator theorem
\citep[][]{MR1814991}. For each particular system, a strategy is
utilized ad hoc: the more complicated is the correlation between
$\mathscr{H}_1$ and $\Gamma_s(\mathscr{H}_2)$, the more difficult is
the strategy. We realized that, if we take suitable advantage of the
fibered structure of the Fock space, the type of interaction between
the spaces is not so relevant. This was a strong motivation to study
the problem from a general perspective. Due to the variety of possible
applications, an effort has been made to formulate the necessary
assumptions in a general form. Roughly speaking, the essential
requirement is that the part of $H_I$ that does not commute with the
number operator of $\Gamma_s(\mathscr{H}_2)$ is at most quadratic with
respect to the creation and annihilation operators. As anticipated,
the space $\mathscr{H}_1$ does not play a particular role, as long as
$H_I$ behaves sufficiently well with respect to $H_{01}$.
\subsection{Paper organization.}
\label{sec:paper-organization}
In Section~\ref{sec:defin-notat} we introduce the notation, and recall
some basic definitions of operators in Fock spaces. In
Section~\ref{sec:assumptions-h} we formulate the necessary assumptions
on the operator $H$. In Section~\ref{sec:essent-self-adjo} we prove
the criterion. In Section~\ref{sec:applications} we outline some of
the most interesting applications. Finally in
Section~\ref{sec:conclusive-remarks-1} we give some conclusive
remarks, and an extension of the criterion to semi-bounded
quartic operators.
\subsection{Definitions and notations.}
\label{sec:defin-notat}
\begin{itemize}[label=\textcolor{darkblue}{\textbullet},itemsep=10pt]
\item Let $\mathscr{K}$ be a separable Hilbert space. Then the
symmetric Fock space $\Gamma_s(\mathscr{K})$ is defined as the
direct sum:
\begin{equation*}
\Gamma_s(\mathscr{K})=\bigoplus_{n=0}^{\infty}\mathscr{K}^{\otimes_s n}\; ,
\end{equation*}
where $\mathscr{K}^{\otimes_s n}$ is the $n$-fold symmetric tensor
product of $\mathscr{K}$, and $\mathscr{K}^{\otimes_s 0}:=\mathds{C}$.
\item Let $h:\mathscr{K}\supseteq D(h)\to \mathscr{K}$ be a densely
defined self-adjoint operator on a separable Hilbert space
$\mathscr{K}$. Its second quantization $d\Gamma(h)$ is the
self-adjoint operator on $\Gamma_s(\mathscr{K})$ defined by
\begin{equation*}
d\Gamma(h)\rvert_{D(h)^{\otimes_s n}}=\sum_{k=1}^n 1\otimes\dotsm\otimes \underbrace{h}_{k}\otimes \dotsm\otimes 1\; .
\end{equation*}
Let $u$ be a unitary operator on $\mathscr{K}$. We define
$\Gamma(u)$ to be the unitary operator on $\Gamma_s(\mathscr{K})$
given by
\begin{equation*}
\Gamma(u)\rvert_{\mathscr{K}^{\otimes_s n}}=\bigotimes_{k=1}^n u\; .
\end{equation*}
If $e^{it h}$ is a group of unitary operators on $\mathscr{K}$,
$\Gamma(e^{it h})=e^{it d\Gamma(h)}$.
\item $N:=d\Gamma (1)$ the number operator of
$\Gamma_s(\mathscr{H}_2)$.
\item $H_0:= H_{01}\otimes 1 + 1\otimes H_{02}$; the free Hamiltonian.
\item If $X$ is a self-adjoint operator on a Hilbert space, we denote
by $D(X)$ its domain, by $q_X(\cdot ,\cdot)$ the form associated
with $X$ and by $Q(X)$ the form domain.
\item Let $\mathscr{K}$ be a Hilbert space;
$\{\mathscr{K}^{(j)}\}_{j\in\mathds{N}}$ a collection of disjoint
subspaces of $\mathscr{K}$; $X$ an operator densely defined on
$\mathscr{K}$. We say that $\{\mathscr{K}^{(j)}\}_{j\in\mathds{N}}$
is invariant for $X$ if $\forall j\in \mathds{N}$, $X$ maps
$D(X)\cap \mathscr{K}^{(j)}\to \mathscr{K}^{(j)}$, and $D(X)\cap
\mathscr{K}^{(j)}$ is dense in $\mathscr{K}^{(j)}$.
\item Let $\mathscr{K}$ be a Hilbert space;
$\{\mathscr{K}^{(j)}\}_{j\in\mathds{N}}$ a collection of disjoint
closed subspaces of $\mathscr{K}$ such that
$\bigoplus_{j\in\mathds{N}}\mathscr{K}^{(j)}=\mathscr{K}$. Then we
call the collection complete, and we define the dense subset
$f_0(\mathscr{K}^{(\cdot)})$ of $\mathscr{K}$ as:
\begin{equation}
\label{eq:3}
f_0(\mathscr{K}^{(\cdot)})=\Bigl\{\phi\in\mathscr{K}, \exists n\in\mathds{N}\text{ s.t. }\phi\in\bigoplus_{j=0}^n\mathscr{K}^{(j)} \Bigr\}\; .
\end{equation}
Also, we denote by $\mathds{1}_j(\mathscr{K}^{(\cdot)})$ the
orthogonal projection on $\mathscr{K}^{(j)}$, by $\mathds{1}_{\leq
n}(\mathscr{K}^{(\cdot)})$ the orthogonal projection on
$\bigoplus_{j=0}^n\mathscr{K}^{(j)}$.
\item Let $\mathscr{K}\ni f,g$ be two elements of a separable Hilbert
space. We define the creation $a^{*}(f)$ and annihilation $a(f)$
operators on $\Gamma_s(\mathscr{K})$ by their action on $n$-fold
tensor products (with $a(f)\phi_0=0$ for any $\phi_0\in
\mathscr{K}^{\otimes_s 0}=\mathds{C}$):
\begin{align*}
a(f)g^{\otimes n}&=\sqrt{n} \; \langle f , g\rangle_{\mathscr{K}} \; g^{\otimes (n-1)}\\
a^{*}(f)g^{\otimes n}&=\sqrt{n+1} \; f\otimes_s g^{\otimes n}\; .
\end{align*}
They extend to densely defined closed operators and are adjoint of
each other: we denote again by $a^{\#}(f)$ their closures. For any
$f\in \mathscr{K}$, $D(a^{*}(f))=D(a(f))$ with
\begin{equation*}
D(a(f))=\Bigl\{\phi\in \Gamma_s(\mathscr{K})\;:\; \sum_{n=0}^{\infty}(n+1)\lVert \langle f(x) , \phi_{n+1}(x,X_n)\rangle_{\mathscr{K}(x)} \rVert_{\mathscr{K}^{\otimes_s n}(X_n)}^2<+\infty \Bigr\}\; ,
\end{equation*}
where $\phi_{n+1}=\phi\bigr\rvert_{\mathscr{K}^{\otimes_s n+1}}$;
also $D(a(f))\supset D(d\Gamma(1)^{1/2})$, $D(a(f))\supset
f_0(\mathscr{K}^{(\cdot)})$. They satisfy the Canonical
Commutation Relations $[a(f_1),a^{*}(f_2)]=\langle f_1 ,
f_2\rangle_{\mathscr{K}}$ on suitable domains
(e.g. $f_0(\mathscr{K}^{(\cdot)})$).
\item We decompose $\Gamma_s(\mathscr{H}_2)$ in its subspaces with
fixed number of particles as usual: $\forall n\in \mathds{N}$,
define $\mathscr{H}_2^{(n)}:=\mathscr{H}_2^{\otimes_s n}$, with the
convention $\mathscr{H}_2^{(0)}=\mathds{C}$. Then
$\{\mathscr{H}^{(n)}_{2}\}_{n\in\mathds{N}}$ is a complete
collection of closed disjoint subspaces of
$\Gamma_{s}(\mathscr{H}_2)$ invariant for $N$.
\item Let $X$ be an operator on $\mathscr{H}$. We say that $X$ is
diagonal if
$\{\mathscr{H}_1\otimes\mathscr{H}^{(n)}_{2}\}_{n\in\mathds{N}}$ is
invariant for $X$; $X$ is non-diagonal if for all $n\in\mathds{N}$
and $\phi\in D(X)\cap \mathscr{H}_1\otimes\mathscr{H}^{(n)}_{2}$,
$X\phi \notin \mathscr{H}_1\otimes\mathscr{H}^{(n)}_{2}$.
\end{itemize}
\section{Assumptions on $H$}
\label{sec:assumptions-h}
In this section we discuss Assumptions~\ref{ass:1}
and~\ref{ass:2}(\ref{ass:3}). In Section~\ref{sec:applications} below
they are checked in concrete examples.
We recall that our Hilbert space $\mathscr{H}$ has the form
\begin{equation*}
\mathscr{H}=\mathscr{H}_1\otimes \Gamma_s(\mathscr{H}_2)\; ;
\end{equation*}
while the operator is
\begin{equation*}
H= H_{01}\otimes 1 + 1\otimes H_{02}+H_I\; .
\end{equation*}
We separate the assumptions on $H_0$ from the ones on $H_I$, to
improve readability. On $H_I$ we require either Assumption~\ref{ass:2}
or Assumption~\ref{ass:3}. In \ref{ass:2} the non-diagonal part of
$H_I$ can be more singular: that restricts the diagonal part to be at
most quadratic in the creation and annihilation operators. In
\ref{ass:3} on the other hand is assumed more regularity on the
non-diagonal part of $H_I$, allowing for a more singular diagonal
part.
\begin{assumption}{A$_{0}$}
\label{ass:1}
$H_{01}$ and $H_{02}$ are semi-bounded self-adjoint operators. We
denote respectively by $-M_1$ and $-M_2$ their lower
bounds. Furthermore, $\forall t\in \mathds{R}$,
$\{\mathscr{H}_2^{(n)}\}_{n\in\mathds{N}}$ is invariant for $e^{it
H_{02}}$.
\end{assumption}
This is quite natural. In physical systems the Hamiltonian is often
split in a part describing the free dynamics (usually a
self-adjoint and positive unbounded operator), and an interaction
part. The invariance of the $n$-particles subspaces is also a usual
feature of free quantum theories: let $h_{02}$ be a semi-bounded
self-adjoint operator on the one-particle space $\mathscr{H}_2$; then
the second quantization $d\Gamma(h_{02})$ is self-adjoint, and the
group $\Gamma(e^{it h_{02}})$ generated by it satisfies the
assumption.
\begin{assumption}{A$_I$}
\label{ass:2}
$H_I$ is a symmetric operator on $\mathscr{H}$, with a domain of
definition $D(H_I)$ such that $D(H_0)\cap D(H_I)$ is dense in
$\mathscr{H}$. Furthermore $\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$,
\begin{equation}
\label{eq:5}
H_I\,\phi\in \bigoplus_{i=-2}^{2}\mathscr{H}_1\otimes \mathscr{H}_2^{(n+i)}\; .
\end{equation}
Also, $H_I$ satisfies the following bound: $\forall
n\in\mathds{N}$ $\exists C>0$ such that $\forall\psi\in
\mathscr{H}$, $\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes
H_{02})\cap\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$:
\begin{equation}
\label{eq:6}
\begin{split}
\lvert \langle \psi , H_I\phi \rangle_{\mathscr{H}}\rvert_{}^2\leq C^2\sum_{i=-2}^2\lVert \psi_{n+i} \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n+i)}}^2 \Bigl[(n+1)^2\lVert \phi \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n)}}^{2}+(n+1)\Bigl(q_{H_{01}\otimes 1}(\phi,\phi)\\+q_{1\otimes H_{02}}(\phi,\phi)+(\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\lVert \phi \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n)}}^{2} \Bigr)\Bigr]\; ;
\end{split}
\end{equation}
where we define
$\psi_n:=1\otimes
\mathds{1}_n(\mathscr{H}_2^{(\cdot)})\psi$.
\end{assumption}
Consider Assumption~\ref{ass:2}. First of all, $H_I$ has to be
sufficiently regular, i.e. relatively bounded by $H_0$ (in some sense)
when restricted to the subspaces
$\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$. Essentially, we require
that $H_I$ is at most quadratic in the annihilation and creation
operators, as reflected by the $n$-dependence in \eqref{eq:6}.
\begin{assumption}{A$_I^{\prime}$}
\label{ass:3}
$H_I$ is a symmetric operator on $\mathscr{H}$, with a domain of
definition $D(H_I)$ such that $D(H_0)\cap D(H_I)$ is dense in
$\mathscr{H}$. Furthermore $\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$,
\begin{equation}
\label{eq:24}
H_I\,\phi\in \bigoplus_{i=-2}^{2}\mathscr{H}_1\otimes \mathscr{H}_2^{(n+i)}\; .
\end{equation}
Also, $H_I=H_{diag}+H_2$ with the following properties:
\begin{enumerate}[label=\textcolor{darkblue}{\roman*)},itemsep=10pt]
\item \label{item:1} $H_{diag}$ is diagonal; $H_2$ is non-diagonal.
\item\label{item:2} $H_{diag}$ satisfies the following bound.
$\forall n\in \mathds{N}$ $\exists C(n)>0$ such that
$\forall\psi \in\mathscr{H}$, $\forall\phi\in Q(H_{01}\otimes
1)\cap Q(1\otimes H_{02})\cap\mathscr{H}_1\otimes
\mathscr{H}_2^{(n)}$:
\begin{equation}
\label{eq:16}
\begin{split}
\lvert \langle \psi , H_{diag}\phi \rangle_{\mathscr{H}}\rvert_{}^2\leq C^2(n)\lVert \psi_{n} \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n)}}^2 \Bigl(q_{H_{01}\otimes 1}(\phi,\phi)+q_{1\otimes H_{02}}(\phi,\phi)+(\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}\\+1)\lVert \phi \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n)}}^{2} \Bigr)\; .
\end{split}
\end{equation}
\item\label{item:3} $H_2$ satisfies the following bound. $\forall
n\in \mathds{N}$ $\exists C>0$ such that $\forall\psi
\in\mathscr{H}$, $\forall\phi\in \mathscr{H}_1\otimes
\mathscr{H}_2^{(n)}$:
\begin{equation}
\label{eq:19}
\begin{split}
\lvert \langle \psi , H_2\phi \rangle_{\mathscr{H}}\rvert\leq C(n+1) \lVert \phi \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n)}}\sum_{\substack{i=-2 \\ i\neq 0}}^2\lVert \psi_{n+i} \rVert_{\mathscr{H}_1\otimes\mathscr{H}_2^{(n+i)}} \; .
\end{split}
\end{equation}
\end{enumerate}
\end{assumption}
Assumption~\ref{ass:3} is similar to Assumption~\ref{ass:2}. However
since the non-diagonal quadratic part $H_2$ is more regular than
before, we can be less demanding on the diagonal part $H_{diag}$: it
has still to be bounded in a suitable sense by $H_0$, but it can be
non-quadratic with respect to the creation and annihilation operators.
\begin{remark}
In some applications, there is a decomposition of $\mathscr{H}_{1}$
invariant for $H$. For example, it may happen that $\mathscr{H}_1$
is also a Fock space but $H$ leaves invariant each sector with fixed
number of particles. In this situation, we can prove essential
self-adjointness with little less regularity on the assumptions. In
particular, Assumption~\ref{ass:2} would be changed in:
\begin{mdquote}
$H_I$ is a symmetric operator on $\mathscr{H}$, with a domain of
definition $D(H_I)$ such that $D(H_0)\cap D(H_I)$ is dense in
$\mathscr{H}$. Furthermore there exists a complete collection
$\{\mathscr{H}_{1}^{(j)}\otimes \Gamma_s(\mathscr{H}_2)
\}_{j\in\mathds{N}}$
invariant for $H_0$ and $H_I$ such that:
$\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1^{(j)}\otimes \mathscr{H}_2^{(n)}$,
\begin{equation*}
H_I\,\phi\in \bigoplus_{i=-2}^{2}\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n+i)}\; .
\end{equation*}
Also, $H_I$ satisfies the following bound:
$\forall j,n\in\mathds{N}$ $\exists C(j)>0$ such that
$\forall\psi\in \mathscr{H}$,
$\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes
H_{02})\cap\mathscr{H}_1^{(j)}\otimes \mathscr{H}_2^{(n)}$:
\begin{equation*}
\begin{split}
\lvert \langle \psi , H_I\phi\rangle_{\mathscr{H}}\rvert_{}^2\leq C^2(j)\sum_{i=-2}^2\lVert\psi_{j,n+i}\rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n+i)}}^2\Bigl[(n+1)^2\lVert \phi\rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n)}}^{2}+(n+1)\Bigl(q_{H_{01}\otimes 1}(\phi,\phi)\\+q_{1\otimes H_{02}}(\phi,\phi)+(\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\lVert \phi\rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n)}}^{2}\Bigr)\Bigr]\; ;
\end{split}
\end{equation*}
where we define
$\psi_{j,n}:=\mathds{1}_j(\mathscr{H}_1^{(\cdot)})\otimes
\mathds{1}_n(\mathscr{H}_2^{(\cdot)})\psi$.
\end{mdquote}
\begin{samepage}
Theorem~\ref{thm:1} would then read:
\begin{mdquote}
Assume \ref{ass:1} and \ref{ass:2}(\ref{ass:3}). Then $H$ is
essentially self adjoint on
$D(H_{01}\otimes 1)\cap D(H_{02}\otimes 1)\cap
f_0(\mathscr{H}_1^{(\cdot)}\otimes \mathscr{H}_2^{(\cdot)})$.
\end{mdquote}
\end{samepage}
\end{remark}
\section{Direct proof of self-adjointness}
\label{sec:essent-self-adjo}
In this section we present the criterion of essential self-adjointness
. The strategy is to prove that $\mathrm{Ran}(H\pm i)$ is dense in
$\mathscr{H}$, by an argument of reductio ad absurdum. As already
discussed, the non-diagonal part of $H_I$ is at most quadratic with
respect to the annihilation and creation operators of
$\Gamma_s(\mathscr{H}_2)$, and that plays a crucial role in the
proof. We prove Theorem~\ref{thm:1} assuming~\ref{ass:2}; the other
case being analogous.
\begin{thm}
\label{thm:1}
Assume \ref{ass:1} and \ref{ass:2}(\ref{ass:3}). Then $H$ is
essentially self adjoint on $D(H_{01}\otimes 1)\cap D(H_{02}\otimes
1)\cap \mathscr{H}_1\otimes f_0( \mathscr{H}_2^{(\cdot)})$.
\end{thm}
\begin{proof}
Let $\psi\in\mathscr{H}$, $z\in \mathds{C}$ with $\Im z\neq
0$. Suppose that $\forall \phi\in D(H_{01}\otimes 1)\cap D(1\otimes
H_{02})\cap \mathscr{H}_1\otimes f_0(\mathscr{H}_2^{(\cdot)})$:
\begin{equation}
\label{eq:7}
\langle \psi , (H-z)\phi\rangle_{\mathscr{H}}=0\; .
\end{equation}
Then it suffices to show that $\psi = 0$. This is done in few
steps. Let $n\in \mathds{N}$ and $\phi_{n}\in D(H_{01}\otimes
1)\cap D(1\otimes H_{02})\cap \mathscr{H}_1\otimes
\mathscr{H}_2^{(n)}$. For all $n\in\mathds{N}$, the space
$Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$ with the scalar
product:
\begin{equation}
\label{eq:8}
\langle \,\cdot \, ,\, \cdot\,\rangle_{\mathscr{X}_{n}}= q_{H_{01}\otimes 1}(\,\cdot\, ,\,\cdot\,)+q_{1\otimes H_{02}}(\,\cdot\, ,\, \cdot\,) +(\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\langle \, \cdot \, ,\, \cdot\,\rangle_{\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}}
\end{equation}
is complete, and therefore a Hilbert space. We denote it by
$\mathscr{X}_{n}$. Then~\eqref{eq:7} together with
Assumption~\ref{ass:1} imply, since $\phi_{n}\in D(H_{01}\otimes
1)\cap D(1\otimes H_{02})$:
\begin{equation}
\label{eq:9}
\langle \psi_{n}, \phi_{n} \rangle_{\mathscr{X}_{n}}= (z+\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\langle \psi_{n} ,\phi_{n}\rangle_{\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}}-\langle \psi , H_I\phi_{n} \rangle_{\mathscr{H}}\; .
\end{equation}
Use bound~\eqref{eq:16} and then Riesz's Lemma on
$\mathscr{X}_{n}$: it follows that $\psi_{n}\in Q(H_{01}\otimes
1)\cap Q(1\otimes H_{02})\cap \mathscr{H}_1\otimes
\mathscr{H}_2^{(n)}$ for any $n\in\mathds{N}$.
Let
$\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes f_0(\mathscr{H}_2^{(\cdot)})$.
Then $\exists \{\phi^{(\alpha)}\}_{\alpha\in\mathds{N}}$ such that
$\forall \alpha\in\mathds{N}$,
$\phi^{(\alpha)}\in D(H_{01}\otimes 1)\cap D(1\otimes H_{02})\cap
\mathscr{H}_1\otimes f_0( \mathscr{H}_2^{(\cdot)})$;
and $\forall n\in\mathds{N}$, $\phi_{n}^{(\alpha)}\to \phi_{n}$
in the topology induced by
$\lVert \,\cdot\, \rVert_{\mathscr{X}_{n}}^{}$. Furthermore
$\forall\alpha\in\mathds{N}$:
\begin{equation}
\label{eq:11}
\langle \psi , (H-z)\phi^{(\alpha)} \rangle_{\mathscr{H}}=0\; .
\end{equation}
Since
$\psi_{n}\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes \mathscr{H}_2^{(n)}$,
we can take the limit of~\eqref{eq:11} and obtain,
$\forall \phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1\otimes f_0( \mathscr{H}_2^{(\cdot)})$:
\begin{equation}
\label{eq:12}
q_{H_{01}\otimes 1}(\psi,\phi)+q_{1\otimes H_{02}}(\psi,\phi)+\langle \psi , H_I\phi \rangle_{\mathscr{H}}=z\langle \psi , \phi\rangle_{\mathscr{H}}\; .
\end{equation}
Hence we can choose
$\phi=\psi_{\leq n}:= 1\otimes \mathds{1}_{\leq
n}(\mathscr{H}_2^{(\cdot)})\psi$
in~\eqref{eq:12}. Then, using Assumption~\ref{ass:1} and taking the
imaginary part we obtain:
\begin{equation}
\label{eq:15}
\Im (z) \langle \psi_{\leq n} , \psi_{\leq n}\rangle=\Im (\langle \psi -\psi_{\leq n}, H_{I}\psi_{\leq n} \rangle_{})\; .
\end{equation}
Now, by Assumption~\ref{ass:2} (the equality holds on the suitable
domain):
\begin{equation*}
H_I \bigl(1\otimes \mathds{1}_{\leq n}(\mathscr{H}_2^{(\cdot)})\bigr) =\bigl(1\otimes \mathds{1}_{\leq n+2}(\mathscr{H}_2^{(\cdot)})\bigr)H_I \bigl(1\otimes \mathds{1}_{\leq n}(\mathscr{H}_2^{(\cdot)})\bigr)\; .
\end{equation*}
Furthermore
$1\otimes \mathds{1}_{\leq n+2}(\mathscr{H}_2^{(\cdot)})(\psi -
\psi_{\leq n})= \psi _{n+1}\oplus \psi _{n+2}$.
Then Equation~\eqref{eq:15} becomes:
\begin{equation*}
\Im (z) \langle \psi_{\leq n} , \psi_{\leq n}\rangle= \sum_{i=1}^2\Im(\langle \psi_{n+i} , H_I\psi_{\leq n}\rangle_{})\; .
\end{equation*}
Using the symmetry of $H_I$, and~\eqref{eq:5} we obtain:
\begin{equation}
\label{eq:17}
\Im (z) \langle \psi_{\leq n} , \psi_{\leq n}\rangle=\Im(\langle \psi_{n+2} , H_I\psi_{n}\rangle+\langle \psi_{n+1} , H_I\psi_{n}\rangle+\langle \psi_{n+1} , H_I\psi_{n-1}\rangle )\; .
\end{equation}
Now bound~\eqref{eq:17} using~\eqref{eq:6}; then we obtain
$\forall n\in \mathds{N}$:
\begin{equation}
\label{eq:18}
\begin{split}
\lvert \Im z\rvert\sum_{i=0}^n\lVert \psi_{i} \rVert_{}^2\leq C\Bigl[\lVert \psi_{n+1} \rVert\Bigl((n+1)\bigl(\lVert \psi_{n} \rVert+\lVert \psi_{n-1} \rVert \bigr)+\sqrt{n+1}\bigl(\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}+\lVert \psi_{n-1} \rVert_{\mathscr{X}_{n-1}}\bigr)\Bigr)\\+\lVert \psi_{n+2} \rVert\Bigl((n+1)\lVert \psi_{n} \rVert+\sqrt{n+1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}\Bigr)\Bigr]\\\leq 2C(n+1)\Bigl[ \sum_{i_1=0}^2\lVert \psi_{n+i_1} \rVert_{}^2+\sum_{i_2=-1}^0(n+1)^{-1}\lVert \psi_{n+i_2} \rVert_{\mathscr{X}_{n+i_2}}^2 \Bigr] \; .
\end{split}
\end{equation}
For all $\alpha >0$ define:
\begin{equation*}
S:= \sum_{n=0}^{\infty}\lVert \psi_{n} \rVert_{}^2\; ; \;S_\alpha:=\sum_{n=0}^{\infty}(n+\alpha)^{-1}\lVert \psi_{n}\rVert_{\mathscr{X}_{n}}^2\; .
\end{equation*}
$\psi\in\mathscr{H}$, hence $S$ is finite. We prove that also
$S_{\alpha}$ is finite. Using equation~\eqref{eq:12} with
$\phi=\psi_{n}$ we obtain, for all $n\in\mathds{N}$:
\begin{equation}
\label{eq:38}
(n+\alpha)^{-1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2=(n+\alpha)^{-1}(z+\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\lVert \psi_{n} \rVert_{}^2-(n+\alpha)^{-1}\langle \psi , H_I\psi_{n}\rangle_{}\; .
\end{equation}
Now, we can use bound~\eqref{eq:6} on
$(n+\alpha)^{-1}\lvert \langle \psi ,
H_I\psi_{n}\rangle\rvert_{}^{}$, obtaining
\begin{equation}
\label{eq:39}
\begin{split}
(n+\alpha)^{-2}\lvert \langle \psi , H_I\psi_{n} \rangle_{\mathscr{H}}\rvert_{}^2\leq C^2 (n+\alpha)^{-2}\sum_{i=-2}^2\lVert \psi_{n+i} \rVert^2 \Bigl[(n+1)^2\lVert \psi_{n} \rVert^2+(n+1)\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2\Bigr]\\\leq C^2(\alpha)\sum_{i=-2}^2\lVert \psi_{n+i} \rVert_{}^2 \Bigl[\lVert \psi_{n} \rVert_{}^2+(n+\alpha)^{-1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2\Bigr]\; ,
\end{split}
\end{equation}
for some $C(\alpha)>0$. The only terms we need to deal with are
$(n+\alpha)^{-1}\lVert \psi_{n+i} \rVert_{}^2 \lVert \psi_{n}
\rVert_{\mathscr{X}_{n}}^2$.
We use the fact that for any $\varepsilon,a,b>0$,
$ab\leq \frac{1}{2}(\varepsilon a^2+\frac{1}{\varepsilon}b^2)$,
obtaining
\begin{equation}
\label{eq:40}
(n+\alpha)^{-1}\lVert \psi_{n+i} \rVert_{}^2 \lVert \psi_{n}\rVert_{\mathscr{X}_{n}}^2\leq \frac{1}{2}\Bigl(\varepsilon(n+\alpha)^{-2}\lVert \psi_{n}\rVert_{\mathscr{X}_{n}}^4+\frac{1}{\varepsilon}\lVert \psi_{n+i} \rVert_{}^4\Bigr)\; .
\end{equation}
Combining~\eqref{eq:40} with~\eqref{eq:39}, and applying to
Equation~\eqref{eq:38}, we obtain the following bound: for all
$\varepsilon,\alpha>0$, $\exists
C(\alpha,\varepsilon)>0$ such that
\begin{equation}
\label{eq:41}
(n+\alpha)^{-1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2\leq C(\alpha,\varepsilon)\sum_{i=-2}^2\lVert \psi_{n+i} \rVert_{}^2+\varepsilon(n+\alpha)^{-1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2\; .
\end{equation}
Fix $\varepsilon <1$, then for all $\alpha >0$, $\exists C(\alpha)>0$ such that $\forall \bar{n}\in
\mathds{N}$:
\begin{equation}
\label{eq:4}
\sum_{n=0}^{\bar{n}} (n+\alpha)^{-1}\lVert \psi_{n} \rVert_{\mathscr{X}_{n}}^2\leq C(\alpha) S\; ;
\end{equation}
uniformly in $\bar{n}$. Then we can take the limit
$\bar{n}\to\infty$ and obtain $S_{\alpha}<\infty$.
\begin{remark*}
The bound of Equation~\eqref{eq:41} could seem to follow from an
implicit smallness condition on the interaction $H_I$. As it will
become clearer with the examples of
Section~\ref{sec:applications}, it is not the case. Roughly
speaking, Assumption~\ref{ass:2} allows for interaction parts that
are at most as singular as
$(H_0+\lvert M_1\rvert_{}^{}+\lvert
M_2\rvert_{}^{})^{1/2}(N+1)^{1/2}$.
\end{remark*}
Now return to Equation~\eqref{eq:18}. There exists $n^{*}\in\mathds{N}$ such that $\forall n\geq n^{*}$:
\begin{equation*}
\frac{1}{2}S\leq \sum_{i=0}^{n}\lVert \psi_{i} \rVert_{}^2\leq S\; .
\end{equation*}
Hence summing in $n^{*} \leq n \leq \bar{n}$ on both sides
of~\eqref{eq:18} we obtain for all $\bar{n}>n^{*}$:
\begin{equation*}
\begin{split}
\frac{1}{2}S\sum_{n=n^{*}}^{\bar{n}}(n+1)^{-1}\leq \sum_{n=n^{*}}^{\bar{n}}(n+1)^{-1}\sum_{i=0}^n\lVert \psi_{i} \rVert_{}^2\leq 2 \frac{C}{\lvert \Im z\rvert_{}^{}}(3S+S_1+S_2)\; .
\end{split}
\end{equation*}
The bound on the right hand side is uniform in $\bar{n}$: that is
absurd, unless $S=S_1=S_2=0$
$\Leftrightarrow$ $\psi =0$.
\end{proof}
Once essential self-adjointness is established, it is possible to give
the following characterization of the domain of self-adjointness
$D(H)$.
\begin{proposition}
\label{prop:1}
Assume \ref{ass:1} and \ref{ass:2}(\ref{ass:3}). If exists $K$
self-adjoint operator with domain $D(K)$ such that:
\begin{enumerate}[label=\textcolor{darkblue}{\roman*)},itemsep=10pt]
\item $D(H_{0})\cap D(K)$ is dense in $\mathscr{H}$;
$\mathscr{H}_1\otimes f_0(\mathscr{H}_2^{(\cdot)})$ is
dense in $D(K)$.
\item There exists $ 0 < \varepsilon < 1$ such that $\exists
C(\varepsilon)>0$, $\forall \phi\in D(H_{0})\cap D(K)$:
\begin{equation}
\label{eq:20}
\lVert H_I\phi \rVert_{}^{}\leq \varepsilon\lVert H_{0}\phi \rVert_{}^{}+C(\varepsilon)(\lVert K\phi \rVert_{}^{}+\lVert \phi \rVert_{}^{})\; .
\end{equation}
\end{enumerate}
Then $D(H)\cap D(K)=D(H_{0})\cap D(K)$.
\end{proposition}
\begin{proof}
Using bound~\eqref{eq:20}, we have $\forall \phi \in D(H_{0})\cap
D(K)$:
\begin{equation}
\label{eq:21}
\lVert H\phi \rVert_{}^{}\leq (\varepsilon+1)\lVert H_{0}\phi \rVert_{}^{}+C(\varepsilon)(\lVert K\phi \rVert_{}^{}+\lVert \phi \rVert_{}^{})\; .
\end{equation}
Then $D(H)\supseteq D(H_{0})\cap D(K)$. Now let $\phi\in D(H)\cap
D(K)$: using~\eqref{eq:20}
\begin{equation}
\label{eq:22}
\lVert H_{0}\phi \rVert_{}^{}\leq \varepsilon\lVert H_0\phi \rVert_{}^{} +\lVert H\phi \rVert_{}^{}+C(\varepsilon)(\lVert K\phi \rVert_{}^{}+\lVert \phi \rVert_{}^{})\; ;
\end{equation}
since $\varepsilon <1$, $D(H_0)\supseteq D(H)\cap D(K)$.
\end{proof}
\section{Applications}
\label{sec:applications}
It is possible to apply Theorem~\ref{thm:1} in several situations of
mathematical and physical interest. We present and discuss some of
them in this section; not before a brief discussion of the
``boundaries'' of Theorem~\ref{thm:1}: it may be interesting to see
how its proof fails when we consider operators that are more than
quadratic in the annihilation/creation operators; and to define a
quadratic operator that is not sufficiently regular for
Assumption~\ref{ass:2}(\ref{ass:3}) to hold. According to this
purpose, we will consider simple toy models on
$\Gamma _s(\mathds{C})$. We denote by $a^{\#}$ the corresponding
annihilation/creation operators.
Let's consider a simple trilinear Hamiltonian on $\Gamma _s(\mathds{C})$:
\begin{equation*}
H_3=a^{*}a+a^{*}a^{*}a^{*}+aaa\; .
\end{equation*}
The free part is $H_0=a^{*}a$, and the interaction part is
$H_I=a^{*}a^{*}a^{*}+aaa$. Assumption~\ref{ass:1} is satisfied, and
Assumption~\ref{ass:3} is slightly modified: $i$ now ranges from $-3$
to $3$, and bounds~\eqref{eq:16} and~\eqref{eq:19} are replaced by the
simple bound:
\begin{equation*}
\lvert \langle \psi , H_I\phi \rangle \rvert_{\mathscr{H}}\leq C(n+1)^{3/2}\lVert \phi \rVert_{\mathscr{H}^{(n)}}^{}\bigl(\lVert \psi _{n+3} \rVert_{\mathscr{H}^{(n+3)}}^{}+\lVert \psi _{n-3} \rVert_{\mathscr{H}^{(n-3)}}^{}\bigr)\; .
\end{equation*}
The proof of Theorem~\ref{thm:1} carries on, almost unchanged, up to
Equation~\eqref{eq:18} that would now read
\begin{equation*}
\lvert \Im z \rvert_{}^{}\sum_{i=0}^n \lVert \psi_i \rVert_{}^2\leq C (n+1)^{3/2}\lVert \psi _{n+3} \rVert_{}^2\; .
\end{equation*}
However if we now take the sum in $n$ from $n^{*}$ to $\bar{n}$ (where
$n^{*}$ is such that
$\frac{1}{2}\lVert \psi \rVert_{}^2\leq \sum_{i=0}^n \lVert \psi_i
\rVert_{}^2\leq \lVert \psi \rVert_{}^2$
for all $n\geq n^{*}$) we \emph{cannot} conclude that
$\lVert \psi \rVert_{}^{}$ must be zero, because the series
$\sum_{n=0}^{\infty }(n+1)^{-3/2}$ converges. Hence the proof fails,
and analogously would fail for any higher order polynomial of the
annihilation/creation operators.
On the other hand, we introduce now a quadratic model for which
Assumption~\ref{ass:2}(\ref{ass:3}) fails to hold, and thus
Theorem~\ref{thm:1} cannot be applied. For the following operator on
$L^2 (\mathds{R}^{} )\otimes \Gamma _s(\mathds{C})$
Assumption~\ref{ass:2} \emph{is satisfied}:
\begin{equation*}
H_{\partial a}=-\partial _x^2+a^{*}a-i\partial _x(a^{*}+a)+a^{*}a^{*}+aa\; ,
\end{equation*}
where $H_0=-\partial _x^2+a^{*}a$ and
$H_I=-i\partial _x(a^{*}+a)+a^{*}a^{*}+aa$. If, however, the derivative
operator is coupled with the quadratic term
\begin{equation*}
H_{\partial aa}=-\partial _x^2+a^{*}a-i\partial _x(a^{*}a^{*}+aa)\; ,
\end{equation*}
\ref{ass:2}(\ref{ass:3}) is no longer satisfied. The interaction in
this case would be of type $H_0^{1/2}N$, and therefore too singular:
Theorem~\ref{thm:1} \emph{does not hold} for $H_{\partial aa}$.
Throughout the section we will adopt the following notations, in
addition to the ones of Section~\ref{sec:defin-notat}. Let
$\mathscr{K}$ a Hilbert space; we denote by $\mathcal{L}(\mathscr{K})$
the set of bounded operators on $\mathscr{K}$ and by
$\lvert \,\cdot\,\rvert_{\mathcal{L}(\mathscr{K})}^{}$ the operator
norm. It is also useful to define the annihilation/creation operator
valued distributions $a^{\#}(x)$, $x\in\mathds{R}^d$. Let
$f\in L^2 (\mathds{R}^d)$, $a^{\#}(f)$ the annihilation/creation
operators on $\Gamma_s(L^2 (\mathds{R}^d))$. Then the operator valued
distributions $a^{\#}(x)$ acting on $L^2 (\mathds{R}^d)$, with values
on $\Gamma_s(L^2 (\mathds{R}^d))$, are defined by:
\begin{equation*}
(a^{*},f)\equiv\int_{\mathds{R}^d}a^{*}(x)f(x)dx:=a^{*}(f)\; ;\; (a,f)\equiv\int_{\mathds{R}^d}a(x)\bar{f}(x)dx:=a(f)\; .
\end{equation*}
They satisfy the commutation relations (inherited by the CCR)
$[a(x),a^{*}(y)]=\delta(x-y)$.
\subsection{Hamiltonians of identical bosons.}
\label{sec:hamilt-many-bosons}
The criterion applies to operators in the Fock space
$\Gamma_s(\mathscr{K})$, for any separable Hilbert space
$\mathscr{K}$. Simply choose $\mathscr{H}_1\equiv \mathds{C}$ and
$\mathscr{H}_2\equiv \mathscr{K}$; then $\mathds{C}\otimes
\Gamma_s(\mathscr{K})\approx \Gamma_s(\mathscr{K})$ up to an unitary
isomorphism.
An example is given by the following class of operators. Let
$\mathscr{K}=L^2 (\mathds{R}^d)$; $h_0$ a positive self adjoint
operator on $L^2 (\mathds{R}^{d})$ (the one-particle free
Hamiltonian). Furthermore, let $V_1\in L^2 (\mathds{R}^{d})$, $V_2,
V_3\in L^2 (\mathds{R}^{2d})$, with $V_2=\overline{V}_2$, and
$V_4(\cdot):\mathds{R}^d\to \mathds{R}$, such that $V_4(x)=V_4(-x)$
and $V_4\,(h_0+1)^{-1/2}\in \mathcal{L}(L^2 (\mathds{R}^d))$. Consider
\begin{equation}
\label{eq:32}
\begin{split}
H=d\Gamma(h_0)+\int_{\mathds{R}^d}^{}\Bigl(V_1(x)a^{*}(x)+\overline{V}_1(x)a(x)\Bigr) dx+ \int_{\mathds{R}^{2d}}^{}\Bigl(V_2(x,y)a^{*}(x)a(y)+V_3(x,y)a^{*}(x)\\
a^{*}(y) +\overline{V}_3(x,y)a(x)a(y)\Bigr) dxdy+\frac{1}{2}\int_{\mathds{R}^{2d}}^{}V_4(x-y)a^{*}(x)a^{*}(y)a(x)a(y) dxdy\; .
\end{split}
\end{equation}
We make the following identifications:
$H_{01}\equiv 0$, $H_{02}\equiv d\Gamma(h_0)$, $H_{diag}\equiv\int
(V_4a^{*}a^{*}aa+V_2a^{*}a)$, $H_2\equiv
\int_{}^{}(V_1a^{*}+\overline{V}_1a)+\int
(V_3a^{*}a^{*}+\overline{V}_3aa)$. Assumption~\ref{ass:1} is trivial
to verify; and Assumption~\ref{ass:3} follows from standard estimates
on Fock space: let $\psi\in\Gamma_s(L^2 (\mathds{R}^d))$, $\phi_n\in
L^2_s(\mathds{R}^{nd})\cap Q(d\Gamma(h_0))$, $n\in\mathds{N}$, then
\begin{gather}
\label{eq:35}
\begin{split}
\lvert \langle \psi, H_{diag}\phi_n \rangle_{}\rvert_{}^{}\leq \Bigl(n\lVert V_2 \rVert_2^{}\lVert \phi_n \rVert_{}^{}+\lvert V_4\,(h_{0}+1)^{-1/2}\rvert_{\mathcal{L}(L^2 (\mathds{R}^d))}^{}\bigl(n^{3/2}\lVert (d\Gamma(h_0))^{1/2}\phi_n \rVert_{}^{}\\+n^2\lVert \phi_n \rVert_{}^{}\bigr)\Bigr)\lVert \psi_n \rVert_{}^{}\; ;
\end{split}\\
\label{eq:36} \lvert \langle \psi, H_2\phi_n \rangle_{}\rvert_{}^{}\leq 2\Bigl(\sqrt{n+1}\lVert V_1 \rVert_2^{}+ (n+1)\lVert V_3 \rVert_2^{}\Bigr)\lVert \phi_n \rVert_{}^{}\sum_{\substack{i=-2\\i\neq 0}}^{2}\lVert \psi_{n+i} \rVert_{}^{}\; .
\end{gather}
Hence we can apply Theorem~\ref{thm:1}; and prove essential
self-adjointness of $H$ in $D(d\Gamma (h_0))\cap
f_0(L^2(\mathds{R}^{d})^{(\cdot)})$. We can also apply
Proposition~\ref{prop:1} with $K\equiv N^{3}$, i.e. $D(H)\cap
D(N^{3})=D(d\Gamma(h_0))\cap D(N^{3})$. Observe that if $d=3$,
the well-known many body Hamiltonian with Coulomb pair interaction
\begin{equation*}
H_C=d\Gamma(-\Delta)\pm\frac{1}{2}\int_{\mathds{R}^6}^{}\frac{1}{\lvert x-y\rvert_{}^{}}a^{*}(x)a^{*}(y)a(x)a(y) dxdy\; ,
\end{equation*}
is just the special case $h_0= -\Delta$, $V_1=V_2=V_3=0$ and
$V_4=\pm \lvert x \rvert_{}^{-1}$.
\subsection{Nelson-type Hamiltonians.}
\label{sec:nels-type-hamilt}
We consider now the dynamics of different species of particles (or
fields) interacting. A typical example is the Nelson Hamiltonian. It
was introduced in a rigorous way by \citet{nelson:1190} to describe
nucleons in a meson field, and studied by several authors
\citep[e.g.][]{DG1,MR1814991,MR1809881,GHPS}.
Let $\mathscr{H}=L^2 (\mathds{R}^{pd})\otimes \Gamma_s(L^2
(\mathds{R}^d))$: the first space corresponds to $n$ non-relativistic
particles; the second to a scalar relativistic field. Let $\omega$ be
a positive self-adjoint operator on $L^2 (\mathds{R}^d)$ (the
dispersion relation of the relativistic field), $V\in L^2_{loc}
(\mathds{R}^d, \mathds{R}_{+})$ an external potential acting on the
particles. The interaction between the particles and the field is
linear in the creation and annihilation operators $a^{\#}$
corresponding to the field. Let $v:\mathds{R}^{2d}\to \mathds{C}$ such
that
\begin{itemize}[label=\textcolor{darkblue}{\textbullet},itemsep=10pt]
\item $(1-\Delta_x)^{-1/2}\lVert v(x,\cdot) \rVert_{L^2_{(k)}
(\mathds{R}^d)}^2(1-\Delta_x)^{-1/2}\in \mathcal{L}(L^2_{(x)}
(\mathds{R}^{d}))$;
\item for all $k\in\mathds{R}^d$, $v(x,k)(1-\Delta_x)^{-1/2}\in
\mathcal{L}(L^2_{(x)} (\mathds{R}^d))$, with $\lvert
v(x,\cdot)(1-\Delta_x)^{-1/2}\rvert_{\mathcal{L}(L^2_{(x)}
(\mathds{R}^{d}))}^{}\in L^2_{(k)} (\mathds{R}^d)$.
\end{itemize}
Then we define the Nelson Hamiltonian:
\begin{equation}
\label{eq:23}
H_N=\Bigl(\sum_{i=1}^p -\Delta_{x_i}+V(x_i)\Bigr)\otimes 1+1\otimes d\Gamma(\omega)+\sum_{i=1}^pa^{*}(v(x_{i},\cdot))+a(v(x_i,\cdot))\; .
\end{equation}
The function $v$ describes the coupling between the particles and the
relativistic field. The assumptions above imply that it has a good
behaviour both for high and small momenta; in particular in
three-dimensions it acts as an UV cutoff function.
\begin{remark*}
The model of \citet{nelson:1190} was much more specific: $d=3$,
$\omega(k)=\sqrt{k^2+\mu^2}$ with $\mu>0$, $V=0$ and
$v(x,k)=\lambda(2\pi)^{-3/2}(2\omega(k))^{-1/2}e^{-ik\cdot
x}\mathds{1}_{\lvert \,\cdot\,\rvert \leq \sigma}(k)$ with
$\lambda,\sigma >0$. With these assumptions, $v\in L^{\infty}
(\mathds{R}^3,L^2 (\mathds{R}^3))$, $\omega^{-1/2}v\in L^{\infty}
(\mathds{R}^3,L^2 (\mathds{R}^3))$; then $H_N$ (the Nelson model
with UV cut off) is self-adjoint by the Kato-Rellich
Theorem. However, if we consider $d=2$ and $\mu=0$ (massless
relativistic field), the Kato-Rellich Theorem is not applicable
because $\omega^{-1/2}v\notin L^{\infty} (\mathds{R}^2,L^2
(\mathds{R}^2))$ due to an infrared divergence. Instead assumptions
\ref{ass:1} and \ref{ass:3} are still satisfied, thus
Theorem~\ref{thm:1} can be used.
\end{remark*}
In order to check Assumptions~\ref{ass:1} and~\ref{ass:2}
on~\eqref{eq:23}, we make the (straightforward) identifications:
$\mathscr{H}_1\equiv L^2 (\mathds{R}^{pd})$, $\mathscr{H}_2\equiv L^2
(\mathds{R}^d)$, $H_{01}\equiv\sum_i-\Delta_{x_i}+V(x_i)$,
$H_{02}\equiv d\Gamma(\omega)$, $H_I\equiv
\sum_ia^{*}(v(x_i,\cdot))+a(v(x_i,\cdot))$. We do not need to
introduce a decomposition of $\mathscr{H}_1$. Assumption~\ref{ass:1}
is satisfied: for all $V\in L^2_{loc} (\mathds{R}^d,\mathds{R}_{+})$,
$-\Delta+V(\cdot)$ is a positive self-adjoint operator, and the
vectors with fixed number of particles are invariant for the evolution
associated with the positive self-adjoint operator
$d\Gamma(\omega)$. Furthermore, since $H_{01}\otimes 1$ and $1\otimes
H_{02}$ are positive self-adjoint commuting operators, $H_0$ is a
positive self-adjoint operator with domain $D(H_0)= D(H_{01}\otimes
1)\cap D(1\otimes H_{02})$. Assumption \ref{ass:2} is also satisfied
by usual estimates: $\forall \psi \in \mathscr{H}$, $\forall\phi_n\in
L^2 (\mathds{R}^{pd})\otimes L^2_s(\mathds{R}^{nd})\cap
Q(H_{01}\otimes 1)$, $n\in\mathds{N}$,
\begin{equation}
\label{eq:13}
\begin{split}
\lvert \langle \psi , H_I\phi_n \rangle_{}\rvert\leq \sqrt{2p}\bigl( 2 \sqrt{n} \lVert \lvert v(x,\cdot)(1-\Delta_x)^{-1/2}\rvert_{\mathcal{L}(L^2_{(x)})}^{} \rVert_{L^2_{(k)}} + \lvert (1-\Delta_x)^{-1/2}\lVert v(x,\cdot) \rVert_{L^2_{(k)}}^2\\(1-\Delta_x)^{-1/2}\rvert_{\mathcal{L}(L^2_{(x)})}^{1/2}\bigr)\Bigl(\Bigl\lVert \bigl(\sum_{i=1}^p-\Delta_{x_i}\bigr)^{1/2}\phi_n \Bigr\rVert+\sqrt{p} \lVert \phi_n \rVert\Bigr) \sum_{\substack{i=-1\\i\neq 0}}^{1}\lVert \psi_{n+i} \rVert\; .
\end{split}
\end{equation}
Then $H_N$ is essentially self-adjoint on $D(H_0)\cap f_0(L^2
(\mathds{R}^{pd})\otimes L^2(\mathds{R}^{d})^{(\cdot)})$.
Let $H_N\rvert_s$ be the restriction of $H_N$ to $L^2_s
(\mathds{R}^{pd})\otimes \Gamma_s(L^2 (\mathds{R}^d))$. It is possible
to extend $H_N\rvert_s$ to $\Gamma_s(L^2 (\mathds{R}^d))\otimes
\Gamma_s(L^2 (\mathds{R}^d))$ in the following way. Define
\begin{equation}
\label{eq:25}
\widetilde{H}_N=d\Gamma(-\Delta+V)\otimes 1+1\otimes d\Gamma(\omega)+\int_{\mathds{R}^d}^{} \psi^{*}(x)\bigl(a^{*}(v(x,\cdot))+a(v(x,\cdot))\bigr)\psi(x)dx\; ,
\end{equation}
where $\psi^{\#}$ are the creation and annihilation operators
corresponding to the first Fock space. Then $H_N\rvert_s$ and
$\widetilde{H}_N$ agree on the $p$-particle sector $L^2_s
(\mathds{R}^{pd})\otimes \Gamma_s(L^2 (\mathds{R}^d))$ of
$\Gamma_s(L^2 (\mathds{R}^d))\otimes \Gamma_s(L^2
(\mathds{R}^d))$. The self-adjointness of $\widetilde{H}_N$ still
follows from Theorem~\ref{thm:1} using the bound~\eqref{eq:13}: it is
sufficient to choose for $\mathscr{H}_1\equiv \Gamma_s(L^2
(\mathds{R}^d))$ the decomposition in finite particle vectors
$\{\mathscr{H}_{1}^{(j)}\otimes \Gamma_s(\mathscr{H}_2)
\}_{j\in\mathds{N}}\equiv \{L^2_s (\mathds{R}^{jd})\otimes
\Gamma_s(\mathscr{H}_2) \}_{j\in\mathds{N}}$. Let
$H_0\equiv d\Gamma(-\Delta+V)\otimes 1 +1\otimes d\Gamma(\omega)$, then the
domain of essential self-adjointness for $\widetilde{H}_N$ is
$D(H_0)\cap f_0(L^2(\mathds{R}^{d})^{(\cdot)}\otimes
L^2(\mathds{R}^{d})^{(\cdot)})$. Let $N_1$ and $N_2$ be the number
operators corresponding to the first and second Fock space
respectively. Then applying Proposition~\ref{prop:1} we also obtain
$D(\widetilde{H}_N)\cap D(N_1^2+N_2^2)=D(H_0)\cap D(N_1^2+N_2^2)$.
\subsection{Pauli-Fierz Hamiltonian.}
\label{sec:pauli-fierz-hamilt}
The last example considered is an operator describing the dynamics of
rigid charges and their radiation field interacting. The model was
introduced by \citet{pauli1938theorie}, and has been extensively
studied by a mathematical standpoint. See \citet[][and references
thereof contained]{MR2097788} for a detailed presentation.
Let $\mathscr{H}^{(spin)}=(\otimes^p
\mathds{C}^{2[\frac{d}{2}]})\otimes L^2 (\mathds{R}^{pd})\otimes
\Gamma_s(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))$, $\mathscr{H}=
L^2 (\mathds{R}^{pd})\otimes \Gamma_s(\mathds{C}^{d-1}\otimes L^2
(\mathds{R}^d))$: the first space corresponds to $p$
spin-$\frac{1}{2}$ particles, the second to spinless particles. Let
$\chi\in L^2 (\mathds{R}^d)$, $V\in L^2_{loc}
(\mathds{R}^{pd},\mathds{R}_{+})$, $\omega=\lvert k\rvert $, $m_j>0$,
$q_j\in \mathds{R}$ for all $j=1,\dotsc,p$. Furthermore, let
$e_{\lambda}: \mathds{R}^d\to \mathds{R}^d$ such that for almost all
$k\in \mathds{R}^d$, $k\cdot e_{\lambda}(k)=0$ and $e_{\lambda}(k)
\cdot e_{\lambda'}(k)=\delta_{\lambda\lambda'}$ for all
$\lambda,\lambda'=1,\dotsc, d-1$. Then we define the electromagnetic
vector potential in the Coulomb gauge as
\begin{equation}
\label{eq:26}
A(x)=\sum_{\lambda=1}^{d-1}\int_{\mathds{R}^d}^{}e_{\lambda}(k)\Bigl( a^{*}_{\lambda}(k)\chi(k)e^{ik\cdot x}+a_{\lambda}(k)\bar{\chi}(k)e^{ik\cdot x} \Bigr) dk\; ;
\end{equation}
where $a_{\lambda}^{\#}$ are the creation and annihilation operators
of $\Gamma_s(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))$ satisfying
the canonical commutation relations
$[a_{\lambda}(k),a^{*}_{\lambda'}(k')]=\delta_{\lambda\lambda'}\delta(k-k')$;
the (spinless) Pauli-Fierz Hamiltonian on $\mathscr{H}$ is then
\begin{equation}
\label{eq:27}
H_{PF}= \sum_{j=1}^p \frac{1}{2m_j}\bigl(-i\nabla_j\otimes 1+q_j A(x_j)\bigr)^2 + V(x_1,\dotsc, x_p)\otimes 1 + 1\otimes \sum_{\lambda=1}^{d-1}\int_{\mathds{R}^d}^{}\omega(k)a^{*}_{\lambda}(k)a_{\lambda}(k) dk\; .
\end{equation}
The function $\chi$ plays the role of an ultraviolet cut off in the
interaction, and is usually interpreted as the
Fourier transform of the particles' charge distribution. Let
$\{\sigma^{(\mu)}\}_{\mu=1}^d$ the $2^{[\frac{d}{2}]}\times
2^{[\frac{d}{2}]}$ matrices satisfying
$\sigma^{(\mu)}\sigma^{(\nu)}+\sigma^{(\nu)}\sigma^{(\mu)}=2\delta_{\mu\nu}\mathrm{Id}$. Also,
denote by $\sigma_j^{(\mu)}$, $j=1,\dotsc, p$ the operator on
$(\otimes^p \mathds{C}^{2[\frac{d}{2}]})$ acting as $\sigma^{(\mu)}$
on the $j$-th space of the tensor product. Then the spin-$\frac{1}{2}$
Pauli-Fierz Hamiltonian on $\mathscr{H}^{(spin)}=(\otimes^p
\mathds{C}^{2[\frac{d}{2}]})\otimes \mathscr{H}$ can be written as:
\begin{equation}
\label{eq:28}
H_{PF}^{(spin)}=1\otimes H_{PF} +\frac{i}{2}\sum_{j=1}^pq_j\sum_{1\leq \mu<\nu\leq d}^{}\sigma_j^{(\mu)}\sigma_j^{(\nu)}\otimes \Bigl(\partial_j^{(\mu)}A^{(\nu)}(x_j)-\partial_j^{(\nu)}A^{(\mu)}(x_j)\Bigr)\; ;
\end{equation}
where $A^{(\mu)}(x)$ is the $\mu$-th component of the vector $A(x)$.
The quadratic form corresponding to the Pauli-Fierz Hamiltonian is
bounded from below, so it is possible to define at least one
self-adjoint extension by means of the Friedrichs Extension
Theorem. This type of information is not completely satisfactory,
since infinitely many extensions may exist, each one dictating a
different dynamics for the system. For small values of the ratios
$q^2_j/m_j$ between charge and mass of the particles, and if
$\chi,\chi/\sqrt{\omega}\in L^2 (\mathds{R}^d)$, a unique self-adjoint
extension is given by KLMN Theorem. For arbitrary values of the ratios
$q^2_j/m_j$, it is possible to prove essential self-adjointness of
both $H_{PF}$ and $H_{PF}^{(spin)}$ (for the spin operator we need in
addition $\omega\chi\in L^2 (\mathds{R}^d)$) by means of
Theorem~\ref{thm:1}, under the sole assumption $\chi\in L^2
(\mathds{R}^d)$. As discussed in Section~\ref{sec:introduction}, an
analogous result (on a slightly different domain) has been obtained
with an argument of functional integration by \citet{MR1891842}. If
the dependence on $x$ of $A(x)$ is more general, functional
integration methods may not be applicable; however Theorem~\ref{thm:1}
still holds.
In the following discussion we will focus on a simplified model, for
the sake of clarity. Assumptions~\ref{ass:1} and~\ref{ass:2} are
checked on $H_{PF}$ with $p=1$, $m=1/2$ and $q=-1$, i.e.:
$\mathscr{H}\equiv L^2 (\mathds{R}^d)\otimes
\Gamma_s(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))$ and
\begin{equation}
\label{eq:29}
H\equiv \bigl(i\nabla_x\otimes 1+A(x)\bigr)^2 + V(x)\otimes 1 + 1\otimes \sum_{\lambda=1}^{d-1}\int_{\mathds{R}^d}^{}\omega(k)a^{*}_{\lambda}(k)a_{\lambda}(k) dk\; .
\end{equation}
Observe that, since we are in the Coulomb gauge, $\nabla_x\cdot
A(x)=0$ hence $[-i\nabla_x\otimes 1, A(x)]=0$ on a suitable dense
domain. Rewrite $H$ in the following form, to identify the free and
interaction parts:
\begin{equation}
\label{eq:30}
\begin{split}
H= \bigl(-\Delta_x+V(x)\bigr)\otimes 1+1\otimes \sum_{\lambda=1}^{d-1}\int_{\mathds{R}^d}^{}\omega(k)a^{*}_{\lambda}(k)a_{\lambda}(k) dk +2i A(x)\cdot(\nabla_x\otimes 1)+ A^2(x)\; .
\end{split}
\end{equation}
We identify $H_{01}\equiv -\Delta+V$, $H_{02}\equiv
\sum_{\lambda}\int_{\mathds{R}^d}^{}\omega a^{*}_{\lambda}a_{\lambda}$
and $H_I\equiv 2i A\cdot(\nabla\otimes 1)+
A^2$. Assumption~\ref{ass:1} is satisfied, as in the Nelson
model~\eqref{eq:23} above. For the interaction part, we have the
following bounds: $\forall\psi\in\mathscr{H}$, $\forall\phi_n\in L^2
(\mathds{R}^d)\otimes (\mathds{C}^{d-1}\otimes L^2
(\mathds{R}^d))^{\otimes_s n}\cap Q(H_{01}\otimes 1)$,
$n\in\mathds{N}$,
\begin{equation}
\label{eq:33}
\begin{aligned}
\lvert \langle \psi , &A(x)\cdot (\nabla_x\otimes 1)\phi_n\rangle_{}\rvert\leq \sqrt{2(d-1)}\lVert \chi \rVert_2^{} \sqrt{n+1}\lVert (\lvert \nabla_x\rvert_{}^{}\otimes 1) \phi_n\rVert_{}^{}\sum_{\substack{i=-1\\i\neq 0}}^1\lVert \psi_{n+i} \rVert_{}^{}\; ;\\
\lvert \langle \psi , &A^2(x)\phi_n\rangle_{}\rvert\leq 2(d-1) \lVert \chi \rVert_2^{}(n+1)\lVert \phi_n \rVert_{}^{}\sum_{i=-2}^{2}\lVert \psi_{n+i} \rVert_{}^{}\; .
\end{aligned}
\end{equation}
Hence Assumption~\ref{ass:2} is satisfied. Then $H$ is essentially
self-adjoint on $D(H_0)\cap f_0(L^2 (\mathds{R}^{d})\otimes
(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))^{(\cdot)})$.
\begin{remark*}
\label{rem:1}
Neither non-negativity of the Pauli-Fierz operator nor smallness of
the coupling constant are necessary to prove essential
self-adjointness by means of Theorem~\ref{thm:1}. Using operator
methods (commutator estimates), self-adjointness of $H_{PF}$ with
$V=0$ has been proved for general coupling constants
in~\citep{MR2436496}, but the non-negativity was needed to
associate a unique self-adjoint operator to the quadratic
form. Theorem~\ref{thm:1} relies on different assumptions, and takes
advantage of the fibered structure of the Fock space: boundedness
from below of the operator is, in general, not necessary. In fact,
the Hamiltonians considered in Sections~\ref{sec:hamilt-many-bosons}
and~\ref{sec:nels-type-hamilt} are possibly unbounded from below, as
well as the following extension~\eqref{eq:31} of the Pauli-Fierz
Hamiltonian to infinite degrees of freedom (for the particles). As
outlined in Section~\ref{sec:conclusive-remarks-1}, if we assume
boundedness from below, Theorem~\ref{thm:1} can be extended to
operators quartic in the creation/annihilation operators (see
Assumptions~\ref{ass:5},~\ref{ass:6} and Theorem~\ref{thm:2}).
\end{remark*}
Let $m_j=1/2$, $q_j=-1$ and
$V=\sum_{i=1}^pV_{ext}(x_i)+\sum_{i<j}V_{pair}(x_i-x_j)$ such that
$V_{ext}\in L^2_{loc}(\mathds{R}^d,\mathds{R}_{+})$,
$V_{pair}(x)=V_{pair}(-x)$ and
$V_{pair}(1-\Delta)^{-1/2}\in\mathcal{L}(L^2 (\mathds{R}^d))$. Under
these assumptions define $H_{PF}\rvert_s$ as the restriction
of~\eqref{eq:27} to $L^2_s (\mathds{R}^{pd})\otimes
\Gamma_s(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))$. The physical
interpretation is a system of $p$ identical bosonic charges subjected
to an external potential, interacting via pair interaction and with
their radiation field. As we did for the Nelson model
in~\eqref{eq:25}, we can extend $H_{PF}\rvert_s$ to $\Gamma_s(L^2
(\mathds{R}^d))\otimes \Gamma_s(\mathds{C}^{d-1}\otimes L^2
(\mathds{R}^d))$:
\begin{equation}
\label{eq:31}
\begin{split}
\widetilde{H}_{PF}=\int_{\mathds{R}^d}^{}\psi^{*}(x)\Bigl\{\bigl(i\nabla_x\otimes 1 + A(x)\bigr)^2+V_{ext}(x) \Bigr\}\psi(x) dx + \frac{1}{2}\int_{\mathds{R}^{2d}}^{}V_{pair}(x-y)\psi^{*}(x)\psi^{*}(y)\\\psi(x)\psi(y) dxdy +1\otimes \sum_{\lambda=1}^{d-1}\int_{\mathds{R}^d}^{}\omega(k)a^{*}_{\lambda}(k)a_{\lambda}(k) dk\; .
\end{split}
\end{equation}
We would like to prove essential self-adjointness by means of
Theorem~\ref{thm:1}. Identify $H_{01}\equiv \int_{}^{}\psi^{*}(-\Delta
+V_{ext})\psi $; $H_{02}\equiv
\sum_{\lambda}\int_{\mathds{R}^d}^{}\omega
a^{*}_{\lambda}a_{\lambda}$; $H_I\equiv \int \psi^{*}(2i
A\cdot(\nabla\otimes 1)+
A^2)\psi+\frac{1}{2}\int_{}^{}V_{pair}\psi^{*}\psi^{*}\psi\psi $; and
$\{\mathscr{H}_{1}^{(j)}\otimes \Gamma_s(\mathscr{H}_2)
\}_{j\in\mathds{N}}\equiv \{L^2_s (\mathds{R}^{jd})\otimes
\Gamma_s(\mathds{C}^{d-1}\otimes L^2 (\mathds{R}^d))
\}_{j\in\mathds{N}}$. Then Assumptions~\ref{ass:1} and~\ref{ass:2} are
satisfied using bounds analogous to~\eqref{eq:33} and~\eqref{eq:35}
(for $V_{pair}$), for each fixed $j\in \mathds{N}$. Hence
$\widetilde{H}_{PF}$ is essentially self-adjoint on $D(H_{01}\otimes
1)\cap D(1\otimes H_{02})\cap f_0(L^2
(\mathds{R}^{d})^{(\cdot)}\otimes (\mathds{C}^{d-1}\otimes L^2
(\mathds{R}^d))^{(\cdot)})$.
\section{Conclusive remarks}
\label{sec:conclusive-remarks-1}
The examples of the preceding section are not exhaustive: we focused
on them because of their relevance in physical and mathematical
literature. The application to operators on curved space-time, or to
anti-symmetric systems may also lead to results of interest.
The Assumptions~\ref{ass:1}, \ref{ass:2} and \ref{ass:3} are easy to
check: in the examples above follow from basic estimates of creation
and annihilation operators. The proof of Theorem~\ref{thm:1} itself is
not complicated, and relies on the direct sum decomposition of
$\Gamma_s(\mathscr{H}_2)$ and the structure of the interaction with
respect to the latter. Hence this criterion gives, in our opinion, a
simple yet powerful tool to prove essential self-adjointness in Fock
spaces, tailored to take maximum advantage of their structure.
If we assume that $H$ is bounded from below, we can take inspiration
from \citet{masson1971} and extend our criterion to accommodate quartic
operators. The modified assumptions and theorem would then read:
\begin{assumption}{B$_{H}$}
\label{ass:5}
$H$ is a densely defined symmetric operator on
$\mathscr{H}=\mathscr{H}_1\otimes \Gamma_s(\mathscr{H}_2)$ bounded
from below. $H_{01}$ and $H_{02}$ are self-adjoint operators bounded
from below such that $\forall t\in \mathds{R}$,
$\{\mathscr{H}_2^{(n)}\}_{n\in\mathds{N}}$ is invariant for $e^{it
H_{02}}$.
\end{assumption}
\begin{assumption}{B$_I$}
\label{ass:6}
$H_I$ is a symmetric operator on $\mathscr{H}$, with a domain of
definition $D(H_I)$ such that $D(H_0)\cap D(H_I)$ is dense in
$\mathscr{H}$. Furthermore exists a complete collection
$\{\mathscr{H}_{1}^{(j)}\otimes \Gamma_s(\mathscr{H}_2)
\}_{j\in\mathds{N}}$ invariant for $H_0$ and $H_I$ such that:
$\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes H_{02})\cap
\mathscr{H}_1^{(j)}\otimes \mathscr{H}_2^{(n)}$,
\begin{equation}
\label{eq:34}
H_I\,\phi\in \bigoplus_{i=-4}^{4}\mathscr{H}_1^{(j)}\otimes \mathscr{H}_2^{(n+i)}\; .
\end{equation}
Also, $H_I$ satisfies the following bound: $\forall
j,n\in\mathds{N}$ $\exists C(j)>0$ such that $\forall\psi\in
\mathscr{H}$, $\forall\phi\in Q(H_{01}\otimes 1)\cap Q(1\otimes
H_{02})\cap\mathscr{H}_1^{(j)}\otimes \mathscr{H}_2^{(n)}$:
\begin{equation}
\label{eq:37}
\begin{split}
\lvert \langle \psi , H_I\phi \rangle_{\mathscr{H}}\rvert_{}^2\leq C^2(j)\sum_{i=-4}^4\lVert \psi_{j,n+i} \rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n+i)}}^2 \Bigl[(n+1)^4\lVert \phi \rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n)}}^{2}+(n+1)^{2}\Bigl(q_{H_{01}\otimes 1}(\phi,\phi)\\+q_{1\otimes H_{02}}(\phi,\phi)+(\lvert M_1\rvert_{}^{}+\lvert M_2\rvert_{}^{}+1)\lVert \phi \rVert_{\mathscr{H}_1^{(j)}\otimes\mathscr{H}_2^{(n)}}^{2} \Bigr)\Bigr]\; .
\end{split}
\end{equation}
\end{assumption}
\begin{thm}
\label{thm:2}
Assume \ref{ass:5} and \ref{ass:6}. Then $H$ is essentially self
adjoint on $D(H_{01}\otimes 1)\cap D(H_{02}\otimes 1)\cap
f_0(\mathscr{H}_1^{(\cdot)}\otimes \mathscr{H}_2^{(\cdot)})$.
\end{thm}
\begin{remark*}
An attempt to extend the results of \citep{masson1971} can be
found in \citep{arai1991}. Theorem~\ref{thm:2} is a generalization
of both: it can be applied to more singular situations and a more
general class of spaces.
\end{remark*}
The proof of Theorem~\ref{thm:1} can be adapted to
Theorem~\ref{thm:2}, making use of the inferior bound for $H$. We
remark that Assumption~\ref{ass:5}, by itself, implies that $H$ has at
least one self-adjoint extension: it may be tricky to prove for
general operators. Theorem~\ref{thm:2} essentially states that for
regular enough quartic interactions, existence of a particular
self-adjoint extension (the Friedrichs one) is equivalent to its
uniqueness. It may have interesting applications in CQFT: e.g. the
$d$-dimensional (bounded from below) $Y_d$ and
$(\lambda\varphi(x)^4)_d$ models with cut offs have interactions that
are at most quartic and regular. It is our hope that the ideas
utilized in this paper could contribute to improve the mathematical
insight on interacting quantum field theories, and could be developed
to study self-adjointness of more singular systems.
\begin{acknowledgments}
This work has been supported by the Centre Henri Lebesgue (programme
``Investissements d'avenir'' --- ANR-11-LABX-0020-01). The author
would like to thank Giorgio Velo, that has suggested to him the idea
of a direct proof of self-adjointness on Fock spaces. Also, he would
like to thank Zied Ammari and Francis Nier for precious advices and
stimulating discussions during the redaction of the paper.
\end{acknowledgments}
|
\section{Introduction}
Cohomological field theories (CohFT for brevity) were introduced in the early 90s by Kontsevich and Manin in \cite{KM}. They appeared to play an important role in many different subjects of mathematics - they are key objects in the mirror symmetry conjecture \cite{FJR}, integrable hierarchies \cite{DZ, FSZ} and even geometry of the moduli space of curves \cite{PPZ}. An important tool to work with CohFTs that is used in all the aspects listed is Givental's action. However in some cases (and in singularity theory in particular) it does not give the feeling of the initial object geometry where another - $ {\rm SL(2,\mathbb C)} $ action is defined naturally. Based on the examples coming from singularity theory we introduce analytically the $ {\rm SL(2,\mathbb C)} $ action on the genus zero part of an arbitrary CohFT and write it in terms of Givental's action to act on the higher genera of the CohFT too.
\subsection{Cohomological Field Theory}
Denote by $\M{g}{k}$ the moduli space of stable genus $g$ curves with $k$ marked points. Let $V = \langle e_1, \dots, e_n \rangle$ be a finite dimensional vector space with a non--degenerate scalar product $\eta$. A Cohomological field theory on the state space $(V, \eta)$ is a system of linear maps
$$
\Lambda_{g,k}: V^{\otimes k} \rightarrow H^*(\M{g}{k}, \mathbb C)
$$
for all $g,k$ such that $\M{g}{k}$ exists and is non--empty. It is required to satisfy certain axioms that arise naturally from the geometry of the moduli space of curves.
Let $\psi_l \in H^*(\M{g}{k})$, $1 \le l \le k$ be Mumford--Morita--Miller classes. The genus $g$ correlators of the CohFT are the following numbers:
$$
\langle \tau_{d_1}(e_{\alpha_1}) \dots \tau_{d_k}(e_{\alpha_k}) \rangle_g := \int_{\M{g}{k}} \Lambda_{g,k}(e_{\alpha_1} \otimes \dots \otimes e_{\alpha_k}) \psi_1^{d_1} \dots \psi_k^{d_k}.
$$
Denote by ${\mathcal F}_g$ the generating function of the genus $g$ correlators, called genus $g$ potential of the CohFT:
$$
{\mathcal F}_g := \sum \frac{\langle \tau_{d_1}(e_{\alpha_1}) \dots \tau_{d_k}(e_{\alpha_k}) \rangle_g}{\mathrm{Aut}( \boldsymbol \alpha, \bf d)} \ t^{d_1,\alpha_1} \dots t^{d_k,\alpha_k}.
$$
It depends on the formal variables $t^{k, \alpha}$ for $1 \le \alpha \le n = \dim(V)$ and $k \ge 0$.
The restriction of ${\mathcal F}_g$ to $t^{k, \alpha} = 0, \ k \ge 1$ is called restriction to the small phase space.
Introduce the notation:
$$
F_g := {\mathcal F}_g \mid_{t^{k, \alpha} = 0, \ k \ge 1}.
$$
We have $F_g = F_g(t^{0,1}, \dots, t^{0,n})$. These generating functions will be called genus $g$ small phase space potentials.
It is useful to assemble the correlators into a generating function called partition function of the CohFT:
$$
\mathcal{Z} := \exp \left( \sum_{g \ge 0} \hbar^{g-1} \mathcal F_g \right).
$$
\subsection{Frobenius manifold of a CohFT}
The notion of Frobenius manifold was introduced by B. Dubrovin in 90s. It provides a generalization of the flat structures of K. Saito introduced in the early 80s and is a crucial step towards the Saito--Givental CohFT of a singularity.
The structure of a Frobenius manifold $M$ is defined by the so-called Frobenius potential $F^M \in \mathbb C[[t^1, \dots, t^n]]$ that is subject to the non-linear system of PDE's called WDVV equation.
Our interest in it comes from the fact that every unital CohFT $\Lambda_{g,k}$ defines a (formal) Frobenius manifold. Namely the generating function $F_0$ restricted to the small phase space is a solution to the WDVV equation and considered as Frobenius potential it defines a certain Frobenius structure:
$$
F^M(t^1,\dots,t^n) := {\mathcal F}_{g = 0} \mid_{t^{k,\alpha} = 0, \ k \ge 1} = F_0,
$$
where we associate $t^\alpha$ on the LHS with the $t^{0,\alpha}$ on the RHS.
The formalism of Dubrovin allows us to consider the action of $A \in {\rm SL(2,\mathbb C)} $ analytically on the Frobenius manifold of a CohFT and to extend it to the higher genera via the technique of Givental.
\begin{example}[Appendix~C in \cite{D}]\label{example: chazy}
Consider the potential of the 3-dimensional Frobenius manifold:
$$
F(\textbf t) = \frac{1}{2} (t^1)^2t^3 + \frac{1}{2} t^1 (t^2)^2 - \frac{(t^2)^4}{16} \gamma(t^3).
$$
The WDVV equation on $F$ is equivalent to the Chazy equation on $\gamma$:
$$
\gamma^{\prime \prime \prime} = 6 \gamma \gamma^{\prime\prime} - 9 (\gamma^\prime)^2.
$$
It is a well-known property of the Chazy equation that for $A \in \rm SL(2, \mathbb C)$ the function $\gamma^A$ solves the Chazy equation too:
$$
\gamma^A(t) := \gamma \left(\frac{at + b}{ct + d}\right) - \frac{c}{2(ct + d)}.
$$
As the consequence of this property of the Chazy equation we get an $\rm SL(2, \mathbb C)$ action on the space of 3-dimensional Frobenius manifolds.
\end{example}
\subsection{Definition of the $\rm SL(2, \mathbb C)$ action}
Consider the action of $A \in \rm SL(2, \mathbb C)$ on the variable $t^{0,n}$:
$$
A \cdot t^{0,n} = \frac{at^{0,n} + b }{c t^{0,n} + d},
\quad
A = \left(
\begin{array}{c c}
a & b \\
c & d
\end{array}
\right).
$$
We would like to ``quantize'' this action on the variable $t^{0,n}$ to the action $\hat A$ on the CohFT partition function. We require it to satisfy the following conditions:
\begin{itemize}\label{quantization conditions}
\item $\hat A$ acts on the space of partition functions. Namely $\hat A \cdot \mathcal Z$ is a partition function of some CohFT.
\item By the action $\hat A$ the variable $t^{0,n}$ is transformed by $t^{0,n} \rightarrow A \cdot t^{0,n}$.
\end{itemize}
Note that the first requirement of the quantization rules means also that $\hat A$ should act on the space of WDVV solutions too. Namely $(\hat A \cdot \mathcal Z)_{g=0}$ should solve WDVV.
The potential of a Frobenius manifold can be written in coordinates as:
\begin{equation}\label{eq:WDVV potential}
F(\textbf{t}) = F(t^1, \dots, t^n) = \frac{1}{2} (t^1)^2 t^n + \frac{1}{2} t^1 \sum_{i=2}^{n-1} t^i t^{n+1-i} + H(t^2, \dots, t^n),
\end{equation}
for some function $H$ depending on $t^2,\dots,t^n$ only.
\begin{definition}
Let $F$ be a Frobenius potential written as in \eqref{eq:WDVV potential}. For $A \in {\rm SL}(2, \mathbb C)$ define the function:
\begin{equation}\label{eq:SLAction}
\begin{aligned}
F^A(\textbf{t}) := \frac{1}{2} (t^1)^2t^n &+ \frac{1}{2}t^1 \sum_{i =2}^{n-1} t^i t^{n+1-i} + \frac{c}{8(ct^n + d)} (t^2 t^{n-1} + \dots + t^{n-1}t^2)^2
\\
&+ (ct^n+d)^2 H \left(\frac{t^2}{ct^n+d}, \dots, \frac{t^{n-1}}{ct^n+d}, \frac{at^n + b}{ct^n+d} \right).
\end{aligned}
\end{equation}
\end{definition}
In what follows we shows that $\hat A \cdot F := F^A$ can be considered as the quantization of $ {\rm SL(2,\mathbb C)} $ action in genus zero.
\subsection{Givental's action} A. Givental introduced in \cite{G} two group actions on the space of partition functions of the CohFT's. These are now known as the actions of the \textit{lower-triangular} and \textit{upper-triangular} Givental groups or $S$- and $R$-actions respectively.
Let $\Lambda_{g,k}$ be a CohFT on $(V,\eta)$. Consider
$$
r(z) \in {\rm Hom}(V,V) \otimes \mathbb C[z],
\quad
s(z) \in {\rm Hom}(V,V) \otimes \mathbb C[z^{-1}],
$$
such that $r(z) + r(-z)^* = 0$ and $s(z) + s(-z)^* = 0$ (where the star means dual w.r.t. $\eta$).
Following Givental define:
$$
\hat R := \exp( \widehat {r(z)} ), \ \hat S := \exp( \widehat {s(z)} ),
$$
where $\widehat {r(z)}$ and $\widehat {s(z)}$ are certain differential operators.
The importance of Givental's action is given by the following proposition.
\begin{proposition}
Differential operators $\hat R$ and $\hat S$ obtained by the quantization of the Givental group elements $R$ and $S$ act on the space of partition functions of the Cohomological Field Theories.
\end{proposition}
We can also consider the action of the Givental's group element $R$ on the Frobenius manifold $M$ with the potential $F(\textbf{t})$. Doing this we act on the CohFT partition function $\mathcal Z$ such that $F = [\hbar^{-1}] \mathcal{Z}$ and take again $\hat R \cdot F := [\hbar^{-1}] \left( \hat R \cdot \mathcal{Z} \right)$. The function $\hat R \cdot F$ will be potential of some Frobenius manifold that we will denote by $\hat R \cdot M$.
\subsection{$ {\rm SL(2,\mathbb C)} $ group action on the genus $g$ potential of the CohFT}
Let $F(\textbf{t})=F(t^1, \dots, t^n)$ be the potential of a Frobenius manifold $M$. In what follows for $\textbf{s} \in M$ we denote by $M \mid_{\textbf{t} = \textbf{s}}$ a neighborhood in $M$ of the point $\textbf{s}$. The structure of the Frobenius manifold $M$ at the point $\textbf{s}$ is given by the expansion of $F(\textbf{t})$ at $\textbf{t} = \textbf{s}$.
The first theorem of this paper states:
\begin{theorem}\label{th:FrobeniusManifold}
Let
$
A = \left(
\begin{array}{c c}
a & b \\
c & d
\end{array}
\right)
\in {\rm SL}(2, \mathbb C).
$
\begin{itemize}
\item[a.] The function $F^A(\textbf{t} )$ defined in \eqref{eq:SLAction} is a solution to the WDVV equation,
\item[b.] Let $F$ be analytic at $\textbf{t} = (0,\dots,0, A \cdot \tau) \in M$ for some $\tau$ and $M^A$ be the Frobenius manifold defined by $F^A$. Then we have an isomorphism:
$$
M^A \mid_{\textbf{t} = (0,\dots,0,\tau)} \ \cong \ \hat R^\sigma \cdot M \mid_{\textbf{t} = (0,\dots, 0, A\cdot \tau)}
$$
where
$$
R^\sigma(z) :=
\exp( \left(
\begin{array}{c c c}
0 & \dots & \sigma \\
\vdots & 0 & \vdots \\
0 & \dots & 0
\end{array}
\right) z), \quad \sigma := -c(c\tau+d).
$$
\end{itemize}
\end{theorem}
\begin{proof}
The statements of Propositions \ref{prop:Composition}, \ref{prop:A_g to A} and \ref{prop:Action via graphs} sum up to the proof of the theorem.
\end{proof}
This theorem sets up the bridge between the analytical action of $ {\rm SL(2,\mathbb C)} $ on the genus $0$ part of a CohFT and the particular Givental's action on a CohFT partition function. Let $F_g(\textbf{t}) = F_g(t^1,\dots,t^n)$ be the genus $g$ small phase space potentials of a CohFT. For $g \ge 1$ consider:
\begin{equation}\label{equation: FgA definition}
F_g^A(\textbf{t}) := (ct^n+d)^{2-2g} F_g \left(\frac{t^1}{ct^n+d}, \dots, \frac{t^{n-1}}{ct^n+d}, \frac{at^n + b}{ct^n+d} \right) - \delta_{1,g} \log \left( \frac{ct^n + d}{c\tau +d} \right),
\end{equation}
The following theorem extends the $ {\rm SL(2,\mathbb C)} $ action to the higher genera.
\begin{theorem}\label{th:HigherGenera}
Let $F_g(\textbf{t})$ be analytic at $p_1 := (0,\dots,0,A \cdot \tau)$. Let $\mathcal{Z}_{p_1}$ and $(F_g^A)_{p_2}$ be expansions of $\exp(\sum_{g \ge 0} \hbar^{g-1} F_g)$ and $F_g^A$ at the points $p_1$ and $p_2 := (0,\dots,0,\tau)$ respectively. Then
$$
[\hbar]^{g-1} \log \left( \hat R^\sigma \cdot \mathcal{Z}_{p_1} \right)(\tilde \textbf{t}) = (c\tau+d)^{2-2g} \ \left( F_g^A (\textbf{t}) \right)_{p_2}
$$
where
$
\tilde \textbf{t} = \left(t^1, (c\tau+d) \ t^2, \dots, (c\tau+d) \ t^{n-1}, (c\tau+d)^2 \ t^n \right).
$
\end{theorem}
This theorem is proved in subsection~\ref{subsection: proofOfTheorem2}.
\subsection{$ {\rm SL(2,\mathbb C)} $ group action on the total ancestor potential of the singularity}
In singularity theory the structure of a Frobenius manifold appears naturally on the base space of a singularity unfolding. It was defined in the early 80s by Kyoji Saito \cite{S}. Given a singularity $W(\textbf{x})$ one of the main objects of Saito's theory needed to build up the Frobenius manifold is the so--called primitive form of $W(\textbf{x})$. The choice of a primitive form for the fixed singularity is generally not unique.
Given a primitive form $\zeta$ of the singularity $W(\textbf{x})$ one can construct much more than just a Frobenius manifold. Namely one can construct a particular CohFT, called nowadays Saito--Givental's CohFT. The partition function of this CohFT is called the total ancestor potential of the singularity $W(\textbf{x})$ and is denoted by $\mathcal A_{\zeta,\textbf{s}} (\hbar,\textbf{t})$.
For the simple elliptic singularities using certain Givental's action T.Milanov and Y.Ruan gave in \cite{MR} the formula connecting $\mathcal{A}_{\zeta_1,\textbf{s}}$ and $\mathcal{A}_{\zeta_2,\textbf{s}^\prime}$ with a two different primitive forms $\zeta_1$,$\zeta_2$ of the same singularity $W(\textbf{x})$.
In \cite{BT} the authors proposed certain $ {\rm SL(2,\mathbb C)} $ action in the form of \eqref{eq:SLAction} to write down explicitly the effect of the primitive form change on the Frobenius manifold potential. However this action is written in a completely different form comparing to the formula of Milanov--Ruan and was not extended to the total ancestor potential. Using Theorem \ref{th:FrobeniusManifold} we show that the two approaches agree.
\begin{theorem*}[Theorem~\ref{theorem: MR-BT equivalence}]\label{th:theorem3}
Let $F$ be the Frobenius manifold potential of a simple elliptic singularity $W(\textbf{x})$. Then the formula of Milanov--Ruan connecting the total ancestor potentials of $W(\textbf{x})$ with two different primitive forms is equivalent to the $ {\rm SL(2,\mathbb C)} $ action on $F$ given by formula \eqref{eq:SLAction}.
\end{theorem*}
This theorem allows one to use the quasi--modularity of the total ancestor potential of a simple--elliptic singularity and certain mirror symmetry theorem to compute explicitly the genus zero Gromov--Witten potentials of the orbifolds ${\mathbb{P}}^1_{3,3,3}$, ${\mathbb{P}}^1_{4,4,2}$ and ${\mathbb{P}}^1_{6,3,2}$. The first one was already computed ba Sakate and Takahashi in \cite{ST}, while the explicit expression for the other two is new.
\subsection{Organization of the paper}
In section~\ref{section: AnalyticalQuantization} we give an analytical approach to the $ {\rm SL(2,\mathbb C)} $ action on the genus $0$ part of the CohFT.
We recall basic facts about Givental's action in section~\ref{section: GiventalsAction}.
In section~\ref{section: GiventalsQuantization} we define an analog of the $ {\rm SL(2,\mathbb C)} $ action that is later approved by means of graphical calculus to coincide with tha analytical approach and also with the action of $\hat R^\sigma$ as above.
We compute $\hat R^\sigma$ action on a CohFT higher genera data and prove Theorem~\ref{th:HigherGenera} in section~\ref{section: HigherGenus}.
In section~\ref{section: coordinateFree} we write the $ {\rm SL(2,\mathbb C)} $ action in terms of Givental's group action only.
We discuss singularity theory applications in section~\ref{section: totalAncestorPotential} giving the proof of Theorem~\ref{theorem: MR-BT equivalence}. Using this singularity theory result we compute the genus zero Gromov--Witten potentials of the orbifolds ${\mathbb{P}}^1_{4,4,2}$ and ${\mathbb{P}}^1_{6,3,2}$.
\subsection{Acknowledgement}
The author is grateful to Sergey Shadrin for his help with Givental's action, Claus Hertling for many useful comments and to Davide Veniani for editorial help.
\section{Analytical quantization of $ {\rm SL(2,\mathbb C)} $ action}\label{section: AnalyticalQuantization}
In this section we develop an analytical approach to the quantization of the $ {\rm SL(2,\mathbb C)} $ action on the CohFT via the genus zero data of it --- Frobenius manifold.
\subsection{Frobenius manifolds}
Let $M$ be a domain in ${\mathbb{C}}^n$. Assume its tangent space ${\mathcal T}_M$ to be endowed with the constant non-degenerate bilinear form $\eta$,
$$
\eta: {\mathcal T}_M \times {\mathcal T}_M \to {\mathbb{C}}.
$$
Let $t^1, \dots, t^n$ be coordinates on $M$. We associate the basis of ${\mathcal T}_M$ with the vectors $\partial /\partial t^i$ and consider $\eta_{pq}$ as components of $\eta$ in this basis.
Consider the function $F(\textbf{t}) = F(t^1, \dots, t^n)$ on $M$, represented by a convergent power series in $t^1, \dots, t^n$.
The function $F(\textbf{t})$ is said to satisfy WDVV equation if for every fixed $1 \le i,j,k,l \le n$ holds:
$$
\sum_{p,q} \frac{\partial^3 F }{\partial t^i \partial t^j \partial t^p} \ \eta^{pq} \ \frac{\partial^3 F}{\partial t^q \partial t^k \partial t^l}
=
\sum_{p,q} \frac{\partial^3 F}{\partial t^i \partial t^k \partial t^p} \ \eta^{pq} \ \frac{\partial^3 F}{\partial t^q \partial t^j \partial t^l},
$$
where $\eta^{pq} = (\eta^{-1})^{p,q}$.
Using the function $F$ define an algebra structure on ${\mathcal T}_M$. Let $c_{ij}^k(\textbf{t}) $ be the structure constants of the multiplication $\circ: {\mathcal T}_M \times {\mathcal T}_M \to {\mathcal T}_M$ defined by $ c_{ij}^k(\textbf{t}):= \sum_p c_{ijp}(\textbf{t}) \eta^{pk}$, where
$$
c_{ijk}(\textbf{t}) := \frac{\partial ^3 F}{\partial t^i \partial t^j \partial t^k}, \quad 1 \le i,j,k \le n,
$$
and $\eta^{ij} := \sum_{p,q} \eta_{pq} \delta^{pi} \delta^{qj}$.
The structure constants $c_{ij}^k(\textbf{t})$ define a commutative algebra structure by the construction, while the associativity is equivalent to the WDVV equation on $F(\textbf{t})$.
Assume in addition that $F(\textbf{t})$ is such that $\partial /\partial t^1$ is the unit of the algebra. Therefore we have:
$$
\eta_{ij} = \frac{\partial ^3 {\mathcal F}}{\partial t^1 \partial t^i \partial t^j}.
$$
Then $\eta_{ij}$ together with $c_{ij}^k(\textbf{t})$ define the {\it Frobenius algebra } structure:
$$
\eta \left(\frac{\partial }{\partial t^i} \circ \frac{\partial }{\partial t^j}, \frac{\partial }{\partial t^k} \right) = \eta \left( \frac{\partial }{\partial t^i}, \frac{\partial }{\partial t^j} \circ \frac{\partial }{\partial t^k} \right).
$$
Note that different points of the Frobenius manifold $M$ give generically different potentials $F$.
\begin{definition}
The data $\eta$ and $\circ$ satisfying conditions as above define a Frobenius manifold structure on $M$. The function $F$ is called {\it Frobenius potential} of $M$ and the coordinates $\textbf{t}$ -- {\it flat coordinates}.
\end{definition}
\begin{remark}
Usually the potential of a Frobenius manifold is assumed to satisfy certain quasi--homogeneity condition w.r.t. a specially chosen vector field called \textit{Euler vector field}. We do not make such an assumption here, working with a larger class of the Frobenius manifolds.
\end{remark}
Sometimes we are given first the function $F$ satisfying WDVV equation without any underlying space $M$ and holomorphicity property. In these ocasions $F$ could anyway define a Frobenius manifold structure that is called sometimes \textit{formal}. We will drop this word assuming it to be clear from the context.
\begin{definition}
Two Frobenius manifold $M_1$ and $M_2$ are called (locally) isomorphic if there is a diffeomorphism $\phi: M_1 \to M_2$ such that for some fixed $\textbf{t} \in M_1$ and $\phi(\textbf{t}) \in M_2$ holds:
\begin{itemize}
\item $\phi$ is linear conformal transformation of the metrics of $M_1$ and $M_2$,
\item the differential of $\phi$ is an isomorphism of the algebras $T_{\textbf{t}}M_1$ and $T_{\phi(\textbf{t})}M_2$.
\end{itemize}
In this case we write $M_1 \mid_\textbf{t} \ \cong \ M_2 \mid_{\phi(\textbf{t})}$.
\end{definition}
In what follows we assume the scalar product of the Frobenius manifold to have a particular form: $\eta_{i,j} = \delta_{i+j,n+1}$. Note however that our results are easily translated to the arbitrary choice of $\eta$. Choosing appropriate coordinates $F$ can be written as follows:
$$
F(\textbf{t}) = F(t^1, \dots, t^n) = \frac{1}{2} (t^1)^2 t^n + \frac{1}{2} t^1 \sum_{i=2}^{n-1} t^i t^{n+1-i} + H(t^2, \dots, t^n),
$$
for some function $H(t^2, \dots, t^n)$ that does not depend on $t^1$.
Motivated by this presentation of the potential we introduce the notation for the summands of $F^A$ defined in \eqref{eq:SLAction}.
\begin{notation}
Consider the formula \eqref{eq:SLAction}. We adopt the notation:
$$
\text{$H$--term of $F^A$} \quad := \quad (ct^n+d)^2 H \left(\frac{t^2}{ct^n+d}, \dots, \frac{t^{n-1}}{ct^n+d}, \frac{at^n + b}{ct^n+d} \right)
$$
and
$$
\text{$Q$--term of $F^A$} \quad := \quad \frac{c}{8(ct^n + d)} (\sum_{k \neq 1} t^k t^{n+1 -k})^2.
$$
Hence $F^A =$ H--term $+$ Q--term $+$ cubic terms.
\end{notation}
It was noted by Dubrovin (cf. Appendix~B in \cite{D}) that there is a non-trivial symmetry of the WDVV equation such that for $F(\textbf{t})$ as above the function $F^I(\hat \textbf{t})$ solves WDVV too:
\begin{equation}\label{equation:Inversion}
F^I (\hat \textbf{t} ) = (t^n)^{-2} \left[ F(\textbf{t}) - \frac{1}{2} t^1 \sum_{i=0}^n t^i t^{n+1-i} \right],
\end{equation}
where for $1 < \alpha < n$:
$$
\hat t^1 := \frac{1}{2} \frac{\sum_i {t^i t^{n+1-i}}}{t^n}, \ \hat t^\alpha := \frac{t^\alpha}{t^n}, \ \hat t^n := -\frac{1}{t^n}.
$$
We will call it ``Inversion transformation'' and write:
$$
F^I ( \hat \textbf{t}) = \hat I \cdot F(\textbf{t}).
$$
\subsection{Analytical quantization via composition}
Consider $A = \begin{pmatrix}
a & b \\ c & d
\end{pmatrix} \in {\rm SL}(2, \mathbb C)$.
The action of $A$ on the variable $t^n$ can be decomposed as follows.
$$
A \cdot t^n = \frac{a t^n + b}{c t^n + d} = T_2 \cdot S_{c^{2}} \cdot I \cdot T_1 \cdot t^n,
$$
where we have:
\begin{itemize}
\item $T_1$ is one more shift $t^n \rightarrow t^n + \frac{a}{c}$,
\item $S_{c^{2}}$ is the scaling $t^n \rightarrow c^{2} t^n $,
\item $T_2$ is the shift $t^n \rightarrow t^n + \frac{d}{c}$,
\item $I: t^n \rightarrow -1 / t^n$.
\end{itemize}
The shifts $T_i$ can be easily quantized to the operators on the space of the WDVV equation solutions because linear change of variables preserves the WDVV equation. More complicated is the quantization of the scaling $S_{c^2}$. Let $F$ be a Frobenius manifold potential. The operator $\hat S_c$ acts in the following way.
\begin{equation}\label{equation: S0 action}
\hat S_c \cdot F(\textbf{t}) = \frac{1}{2}(t^1)^2 t^n + t^n \sum_{p=2}^{n-1} t^p t^{n+1-p} + c H(t^2, \dots, t^{n-1}, c t^n).
\end{equation}
Finally, the action of $I$ is quantized by the Inversion transformation $\hat I$.
\begin{definition}Define the action $\hat A$ on the space of WDVV solutions:
$$
\hat A = \hat T_2 \cdot \hat S_{c} \cdot \hat I \cdot \hat S_{c^{-1}} \cdot \hat T_1.
$$
\end{definition}
\begin{proposition}\label{prop:Composition} For the $\hat A$ as above we have:
\begin{itemize}
\item The action $\hat A$ satisfies the quantization condition in genus $0$.
\item The action $\hat A$ agrees with the formula \eqref{eq:SLAction} up to quadratic terms:
$$
\hat A \cdot F = F^A + \text{quadratic terms}.
$$
\end{itemize}
\end{proposition}
\begin{proof}
The quantization presented coincides with action of $A$ on $t^n$ by the construction and also $\hat A$ acts on the space of WDVV solution as the composition of operators acting of the space of WDVV solutions.
For the second part note that applying $\hat S_{c} \cdot \hat I \cdot \hat S_{c^{-1}}$ to $H(t^2, \dots, t^n)$ both $S_c$ and $S_{c^{-1}}$ add their factors that cancel out. Hence for the $H$-terms the proposition follows. For the $Q$-terms we have:
$$
Q-\text{terms of } \hat S_{c} \cdot \hat I \cdot \hat S_{c^{-1}} \cdot \hat T_1 = \frac{c \left( \sum_{p \neq 1, n} t^p t^{n-p}\right)^2}{c t_n},
$$
where the factor $c$ comes twice because of the quantization rule of $\hat S_c$. In the denominator because $t^n$ is multiplied by $c$ and in the nominator because the function $H$ in the presentation \eqref{eq:WDVV potential} is multiplied by $c$. At the same time by the definition of $\hat I$ the action of $\hat S_{c^{-1}} \cdot \hat T_1$ does not affect these terms.
Applying the shift $t^n \rightarrow t^n + \frac{d}{c}$ we get exactly the $Q$-term of $F^A$.
\end{proof}
This proves $a$-part of Theorem \ref{th:FrobeniusManifold}.
\subsection{Example: Gromov--Witten theory of ${\mathbb{P}}^1_{2,2,2,2}$}
Consider the so--called theta--constants $\vartheta_2(\tau),\vartheta_3(\tau),\vartheta_4(\tau)$ that are the values at $z = 0$ of the Jacobi theta functions $\vartheta_k(z,\tau)$:
\begin{equation*}
\begin{aligned}
\vartheta_2(\tau) = \sum_{n = -\infty}^\infty & q^{(n-1/2)^2 }, \quad
\vartheta_3(\tau) = \sum_{n = -\infty}^\infty q^{n^2 } ,
\\
\vartheta_4(\tau) &= \sum_{n = -\infty}^\infty (-1)^n q^{n^2 }, \quad q := \exp(\pi \sqrt{-1} \tau).
\end{aligned}
\end{equation*}
Let $X_k^\infty(\tau)$ and $X_k^\infty(q)$ be the logarithmic derivatives of the theta--constants:
$$
X_k^\infty(\tau) := 2 \frac{\partial}{\partial \tau} \log \vartheta_k(\tau), \quad X_k^\infty(q) := \frac{1}{\pi \sqrt{-1}} X_k^\infty \left( \frac{\tau}{\pi \sqrt{-1}} \right).
$$
The functions $\vartheta_k(\tau)$ and $X_k^\infty(\tau)$ are examples of the modular and quasi--modular forms.
\begin{definition}
Let $\Gamma \subset {\rm SL}(2,{\mathbb{Z}})$ be a finite index subgroup. Let $k \in {\mathbb{N}}_{\ge 0}$ and $f(\tau)$ --- a holomorphic function on ${\mathbb{H}}$.
\begin{itemize}
\item $f(\tau)$ is called a {\it modular form of weight $k$} if it satisfies the following condition:
\begin{equation*}
\frac{1}{(c\tau+d)^k}f\left(\frac{a\tau+b}{c\tau+d}\right) = f(\tau) \text{ for any }
\begin{pmatrix}
a &b\\
c & d
\end{pmatrix}
\in \Gamma.
\end{equation*}
\item $f(\tau)$ is called a {\it quasi-modular form of weight $k$ and depth $m$} if there are functions $f_0(\tau), \dots, f_m(\tau)$, holomorphic in ${\mathbb{H}}$ s.t.:
\begin{equation*}
\frac{1}{(c\tau+d)^k}f\left(\frac{a\tau+b}{c\tau+d}\right) = \sum_{l = 0}^m f_k(\tau) \left( \frac{c}{c \tau + d}\right)^k \text{ for any }
\begin{pmatrix}
a &b\\
c & d
\end{pmatrix}
\in \Gamma.
\end{equation*}
\end{itemize}
\end{definition}
It's clear that every modular form is quasi--modular too with all functions $f_1, \dots, f_m$ identically zero.
\begin{proposition}[cf. \cite{Z}]
The functions $(\vartheta_2(\tau))^2$, $(\vartheta_3(\tau))^2$, $(\vartheta_4(\tau))^2$ are modular forms of weight $1$ on $\Gamma(2) = \{A \in \mathrm{SL}(2,{\mathbb{Z}}) \ | \ A \equiv 1 \text{ mod } 2\}$.
\end{proposition}
Frobenius manifold potential of the Gromov--Witten theory of the orbifold ${\mathbb{P}}^1_{2,2,2,2}$ was found in \cite{ST} by I.Satake and A.Takahashi. It reads:
\begin{align*}
F^{{\mathbb{P}}^1_{2,2,2,2}}_0 & (t_0,t_1,t_2,t_3,t_4,\tau) = \frac{1}{2} t_0^2 \tau + \frac{1}{4}t_0 \left( \sum_{i=1}^4 t_i^2 \right) + \frac{1}{8} (t_1t_2t_3t_4) \left( X_3^\infty(q) - X_4^\infty(q) \right) \\
& - \frac{1}{192} \left(\sum_{i=1}^4 t_i^4 \right) \left( 4X_2^\infty(q) + X_3^\infty(q) + X_4^\infty(q) \right) - \frac{1}{32} \left( \sum_{i<j} t_i^2t_j^2 \right) \left( X_3^\infty(q) + X_4^\infty(q)\right).
\end{align*}
The modularity of $(\vartheta_k)^2$ allows us to consider the action of $\Gamma(2)$ on the coefficients of the function $F_0^{{\mathbb{P}}^1_{2,2,2,2}}$. The (a)--part of Theorem~\ref{th:FrobeniusManifold} extends this action to the action on the Frobenius manifold potential.
\begin{proposition}
Let $A \in \Gamma(2)$, consider the action of it by formula~\eqref{eq:SLAction}. Then we have:
$$
\left( F_0^{{\mathbb{P}}^1_{2,2,2,2}} \right)^A = F_0^{{\mathbb{P}}^1_{2,2,2,2}}.
$$
\end{proposition}
\begin{proof}
It's easy to see from the proposition above that the functions $X_k^\infty(\tau)$ are quasi--modular of weight $2$. Hence the coefficient $[t_1t_2t_3t_4] F_0^{{\mathbb{P}}^1_{2,2,2,2}}$, considered as the function in $\tau$ is modular of weight $2$ on $\Gamma(2)$ and the coefficients $[t_k^4] F_0^{{\mathbb{P}}^1_{2,2,2,2}}$, $[t_k^2t_l^2] F_0^{{\mathbb{P}}^1_{2,2,2,2}}$ are quasi--modular of weight $2$ on $\Gamma(2)$.
Some easy computations show that by the transformation $\tau \to A \cdot \tau$ this quasi--modularity gives exactly additional summands annihilating the $Q$--terms of the function $\left( F_0^{{\mathbb{P}}^1_{2,2,2,2}} \right)^A$.
\end{proof}
It's important to note that taking $A \in \mathrm{SL}(2,{\mathbb{Z}})$ such that $A \not\in \Gamma(2)$ the genus zero potential of ${\mathbb{P}}^1_{2,2,2,2}$ is transformed differently. For example the following formulae are straightforward from the definition of $X_k^\infty(\tau)$:
\begin{align*}
X_2^\infty(\tau+1) = X_2^\infty(\tau), \ X_3^\infty(\tau+1) = X_4(\tau), \ X_4^\infty(\tau+1) = X_3^\infty(\tau),
\\
X_2^\infty \left(-\frac{1}{\tau}\right) = \tau^2 X_4^\infty(\tau) + \tau, \ X_3^\infty\left(-\frac{1}{\tau}\right) = \tau^2 X_3(\tau) + \tau, \ X_4^\infty\left(-\frac{1}{\tau}\right) = \tau^2 X_2^\infty(\tau) + \tau.
\end{align*}
Hence taking $A = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix}$ or $A = \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}$ we get $ \left( F_0^{{\mathbb{P}}^1_{2,2,2,2}}(\textbf{t}) \right)^A = F_0^{{\mathbb{P}}^1_{2,2,2,2}}(\tilde \textbf{t})$ for $\tilde \textbf{t}$ differing from $\textbf{t}$ by a permutation of the variables $t_1,\dots,t_4$. Hence the Frobenius manifold obtained by such an action is still isomorphic to the initial one. We can not expect such a behaviour for an arbitrary $A \in {\rm SL(2,\mathbb C)} $.
\subsection{Example: Hurwitz--Frobenius manifolds}
Consider the space of meromorphic functions $\lambda: C \to \mathbb P^1$ on the compact genus $g$ Riemann surface $C$. Fix the pole orders of $\lambda$ to be $\textbf k := \{k_1, \dots, k_m\}$:
$$
\lambda^{-1}(\infty) = \{ \infty_1, \dots, \infty_m \}, \quad \infty_p \in C,
$$
so that locally at $\infty_p$ we have $\lambda(z) = z^{k_p}$.
Such meromorphic functions define the ramified coverings of $\mathbb P^1$ by $C$ with the ramification profile $\textbf k$ over $\infty$. Assume further that $\lambda$ has only simple ramification points at $P_i \in \mathbb P^1 \backslash \{0\}$. On the space of the pairs $(C, \lambda)$, considered up to a certain equivalence, B. Dubrovin introduced in \cite[Lecture 6]{D} a Frobenius manifold structure that is now known under the name \textit{Hurwitz--Frobenius} manifold and is denoted by $\mathcal H_{g; \textbf{k}}$.
When $g=1$ the ramified covering $\lambda$ is written via the elliptic functions and one of the parameters of it (and hence of the Hurwitz--Frobenius manifold) is $\tau \in {\mathbb{H}}$, that stands for the modulus of an elliptic curve. For $\textbf{k} = \{2,2,2,2\}$ it has the following form:
$$
\lambda(z) = \sum_{i = 1}^4 \left( \wp(z- a_i, \tau) u_i + \frac{1}{2} \frac{\wp^\prime(z - a_i, \tau)}{\wp(z-a_i, \tau)} s_i \right) + c,
$$
where $\wp(z,\tau)$ is the Weierstrass function and $a_i$,$u_i$,$s_i$,$c$ are complex parameters. The corresponding Frobenius potential is also written in terms of a certain quasi--modular forms (see \cite{B_hf}).
One can consider the Frobenius manifold structure on $\mathcal{H}_{1,\textbf{k}}$ at different points $p_1$ and $p_2$.
Because one of the parameters of $p_1$ and $p_2$ is the modulus of the corresponding elliptic curve, it's natural for two to be connected by a certain $ {\rm SL(2,\mathbb C)} $ action.
Given a Frobenius potential $F_1(\textbf{t})$ encoding the algebra structure at $p_1$ one can consider $F_2(\textbf{t}) := F_1(\textbf{t} + p_2-p_1)$. However such a shift applied to the function $f(z)$, holomorphic in ${\mathbb{H}}$ (like for example the functions $X_k^\infty(\tau)$), reduces drastically the domain of the holomorphicity. In order to keep the domain of holomorphicity big, one should apply not the Taylor series shift, but the following action instead (\cite{Z}):
$$
f(z) \to f \left( \frac{\tau_0 - \bar \tau_0 z}{1-z} \right).
$$
This action can be realized by the composition of the rescaling and the $ {\rm SL(2,\mathbb C)} $ action developed above.
\section{Cohomological field theories and Givental's action}\label{section: GiventalsAction}
We briefly recall some basic facts of the CohFTs and Givental's action on the partition function of it. We introduce Givental's action in two forms -- via the symplectic formalism and in the infinitesimal form. The first one is used in the last section and is more natural from the point of view of singularity theory. The second approach is more explicit and allows us to make explicit computations via the graphical calculus introduced in \cite{DbSS}.
\subsection{Cohomological Field Theory axioms}
Let $(V,\eta)$ be a finite--dimensional vector space with a non--degenerate bilinear form on it.
Consider a system of linear maps
$$
\Lambda_{g,k}: V^{\otimes k} \rightarrow H^*(\M{g}{k}),
$$
defined for all $g,k$ such that $\M{g}{k}$ exists.
It is called Cohomological field theory on $(V,\eta)$ if it satisfies the following axioms.
\begin{itemize}
\item[\textbf A1:] $\Lambda_{g,k}$ is equivariant w.r.t. the $S_k$--action permuting the factors in the tensor product and the numbering of marked points in $\M{g}{k}$.
\item[\textbf A2:] For the gluing morphism $\rho: \M{g_1}{k_1+1} \times \M{g_2}{k_2+1} \rightarrow \M{g_1+g_2}{k_1+k_2}$ we have:
$$
\rho^* \Lambda_{g_1+g_2,k_1+k_2} = (\Lambda_{g_1, k_1+1} \cdot \Lambda_{g_2,k_2+1}, \eta^{-1}),
$$
where we contract with $\eta^{-1}$ the factors of $V$ that correspond to the node in the preimage of $\rho$.
\item[\textbf A3:] For the gluing morphism $\sigma: \M{g}{k+2} \rightarrow \M{g+1}{k}$ we have:
$$
\sigma^* \Lambda_{g+1,k} = (\Lambda_{g, k+2}, \eta^{-1}),
$$
where we contract with $\eta^{-1}$ the factors of $V$ that correspond to the node in the preimage of $\sigma$.
\end{itemize}
In this paper we further assume that the CohFT $\Lambda_{g,k}$ is unital --- there is a fixed vector $\textbf 1 \in V$ called unit such that the following axioms are satisfied.
\begin{itemize}
\item[\textbf U1:] For every $a,b \in V$ we have: $\eta(a, b) = \Lambda_{0,3}(\textbf{1} ,a,b)$.
\item[\textbf U2:] Let $\pi: \M{g}{k+1} \rightarrow \M{g}{k}$ be the map forgetting the last marking, then:
$$
\pi^* \Lambda_{g,k}(a_1 \otimes \dots \otimes a_n) = \Lambda_{g,k+1}(a_1 \otimes \dots \otimes a_k \otimes \textbf{1}).
$$
\end{itemize}
In what follows we will denote the CohFT just by $\Lambda$ rather than $\Lambda_{g,k}$ when there is no ambiguity.
The correlators of an arbitrary CohFT satisfy the following equations: Dilaton equation:
\begin{equation}\label{eq: dilaton}
\left\langle \tau_1(e_1) \prod_{k=1}^l \tau_{d_k}(e_{i_k}) \right\rangle_g = (2g-2+l) \left\langle \prod_{k=1}^l \tau_{d_k}(e_{i_k}) \right\rangle_g.
\end{equation}
and String equation:
\begin{equation}
\left\langle \tau_0(e_1) \prod_{k=1}^l \tau_{d_k}(e_{i_k}) \right\rangle_g = \sum_{m = 1, d_m \neq 0}^l \left\langle \prod_{k \neq m} \tau_{d_k}(e_{i_k}) \cdot \tau_{d_m -1}(e_{i_m}) \right\rangle_g.
\end{equation}
\begin{example}
Let $V$ be $1$-dimensional vector space generated by $e_1$. The map
$$
\alpha_{g,k} : V^{\otimes k} \rightarrow 1 \in H^*(\M{g}{k})
$$ satisfies the axioms of the unital CohFT and is called trivial CohFT.
\end{example}
\subsection{Symplectic geometry of Givental's action}\label{section: GiventalSymplectic}
For every CohFT $\Lambda_{g,k}$ on the vector space $(V,\eta)$ Givental introduced in \cite{G} the following symplectic formalism. Consider the space of formal Laurent series $\mathcal H := V((z))$ with the symplectic structure $\Omega$:
$$
\Omega(f(z),g(z)) := \mathrm{res}_{z=0} \ \eta \left( f(-z),g(z) \right) dz, \quad f,g \in \mathcal{H}.
$$
There is a natural polarization $\mathcal{H} = \mathcal{H}_q \oplus \mathcal{H}_p$ for $\mathcal{H}_q = V[z]$ and $\mathcal{H}_p = z^{-1} V[[z^{-1}]]$. Together with the symplectic form $\Omega$ the space $\mathcal H$ can be identified with the cotangent bundle $T^*\mathcal{H}_q$.
Let the vectors $\partial_k$ for $1 \le k \le n$ build up the basis of $V$ and $dt_k$ constitute the dual basis w.r.t. $\eta$.
For any $f \in \mathcal{H}$ define $p_{k,i}$ and $q^i_k$ to be Darboux coordinates on $\mathcal H$ by the following equation:
$$
f(z) = \sum_{k = 0}^\infty \sum_{i=1}^{n} \left( q_k^i \partial_i z^k + p_{k,i} dt_i (-z)^{-k-1} \right).
$$
In the coordinates $\{p_{k,i}, q^i_k\}$ the symplectic form takes the canonical form $\Omega = \sum_{i,k} dp_{k,i} \wedge dq_{k}^i$.
\subsubsection{Quantization of quadratic Hamiltonians}
Let $A(z)$ be an infinitesimal symplectic operator on $\mathcal H$:
$$
\Omega(A \cdot f,g) + \Omega(f,A \cdot g) = 0.
$$
Givental associates to it a quadratic Hamiltonian $P(A)$ by the following equation:
$$
P(A)(f) = \frac{1}{2} \Omega \left(A \cdot f, f \right), \quad \forall f \in \mathcal{H}.
$$
Quadratic Hamiltonian $P$ can be quantized to a differential operator $\hat P$ by the following Weyl quantization rules:
\begin{align*}
& \hat 1 = 1, \ \hat p_k^i = \sqrt{\hbar} \frac{\partial}{\partial q_k^i}, \ \hat q_k^i = \frac{q_k^i}{\sqrt{\hbar}},
\\
& (p_{k,i} p_{l,j}) \hat{\ } = \hbar \frac{\partial^2}{\partial q_k^i \partial q_l^j}, \ (p_{k,i}q_l^j) \hat{\ } = q_l^j \frac{\partial}{\partial q_k^i}, \ (q_k^iq_l^j)\hat{\ } = \frac{1}{\hbar} q_k^iq_l^j.
\end{align*}
This allows one to define the quantization of $\exp(A(z))$ by the formula:
$$
\exp(A(z)) \hat{\ } := \exp( \hat P(A)).
$$
\subsubsection{CohFT partition function as the function on $\mathcal{H}_q$}
Let $\mathcal Z (\bf t)$ be the partition function of the CohFT $\Lambda$. We can consider $\mathcal Z$ as the function of $\mathcal{H}_q$ by applying the so--called \textit{dilaton shift} coordinate transformation:
$$
t^{d,i} = q_d^i + \delta_{d,1}\delta_{i,1}.
$$
Because of this identification the action of $\hat R$ arising from the quantized quadratic Hamiltonian $P(r)$ can be applied to the partition function $\mathcal Z$ too.
The quantization procedure is of big importance for us because of the following theorem essentially due to Givental.
\begin{theorem}[\cite{K,T,FSZ}]
Let $R = R(z) \in \mathrm{Hom}(V,V)[[z]]$ be such that $R^*(-z)R(z) = 1$. Then the quantized operator $\hat R$ acts on the space of partition functions of the cohomological field theories.
\end{theorem}
The following proposition due to Givental gives explicit formula for a $\hat R$ --action in the symplectic setting.
\begin{proposition}[Proposition~7.3 in \cite{G}]\label{proposition: action in symplectic setting}
Let $R = R(z) \in \mathrm{Hom}(V,V)[[z]]$ be such that $R^*(-z)R(z) = 1$.
The action of $\hat R$ on $F(\bf q) \in \mathbb{C}[[\bf q]]$ reads:
\begin{equation}\label{equation: action in symplectic setting}
\hat R \cdot F(\mathbf{q}) := \left( \exp \left( \frac{\hbar}{2} \sum_{k,l} \eta(\partial^a, \mathrm{V^s}_{k,l} \partial^b) \partial_{q_k^a} \partial_{q_l^b} \right) F(\mathbf{q}) \right) \mid_{\mathbf{q} \to R^{-1} \mathbf{q}},
\end{equation}
for the operators $\mathrm{V^s}_{k,l} \in \mathrm{Hom}(V)$ defined by:
$$
\sum_{k,l = 0}^\infty \mathrm{V^s}_{k,l}(-z)^k(-w)^l = \frac{R^*(z)R(w) - 1}{z+w}.
$$
\end{proposition}
\subsection{Inifinitesimal version of Givental's action}
In this subsection we introduce Givental's group action on the partition function of a CohFT via the inifinitesimal action computed by Y.P.~Lee \cite{L}. Let $\Lambda_{g,k}$ be a unital CohFT on $(V, \eta)$ with the unit $e_1 \in V$.
\subsubsection{Upper--triangular group} Consider
$$
r(z) = \sum_{l \ge 1} r_l z^l \in {\rm Hom}(V,V) \otimes \mathbb C[z],
$$
such that $r(z) + r(-z)^* = 0$ (where the star means dual w.r.t. $\eta$). Following Givental, we define:
$$
\hat R := \exp( \sum_{l=1} \widehat{r_l z^l}),
$$
where
\begin{equation*}
\begin{aligned}
\widehat {r_lz^l} := & -(r_l)_1^\alpha \frac{\partial}{\partial t^{l+1,\alpha}} + \sum_{d =0}^\infty t^{d, \beta} (r_l)_\beta^\alpha \frac{\partial}{\partial t^{d+l, \alpha}}
\\
& + \frac{\hbar}{2} \sum_{i+j=l-1} (-1)^{i+1} (r_l)^{\alpha,\beta} \frac{\partial^2}{\partial t^{i,\alpha} t^{j,\beta}},
\end{aligned}
\end{equation*}
and $(r_l)^{\alpha,\beta} = (r_l)^\alpha_\sigma \eta^{\sigma, \beta}$.
\begin{definition}
The action of the differential operator $\hat R$ on the partition function of the CohFT is called Givental's $R$--action or upper--triangular Givental's group action.
\end{definition}
\subsubsection{Lower--triangular group} Consider
$$
s(z) = \sum_{l \ge 1} s_l z^{-l} \in {\rm Hom}(V,V) \otimes \mathbb C[z^{-1}],
$$
such that $s(z) + s(-z)^* = 0$. Following Givental, we define:
$$
\hat S := \exp( \sum_{l=1}^\infty \widehat{s_l z^{-l}}),
$$
where
\begin{align*}
\sum_{l=1}^\infty (s_lz^{-l})\hat{\ } & = -(s_1)_1^\alpha \frac{\partial}{\partial t^{0,\alpha}}
+ \frac{1}{\hbar} \sum_{d=0}^{\infty} (s_{d+2})_{1,\alpha} \, t^{d,\alpha}
\\
& + \sum_{ \substack{d=0\\ l=1} }^\infty
(s_l)_\beta^\alpha \, t^{d+l,\beta} \frac{\partial}{\partial t^{d,\alpha}}
+ \frac{1}{2 \hbar} \sum_{ \substack{d_1,d_2 \\ \alpha_1,\alpha_2} }
(-1)^{d_1} (s_{d_1+d_2+1})_{\alpha_1,\alpha_2} \, t^{d_1,\alpha_1} t^{d_2,\alpha_2}.\notag
\end{align*}
In what follows we also need the $\hat S_0$ action, which has to be treated exclusively. It is actually clear that for $s_0 \in {\rm Hom} (V,V)$ such that $s_0 + s_0^* = 0$ the action of $S_0 := \exp(s_0)$ defined by:
$$
\tilde t^\alpha = (S_0)_\beta^\alpha t^\beta
$$
preserves the WDVV equation. We will comment more on the $S_0$ action on the full CohFT partition function later.
\begin{definition}
The action of the differential operator $\hat S$ on the partition function of the CohFT is called Givental's $S$--action or lower--triangular Givental's group action.
\end{definition}
The action of the differential operators $\hat R$ and $\hat S$ as above are indeed just another language of the same action introduced via the quantization of quadratic Hamiltonians. However there is a small conventional difference in the two defintions that will play an important role in the last section of this paper.
\begin{proposition}\label{proposition: two actions}
Let $R = \exp(\sum_k r_kz^k)$ be a an element of the upper--triangular group of Givental. Then symplectic form of the action \eqref{equation: action in symplectic setting} is equivalent to the infinitesimal form of the action of $\tilde R := \exp(- \sum_k r_kz^k)$.
\end{proposition}
\begin{proof}
The operator $\hat R$ can be decomposed as follows (see Section~2 in \cite{DbSS}):
\begin{equation}\label{equation: givGraphDecomposition}
\exp \left( \sum_k r_kz^k \right) = \exp \left( \sum_{d \ge 0,l \ge 1} t^{d,\nu} (r_l)^\mu_\nu \frac{\partial}{\partial t^{d+l,\mu}} \right) \exp \left( \sum_{k,l \ge 0} \left(\mathrm{V}_{k,l}\right)^{\mu \nu} \frac{\partial^2}{\partial t^{k,\mu} \partial t^{l,\nu}} \right),
\end{equation}
where $\mathrm{V}_{k,l}$ is defined by:
$$
\sum_{k,l} \mathrm{V}_{k,l}z^kw^l = - \frac{\hbar}{2} \frac{ \exp \left( -r(-z)\right) \exp \left( r(w)\right) - 1}{z+w}.
$$
It's not hard to see that the first multiple in equation \eqref{equation: givGraphDecomposition} corresponds to the (dilaton--shifted) change of variables and the second multiple is exactly certain quadratic differential operator.
The differential operator of equation \eqref{equation: givGraphDecomposition} defined by $\mathrm{V}_{k,l}$ is easily matched with the action of the operator $\mathrm{V}^s_{k,l}$ in the formula \eqref{equation: action in symplectic setting}. Note however that these two a related by a sign change: $\mathrm{V}^s_{k,l} = - \mathrm{V}_{k,l}$.
Another difference of the formulas \eqref{equation: givGraphDecomposition} and \eqref{equation: action in symplectic setting} is use of the inverse operator $R^{-1}$ in the latter one. Because for an infinitesimal version the inverse is obtained by changing the sign we get the proposition.
\end{proof}
\subsection{Givental graph}
We briefly introduce here the construction of \cite{DbSS} for the graphical computation of the Givental's $R$--action \footnote{See also \cite{DbOSS} for the more general framework.} on a CohFT partition function. From now on let us consider particular upper--triangular group element $R = \exp(\sum r_lz^l)$ acting on the partition function of the fixed unital CohFT $\Lambda_{g,k}$.
Let $\gamma$ be a connected graph with leaves and oriented edges. Introduce the decoration of it:
\subsubsection{Leaves} Leaves are decorated either with the vector:
$$
\mathcal L := \exp \left( \sum_{l=1}^\infty r_l z^l \right) \left( \sum_{d=0}^\infty \sum_{\alpha = 1}^n e_\alpha t^{d,\alpha} z^d \right) \in V[[z]],
$$
or with the vector:
$$
\mathcal L_0 := -z \left( \exp (\sum_{l=1}^\infty r_l z^l) - 1 \right) (e_1) \in V[[z]].
$$
\subsubsection{Edges} Let every edge be oriented and consist of two leaves. The decoration of the edge is given by $\mathcal E \in V[[z]] \otimes V[[w]]$. The factor of $V$ with the variable $z$ corresponds to the input leaf and the factor with the variable $w$ --- to the output leaf of the edge.
For $\tilde {\mathcal E} \in {\rm{Hom}}(V,V)[[z,w]]$ given by:
$$
\tilde{\mathcal E} := - \hbar \frac{\exp(-r(-z)) \exp(r(w)) - 1}{z+w},
$$
define $\mathcal E$ by the equality:
$$
\mathcal E = \tilde{\mathcal E} \eta.
$$
\subsubsection{Vertices} The vertex with the valence $n$ is decorated with:
$$
\mathcal V [n] := \sum_{g \ge 0} \hbar^{g-1} \mathcal V_g[n],
$$
where the tensor $\mathcal V_g[n] \in (V^*[[z]])^{\otimes n}$ is defined by its values on the basis:
$$
\mathcal V_g [n] \left( e_{\alpha_1} z_1^{d_1} \otimes \dots \otimes e_{\alpha_n} z_n^{d_n} \right) := \langle \tau_{d_1}(\alpha_1) \dots \tau_{d_n}(\alpha_n) \rangle_g.
$$
The correlators taken are those of the CohFT fixed.
\subsubsection{Contraction of tensors}
Let $\Gamma$ be the set of all connected graphs. Introduce an orientation of its edges. Denote by $V(\gamma)$ the set of vertices of the graph $\gamma$ and for $v \in V(\gamma)$ denote by $H(v)$ the set of half-edges and leaves adjacent to $v$. Denote by $\mathcal V_v$ the decoration of the vertex $v$. Denote by $D(h)$ for $h \in H(h)$ the decoration of $h$. It is either $\mathcal L$, $\mathcal L_0$ or $\mathcal E$.
\begin{theorem}[Section~2 in \cite{DbSS}]\label{eq:CorellatorsInGraph}
$$
\log (\hat R \cdot Z) = \sum_{\gamma \in \Gamma} \frac{1}{ \mathrm{Aut}(\gamma)} \prod_{v \in V(\gamma)} \mathcal V_v \left( \bigotimes_{h \in H(v)} D(h) \right).
$$
\end{theorem}
\begin{remark}\label{remark:orientation of Givental graph}
The orientation of the graph edges does not appear to contribute to the contraction of tensors -- $\mathcal V_v$ . However it plays important role in the automorphisms counting.
\end{remark}
Note that the decoration of the graph has to be consistent in order to give non-zero contribution to the the formula above. For example contracting the tensors at the particular vertex only one summand $\mathcal V_g$ turns out to give a non-zero value. We will use this fact later computing the $R$-action explicitly.
\subsection{Givental's action on the small phase space} Let $\mathcal Z^\prime : = \hat R \cdot \mathcal Z$. We are interested in the functions:
$$
F_g^\prime = [\hbar^{g-1}] \ \log \left( \hat R \cdot \mathcal Z \right) \mid_{ t^{k \ge 1, \alpha} = 0 },
$$
where $[\hbar^{g-1}]$ means taking the coefficient of $\hbar^{g-1}$ in the formal series.
Recall that the action of $\hat R$ is given by the exponentiation of the differential operator $\hat r(z)$.
To compute $F_g^\prime$ it is not enough to consider the operator $\hat r(z)$ restricted to the small phase space.
The first problem is that in $\hat R$ the differential operator $\hat r$ can act on itself --- Givental's action $r(z)$ introduces higher order variables and can ``eat'' them by a subsequent differentiation. The second problem is that only the derivatives of $\mathcal F_g$ and not $F_g$ are used by a Givental's action.
The second problem can not be resolved, but the first one can be simplified when applying Givental graph calculus. To compute $\mathcal Z^\prime$ on the small phase space we should consider the restriction of the decoration $\mathcal L$ to the small phase space.
\begin{proposition}\label{prop: small phase space leaves}
Let $\mathcal Z^\prime : = \hat R \cdot \mathcal Z$. In order to compute the restriction of $\mathcal Z^\prime$ to the small phase space it is enough to consider the graphs with the leaves decorated by
$$
\mathcal L_s = \mathcal L \mid_{t^{k,\alpha} =0, \ k \ge 1}.
$$
\end{proposition}
\begin{proof}
It's clear from formula~\eqref{equation: givGraphDecomposition} and the definition of the graphical calculus that these are indeed the leaves, that ``introduce'' the variables of a $R$--transformed partition function.
\end{proof}
\section{Givental's quantization of the $ {\rm SL(2,\mathbb C)} $ action}\label{section: GiventalsQuantization}
The expression of $\hat A$ given in Proposition \ref{prop:Composition} motivates considering the following ``composition'' action $\hat A_C$ defined in a similar way.
Define:
\begin{equation}\label{equation: A_C}
\hat A_C := \hat S^{-\beta} \cdot \hat S_0^{c} \cdot \hat R^1 \cdot \hat S_0^{c^{-1}} \cdot \hat S^\alpha,
\end{equation}
where $\alpha = \frac{a+1}{c}$ and $\beta = -\frac{d+1}{c}$, ${S^a = S^a(z) = \exp (s_1^a z^{-1})}$, $S_0^c$ are Givental's $S$--actions for:
$$
s_1^a :=
\left(
\begin{array}{c c c}
0 & \dots & 0 \\
\vdots & 0 & \vdots \\
a & \dots & 0
\end{array}
\right),
\quad
S_0^c :=
\left(
\begin{array}{c c c}
c & \dots & 0 \\
\vdots & I_{n-2} & \vdots \\
0 & \dots & c^{-1}
\end{array}
\right),
$$
and $R^1$ is the Givental's $R$--action ``corresponding'' to the Inversion of Dubrovin.
Explicitly this $R$--action was found in \cite{DbSS}:
\begin{theorem}[Theorem~3.1 in \cite{DbSS}] The Inversion transformation on the Frobenius manifold with potential $F$ coincides with Givental's $R$-action given by:
$$
R^1(z) =
\exp( \left(
\begin{array}{c c c}
0 & \dots & 1 \\
\vdots & 0 & \vdots \\
0 & \dots & 0
\end{array}
\right) z).
$$
Namely $\hat R^1$--transformed Frobenius structure of $F$ at the point $(0, \dots, 0, 1)$ coincides with the Frobenius structure of $F^I$ at the point $(0, \dots, 0, -1)$.
\end{theorem}
Note that unlike the ``analytical'' Inversion $\hat I$ that we have in the formula for $\hat A$, Givental's analog $\hat R^1$ also applies certain shift of variables on both sides. In order for two quantizations to agree these additional shifts have to be absorbed.
\begin{proposition}\label{prop:A_g to A}
The action $\hat A_G$ agrees with the action $\hat A$.
\end{proposition}
\begin{proof}
It is clear by construction that both $\hat A_G$ and $\hat A$ act similarly. We only have to show that the shifts of the coordinates applied on both sides agree.
Let $t^s$ and $t^e$ be the coordinates before and respectively after the operator $\hat A$ is applied. Let $t^{\rm AI}$ and $t^{\rm BI}$ be the coordinates after and respectively before the inversion operator $I$.
The action of $\hat A$ changes the coordinates in the following way.
$$
S_{c^{-1}} \cdot T_1: t^{\rm BI} = \frac{t^s}{c} + \frac{a}{c}, \quad I: t^{\rm BI} = - \frac{1}{t^{\rm AI}}, \quad T_2 \cdot S_c: t^{\rm AI} = c(t^e + \frac{d}{c}).
$$
On the $\hat A_G$ side $t^s$ and $t^e$ are the coordinates before and respectively after $\hat A_G$ is applied. Let $\tilde t^{\rm AI}$ and $\tilde t^{\rm BI}$ be the coodinates after and respectively before $\hat R^1$ is applied.
$$
S_0^{c^{-1}} \cdot S^\alpha: \tilde t^{\rm BI} = \frac{t^s}{c} + \frac{a+1}{c}, \quad S^\beta \cdot S_0^c: \tilde t^{\rm AI} = c(t^e + \frac{d+1}{c}).
$$
Due to the theorem above the coordinates $\tilde t^{\rm AI}$ and $t^{\rm AI}$ are related by:
$$
-1 + \tilde t^{\rm AI} = t^{\rm AI}.
$$
This approves the shift $\beta$. In order to approve the shift $\alpha$ it's enough to check that
$$
A \cdot \beta = \frac{a\beta + b}{c\beta + d} = \alpha.
$$
\end{proof}
\begin{corollary}
$\hat A_C$--transformed Frobenius structure of $F$ at the point $(0, \dots, 0, \frac{a+1}{c})$ coincides with the Frobenius structure of $F^A$ at the point $(0, \dots, 0, -\frac{d+1}{c})$.
\end{corollary}
Note that the decomposition of the form \eqref{equation: A_C} does not allow one to make a statement about a more general $R$--action because the actions of the lower-- and upper--triangular groups do not commute. Therefore we are not allowed to multiply the matrices $S_0^c$ and $R^1$ to get a new $R$--action.
Another disappointment of the action $\hat A_C$ is that the choice of the points of the Frobenius structure expansion on both sides is very restrictive due to the corollary above. This restriction can not be relaxed while we use the Givental's analogue of the Inversion -- the action of $\hat R^1$. The following proposition makes more general statement without a special choice of the Frobenius structure expansion points.
\begin{proposition}\label{prop:Action via graphs}
For every $A \in {\rm SL(2,\mathbb C)} $ and $\textbf{s} = (0,\dots,0,\tau) \in \mathbb{C}^n$ such that $F(\textbf{t})$ is analytic at $A\cdot\textbf{s} = (0,\dots,0,A\cdot \tau)$ consider the upper--triangular group element $R^\sigma$:
$$
R^\sigma(z) =
\exp( r^\sigma z), \quad
r^\sigma :=
\left(
\begin{array}{c c c}
0 & \dots & \sigma \\
\vdots & 0 & \vdots \\
0 & \dots & 0
\end{array}
\right), \quad \sigma = -c(c\tau+d).
$$
Then we have:
\begin{itemize}
\item The action $\hat A$ on $F(\textbf{t})$ given by \eqref{eq:SLAction} coincides with the action of $\hat R^\sigma(z)$ up to the rescaling:
$$
F^A \to (c\tau+d)^2 F^A, \quad t^n \to (c\tau+d)^2 t^n, \quad t^k \to (c\tau + d) t^k, \ 1 < k < n.
$$
\item The action of $\hat R^\sigma$ identifies Frobenius manifold structure of $F^A$ at the point $(0, \dots, 0, \tau)$ with the $\hat R^\sigma$--transformed Frobenius manifold structure of $F$ at the point ${(0, \dots, 0, A \cdot \tau)}$.
\end{itemize}
\end{proposition}
Proposition \ref{prop:Action via graphs} together with Proposition~\ref{prop:A_g to A} complete the proof of Theorem~\ref{th:FrobeniusManifold}.
While the formulation of Proposition~\ref{prop:Action via graphs} is enough for the isomorphism of the Frobenius manifolds given by $F^A$ and $\hat R^\sigma \cdot F$, we consider in more details the rescaling used in the next sections where it will play an important role.
The rest of this section is devoted to the proof of Proposition~\ref{prop:Action via graphs}.
\subsection{Graphical calculus of $\hat R^\sigma$ in genus $0$}\label{section: GraphsCalculus}
From now on we consider a particular $R$-action given by $R^\sigma$ on the CohFT whose Frobenius structure is given by the potential $F$.
In this subsection fix the notation:
$$
R^\sigma(z) =
\exp( r^\sigma z), \quad
r^\sigma :=
\left(
\begin{array}{c c c}
0 & \dots & \sigma \\
\vdots & 0 & \vdots \\
0 & \dots & 0
\end{array}
\right), \quad \sigma \in \mathbb C \backslash \{0\}.
$$
We further assume $F$ to be given in the form \eqref{eq:WDVV potential}. In particular we assume $\eta_{i,j} = \delta_{i+j,n+1}$. To simplify the formulae introduce the notation:
\begin{notation}
For $1 \le \alpha_i \le n$ define:
$$
\langle \tau_{d_1}(\alpha_1) \dots \tau_{d_k}(\alpha_k) \rangle_g := \langle \tau_{d_1}(e_{\alpha_1}) \dots \tau_{d_k}(e_{\alpha_k}) \rangle_g.
$$
We also denote with the capital $G$ subscript the correlators of Givental-transformed partition function. Namely for $\langle \cdot \rangle_g$ --- correlators of $\mathcal Z$ we denote by $\langle \cdot \rangle_g^G$ the correlators of $ \hat R^\sigma \cdot \mathcal Z$:
$$
\left\langle \prod_k \tau_{d_p}(\alpha_p) \right\rangle_g^G := \prod_p \frac{\partial}{\partial t^{d_p, \alpha_p}} \ [\hbar^{g-1} ] \log \left( \hat R^\sigma \cdot \mathcal Z \right) \ \mid_{\textbf{t} = 0}.
$$
For $e_\alpha \in V$ define $e_{\bar \alpha} \in V$ to be the vector such that $\eta(e_\alpha, e_{\bar \alpha}) = 1$. With our choice of $\eta$ it is equivalent to:
$$
\bar \alpha := n+1-\alpha \quad \text{ for } \quad 1 \le \alpha \le n.
$$
\end{notation}
We classify the graphs giving non--trivial contribution to the action of $R^\sigma$ in the following proposition.
\begin{proposition}\label{prop:GraphFamilies}
The only decorated graphs giving non-trivial contribution to the Frobenius structure of $\hat R^\sigma \cdot \mathcal Z$ are those from the following two series:
\\
\begin{tabular}{ll}
\raisebox{-.5\height}
{
\hbox{
\begin{tikzpicture}[y=1.1pt, x=1.1pt,yscale=-1]
\path[draw=black,line width=0.634pt] (42,79) -- (65,57);
\path[draw=black,line width=0.634pt] (42,27) -- (65,49);
\path[draw=black,line width=0.634pt] (37,34) -- (65,49);
\path[draw=black,line width=0.634pt] (96,26) -- (73,49);
\path[draw=black,line width=0.634pt] (96,79) -- (73,57);
\path[draw=black,line width=0.634pt] (102,48) -- (75,52);
\path[draw=black,line width=0.634pt] (102,57) -- (74,54);
\path[cm={{1.056,0.0,0.0,1.056,(11.36937,-49.61728)}}, draw=black, fill opacity=0.000,line width=0.634pt]
(60,97)arc(0:180:5)arc(-180:0:5) -- cycle;
\path[cm={{1.00447,0.10339,-0.16196,1.24445,(-8.229,-14.22512)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:196:22 and 20);
\path[cm={{-0.59464,-0.32663,0.27113,-1.18207,(114.92367,126.65456)}}, draw=black, dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:157:22 and 20);
\path[cm={{-0.35725,-0.57676,0.81721,-0.89609,(68.3994,154.12582)}}, draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:152:22 and 20);
\path[fill=black] (16,39) node[above right] {$e_{\alpha_2}$};
\path[fill=black] (30,92) node[above right] {$e_{\alpha_N}$};
\path[fill=black] (29,28) node[above right] {$e_{\alpha_1}$};
\path[fill=black] (98,28) node[above right] {$z e_1$};
\path[fill=black] (102,50) node[above right] {$z e_1$};
\path[fill=black] (128,46) node[above right] {$\Big \} k$ times};
\path[fill=black] (102,66) node[above right] {$e_n$};
\path[fill=black] (94,92) node[above right] {$e_n$};
\path[fill=black] (128,86) node[above right] {$\Big \} p$ times};
\end{tikzpicture}
}
}
& \quad $1 \le \alpha_i < n$, \quad $0 \le k,p$, \quad $3 \le N+p$.
\vspace*{5 mm}
\end{tabular}
\begin{tabular}{ll}
\raisebox{-.6\height}
{
\hbox{
\begin{tikzpicture}[y=1.1pt, x=1.1pt,yscale=-1]
\path[fill=black] (-20,52) node[above right] {$e_{\bar \alpha}$};
\path[fill=black] (-9,25) node[above right] {$e_\alpha$};
\path[fill=black] (2,98) node[above right] {$ze_1$};
\path[fill=black] (39,98) node[above right] {$z e_1$};
\path[fill=black] (35,47) node[above right] {$e_1$};
\path[fill=black] (64,118) node[above right] {$\underbrace{\hspace{65pt}}$};
\path[fill=black] (78,138) node[above right] {$l$ times};
\path[fill=black] (63,90) node[above right] {$z e_1$};
\path[fill=black] (90,90) node[above right] {$z e_1$};
\path[fill=black] (63,47) node[above right] {$e_1$};
\path[fill=black] (91,47) node[above right] {$e_1$};
\path[fill=black] (71,98) node[above right] {$\underbrace{\hspace{12pt}}$};
\path[fill=black] (75,108) node[above right] {$m_i$};
\path[fill=black] (17,104) node[above right] {$\underbrace{\hspace{18pt}}$};
\path[fill=black] (20,114) node[above right] {$m_s$};
\path[fill=black] (212,52) node[above right] {$e_{\bar \beta}$};
\path[fill=black] (204,25) node[above right] {$e_{\beta}$};
\path[fill=black] (188,98) node[above right] {$z e_1$};
\path[fill=black] (154,47) node[above right] {$e_1$};
\path[fill=black] (78,25) node[above right] {$e_n$};
\path[fill=black] (152,98) node[above right] {$z e_1$};
\path[fill=black] (166,104) node[above right] {$\underbrace{\hspace{18pt}}$};
\path[fill=black] (166,114) node[above right] {$m_e$};
\path[draw=black,line width=0.634pt] (3.3834,25.5017) -- (26.3404,47.9327);
\path[draw=black,line width=0.634pt] (45.6274,85.2293) -- (32.8873,55.9217);
\path[draw=black,line width=0.634pt] (13.9659,85.2293) -- (26.7061,55.9217);
\path[draw=black,line width=0.634pt] (35.5090,52.0541) -- (59.6537,52.0541);
\path[draw=black,line width=0.634pt] (-6.5525,52.0541) -- (24.5093,52.0541);
\path[draw=black,dash pattern=on 1pt off 1pt,line width=0.634pt] (145.5525,52.0541) -- (115.5093,52.0541);
\path[cm={{0.72134,0.70662,-0.90194,0.87257,(14.04218,-50.16265)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000, line width=0.634pt] (61,64)arc(119:166:22 and 20);
\path[cm={{0.51685,-0.86747,1.06045,0.67106,(-61.47482,90.94997)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000, line width=0.634pt] (61,64)arc(119:166:22 and 20);
\path[cm={{1.056,0.0,0.0,1.056,(-28.07013,-51.23201)}},draw=black, fill opacity=0.000, line width=0.634pt]
(60,97)arc(0:180:5)arc(-180:0:5) -- cycle;
\path[cm={{1.056,0.0,0.0,1.056,(26.57106,-51.23201)}},draw=black, fill opacity=0.000,line width=0.634pt]
(60,97)arc(0:180:5)arc(-180:0:5) -- cycle;
\path[cm={{-0.72134,0.70662,0.90194,0.87257,(194.7515,-50.16267)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:166:22 and 20);
\path[cm={{-0.51685,-0.86747,-1.06045,0.67106,(270.26849,90.94994)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:166:22 and 20);
\path[cm={{0.44848,-0.88233,0.92016,0.68255,(5.83693,85.10118)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:166:22 and 20);
\path[cm={{-1.056,0.0,0.0,1.056,(236.86381,-51.23204)}},draw=black,line width=0.634pt] (60,97)arc(0:180:5)arc(-180:0:5) -- cycle;
\path[draw=black,line width=0.634pt] (96.8314,78.0787) -- (87.4289,55.8222);
\path[draw=black,line width=0.634pt] (72.7407,78.0895) -- (81.4577,55.8113);
\path[draw=black,line width=0.634pt] (90.0705,52.0541) -- (109.4266,52.0541);
\path[draw=black,line width=0.634pt] (85.0705,46.0541) -- (85.0705,25.0541);
\path[draw=black,line width=0.634pt] (59.7007,52.0541) -- (79.3957,52.0541);
\path[draw=black,line width=0.634pt] (205.4103,25.5017) -- (182.4532,47.9327);
\path[draw=black,line width=0.634pt] (163.1663,85.2293) -- (175.9064,55.9217);
\path[draw=black,line width=0.634pt] (194.8277,85.2293) -- (182.0876,55.9217);
\path[draw=black,line width=0.634pt] (173.2847,52.0541) -- (149.1400,52.0541);
\path[draw=black,line width=0.634pt] (215.3461,52.0541) -- (184.2843,52.0541);
\end{tikzpicture}
}
}
& \begin{tabular}{@{}c@{}} $1 \le \alpha, \beta \le n$, \\ $0 \le l$, $0 \le m_k$. \end{tabular}
\vspace*{4 mm}
\end{tabular}
\\
with the vertices decorated by $\mathcal V_{g=0}[\cdot]$.
\end{proposition}
\begin{notation}
We will denote by
$$
\langle \tau_0(\alpha_1) \dots \tau_0(\alpha_l) (\tau_0(n))^k \rangle_H^G
$$
the genus $0$ correlators of $\hat R \cdot Z$ obtained by graph calculus from the first (one-vertex) series above and by
$$
\langle \tau_0(\alpha) \tau_0(\bar \alpha) \tau_0(\beta) \tau_0(\bar \beta) (\tau_0(n))^k \rangle_Q^G
$$
the correlators obtained from the second series above.
\end{notation}
\begin{proof}[Proof of the proposition \ref{prop:GraphFamilies}]
Note that the only non-zero element of $(r^\sigma)^{ij}$ is $(r^\sigma)^{11} = \sigma$. Therefore there is a unique choice for an edge decoration:
$$
\mathcal E = - (r^\sigma)^{11} e_1 \otimes e_1.
$$
At the same time $r^\sigma (e_\alpha) \neq 0 \iff \alpha = n$ and for the leaves decoration we have:
$$
\mathcal L = \sum_{d \ge 0} \sigma e_1 t^{d,n} z^{d+1} + \sum_{d=0}^\infty \sum_{\alpha = 1}^n e_\alpha t^{d,\alpha} z^d, \quad \mathcal L_0 = 0.
$$
Restricting ourselves to the Frobenius structure defined by the CohFT $\hat R^\sigma \cdot Z$ we consider only genus zero graphs decorated with (recall Proposition~\ref{prop: small phase space leaves}):
$$
\mathcal L_s = \sigma e_1 t^{0,n} z^{1} + \sum_{\alpha = 1}^n e_\alpha t^{0, \alpha} z^0, \quad \mathcal L_0 = 0.
$$
Hence the graphs giving non-trivial contribution are genus $0$ graphs obtained by $\mathcal E$-gluing from the following two types of graphs:
\begin{tabular}{ll}
\raisebox{-.5\height}
{
\hbox{
\begin{tikzpicture}[y=1.1pt, x=1.1pt,yscale=-1]
\path[draw=black,line width=0.634pt] (35,51) -- (75,51);
\path[draw=black,line width=0.634pt] (3,78) -- (26,55);
\path[draw=black,line width=0.634pt] (3,25) -- (26,47);
\path[cm={{1.056,0.0,0.0,1.056,(-28.07013,-51.23201)}},draw=black, fill opacity=0.000,line width=0.634pt]
(60,97)arc(0:180:5)arc(-180:0:5) -- cycle;
\path[fill=black] (-9,90) node[above right] {$e_{\bar \alpha}$};
\path[fill=black] (-9,26) node[above right] {$e_\alpha$};
\path[fill=black] (60,50) node[above right] {$e_1$};
\end{tikzpicture}
\qquad
\begin{tikzpicture}[y=1.1pt, x=1.1pt,yscale=-1]
\path[draw=black,line width=0.634pt] (42,79) -- (65,57);
\path[draw=black,line width=0.634pt] (42,27) -- (65,49);
\path[draw=black,line width=0.634pt] (37,34) -- (65,49);
\path[draw=black,line width=0.634pt] (96,26) -- (73,49);
\path[draw=black,line width=0.634pt] (96,79) -- (73,57);
\path[draw=black,line width=0.634pt] (95,67) -- (74,56);
\path[cm={{1.00447,0.10339,-0.16196,1.24445,(-8.229,-14.22512)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:196:22 and 20);
\path[cm={{-0.59464,-0.32663,0.27113,-1.18207,(114.38882,128.80704)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:196:22 and 20);
\path[cm={{1.056,0.0,0.0,1.056,(11.36937,-49.61728)}},draw=black,fill opacity=0.000,line width=0.634pt]
(60,97)arc(0:180:5)arc(-180:0:5) -- cycle;
\path[fill=black] (30,92) node[above right] {$e_{i_l}$};
\path[fill=black] (29,23) node[above right] {$e_{i_1}$};
\path[fill=black] (22,37) node[above right] {$e_{i_2}$};
\path[fill=black] (98,26) node[above right] {$z r(e_n)$};
\path[fill=black] (97,97) node[above right] {$z r(e_n)$};
\path[fill=black] (140,68) node[above right]{$\Bigg{ \}} k$ times};
\end{tikzpicture}
}
}
&
\begin{tabular}{@{}c@{}} $1 \le \alpha, i_p \le n$, \\ $3 \le l$, \\ $ 0 \le k$. \end{tabular}
\vspace*{5 mm}
\end{tabular}
\\
However the gluing is constituted by the $e_1 \otimes e_1$--decorated edges only. Because of the Dilaton equation we arrive at the series presented in the statement of the proposition.
\end{proof}
\begin{corollary}\label{cor:FGpotential}
Frobenius potential $F^G$ of the CohFT $\hat R^\sigma \cdot \mathcal Z$ reads:
\begin{equation*}
\begin{aligned}
F^G(\textbf t) &= \frac{(t^1)^2t^n}{2} + \frac{t^1}{2} \sum_{p=2}^{n-1} t^{p} t^{\bar p} + \sum_{\{ \alpha \}, k} \frac{(t^n)^k \prod t^{\alpha_i} }{k! \mathrm{Aut} (\{ \alpha \})} \left \langle (\tau_0(n))^k \prod \tau_0(\alpha_i) \right \rangle_H^G
\\
& + \sum_{\alpha, \beta, k} \frac{t^\alpha t^\beta t^{\bar \alpha} t^{\bar \beta} (t^n)^k}{k! \mathrm{Aut} (\alpha, \beta, \bar \alpha, \bar \beta)} \left \langle \tau_0(\alpha) \tau_0(\bar \alpha) \tau_0(\beta) \tau_0(\bar \beta) (\tau_0(n))^k \right \rangle_Q^G,
\end{aligned}
\end{equation*}
where the summations are taken over $\alpha_i,\alpha,\beta \neq 1$.
\end{corollary}
\begin{proof}
Indeed the $R$-action does not change the algebra structure at $\textbf t = 0$. Hence the third order terms of $F^G$ coincide with the third order terms of $F$.
We do not consider lower order terms (that are indeed different) because they do not change the Frobenius structure.
The summands written obviously come from the graphical calculus of the Givental graphs and the proposition above.
However the correlators $\langle \cdot \rangle_H^G$ and $\langle \cdot \rangle_Q^G$ are allowed to have $\tau_0(e_1)$ insertions both contributing to $F^G$ as the coefficients of the monomials containing $t^1$. But such monomials could only appear in the third order terms of the Frobenius potential that are preserved by the $R$--action.
\end{proof}
\begin{proposition}\label{prop:HCorellator}
\begin{equation*}
\begin{aligned}
\langle \tau_0(\alpha_1) & \dots \tau_0(\alpha_N) \left( \tau_0(e_n) \right)^q \rangle^G_H
\\
&= \sum_{p+k = q} \frac{q!}{p!} { N + k + p -3 \choose k} \langle \tau_0(\alpha_1) \dots \tau_0(\alpha_N) \left( \tau_0(e_n) \right)^p \rangle \cdot \sigma^k.
\end{aligned}
\end{equation*}
\end{proposition}
\begin{proof}
Recall that $r(e_n) = \sigma e_1$. From the graphical calculus we have:
\\
\begin{tabular}{ll}
$\left\langle \tau_0(\alpha_1) \dots \tau_0(\alpha_N) \left( \tau_0(n) \right)^q \right\rangle^G_H = $
\larger
$\sum_{p+k=q} {q \choose k} \sigma^k $
&
\raisebox{-.4\height}
{
\hbox{
\begin{tikzpicture}[y=1.1pt, x=1.1pt,yscale=-1]
\path[draw=black,line width=0.634pt] (42,79) -- (65,57);
\path[draw=black,line width=0.634pt] (42,27) -- (65,49);
\path[draw=black,line width=0.634pt] (37,34) -- (65,49);
\path[draw=black,line width=0.634pt] (96,26) -- (73,49);
\path[draw=black,line width=0.634pt] (96,79) -- (73,57);
\path[draw=black,line width=0.634pt] (102,48) -- (75,52);
\path[draw=black,line width=0.634pt] (102,57) -- (74,54);
\path[cm={{1.056,0.0,0.0,1.056,(11.36937,-49.61728)}}, draw=black, fill opacity=0.000,line width=0.634pt]
(60,97)arc(0:180:5)arc(-180:0:5) -- cycle;
\path[cm={{1.00447,0.10339,-0.16196,1.24445,(-8.229,-14.22512)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:196:22 and 20);
\path[cm={{-0.59464,-0.32663,0.27113,-1.18207,(114.92367,126.65456)}}, draw=black, dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:157:22 and 20);
\path[cm={{-0.35725,-0.57676,0.81721,-0.89609,(68.3994,154.12582)}}, draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:152:22 and 20);
\path[fill=black] (16,39) node[above right] {$e_{\alpha_2}$};
\path[fill=black] (30,92) node[above right] {$e_{\alpha_N}$};
\path[fill=black] (29,28) node[above right] {$e_{\alpha_1}$};
\path[fill=black] (98,28) node[above right] {$z e_1$};
\path[fill=black] (102,50) node[above right] {$z e_1$};
\path[fill=black] (128,46) node[above right] {$\Big \} k$ times};
\path[fill=black] (102,66) node[above right] {$e_n$};
\path[fill=black] (94,92) node[above right] {$e_n$};
\path[fill=black] (128,86) node[above right] {$\Big \} p$ times};
\end{tikzpicture}
}
}
\vspace*{5 mm}
\end{tabular}
\\
where the decoration of the vertices on the right hand side is given by the partial derivatives of the function $F$. In terms of correlators the equation above reads:
\begin{equation*}\label{HCorellator}
\begin{aligned}
\langle \tau_0(\alpha_1) & \dots \tau_0(\alpha_N) \left( \tau_0(n) \right)^q \rangle^G_H
\\
&= \sum_{p+k = q} \frac{q!}{k! p!} \left\langle \tau_0(\alpha_1) \dots \tau_0(\alpha_N) \left( \tau_0(n) \right)^p \left( \tau_0(1) \right)^k \right\rangle \cdot \sigma^k.
\end{aligned}
\end{equation*}
Using the Dilaton equation we get:
\begin{equation*}
\begin{aligned}
\langle \tau_0(\alpha_1) & \dots \tau_0(\alpha_N) \left( \tau_0(n) \right)^q \rangle^G_H
\\
&= q! \sum_{p+k = q} \frac{(N + k + p -3)!}{p!k! (N + p-3)!} \cdot \langle \tau_0(\alpha_1) \dots \tau_0(\alpha_N) \left( \tau_0(n) \right)^p \rangle \cdot \sigma^k.
\end{aligned}
\end{equation*}
The proposition follows.
\end{proof}
\begin{definition}
For any $1 \le \alpha, \beta \le n$ define:
$$
\mathrm{Aut}_2 (\alpha, \beta) :=
\begin{cases}
8 & \text{if} \quad \alpha = \beta = \bar \alpha = \bar \beta,
\\
1 & \text{if} \quad \text{all} \ \alpha, \beta, \bar \alpha, \bar \beta \ \text{are different},
\\
2 & \text{otherwise}.
\end{cases}
$$
\end{definition}
\begin{proposition}\label{prop:QCorellator} For $\alpha, \beta \neq 1,n$ we have:
\begin{equation*}
\left\langle \tau_0(\alpha) \tau_0(\beta) \tau_0(\bar \alpha) \tau_0(\bar \beta) \left( \tau_0(n) \right)^k \right\rangle^G_Q
=
- \frac{k! | \mathrm{Aut}((\alpha,\beta,\bar \alpha, \bar \beta))| }{| \mathrm{Aut}_2(\alpha,\beta)|} \sigma^{k+1}.
\end{equation*}
\end{proposition}
\begin{proof}
From graphical calculus (keeping in mind Remark~\ref{remark:orientation of Givental graph}) we have:
\begin{align*}
& \left \langle \tau_0(\alpha) \tau_0(\bar \alpha) \tau_0(\beta) \tau_0(\bar \beta) \left( \tau_0(n) \right)^k \right\rangle^G_Q =
\end{align*}
\\
\begin{tabular}{ll}
$=$ \larger $\sum_{l=0}^{k} \sum_{ \substack{ m_i \ge 0 \\ |\textbf m| = k-l}} \frac{(-1)^{l+1} \sigma^{k+1}}{ m_s! m_e! \prod_{i=1}^l m_i !} $
&
\raisebox{-.65\height}
{
\hbox{
\begin{tikzpicture}[y=1.1pt, x=1.1pt,yscale=-1]
\path[fill=black] (-20,52) node[above right] {$e_{\bar \alpha}$};
\path[fill=black] (-9,25) node[above right] {$e_\alpha$};
\path[fill=black] (2,98) node[above right] {$ze_1$};
\path[fill=black] (39,98) node[above right] {$z e_1$};
\path[fill=black] (35,47) node[above right] {$e_1$};
\path[fill=black] (64,118) node[above right] {$\underbrace{\hspace{65pt}}$};
\path[fill=black] (98,134) node[above right] {$l$};
\path[fill=black] (63,90) node[above right] {$z e_1$};
\path[fill=black] (90,90) node[above right] {$z e_1$};
\path[fill=black] (63,47) node[above right] {$e_1$};
\path[fill=black] (91,47) node[above right] {$e_1$};
\path[fill=black] (71,98) node[above right] {$\underbrace{\hspace{12pt}}$};
\path[fill=black] (75,108) node[above right] {$m_i$};
\path[fill=black] (17,104) node[above right] {$\underbrace{\hspace{18pt}}$};
\path[fill=black] (20,114) node[above right] {$m_s$};
\path[fill=black] (212,52) node[above right] {$e_{\bar \beta}$};
\path[fill=black] (204,25) node[above right] {$e_{\beta}$};
\path[fill=black] (188,98) node[above right] {$z e_1$};
\path[fill=black] (154,47) node[above right] {$e_1$};
\path[fill=black] (78,25) node[above right] {$e_n$};
\path[fill=black] (152,98) node[above right] {$z e_1$};
\path[fill=black] (166,104) node[above right] {$\underbrace{\hspace{18pt}}$};
\path[fill=black] (166,114) node[above right] {$m_e$};
\path[draw=black,line width=0.634pt] (3.3834,25.5017) -- (26.3404,47.9327);
\path[draw=black,line width=0.634pt] (45.6274,85.2293) -- (32.8873,55.9217);
\path[draw=black,line width=0.634pt] (13.9659,85.2293) -- (26.7061,55.9217);
\path[draw=black,line width=0.634pt] (35.5090,52.0541) -- (59.6537,52.0541);
\path[draw=black,line width=0.634pt] (-6.5525,52.0541) -- (24.5093,52.0541);
\path[draw=black,dash pattern=on 1pt off 1pt,line width=0.634pt] (145.5525,52.0541) -- (115.5093,52.0541);
\path[cm={{0.72134,0.70662,-0.90194,0.87257,(14.04218,-50.16265)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000, line width=0.634pt] (61,64)arc(119:166:22 and 20);
\path[cm={{0.51685,-0.86747,1.06045,0.67106,(-61.47482,90.94997)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000, line width=0.634pt] (61,64)arc(119:166:22 and 20);
\path[cm={{1.056,0.0,0.0,1.056,(-28.07013,-51.23201)}},draw=black, fill opacity=0.000, line width=0.634pt]
(60,97)arc(0:180:5)arc(-180:0:5) -- cycle;
\path[cm={{1.056,0.0,0.0,1.056,(26.57106,-51.23201)}},draw=black, fill opacity=0.000,line width=0.634pt]
(60,97)arc(0:180:5)arc(-180:0:5) -- cycle;
\path[cm={{-0.72134,0.70662,0.90194,0.87257,(194.7515,-50.16267)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:166:22 and 20);
\path[cm={{-0.51685,-0.86747,-1.06045,0.67106,(270.26849,90.94994)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:166:22 and 20);
\path[cm={{0.44848,-0.88233,0.92016,0.68255,(5.83693,85.10118)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:166:22 and 20);
\path[cm={{-1.056,0.0,0.0,1.056,(236.86381,-51.23204)}},draw=black,line width=0.634pt] (60,97)arc(0:180:5)arc(-180:0:5) -- cycle;
\path[draw=black,line width=0.634pt] (96.8314,78.0787) -- (87.4289,55.8222);
\path[draw=black,line width=0.634pt] (72.7407,78.0895) -- (81.4577,55.8113);
\path[draw=black,line width=0.634pt] (90.0705,52.0541) -- (109.4266,52.0541);
\path[draw=black,line width=0.634pt] (85.0705,46.0541) -- (85.0705,25.0541);
\path[draw=black,line width=0.634pt] (59.7007,52.0541) -- (79.3957,52.0541);
\path[draw=black,line width=0.634pt] (205.4103,25.5017) -- (182.4532,47.9327);
\path[draw=black,line width=0.634pt] (163.1663,85.2293) -- (175.9064,55.9217);
\path[draw=black,line width=0.634pt] (194.8277,85.2293) -- (182.0876,55.9217);
\path[draw=black,line width=0.634pt] (173.2847,52.0541) -- (149.1400,52.0541);
\path[draw=black,line width=0.634pt] (215.3461,52.0541) -- (184.2843,52.0541);
\end{tikzpicture}
}
}
\end{tabular}
\\
where we denote $|m| := m_s + m_e + \sum_{i=1}^l m_i$ and assume $0! = 1$.
Using Dilaton equation and writing in correlators the equation above reads:
\begin{equation*}
\begin{aligned}
\langle \tau_0(\alpha) \tau_0(\bar \alpha) & \tau_0(\beta) \tau_0(\bar \beta) \left( \tau_0(n) \right)^k \rangle^G_Q
\\
&= \frac{k! | \mathrm{Aut}((\alpha,\beta,\bar \alpha, \bar \beta))| }{| \mathrm{Aut}_2(\alpha,\beta)|} \sigma^{k+1} \sum_{p = 1}^{k+1} (-1)^p {k+1 \choose p}
\\
&= - \frac{k! | \mathrm{Aut}((\alpha,\beta,\bar \alpha, \bar \beta))| }{| \mathrm{Aut}_2(\alpha,\beta)|} \sigma^{k+1}.
\end{aligned}
\end{equation*}
\end{proof}
\subsection{From $\hat R^c \cdot \mathcal Z$ to $F^A_0$}\label{subsec: Analysis to graphs}
Following the method of \cite{DbSS} we prove Proposition~\ref{prop:Action via graphs}.
\begin{proof}[Proof of Proposition~\ref{prop:Action via graphs}.]
Consider the potential $F^A$. We show explicitly that the coefficients of its series expansion coincide with the correlators obtained by Givental's group action.
Let $(0,\dots,0,\tau)$ be a point on the Frobenius manifold of $F^A$.
Introduce the constants:
$$
\gamma := \frac{a\tau + b}{c\tau + d}, \quad \mu := \tau.
$$
\subsubsection*{\textbf {H-correlators}}
Write the series expansion of the function $H$ (recall the formula \eqref{eq:SLAction}) of $F^A$ at the point $(0, \dots, 0, \gamma)$. We have:
\begin{equation*}
\begin{aligned}
{\mathcal K} := & (ct^n+d)^2 H \left(\frac{t^2}{ct^n+d}, \dots, \frac{t^{n-1}}{ct^n+d}, \frac{at^n + b}{ct^n+d} \right)
\\
&= (ct^n+d)^2\sum_{p,\alpha_1, \dots, \alpha_N} \frac{t^{\alpha_1} \dots t^{\alpha_N}}{| \mathrm{Aut}(\alpha)| p!} \left( \frac{at^n + b}{ct^n + d} - \gamma \right)^p \frac{H_{\alpha_1 \dots \alpha_N, n^p}}{(ct^n + d)^N},
\end{aligned}
\end{equation*}
where we denote
$$
H_{\alpha_1 \dots \alpha_N, n^p} := \frac{\partial^{N+p} H (t^2, \dots, t^n)}{\partial t^{\alpha_1} \dots \partial t^{\alpha_N} (\partial t^n)^p} \mid_{\textbf t = (0, \dots, 0, \gamma) }.
$$
Let $\tilde t^n = t^n - \mu$. The function rewrites:
\begin{align*}
\mathcal K &= \sum_{p,\alpha_1, \dots, \alpha_N} \frac{t^{\alpha_1} \dots t^{\alpha_N}}{| \mathrm{Aut}(\alpha)| p!}
\frac{ \left( \tilde t^n \right)^p }{ (c \tau + d)^p }
\frac{ H_{\alpha_1 \dots \alpha_N, n^p} }{ (c \tilde t^n + (c \tau + d))^{N+p-2}}
\\
&= \sum_{p,\alpha_1, \dots, \alpha_N} \frac{t^{\alpha_1} \dots t^{\alpha_N}}{| \mathrm{Aut}(\alpha)| p!}
\frac{ \left( \tilde t^n \right)^p }{ (c \tau + d)^p}
\frac{1}{(c \tau + d)^{N+p-2}}\frac{ H_{\alpha_1 \dots \alpha_N, n^p} }{ (c (c \tau + d)^{-1}\tilde t^n + 1)^{N+p-2}}
\end{align*}
Taking $\hat t^{\alpha_k} := t^{\alpha_k} / (c\tau + d)$, $\hat t^n := \tilde t^n / (c\tau + d)^2$ and $\sigma = -c(c \tau + d)$ the expression simplifies to:
\begin{align*}
\mathcal K &= (c\tau+d)^2\sum_{p,\alpha_1, \dots, \alpha_N} \frac{\hat t^{\alpha_1} \dots \hat t^{\alpha_N}}{| \mathrm{Aut}(\alpha)| p!}
\frac{ \left( \hat t^n \right)^p }{ ( 1 - \sigma \hat t^n)^{N+p-2}} H_{\alpha_1 \dots \alpha_N, n^p}
\\
&= (c\tau+d)^2 \sum_{k \ge 0} \sum_{N+p \ge 3} \sum_{\alpha_1, \dots, \alpha_N} {N + k + p -3 \choose k} \frac{\hat t^{\alpha_1} \dots \hat t^{\alpha_N}}{| \mathrm{Aut}(\alpha)| p!} (\hat t^n)^{p+k} H_{\alpha_1 \dots \alpha_N, n^p} \left( \sigma \right)^k
\end{align*}
Using Proposition~\ref{prop:HCorellator} we get in new variables exactly one of the higher than three order summands of the Frobenius potential written in the Corollary~\ref{cor:FGpotential}.
\subsubsection*{\textbf {Q-correlators}}
In an analogous way we write:
\begin{equation*}
\begin{aligned}
\frac{c}{8(ct^n + d)} (t^2t^{n-1} +& \dots + t^{n-1}t^2)^2 =
\\
\frac{c}{8(c \tilde t^n + c\tau + d )} (t^2t^{n-1} +& \dots + t^{n-1}t^2)^2 =
\\
(c\tau+d)^2 \frac{-\sigma}{8(1 -\sigma \hat t^n )} (\hat t^2 \hat t^{n-1} +& \dots + \hat t^{n-1} \hat t^2)^2 =
\\
&= - (c\tau+d)^2 \sum_{k \ge 0} \frac{1}{| \mathrm{Aut}_2(\alpha, \beta)|} \hat t^\alpha \hat t^{\bar \alpha} \hat t^{\beta} \hat t^{\bar \beta} \left( \sigma \right)^{k+1} (\hat t^n)^k.
\end{aligned}
\end{equation*}
Using Proposition~\ref{prop:QCorellator} we recognize here the second higher than three order summand of the Frobenius potential of Corollary~\ref{cor:FGpotential}.
\end{proof}
Note that the rescaling applied in the proof above can be written as the Givental's $S$--action. In what follows we will adopt the notation $S_0^A$ for it:
\begin{equation}\label{equation: S0M definition}
S_0^A :=
\left(
\begin{array}{c c c}
1 & \dots & 0 \\
\vdots & (c\tau+d) I_{n-2} & \vdots \\
0 & \dots & (c\tau+d)^2
\end{array}
\right).
\end{equation}
Applied to the potential of a Frobenius manifold $F(\textbf{t})$, $S_0^A$ acts by the change of variables. However this would change also the pairing of the Frobenius manifold, hence some additional factors are needed:
$$
S_0^A \cdot F = (c\tau + d)^{-2} \left( \frac{1}{2} (\tilde t^1)^2 \tilde t^n + \frac{1}{2} \tilde t^1 \sum_{k=2}^{n-1} \tilde t^k \tilde t^{n+1-k} + H(\tilde t^2, \dots, \tilde t^n) \right),
$$
where $\tilde \textbf{t} = S_0^A \textbf{t}$ and we assume $F$ to be written in the form of equation \eqref{eq:WDVV potential}.
\begin{corollary}\label{corollary: sigma-sigmaPrime}
Denote by $F^A_{\tau}$ and $F_{A \cdot \tau}$ the local expansions of $F^A$ and $F$ at the points $(0,\dots,0,\tau)$ and $(0,\dots,0,A\cdot \tau)$ respectively. For $\sigma := -c(c\tau+d)$, $\sigma^\prime := -c/(c\tau+d)$ and $S_0^A$ as above holds:
\begin{align*}
F^A_{\tau} &= \left( \hat S_0^A \right)^{-1} \cdot \hat R^{\sigma} \cdot F_{A \cdot \tau},
\\
F^A_{\tau} &= \hat R^{\sigma^\prime} \cdot \left( \hat S_0^A \right)^{-1}\cdot F_{A \cdot \tau}.
\end{align*}
\end{corollary}
\begin{proof}
The first equation is the Givental action form of Proposition~\ref{prop:Action via graphs} and was proved above. We only have to prove the second equation.
Sligthly altering the proof of the proposition above we write:
\begin{equation*}
\begin{aligned}
\frac{c}{8(ct^n + d)} (t^2t^{n-1} +& \dots + t^{n-1}t^2)^2 =
\\
\frac{-\sigma^\prime}{8(1 - \sigma^\prime \tilde t^n )} (t^2 t^{n-1} +& \dots + t^{n-1} t^2)^2 =
\\
&= - \sum_{k \ge 0} \frac{1}{| \mathrm{Aut}_2(\alpha, \beta)|} t^\alpha t^{\bar \alpha} t^{\beta} t^{\bar \beta} \left( \sigma^\prime \right)^{k+1} (\tilde t^n)^k,
\end{aligned}
\end{equation*}
and
\begin{equation*}
\begin{aligned}
{\mathcal K} := & (ct^n+d)^2 H \left(\frac{t^2}{ct^n+d}, \dots, \frac{t^{n-1}}{ct^n+d}, \frac{at^n + b}{ct^n+d} \right)
\\
&= \sum_{p,\alpha_1, \dots, \alpha_N} \frac{t^{\alpha_1} \dots t^{\alpha_N}}{| \mathrm{Aut}(\alpha)| p!}
\frac{ \left( \tilde t^n \right)^p }{ (c \tau + d)^p}
\frac{1}{(c \tau + d)^{N+p-2}}\frac{ H_{\alpha_1 \dots \alpha_N, n^p} }{ ( 1 + c (c \tau + d)^{-1}\tilde t^n)^{N+p-2}}
\\
&= \sum_{k \ge 0} \sum_{N+p \ge 3} \sum_{\alpha_1, \dots, \alpha_N} {N + k + p -3 \choose k} \frac{t^{\alpha_1} \dots t^{\alpha_N}}{| \mathrm{Aut}(\alpha)| p!} \frac{(\tilde t^n)^{p+k} H_{\alpha_1 \dots \alpha_N, n^p}}{(c\tau+d)^{N+2p-2} } \left( \sigma^\prime \right)^k.
\end{aligned}
\end{equation*}
Hence by rescaling the variables in the function $\mathcal K$ only we get the equality with the local expansion $F^A_{\tau}$. But this is exactly what we need because the action of $S_0^A$ does not change cubic terms of the potential that build up the $Q$--term computed above.
\end{proof}
\section{Graphical calculus of $\hat R^\sigma$ in higher genera}\label{section: HigherGenus}
In this section we prove Theorem~\ref{th:HigherGenera} using graphical calculus. Let $R^\sigma$ and $r^\sigma$ be as in Subsection~\ref{section: GraphsCalculus} with $\sigma \in {\mathbb{C}} \backslash \{0\}$, a complex parameter.
\subsection{Graphs of the $R^\sigma$ action in genus $g \ge 1$}
We start by analyzing the graphs contributing to the graphical calculus of Givental's $R^\sigma$ action in higher genera.
\begin{proposition}\label{prop: HigherGeneraGraphs}
Let $\mathcal Z$ be the partition function of the fixed CohFT $\Lambda_{g,k}$.
The graphs contributing to the genus $g$ potential of $R^\sigma \cdot \mathcal Z$ restricted to the small phase space are the following:
\\
\begin{tabular}{ll}
\raisebox{-.5\height}
{
\hbox{
\begin{tikzpicture}[y=1.1pt, x=1.1pt,yscale=-1]
\path[draw=black,line width=0.634pt] (42,79) -- (65,57);
\path[draw=black,line width=0.634pt] (42,27) -- (65,49);
\path[draw=black,line width=0.634pt] (37,34) -- (65,49);
\path[draw=black,line width=0.634pt] (96,26) -- (73,49);
\path[draw=black,line width=0.634pt] (96,79) -- (73,57);
\path[draw=black,line width=0.634pt] (102,48) -- (75,52);
\path[draw=black,line width=0.634pt] (102,57) -- (74,54);
\path[cm={{1.056,0.0,0.0,1.056,(11.36937,-49.61728)}}, draw=black, fill opacity=0.000,line width=0.634pt]
(60,97)arc(0:180:5)arc(-180:0:5) -- cycle;
\path[cm={{1.00447,0.10339,-0.16196,1.24445,(-8.229,-14.22512)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:196:22 and 20);
\path[cm={{-0.59464,-0.32663,0.27113,-1.18207,(114.92367,126.65456)}}, draw=black, dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:157:22 and 20);
\path[cm={{-0.35725,-0.57676,0.81721,-0.89609,(68.3994,154.12582)}}, draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61,64)arc(119:152:22 and 20);
\path[fill=black] (16,39) node[above right] {$e_{\alpha_2}$};
\path[fill=black] (30,92) node[above right] {$e_{\alpha_N}$};
\path[fill=black] (29,28) node[above right] {$e_{\alpha_1}$};
\path[fill=black] (98,28) node[above right] {$z e_1$};
\path[fill=black] (102,50) node[above right] {$z e_1$};
\path[fill=black] (128,46) node[above right] {$\Big \} p$ times};
\path[fill=black] (102,66) node[above right] {$e_n$};
\path[fill=black] (94,92) node[above right] {$e_n$};
\path[fill=black] (128,86) node[above right] {$\Big \} q$ times};
\end{tikzpicture}
}
}
&
\raisebox{-.5\height}
{
\hbox{
\begin{tikzpicture}[y=1.3pt,x=1.3pt,yscale=-1]
\draw[line width=0.634pt] (287,74) circle (40pt);
\draw[fill=white, fill opacity=1.000, line width=0.634pt] (287,44) circle (5pt);
\path[cm={{-0.24605,0.58095,-0.61658,-0.56766,(337.05741,30.01111)}},draw=black,
dash pattern=on 0.48pt off 0.48pt, line width=0.634pt] (61,64) arc (119:157:22 and 20);
\path[draw=black,line width=0.634pt] (277,26) -- (284,41.5);
\path[draw=black,line width=0.634pt] (297,26) -- (290,41.5);
\path[draw=black,line width=0.634pt] (287,48) -- (287,60);
\path[draw=black,line width=0.634pt] (265,60) -- (278,67.5);
\path[draw=black,line width=0.634pt] (313,75) -- (301,74);
\path[fill=black] (280.43311,6.3248138) node[above right] (text4339-5) {$m_1$};
\path[fill=black] (276,18) node[above right] (text4339-5) {$\overbrace{\hspace{12pt}}$};
\begin{scope}[cm={{-0.00698,0.99998,-0.99998,-0.00698,(366.2248,-218.50883)}}]
\path[cm={{0.75294,0.0,0.0,0.75294,(252.63785,-26.03223)}},draw=black,fill=white,line
join=round,miter limit=4.00,fill opacity=1.000,line width=0.634pt]
(60.0000,97.3622) .. controls (60.0000,100.1236) and (57.7614,102.3622) ..
(55.0000,102.3622) .. controls (52.2386,102.3622) and (50.0000,100.1236) ..
(50.0000,97.3622) .. controls (50.0000,94.6008) and (52.2386,92.3622) ..
(55.0000,92.3622) .. controls (57.7614,92.3622) and (60.0000,94.6008) ..
(60.0000,97.3622) -- cycle;
\path[cm={{-0.24605,0.58095,-0.61658,-0.56766,(344.13341,31.65506)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61.6611,64.7696) .. controls
(57.1711,62.4903) and (53.6626,58.8945) .. (51.7078,54.5685);
\path[draw=black,line width=0.634pt] (285.0158,28.9543) -- (291.7986,43.6211);
\path[draw=black,line width=0.634pt] (301.6869,28.9443) -- (295.8844,44.1313);
\end{scope}
\begin{scope}[cm={{0.57501,-0.81815,0.81815,0.57501,(54.42533,272.07611)}}]
\path[cm={{0.75294,0.0,0.0,0.75294,(252.63785,-26.03223)}},draw=black, fill=white, fill opacity=1.000,line width=0.634pt]
(60.0000,97.3622) .. controls (60.0000,100.1236) and (57.7614,102.3622) ..
(55.0000,102.3622) .. controls (52.2386,102.3622) and (50.0000,100.1236) ..
(50.0000,97.3622) .. controls (50.0000,94.6008) and (52.2386,92.3622) ..
(55.0000,92.3622) .. controls (57.7614,92.3622) and (60.0000,94.6008) ..
(60.0000,97.3622) -- cycle;
\path[cm={{-0.24605,0.58095,-0.61658,-0.56766,(344.13341,31.65506)}},draw=black,dash pattern=on 0.48pt off 0.48pt,
fill opacity=0.000,line width=0.634pt] (61.6611,64.7696) .. controls
(57.1711,62.4903) and (53.6626,58.8945) .. (51.7078,54.5685);
\path[draw=black,line width=0.634pt] (285.0158,28.9543) -- (291.7986,43.6211);
\path[draw=black,line width=0.634pt] (301.6869,28.9443) -- (295.8844,44.1313);
\end{scope}
\path[cm={{-1.28188,-1.27443,1.5773,-1.46237,(205.71404,327.23285)}},draw=black,dash pattern=on 0.29pt off 0.49pt,
fill opacity=0.000, line width=0.634pt]
(50,119) arc (145:337:25 and 21);
\path[fill=black] (281.5813,25.27486) node[above right] {$ze_1$};
\path[fill=black] (340.35062,76.041534) node[above right] {$ze_1$};
\path[fill=black] (225.06505,50.034634) node[above right] {$ze_1$};
\path[fill=black] (203,50) node[above right] {$m_p$};
\path[fill=black] (217,52) node[above right] {$\big\{$};
\path[fill=black] (287,62) node[above right] {$e_n$};
\path[fill=black] (296,85) node[above right] {$e_n$};
\path[fill=black] (264,75) node[above right] {$e_n$};
\path[fill=black] (259,52) node[above right] {$e_1$};
\path[fill=black] (267,47) node[above right] {$e_1$};
\path[fill=black] (245,73) node[above right] {$e_1$};
\path[fill=black] (302,48) node[above right] {$e_1$};
\path[fill=black] (316,64) node[above right] {$e_1$};
\path[fill=black] (314,98) node[above right] {$e_1$};
\path[fill=black] (358,85) node[above right] {$\Big\}$};
\path[fill=black] (367,78) node[above right] {$m_2$};
\end{tikzpicture}
}
}
\end{tabular}
\\
where $p,q \ge 0$, $m_i \ge 0$, $\sum_{i=1}^p m_i + p = k \ge 1$.
\begin{notation}
We will denote the first series by $\Gamma_{p,q;\alpha}^g$ and the second series by $\Gamma_{k,p}^\circ,$.
\end{notation}
\end{proposition}
\begin{proof}
As in Proposition~\ref{prop:GraphFamilies} the only possible decorations of leaves restricted to the small phase space (recall Proposition \ref{prop: small phase space leaves}) are:
$$
\mathcal L_s = \sigma e_1 t^{0,n} z^{1} + \sum_{\alpha = 1}^n e_\alpha t^{0, \alpha} z^0, \quad \mathcal L_0 = 0.
$$
The edges are decorated by:
$$
\mathcal E = - (r^\sigma)^{11} e_1 \otimes e_1.
$$
Consider all possible genus $g$ graphs with the decoration given. Contraction of leaves in the vertices are given by the following correlators that can be simplified using the Dilaton equation:
$$
\left\langle \left(\tau_0(e_1) \right)^k \left(\tau_1(e_1)\right)^l \prod_{i=1}^N \tau_0(\alpha_i) \right\rangle_{\tilde g}
=
\frac{(2\tilde g - 2 + k +l -1 + N)!}{(2\tilde g - 2 + k + N)!} \left\langle \left(\tau_0(e_1) \right)^k \prod_{i=1}^N \tau_0(\alpha_i) \right\rangle_{\tilde g}.
$$
It is clear from the axioms of the unital CohFT that the correlator on the right hand side vanishes unless $k=0$ or $\tilde g = 0$.
Taking graphs with only one vertex we get exactly the first series of the proposition.
In order to build genus $g$ graph having more than one vertex we have to use vertices with at least two half-edges decorated by $e_1$. By the reasoning above such vertices need to be of genus $0$. At the same time we know that in genus $0$ the correlators with more than two $\tau_0(e_1)$ insertions vanish. Hence we can only build the genus $1$ graph using the vertices of the lower genus.
Due to the reasoning of Proposition~\ref{prop:GraphFamilies} the second series represents an arbitrary decoration of such graphs.
\end{proof}
\begin{remark}
According to the decoration $\mathcal V[k]$ assigned to the vertex its contraction with the half-edges could give the contribution to the correlators of any genus. It is clear however from the proof of the proposition that the graphs from $\Gamma_{p,q;\alpha}^g$ contribute to arbitrary genus $g$ with the vertex decorated by $\mathcal V_g$ only and the graphs from $\Gamma^\circ_{k,p}$ contribute to the genus $1$ only with all the vertices decorated by $\mathcal V_0$ only.
\end{remark}
In order to use graphical calculus we have to compute the automorphisms of the graphs presented above. It is not a problem at all for the series $\Gamma^g_{p,q;\alpha}$ but some additional analysis is needed for the graphs from $\Gamma^\circ_{k,p}$.
\subsubsection{Graphs counting}
By setting $a_i = m_i + 1$ we associate to every graph $\gamma \in \Gamma^\circ_{k,p}$ the ordered sequence of numbers such that $a_1 + \dots + a_p = k$ and $k \ge a_i \ge 1$ . Obviously it is defined up to cyclic shift. Such partitions of $k$ are called cyclic compositions.
\begin{definition}
For $\gamma \in \Gamma^\circ_{k,p}$ define by $ \mathrm{Cy} (\gamma)$ corresponding cyclic composition $(1+m_1) + \dots + (1 + m_p) = k$.
\end{definition}
\begin{definition}
For $p,k \in \mathbb N$ and $p \le k$ define:
$$
\CyN{k}{p} := \mid \Gamma^\circ_{k,p}\mid,
\quad
\CyN{k}{p}_{prim} := \# \{ \gamma \in \Gamma^\circ_{k,p} \mid | \mathrm{Aut}(\gamma)| = 1 \}.
$$
\end{definition}
These numbers are called cyclic compositions and primitive cyclic compositions numbers respectively (cf. \cite{KR,FS}).
\begin{proposition}[Theorem~2 in \cite{KR}]
The number of cyclic compositions is equal to:
\begin{equation}\label{eq:cyclic compositions}
\CyN{n}{p} = \frac{1}{n} \sum_{a | \mathrm{gcd} (n,p)} \varphi(a) {n/a \choose p/a},
\end{equation}
where $\varphi$ is Euler's totient function.
\end{proposition}
We adopt the proof of \cite{KR} to compute the number of primitive cyclic compositions.
\begin{proposition}
The number of primitive cyclic compositions is equal to:
\begin{equation}\label{eq:primitive cyclic compositions}
\CyN{n}{p}_{prim} = \frac{1}{n} \sum_{a | \mathrm{gcd} (n,p)} \mu(a) {n/a \choose p/a},
\end{equation}
where $\mu(a)$ is M\"obius function.
\end{proposition}
\begin{proof}
Consider the generating function:
$$
PC(z,u) := \sum \CyN{k}{l}_{prim} z^k u^l.
$$
It can be expressed (Appendix A.4 in \cite{FS}):
$$
PC(z,u) = \sum_{k \ge 1} \frac{\mu(k)}{k} \log \left(1 - \frac{u^k z^k}{1 - z^k} \right)^{-1}.
$$
Expanding further we get:
$$
\begin{aligned}
PC (z,u) &= \sum_{k \ge 1} \frac{\mu(k)}{k} \left( - \sum_{m\ge 1} \frac{z^{km}}{m} + \sum_{m\ge 1} \frac{(1+u^k)z^{km}}{m} \right)
\\
&= \sum_{k \ge 1} \sum_{m\ge 1} \frac{\mu(k)}{km} z^{km} \left((1+u^k) -1 \right).
\end{aligned}
$$
Setting $l = km$ one gets the formula.
\end{proof}
\begin{lemma}
\begin{equation}\label{eq:Cyclic graphs counting}
\sum_{p = 1}^q \sum_{\gamma \in \Gamma^0_{q,p}} \frac{(-1)^p}{ \mathrm{Aut}( \mathrm{Cy} (\gamma))} = - \frac{1}{q}.
\end{equation}
\end{lemma}
\begin{proof}
All automorphisms of $ \mathrm{Cy} (\gamma)$ for $\gamma \in \Gamma^\circ_{q,p}$ are given by the shifts, which correspond to the rotations of the graph $\gamma$. Therefore we can write:
$$
\sum_{\gamma \in \Gamma^\circ_{q,p}} \frac{1}{ \mathrm{Aut}( \mathrm{Cy} (\gamma))} = \sum_{d|p, \ d|q} \frac{1}{d} \CyN{q/d}{p/d}_{prim}.
$$
Using this and \eqref{eq:primitive cyclic compositions} we write:
$$
\sum_{p = 1}^q \sum_{\gamma \in \Gamma^\circ_{q,p}} \frac{(-1)^p}{ \mathrm{Aut}( \mathrm{Cy} (\gamma))} = \sum_{p = 1}^q (-1)^p \sum_{d|p, \ d|q} \frac{1}{q} \sum_{a | \mathrm{gcd} (q/d,p/d)} \mu(a) {q/ad \choose p/ad}.
$$
Changing the summation by introducing $l = ad$ we get:
$$
\sum_{p = 1}^q \sum_{\gamma \in \Gamma^\circ_{q,p}} \frac{(-1)^p}{ \mathrm{Aut}( \mathrm{Cy} (\gamma))} = \frac{1}{q} \sum_{p = 1}^q (-1)^p \sum_{l | \mathrm{gcd} (q,p)} {q/l \choose p/l} \sum_{b | l} \mu(b).
$$
Due to:
$$
\sum_{d | n} \mu(d) =
\begin{cases}
1 & \mbox{ if } n=1
\\
0 & \mbox{ if } n>1.
\end{cases}
$$
we conclude:
$$
\sum_{p = 1}^q \sum_{\gamma \in \Gamma^\circ_{q,p}} \frac{(-1)^p}{ \mathrm{Aut}( \mathrm{Cy} (\gamma))} = \frac{1}{q} \sum_{p = 1}^q (-1)^p {q \choose p} = - \frac{1}{q}.
$$
\end{proof}
\subsection{From $\hat R^c \cdot \mathcal Z$ to $F^A_g$}
From graphical calculus we get:
\begin{proposition}
The correlators of the genus $g$ potential of $R^\sigma \cdot \mathcal Z$ read:
\begin{equation}\label{eq:Graphical counting in Fg}
\begin{aligned}
\langle \tau_0 & (\alpha_1) \dots \tau_0(\alpha_N) \left(\tau_0(e_n)\right)^q\rangle_g^G =
\\
& \sum_{p+k=q} \frac{q!}{p!} {N+q+2g-3 \choose k} \langle \tau_0(\alpha_1) \dots \tau_0(\alpha_N) \left(\tau_0(e_n)\right)^p\rangle_g \cdot \sigma^k -\delta_{g,1} \frac{\sigma^q}{q} .
\end{aligned}
\end{equation}
\end{proposition}
\begin{proof}
We compute the contribution of the graphs listed in Proposition~\ref{prop: HigherGeneraGraphs} (see also the remark after it).
For $g \ge 2$ these are only $\gamma \in \Gamma^g_{k,p; \alpha}$. The contribution of these graphs reads:
\begin{equation*}
\begin{aligned}
\langle \tau_0(\alpha_1) & \dots \tau_0(\alpha_N) \left(\tau_0(e_n)\right)^q\rangle_g^G =
\\
& \sum_{p+k=q} \frac{q!}{p!} \langle \tau_0(\alpha_1) \dots \tau_0(\alpha_N) \left(\tau_0(e_1)\right)^p \left(\tau_1(e_1)\right)^k\rangle_g \cdot \sigma^k
\end{aligned}
\end{equation*}
Using the Dilaton equation we get:
\begin{equation*}
\begin{aligned}
\langle \tau_0(\alpha_1) & \dots \tau_0(\alpha_N) \left(\tau_0(e_n)\right)^q\rangle_g^G =
\\
& q! \sum_{p+k=q} \frac{\left( N + k+ p + 2g-3 \right) !}{p! \left( N + p + 2g - 3\right)!} \langle \tau_0(\alpha_1) \dots \tau_0(\alpha_N) \left(\tau_0(e_n)\right)^p \rangle_g \cdot \sigma^k.
\end{aligned}
\end{equation*}
For $g=1$ we get the contribution as above with the additional contribution of the graph $\gamma^\circ \in \Gamma^\circ_{k,p}$.
Note that the relation between $\mathrm{Aut}(\gamma^\circ)$ and $\mathrm{Aut}( \mathrm{Cy} (\gamma^\circ))$ is given by:
$$
\mathrm{Aut}(\gamma^\circ) = \mathrm{Aut}( \mathrm{Cy} (\gamma^\circ)) \prod_{i=1}^p m_i !.
$$
Using the Dilaton equation for the graph $\gamma^\circ$ the factor $\prod_{i=1}^p m_i$ cancels outs and the contribution of $\gamma^\circ \in \Gamma^\circ_{k,p}$ for $g=1$ reads:
\begin{equation*}
\begin{aligned}
\langle \tau_0(\alpha_1) & \dots \tau_0(\alpha_N) \left(\tau_0(e_n)\right)^q\rangle_1^G =
\\
q! & \sum_{p+k=q} \frac{\left( N + k+ p + 1 \right) !}{p! \left( N + p + 1 \right)!} \langle \tau_0(\alpha_1) \dots \tau_0(\alpha_N) \left(\tau_0(e_n)\right)^p \rangle_1 \cdot \sigma^k
\\
& + \sigma^{q} \sum_{k=1}^{q} \sum_{\gamma \in \Gamma^0_{q,k}} \frac{(-1)^k}{\mathrm{Aut}( \mathrm{Cy} (\gamma))}.
\end{aligned}
\end{equation*}
Because of the graphs counting formula \eqref{eq:Cyclic graphs counting} the proposition follows.
\end{proof}
We illustrate the graph computations of the proposition above in the Figure~\ref{figure: graphs counting} on page \pageref{figure: graphs counting}.
\begin{figure}
\hbox{
\begin{tikzpicture}[y=1.3pt,x=1.3pt,yscale=-1]
\draw[line width=0.634pt] (23,23) circle (23pt);
\path[draw=black,line width=0.634pt] (23,5) -- (23,-15);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23,5) circle (4pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (23,41) circle (2pt);
\path[fill=black] (23,-5) node[above right] {$e_n$};
\path[fill=black] (23,75) node[] {$-\sigma$};
\draw[line width=0.634pt] (23+120,23) circle (23pt);
\path[draw=black,line width=0.634pt] (23+120,5) -- (23+120,-15);
\path[draw=black,line width=0.634pt] (23+120,41) -- (23+120,61);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+120,5) circle (4pt);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+120,41) circle (4pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (5+120,23) circle (2pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (41+120,23) circle (2pt);
\path[fill=black] (23+120,-5) node[above right] {$e_n$};
\path[fill=black] (23+120,60) node[above right] {$e_n$};
\path[fill=black] (23+120,75) node[] {$\dfrac{\sigma^2}{2}$};
\draw[line width=0.634pt] (23+60+120,23) circle (23pt);
\path[draw=black,line width=0.634pt] (23+60+120,5) -- (13+60+120,-15);
\path[draw=black,line width=0.634pt] (23+60+120,5) -- (33+60+120,-15);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+60+120,5) circle (4pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (23+60+120,41) circle (2pt);
\path[fill=black] (13+60+120,-5) node[above right] {$e_n$};
\path[fill=black] (32+60+120,-5) node[above right] {$z e_1$};
\path[fill=black] (23+60+120,75) node[] {$-\sigma^2$};
\end{tikzpicture}
}
\hbox{
\begin{tikzpicture}[y=1.3pt,x=1.3pt,yscale=-1]
\draw[line width=0.634pt] (23,23) circle (23pt);
\path[draw=black,line width=0.634pt] (23,5) -- (23,-15);
\path[draw=black,line width=0.634pt] (6,29) -- (-9,38);
\path[draw=black,line width=0.634pt] (40,29) -- (55,38);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23,5) circle (4pt);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (6,29) circle (4pt);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (40,29) circle (4pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (23,41) circle (2pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (8,13) circle (2pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (38,13) circle (2pt);
\path[fill=black] (23,-5) node[above right] {$e_n$};
\path[fill=black] (47,37) node[above right] {$e_n$};
\path[fill=black] (-16,37) node[above right] {$e_n$};
\path[fill=black] (23,75) node[] {$-\dfrac{\sigma^3}{3}$};
\draw[line width=0.634pt] (23+60,23) circle (23pt);
\path[draw=black,line width=0.634pt] (23+60,5) -- (13+60,-15);
\path[draw=black,line width=0.634pt] (23+60,5) -- (33+60,-15);
\path[draw=black,line width=0.634pt] (23+60,41) -- (23+60,61);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+60,5) circle (4pt);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+60,41) circle (4pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (5+60,23) circle (2pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (41+60,23) circle (2pt);
\path[fill=black] (13+60,-5) node[above right] {$e_n$};
\path[fill=black] (32+60,-5) node[above right] {$z e_1$};
\path[fill=black] (23+60,60) node[above right] {$e_n$};
\path[fill=black] (23+60,75) node[] {$-\sigma^3$};
\draw[line width=0.634pt] (23+120,23) circle (23pt);
\path[draw=black,line width=0.634pt] (23+120,5) -- (8+120,-15);
\path[draw=black,line width=0.634pt] (23+120,5) -- (30+120,-15);
\path[draw=black,line width=0.634pt] (23+120,5) -- (40+120,-15);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+120,5) circle (4pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (23+120,41) circle (2pt);
\path[fill=black] (9+120,-7) node[above right] {$e_n$};
\path[fill=black] (28+120,-12) node[above right] {$z e_1$};
\path[fill=black] (40+120,-5) node[above right] {$z e_1$};
\path[fill=black] (23+120,75) node[] {$\dfrac{2\sigma^3}{2}$};
\end{tikzpicture}
}
\hbox{
\begin{tikzpicture}[y=1.3pt,x=1.3pt,yscale=-1]
\draw[line width=0.634pt] (23,23) circle (23pt);
\path[draw=black,line width=0.634pt] (9,41-29) -- (9-15,41-29-14);
\path[draw=black,line width=0.634pt] (37,41-29) -- (37+15,41-29-14);
\path[draw=black,line width=0.634pt] (9,41-7) -- (9-15,41-7+14);
\path[draw=black,line width=0.634pt] (37,41-7) -- (37+15,41-7+14);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (9,41-29) circle (4pt);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (37,41-29) circle (4pt);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (9,41-7) circle (4pt);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (37,41-7) circle (4pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (23,41) circle (2pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (23,5) circle (2pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (5,23) circle (2pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (41,23) circle (2pt);
\path[fill=black] (-7,0) node[above right] {$e_n$};
\path[fill=black] (39,0) node[above right] {$e_n$};
\path[fill=black] (47,47) node[above right] {$e_n$};
\path[fill=black] (-16,47) node[above right] {$e_n$};
\path[fill=black] (23,80) node[] {$\dfrac{\sigma^4}{4}$};
\draw[line width=0.634pt] (23+60+20,23) circle (23pt);
\path[draw=black,line width=0.634pt] (23+60+20,5) -- (13+60+20,-15);
\path[draw=black,line width=0.634pt] (23+60+20,5) -- (33+60+20,-15);
\path[draw=black,line width=0.634pt] (6+60+20,29) -- (-9+60+20,38);
\path[draw=black,line width=0.634pt] (40+60+20,29) -- (55+60+20,38);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+60+20,5) circle (4pt);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (6+60+20,29) circle (4pt);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (40+60+20,29) circle (4pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (23+60+20,41) circle (2pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (8+60+20,13) circle (2pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (38+60+20,13) circle (2pt);
\path[fill=black] (13+60+20,-5) node[above right] {$e_n$};
\path[fill=black] (32+60+20,-5) node[above right] {$z e_1$};
\path[fill=black] (47+60+20,37) node[above right] {$e_n$};
\path[fill=black] (-16+60+20,37) node[above right] {$e_n$};
\path[fill=black] (23+60+20,80) node[] {$-\sigma^4$};
\draw[line width=0.634pt] (23+60+60+20,23) circle (23pt);
\path[draw=black,line width=0.634pt] (23+60+60+20,5) -- (13+60+60+20,-15);
\path[draw=black,line width=0.634pt] (23+60+60+20,5) -- (33+60+60+20,-15);
\path[draw=black,line width=0.634pt] (23+60+60+20,41) -- (13+60+60+20,41+20);
\path[draw=black,line width=0.634pt] (23+60+60+20,41) -- (33+60+60+20,41+20);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+60+60+20,5) circle (4pt);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+60+60+20,41) circle (4pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (5+60+60+20,23) circle (2pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (41+60+60+20,23) circle (2pt);
\path[fill=black] (13+60+60+20,-5) node[above right] {$e_n$};
\path[fill=black] (32+60+60+20,-5) node[above right] {$z e_1$};
\path[fill=black] (13+60+60+20,41+25) node[above right] {$e_n$};
\path[fill=black] (32+60+60+20,41+25) node[above right] {$z e_1$};
\path[fill=black] (23+60+60+20,80) node[] {$\dfrac{\sigma^4}{2}$};
\draw[line width=0.634pt] (23+120+60+20,23) circle (23pt);
\path[draw=black,line width=0.634pt] (23+120+60+20,5) -- (8+120+60+20,-15);
\path[draw=black,line width=0.634pt] (23+120+60+20,5) -- (30+120+60+20,-15);
\path[draw=black,line width=0.634pt] (23+120+60+20,5) -- (40+120+60+20,-15);
\path[draw=black,line width=0.634pt] (23+120+60+20,41) -- (23+120+60+20,61);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+120+60+20,5) circle (4pt);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+120+60+20,41) circle (4pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (5+120+60+20,23) circle (2pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (41+120+60+20,23) circle (2pt);
\path[fill=black] (9+120+60+20,-7) node[above right] {$e_n$};
\path[fill=black] (23+120+60+20,60) node[above right] {$e_n$};
\path[fill=black] (28+120+60+20,-12) node[above right] {$z e_1$};
\path[fill=black] (40+120+60+20,-5) node[above right] {$z e_1$};
\path[fill=black] (23+120+60+20,80) node[] {$\dfrac{2\sigma^4}{2}$};
\draw[line width=0.634pt] (23+180+60+20,23) circle (23pt);
\path[draw=black,line width=0.634pt] (23+180+60+20,5) -- (8+180+60+20,-15);
\path[draw=black,line width=0.634pt] (23+180+60+20,5) -- (30+180+60+20,-15);
\path[draw=black,line width=0.634pt] (23+180+60+20,5) -- (40+180+60+20,-15);
\path[draw=black,line width=0.634pt] (23+180+60+20,5) -- (45+180+60+20,-5);
\draw[line width=0.634pt, fill=white, fill opacity=1.000] (23+180+60+20,5) circle (4pt);
\draw[line width=0.634pt, fill=black, fill opacity=1.000] (23+180+60+20,41) circle (2pt);
\path[fill=black] (9+180+60+20,-7) node[above right] {$e_n$};
\path[fill=black] (25+180+60+20,-14) node[above right] {$z e_1$};
\path[fill=black] (40+180+60+20,-8) node[above right] {$z e_1$};
\path[fill=black] (42+180+60+20,2) node[above right] {$z e_1$};
\path[fill=black] (23+180+60+20,80) node[] {$-\dfrac{3!\sigma^4}{3!}$};
\end{tikzpicture}
}
\caption{The contribution of the graphs from $\Gamma_{k,p}^\circ$ in Equation~\eqref{eq:Graphical counting in Fg}. We do not decorate the half--edges denoting by the black dot the gluing of two $(e_1,e_1)$ half--edges by $r^{11}$. Summing up the numbers beneath the graphs one gets the contribution $-\sigma^q/q$ as in Equation~\eqref{eq:Graphical counting in Fg}.}
\label{figure: graphs counting}
\end{figure}
\subsection{Proof of Theorem~\ref{th:HigherGenera}}\label{subsection: proofOfTheorem2}
Let
$
A = \left(
\begin{array}{c c}
a & b \\
c & d
\end{array}
\right)
\in {\rm SL}(2, \mathbb C)
$. With the help of graphical calculus we give the formulae for the action of $A$ on the higher genus potential of the CohFT partition function.
\begin{proof}
We show explicitly that the coefficients of $F_g^A(\textbf{t})$ series expansion coincide with the correlators obtained by Givental's group action.
As in the proof of Proposition~\ref{prop:Action via graphs} consider the constants:
$$
\gamma := \frac{a \tau + b}{c \tau + d}, \quad \mu := \tau.
$$
Write the series expansion of the function $F_g^A$ at the point $(0, \dots, 0, \gamma)$. We have:
\begin{equation*}
\begin{aligned}
\mathcal K := &(ct^n+d)^{2-2g} F_g \left(\frac{t^2}{ct^n+d}, \dots, \frac{t^{n-1}}{ct^n+d}, \frac{at^n + b}{ct^n+d} \right)
\\
&= (ct^n+d)^{2-2g} \sum \frac{t^{\alpha_1} \dots t^{\alpha_N}}{|\mathrm{Aut}(\alpha)| p!} \left( \frac{at^n + b}{ct^n + d} - \gamma \right)^p \frac{(F_g)_{\alpha_1 \dots \alpha_N, n^p}}{(ct^n + d)^N},
\end{aligned}
\end{equation*}
Let $\tilde t^n = t^n - \mu$. The series expansion rewrites:
\begin{align*}
\mathcal K &= \sum_{p,\alpha_1, \dots, \alpha_N} \frac{t^{\alpha_1} \dots t^{\alpha_N}}{|\mathrm{Aut}(\alpha)| p!}
\frac{ \left( \tilde t^n \right)^p }{ (c \tilde t^n + (c\tau + d))^{N+p+2g-2}} \frac{(F_g)_{\alpha_1 \dots \alpha_N, n^p}}{(c \tau + d)^p}
\\
&= \sum_{p,\alpha_1, \dots, \alpha_N} \frac{t^{\alpha_1} \dots t^{\alpha_N}}{|\mathrm{Aut}(\alpha)| p!}
\frac{ \left( \tilde t^n \right)^p }{ (c (c\tau + d)^{-1} \tilde t^n + 1)^{N+p+2g-2}} \frac{(F_g)_{\alpha_1 \dots \alpha_N, n^p}}{(c \tau + d)^{N + 2p + 2g - 2}}.
\end{align*}
Taking $\hat t^{\alpha_k} := t^{\alpha_k} / (c\tau + d)$, $\hat t^n := \tilde t^n / (c\tau + d)^2$ and $\sigma = -c(c \tau + d)$ we rewrite:
\begin{align*}
\mathcal K = (c\tau + d)^{2-2g} \sum_{p,\alpha_1, \dots, \alpha_N} & \frac{\hat t^{\alpha_1} \dots \hat t^{\alpha_N}}{|\mathrm{Aut}(\alpha)| p!}
\frac{ \left( \hat t^n \right)^p }{ (1 -\sigma \hat t^n )^{N+p+2g-2}} (F_g)_{\alpha_1 \dots \alpha_N, n^p}
\\
= (c\tau + d)^{2-2g} \sum_{k \ge 0} & \sum_{N + p + 2g-3 \ge 0} \sum_{\alpha_1, \dots, \alpha_N}
\\
& {N + p+k +2g -3 \choose k} \frac{\hat t^{\alpha_1} \dots \hat t^{\alpha_N}}{|\mathrm{Aut}(\alpha)| p!} (\hat t^n)^{p+k} (F_g)_{\alpha_1 \dots \alpha_N, n^p} \ \sigma^k.
\end{align*}
Using equation~\eqref{eq:Graphical counting in Fg} we get in the new variables exactly the Givental-transformed genus $g$ potential for $g \ge 2$. For the genus $1$ potential $F_1^A$ we get also the additional term:
$$
- \log \left( \frac{ct^n + d}{c\tau+d} \right) = - \log \left(1 + c (c\tau+d)\hat t^n \right) = -\sum_{k=1}^\infty \frac{\sigma^k}{k} (\hat t^n)^k.
$$
This coincides with formula~\eqref{eq:Graphical counting in Fg} of $R^\sigma$ -- transformed genus $1$ potential correlators obtained by graphical calculus.
\end{proof}
Extend the action of $S_0^A$ to the partition function $\mathcal{Z} = \expandafter (\sum_{g \ge 0} \hbar^{g-1} F_g(\textbf{t}))$ by the following:
\begin{equation}\label{equation: S0A in higher genera}
\hat S_0^A \cdot \mathcal{Z}(\hbar,\textbf{t}) = \mathcal{Z} \left((c\tau+d)^2 \hbar, \tilde \textbf{t} \right), \quad \tilde \textbf{t} = S_0^A \textbf{t}.
\end{equation}
Similarly to Corollary~\ref{corollary: sigma-sigmaPrime} one can show the following corollary.
\begin{corollary}\label{corollary: sigma-sigmaPrime higher genera}
Denote by $\left( F^A_g \right)_{\tau}$ and $\left( F_g \right)_{A \cdot \tau}$ the local expansions of $F^A_g$ and $F_g$ at the points $(0,\dots,0,\tau)$ and $(0,\dots,0,A\cdot \tau)$ respectively. For $\sigma := -c(c\tau+d)$, $\sigma^\prime := -c/(c\tau+d)$ and $S_0^A$ as in \eqref{equation: S0M definition} holds:
\begin{align*}
\left( F^A_g \right)_{\tau} &= \left( \hat S_0^A \right)^{-1} \cdot \hat R^{\sigma} \cdot \left( F_g\right)_{A \cdot \tau},
\\
\left( F^A_g \right)_{\tau} &= \hat R^{\sigma^\prime} \cdot \left( \hat S_0^A \right)^{-1}\cdot \left( F_g \right)_{A \cdot \tau}.
\end{align*}
\end{corollary}
\begin{proof}
This is obtained by some easy computation similar to the proof of Corollary~\ref{corollary: sigma-sigmaPrime}
\end{proof}
\section{Coordinate--free form of the $ {\rm SL(2,\mathbb C)} $ action}\label{section: coordinateFree}
Theorems~\ref{th:FrobeniusManifold} and \ref{th:HigherGenera} make use of the functions expanded at the certain points. We would like to get rid of this special requirement. The statement of the above mentioned theorems can be rewritten as:
\begin{equation*}
\hat S_0^A \cdot \hat S^{\tau} \cdot F^A_g = \hat R^\sigma \cdot \hat S^{A \cdot \tau} \cdot F_g,
\end{equation*}
for
$$
A = \begin{pmatrix}
a & b \\ c & d
\end{pmatrix}
\quad \text{and} \quad \sigma = -c(c\tau+d),
$$
with $S_0^A$, $S^{\tau}$ and $S^{A\cdot \tau}$ defined in Section~\ref{section: GiventalsQuantization} and above. This equation suggests a Givental analogue of the $ {\rm SL(2,\mathbb C)} $--action to be defined by:
\begin{align*}
\hat A_G := &\left( \hat S^{\tau}\right)^{-1} \left( \hat S_0^A \right)^{-1} \cdot \hat R^\sigma \cdot \hat S^{A \cdot \tau}
\\
& = \hat S^{-\tau} \left( \hat S_0^A \right)^{-1} \cdot \hat R^\sigma \cdot \hat S^{A \cdot \tau}.
\end{align*}
It's easy to check by hands that $ {\rm SL(2,\mathbb C)} $ action defined in the analytic form \eqref{eq:SLAction} and \eqref{equation: FgA definition} is indeed a group action. However this is not clear on the Givental's side.
In what follows we show that the action $\hat A_G$ forms a group. This is an interesting result because the upper--triangular and lower--triangular actions form two different groups, whose elements do not commute, while the action $\hat A_G$ makes use of both.
\begin{proposition}
For any $A \in {\rm SL(2,\mathbb C)} $ the following formula holds:
$$
\hat A_G \cdot \left( A^{-1}\right)\widehat{\ }_G = \mathrm{Id}.
$$
\end{proposition}
\begin{proof}
Let $A = \begin{pmatrix}a & b \\ c & d\end{pmatrix} \in {\rm SL(2,\mathbb C)} $ then $A^{-1} = \begin{pmatrix} d & -b \\ -c & a\end{pmatrix}$.
Denote by $\sigma_A$ and $\sigma^\prime_A$ the numbers:
$$
\sigma_A := - c(c\tau_0 + d) \quad \sigma^\prime_A := - c/(c\tau_0+d).
$$
Let $A\cdot \tau_0 = \tau_1$. Then we have:
$$
\sigma_A = -\sigma^\prime_{A^{-1}}, \ \sigma^\prime_A = -\sigma_{A^{-1}}, \quad S_0^A = \left( S_0^{A^{-1}} \right)^{-1}.
$$
Using these equalities together with Corollary~\ref{corollary: sigma-sigmaPrime} the composition of Givental's actions reads:
\begin{align*}
\hat A_G \cdot \left( A^{-1}\right)\widehat{\ }_G & = \hat S^{-\tau_0} \left( \hat S_0^A \right)^{-1} \cdot \hat R^{\sigma_A} \cdot \hat S^{A \cdot \tau_0} \cdot \hat S^{-\tau_1} \left( \hat S_0^{A^{-1}} \right)^{-1} \cdot \hat R^{\sigma_{A^{-1}}} \cdot \hat S^{A^{-1} \cdot \tau_1}
\\
& = \hat S^{-\tau_0} \left( \hat S_0^A \right)^{-1} \cdot \hat R^{\sigma_A} \cdot \left( \hat S_0^{A^{-1}} \right)^{-1} \cdot \hat R^{\sigma_{A^{-1}}} \cdot \hat S^{\tau_0}
\\
& = \hat S^{-\tau_0} \left( \hat S_0^A \right)^{-1} \cdot \hat R^{\sigma_A} \cdot \hat R^{\sigma^\prime_{A^{-1}}} \cdot \left( \hat S_0^{A^{-1}} \right)^{-1} \cdot \hat S^{\tau_0} = \mathrm{Id}.
\end{align*}
\end{proof}
\begin{proposition}
Let $A,B \in {\rm SL(2,\mathbb C)} $ then $\hat A_G \cdot \hat B_G = \hat C_G$ for $C = BA$.
\end{proposition}
\begin{proof}
Let $A = \begin{pmatrix}
a_{11} & a_{21} \\ a_{21} & a_{22}
\end{pmatrix}$,
$B = \begin{pmatrix}
b_{11} & b_{21} \\ b_{21} & b_{22}
\end{pmatrix}$ and $\tau_1 = A\cdot \tau_0$. Then we can write:
$$
\hat A_G \cdot \hat B_G = S^{- \tau_0} \left( S_0^A \right)^{-1} R^{\sigma_A} \left(S_0^B \right)^{-1} R^{\sigma_B} S^{B\tau_1}.
$$
Commuting $\left(S_0^B \right)^{-1}$ with $R^{\sigma_A}$ as in Corollary~\ref{corollary: sigma-sigmaPrime} and observing that $S_0^B S_0^A = S_0^{C}$.
$$
\hat A_G \cdot \hat B_G = S^{- \tau_0} \left( S_0^C \right)^{-1} R^{\tilde \sigma + \sigma_B} S^{C\tau_0},
$$
for $\tilde \sigma = \sigma_A (b_{21} \tau_1 + b_{22})^2$. Some easy but long computations give the needed equality $\tilde \sigma + \sigma_B = \sigma_C$.
\end{proof}
\subsection{The calibration of $F^A$}\label{section: calibration}
For an arbitrary Givental's lower-triangular group element $S(z) = 1 + \sum_{k \ge 1} S_{-k}z^k$ the action of $\hat S$ on the partition function can be written as (see \cite{G, L}):
$$
\hat S \cdot \mathcal Z(\textbf{t}) = \mathcal Z(\tilde \textbf{t}(\textbf{t})) + \frac{1}{2} W(\textbf{t}),
$$
for the certain function $W(\textbf{t})$, quadratic in $\textbf{t}$ and the certain linear change of variables $\tilde \textbf{t}(\textbf{t})$.
Let $q^k, k \ge 0$ be infinite set of vectors defined in coordinates by $\left(q^k\right)^l = t^{k,l} - \delta_{k,1} $. We have:
$$
\tilde t^{k,l} = \left(\sum_{p=k}^\infty S_{p-k} q^p \right)^l,
$$
where the expression under the summation is understood as the action of the operator on the vector and the superscript $l$ takes the corresponding coordinate of the vector obtained. In particular restricted to the small phase space we get $\tilde t^{0,l} = t^{0,l} - \left(S_{-1}\right)_m^l {\bf 1}$, where ${\bf 1} = (1,\dots,1)$.
The function $W$ is defined by:
$$
W(\textbf{t}) := \sum_{i,j} (W_{i,j} q^i, q^j), \quad \sum_{i,j} W_{i,j}z^{-i} w^{-j} = \frac{S(z)S^\ast(w) - 1}{z^{-1}+w^{-1}}.
$$
Note that the action of $\hat S$ does not change the Frobenius structure of the CohFT - quadratic terms do not affect the structure constants of the algebra. But this quadratic terms of the Frobenius potential play an important role in particular applications (cf. \cite{G, DZ}). They give the so--called \textit{calibration} of the Frobenius manifold.
\begin{proposition}
The action of $\hat A_G$ on the calibration of $F_0$ is given by the addition of the following terms:
$$
\left( \hat A_G \cdot F_0 \right) (\textbf t)= F^A_0(\textbf t) + \left( A\cdot \tau - \tau \right) \left( t^1 \right)^2.
$$
\end{proposition}
\begin{proof}
In the definition of $\hat A_G$ it was indeed the purpose of the introduction of the $S$--action to apply certain change of variables by $S^{-\tau_0}$ and $S^{A\tau_0}$. Let's compute the quadratic form $W$ for both. It's easy to see that in both cases $W_{i,j}$ is only non--zero when $i = 0$ and $j = 0$. We have:
$$
W_{0,0}^\alpha=
\left(
\begin{array}{c c c}
0 & \dots & 0 \\
\vdots & 0 & \vdots \\
\alpha & \dots & 0
\end{array}
\right)
\quad
\Rightarrow
\quad
W^\alpha(t) = \alpha (t^1)^2,
$$
where $\alpha$ is either $A \cdot \tau_0$ or $- \tau_0$.
Hence the actions of $S^{A\tau_0}$ and $S^{-\tau_0}$ add quadratic term $(A\cdot\tau_0 - \tau_0) \left( t^1 \right)^2$. The action of $S_0^A$ does not rescale the coordinate $t^1$ and we only have to take care of the $R^\sigma$--action.
Generally a $R$--action does not change any lower than cubic terms of the potential.
However in the formula of $\hat A_G$ we apply $R^\sigma$ to the $S$--transformed partition function. In particular $\hat R^\sigma$ can act on the quadratic term added by $S^{A\tau_0}$. This is clear however that in our particular case this does not happen on the level of Frobenius manifold giving higher order corrections only.
\end{proof}
\section{$ {\rm SL(2,\mathbb C)} $ action in singularity theory}\label{section: totalAncestorPotential}
This section is devoted to the connection of the $ {\rm SL(2,\mathbb C)} $ action developed above to the action on the total ancestor potential of a simple elliptic singularity given by \cite{MR}.
We do not give all the details of Saito's (Frobenius) structures of the singularity and total ancestor potential referring the interested reader to \cite{G} and \cite{H}.
\subsection{Frobenius manifold of the hypersurface singularity} Let $W(\textbf{x}) = W(x_1, \dots ,x_N)$ define an isolated singularity at $0 \in \mathbb C^N$. The unfolding of the singularity is the following function:
$$
W(\textbf{x}, \textbf{s}) := W(\textbf{x}) + \sum_{i=1}^\mu \phi_i(\textbf{x}) s^i,
$$
where $\mu$ is the Milnor number of $W(\textbf{x})$ and the functions $\phi_i(\textbf{x}) \in \mathcal O_{\mathbb C^N,0}$ generate the Milnor algebra of the singularity: $\langle \phi_1, \dots, \phi_n \rangle = \mathcal O_{\mathbb C^N, 0} / \left\langle \partial_{x_1} W, \dots, \partial_{x_N}W \right\rangle$. We assume further that $\phi_1 = 1$.
Consider the base space $\mathcal S \subset \mathbb C^\mu$ of the singularity unfolding: $(s_1, \dots, s_\mu) \in \mathcal S$. For some volume form $\omega = f(\textbf{x}, \textbf{s}) dx_1 \dots dx_N$ define:
$$
c_{ijk}(\textbf{s}) := \int_{\Gamma_\epsilon} \frac{\partial_{s_i} W(\textbf{x}, \textbf{s}) \partial_{s_j} W(\textbf{x}, \textbf{s}) \partial_{s_k} W(\textbf{x}, \textbf{s})}{\partial_{x_1} W(\textbf{x}, \textbf{s}) \dots \partial_{x_N} W(\textbf{x}, \textbf{s})} \omega, \quad \eta_{ij}(\textbf{s}) := c_{1ij}(\textbf{s}),
$$
and $\Gamma_\epsilon = \Gamma_\epsilon(\textbf{s})$ is supported on $\{\textbf{x} \in \mathbb C^N | |\partial_{x_1} W(\textbf{x},\textbf{s})| = \dots = |\partial_{x_N} W(\textbf{x},\textbf{s})| = \epsilon\}$ for some $\epsilon$ small enough.
Defined in this way $\eta^{ij}$ will be structure constants of a non--degenerate bilinear form on $\mathcal{T}_{\mathcal S}$ (cf. \cite{AGV}). As in Section~\ref{section: AnalyticalQuantization} the multiplication can be defined on $\mathcal{T}_{\mathcal S}$ by the structure constants $c_{ij}^k(\textbf{s}) := c_{ijp}(\textbf{s}) \eta^{pk}(\textbf{s})$. This multiplication together with $\eta$ define a Frobenius algebra structure.
However special choice of the volume form is needed to make $\eta$ flat and get the Frobenius manifold structure.
\begin{theorem}[\cite{S}]
There is a choice of the volume form $\omega$ such that the pairing $\eta$ is flat and defines together with $c_{ij}^k$ a structure of the Frobenius manifold (of dimension $\mu$) on $\mathcal S$.
\end{theorem}
The volume form as above is called a \textit{primitive form} of Saito and its existence is another complicated question (cf. \cite{SM} and \cite{H}). The choice of the primitive form is generally not unique. In what follows we will denote by $\textbf{t} = \textbf{t}(\textbf{s})$ the coordinates on $\mathcal S$ that are flat for $\eta$.
We will denote by $M_{W,\zeta}$ the Frobenius manifold structure of $W(\textbf{x})$ with the choice of the primitive form $\zeta$.
\subsection{Total ancestor potential}
Let $F(\textbf t) = F(t^1, \dots, t^\mu)$ be a potential of the Frobenius manifold $M_{W,\zeta}$.
Consider the change of variables $u^p = u^p(\textbf t)$ for $1 \le p \le \mu$ such that
$$
c_{ij}^k(\textbf u) = \Delta_i \delta_{i,j} \delta_{j.k}, \quad 1 \le i,j,k \le n,
$$
for some functions $\Delta_i = \Delta_i(\textbf u)$.
Such coordinates $\textbf u$ are called canonical. Denote by $\Psi$ the transformation matrix:
$$
\Psi = \Psi(\textbf t): \langle \partial / \partial t^i \rangle \rightarrow \langle \partial/\partial u^i \rangle.
$$
The point $\tau \in M_{W,\zeta}$ is called \textit{semi--simple} if $\Psi(\tau)$ is non--degenerate.
Canonical coordinates $u^i(\textbf{t})$ as the functions of the flat coordinates are found as the critical values of the singularity unfolding (written in flat coordinates):
$$
u^i(\textbf{t}) = W(P_i,\textbf{t}), \quad 0 = \frac{\partial W(\textbf{x}, \textbf{t})}{\partial x_1} \mid_{\textbf{x} = P_i} = \dots = \frac{\partial W(\textbf{x}, \textbf{t})}{\partial x_N} \mid_{\textbf{x} = P_i}.
$$
In what follows define $U := \text{diag} (u^1, \dots, u^\mu)$.
\subsubsection{Lefschetz thimbles}
Let $W(\textbf{x}, \textbf{s})$ be the unfolding of a fixed isolated quasi--homogeneous singularity $W(\textbf{x}) = W(x_1, \dots, x_N)$ with the Milnor number $\mu$.
Fix some positive $\rho$, $\delta$ and $\nu$. Let $B^N_\rho \subset \mathbb{C}^N$ , $B^1_\delta \subset \mathbb{C}$ and $B^\mu_\nu \subset \mathbb{C}^\mu$ be respectively the balls of radii $\rho$, $\delta$ and $\nu$ centered at the origin. Consider the space:
$$
\mathcal{X}_{\rho,\delta,\nu} := \left( B^N_\rho \times B^\mu_\nu \right) \cap \varphi^{-1} \left( B^1_\delta \times B^\mu_\nu \right) \subset \mathbb{C}^N \times \mathbb{C}^\mu.
$$
Taking $\rho$ such that $X_{0,0}$ is intersected transversally by $\partial B^N_r$ for all $r: \ 0 < r \le \rho$ and $\delta, \nu$ such that $X_{\lambda,\textbf{s}}$ is intersected transversally by $\partial B^N_\rho$ for all $(\lambda,\textbf{s}) \in B^1_\delta \times B^\mu_\nu$ we get the following proposition.
\begin{proposition}[\cite{AGV}]
There is $D \subset B^1_\delta \times B^\mu_\nu$ such that for $\mathcal{X}^\prime := \mathcal{X}_{\rho, \delta,\nu} \backslash \varphi^{-1}(D)$ the map $\varphi: \mathcal{X}^\prime \to \left( B^1_\delta \times B^\mu_\nu \right) \backslash D $ is a locally trivial fibration with a generic fibre homeomorphic to a bouquet of $\mu$ spheres of dimension $N-1$.
\end{proposition}
For $m \in \mathbb{R}$ consider the space $\mathcal{X}^-_{m} \subset \mathcal{X}_{\rho,\delta,\nu}$ defined by:
$$
\mathcal{X}^-_{m} := \left\lbrace (\textbf{x},\textbf{s}) \in \mathcal{X}^\prime \ | \ \mathrm{Re}\left( W(\textbf{x},\textbf{s})/z \right) \le -m \right\rbrace.
$$
Because of the proposition above and exact sequence of the pair we get the following isomorphisms:
\begin{equation*}
H_N(\mathcal{X}^\prime, \mathcal{X}^-_{m}) \cong H_{N-1}(X_{\lambda,\textbf{s}}) \cong \mathbb{Z}^\mu.
\end{equation*}
Define the cycles
\footnote{
These cycles have first appeared in \cite{G2} and used later for example in \cite{MR} in the following form. For $(\mathbb{C}^N)_m := \{\textbf{x} \in \mathbb{C}^N | {\rm Re}(W(\textbf{s},\textbf{x})/z) \le -m\}$. Define $ \mathcal{A} \in \lim_{m \to \infty} H_N(\mathbb{C}^N, (\mathbb{C}^N)_{-m}; \mathbb{C}) \cong \mathbb{C}^\mu$. However such a definition should be considered more like a notation based on the fact that $\mathcal{X}^\prime$ is contractible while the similarity of $\mathcal{X}^-_m$ and $(\mathbb{C}^N)_m$ is clear.
}:
$$
\mathcal{A}_{\textbf{s},z} \in \lim_{m \to \infty} H_N(\mathcal{X}^\prime, \mathcal{X}^-_{m}; \mathbb{C}) \cong \mathbb{C}^\mu.
$$
Slightly obusing the notation in what follows we denote by $\mathcal A_1, \dots, \mathcal A_\mu$ the basis of the relative homology group above.
\subsubsection{Oscillatory integrals} Having built the cycles $\mathcal{A}_k$ we are going to integrate the function $\exp(W(\textbf{x},\textbf{s})/z)$ against them.
Considered as the functions of $\textbf{s}$ the stationary phase asymptotic of these integrals by taking $z \rightarrow 0$ reads:
$$
\int_{\mathcal {A}_k} \exp(W(\textbf{x},\textbf{s})/z) \zeta \sim \frac{e^{u^k/z}}{\sqrt{\Delta_k}}(1 + R_1 z + R_2 z + R_3z^3 + \dots)(e_k), \quad z \to 0,
$$
where $u^k, \Delta_k$ are as in the previous subsection, $e_k$ is a basis vector from $T_{\mathcal S}$ and the matrices $R_p$ on the RHS are functions of $\textbf{s}$. We define $R_{\textbf{s}} := 1 + R_1 z + R_2 z^2 + R_3z^3 + \dots$. It's not hard to check this is indeed an upper--triangular Givental's group element.
Note that the matrix $R_\textbf{s}$ is only defined for a semi--simple $\textbf{s} \in \mathcal{S}$.
\subsubsection{Total ancestor potential}
Consider the formal coordinates $\textbf u^p := \{ u^{d,p} \}$ for a fixed $p$ s.t. $1 \le p \le \mu$ and all integer $d \ge 0$. Define:
$$
\mathcal D_{KdV}^{\{p\}}(\hbar, \textbf u^p) := \exp \left( \sum \frac{\hbar^{g-1}}{n!} \int_{\overline {\mathcal M}_{g,n}} \prod_{i=1}^n \left( \sum_{d \ge 0} u^{d,p} (\psi_i)^d + \psi_i \right) \right).
$$
This is the partition function of the trivial CohFT with the Dilaton shift applied.
\begin{definition}
The partition function $\mathcal A_{\textbf{s}}(\hbar,\textbf{t})$ formally defined as follows is called total ancestor potential.
$$
\mathcal A_{\textbf{s}}(\hbar,\textbf{t}) := \hat \Psi_{\textbf{s}} \hat R_{\textbf{s}} \prod_{p=1}^\mu \mathcal D_{KdV}^{\{p\}}(\hbar \Delta_p, \textbf u^p \sqrt{\Delta_p}),
$$
where $\hat R_{\textbf{s}}$ acts as the Givental's upper triangular group element, $\hat \Psi_{\textbf{s}}$ acts by the change of variables $u^{d,i} = \left( \Psi (\textbf{s}) \right)^i_j t^{d,j}$.
\end{definition}
In the formula above it's very important to distinguish the dependance of the total ancestor potential on $\textbf{t}$ and on $\textbf{s}$. The first one is a formal variable, while $\textbf{s}$ stands for a choice of the point of $\mathcal S$ around wich the structure is defined. This is best seen on the connection of the total ancestor potential to the Frobenius manifold $M_{W,\zeta}$. The function
$$
F_\textbf{s}(\textbf{t}) := \left[ \hbar^{-1} \right] \ \log \mathcal{A}_{\textbf{s}} (\hbar, \textbf{t}) |_{\textbf{t}^{d,p} = 0, \ d \ge 1}
$$
is the potential of the Frobenius manifold structure of $M_{W,\zeta}$ in the neighborhood of the point $\textbf{s}$.
Hence the variables $t^{0,p}$ correspond indeed to the flat coordinates discussed at the start of this section while for $\textbf{s}$, considered as the parameter, it's even irrelevant to speak about the flatness.
\subsection{Choice of the primitive form}
All the elements building up both total ancestor potential and Frobenius manifold of the singularity $W(\textbf{x})$ depend heavily on the choice of the primitive form.
Let $W(\textbf{x})$ be one of the following polynomials:
\begin{align*}
& x_1^3 + x_2^3 + x_3^3 + \sigma x_1x_2x_3,
\\
& x_1^2x_3 + x_1x_2^3 + x_3^2 + \sigma x_1x_2x_3,
\\
& x_1^3x_3 + x_2^3 + x_3^2 + \sigma x_1x_2x_3,
\end{align*}
that depend also on the additional complex parameter $\sigma$. These polynomials $W(\textbf{x}) = W_\sigma(\textbf{x})$ defines so--called simple--elliptic singularities.
Considering
$$
E_\sigma := \{ W_\sigma(\textbf{x}) = 0 \} \subset \mathbb P^2(c_1,c_2,c_3)
$$
for some $c_i$ one gets a family of elliptic curves $E_\sigma \to \mathbb C \backslash D$, where ${D := \{ \sigma \in \mathbb C \ | \ E_\sigma \text{ is singular}\} }$.
Let $\Omega_\sigma$ be a holomorphic volume form on $E_\sigma$ and $A_\sigma \in H_1(E_\sigma,\mathbb C)$ -- flat family of cycles.
The primitive form for simple elliptic singularities was found by K. Saito (Examples in \cite[Paragraph 3]{S}):
$$
\zeta = \frac{d^3\textbf{x}}{\pi_A(\sigma)} \quad \text{ for } \quad \pi_A(\sigma) := \int_{A_\sigma} \Omega_\sigma.
$$
The period $\pi_A(\sigma)$ is a solution to the Picard-Fuchs equation and is fixed by the choice of a particular $A_{\sigma_0} \in H_1(E_{\sigma_0})$ for some fixed $\sigma_0$.
This property was explored differently in \cite{MR} and \cite{BT} to consider different primitive forms of the fixed simple elliptic singularity.
\subsubsection{Analytical action on the 3--dimensional Frobenius manifolds} Let $\gamma(t)$ be a Chazy equation solution $\gamma^{\prime\prime\prime} = 6 \gamma \gamma^{\prime\prime} - 9(\gamma^\prime)^2$. It's easy to see that for any
$A = \begin{pmatrix}
a & b
\\
c & d
\end{pmatrix} \in \mathrm{GL}(2,\mathbb C)
$ the following function is solution to Chazy equation too.
$$
\gamma^A(t) := \frac{\det(A)}{(ct + d)^2} \gamma \left(\frac{at + b}{ct + d}\right) - \frac{c}{2(ct + d)}.
$$
Hence for any $A \in \mathrm{GL}(2,\mathbb C)$ we get the potential of the $3$--dimensional Frobenius manifold (recall Example~\ref{example: chazy}):
$$
F^A(t^1,t^2,t^3) := \frac{1}{2} (t^1)^2t^3 + \frac{1}{2} t^1 (t^2)^2 - \frac{(t^2)^4}{16} \gamma^A(t^3).
$$
However it's clear that for
$$
A^\prime := \begin{pmatrix}
a/\det(A) & b
\\
c/\det(A) & d
\end{pmatrix} \in \mathrm{SL}(2,\mathbb C)
$$
the potentials $F^A$ and $F^{A^\prime}$ differ by the following rescaling:
$$
\det(A) \ F^{A^\prime} \left( \frac{t^1}{\det(A)}, t^2 , \det(A) t^3 \right) = F^A(t^1,t^2,t^3).
$$
In \cite{BT} the authors proposed the following $\mathrm{GL}(2,\mathbb C)$ action as the model for the primitive form change. For any $\tau_0 \in \mathbb H$ and $\omega_0 \in \mathbb{C}^*$ define:
$$
A^{(\tau_0, \omega_0)} :=
\begin{pmatrix}
\dfrac{\bar{\tau}_0}{4\pi\omega_0{\rm Im}(\tau_0)} & \omega_0 \tau_0
\\
\dfrac{1}{4\pi\omega_0{\rm Im}(\tau_0)} & \omega_0
\end{pmatrix}.
$$
For any $\tau_0 \in \mathbb H$ and $\omega_0 \in \mathbb C^*$ we have $\det(A^{(\tau_0,\omega_0)}) = 1/(2 \pi \sqrt{-1})$ and because of the reasoning above we can consider unique $\mathrm{SL}(2,\mathbb{C})$ action defined by $(\tau_0,\omega_0)$. Fixing some particular solution to Chazy equation $\gamma^\infty(t)$ consider the Frobenius manifold potential
$$
F^{(\tau_0,\omega_0)}(t^1,t^2,t^3) := \frac{1}{2} (t^1)^2t^3 + \frac{1}{2} t^1 (t^2)^2 - \frac{(t^2)^4}{16} \gamma^{(\tau_0,\omega_0)}(t^3),
$$
where $\gamma^{(\tau_0,\omega_0)}$ is obtained from $\gamma^\infty$ by the $\mathrm{SL}(2,\mathbb C)$--action of $(\tau_0,\omega_0)$.
The action of $A^{(\tau_0,\omega_0)}$ on the space of $3$--dimensional Frobenius manifolds can be considered as the model for the change of the primitive form of a simple elliptic singularity (cf. Section~2.5 in \cite{BT}). We are going to compare it with the completely different approach of Milanov--Ruan.
\subsubsection{Approach of Milanov--Ruan}
From now on we fix $n := \mu$, the Milnor number of the singularity.
Let $e_1, \dots, e_n$ be the basis vectors of $T_{\textbf{s}}\mathcal{S}$ at ${\textbf{s}} = \{s^1,\dots,s^n\}$.
For any
$A \in {\rm SL(2,\mathbb C)} $ let the linear operator $J: T_\textbf{s} \mathcal{S} \to T_\textbf{s} \mathcal{S}$ be defined by:
$$
J (e_1) = e_1, \ J(e_n) = (cs^n + d)^2 e_n, \ J(e_p) = (cs^n + d) e_p, \quad 1 < p < n, \quad A = \begin{pmatrix}
a & b
\\
c & d
\end{pmatrix}.
$$
Let $\textbf{q} = \{q_k^i\}$ be as in Section~\ref{section: GiventalSymplectic} with the upper index corresponding to the vector number in the basis fixed above. Introduce the action of $J$ on $\textbf{q}$ by:
$$
J: q^j_d \rightarrow \sum_{i=1}^n J_i^j q^i_d, \quad \forall d \ge 0.
$$
Consider Givental's upper--triangular group element $X_{\textbf v}(z)$ defined by:
$$
\left( X_\textbf{s}(z) \right)_i^j = \delta_i^j + - z \frac{c}{c s^n + d} \delta_{i,n}\delta^{j,1},
$$
\begin{theorem}[Theorem~4.4 in \cite{MR}]\label{theorem: Milanov-Ruan}
Consider total ancestor potentials of a simple elliptic singularity
\footnote{Only one particular case of simple elliptic singularity $P_8$ was considered in \cite{MR}. However the technique used is extended in a straightforward way to all other cases too.}
with the primitive forms $\zeta_1$ and $\zeta_2$. Then there is $A \in {\rm SL}(2,\mathbb C)$ such that the corresponding total ancestor potentials are connected by the following transformation:
$$
\mathcal A_{\zeta_1, A \cdot \textbf{s}}( \hbar, \textbf{q}) = (\hat X_\textbf{s} \cdot \mathcal A_{\zeta_2, \textbf{s}}) \left( (c \tau + d)^2 \hbar, J \textbf{q} \right),
$$
where the action of $\hat X_\textbf{s}$ is applied as in Proposition~\eqref{proposition: action in symplectic setting},
$\textbf{s} = (0, \dots, 0, \tau)$ and $A \cdot \textbf{s} = (0, \dots, 0, A \cdot \tau)$.
\end{theorem}
In \cite{MR} in order to fix particular primitive form the authors choose particular solution to the Picard--Fuchs equation. In this way it's very hard to present particular $ {\rm SL(2,\mathbb C)} $ matrix of the theorem above.
\subsection{Equivalence of the approaches}
Comparing the total ancestor potential formula of Milanov--Ruan with the $ {\rm SL(2,\mathbb C)} $ action developed in this paper we get the theorem.
\begin{theorem}\label{theorem: MR-BT equivalence}
Let $F$ be the Frobenius manifold potential of a simple elliptic singularity. Then the transformation of Milanov--Ruan is equivalent to the $ {\rm SL(2,\mathbb C)} $ action on $F$ given by \eqref{eq:SLAction}.
\end{theorem}
\begin{proof}
Consider particular pair of primitive forms $\zeta_1$ and $\zeta_2$ with the corresponding $ {\rm SL(2,\mathbb C)} $--matrix $A$ as in Theorem~\ref{theorem: Milanov-Ruan}. We show that the transformation of Milanov-Ruan acts in the same way as Givental's form of our $ {\rm SL(2,\mathbb C)} $ action.
Denote by $\hat J$ the rescaling action $(\hbar, \textbf{q}) \to ((c\tau+d)^2 \hbar, J \textbf{q})$. Note that its action is the same as $S_0^A$ action of Corollary~\ref{corollary: sigma-sigmaPrime}. Recall $R^\sigma$--action of Section~\ref{section: GraphsCalculus}. Because of its particular form we have:
$$
R^\sigma(z) = \exp(r^\sigma z) = 1 + r^\sigma z.
$$
By Corollary~\ref{corollary: sigma-sigmaPrime} we know that for $\sigma := -c(c \tau + d)$ the following equality holds:
$$
\hat R^\sigma \cdot \hat S_0^A = \hat J \cdot \hat X_\textbf{s}.
$$
Hence the transformation of Milanov--Ruan is represented by the $ {\rm SL(2,\mathbb C)} $ action we consider. We only have to compare the points of the Frobenius manifolds on both sides. Using Theorem~\ref{theorem: Milanov-Ruan}, Theorem~\ref{th:FrobeniusManifold} and Proposition~\ref{proposition: two actions} we write:
$$
F_{\zeta_1, A \cdot \tau} = \left( \hat R^\sigma \right)^{-1} \hat S_0^A F_{\zeta_2, \tau} = \hat R^{-\sigma} \hat S_0^A F_{\zeta_2, \tau}
= \hat S_0^A \hat X_\textbf{s} F_{\zeta_2, \tau}.
$$
This completes proof of the theorem.
\end{proof}
\begin{corollary}
Two approaches of \cite{MR} and \cite{BT} for the action changing the primitive form of the simple elliptic singularity coincide.
\end{corollary}
This result shows that taking all the possible primitive forms for a fixed simple elliptic singularity there are only two ``geometrically'' different Frobenius manifolds --- $F_{\zeta_1}$ and $\hat I \cdot F_{\zeta_1}$ (where $I$ is the Inversion transformation of Dubrovin). All the other Frobenius manifolds fixed by other choices of the primitive forms are obtained from these two by taking linear changes of the variables.
\subsection{Application to the Gromov--Witten theory}
The change of the primitive form for a given singularity $W(\textbf{x})$ is of big importance in Mirror symmetry. For $W(\textbf{x})$ defining a hypersurface simple--elliptic singularity it was proved by Milanov--Ruan and Milanov--Shen (cf. \cite{MR,MS1}) that for a special choice of the primitive form $\zeta_\infty$ the total ancestor potential of $W$ with this primitive form coincides with the partition function of the (orbifold) Gromov--Witten theory of $\mathcal{X}$, for $\mathcal{X}$ one of $\mathbb{P}^1_{3,3,3}$, $\mathbb{P}^1_{4,4,2}$, $\mathbb{P}^1_{6,3,2}$.
Another amazing result of Milanov--Ruan later extended by Milanov--Shen is the following theorem proved via the analysis of the oscillatory integrals:
\begin{theorem}[Theorem~4.2 in \cite{MR}, Proposition~3.4 in \cite{MS2}]
For any choice of the primitive form $\zeta$ for a hypersurface simple--elliptic singularity there are $A \in \mathrm{SL}(2,{\mathbb{Z}})$ such that:
$$
(\hat X_\textbf{s} \cdot \mathcal A_{\zeta, \textbf{s}}) \left( (c \tau + d)^2 \hbar, J \textbf{q} \right) = \mathcal A_{\zeta, A \cdot \textbf{s}}( \hbar, \textbf{q}) ,
$$
where the action of $\hat X_\textbf{s}$ is applied as in Proposition~\eqref{proposition: action in symplectic setting},
$\textbf{s} = (0, \dots, 0, \tau)$ and $A \cdot \textbf{s} = (0, \dots, 0, A \cdot \tau)$.
\end{theorem}
Namely, comparing to Theorem~\ref{theorem: Milanov-Ruan} the primitive form on the both sides of the equality above is the same. Such transformtaions were called \textit{modular}.
Applied to the total ancestor potential with the primitive form $\zeta_\infty$ the parititon functions of the Gromov--Witten theories of the elliptic orbifolds were proved to be quasi--modular forms (recall the definition from Section~\ref{section: AnalyticalQuantization}) in \cite{MR,MS2}. We give the quasi--modularity result mentioned in a simplified --- genus $0$ version.
\begin{corollary}[Corollary~6.7 in \cite{MR}, Corollary~1.3 in \cite{MS2}]
Consider the genus zero potential $F^{\mathcal X}(\textbf{t})$ of an elliptic orbifold ${\mathcal X}$. Let $f_{\bf k}(t^n)$ be its coefficients in $t^2,\dots,t^{n-1}$:
$$
f_{\bf k}(t^n) = [t^{k_1} \cdots t^{k_p}] \ F(\textbf{t}), \quad {\bf k} = \{k_1,\dots, k_p\},
$$
for a different index set $\bf k$. Then the functions $f_{\bf k}(t^n)$ are quasi--modular forms.
\end{corollary}
Because we have identified the transformation of Theorem~\ref{theorem: Milanov-Ruan} with the analytical $ {\rm SL(2,\mathbb C)} $ action on the Frobenius manifold potential we can make the following statement:
\begin{proposition}\label{proposition: modQuasiMod}
Consider the Frobenius manifold potential $F^{\mathcal X}(\textbf{t})$ of an elliptic orbifold ${\mathcal X}$. Let $f_{\bf k}(t^n)$ be its coefficients in $t^1,\dots,t^{n-1}$. Then holds:
\begin{itemize}
\item[(a)] $f_{\bf k}$ are quasi--modular of weight $2$ if
$$
\frac{\partial ^4}{\partial t^{k_1} \partial t^{k_2} \partial t^{k_3} \partial t^{k_4}} \left(\frac {\partial F^{\mathcal X}(\textbf{t})}{\partial t^1}\right)^2 \neq 0,
$$
\item[(b)] $f_{\bf k}$ is modular of weight $|k|-2$ otherwise.
\end{itemize}
\end{proposition}
\begin{proof}
The proposition follows immediately by comparing the coefficients in $t^2,\dots,t^{n-1}$ of the equality $\left(F^{\mathcal X} \right)^A = F^{\mathcal X}$. In particular the $Q$--term has to be annihilated by some $f_{\bf k}$, that hence turn out to be quasi--modular.
\end{proof}
The genus $0$ potentials of the elliptic orbifolds are not polynomial --- the functions $f_{\bf k}$ in the theorem above have infinite Fourier series. However their Fourier series can be computed up to any order needed (using for example the reconstruction theorems~3.1 and 4.2 in \cite{IST}). It's well known that if two functions $g_1(\tau)$ and $g_2(\tau)$ are modular, then only finite number of terms (given by the Stein bound) in their Fourier expansions are to be identified to check if $g_1(\tau) \equiv g_2(\tau)$.
Because of Proposition~\ref{proposition: modQuasiMod} we know that majority of functions $f_{\bf k}$ in the theorem above are indeed modular and we can apply the Stein bound argument.
For those $f_{\bf k}$ that are not modular amazing theorem of Kaneko--Zagier (\cite{KZ}) can be applied, giving that any quasi--modular form $g(\tau)$ of weight $2$ can be expressed as the product of the second Eisenstein series $E_2(\tau)$ and a weight $0$ modular form. Hence we can apply the Stein bound argument again. Using the computer program we have computed the genus $0$ potentials of the elliptic orbifolds in the closed form using this method. The expressions are too big and therefore we put them on the webpage \cite{B_hp}.
Using the explicit characterization of the rign of modular forms, similar computations were done by Y.Shen and J.Zhou in \cite{SZ}. They have computed explicitly some of the correlators of the Gromov--Witten theory of the elliptic orbifolds, that are enough to reconstruct all genera data. Our result agree, however they do not give a closed formula for the genus $0$ potential and we do not have a formula for the higher genera potentials.
|
\section{Introduction}
Observations suggest that structure formation in the Universe proceeds
hierarchically, with the smallest structures collapsing first and then
later merging to form larger structures. An important property of mergers
is that if two halos with identical density profiles merge, the merger remnant
will have a density profile with the same inner and outer slopes as the initial
profiles \citep{2006ApJ...641..647K}. It has been proposed that this behaviour may be understood from considerations about mixing of collisionless
systems \citep{Dehnen2005}. Further merger remnants have been found to be
triaxial with a major axis along the collision axis~\citep{2004MNRAS.354..522M},
and the velocity anisotropy profiles
are typically radial, in the sense $\sigma_{\rm r}^2 \ge \sigma_{\rm tan}^2$,
whether calculated in spherical or
elliptical bins. The velocity anisotropy profiles of merger remnants
are, however, not simple functions; they are asymmetric in the sense
that a very different behaviour is seen along different axes
\citep{2012JCAP...07..042S}. It has been pointed out~\citep{2013arXiv1303.2056W}
that the velocity ellipsoids of a halo are typically
aligned with the major axis, and it means that the velocity anisotropy
parameter gives a misleading description of triaxial halos.
The relaxation and mixing processes in collisionless mergers have been
examined in \citep{2007ApJ...658..731V}, where it was found that mixing of the
6-dimensional phase space distribution function mainly occurs during
the tidal shocking arising when the center of the merging halos pass
through each other, and that about
40 \% of the particles
from the merging halos are located outside the virial radius of the
remnant. In controlled numerical galaxy collisions it has been
known for a long time that some particles are ejected with positive
energies (see e.g. \citep{hernquist92}). In cosmological simulations it has
been found that unbound particles are abundant in halos which have
recently undergone a major merger~\citep{2012arXiv1208.0334B}.
Why particles are ejected in the course of a merger is straightforward
to understand on energetic grounds. When a small structure is engulfed by
a larger one, it will be ripped apart dynamically, and a new equilibrium
state will be reached. At the end of this equilibration, the virial
theorem must hold for the bound particles, i.e.,
\begin{equation}
2 K + W = 0 \, ,
\end{equation}
where $K$ and $W$ are the total kinetic and potential energy of the
bound particles (with $K$ calculated relative to their centre of mass).
The total energy of the virialized particles is thus $E_b=-K$.
For the case of two virialized structures initially
very far from one another, the total energy can be written as
$E= -K_1-K_2+K_{CM}$ where $K_1$ and $K_2$ are the
initial kinetic energies of the two systems (relative to their centre of
masses) and $K_{CM}$ is the kinetic energy of the centre of
masses of the two halos when they are far apart.
By energy conservation we therefore have that particles must be
ejected if $K_{CM}$ in the centre of mass frame, i.e.,
the kinetic energy of the initial relative motion, is sufficiently
large to make the total initial energy positive: dark matter,
unlike baryons which can radiate, have no other way of
discarding the excess energy in the system. In the case of
cold dark matter, such relative motion is typically small, and
we will in fact assume it to be zero, i.e., the structures start
far apart without initial relative motion.
In this case, from a global energetic point of
view, ejection {\it may or may not} take place. However, in
the case of a small structure which encounters a very
large structure it is not difficult to see why such ejection
{\it does} in fact typically take place, and why the ejected
particles are (as we will verify in detail below) those
in the small structure. Neglecting completely the
potential energy due to the small structure, we
can consider the particles belonging to it as
unbound particles falling into the fixed potential
of the large structure. The condition of
virialization in this fixed potential then imposes that
energy must be ejected:
for example, if two
particles arrive from far away
with zero energy, and virialize
in this way, the combination of the virial
condition and energy conservation
would then give
$2(K+k_a+k_b) + (W+w_a+w_b)=
k_a+k_b=0$, where $k_a$ ($w_a$)
and $k_b$ ($w_b$) are the average
kinetic (potential) energies of the
particles, a condition which cannot
be satisfied. The system, however, can reach
a virial equilibrium without any
signification modification of the large
structure by ejecting particles carrying
away excess energy.
From a dynamical point of view, the particles in
the small structure are much less bound and the potential
fluctuations induced by the merger are of order
the initial potential energy of this structure,
precisely of the order required to eject them.
In this paper we perform numerical simulations to study
and characterize the ejection of particles in mergers and
to understand the mechanism responsible for this ejection.
We use $N$-body simulations to study the ejection
of particles in minor mergers, and we use single particle simulations
in smooth potentials to demonstrate that the ejection mechanism is
indeed a mean-field effect. Particle (and energy) ejection during the
process of violent relaxation has been previously discussed
at length in~\cite{2009MNRAS.397..775J}, and studied in detail with $N$-body
simulations for the case of cold quasi-spherical initial conditions.
Typically about $15\%$ of the particles are found to ejected in
this case, and it is shown that these are particles initially in
the outer shells which ``arrive late" at turnaround and pick up a
positive energy kick as they travel
in the potential sourced
by the re-expanding
potential of the bulk of the particles. Despite the quite different
initial conditions studied here, we will see that the
mechanisms at play in the ejection are in fact very similar.
\section{$N$-body simulations}
\label{sub:nbody}
In this section we will describe our $N$-body simulations of minor
mergers, which have been performed using
the \texttt{GADGET-2} code \citep{2005MNRAS.364.1105S,springel_gadget}.
All the simulations were pure dark matter simulations, where dark matter
is modelled as collisionless particles, and in a non-expanding universe
with open boundary conditions.
\begin{figure}
\includegraphics[scale=0.8]{snap_plot.pdf}
\caption{The positions in $x$ and $y$ of the particles at different
time steps of the simulation. The yellow dots are particles
belonging to the major halo, the blue triangles are the minor halo
particles that stays bound throughout the simulation, and the red
stars are the particles which get ejected. At $t=64$, it can be
seen that the ejected particles from the minor halo arrive at the
plane of the center of the major halo slightly later than the center of the
minor halo does.}
\label{fig:snaps}
\end{figure}
\subsection{Simulation setup}
\label{sub:setup1}
We study mergers between halos with a mass ratio of 1:10,
representative of typical mergers encountered in cosmological
simulations. In our simulations we have $10^5$ particles in the
big halo and $10^4$ particles in the small halo.
We work in units in which the gravitational
constant $G$ is unity. Further we take both the mass of the larger halo,
$M_1$, and its scale radius, $r_\text{s}$, to be unity. All particles have
the same mass.
We use a gravitational softening given by $0.023$ in the
\texttt{GADGET-2} parameter file, and force and time integration
accuracies are fixed by \texttt{ErrTolForceAcc} $=0.005$ and
\texttt{ErrTolIntAccuracy} $=0.025$, respectively. We will discuss
further below tests we have performed on the stability of our results
to variation of these parameters.
The positions and the velocities of the particles are chosen so that
each structure, treated as an isolated system, is in a steady state
in virial equilibrium. Positions are assigned from the
Hernquist mass profile \citep{1990ApJ...356..359H},
\begin{align*}
\rho (r) = \frac{1}{r/r_\text{s}} \frac{\rho_0}{(1+r/r_\text{s})^3},
\end{align*}
and the velocities are selected from a Gaussian PDF with the initial
isotropic velocity dispersion derived from the Jeans equation. The
initial velocities are truncated at $0.95 v_\text{esc}$. We ran
simulations with different scale lengths, $r_{s2}$, of the small halo
(but with fixed mass).
The centres of mass of the two halos are placed at $y=0$ and $z=0$ in
our cartesian coordinate system. Instead in the $x$ direction we place
the big halo at $0$ and the small at $15$. The latter value is a rough
estimate for the turnaround radius of the big structure
\citep{2006ApJ...645.1001P,2008MNRAS.389..385C}, assuming it
to have a typical concentration of galaxies \citep{2008MNRAS.391.1940M}.
Having chosen to place the minor halo at the turnaround radius of the major,
we start the simulation start with both halos at rest.
The small structure starts approaching the big one, pulled by
its gravitational attraction.
After the merger we continue the simulation for at least $10$ more
dynamical times, where we define a typical dynamical time from the
circular velocity at $r_4=4\,r_{s1}$, $\tau_\text{dyn}=r_4/v_c(r_4)$.
When we observe that the total energy of the new structure (without the
ejected particles) is constant, we deduce the system has reached a
new equilibrium.
\subsection{Results}
\subsubsection{Identifying the ejected particles}
\label{sub:example}
Our first simulation uses a scale radius $r_{s2} = 0.3$ for the small
halo. The positions of the particles as a function of time is shown in
figure \ref{fig:snaps}. The red stars in the plot are particles which
start in the minor halo and are ejected during the simulation, where
we label a particle \emph{ejected} if its total energy in the
final snap-shot is positive.
The colour code allows us to follow the ejected particles in red from
the beginning of the run, to distinguish them from the
other particles belonging initially to the small structure, in blue,
and those belonging to the big structure, in yellow. We see
that, at $t=50$, the small halo has not yet crossed the big halo core and
the red particles are well mixed with the blue ones. As the
small halo enters the large halo's core at $t=64$, we see that
the red particles are those which \textquotedblleft lag
behind\textquotedblright and are the last ones to cross the core. The
next snapshot describes the subsequent ejection, and in the last
one at $t=200$ ($\sim 23$ dynamical times) there are no longer
red particles within 15 times the scale radius.
\subsubsection{Numerical tests}
As pointed out in the introduction, we expect ejection to happen generically in this kind of merger.
Nevertheless to check that the mass ejection we observe numerically is not the consequence of
some other effect, notably ill-posed initial conditions (i.e. particles not in equilibrium to begin with),
or accumulated integration errors in the particle orbits, we run additional specific test simulations.
We test first that our results are insensitive to the use of the Gaussian
approximation of the initial velocities. Specifically, we use the
Eddington inversion method \citep{binneytremaine} to set up initial equilibrated
structures (using an implementation which was tested in
\citep{attractor,2012JCAP...10..049}), and we find that merger time and number of
particles ejected are identical within a few percent. Also the
interesting feature of figure~\ref{fig:snaps}, that the ejected particles
are the ones arriving ``late'' is exactly the same for initial structures created
using the Eddington method.
Figure~\ref{fig:snaps} keeps its characteristics also when we run more accurate simulations,
taking \texttt{ErrTolForceAcc} $=0.001$ and \texttt{ErrTolIntAccuracy} $=0.005$ (i.e. a factor of
five smaller than in our fiducial simulations). In this case too, merger time and number of
particles ejected are stable within a few percent.
\subsubsection{Which particles are ejected?}
The mechanism of ejection is in fact simply that of violent
relaxation in general, as originally described by
\citep{1967MNRAS.136..101L} for stellar systems:
starting from an initial configuration which is far
from dynamical equilibrium, such a system can
relax precisely because, in the time dependent
gravitational potential, particles' energies can
change rapidly (i.e. on mean field time scales).
If the fluctuations of potential are sufficiently
violent, particles can reach and surpass
the escape velocity of the system.
\begin{figure}
\includegraphics[scale=0.8]{energies.pdf}
\caption{Time evolution of the total energy (upper panel) and potential
energy (lower panel) of all the particles belonging to the small
halo (green), of the particles which remain bounded (blue), and of
those which are ejected (red).}
\label{fig:energies}
\end{figure}
More specifically it is the particles that just happen
to arrive later in the region at the centre of the merger which
pick up a large positive kick to their energy in a short time as they
pass through the time-dependent potential well created by the
rest of the mass. To see that this is the case, we plot in
the lower panel of figure \ref{fig:energies} the average potential
energy of the blue (bounded) and red (escaping) particles
as defined in the previous figure, as well as the average
over all particles (in green). It can be seen clearly that the red particles pass on average through the minimum of the potential well slightly later than the blue ones.
Further from the plot in the upper right panel of figure~\ref{fig:snaps}, it
can be seen that all the ejected particles were on the right side of the
minor halo at $t=64$.
\begin{figure}
\includegraphics[scale=0.8]{dip.pdf}
\caption{The distribution of ``dip-times'' for all the particles which
are either ejected or bound. The dip-time is defined as the time
when the potential energy of the particle reaches its minimum. The
particles which are ejected are seen statistically to have a later
dip-time than the particles which are bound.}
\label{fig:diptime}
\end{figure}
In order to show that the ejected particles are the late arrivers, we
follow the orbit of each particle from the small halo, and we find the
time when the potential energy is at a minimum for each particle. For
this figure we include only particles inside 5 times the scale radius.
We can now plot the distribution of these ``dip-in-energy" times for
the particles which will eventually be ejected, and compare this to
the particles which will be bound, and we clearly see the difference
in Figure~\ref{fig:diptime}. We see that the peak of the ejected
particles is later than the peak for the particles which remain bound.
\begin{figure}
\includegraphics[scale=0.8]{v_r.pdf}
\caption{Radius $r$ and radial velocity $v$ normalised with the
related virial quantities for $2$ different scale radii for the
small structure. The red circles represent the particles which will
later be ejected, the blue dots are the particles which remain bound.}
\label{fig:vr}
\end{figure}
In figure \ref{fig:vr} we plot the radial velocities and the radii of
the particles in the small halo before the merger happens, for two
different values of $r_\text{s}$. The red circles label the particles that will
subsequently be ejected. They are distributed almost as the other particles:
in the sense that at any given radius the velocity distribution
of the ejected particles is similar to that of the particles
which remain bound. Furthermore, the ejected particles
are neither the most energetic nor those furthest from the centre
of the structure. They are not, on the other hand, not the particles
orbiting at the smallest radii, and hence not the most bound particles.
Further plotting other quantities such as the angular
momentum (modulus and direction) we have not identified
any particular feature that characterises the particles in question.
The decisive factor thus seems to be whether or not a particle
is falling early or late into the combined potential of the cores
of the halos.
\subsubsection{Analysis of particle energies}
The particles which are ejected
are those of which the energy is positive after the merger.
Figure \ref{fig:example} shows the evolution of
the energy of a few randomly selected particles. In
the case of the big structure, its particles' energies
stay roughly constant in a range of values well
represented by the total energy of the halo
averaged over the number of its particles.
None of the particles from the big halo are ejected during the merger.
Among the
chosen particles belonging to the small halo, there are a couple
for which the total energy becomes positive after the merger,
roughly in a time window between $t=60$ and $t=70$.
None of the particles are ejected before $t=50$.
The mean particle energy of the small halo displays a sharp peak
at the moment of the merger and then falls back to a roughly
constant value, which is higher than the initial one.
Thus, during a collision in this simulation $11\%$ of the
particles from the small halo are freed from the
system during the merger, and will never return.
\begin{figure}
\includegraphics[scale=0.8]{example.pdf}
\caption{Time variation of the total energy. On the left we see the
time evolution of five different particles, and on the right the
average over all the particles. The upper plots are for the big halo,
the lower ones for the small halo. The red dotted line shows the zero
point of the energy.}
\label{fig:example}
\end{figure}
\subsubsection{Mechanism of ejection}
\label{sec:ejection}
The observation that it is the particles coming in later which
are ejected is similar to the case of cold uniform spherical collapse
in \citep{2009MNRAS.397..775J,sylos}, where the ejected particles were
found to be those starting out in the outer shells. In this case
close analysis of the particle energies shows that those which
escape pick up the energy leading to their ejection when they
pass through the potential generated by the bulk of the
mass which has already turned around and starting
re-expanding: as the time derivative of the potential
is then positive, they gain energy.
A detailed view of the time evolution of particle energies here
reveals that what is happening in the minor merger is very
similar. In figure \ref{fig:rdm} we zoom in on the time steps during
the merging of the small structure. We follow again the evolution of
five particles, where the green and red (dashed lines) are ejected
particles, whereas the blue and purple (dotted lines) are finally
bounded. Further we show the energy of the most bound
particle from the small structure (stars), and the same from the big
structure (triangles). Inspecting the time steps between 62 and
67, we see that the potential of both the small and the large
structures get deeper. This is natural because during the first
passage when the 2 structures overlap for the first time, the
potentials deepen. However, during the time-steps from 67 to
70 the potential falls back to a smaller absolute value. This is
just after the first passage of the small structure. After this point, the
potential flattens out close to the final value.
Now, comparing the time of passage of the particles which will
remain bounded (black and purple dots), we see that they pass the center
during the deepening of the potential of the structure. The
particles which are ejected, on the other hand, arrive later,
during the phase when the potential happens to be
weakening (i.e. increasing towards less negative values).
As the sign of the time derivative of the mean field
potential is positive the energy of the particles increase,
since the time variation in particle energy
along a trajectory is equal to the time derivative of
potential energy~\citep{1967MNRAS.136..101L}, $dE/dt = \partial \Phi/ \partial t$.
While we have shown here the time evolution of only two ejected particles,
we underline that figure 1 shows that of the roughly 1000
ejected particles, essentially all of which arrive late during the
merger as we have described. We note that this effect is analogous
to the so-called late-time integrated Sachs-Wolfe effect, in which photons
gain energy because of the decay of the potential they are
traversing.
We note that temporal variation of the mean field potential is
indeed well known to induce changes in the energy of individual particles,
potentially even leading to evaporation of particles, e.g. discussed as
tidal forces \citep{1958ApJ...127...17S}, and as gravitational shocks during merging
\citep{1972ApJ...176L..51O}. Very large changes in energy are
also known to occur during impulsive tidal shocking
\citep{2007ApJ...658..731V} which
exactly happens during pericenter passages.
Whereas these previous studies have considered average changes in
energy, we emphasize here that , like in the case of spherical cold collapse
\citep{2009MNRAS.397..775J}, the particles which
pick up enough energy to be ejected are those specific ones which
arrive late relative to the time at which the total potential reaches
its minimum. These particles are thus ejected because of a coincidence
between their individual orbits combined with the potential changes
during merging.
\begin{figure}
\includegraphics[scale=0.8]{rdm_parts.pdf}
\caption{Zoom in on the time variation of potential and total energy. The particles
which will remain bound (blue and purple dotted lines) pass through
the central region while the potential due to both the small and large
structures (symbols) are deepening (around time step 66). This contrasts with
the particles which will be ejected (red and green dashed lines), which arrive
a little later (time step 67) and pass the centre when the potential
is decaying.}
\label{fig:rdm}
\end{figure}
\subsubsection{What determines the fraction of ejected particles?}
After having considered the phenomenon of ejected particles in mergers,
we proceed by performing simulations with different sizes of the
minor halo. We run different simulations keeping the scale radius of
the big halo set to $r_\text{s} = 1$, while changing the small halo $r_\text{s}$
within the range $0.1 \leq r_{s2} \leq 0.7$. The ratio of the halo
masses is unchanged (equal to $0.1$).
In figure \ref{fig:fpt} we plot the fraction of the minor halo
particles that get ejected for the different values of the minor halo's
scale radius. We see that the number of ejected particles grows with
the dynamical time up to a peak that corresponds roughly to a small
structure with a dynamical time $t_{\text{dyn}}(\text{small}) \simeq 0.7 t_{\text{dyn}}(\text{big}) $,
and then beyond this point the number decreases monotonically again.
\begin{figure}
\includegraphics[scale=0.8]{t_fp.pdf}
\caption{The fraction of particles initially in the small structure which
are ejected , $f^p$, as a function of the ratio between the dynamical times
of the small structure and the big structure.}
\label{fig:fpt}
\end{figure}
This behaviour is expected from the considerations above. Basically we
need to compare two timescales, namely the time it takes the small
structure to cross the big structure, and the time for a
typical orbit in the small structure. The first timescale corresponds
to the crossing time for the big structure, which is proportional to
our definition of the dynamical time, $\tau_\text{dyn} = 4r_\text{s}
/v_c(4r_\text{s})$. The second timescale is similar, but defined for the
small structure.
Thus, for a very compact small structure, the particles in the small
structure will make many orbits while crossing the big structure, and
fewer receive sufficient increase in energy to leave the structure. On
the other hand, for a very dilute small structure, the particles in
the small structure will perform much less than one orbit while
crossing the big structure, rendering the motion almost adiabatic.
An extrapolation of this finding is that smooth accretion should not
lead to particle ejection.
\section{Tracing particles in an analytical potential}
\label{sec:MFE}
In order to demonstrate that the ejection of particles during mergers
is a mean field effect, we implement a toy-model in which we consider
one particle moving in an analytical time-dependent potential
approximating that of the merging structures. In this
way we eliminate dynamical friction entirely. We also eliminate any two-body effects
which may be a possible spurious source of particle ejection.
\subsection{Simulation Set-Up}
\label{sub:setup2}
We simulate a merger with the same characteristics as the one
described in section \ref{sub:example}, but now
using a smooth potential instead of the $N$-body approach used earlier.
We then follow the orbits in this potential of a test particle which
initially is bound to the small structure\footnote{This approach is similar to that of \citep{teyssier}, but differs
in that the latter considered the sum of particles on large orbits
(wandering stars) and truly ejected particles. Since we here
consider separately the ejected particles (and don't consider the
particles merely on large orbits), a direct comparison is not
possible.}.
For both structures we use
analytical potentials for the same Hernquist model
\citep{1990ApJ...356..359H}
\begin{equation}
\Phi(r) = - \frac{GM/r_\text{s}}{1 + r/r_\text{s}} \, ,
\end{equation}
where the large structure has a scale radius $r_{s1} = 1$ and mass
$M_1=1$, and the smaller one has $r_{s2}= 0.3$ and $M_2=0.1$.
The two potentials are initially at rest and centered at $15 r_{s1}$
apart along the $x$-axis, which is the turnaround radius of the big
structure, just as discussed in section \ref{sub:setup1}. For each time
step the potential
in which the test particles move is then evolved as follows.
The large structure is kept fixed at all times, while the centre
of the small structure follows the free fall trajectory induced by
the large large structure until the time when
their centres coincide. For the subsequent evolution
we consider four extreme cases for the interaction between
the two potentials, representing extremes between which the
full $N$-body simulation would most likely be.
In the first set of simulations (Case 1) the small structure stops
instantaneously when its centre overlaps the centre of the bigger
structure; in the second set (Case 2) we let the small structure
continue its motion in the $x$ direction, following still its
free fall trajectory due to the gravity of the large structure.
In the third and fourth cases, we do as for the first
and second cases respectively, but now include a small
impact parameter in order to model non-head-on collisions.
We achieve this by shifting the velocity direction by hand
when the small halo is at $x = 5$.
In each simulation we follow the motion of one particle
initially bound to the small potential. We choose its initial
position and velocity similarly to how this was done in setting
up the $N$-body simulations, with its radius sampled from
the Hernquist mass profile. We assign an initial velocity by determining the
typical speed at the radius just chosen, using an analytical
expression for the radial component of the velocity dispersion. We
then let the system evolve. For each time step the gravitational
attraction of the big structure on the small one, and of both
structures on the particle, are calculated using a leapfrog
integrator.
\subsection{Particle trajectories}
\label{sub:results}
\begin{figure}[p!]
\begin{minipage}[b]{0.47\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{case1.pdf}
\caption{ Positions of the centre of the two potentials, and
orbits of two particles projected in the spatial coordinates $(x,y)$.
Case 1 is a head-on collision where the small potential
stops instantaneously as it reaches the big potential, which stays fixed at $x =
15$ and $y=0$ (magenta square). Black dots are equal time
intervals of the positions in $x$ and $y$ of the centre of the
small potential. The blue orbit shows a particle that happens to
be bound in the final resulting potential, and the red orbit
shows a particle that gets ejected from the system.}
\label{fig:orbits1}
\end{minipage}
\hfill
\begin{minipage}[b]{0.47\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{case2.pdf}
\caption{Case 2 is a head-on collision where the small potential
continues moving after passing the centre of the big structure.
Black dots are equal time intervals of the positions in $x$ and $y$
of the centre of the small potential. The blue orbit shows a
particle that happens to be caught in the final resulting potential,
and the red orbit shows a particle that gets ejected from the
system. The green orbit shows a particle that gets bounded to the
larger potential after the merger.\vspace*{0.4cm}}
\label{fig:orbits2}
\end{minipage}
\end{figure}
\begin{figure}[p!]
\begin{minipage}[b]{0.47\linewidth}
\includegraphics[width=0.95\linewidth]{case3.pdf}
\caption{ Case 3 is a collision with non-zero impact parameter, where
the small potential stops as it reaches the big potential, which
stays fixed at $x = 15$ and $y=0$ (magenta square). Black dots are
equal time intervals of the positions in $x$ and $y$ of the centre
of the small potential. The blue orbit shows a particle that
happens to be caught in the final resulting potential, and the red
orbit shows a particle that gets ejected from the system.\vspace*{0.5cm}}
\label{fig:orbits3}
\end{minipage}
\hfill
\begin{minipage}[b]{0.47\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{case4.pdf}
\caption{ Case 4 is a collision with non-zero impact parameter, where
the small potential continues moving after passing the centre of the
big structure. Black dots are equal time intervals of the positions
in $x$ and $y$ of the centre of the small potential. The blue orbit
shows a particle that happens to be caught in the final resulting
potential, and the red orbit shows a particle that gets ejected from
the system. The green orbit shows a particle that gets bounded to
the major potential after the merger.}
\label{fig:orbits4}
\end{minipage}
\end{figure}
In figure \ref{fig:orbits1} we plot the orbits of two different
particles, hence two different simulations, belonging to the first set
of simulations,
in which the potential of the small structure stops (forcing its velocity to zero) when its
centre reaches that of the large structure.
The red particle is ejected from the system, the blue one
remains bounded in the resulting structure and starts orbiting around
it.
In figure
\ref{fig:orbits2}-\ref{fig:orbits4}
we present similar plots for the other three cases, picking for
each case a particle that
remains bounded and another that gets ejected.
To get statistics on the particle ejection, we have run $1000$ simulations
for each case, choosing the test particle randomly as described above.
To check for ejection we control the
sign and constancy of the total energy of the particle. In the second
and in the fourth cases, when the small potential keeps moving in free fall after passing through the big one, it can
happen that the particle neither follows the small potential nor
gets ejected, but it gets trapped by the bigger potential. If this is
the case, the kinetic energy of the particle has to be calculated with
respect to the potential of the large structure. We show examples in figures
\ref{fig:orbits2} and \ref{fig:orbits4} (in green) of
orbits of particles that get bound to the big potential after the
collision.
\clearpage
\subsection{Fraction of ejected particles}
We summarise the results of the $4000$ simulations in table
\ref{tab:stat}. For the simulation with the same
parameters calculated with the $N$-body code and described in section
\ref{sub:example}, the percentage of the ejected particles is $11\%$,
which lies between the two values of cases $1$ and $2$.
We notice that in the case of non-head-on collisions the fraction
of the mass ejected increases slightly.
\begin{table}
\centering
\caption{\label{tab:stat} Statistics for particle ejection. $1000$ simulations were run for each case. In cases 1 and 3, after the potentials of the small and large structures overlap, the simulation continues only for the test particle, leaving the potentials fixed. In cases 2 and 4 the small potential continues through the large potential, and hence the particle may end up in either of the two, or be ejected.}
\begin{tabular}{ lccc}
& Ejected & Small& Big\\
& & halo & halo\\
\hline
{\bf Case 1} & & & \\
Halo stops. & $14 \%$&$86 \%$ & \\
Head-on. & & & \\
& & & \\
{\bf Case 2}& & & \\
Halo doesn't stop. & $5.3\%$&$82.4 \%$ &$12.3\%$ \\
Head-on. & & & \\
& & & \\
{\bf Case 3} & & & \\
Halo stops. & $17 \%$&$83\%$ & \\
Impact parameter. & & & \\
& & & \\
{\bf Case 4}& & & \\
Halo doesn't stop.& $7.2\%$&$76.7\%$ &$16.1\%$ \\
Impact parameter. & & & \\
\end{tabular}
\end{table}
\section{The relation between scale radius and fraction of ejected particles}
We will now study how the fraction of ejected particles from the small structure depends on the size of the minor halo in the merger, and compare the dependence in the $N$-body simulations with the dependence in simulations, where individual particles are tracked in analytic potentials.
Figure \ref{fig:fp_vs_rs} shows the ejected fraction for the head-on mergers for the $N$-body simulations (from section~\ref{sub:nbody}) and the analytical models, where we track particles in an analytic potential (from section~\ref{sec:MFE}).
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{10000_11.pdf}
\caption{The fraction of ejected particles for the head-on mergers for the $N$-body simulations (blue), the analytical potential model where the minor halo is stopped when reaching the big one (red dotted), and analytical model where the minor halo continues with free fall velocity through the major halo (black dashed).}
\label{fig:fp_vs_rs}
\end{figure}
The first thing we notice is that the fraction of ejected particles in the $N$-body simulations falls within the range spanned by the two analytical models. This result is independent of the scale radius of the minor halo. It is not a surprising result since the way the minor halos evolves in the $N$-body simulation falls somewhere in between the extreme cases studied in the analytical models, where the minor halo potential is either totally unaffected by the major halo, or it is stopped instantly when it reaches the center of the major halo.
Another result is that the $N$-body simulation and both analytical models all show a decline in the fraction of ejected particles when the minor halos scale radius is larger than $r_\text{s}\simeq 0.35$ (i.e. 35\% of the major halos scale radius). The similarities (the decline in fraction of ejected particles for $r_\text{s}>0.35$) and differences (i.e. the different overall normalization in the different models) between the analytical models and the $N$-body simulations suggest several conclusions
\begin{itemize}
\item The detailed destruction history of the minor halo is important for the overall normalisation of the fraction of ejected particles.
\item In the setups studied in this work, there is a decline in fraction of ejected particles when the minor halos scale radius reaches 35\% of the scale radius of the major halo. This results holds when the mass ratio of the two halos are fixed at 1:10, and will likely change for different mass ratios.
\item In the models, where particles are tracked in an analytical potential, dynamical friction between the two merging halos is not present. Because of the similarities between these models and the $N$-body simulations, we conclude that it is possible to eject particles without dynamical friction. It is still possible that dynamical friction can affect the exact number of ejected particles, but we have shown that it is not the main driver of ejection in minor mergers.
\end{itemize}
\section{Discussion and Conclusion}
We have shown that during minor mergers approximately $5-15\%$ of the particles from the minor halo are ejected, making the phenomenon quantitatively important in structure formation scenarios. In analytical models (with dynamical friction turned off) and $N$-body simulations, there exists similar relations between the fraction of ejected particles as a function of the ratio between the scale radii of the two halos in the merger. This similarity shows that dynamical friction is not the driving mechanisms for the ejection of particles.
Instead we find that the relevant mechanism is the increase in the total energy of individual particles arising from the time-dependence of the mean field potential during the merger process. The ejected particles are those that travel in a deep potential as they move in towards the plane of the center of the main halo, and a shallower potential as they move out. In a minor merger these are the particles which originate in the small halo and cross the plane of the center of the large halo slightly later than the center of the halo they originally are bound to does.
In this study we have only considered collisionless
systems. It is unclear how the presence of baryons will affect the
number of ejected particles. It has been proposed that the potential
variations due to supernova feedback can turn galaxy cusps into
cores \citep{Pontzen2013,Amorisco2013}, and it is
likely that these potential changes can also affect the number of
particles ejected during a merger. It is expected that the merging
of two galaxies leads to a significant increase in the star
formation rate \citep{Hayward2013}, so an increase in the potential
variation due to feedback processes is expected during a
merger. The exact role of feedback processes will of course depend
on the actual feedback model used in the simulation (many different
feedback models exist, e.g. \citep{Agertz2013,Vogelsberger2013,Stinson2013,Marinacci2014}).
Our finding provides an explanation for the origin of high-velocity
component of dark matter particles observed in cosmological $N$-body
simulations. This component of high-velocity particles is important
since it potentially may give a clear signature in underground dark
matter detectors~\citep{2012arXiv1208.0334B}. This is because
particles ejected throughout the merging history of the
Universe should permeate space, and also be present in the
Earth's neighbourhood, at energies significantly higher than the
equilibrated dark matter component. Thus, even though the ejected particles
only contribute around $3\%$ of the dark matter near Earth, then they could
still induce a peaked signal on top of the broad bump from
the thermalized dark matter component.
\acknowledgments
The Dark Cosmology Centre is funded by the Danish National Research Foundation.
MJ thanks the Dark Cosmology Centre for financial support as a visitor at the
centre.
\defAJ{AJ}
\defARA\&A{ARA\&A}
\defApJ{ApJ}
\defApJ{ApJ}
\defApJS{ApJS}
\defAp\&SS{Ap\&SS}
\defA\&A{A\&A}
\defA\&A~Rev.{A\&A~Rev.}
\defA\&AS{A\&AS}
\defMNRAS{MNRAS}
\defNew Astronomy{New Astronomy}
\defJCAP{JCAP}
\defNature{Nature}
\defPASP{PASP}
\defAstrophys.~Lett.{Astrophys.~Lett.}
\defPhysical Reviews{Physical Reviews}
\bibliographystyle{JHEP}
\providecommand{\href}[2]{#2}\begingroup\raggedright |
\section{Introduction}
The theory of nonlinear electrodynamics was first introduced in
$1934$ by Born and Infeld for the purpose of solving various
problems of divergence appearing in the Maxwell theory \cite{BI}.
In recent years, the study of nonlinear electrodynamics has got a
new impetus. Strong motivation comes from developments in
string/M-theory, which is a promising approach to quantum gravity
\cite{BIString}. It has been shown that the Born-Infeld (BI)
theory naturally arises in the low energy limit of the open string
theory \cite{Frad,Cal}. Another motivation originates from the
fact that most physical systems in the nature, including the field
equations of the gravitational systems, are intrinsically
nonlinear. The nonlinear BI electromagnetic theory was designed to
regulate the self-energy of a point-like charge \cite{BI}. Various
aspects of black hole solutions coupled to nonlinear BI gauge
field has been studied. Exact solutions of the Einstein BI
theory with or without cosmological constant have been constructed in \cit
{Fern,Tamaki,Dey,Cai1,MHD}. In the scalar-tensor theories of gravity, black
object solutions coupled to a Born-Infeld nonlinear electrodynamics have
also been studied widely in the literature \cite{Tam,SheyBI}.
However, BI theory is not the only nonlinear electrodynamics theory which
can remove the divergence of the electric field at $r = 0$. In particular,
in recent years, other types of nonlinear electrodynamics in the context of
gravitational field have been introduced. In \cite{Soleng} exact solution
for a static spherically symmetric field outside a charged point particle is
found in a nonlinear $U(1)$ gauge theory with a logarithmic Lagrangian.
While this particular theory appears to have no direct relation to
superstring theory, it serves as a toy model illustrating that certain
nonlinear field theories can produce particle-like solutions which can
realize the limiting curvature hypothesis also for gauge fields \cite{Soleng
. In addition to BI and logarithmic types for nonlinear gauge
fields, very recently one of the present authors proposed an
exponential form of nonlinear electromagnetic Lagrangian
\cite{HendiJHEP}. Although a logarithmic form of the
electrodynamic Lagrangian, like BI electrodynamics, removes
divergences in the electric field, the exponential form of
nonlinear electromagnetic Lagrangian does not cancel the
divergency of the electric field at $r = 0$, however, its
singularity is much weaker than in the Einstein-Maxwell theory.
Other studies on the gravitational systems coupled to nonlinear
electrodynamics gauge fields have been carried out in \cit
{PMIpaper,Oliveira,Soleng2}.
The extension of the Maxwell field to the nonlinear
electromagnetic gauge field provides powerful tools for
investigation of black object solutions. In the present work, we
would like to turn the investigations on the nonlinear
electrodynamics to the rotating black string solutions with one
rotation parameter in the anti-de Sitter (AdS) background. We will
consider the four dimensional action of Einstein gravity with two
kinds of BI type nonlinear electromagnetic gauge fields.
The structure of this paper is as follows. In the next section we
present the basic field equations as well as the Lagrangian of two
types of nonlinear electrodynamic fields. We will also solve the
equations for the rotating black string spacetime and study the
physical properties of the solutions. In Sec. III, we calculate
the conserved and thermodynamic quantities of the black string
solutions and verify the first law of thermodynamics. We finish
our paper with closing remarks in the last section.
\section{Basic equations and solutions}
We consider a model of a gravitating electromagnetic field in the presence
of cosmological constant. The Lagrangian for this system is chosen a
\begin{equation}
\mathcal{L}=R-2\Lambda +L(\mathcal{F}), \label{Lagrangian}
\end{equation
where ${R}$ is the Ricci scalar, $\Lambda $ refers to the cosmological
constant and $L(\mathcal{F})$ is a general Lagrangian of electromagnetic
field in which $\mathcal{F}=F_{\mu \nu }F^{\mu \nu }$ is Maxwell invariant,
F_{\mu \nu }=\partial _{\mu }A_{\nu }-\partial _{\nu }A_{\mu }$ and $A_{\mu
} $ is the gauge potential. Here, we assume a kind of rotating metric whose
t$ =constant and $r$ =constant boundary has the topology $R\times S^{1}$
\cite{Lem}
\begin{equation}
ds^{2}=-f(r)\left( \Xi dt-ad\phi \right) ^{2}+\frac{r^{2}}{l^{4}}\left(
adt-\Xi l^{2}d\phi \right) ^{2}+\frac{dr^{2}}{f(r)}+\frac{r^{2}}{l^{2}}
dz^{2}, \label{Metric}
\end{equation}
where the functions $f(r)$ should be determined from gravitational field,
\Xi =\sqrt{1+a^{2}/l^{2}}$, $a$ is the rotation parameter, $0\leq \phi <2\pi
$ and $-\infty <z<\infty $.
In order to investigate the properties of the electromagnetic
field, one may consider a suitable $L(\mathcal{F})$ and solve the
Maxwell-like field equation. In this paper, we consider two new
classes of nonlinear electromagnetic fields, namely exponential
form of nonlinear electromagnetic (ENE) Lagrangian and logarithmic
form of nonlinear electromagnetic (LNE) Lagrangian, whose
Lagrangians are
\begin{equation}
L(\mathcal{F})=\left\{
\begin{array}{ll}
\beta ^{2}\left( \exp (-\frac{\mathcal{F}}{\beta ^{2}})-1\right) , & \text
ENE} \\
-8\beta ^{2}\ln \left( 1+\frac{\mathcal{F}}{8\beta ^{2}}\right) , & \text{LN
\end{array
\right. , \label{L(F)}
\end{equation
where $\beta $ is called the nonlinearity parameter. In the limit $\beta
\rightarrow \infty $, the mentioned $L(\mathcal{F})$'s reduce to the
Lagrangian of the standard Maxwell field
\begin{equation}
\left. L({\mathcal{F}})\right\vert _{\text{large }\beta }=-{\mathcal{F}}+O(
\mathcal{F}}^{2}) \label{LBIexpand}
\end{equation}
Considering a strong electromagnetic field in regions near to
point-like charges, Dirac suggested that one may have to use
generalized nonlinear Maxwell theory in those regions
\cite{Dirac64}. Similar behavior may have occurred in the vicinity
of neutron stars and black objects and so it is expected to
consider nonlinear electromagnetic fields with an astrophysical
motive \cite{AstroNON}. In addition, within the framework of
quantum electrodynamics, it was shown that quantum corrections
lead to nonlinear properties of vacuum which affect the photon
propagation \cite{H-E,Schwinger,Stehle,Delphenich}.
Although in the context of nonlinear electrodynamics, BI theory is
a specific model, the recent interest in the nonlinear
electrodynamics theories is mainly due to their emergence in the
context of the low-energy limit of heterotic string theory, where
a quartic correction of Maxwell field strength appear \cite{Kats}.
In other words, it was shown that all order loop corrections to
gravity may be added up as a Born-Infeld type Lagrangian
\cite{BIString,Frad,Fradkin}. Any nonlinear electrodynamics that
satisfies the weak field limit (\ref{LBIexpand}) is said to be of
the Born-Infeld type \cite{BItype}.
For completeness, we should note that working in the context of
AdS/CFT correspondence, it is worth investigating the effects of
nonlinear electrodynamic fields on the dynamics of the strongly
coupled dual theory \cite{CaiSun}.
Motivated by the recent results mentioned above, we consider the
mentioned Born-Infeld type theory and investigate their
properties.
The equation of motion for the gauge field can be written as
\begin{equation}
\partial _{\mu }\left( \sqrt{-g}L_{\mathcal{F}}F^{\mu \nu }\right) =0,
\label{MaxEq}
\end{equation
where $L_{\mathcal{F}}=\frac{dL(\mathcal{F})}{d\mathcal{F}}$.
Considering Eq. (\ref{Metric}) with Eq. (\ref{MaxEq}), we find
that the consistent gauge potential is
\begin{equation}
A_{\mu }=h(r)\left( \Xi \delta _{\mu }^{0}-a\delta _{\mu }^{\phi }\right) ,
\label{Amu}
\end{equation
in which the radial function $h(r)$ can be written as
\begin{equation}
h(r)=\left\{
\begin{array}{ll}
\frac{-\beta r\sqrt{L_{W}}}{2}\left[ 1+\frac{L_{W}}{5}\digamma {\left( \left[
1\right] ,\left[ \frac{9}{4}\right] ,\frac{L_{W}}{4}\right) }\right] , &
\text{ENE}\vspace{0.1cm} \\
\frac{-2q}{3r}\left[ 2\digamma {\left( \left[ \frac{1}{4},\frac{1}{2}\right]
,\left[ \frac{5}{4}\right] ,1-\Gamma ^{2}\right) -}\frac{1}{\left( 1+\Gamma
\right) }\right] , & \text{LNE
\end{array
\right. , \label{h(r)}
\end{equation
where $q$ is an integration constant which is related to the electric charge
of the black string, $L_{W}=LambertW\left( \frac{4q^{2}}{\beta ^{2}r^{4}
\right) $ which satisfies $LambertW(x)\exp \left[ LambertW(x)\right] =x$,
\digamma ([a],[b],z)$ is hypergeometric function and $\Gamma =\sqrt{1+\frac
q^{2}}{r^{4}\beta ^{2}}}$ (for more details, see \cite{Lambert}).
Using Eq. (\ref{Amu}), we find that the only nonzero components of
the electromagnetic field tensor are $F_{tr}$ and $F_{\phi r}$,
\begin{equation}
F_{\phi r}=-\frac{a}{\Xi }F_{tr},\;{~\;~}F_{tr}=\frac{\Xi q}{r^{2}}\times
\left\{
\begin{array}{ll}
e^{-\frac{L_{W}}{2}}, & \text{ENE} \\
\frac{2}{\Gamma +1}, & \text{LNE
\end{array
\right. . \label{El}
\end{equation
Expanding $F_{tr}$ for large value of $r$, we arrive a
\begin{equation}
F_{tr}=\frac{\Xi q}{r^{2}}+-\frac{\xi \Xi q^{3}}{4\beta ^{2}r^{6}}+O\left(
\frac{1}{r^{10}}\right) , \label{Ftrexpand}
\end{equation}
where $\xi =1,8$ for LNE and ENE branches, respectively. We find
that for large distances, the first term in Eqs. (\ref{Ftrexpand})
dominates and the electric field of Maxwell theory is recovered.
We have plotted $F_{tr}$ as a function of $r$ in Fig. \ref{E}.
From this figure it can be seen that for
large values of $r$ the electric fields vanish as one expected. Besides, as
r\rightarrow \infty $, both ENE and LNE behave like the linear
Maxwell field. This implies that the nonlinearity of these fields
make sense only near the origin. It is worth mentioning that, in
contrast to the standard Maxwell and ENE fields, the electric
field of LNE is finite at the origin. We should note that the
divergency of the electric field of ENE is much slower than the
divergency of the standard Maxwell field at $r=0$.
\begin{figure}[tbp]
\begin{array}{cc}
\epsfxsize=8cm \epsffile{FigFtr.eps} &
\end{array}
\caption{The behavior of $F_{tr}$ versus $r$ for $q=1$, $\protect\beta=4$
and $\Xi=1.2$. ENE (Bold line), LNE (solid line), and linear Maxwell field
(dotted line).}
\label{E}
\end{figure}
Now, we are in a position to discuss the geometric view of
spacetime. To do this, we should first obtain the metric function
$f(r)$. Gravitational field equation of Einstein--$\Lambda
$--nonlinear electromagnetic theory may be written as
\cite{HendiJHEP}
\begin{equation}
R_{\mu \nu }-\frac{1}{2}g_{\mu \nu }\left( R-2\Lambda \right) =\frac{1}{2
g_{\mu \nu }L(\mathcal{F})-2L_{\mathcal{F}}F_{\mu \lambda }F_{\nu
}^{\;\lambda }. \label{GravEq}
\end{equation
In order to obtain $f(r)$, we should consider the field Eq. (\re
{GravEq}) with Eqs. (\ref{Metric}), (\ref{L(F)}) and (\ref{El}). It is easy
to show that the $rr$ component of Eq. (\ref{GravEq}) can be written in the
following for
\begin{equation}
rf^{\prime }(r)+f(r)+\Lambda r^{2}+\beta ^{2}r^{2}g(r)=0, \label{rrEq}
\end{equation
where
\begin{equation}
g(r)=\left\{
\begin{array}{ll}
\frac{\exp (8\Psi )}{2}\left[ 1-\left( 1-16\Psi \right) \right] , & \text{EN
}\vspace{0.2cm} \\
4\ln \left( 1-\Psi \right) +\frac{8\Psi }{1-\Psi }, & \text{LNE
\end{array
\right. , \label{g(r)Eq}
\end{equation
and $\Psi =\left( \frac{F_{tr}}{2\Xi \beta }\right) ^{2}$. After some
cumbersome calculations, the solutions of Eq. (\ref{rrEq}) can be obtained a
\begin{equation}
f(r)=\frac{2m}{r}-\frac{\Lambda r^{2}}{3}+\beta W(r), \label{f(r)}
\end{equation
wit
\[
W(r)=\left\{
\begin{array}{ll}
\frac{q}{3\sqrt{L_{W}}}\left[ 1+L_{W}+\frac{4}{5}L_{W}^{2}\digamma {\left(
[1],[\frac{9}{4}],\frac{L_{W}}{4}\right) }\right] , & \text{ENE}\vspace{0.2c
} \\
\frac{8q^{2}\digamma {\left( [\frac{1}{2},\frac{1}{4}],[\frac{5}{4
],1-\Gamma ^{2}\right) }}{3\beta r^{2}}-\frac{4\beta r^{2}\left( \Gamma -\ln
[\frac{(\Gamma ^{2}-1)}{2e^{-7/3}}]\right) }{3}-\frac{4\beta \int r^{2}\ln
\left( \Gamma -1\right) dr}{r}, & \text{LNE
\end{array
\right. ,
\
where $m$ is the integration constant which is the total mass of spacetime,
\digamma ([a],[b],z)$ is hypergeometric function and the last term of LNE
can be calculated a
\begin{eqnarray}
\int r^{2}\ln \left( \Gamma -1\right) dr &=&-\frac{q^{3/2}(\Gamma -1)^{1/4}}
2^{3/4}\beta ^{3/2}}\left[ \frac{14}{3}\digamma {\left( \left[ \frac{1}{4}
\frac{1}{4},\frac{11}{4}\right] ,\left[ \frac{5}{4},\frac{5}{4}\right]
\frac{1-\Gamma }{2}\right) }-\right. \nonumber \\
&&\left. \frac{14}{25}(\Gamma -1)\digamma {\left( \left[ \frac{5}{4},\frac{
}{4},\frac{11}{4}\right] ,\left[ \frac{9}{4},\frac{9}{4}\right] ,\frac
1-\Gamma }{2}\right) -}\right. \nonumber \\
&&\left. \frac{\left[ 4+3\ln (\Gamma -1)\right] }{9(\Gamma -1)}\digamma
\left( \left[ \frac{-3}{4},\frac{7}{4}\right] ,\left[ \frac{1}{4}\right]
\frac{1-\Gamma }{2}\right) }+\right. \nonumber \\
&&\left. \left[ -4+\ln (\Gamma -1)\right] \digamma {\left( \left[ \frac{1}{4
,\frac{7}{4}\right] ,\left[ \frac{5}{4}\right] ,\frac{1-\Gamma }{2}\right)
\right] . \label{integral}
\end{eqnarray}
We should note that obtained solutions given by Eq. (\ref{f(r)})
satisfy all the components of the field equations (\ref{GravEq}).
\subsection*{Properties of the solutions}
Here we are going to study the physical properties of the solutions as well
as the asymptotic behavior of the spacetime. Expanding the metric functions
for large $r$, we hav
\begin{equation}
f(r)=\frac{2m}{r}-\frac{\Lambda r^{2}}{3}+\frac{q^{2}}{r^{2}}-\frac{\xi q^{4
}{40r^{6}\beta ^{2}}+O\left( \frac{1}{r^{10}}\right) , \label{f(r)Expand}
\end{equation
where for $\beta \rightarrow \infty $, one can recover the rotating black
string solutions in Einstein-Maxwell gravity \cite{Lem}. Next, we calculate
the curvature scalars of this spacetime. It is easy to show that the Ricci
scalar and the Kretschmann invariant of the spacetime are
\begin{eqnarray}
R &=&-f^{\prime \prime }(r)-\frac{4f^{\prime }(r)}{r}-\frac{2f(r)}{r^{2}},
\label{R} \\
R_{\mu \nu \rho \sigma }R^{\mu \nu \rho \sigma } &=&f^{\prime \prime
2}(r)+\left( \frac{2f^{\prime }(r)}{r}\right) ^{2}+\left( \frac{2f(r)}{r^{2}
\right) ^{2}. \label{RR}
\end{eqnarray
where the prime denotes derivative with respect to $r$. Since other
curvature invariants of the spacetime such as the Ricci square are only the
functions of $f^{\prime \prime }$, $f^{\prime }/r$ and $f/r^{2}$, thus we
only consider the Ricci scalar and the Kretschmann invariant. Substituting
the metric functions (\ref{f(r)}) in (\ref{R}) and (\ref{RR}), we find
\begin{eqnarray}
\lim_{r\longrightarrow 0^{+}}R &=&\infty , \label{Rorigin} \\
\lim_{r\longrightarrow 0^{+}}R_{\mu \nu \rho \sigma }R^{\mu \nu \rho \sigma
} &=&\infty . \label{RRorigin}
\end{eqnarray
This indicates that we have an essential singularity located at $r=0$. On
the other side, we can expand Eqs. (\ref{R}) and (\ref{RR}) for large $r$
and keep the first order nonlinear correction ter
\begin{eqnarray}
\lim_{r\longrightarrow \infty }R &=&4\Lambda +\frac{\xi Q^{4}}{2\beta
^{2}r^{8}}+O\left( \frac{1}{r^{12}}\right) , \label{Rinf} \\
\lim_{r\longrightarrow \infty }R_{\mu \nu \rho \sigma }R^{\mu \nu \rho
\sigma } &=&\frac{8}{3}\Lambda ^{2}+\frac{48m^{2}}{r^{6}}-\frac{96mQ^{2}}
r^{7}}+\frac{56Q^{4}}{r^{8}}-\frac{2\xi Q^{4}}{\beta ^{2}l^{2}r^{8}}+O\left(
\frac{1}{r^{11}}\right) , \label{RRinf}
\end{eqnarray
where we conclude that the spacetime is asymptotically anti-de Sitter.
\begin{figure}[tbp]
\begin{array}{cc}
\epsfxsize=7cm \epsffile{Figexp.eps} & \epsfxsize=7cm \epsffile{Figlog.eps
\end{array}
\caption{$f(r)$ versus $r$ for $\Lambda=-1$, $q=1$, $m=1$, and $\protec
\beta > \protect\beta_{c}$ (solid line) and $\protect\beta <\protect\bet
_{c} $ (dashed line) in ENE branch (left figure) and LNE branch (right
figure). }
\label{Figf}
\end{figure}
Furthermore, such as black string with linear Maxwell source, one
expects to obtain a black string with an outer and an inner
horizon, an extreme black string or a naked singularity. We should
note that for the obtained black string with nonlinear source, a
new interesting situation appears. This new situation appears for
small values of the nonlinearity parameter, in which the black
string has one non-extreme horizon as it happens for Schwarzschild
solution (uncharged solution). In other words, we find that there
is a critical nonlinearity parameter $\beta _{c}$ in which for
$\beta <\beta _{c}$ the metric function may be negative near the
origin. This means that the singularity is timelike for $\beta
>\beta _{c}$, but for $\beta <\beta _{c}$ it is spacelike (see
Fig. \ref{Figf} for more details). Although we cannot obtain
$\beta _{c}$ analytically, we find that it is a function of other
parameters. We obtain $\beta _{c}$, numerically, from the
following
condition
\[
\lim_{r\longrightarrow 0^{+}}f(r)\longrightarrow \left\{
\begin{array}{ll}
+\infty , & \beta >\beta _{c} \\
-\infty , & \beta <\beta _{c
\end{array
\right. .
\]
For more clarifications, we give a numerical method for obtaining $\beta
_{c} $. At first we should fix metric parameters $m$, $q$ and $\Lambda $,
and we can check that for large $\beta $ the metric function is positive
near the origin. Then, we reduce $\beta $ until the sign of the metric
function switch to negative. One can use this method to obtain $\beta _{c}$
with ideal accuracy. Numerical calculations show that $\beta _{c}$ change
when we alter at least one of the metric parameters $m$, $q$ and $\Lambda$
(see table A for more details).
\begin{center}
\begin{tabular}{ccc}
\hline\hline
$f_{ENE}(r)\longrightarrow $ & $-\infty $ & ~~ $+\infty $ \\ \hline
$\beta =$ & ~~ $0.171167$ & ~~ $0.171168$ \\ \hline\hline
\end{tabular}
\\[0pt]
\vspace{0.5cm}
\begin{tabular}{ccc}
\hline\hline
$f_{LNE}(r)\longrightarrow $ & ~~ $-\infty $ & ~~ $+\infty $ \\ \hline
$\beta =$ & ~~ $0.092042$ & ~~ $0.092043$ \\ \hline\hline
\end{tabular}
\\[0pt]
\vspace{0.3cm} Table A: $f(r)$ for $m=1$, $q=1$, $\Lambda =-1$,
r\longrightarrow 0^{+}$ and \\[0pt]
$\beta _{c}\approx 0.171$ and $0.092$ for ENE and LNE branches, respectively.
\end{center}
\section{Conserved quantities and the first law of thermodynamics \label{Therm}}
At the first step, we use the definition of quasilocal energy \cit
{BY,Kraus,Mal} to compute the conserved charges of our solutions. Following
the counterterm method, the divergence-free boundary stress-tensor can be
written as
\begin{equation}
T^{ab}=\frac{\Theta ^{ab}-\left( \Theta +\frac{2}{l}\right) \gamma ^{ab}}
8\pi }, \label{Stres}
\end{equation
where the last term comes from counterterm procedure. Considering a Killing
vector field $\mathcal{\xi }$ on the boundary $\mathcal{B}$, it is known
that its quasilocal conserved quantity may be obtain from the following
relatio
\begin{equation}
Q(\mathcal{\xi )}=\int_{\mathcal{B}}d^{2}x\sqrt{\sigma }T_{ab}n^{a}\mathcal
\ \xi }^{b}, \label{TotalCharge}
\end{equation
where $\sigma $ is the determinant of the Arnowitt-Deser-Misner form of
boundary metric $\sigma _{ij}$ and $n^{a}$ is the unit normal vector on the
boundary $\mathcal{B}$. It is easy to find that boundary $\mathcal{B}$ has
two killing vector fields. They are timelike ($\partial /\partial t$) and
rotational ($\partial /\partial \varphi $) Killing vector fields, in which
their corresponding conserved charges are the quasilocal mass and angular
momentum. One can find that the mass and angular momentum per unit length of
the string when the boundary $\mathcal{B}$ goes to infinity can be
calculated a
\begin{equation}
{M}=\frac{1}{16\pi l}\left( 3\Xi ^{2}-1\right) m, \label{M}
\end{equation
\begin{equation}
{J}=\frac{3}{16\pi l}\Xi ma. \label{J}
\end{equation
As one can find the angular momentum per unit length vanishes for $a=0$ (
\Xi =1$) and therefore, $a$ is the rotational parameter,
correctly.
The second step is devoted to calculation of the thermodynamic
quantities. It was known that the universal area law of the
entropy can apply to all types of black objects in Einstein
gravity \cite{Bekenstein,hunt}. Therefore, the entropy per unit
length of the black string is
\begin{equation}
{S}=\frac{\Xi r_{+}^{2}}{4l}. \label{S}
\end{equation
In order to obtain angular velocity $\Omega $ and Hawking temperature of the
black string at the event horizon, we use the method of analytic
continuation of the metric. One can obtain the Euclidean section of the
metric by use of a transformation ($t\rightarrow i\tau $ and $a\rightarrow
ia $) and regularity at $r=r_{+}$ requires that $\tau $ and $\phi $\ should,
respectively, identify with $\tau +\beta _{+}$ and $\phi +i\Omega \beta _{+}
, where $\beta _{+}$ is the inverse of the Hawking temperature. So, it is
easy to find tha
\begin{equation}
\Omega =\frac{a}{\Xi l^{2}}, \label{Omega}
\end{equation
an
\begin{equation}
T_{+}=-\frac{\Lambda r_{+}}{4\pi }+\left\{
\begin{array}{ll}
\frac{\beta q\left( 1-L_{W_{+}}\right) }{4\pi r_{+}\sqrt{L_{W_{+}}}}-\frac
\beta ^{2}r_{+}}{8\pi }, & \text{ ENE}\vspace{0.2cm} \\
-\frac{q^{2}\left( \Gamma _{+}-2\right) }{\pi r_{+}^{3}\Gamma _{+}\left(
\Gamma _{+}-1\right) }+\frac{\beta ^{2}r_{+}\left( \ln (\frac{\Gamma
_{+}^{2}-1}{2})-\frac{2}{\Gamma _{+}}\right) }{\pi }, & \text{LNE
\end{array
\right. , \label{T}
\end{equation
where $\Gamma _{+}=\sqrt{1+\frac{q^{2}}{r_{+}^{4}\beta ^{2}}}$ and
L_{W_{+}}=LambertW\left( \frac{4q^{2}}{\beta ^{2}r_{+}^{4}}\right) $.
In order to examine the first law of thermodynamics, we should calculate the
electric charge and potential of the black string. We should use the Gauss'
law and calculate the flux of the electromagnetic field at infinity to
obtain the electric charge per unit length of black string
\begin{equation}
{Q}=\frac{\Xi q}{4\pi l}. \label{Q}
\end{equation
In addition, the electric potential $U$, measured at infinity with respect
to the event horizon $r_{+}$, is defined by \cite{Cvetic}
\begin{equation}
U=A_{\mu }\chi ^{\mu }\left\vert _{r\rightarrow \infty }-A_{\mu }\chi ^{\mu
}\right\vert _{r=r_{+}}, \label{U1}
\end{equation
where $\chi =\partial _{t}+\Omega \partial _{\phi }$ is the null generator
of the event horizon. It is easy to show tha
\begin{equation}
U=\frac{1}{\Xi }\times \left\{
\begin{array}{ll}
\frac{\beta r\sqrt{L_{W}}}{2}\left[ 1+\frac{L_{W_{+}}}{5}\digamma {\left(
\left[ 1\right] ,\left[ \frac{9}{4}\right] ,\frac{L_{W_{+}}}{4}\right)
\right] , & \text{ENE}\vspace{0.1cm} \\
\frac{2q}{3r_{+}}\left[ 2\digamma {\left( \left[ \frac{1}{4},\frac{1}{2
\right] ,\left[ \frac{5}{4}\right] ,1-\Gamma _{+}^{2}\right) -}\frac{1}
\left( 1+\Gamma _{+}\right) }\right] , & \text{LNE
\end{array
\right. . \label{U}
\end{equation}
Now, we are in a position to check the first law of black string
thermodynamics. Using the Smarr-type formula, it is straightforward to
calculate the temperature, angular velocity and electric potential in the
following manner
\begin{equation}
T=\left( \frac{\partial M}{\partial S}\right) _{J,Q},\text{ \ \ \ \ \ \
\Omega =\left( \frac{\partial M}{\partial J}\right) _{S,Q},\text{ \ \ \ \ \
\ }U=\left( \frac{\partial M}{\partial Q}\right) _{S,J}. \label{TOmegaU}
\end{equation
One can find that the quantities calculated by Eq. (\ref{TOmegaU}) coincide
with Eqs. (\ref{T}), (\ref{Omega}) and (\ref{U}). Hence we conclude that the
first law of thermodynamics is satisfied in the following form
\begin{equation}
dM=TdS+\Omega d{J}+Ud{Q}. \label{firstLaw}
\end{equation}
Since the asymptotic behavior of the solutions is the same as
linear Einstein-Maxwell black string, it is expected that the
nonlinearity does not affect on the electrical charge, mass and
angular momentum. Nevertheless, we find that the nonlinearity
affects on the other quantities in which calculated at the
horizon. Although the nonlinear electromagnetic field changes some
of the conserved and thermodynamic quantities, as we expected
\cite{1Law}, these quantities satisfy the first law of black hole
mechanics.
\section{Closing remarks}
Many physical systems in the nature have nonlinear behavior.
Einstein field equations of general relativity is also a system of
nonlinear gravitational field equations which can be applied for
describing various gravitational objects. In order to solve the
gravitational field equations in the presence of a matter field,
one can consider either the linear gauge field such as the Maxwell
electrodynamics or the nonlinear matter field such as the BI
electrodynamics. Static and stationary black object solutions of
these theories have been established and their thermodynamics have
been studied during the past decades (see
\cite{Fern,Tamaki,Dey,Cai1,MHD,EM} and references therein). The
advantages of the nonlinear electrodynamics in comparison to the
Maxwell field is that it avoids the divergences at the origin and
leads to a finite electric field on the point particles.
In this paper as a new step, we considered two types of nonlinear
electrodynamic Lagrangians as source. The first one is called the
logarithmic form and the second one named the exponential form.
Then, we constructed new four-dimensional charged rotating black
string solutions with horizon topology $R\times S^{1}$ coupled to
the nonlinear electrodynamic field. These solutions are
asymptotically anti-de Sitter. If one expand the nonlinear
electromagnetic fields for large $r$, one finds that the
asymptotic behavior of them is similar to the linear Maxwell
field. We also calculated the curvature invariants of the
spacetime and showed that there is indeed a curvature singularity
located at $r=0$. Furthermore, we found that, unlike
Einstein-Maxwell black string solutions, for small values of the
nonlinearity parameter, one can obtain a black string with a
non-extreme horizon. We also calculated the conserved quantities
of the rotating black string such as the mass and the angular
momentum as well as the thermodynamic quantities such as the
temperature and entropy associated with the horizon and checked
that the obtained conserved and thermodynamic quantities satisfy
the first law of black hole thermodynamics.
Finally, it is worthwhile to study the dynamic as well as
thermodynamic stability of the solutions, and investigate the
effects of nonlinearity parameter, $\beta$, on the stability of
the presented solutions. We leave these problems for the future
studies.
\acknowledgments{ We thank Shiraz University Research Council.
This work has been supported financially by Research Institute for
Astronomy \& Astrophysics of Maragha (RIAAM), Iran.}
|
Subsets and Splits